Datasets:

markhneedham

/

blog

like 0

2010-01-07 23:31:37

F#: Refactoring to active patterns

I've been playing around with more F# code and after realising that I'd peppered the code with if statements I thought it would be interesting to try and refactor it to make use of active patterns. The code is part of my F# solution to http://osherove.com/tdd-kata-1/[Roy Osherove's TDD Kata] and is used to parse the input string and find which delimeters are being used. This is the original code: [source,ocaml] ---- let hasCustomDelimeter (value:string) = value.Length > 2 && "//".Equals(value.Substring(0, 2)) let hasMultipleDelimeters (value:string) = hasCustomDelimeter value && "[".Equals(value.Substring(2,1)) let delimeter value = if (hasCustomDelimeter value) then value.Substring(2, value.IndexOf("\n") - 2) else "," let delimeters (value:string) = Regex.Matches(value, "\[(.)\]") |> Seq.cast |> Seq.map (fun (x:Match) -> x.Groups) |> Seq.map (fun x -> x |> Seq.cast<Group> |> Seq.nth 1) |> Seq.map (fun x -> x.Value) |> Seq.to_array let digits value = let delimeter = delimeter value if (hasMultipleDelimeters value) then value.Substring(value.IndexOf("\n")) |> split (delimeters value) |> Array.map parse else if (hasCustomDelimeter value) then value.Substring(value.IndexOf("\n")) |> split [| delimeter |] |> Array.map parse else value.Replace("\n", delimeter) |> split [|delimeter |] |> Array.map parse ---- The active pattern that we want to use is known as a 'multi case active pattern' which http://blogs.msdn.com/chrsmith/archive/2008/02/21/Introduction-to-F_2300_-Active-Patterns.aspx[Chris Smith] defines as "partitioning the entirety of the input space into different things". In this case given an input string we want to determine what type of delimeters are contained within that string. This is the code after I'd created the active pattern: [source,ocaml] ---- let (|MultipleCustomDelimeters|SingleCustomDelimeter|NoCustomDelimeter|) (value:string) = if (value.Length > 2 && "//".Equals(value.Substring(0, 2))) then if ("[".Equals(value.Substring(2,1))) then MultipleCustomDelimeters(delimeters value) else SingleCustomDelimeter(delimeter value) else NoCustomDelimeter(delimeter value) let digits value = match value with | SingleCustomDelimeter(delimeter) -> value.Substring(value.IndexOf("\n")) |> split [| delimeter |] |> Array.map parse | MultipleCustomDelimeters(delimeters) -> value.Substring(value.IndexOf("\n")) |> split delimeters |> Array.map parse | NoCustomDelimeter(delimeter) -> value.Replace("\n", delimeter) |> split [|delimeter |] |> Array.map parse ---- One thing I didn't realise is that you can set different types for the constructor values of each of the active patterns. In this case I want to match single or multiple delimeters and then return those delimeters. We can define one active pattern which matches a single delimeter and just returns that and then another one which returns an array of delimeters which is quite neat. The way it works is quite similar to http://tomasp.net/articles/fsharp-ii-functional.aspx[discriminated unions]. I found that that having all the code around parsing the input in the same function made it easier for me to understand that code and I quite like that it's possible to match the pattern and also get the delimeter/delimeters in one expression. Although there are more lines of code I think this code is more expressive and it wouldn't be too hard to add in another active pattern if there's another delimeter type that needs to be handled. I can't decide whether the 'value.SubString' and 'value.Replace' code should also be contained within the active pattern or not. At the moment I'm thinking perhaps not because it's not related to the actual delimeters.

f-refactoring-to-active-patterns

https://markhneedham.com/blog/2010/01/07/f-refactoring-to-active-patterns

2010-01-31 14:05:05

DDD8: Mixing functional and object oriented approaches to programming in C#

I did a presentation titled 'Mixing functional and object oriented approaches to programming in C#' at the http://developerdeveloperdeveloper.com/ddd8/Schedule.aspx[Developer Developer Developer conference] in Reading. The slides from the talk are below: http://www.slideshare.net/markhneedham/mixing-functional-and-object-oriented-approaches-to-programming-in-c[Mixing functional and object oriented approaches to programming in C#] I've not done many technical talks so far. My only previous attempt was a talk on http://www.markhneedham.com/blog/2009/06/30/f-what-ive-learnt-so-far/[F# one at the Sydney Alt.NET user group] last year so I'm still learning how to do this effectively. It was quite interesting reading through some of the feedback from twitter and several people pointed out that the content was too basic which was something I was concerned about while writing the talk: ____ OK. 30 mins into #ddd8 talk of functional vs OO approaches and I've been told one small thing I didn't know. ____ ____ #ddd8 finished @markhneedham talk\... Good speaker,but but basic For me, the last 20% of the content should have been 80% of the talk ____ ____ "If you want to learn more\..." Er\...yeah. I certainly want to learn more than I just did. #ddd8 ____ ____ Mark did a good job but difficulty lvl would help as it was quite basic. Missed caching bit looking at twitter. Doh! ____ ____ Mark Needham's talk on functional and OO at #DDD8 was OK, but the first 2/3 was too much of a simple intro. ____ The feedback I got in person was that the talk worked alright and there were a couple more positive comments on twitter too: ____ Great day at #ddd8 - really enjoyed Mark Needham's functional programming session & testing ASP.Net with Ruby ____ ____ Listening to @markneedham at #ddd8. He has a fantastic abilty to break things down and explain them :) ____ I think there were probably too many slides in the talk - I found myself pausing unnaturally so that I could move to the next slide which was also pointed out on twitter: ____ In my first session at #ddd8 on functional programming - loadsa slides - someone get the guy a remote! ____ I think I'll probably go with less slides the next time - some of them weren't adding much value anyway so I don't think it would hurt too much to drop them. Overall though it was a good experience for me getting the chance to put this talk together and I think I learnt a little more about better ways to understand the way we can use functional programming in C# by having the opportunity to prepare this.

ddd8-mixing-functional-and-object-oriented-approaches-to-programming-in-c

https://markhneedham.com/blog/2010/01/31/ddd8-mixing-functional-and-object-oriented-approaches-to-programming-in-c

2010-01-01 14:32:58

The Last Lecture - Randy Pausch

I recently watched Randy Pausch's 'http://www.youtube.com/watch?v=ji5_MqicxSo[Last Lecture: Achieving Your Childhood Dreams]' and read the http://www.amazon.com/gp/product/1401323251?ie=UTF8&tag=marneesblo-20&linkCode=as2&camp=1789&creative=390957&creativeASIN=1401323251[corresponding book] and although it's not directly related to software development I think that some of the points that he makes are really intriguing. These were some of the parts that particularly stood out for me: * *Introduce the elephant in the room* - whatever it is that people are really thinking about, put it out in the open. I think on development teams there is often a distrust of the way that other people write code because it's different to the way that we do. If we can get this out in the open more frequently and agree on a shared approach then it should result in a http://www.markhneedham.com/blog/2009/11/04/consistency-in-the-code-base/[more consistent code base]. * *Get the fundamentals down* - Pausch suggests that we need to understand the fundamentals because otherwise the fancy stuff isn't going to work. In Apprenticeship Patterns Ade Oshineye and Dave Hoover suggest that we should 'http://my.safaribooksonline.com/9780596806842/study_the_classics[Study the Classics]' which is a similar idea. While I think they're certainly right I'm not sure that http://www.markhneedham.com/blog/2008/02/09/learning-theory-first/[learning theory before putting it into practice] is the most effective way to learn. I think we need to do a bit of both at the same time alternating between the two so that we actually have a frame of reference when we're learning the fundamentals. * *The brick walls are there for a reason* - Pausch describes how we'll often come across obstacles stopping or blocking us from what we want to do but that we shouldn't be discouraged, they are there so that we can see how much we really want something. I think this is just generally true and not just related to software development. * *The skill of self assessment* - Pausch suggests that we need know what we don't know which I think is perhaps the best advice in the book. We need to recognise our true abilities but also have a sense of our flaws and create feedback loops to allow this to happen. My colleague Liz Douglass has come up with the idea of http://lizdouglass.wordpress.com/2009/11/16/friday-feedback-frenzies/[Feedback Frenzies] to help create this feedback loop and this is one of the best ideas I've come across for doing this. * One of my favourite quotes from the book is the following which I believe was originally from Dan Stanford: + ____ Experience is what you get when you didn't get what you wanted \... It's a phrase worth considering at every brick wall we encounter, and at every disappointment. It's also a reminder that *failure is not just acceptable, it's often essential*. ____ + I think we can do this much more in the software world. There sometimes seems to be a stigma with identifying things which fail but I think it's really useful for others to learn from mistakes that have been made. The Last Lecture is one of the best presentations I've had the chance to watch - I'd certainly recommend it.

the-last-lecture-randy-pausch

https://markhneedham.com/blog/2010/01/01/the-last-lecture-randy-pausch

2010-01-06 00:08:14

TDD: Hungarian notation for mocks/stubs

A fairly common discussion that I've had with several of my colleagues is around the way that we name the variables used for mocks and stubs in our tests. There seems to be about a 50/50 split between including 'Stub' or 'Mock' on the end of those variable names and not doing so. In a simple example test using Rhino Mocks as the testing framework this would be the contrast between the two approaches: [source,csharp] ---- [Test] public void ShouldDoSomething() { var someDependency = MockRepository.CreateMock<ISomeDependency>(); someDependency.Expect(x => x.SomeMethod()).Return("someValue"); var myController = new MyController(someDependency); myController.DoSomething() someDependency.VerifyAllExpectations(); } ---- [source,csharp] ---- [Test] public void ShouldDoSomething() { var someDependencyMock = MockRepository.CreateMock<ISomeDependency>(); someDependencyMock.Expect(x => x.SomeMethod()).Return("someValue"); var myController = new MyController(someDependencyMock); myController.DoSomething() someDependencyMock.VerifyAllExpectations(); } ---- I favour the former where we don't specify this information because I think it adds unnecessary noise to the name and is a detail about the implementation of the object behind that variable which I don't care about when I'm reading the test. I do care about it if I want to change something but if that's the case then I can easily see that it's a mock or stub by looking at the place where it's instantiated. From my experience we often tend to end up with the situation where the variable name suggests that something is a mock or stub and then it's used in a different way: [source,csharp] ---- [Test] public void ShouldDoSomething() { var someDependencyMock = MockRepository.CreateMock<ISomeDependency>(); someDependencyMock.Stub(x => x.SomeMethod()).Return("someValue"); var myController = new MyController(someDependencyMock); myController.DoSomething() } ---- That then becomes a pretty misleading test because the reader is unsure whether the name is correct and the stub call is incorrect or whether it should in fact be a stub and the name is wrong. The one time that I've seen that extra information being useful is when we have really long tests - perhaps when writing tests around legacy code which is tricky to get under test. In this situation it is very nice to be able to easily see exactly what we're doing with each of our dependencies. Hopefully this is only a temporary situation before we can work out how to write simpler tests which don't require this extra information.

tdd-hungarian-notation-for-mocksstubs

https://markhneedham.com/blog/2010/01/06/tdd-hungarian-notation-for-mocksstubs

2010-01-24 01:13:57

TDD: Removing the clutter

I got the chance to work with http://fragmental.tw/[Phil] for a couple of weeks last year and one of the most interesting things that he started teaching me was the importance of reducing the clutter in our tests and ensuring that we take some time to refactor them as well as the code as part of the 'red-green-refactor' cycle. I'm still trying to work out the best way to do this but I came across a really http://blog.thecodewhisperer.com/post/333781027/what-your-tests-dont-need-to-know-will-hurt-you[interesting post by J.B. Rainsberger where he describes how he removes irrelevant details from his tests]. Since I worked with Phil I've started noticing some of the ways that we can simplify tests so that they are more useful as documentation of how our system works. == Wrapping methods around irrelevant test builders One thing I've noticed in tests recently is that generally most of the setup code for a test is irrelevant and is just there to get the test to actually run. There's very little that's actually interesting and more often than not it ends up getting hidden amongst the other irrelevant stuff. The http://www.markhneedham.com/blog/2009/01/21/c-builder-pattern-still-useful-for-test-data/[test builder pattern] is a really useful one for allowing us to easily setup test data but I feel that if we're not careful it contributes to the clutter. To describe a contrived example: [source,csharp] ---- [Test] public void SomeTest() { var bar = BarBuilder.Build.WithBaz("baz").BuildBar(); var foo = FooBuilder.Build.Bar(bar).BuildFoo(); var someObject = new SomeObject(); var result = someObject.SomeMethod(foo); Assert.That(result.Baz, Is.EqualTo("baz"); } ---- In this example 'Foo' is actually not important at all. What we're really interested in is Baz which happens to be accessed via Foo. I've started refactoring tests like that into the following style: [source,csharp] ---- [Test] public void SomeTest() { var bar = BarBuilder.Build.WithBaz("baz").BuildBar(); var someObject = new SomeObject(); var result = someObject.SomeMethod(FooWith(bar)); Assert.That(result.Baz, Is.EqualTo("baz"); } private Foo FooWith(Bar bar) { return FooBuilder.Build.Bar(bar).BuildFoo(); } ---- In this example it doesn't make much different in terms of readability but when there's more dependencies it works quite well for driving the test into a state where all we see are the important details that form part of the 'Arrange - Act - Assert' pattern. == New up object under test inline Object initialisation is often not worthy of a variable in our test since it doesn't really add anything to our understanding of the test. I only really break this rule when we need to call one method on the object under test and then need to call another method to verify whether or not the expected behaviour happened. More often than not this is only the case when dealing with framework code. It's much easier to avoid this in our own code. In the example that I started with we can inline the creation of 'SomeObject' without losing any of the intent of the test: [source,csharp] ---- [Test] public void SomeTest() { var bar = BarBuilder.Build.WithBaz("baz").BuildBar(); var result = new SomeObject().SomeMethod(FooWith(bar)); Assert.That(result.Baz, Is.EqualTo("baz"); } ---- The only time I don't do this is when the constructor takes in a lot of dependencies and keeping it all inlined would take the code off the right side of the screen. In any case it's a sign that something's gone wrong and the object probably has too many dependencies so we need to try and fix that. == Pull up static dependencies into fields Another technique I've been trying is pulling static dependencies i.e. ones whose values are not mutated in the test up into fields and initialising them there. A typical example would be in http://www.markhneedham.com/blog/2010/01/15/tdd-thoughts-on-using-a-clock-in-tests/[tests that have a clock]. [source,csharp] ---- [Test] public void ShouldShowFoosOlderThanToday() { var clock = new ControlledClock(new DateTime(2010,1,16)); var fooService = MockRepository.GenerateStub<IFooService>(); var fooFromYesterday = new Foo { Date = 1.DayBefore(clock) }); var aCollectionOfFoos = new List<Foo> { fooFromYesterday }; fooService.Stub(f => f.GetFoos()).Return(aCollectionOfFoos); var oldFoos = new FooFinder(clock, fooService).GetFoosFromEarlierThanToday(); Assert.That(oldFoos.Count, Is.EqualTo(1)); // and so on } ---- I would pull the clock variable up to be a field since the value we want to return for it is going to be the same for the whole test fixture. [source,csharp] ---- [TestFixture] public class TheTestFixture { private readonly IClock Clock = new ControlledClock(new DateTime(2010,1,16)); [Test] public void ShouldShowFoosOlderThanToday() { var fooService = MockRepository.GenerateStub<IFooService>(); var fooFromYesterday = new Foo { Date = 1.DayBefore(clock) }); var aCollectionOfFoos = new List<Foo> { fooFromYesterday }; fooService.Stub(f => f.GetFoos()).Return(aCollectionOfFoos); var oldFoos = new FooFinder(Clock, fooService).GetFoosFromEarlierThanToday(); Assert.That(oldFoos.Count, Is.EqualTo(1)); // and so on } } ---- I'm less certain what I would do with 'fooService'. I've run into problems previously by pulling these types of dependencies up into a setup method if we've also moved the corresponding 'Stub' or 'Mock' call as well. With that setup the intent of the test is now in two places which makes it more difficult to understand. == In Summary It's really interesting to read about the way that others are trying to write better tests and http://www.exampler.com/blog/2010/01/13/mocks-the-removal-of-test-detail-and-dynamically-typed-languages/[Brian Marick also has a post where he describes how he is able to create even more intention revealing tests in a dynamic language]. It'd be cool to here some more ideas around this.

tdd-removing-the-clutter

https://markhneedham.com/blog/2010/01/24/tdd-removing-the-clutter

2010-01-23 14:45:59

Coding: The collecting parameter pattern

The http://www.industriallogic.com/xp/refactoring/accumulationToCollection.html[collecting parameter pattern] is one of my favourite ones when used well but I've noticed recently that it can lead to quit misleading APIs as well. One way that we used it quite effectively was when http://www.markhneedham.com/blog/2009/03/10/oo-micro-types/[getting objects to render themselves to a ViewData container] which was then used to populate the view. [source,csharp] ---- public class Micro { private string micro; public Micro(string micro) { this.micro = micro; } public void renderTo(ViewData viewData) { viewData.add(micro); } } ---- I think it's quite obvious from the method name what it does and we can easily test this method by checking on the contents of the value passed in after the method's been executed. The problem I've noticed with this pattern is when the method isn't so explicit about what it's doing or where a method is using a collecting parameter as well as returning a value. For example: [source,csharp] ---- public SomeModel GetModel(ViewData viewData) { viewdata["someKey"] = "someValue; // do some other stuff return new SomeModel(...); } ---- It's not entirely obvious from reading the method signature that it would be mutating the 'ViewData' that's been passed in. I would actually expect that 'SomeModel' is being constructed from 'ViewData' if I just read the method signature in isolation. It's almost like it's following the collecting parameter pattern by accident. In a way I find that API as misleading as ones which make use of an 'out' parameter to allow multiple values to be returned - it doesn't do what you expect, its *intent is not clear*. Since I've started playing around with functional programming languages I've become more keen on the idea that functions take in inputs and return outputs and this pattern goes against that idea by encouraging mutation of state. It seems to me that for the most part writing functions which return a value and don't do anything else reveal their intent more and for the less frequent occasions where we want to keep encapsulation but still find out some data about an object the collecting parameter pattern works pretty well. It should be used sparingly though.

coding-the-collecting-parameter-pattern

https://markhneedham.com/blog/2010/01/23/coding-the-collecting-parameter-pattern

2010-01-15 21:44:36

TDD: Hand written stubs vs Framework generated stubs

A few months ago http://blog.objectmentor.com/articles/2009/10/28/manual-mocking-resisting-the-invasion-of-dots-and-parentheses[Uncle Bob wrote a post about TDD where he suggested that he preferred to use hand created stubs in his tests] wherever possible and only resorted to using a Mockito created stub as a last resort. I've tended to use framework created ones but my colleague Matt Dunn and I noticed that it didn't seem to work out too well for us writing some tests around a controller where the majority of our tests were making exactly the same call to that repository and expected to receive the same return value but a few select edge cases expected something different. The edge cases came along later on by which time we'd already gone past the stage of putting the stub expectation setup in every test and had moved it up into a setup method which ran before all test. We tried to keep on using the same stub instance of the dependency which meant that we were trying to setup more than one stub expectation on the same method with different return values. [source,csharp] ---- [TestFixture] public SomeControllerTests { private ITheDependency theDependency; [SetUp] public void Before() { theDependency = MockRepository.GenerateStub<ITheDependency>(); theDependency.Stub(t => t.SomeMethod()).Return(someResult); controller = new Controller(theDependency); } .... [Test] public void SomeAnnoyingEdgeCaseTest() { theDependency.Stub(t => t.SomeMethod().Return(someOtherResult); controller.DoSomething(); // and so on... } } ---- We were struggling to remember where we had setup the various return values and could see a few ways to try and reduce this pain. The first was to extract those edge case tests out into another test fixture where we would setup the stub expectation on a per test basis instead of using a blanket approach for all of them. This generally works pretty well although it means that we now have our tests in more than one place which can be a bit annoying unless you know that's what's being done. Another way which Matt suggested is to *make use of a hand written stub for the tests which aren't bothered about the return value and then use a framework created one for specific cases*. The added benefit of that is that it's more obvious that the latter tests care about the result returned because it's explicitly stated in the body of the test. We found that our tests were easier to read once we'd done this and we spent much less time trying to work out what was going on. I think framework generated stubs do have thier place though and as http://blog.objectmentor.com/articles/2009/10/28/manual-mocking-resisting-the-invasion-of-dots-and-parentheses[Colin Goudie points out in the comments on Uncle Bob's post] it probably makes more sense to make use of a framework generated stub when each of our tests returns different values for the same method call. If we don't do that then we end up writing lots of different hand written stubs which I don't think adds much value. I still start out with a framework generated stub but if it's becoming overly complicated or repetitive then I'm happy to switch to a hand written one. It'd be interesting to hear others' approaches on this.

tdd-hand-written-stubs-vs-framework-generated-stubs

https://markhneedham.com/blog/2010/01/15/tdd-hand-written-stubs-vs-framework-generated-stubs

2010-01-15 23:23:58

C#: A functional solution to a modeling problem

We were working on some refactoring today where we http://www.markhneedham.com/blog/2009/11/11/coding-pushing-the-logic-back/[pushed some logic back] from a service and onto a domain object and I noticed that we were able to use functions quite effectively to reduce the amount of code we had to write while still describing differences in behaviour. The class we want to write needs to take in two integers which represent two different situations related to Foo. Depending upon whether we have 'Situation 1', 'Situation 2' or both situations we will display the results slightly differently. [source,csharp] ---- public class Foo { private readonly int situationOneValue; private readonly int situationTwoValue; private readonly Func<int, string> situationOneDisplayFunc; private readonly Func<int, string> situationTwoDisplayFunc; private Foo(int situationOneValue, int situationTwoValue, Func<int, string> situationOneDisplayFunc, Func<int, string> situationTwoDisplayFunc) { this.situationOneValue = situationOneValue; this.situationTwoValue = situationTwoValue; this.situationOneDisplayFunc = situationOneDisplayFunc; this.situationTwoDisplayFunc = situationTwoDisplayFunc; } public static Foo ForSituation1(int situationOneValue, int situationTwoValue) { return new Foo(situationOneValue, situationTwoValue, Format, _ => "Irrelevant value"); } public static Foo ForSituation2(int situationOneValue, int situationTwoValue) { return new Foo(situationOneValue, situationTwoValue, _ => "Irrelevant value", Format); } public static Foo ForSituation1And2(int situationOneValue, int situationTwoValue) { return new Foo(situationOneValue, situationTwoValue, Format, Format); } public string Situation1() { return situationOneDisplayFunc(situationOneValue); } public string Situation2() { return situationTwoDisplayFunc(situationTwoValue); } private string Format(int value) { return string.Format("Formatted Value {0}"); } } ---- The way that it's used is that we get a value coming in from another system as a string which represents which situation we have to deal with. As a result in the code which processes this data we have a dictionary which maps from these strings to the corresponding static method defined above: [source,csharp] ---- private static Dictionary<string, Func<int, int, Foo>> Foos = new Dictionary<string, Func<int, int, Foo>> { { "Situation 1", Foo.ForSituation1 }, { "Situation 2", Foo.ForSituation2 }. { "Situation 1 and 2", Foo.ForSituation1And2 }; ---- This is a neat little idea which I learnt from some Scala code that my colleague http://elhumidor.blogspot.com/[John Hume] posted on an internal mailing list a few weeks ago where he'd done something similar. In our code which parses the incoming data we have something like the following: [source,csharp] ---- var foo = Foos[incomingMessage.Situation](incomingMessage.Situation1Value, incomingMessage.Situation2Value); // and so on ---- If we wanted to solve this without using functions then for the first part I think we would probably need to use inheritance to create three different Foos, probably with a common interface, and then define the different behaviour on those classes. The second part of the solution might end up being solved with some if statements or something similar. I was a bit unsure of how easy it was to understand this code when we first came up with it but it's growing on me.

c-a-functional-solutional-to-a-modeling-problem

https://markhneedham.com/blog/2010/01/15/c-a-functional-solutional-to-a-modeling-problem

2010-01-15 21:56:48

TDD: Thoughts on using a clock in tests

A few months ago http://blog.objectmentor.com/articles/2009/10/28/manual-mocking-resisting-the-invasion-of-dots-and-parentheses[Uncle Bob wrote a post about TDD where he suggested that he preferred to use hand created stubs in his tests] wherever possible and only resorted to using a Mockito created stub as a last resort. I http://www.markhneedham.com/blog/2010/01/15/tdd-hand-written-stubs-vs-framework-generated-stubs/[wrote previously about my thoughts of where to use each of the two approaches] and one example of where hand written stubs seems to make sense is the clock. I wonder if this ties in with http://blog.thecodewhisperer.com/post/333781027/what-your-tests-dont-need-to-know-will-hurt-you[J.B. Rainsberger's theory of irrelevant details] in the tests which make use of it. We would typically define an interface and stub version of the clock like so: [source,csharp] ---- public interface IClock { DateTime Now(); } ---- [source,csharp] ---- public class ControlledClock : IClock { private readonly DateTime dateTime; public ControlledClock(DateTime dateTime) { this.dateTime = dateTime; } public DateTime Now() { return dateTime; } } ---- I forgot about it to start with and was stubbing it out using Rhino Mocks but I realised that every single test needed something similar to the following code: [source,csharp] ---- var theCurrentTime = new DateTime(2010, 1, 16); var clock = MockRepository.GenerateStub<IClock>(); clock.Stub(c => c.Now()).Return(theCurrentTime); ---- We can extract out the creation of the first two lines into fields but the third line remains and typically that might end up being pushed into a setup method which is run before each test. With a clock it's maybe not such a big deal but with other dependencies from my experience it can become very difficult to follow where exactly the various return values are being setup. When we use a hand written stub we only have to write the following code and then the date is controlled everywhere that calls 'Now()': [source,csharp] ---- private readonly IClock clock = new ControlledClock(new DateTime(2010,1,16)); ---- Following on from that my colleague http://twitter.com/mikewagg[Mike Wagg] suggested the idea of creating extension methods on integers to allow us to fluently define values relative to the clock. [source,csharp] ---- [Test] public void ShouldShowFoosOlderThanToday() { var clock = new ControlledClock(new DateTime(2010,1,16)); var fooService = MockRepository.GenerateStub<IFooService>(); var fooFromYesterday = new Foo { Date = 1.DayBefore(clock) }); var aCollectionOfFoos = new List<Foo> { fooFromYesterday }; fooService.Stub(f => f.GetFoos()).Return(aCollectionOfFoos); var oldFoos = new FooFinder(clock, fooService).GetFoosFromEarlierThanToday(); Assert.That(oldFoos.Count, Is.EqualTo(1)); // and so on } ---- The extension method on integer would be like this: [source,csharp] ---- public static class IntegerExtensions { public static DateTime DayBefore(this int value, IClock clock) { return clock.Now.Subtract(TimeSpan.FromDays(value)); } } ---- It reads pretty well and seems more intention revealing than any other approaches I've tried out so I think I'll be continuing to use this approach.

tdd-thoughts-on-using-a-clock-in-tests

https://markhneedham.com/blog/2010/01/15/tdd-thoughts-on-using-a-clock-in-tests

2010-01-12 01:33:58

F#: Refactoring to pattern matching

I was looking through some of the F# code I've written recently and I realised that I was very much writing C# in F# with respect to the number of if statements I've been using. I thought it would be interesting to see what the code would look like if I was able to refactor some of that code to make use of pattern matching instead which would be a more idiomatic way of solving the problem in F#. The first example of if statements is in http://www.markhneedham.com/blog/2010/01/10/roy-osheroves-tdd-kata-an-f-attempt/[my post about my F# solution to Roy Osherove's TDD Kata]. I originally wrote a parse function which was able to parse a string and give it's decimal value or 0 if it couldn't be parsed. [source,ocaml] ---- let parse value = let (itParsed, value) = Decimal.TryParse value if (itParsed) then value else 0.0m ---- If we use a pattern match expression we'd end up with the following: [source,ocaml] ---- let parse value = match Decimal.TryParse value with | (true, value) -> value | (false, _) -> 0.0m ---- The neat thing about this approach is that we don't need to store the result of the 'Decimal.TryParse' function as we did in my original version. We could in theory also write it like this... [source,ocaml] ---- let parse value = match Decimal.TryParse value with | (true, value) -> value | (_, _) -> 0.0m ---- ...but while that is slightly less code I quite like the other version because it's a bit more intention revealing that 0 will be returned if we fail to parse the string. I think there's less thinking needed to understand the code. Another example is this bit of code: [source,ocaml] ---- let add value = if ("".Equals(value) or "\n".Equals(value)) then 0.0m else match digits value |> Array.filter (fun x -> x < 1000m) with | ContainsNegatives(negatives) -> raise (ArgumentException (buildExceptionMessage negatives)) | NoNegatives(digits) -> digits |> Array.sum ---- If we convert that to use pattern matching we would get this: [source,ocaml] ---- let add value = match value with | "" -> 0.0m | "\n" -> 0.0m | value -> match digits value |> Array.filter (fun x -> x < 1000m) with | ContainsNegatives(negatives) -> raise (ArgumentException (buildExceptionMessage negatives)) | NoNegatives(digits) -> digits |> Array.sum ---- That's maybe slightly easier to read mainly because of the splitting up of the two inputs which lead to a 0 result. The final example of using if statements is the following bit of code: [source,ocaml] ---- let (|CustomDelimeter|NoCustomDelimeter|) (value:string) = ... if (value.Length > 2 && "//".Equals(value.Substring(0, 2))) then if ("[".Equals(value.Substring(2,1))) then CustomDelimeter(delimeters value) else CustomDelimeter([| value.Substring(2, value.IndexOf("\n") - 2) |]) else NoCustomDelimeter(",") ---- The only way I could see how to make this use active patterns is on the boolean statements like so: [source,ocaml] ---- let (|CustomDelimeter|NoCustomDelimeter|) (value:string) = ... match (value.Length > 2 && "//".Equals(value.Substring(0, 2))) with | true -> match ("[".Equals(value.Substring(2,1))) with | true -> CustomDelimeter(delimeters value) | false -> CustomDelimeter([| value.Substring(2, value.IndexOf("\n") - 2) |]) | false -> NoCustomDelimeter(",") ---- I quite like that it lines up the return values of the active pattern which seems to make it a bit more readable. I think overall maybe the pattern matching versions are slightly more readable but maybe not by much. What do you think?

f-refactoring-to-pattern-matching

https://markhneedham.com/blog/2010/01/12/f-refactoring-to-pattern-matching

2010-01-13 01:37:15

C# Test Builder Pattern: My current thinking

I've written previously about the http://www.markhneedham.com/blog/2009/01/21/c-builder-pattern-still-useful-for-test-data/[test builder pattern in C#] and having noticed some different implementations of this pattern I thought it'd be interesting to post my current thinking on how to use it. One thing I've noticed is that we often end up just creating methods which effectively act as setters rather than easing the construction of an object. This seems to happen most commonly when the value we want to set is a boolean value. The following is quite typical: [source,csharp] ---- public CarBuilder WithSomething(boolean value) { this.something = value; return this; } ---- The usage would be like so: [source,csharp] ---- new CarBuilder().WithSomething(true).Build(); ---- It doesn't read too badly but it seems to be unnecessary typing for the user of the API. If we're going to use the fluent interface approach then I would prefer to have two methods, defined like so: [source,csharp] ---- public CarBuilder WithSomething() { this.something = true; return this; } public CarBuilder WithoutSomething() { this.something = false; return this; } ---- We could then use this like so: [source,csharp] ---- new CarBuilder().WithSomething().Build(); ... new CarBuilder().WithoutSomething().Build(); ---- That requires more code but I think it's a bit more expressive and makes life easier for the user of the API. An alternative approach which my colleague Lu Ning showed me and which I think is actually better is to http://www.markhneedham.com/blog/2009/08/15/builders-hanging-off-class-vs-builders-in-same-namespace/[make use of the object initializer syntax if all we have are setter methods on a builder]. We might therefore end up with something like this: [source,csharp] ---- public class FooBuilder { public string Something = "DefaultSomething"; public boolean SomethingElse = false; public Foo Build() { return new Foo(Something, SomethingElse); } } ---- [source,csharp] ---- new FooBuilder { SomethingElse = true }.Build(); ---- With this approach we end up writing less code and although we use public fields on the builder I don't think it's a big deal since it allows us to achieve our goal more quickly. If we need other methods that take out the complexity of construction then we can easily just add those as well. Another thing to note when using this pattern is that *we don't need to override all the object's attributes on every single test*. We only need to override those ones which we are using in our test. The rest of the values can just be defaulted. [source,csharp] ---- [Test] public void ShouldDoSomething() { var foo = new FooBuilder { Bar = "myBar", Baz = "myBaz", SomethingElse = "mySE", AndAgain = "..." }.Build(); // and then further on we only check the value of Bar for example Assert.That(someValue, Is.EqualTo("myBar"); } ---- We don't need to specifically set any of the values except 'Bar' because they are irrelevant and create a clutter in the test which means it takes longer to understand what's going on. This would be preferable: [source,csharp] ---- [Test] public void ShouldDoSomething() { var foo = new FooBuilder { Bar = "myBar" }.Build(); // and then further on we only check the value of Bar for example Assert.That(someValue, Is.EqualTo("myBar"); } ---- Something which I've been wondering about recently is understanding the best way to describe the case where we don't want to define a specific value i.e. we want it as null. I'd normally just set it to null like so: [source,csharp] ---- new FooBuilder { Bar = null }.Build(); ---- But we could make the API have a specific method and I haven't really decided whether there's much value to doing so yet: [source,csharp] ---- new FooBuilder().WithNoBar().Build(); ----

c-test-builder-pattern-my-current-thinking

https://markhneedham.com/blog/2010/01/13/c-test-builder-pattern-my-current-thinking

2010-01-14 08:01:29

F#: Refactoring to sequence/for expressions

Since I started playing around with F# one of the things I've been trying to do is not use the 'for' keyword because I was trying to avoid writing code in an imperative way and for loops are a big part of this for me. Having read http://fsharpnews.blogspot.com/2009/12/zach-cox-word-count-challenge.html[Jon Harrop's solution] to the http://www.markhneedham.com/blog/2009/12/20/f-word-count-using-a-dictionary/[word count problem] where he made use of both sequence and for expressions I thought it'd be intersting to see what some of the code I've written would look like using that approach. An example of a function that I wrote which could be rewritten the other way is the following: [source,ocaml] ---- let delimeters (value:string) = Regex.Matches(value, "\[([^]]*)\]") |> Seq.cast |> Seq.map (fun (x:Match) -> x.Groups) |> Seq.map (fun x -> x |> Seq.cast<Group> |> Seq.nth 1) |> Seq.map (fun x -> x.Value) ---- This could be written like this if we used a sequence expression instead of chaining map operations: [source,ocaml] ---- let delimeters (value:string) = seq { for m in Regex.Matches(value, "\[([^]]*)\]") do yield m.Groups.Item(0).Value } ---- One interesting thing I found about writing it like this was that I noticed that 'GroupCollection' had the 'Item' property on it which would let me get the match much more easily. I completely missed that when I was writing the first solution so I'm not sure if that was just due to my lack of knowledge of that part of the API or whether the second approach actually encouraged me to explore more and therefore end up with a simpler solution. Another example I found was this expression for getting the matches for a regular expression: [source,ocaml] ---- let regex pattern input = Regex.Matches(input, pattern) |> Seq.cast |> Seq.map (fun (x:Match) -> x.Value) ---- That can be simplified to the following with a sequence expression: [source,ocaml] ---- let regex pattern input = seq { for m in Regex.Matches(input, pattern) do yield m.Value } ---- One neat thing about using sequence expressions is that we don't need to make use of 'Seq.cast' to convert a value to a typed sequence - we can just use it as it is. The following function can be rewritten to just use a for expression: [source,ocaml] ---- let writeTo (path:string) (values:seq<string * int>) = use writer = new StreamWriter(path) values |> Seq.map (fun (value, count) -> value + " " + count.ToString()) |> Seq.iter (fun x -> writer.WriteLine(x)) ---- Like so: [source,ocaml] ---- let writeTo (path:string) (values:seq<string * int>) = use writer = new StreamWriter(path) for (value,count) in values do writer.WriteLine(value + " " + count.ToString()) ---- We eventually iterate through the sequence anyway so I think it's more intention revealing to just do the iteration and mapping in one step. This is a function from when I was http://www.markhneedham.com/blog/2009/07/12/f-a-day-writing-a-feedburner-graph-creator/[writing the little Feedburner application]: [source,ocaml] ---- let calculateWeeklyAverages = Seq.reverseSequence >> Seq.windowed days >> Seq.map (fun (entries:array<Entry>) -> (entries.[0]).Date , entries |> Array.map (fun e -> e.Circulation |> toDecimal) |> Array.average ) >> Seq.reverseSequence ---- If we use a sequence expression it'd look like this: [source,ocaml] ---- let calculateWeeklyAverages entries = seq { for (e:array<Entry>) in (entries |> Seq.reverseSequence |> Seq.windowed days) do yield ((e.[0]).Date, entries |> Array.map (fun e -> e.Circulation |> toDecimal) |> Array.average) } |> Seq.reverseSequence ---- The resulting code is shorter but it seems to me like the focus when you read the code has moved to the line which yields the tuple whereas in the first version I find that I read the function as a whole. I've not really used sequence expressions that much so it's been interesting going through the code and seeing where they might be useful. I found several places where I'd used lists because I find those easier to pattern match against but I wonder whether it would make sense to use sequences there as well.

f-refactoring-to-sequencefor-expressions

https://markhneedham.com/blog/2010/01/14/f-refactoring-to-sequencefor-expressions

2010-01-22 23:21:56

Automapper: First thoughts

I came across Jimmy Bogard's http://www.codeplex.com/AutoMapper[Automapper] library a while ago but hadn't had the opportunity to try it out on a project until this week. The problem we wanted to solve was relatively simple. We had a domain object and we wanted to create a copy of that with one of the fields changed and all of the ids cleared from the object and any objects contained within it so that we could persist the new web of objects to the database. We had a structure a bit like this: [source,csharp] ---- public class Foo { public Bar Bar { get; set; } public Baz Baz { get; set; } } ---- And wanted to create a copy of this object while changing one of the values on Baz: [source,csharp] ---- Mapper.CreateMap<Foo, Foo>().ForMember(x => x.Id, opts => opts.Ignore()); Mapper.CreateMap<Bar, Bar>().ForMember(x => x.Id, opts => opts.Ignore()); Mapper.CreateMap<Baz, Baz>() .ForMember(x => x.Id, opts => opts.Ignore()) .ForMember(x => x.SomeProperty, opts => opts.MapFrom(source => someNewValue)); ---- It works really well although it took me a little while to realise that the 'Mapper' class was keeping track of what we'd set up via its methods internally. http://www.lostechies.com/blogs/jimmy_bogard/archive/2009/12/23/automapper-dsl-design-post-mortem.aspx[Jimmy refers to this as the static gateway pattern] which seems a little similar to the way that Rhino Mocks keeps track of expectations from what I remember from http://www.markhneedham.com/blog/2009/07/28/reading-code-rhino-mocks/[reading some of the code]. In general though I've got more used to expression builder style DSLs so it was interesting to see one which has been done with a different approach. The next thing we had to work out was where that mapping code should go. Since it's completely about 'Foo' I originally thought it should go inside Foo on a static method which we could use to clone the object perhaps with an API like this: [source,csharp] ---- public Foo CloneWith(string someNewValue) { // mapping code in here } ---- The problem with that is that it means we have a static dependency in our domain model and from reading http://alistair.cockburn.us/Hexagonal+architecture[Alistair Cockburn's Hexagonal Architecture article] I've come to believe that we should try to ensure that objects in our domain model don't have dependencies on anyone else. An alternative might be to have it on the object and then create an interface which represents the ability to clone Foo: [source,csharp] ---- public Foo CloneWith(string someNewValue, ICanCloneAFoo fooCloner) { return fooCloner.Clone(); } ---- The mapping code would then go on the implementer of 'ICanCloneAFoo'. The discussions we had around this reminded me of http://www.markhneedham.com/blog/2009/02/09/oop-what-does-an-objects-responsibility-entail/[a post I wrote about a year ago] and the solution we decided to use was the same. We put the mapping code onto a mapper object. It's not a bad solution as it follows the convention that's been done for other mapping problems in the code although as http://twitter.com/mikewagg[Mike] pointed out, the code to clone the object is not as discoverable as it would be if it was on the object itself. In that mapper we're still calling the Automapper code directly which I think is fine although I'm not sure whether we're quite adhering to the idea of wrapping 3rd party libraries, as suggested in http://www.amazon.com/gp/product/0321503627?ie=UTF8&tag=marneesblo-20&linkCode=as2&camp=1789&creative=390957&creativeASIN=0321503627[Growing Object Oriented Software], and not allowing them to bleed into our code base. It seems like in the case doing that might lead to more problems than it would solve.

automapper-first-thoughts

https://markhneedham.com/blog/2010/01/22/automapper-first-thoughts

2010-01-25 22:06:23

TDD: Rewriting/refactoring tests

I've read several times about the dangers of http://chadfowler.com/2006/12/27/the-big-rewrite[the] http://blog.objectmentor.com/articles/2009/01/09/the-big-redesign-in-the-sky[big] http://www.magpiebrain.com/2010/01/10/the-great-rewrite/[rewrite] when it comes to production code but I've recently been wondering whether or not we should apply the same rules when it comes to test code or not. I worked with http://raphscallion.com/blog/[Raphael Speyer] for a few weeks last year and on the code base we were working on he often spent some time rewriting tests originally written using http://rmock.sourceforge.net/[rMock] to use http://mockito.org/[mockito] which was the framework we were driving towards. One of the benefits that he was able to get from doing this was that he had to understand the test in order to change it which enabled him to increase his understanding of how the code was supposed to work and identify anything that didn't seem quite right. I quite liked this idea at the time and while spending some time recently working with some tests which required quite a lot of setup and were testing several different things in the same test. Unfortunately a few of them were failing and it was quite difficult to work out why that was. My initial approach was to try and work my way through the tests inlining all the test code to start with and then http://www.markhneedham.com/blog/2010/01/24/tdd-removing-the-clutter/[extracting out irrelevant details] to make the tests easier to understand. Despite those attempts I was unable to work out why the test was failing so I worked out what the main things the test was trying to verify and then wrote tests from scratch for each of those cases. I was able to write tests covering everything the original test did in several smaller tests in less time than I had spent trying to debug the original one and with a fair degree of confidence that I'm testing exactly the same thing. As I see it the big danger of rewriting is that we're always playing catch up with the current system which is still being worked on in production and we never quite catch up. I'm not so sure this logic applies in this case because we're only rewriting small bits of code which means that we can replace the original test very quickly. My main driver when working with tests is to ensure that they're *easy to understand* and make it *easy to reason about the code* so if I have to rewrite some tests to make that happen then I think it's a fair trade off. My initial approach would nearly always be to refactor the tests that are already there. Rewriting is something I'd look to do if I was really struggling to refactor effectively.

tdd-rewritingrefactoring-tests

https://markhneedham.com/blog/2010/01/25/tdd-rewritingrefactoring-tests

2010-01-25 21:23:31

TDD: Simplifying a test with a hand rolled stub

I wrote a couple of weeks ago about my thoughts on http://www.markhneedham.com/blog/2010/01/15/tdd-hand-written-stubs-vs-framework-generated-stubs/[hand written stubs vs framework generated stubs] and I noticed an interesting situation where it helped me out while trying to simplify some test code. The code in question was making use of several framework generated stubs/mocks and one in particular was trying to return different values depending on the value passed as a parameter. The test was failing and I spent about half an hour unsuccessfully trying to work out why it wasn't working as expected before I decided to replace it with a hand rolled stub that did exactly what I wanted. This is a simplified version of the test: [source,csharp] ---- [Test] public void SomeTest() { var service = MockRepository.GenerateStub<IService>(); service.Stub(x => x.SomeMethod("call1")).Return("aValue"); service.Stub(x => x.SomeMethod("call2")).Return("anotherValue"); // and so on new SomeObject(service).AnotherMethod(); // some assertions } ---- For the sake of the test I only wanted 'service' to return a value of 'aValue' the first time it was called and then 'anotherValue' for any other calls after that. I therefore wrote the following hand rolled stub to try and simplify the test for myself and plugged it into the original test: [source,csharp] ---- public class AValueOnFirstCallThenAnotherValueService : IService { private int numberOfCalls = 0; public string SomeMethod(string parameter) { if(numberOfCalls == 0) { numberOfCalls++; return "aValue"; } else { numberOfCalls++; return "anotherValue"; } } } ---- [source,csharp] ---- [Test] public void SomeTest() { var service = new AValueOnFirstCallThenAnotherValueService(); new SomeObject(service).AnotherMethod(); // some assertions } ---- I've never tried this particular approach before but it made it way easier for me to identify what was going wrong and I was then able to get the test to work as expected and move onto the next one. In retrospect it should have been possible for me to work out why the original framework generated stub wasn't working but it seemed like the right time to cut my losses and the time to write the hand generated one and get it working was an order of magnitude less.

tdd-simplifying-a-test-with-a-hand-rolled-stub

https://markhneedham.com/blog/2010/01/25/tdd-simplifying-a-test-with-a-hand-rolled-stub

2010-06-04 21:03:38

Ask for forgiveness, not for permission

I gave a presentation at our ThoughtWorks Brazil office in Porto Alegre last way on some of the things that I've learned while working at ThoughtWorks and the first point I made was that it was better to '*ask for forgiveness, not for permission*'. This was something that was taught to me a few years ago and the idea behind this is that if there's some idea we want to try out it makes much more sense to start trying it now and then we can always apologise later on if someone has a problem with us doing that. This is in preference to an approach where we ask whether we can try out an idea and get told no straight away i.e. we should look to be *proactive*. It's really easy to say no to something when you don't really no what it is you're being asked and it seems to be the default answer a lot of the time. An example of something where this rule may apply would be if we wanted to try writing some tests with a different language/framework to see whether it improves the way we're doing things. We don't really know whether the new approach is going to be better than the current one so it makes sense to spend a bit of time playing around with it so that we can find out. Later on in the talk I suggested that on some decisions it makes sense to gather data and then let the client decide what they want to do. An example of this happened on a project I worked on a couple of years ago where the value of writing functional/unit tests was being questioned. My colleague Kristan Vingrys came up with a matrix which showed the risks of not testing various features and then let the client decide whether they wanted to take that risk. http://blog.m.artins.net/measuring-test-effort/[Alex wrote more about this at the time]. After I mentioned this http://twitter.com/caiokf[Caio] pointed out that this idea seemed to contradict the idea of asking for forgiveness and not permission and at the time I couldn't really think of a good answer to why I didn't feel the situations were exactly the same. A week on I think that the difference is that the 'ask for forgiveness' approach applies mainly for new situations where we haven't agreed as a team what we should do and people don't have enough information yet to know whether we should or should not do what is being proposed. The idea of gathering data and letting the client decide applies more when a decision has been made or suggested and we want to try and influence that decision. It doesn't make sense to just do what we want in this situation if it explicitly goes against what the client wants but equally we don't want to do something which goes against what we believe is the best approach without at least pointing that out.

ask-for-forgiveness-not-for-permission

https://markhneedham.com/blog/2010/06/04/ask-for-forgiveness-not-for-permission

2010-06-04 21:05:52

XP2010: Coding Dojo Open Space

I attended an open space hosted by http://twitter.com/emilybache[Emily Bache] at the XP2010 conference in Trondheim, Norway with several other people who have been organising coding dojos around the world. It was really interesting to hear about some of the different approaches that people have taken and how a lot of the issues we had with http://www.markhneedham.com/blog/category/coding-dojo/[the one we used to run in Sydney] were the same as what others had experienced. These were some of the things that I took from the session: * We talked about the difficulty of keeping up the enthusiasm/energy for a coding dojo over a long period of time and Emily suggested that one way to reduce this problem was to *link the dojo with a group which met up for something else* other than just a coding dojo. For example Emily runs a dojo with the Ruby and Python user groups whereby one week someone would do a presentation on something and the next week they would run a dojo. The one we ran was generally internal to ThoughtWorks so we had a different problem in that some people preferred a coding dojo and others preferred to hack on problems alone but with others around. * Another suggestion around this was that *rotating the organiser* between the group was a way to keep enthusiasm up and made it the whole groups' responsibility to keep the dojo running rather than placing the responsibility at the feet of one or two people * An idea suggested by http://twitter.com/hugocorbucci[Mariana] & http://twitter.com/hugocorbucci[Hugo] from their experiences running a dojo in Sao Paulo was how to *introduce a new language in the dojo environment*. They originally tried just running a randori style dojo but found that this didn't work out so well if there wasn't at least one person who had enough experience to at least know how to setup the development environment and write some tests in the language. The approach they took was to have one person do a prepared kata in the new language at the beginning of the session and then after they'd done that then the rest of the group would participate. * Another suggestion was to run a retrospective on each dojo session where people would write down *what they had learnt from the session and what had hindered their learning*. This strikes me as something which would be quite useful to consider even on a normal project, especially near the beginning, and would help us to at least recognise if there are factors which are preventing us learning as much as we'd like. We could then do something to fix that if necessary. * Hugo and Emily both suggested the *'Kake' format* as one which might be better suited in a dojo with people who are quite experienced with the dojo idea. In this format there would be more than one computer being used and pairs would rotate between the stations. At each station the same problem could be being solved in a different language or a different problem in the same language. * I also learnt that we had misunderstood the way we were supposed to pair in a randori dojo. We somehow had the idea that the driver was supposed to drive for the whole 7 minutes and the navigator would just navigate during that time and then they would be the driver next time around. Emily pointed out that the idea is just to *pair normally but have one person rotate after 7 minutes*. This makes much more sense to me and would have made our sessions much more natural than they felt at times while we tried to follow this idea.

xp2010-coding-dojo-open-space

https://markhneedham.com/blog/2010/06/04/xp2010-coding-dojo-open-space

2010-06-20 22:37:32

Coding: Controlled Technical Debt

A couple of months ago I wrote about http://www.markhneedham.com/blog/2010/04/14/agile-slimming-down-stories/[an approach to stories] that Christian has been encouraging on our project whereby we slim stories down to allow us to deliver the core functionality of the application as quickly as possible. In our case we had a requirement to setup a range of different parameters used to lookup reference data used in the different calculations that we have in our application. At the time we created an interface that the rest of the application would interact with so that we could easily substitute the real version in when we needed to:+++<center>+++image:{{<siteurl>}}/uploads/2010/06/technicaldebt.jpg[technicaldebt.jpg,514]+++</center>+++ We released the first version of the application about a month ago and finally last week implemented the story where the data would move from being in memory to being in a database table. One of the requirements which we had delayed by only having these parameters in memory was the ability to easily modify them. Any changes that needed to be made required an update of the code and then redeployment whereas with the database approach we would only have needed to deploy a delta script. In the event there has only been one occasion so far where those parameters needed to be updated so it hasn't proved to be a problem. Discussing this with a colleague on Friday he pointed out that what we'd done originally was to accept technical debt in our solution knowing that at some stage in the future we would need to address that. The interesting thing about this case was that *we knew exactly when we were going to repay that debt* whereas it's often the case that we create technical debt in a code base and vaguely know that at some stage we'll address it. As Martin Fowler points out in http://www.martinfowler.com/bliki/TechnicalDebt.html[his entry on technical debt]: ____ Just as a business incurs some debt to take advantage of a market opportunity developers may incur technical debt to hit an important deadline. ____ We probably saved at least a day's worth of effort by delaying this decision and were able to work on functionality that the business needed by the deadline instead. We then paid back that debt last week when we had more slack in the system. The benefits of getting something into the market place quickly are much greater than I previously imagined and I think we can look at our assumptions of how a solution 'needs' to be designed much more closely to see if we can make these trade offs more frequently.

coding-controlled-technical-debt

https://markhneedham.com/blog/2010/06/20/coding-controlled-technical-debt

2010-06-18 17:36:25

Slack time

Ken Schwaber recently wrote a blog post where he compared http://kenschwaber.wordpress.com/2010/06/10/waterfall-leankanban-and-scrum-2/[the differences between the kanban, lean and scrum approaches to software development] and although I haven't had the same experiences as he has with the first two, one interesting thing he implies is that with a scrum approach we have slack time built in. ____ God help us. People found ways to have slack in waterfall, to rest and be creative. With Lean and Kanban, those hiding places are removed. We now have a progressive death march without pause. ____ The project that I'm currently working on has switched emphasis from being in pure delivery mode into a combination of delivery and handover and it's been quite noticeable how much more slack we have in the system as a result. I find it much more enjoyable when there's a bit of slack such that we're not churning out story after story. It gives you the opportunity to explore the code base a bit more and try out any ideas that you have to see if they're likely to improve the productivity of the team. An example of this on our project is that we've had the time to introduce a calculation descriptions DSL that my colleague http://twitter.com/dermotkilroy[Dermot] had written in his own time. This has helped reduce the number of tests required in certain areas of the code base as well as making the code more intuitive and therefore easy to understand. It can be quite difficult to create this slack time when the team is under big deadline pressure and working on anything which isn't directly related to hitting that deadline isn't considered valuable. I've noticed that the typical slack in the projects I've worked on tends to be towards the end of the day if we finish a story late on. That tends to disappear if the team is working late and people start to become too tired to notice that there might be a better way to do things. It seems to me that it would beneficial to try and work some slack time into projects to allow for the type of innovation that allows us to come up with ideas that allow us to be more productive.

slack-time

https://markhneedham.com/blog/2010/06/18/slack-time

2010-06-29 06:45:11

NHibernate 2nd level cache: Doing it wrong?

I wrote a couple of weeks ago about how we'd been trying to make use of the http://www.markhneedham.com/blog/2010/06/16/fluent-nhibernate-and-the-2nd-level-cache/[NHibernate 2nd level cache] and we were able to cache our data by following the various posts that I listed. Unfortunately when we ran some performance tests we found that the performance of the application was significantly worse than when we just wrote our own 'cache' - an object which had a dictionary containing the reference data items we'd previously tried to lookup and the appropriate values. We don't need to handle cache invalidation. The client's policy is to restart production servers every night so if we want to update any of the reference data then we just need to make sure a database script is run before the servers get restarted. == Explicit transactions when reading There is a post on the NHibernate Profiler website which describes why http://nhprof.com/Learn/Alerts/DoNotUseImplicitTransactions[we should not use implicit transactions] when using NHibernate. Instead we should create explicit ones: ____ When we don't define our own transactions, it falls back into implicit transaction mode, where every statement to the database runs in its own transaction, resulting in a large performance cost (database time to build and tear down transactions), and reduced consistency. Even if we are only reading data, we should use a transaction, because using transactions ensures that we get consistent results from the database. ____ We ended up with something like this: [source,csharp] ---- public class OurRepository { public ReferenceDataObject Find(ReferenceDataObjectKey key) { using(var session = SessionFactory.OpenSession()) { using(var tx = session.BeginTransaction()) { var query = session.Linq<ReferenceDataObject>().Where(r => r.Key == key); query.QueryOptions.SetCachable(true).SetCacheMode(CacheMode.Normal); // and so on } } // return the object } } ---- As well as performance tests, we found that our integration tests became much slower than when we used our own 'cache'. We have some tests which look up 100s of different bits of reference data and the total time taken to run those tests went from around 4 seconds up to 30 seconds. As I understand it, putting a transaction around a query means that we create a transaction with the database on every request even if the query is cached and we're going to retrieve the results from the 2nd level cache rather than the database. We took out the transaction which reduced the time taken to 7 seconds but it was still slower than when we used our hand rolled cache. It seems like we must be doing something wrong or not understand something with respect to the NHibernate 2nd level cache because it seems ridiculous that the performance could be this different?

nhibernate-2nd-level-cache-doing-it-wrong

https://markhneedham.com/blog/2010/06/29/nhibernate-2nd-level-cache-doing-it-wrong

2010-06-16 00:07:43

Fluent NHibernate and the 2nd level cache

We've been trying to cache some objects using NHibernate's second level cache which always proves to be a trickier task than I remember it being the previous time! We're storing some reference data in the database and then using LINQ to NHibernate to query for the specific row that we want based on some user entered criteria. We can cache that query by calling 'SetCacheable' on the 'QueryOptions' property of our query: [source,csharp] ---- public class OurRepository { public ReferenceDataObject Find(string criteria1, string criteria2) { var query = session.Linq<ReferenceDataObject>(); query.QueryOptions.SetCachable(true).SetCacheMode(CacheMode.Normal); // and so on } } ---- That helps to ensure that if we do the exact same query again it won't go to the database again. Instead it will just retrieve the id of the appropriate object from the 2nd level cache and then go and retrieve that. In order to ensure that the 'ReferenceDataObject' does not get retrieved from the database each time we need to ensure that it is cached as well. We can do that by adding the following to its mapping file: [source,csharp] ---- public class ReferenceDataObjectMappings : ClassMap<ReferenceDataObject> { ... Cache.ReadOnly(); // or Cache.ReadWrite(); // depending on whether we want to update the value in the cache based on it changing in the database or not .. } ---- The next step is to ensure that we have the query cache and second level cache turned on in the place where we configure our session factory: [source,csharp] ---- Fluently .Configure() .Database(MsSqlConfiguration.MsSql2000.ConnectionString("connection string") .Cache(c => c.UseQueryCache().ProviderClass(typeof(NHibernate.Caches.SysCache.SysCacheProvider).AssemblyQualifiedName))) .Mappings(m => m.FluentMappings.AddFromAssemblyOf<ICurrentSessionFactory>()) .BuildSessionFactory(); ---- The 'SysCacheProvider' is one we can use for web applications although there are all sorts of others available too. Having got all this setup we wrote an integration test which added an item to the reference data table and then tried to retrieve it several times with the theory being that we would only see one call to the database. [source,csharp] ---- [Test] public void ShouldCacheTheReferenceData() { var myReferenceObject = new ReferenceDataObject(...); using(var session = SessionFactory.OpenSession()) { session.Save(myReferenceDataObject); session.Flush(); } // try and retrieve the object a few times here } ---- Unfortunately we were seeing the database being hit every single time which confused us greatly until we came across http://blogs.hibernatingrhinos.com/nhibernate/archive/2008/11/09/first-and-second-level-caching-in-nhibernate.aspx[the following explanation on Gabriel Schenker's blog]: ____ A common error (It happened to me as well!) is to *forget to commit or omit a transaction when adding or changing an entity/aggregate to the database*. If we now access the entity/aggregate from another session then the 2nd level cache will not be prepared to provide us the cached instances and NHibernate makes an (unexpected round trip to the database). ____ We need to change our test to commit the object in a transaction if we want the object to be propagated down to the 2nd level cache: [source,csharp] ---- [Test] public void ShouldCacheTheReferenceData() { var myReferenceObject = new ReferenceDataObject(...); using(var session = SessionFactory.OpenSession()) { using(var tx = session.BeginTransaction()) { session.Save(myReferenceObject); tx.Commit(); // important otherwise caching does NOT work! } } // try and retrieve the object a few times here } ----

fluent-nhibernate-and-the-2nd-level-cache

https://markhneedham.com/blog/2010/06/16/fluent-nhibernate-and-the-2nd-level-cache

2010-06-28 08:39:10

Intuition and 'quit thinking and look'

Something which http://twitter.com/dermotkilroy[Dermot], http://twitter.com/christianralph[Christian] and I noticed last week is that on our project we've reached the stage where we intuitively know what the underlying problem is for any given error message in the application we're working on. We're pretty much at the stage where we're effectively pattern matching what's going on without needing to think that much anymore. This is a good thing because it saves a lot of time analysing every single message to try and work out what's going on - I think this means that we've reached a higher level of the http://www.markhneedham.com/blog/2009/08/10/dreyfus-model-more-thoughts/[Dreyfus model] when it comes to this particular situation. The problem with getting too used to this approach is that eventually we'll come across a problem that we haven't come across before and if we don't recognise that this is the case we'll end up getting very frustrated. I had the chance to work with http://www.oshineye.com/[Ade Oshineye] a few years ago and he always encouraged me to "*quit thinking and look*" when it came to problem solving. This idea is derived from a book titled 'http://www.amazon.com/exec/obidos/ASIN/0814474578/debuggingrule-20[9 indispensable rules of debugging]' and the thinking behind it is that we often go straight to the solution for a problem without spending the time to understand what the problem actually is. We actually came across a situation like this recently while investigating a problem in production. We were getting an exception which looked pretty similar to other problems that we'd seen previously so we immediately tried that solution without any further investigation. It had no impact at all so we had to go back and actually look at the error message we were receiving before trying something else. As it turned out the time we wasted picking the wrong solution was less than 30 seconds but I think we had got to the point where a bit of *complacency had crept in* and we believed that there weren't any ways the application could go wrong that we hadn't seen before. Talking further about this with Dermot he pointed out that this situation was akin to discovering an unknown unknown i.e. we came across a problem that we didn't know we didn't know about since we thought we knew about all of them! Errol Morris wrote http://opinionator.blogs.nytimes.com/2010/06/20/the-anosognosics-dilemma-1/?hp[an interesting article about this recently in the New York Times] in which referenced Dunning and Kruger's paper 'http://en.wikipedia.org/wiki/Dunning%E2%80%93Kruger_effect[Unskilled and Unaware of It: How Difficulties of Recognizing One's Own Incompetence Lead to Inflated Self-assessments]'. The learning for me from this is that while intuition is very useful it's also important to at least be aware that we probably don't know everything and that we may well come across new situations and will have to approach them as a novice again.

intuition-and-quit-thinking-and-look

https://markhneedham.com/blog/2010/06/28/intuition-and-quit-thinking-and-look

2010-06-17 07:00:09

Using real life metaphors

My colleague http://twitter.com/dermotkilroy[Dermot Kilroy] attended the http://skillsmatter.com/event/design-architecture/ddd-exchange-2010[DDD 2010 Exchange] in London last week and one of the ideas that he's been sharing with us from that is that of thinking how the user would solve a given problem without a technological solution i.e. how was something done before computers existed. This encourages us to take a bigger picture view and can actually lead to a much simpler solution than we'd otherwise come up with. An example that came up recently on the project I'm working on was related to how we should handle the situation where two agents tried to access the same users record at the same time. The agents are most likely in the same building. The original technically focused solution was that we needed to lock the record so that if one agent had already opened the record then if another tried to access it they would be unable to. Of course it's technically possible to do this but it's a bit tricky. If there were no computers then there would probably be a physical folder which would contain all the information about the user. It's not actually possible for two agents to both access that at the same time - if one wants to access it and it's already in use then they need not speak to their colleague and work out when they'll be finished with it. If we consider it from this point of view then it becomes clear that putting in the effort to lock the record is unnecessary and that it would make more sense to just pop up a message on the screen informing the agent of the situation and allowing them to decide what to do. This is the solution that we went with. It took way less time to implement and it's worked out fine so far. --- I've retrospectively looked at this situation with my colleagues and realised that it seems to fit this pattern quite well but the way that it was reached originally was by http://www.markhneedham.com/blog/2010/03/26/finding-the-assumptions-in-stories/[looking at the requirement and finding the (initially hidden) reason for it]. It would be interesting to see whether we can rely on this degree of user integrity on a public facing application, I suspect that we can to a much greater degree than we might think.

using-real-life-metaphors

https://markhneedham.com/blog/2010/06/17/using-real-life-metaphors

2010-06-17 00:43:41

Incremental Refactoring: Create factory method

http://twitter.com/dermotkilroy[Dermot] and I spent a bit of time today refactoring some code where the logic had ended up in the wrong place. The code originally looked a bit like this: [source,csharp] ---- public class LookupService { public LookUp Find(UserData userData) { var param1 = GetParam1From(userData); var param2 = GetParam2From(userData); var param3 = GetParam3From(userData); var lookupKey = new LookUpKey(param1, param2, param3); return lookupRepository.Find(lookupKey); } } ---- [source,csharp] ---- public class LookUpKey { private readonly string param1; private readonly string param2; private readonly string param3; public LookUpKey(string param1, string param2, string param3) { this.param1 = param1; this.param2 = param2; this.param3 = param3; } } ---- We wanted to push the logic used to create the params in 'LookUpKey' into the 'LookUpKey' since it seemed that was a more natural place for that behaviour to live. Unfortunately there were also a few other places that were use LookUpKey's constructor so we couldn't just go and change that to take in a 'UserData' otherwise we'd break other places in the code base. My initial thought was that we could http://www.markhneedham.com/blog/2010/04/25/small-step-refactoring-overload-constructor/[overload the constructor] and temporarily have two constructors until we got rid of the original. Dermot pointed out a better approach which was to add a static factory method to 'LookUpKey' and initially push the logic into that method. [source,csharp] ---- public class LookUpKey { ... public static LookUpKey CreateFrom(UserData userData) { var param1 = GetParam1From(userData); var param2 = GetParam2From(userData); var param3 = GetParam3From(userData); return new LookUpKey(param1, param2, param3); } } ---- The service is now much simplified: [source,csharp] ---- public class LookupService { public LookUp Find(UserData userData) { var lookupKey = LookUpKey.CreateFrom(userData); return lookupRepository.Find(lookupKey); } } ---- We can also move tests that were originally indirectly testing the key creation by checking what was passed to the repository to be directly against the 'LookUpKey'. Eventually we want to make LookUpKey's constructor private but for now we've been able to make our change without breaking the code base elsewhere.

incremental-refactoring-create-factory-method

https://markhneedham.com/blog/2010/06/17/incremental-refactoring-create-factory-method

2010-06-10 07:22:38

Retrospectives: Some thoughts

I've worked on two different teams this year which had quite different approaches to retrospectives. In the first team we had a *retrospective at the beginning of every iteration* i.e. once every two weeks and in the second team we tried out the idea of having a *rolling retrospective* i.e. we put up potential retrospective items on the wall and when there were enough of those we discussed them in the standup. The problem we had with the first approach was that we often had the *same types of discussions in each retrospective* and the same types of issues were raised each week so that it seemed almost pointless to even have it in the first place. Several people ended up disliking the idea of even doing a retrospective in the first place because of this. We came up with the idea of having a vote at the beginning of the iteration to decide whether or not we should have one with the idea being that if it didn't seem necessary then we could skip it for one week but would definitely then have the next one. I think this worked reasonably well although it's slightly different to the approach taken by Rolf Knutsen which he discussed at link:[XP2010] in his session on 'Feedback and Retrospectives in Agile Projects'. Rolf suggested that we should have a retrospective at a given time interval and just cut it short if we don't have anything else to discuss. I think I prefer Rolf's suggestion as it's too easy to come to the conclusion that we don't need a retrospective especially if we think that they're a waste of time. I thought the rolling retrospective approach that we took on the second project would work much better as we could just address any problems as soon as they came up but it hasn't worked quite that way from my experience. Several issues which would normally be addressed in a retrospective seemed to go by unchecked and without being resolved. Having a retrospective is a very useful way to *put people in a reflective mindset* where they look for things that aren't being done well or that could be done differently whereas during the day we might not do that. I hadn't appreciated this benefit of a retrospective session and had felt that problems were often being saved up for a retrospective even if they could have been addressed earlier. Fabio previously wrote about the idea of the http://fabiopereira.me/blog/2009/11/15/future-retro-box/[future retrospective box] to ensure that we don't forget anything that happened early on in an iteration when we come to a retrospective but that still doesn't address the problem of only addressing problems in the retrospective! It seems like a combination of having a pre scheduled retrospective with problem solving on a day to day basis is what we need so I'd be interested to know how others have addressed this.

retrospectives-some-thoughts

https://markhneedham.com/blog/2010/06/10/retrospectives-some-thoughts

2010-06-19 22:14:06

Git/Mercurial: Pushing regularly

I was reading a recent blog post by Gabriel Schenker where he discusses http://feedproxy.google.com/~r/LosTechies/~3/h-tL8ABnNkY/git-and-our-friction-points-and-beginners-mistakes.aspx[how his team is making use of Git] and about half way through he says the following: ____ When using Git as your SCM it is normal to work for quite a while -- maybe for a couple of days -- in a local branch and without ever pushing the changes to the origin. Usually we only push when a feature is done or a defect is completely resolved. ____ We've been using Mercurial on the project I'm currently working on over the past few months and although it's a similar tool we've been following a different approach. We've got it setup the same way we would setup Subversion: image::{{<siteurl>}}/uploads/2010/06/dscm.gif[dscm.gif,675] We've been trying to push to the central repository as frequently as possible, just as we would if we were using Subversion. I don't know the Git workflow that well because I haven't used it on a project yet but we've always found that it's beneficial to integrate with code being written by others on the team as frequently as possible. Not doing this can lead to the problems which Martin Fowler outlines in his post about http://martinfowler.com/bliki/FeatureBranch.html[feature branches]. We've tried to ensure that after every commit the build still passes although we do sometimes have broken versions in the code committed locally because we don't run our full test suite before every local check in. Even if a feature isn't completed I still think it's valuable to have what we've done so far checked in and it also helps remove the problem with needing to backup local repositories: ____ Since we are going to work locally potentially for days without pushing to the origin (our central repository) we might well loose our work if we have a hard disk crash or our office is flooded. Thus we need some backup strategy. ____ We just need to make sure the central repository is being backed up and then the danger of losing our work is significantly reduced.

gitmercurial-pushing-regularly

https://markhneedham.com/blog/2010/06/19/gitmercurial-pushing-regularly

2010-06-26 10:03:25

Is 'be the worst' ever limiting?

One of my favourite patterns from Ade Oshineye and Dave Hoover's 'http://apprenticeship-patterns.labs.oreilly.com/[Apprenticeship Patterns]' is 'http://apprenticeship-patterns.labs.oreilly.com/ch04.html#be_the_worst[Be the worst]' which is described as follows: ____ Surround yourself with developers who are better than you. Find a stronger team where you are the weakest member and have room to grow. Be the Worst was the seminal pattern of this pattern language. It was lifted from some advice that Pat Metheny offered to young musicians: "`Be the worst guy in every band you're in.`" Pat's advice struck a chord with Dave, and was one of the reasons he started writing this book. ____ Since I started working at ThoughtWorks it hasn't been that difficult to follow this pattern and I've been the least experienced developer on the majority of teams that I've worked on. I always thought this was a pretty good thing as since I'm always surrounded by people who know much more than I do about various aspects of software development the opportunity to learn is very high. Recently conversations with several different colleagues leave me questioning whether at some stage we need to move away from this pattern, perhaps temporarily to improve skills in other areas. The authors do cover this a bit in the book: ____ Being in a strong team can make you feel as if you are performing better. *The other members of that team will often prevent you from making mistakes, and help you recover from mistakes so smoothly that you won't realize that you may not be learning as much as you think.* It's only when you work on your own that you will see how much your team increases your productivity and realize how much you have learned. ____ == Making technical decisions One area in which I've noticed that this is true is when it comes to making technical decisions on a project. Quite often I find that even though I have an idea of what the solution to a problem should be I end up deferring the decision to someone more senior in the team. I don't think there's a problem with discussing solutions with others in the team but it certainly seems that it would be a better learning experience to be in a situation where I was forced to make the call and then see how things went as a result of that decision. It's certainly possible to engineer a situation where you have to make that type of decision by working on open source projects but it's still useful to get experience on real projects as well. I guess the easiest way is to be on a team where you aren't the worst so that by default you'll end up in a position where you have to make those calls. An alternative is for more senior people to encourage others to make decisions with the knowledge that at least they will be there to help recover the situation if it goes wrong. I've seen some of my colleagues use this approach and it seems to work reasonably well. It's still not a perfect approach though because often someone more experienced may know intuitively that an approach isn't going to work but will struggle to explain why they know that. == Solving problems The other situation where 'be the worst' is perhaps limiting is when it comes to solving problems on a team. At the first sign of being 'stuck' there is a real temptation to ask someone for help even though you might be able to solve the problem alone given enough time. A colleague in my first job suggested that whenever I got stuck it was worth struggling with it for an hour before asking for help. I don't know if that's too prescriptive but there certainly seems to be some merit in the idea. I'd be interested in hearing others' thoughts on this and whether there is in fact a point at which you can grow more by focusing less on learning all the time from others and more on stretching yourself to find situations in which to take the responsibility for making decisions based on that knowledge.

is-be-the-worst-ever-limiting

https://markhneedham.com/blog/2010/06/26/is-be-the-worst-ever-limiting

2010-06-21 21:30:20

iPad: First thoughts

I've had the iPad for about a month now and since my colleagues http://martinfowler.com/bliki/iPad.html[Martin Fowler], http://memeagora.blogspot.com/2010/06/ipad-good-bad-and-ugly.html[Neal Ford] and http://skizz.biz/blog/2010/05/01/thoughts-on-the-ipad/[Chris Stevenson] have already previously written about their experiences with it I thought I'd share the way I'm using it as well. == Twitter I follow a lot of people involved in software development on twitter and come across a lot of interesting articles/blogs that people link to or write. A lot of the time I don't really want to read those posts when I come across them - it would be much better if I could just save them to read later on. Before I had the iPad the way I did this was just by having loads of tabs open on Chrome which tended to result in my computer grinding to a halt because of the memory being consumed by having all those articles open. On the iPad I've been making use of the http://www.osfoora.com/[Osfoora HD] twitter client which is integrated with http://www.instapaper.com/[Instapaper] - an application which allows you to save web pages and then read them offline. I'm online most of the time so I was initially unsure how much value I'd get out of it. I now *read almost every technical article I come across via Instapaper*. I've been posting to twitter way more because it's now so easy to post links there which is the main way that I use it. == Reading technical material Before I bought the iPad I wanted to make sure that I'd be able to read pdfs on it. I have a lot of technical books as pdfs and I've wanted to buy a device that would allow me to read those more easily since I don't like reading pdfs on my laptop unless they're purely coding books. Luckily the http://www.goodiware.com/goodreader.html[Good Reader] application allows me to do exactly that and the ability to crop the pages to be exactly how you want is fantastic and something that I'd never even thought of until I saw it. I use Google Reader for my RSS feeds so none of the RSS reader applications really appealed to me until I came across http://reederapp.com/[reeder] which basically sits on top of Google Reader and downloads all the unread items for offline viewing. It's integrated with twitter as well so I can easily post any interesting things that I read. == Blogging I'm using the http://blogpress.coollittlethings.com/[Blog Press] application which I've found to be better than the Word Press one because it allows me to cut and paste text around the post - I found that I couldn't do this with the Word Press application. Ben Parr has a more extensive post on http://mashable.com/2010/04/03/can-you-blog-from-an-ipad/[his experiences blogging from the iPad] but overall for me it's still not as good as blogging from http://www.red-sweater.com/marsedit/[Mars Edit] on the Mac. It's a bit of a hassle copying links into posts because you need to manually write the Html to do so which is pretty painful in itself but also because you have to keep switching to other applications to get the links in the first place. I have been using BlogPress to quickly copy/paste any interesting quotes from the blogs that I read but more often than not I end up posting it as a draft and then tidying it up with Mars Edit later on. == Videos Watching videos on the iPad is amazing - the picture is obviously much bigger than it would be on an iPhone and it's much more convenient than it would be on a laptop. I watch technical videos now and then for which the http://www.inmethod.com/air-video/index.html[Air Video] application comes in really handy. It can convert any video files you have on your computer to a format that can be watched on the iPad although there are several formats which work without conversion as well. http://www.infoq.com/[InfoQ] and http://skillsmatter.com/[Skills Matter] both frequently post technical presentations and this week http://search.twitter.com/search?q=ndc2010[NDC 2010] has made a lot of the presentations from that conference available to download.

ipad-first-thoughts

https://markhneedham.com/blog/2010/06/21/ipad-first-thoughts

2010-06-09 15:29:44

XP2010: General thoughts

I had the chance to attend the http://xp2010.org/[XP2010 conference] in Trondheim, Norway for a couple of days last week as I was presenting a lightening talk based on a blog post I wrote last year titled 'http://www.markhneedham.com/blog/2009/08/13/challenging-projects-and-the-five-stages-of-grief/[Tough projects and the Kubler Ross Grief Cycle]'. It was interesting to see the way another conference was organised as the only other conference I've attended was QCon which is a much more technical conference. It seemed that a lot of people I spoke to were coming to the same types of conclusions around different software development issues. Notably that it only makes sense to refactor code mercilessly if we're actually working around a specific piece of code and that having a target velocity is fairly pointless but seems to be necessary when working with stakeholders who aren't used to an agile approach. David Anderson did the day's keynote which was titled 'http://multimedie.adm.ntnu.no/mediasite/SilverlightPlayer/Default.aspx?peid=8c850097b1864449ab2f9b035a786211[Catalyzing Lean: Building a Limited WIP Society in Your Organization]'. There were a couple of standout points from this talk for me: * He talked about using a Work in Progress limit with respect to the number of stories that we can have in each swim lane of the story wall and how it can be used to provoke conversation in the team. For example if we notice that there are no items currently being tested then we need to look back to *the place just before that i.e. in development to see where the bottleneck in our process is*. We can then have a conversation about what we can do to fix that. David pointed out that it still takes leadership to make the conversation happen - just using a WIP limit won't fix all our problems. * He also showed the idea of using opportunity cost of delay diagrams showing what happens if deliver something late so that we can see which features really are a priority. This is linked with the idea of each story having a *class of service*. If the client loses a lot of money if a story isn't completed by a certain date then this would become a very high priority story for us to complete and it would have a high class of service. This seems like a much more useful way of working out the priority of stories as we now understand exactly why something is high priority rather than it just being because someone said it is. Another interesting session that I attended was one run as a series of Pecha Kucha's where several high profile agilists described their 'http://multimedie.adm.ntnu.no/mediasite/SilverlightPlayer/Default.aspx?peid=2cb930a5b6444cf8afd7fcd96b7b0c61[Agile Suitcase]' - a list of items that they would take around with them whenever they have to coach the agile approach to software development. There were lots of cool ideas but the most interesting to me was Joshua Kerievsky's where he mentioned that running an *agile readiness assessment* as something that he likes to do with new clients. I've not come across the idea before but it strikes me at least as an approach that would help to get some data on what level agile coaches should start at when working in this type of role. I don't know much more about what it entails but found http://twitter.com/JoshuaKerievsky/status/8081104517[the following on twitter from Kerievsky]: ____ An Agile Readiness Assessment is a time to demonstrate your agility, assess their resistance, explore their future and gain mutual trust. ____ Another interesting idea from a 'Learning and Improvement' lightning talk about 'Agile teams and rescue dogs' was that of *selecting the team member with the least experience as the team leader*. In the rescue dogs organisation the goal of the leader would be to ensure that the right leader was leading the team depending on the situation. The leader would vary depending on the situation and the expertise required which seems like it should fit quite well in agile teams. I've not worked in a team where this approach was taken but it seems like something interesting to try out.

xp2010-general-thoughts

https://markhneedham.com/blog/2010/06/09/xp2010-general-thoughts

2010-06-30 10:46:20

jQuery: Dynamically updating a drop down list

We recently had a requirement to dynamically update a drop down list based on how the user had filled in other parts of the page. Our initial approach was to populate the drop down with all potential options on page load and then add CSS selectors to the options that we wanted to hide. That worked fine in Chrome and Firefox but Internet Explorer seems to ignore CSS selectors inside a drop down list so none of the options were being hidden. We therefore had to try and dynamically add and remove options from the drop down list instead. The list that we initially loaded onto the page was like this: [source,html] ---- <select id="foo" name="foo"> <option value="A" selected="selected">A</option> <option value="B">B</option> <option value="C">C</option> <option value="Not applicable">Not applicable</option> </select> ---- http://twitter.com/christianralph[Christian] eventually came up with the following solution to hide/show those options and select the appropriate one after several hours of trial and error: [source,javascript] ---- // captured on page load var originalFooOptions = $("#foo > option"); ---- [source,javascript] ---- var foo = $("#foo"); var dependentField = $("#fieldFooIsDependentOn"); if(notApplicable()) { foo.children("option[value!='Not applicable']").remove(); foo.val("Not applicable"); } else { foo.empty(); $(originalFooOptions).each(function() { var newOption = $(this).clone(); if ($(this).val() === dependentField.val()) { newOption.attr("selected", "selected"); } newOption.appendTo(foo); }); } ---- Another approach we tried was to try and dynamically update 'foo' rather than removing all the items and then adding them back. Unfortunately we kept getting an 'unspecified error' on Internet Explorer 6 when we tried to set the selected value with that approach.

jquery-dynamically-updating-a-drop-down-list

https://markhneedham.com/blog/2010/06/30/jquery-dynamically-updating-a-drop-down-list

2010-06-15 23:15:30

Fluent NHibernate: Seeing the mapping files generated

We've been fiddling around with http://fluentnhibernate.org/[Fluent NHibernate] a bit over the last couple of days and one of the things that we wanted to do was output the NHibernate mapping files being generated so we could see if they were as expected. I couldn't figure out how to do it but thanks to the help of http://twitter.com/jagregory/status/16210304123[James Gregory], http://twitter.com/trullock/status/16214631489[Andrew Bullock] and http://twitter.com/MatthewErbs/status/16213970597[Matthew Erbs] on twitter this is the code that you need in order to do that: [source,csharp] ---- Fluently .Configure() .Database(MsSqlConfiguration.MsSql2000.ConnectionString("connection string")) .Mappings(m => m.FluentMappings.ExportTo("c:\\directory-to-output-files-to")) .BuildSessionFactory(); ---- The 4th line is the important one here and we can choose which directory the mappings files get outputted to and then check that everything is getting setup as we'd expect.

fluent-nhibernate-seeing-the-mapping-files-generated

https://markhneedham.com/blog/2010/06/15/fluent-nhibernate-seeing-the-mapping-files-generated

2010-06-13 22:35:14

C#: A failed attempt at F#-ish pattern matching

A few weeks ago we had some http://www.markhneedham.com/blog/2010/04/20/functional-c-an-imperative-to-declarative-example/[C# code around calcuations] which had got a bit too imperative in nature. The code looked roughly like this: [source,csharp] ---- public class ACalculator { public double CalculateFrom(UserData userData) { if(userData.Factor1 == Factor1.Option1) { return 1.0; } if(userData.Factor2 == Factor2.Option3) { return 2.0; } if(userData.Factor3 == Factor3.Option2) { return 3.0 } return 0.0; } } ---- I think there should be a more object oriented way to write this code whereby we push some of the logic onto the 'UserData' object but it struck me that it reads a little bit like http://www.markhneedham.com/blog/2010/01/12/f-refactoring-to-pattern-matching/[pattern matching code you might see in F#]. I decided to drive the code to use a dictionary which would store functions representing each of the conditions in the if statements: [source,csharp] ---- public class ACalculator { private Dictionary<Func<UserData, bool>, double> calculations; public ACalculator() { calculations = new Dictionary<Func<UserData,bool>,double> { {u => u.Factor1 == Factor1.Option1, 1.0}, {u => u.Factor2 == Factor2.Option3, 2.0}, {u => u.Factor3 == Factor3.Option2, 3.0} }; } public double CalculateFrom(UserData userData) { var calculation = calculations.Keys.FirstOrDefault(calc => calc(userData)); if(calculation != null) { return calculations[calculation]; } return 0.0; } } ---- It's less readable than it was before and it's not obvious that the adding of the functions to the dictionary needs to be in that order in order for it to work. I've simplified the real example a bit to show the idea but I don't think it works as the best abstraction in this situation either way although it was an interesting experiment.

c-a-failed-attempt-at-f-ish-pattern-matching

https://markhneedham.com/blog/2010/06/13/c-a-failed-attempt-at-f-ish-pattern-matching

2010-06-13 13:37:39

The Refactoring Dilemma

On several of the projects that I've worked on over the last couple of years we've seen the following situation evolve: * The team starts coding the application. * At some stage there is a breakthrough in understanding and a chance to really improve the code. * However the deadline is tight and we wouldn't see a return within the time left if we refactored the code now * The team keeps on going with the old approach * The project ends up going on for longer than the original deadline * It's now much more work to move towards the new solution In the situations I describe the refactorings could have been done incrementally but doing that would take longer than continuing with the original approach and also http://www.markhneedham.com/blog/2010/05/05/consistency-in-the-code-base-and-incremental-refactoring/[leave the code in an inconsistent state]. I think the reason this situation evolves consistently in this manner is because *although we talk about writing maintainable code, delivery is often considered more important*. Pushing out a delivery date in order to refactor code so that it will be easier to work with in the future isn't something that I've seen happen. Pushing a delivery date out is a cost that we can see straight away. On the other hand it's quite difficult to estimate how much of a gain you'll get by refactoring to a more maintainable/easier to test solution and that gain will not be immediate. We therefore end up in the situation where we tend to make major refactorings only if we're going to see a benefit from doing that refactoring before the project ends. In one sense that seems reasonable because we're trying to ensure that we're adding as much value as possible while the client is paying for our time. On the other hand we're making life harder for future maintainers of the code base which may in fact be us! I'd be keen to hear how others handle these types of situations because it feels like this trade off is quite a common one and the way we've dealt with it doesn't seem optimal.

the-refactoring-dilemma

https://markhneedham.com/blog/2010/06/13/the-refactoring-dilemma

2010-06-14 22:46:00

TDD: Driving from the assertion up

About a year ago I wrote a post about http://www.markhneedham.com/blog/2009/06/20/book-club-the-readability-of-tests-growing-object-oriented-software-steve-freemannat-pryce/[a book club we ran in Sydney covering 'The readability of tests'] from Steve Freeman and Nat Pryce's book in which they suggest that their preferred way of writing tests is to drive them from the assertion up: ____ Write Tests Backwards Although we stick to a canonical format for test code, we don't necessarily write tests from top to bottom. What we often do is: write the test name, which helps us decide what we want to achieve; write the call to the target code, which is the entry point for the feature; write the expectations and assertions, so we know what effects the feature should have; and, write the setup and teardown to define the context for the test. Of course, there may be some blurring of these steps to help the compiler, but this sequence reflects how we tend to think through a new unit test. Then we run it and watch it fail. ____ At the time I wasn't necessarily convinced that this was the best way to drive but we came across an interesting example today where that approach might have been beneficial. The test in question was an integration test and we were following the approach of http://www.markhneedham.com/blog/2008/10/30/testing-hibernate-mappings-setting-up-test-data/[saving the test object directly through the NHibernate session] and then loading it again through a repository. We started the test from the setup of the data and decided to get the mappings and table setup in order to successfully persist the test object first. We didn't write the assertion or repository call in the test initially. Having got that all working correctly we got back to our test and wrote the rest of it only to realise as we drove out the repository code that we actually needed to create a new object which would be a composition of several objects including our original test object. We wanted to retrieve a 'Foo' by providing a key and a date - we would retrieve different values depending on the values we provided for those parameters. This is roughly what the new object looked like: [source,csharp] ---- public class FooRecord { public Foo Foo { get; set; } public FooKey FooKey { get; set; } public DateTime OnDate { get; set; } } ---- 'FooRecord' would need to be saved to the session although we would still retrieve 'Foo' from the repository having queried the database for the appropriate one. [source,csharp] ---- public class FooRepository { public Foo Find(Date onDate, FooKey fooKey) { // code to query NHibernate which retrieves FooRecords // and then filters those to find the one we want } } ---- We wouldn't necessarily have discovered this more quickly if we'd driven from the assertion because we'd still have had to start driving the implementation with an incomplete test to avoid any re-work. I think it would have been more likely that we'd have seen the problem though.

tdd-driving-from-the-assertion-up

https://markhneedham.com/blog/2010/06/14/tdd-driving-from-the-assertion-up

2010-06-22 22:51:10

Leadership and software teams: Some thoughts

Roy Osherove wrote a post about a month ago describing the http://5whys.com/blog/the-3-maturity-stages-of-a-software-team-and-how-scrum-fails.html[different maturity levels of software teams] and the strategies that he uses when leading each of these which I found quite interesting. He describes the following states of maturity for a team: * Chaotic Stage -- the state where a team does not possess the skills, motives or ambition to become a mature self managing team. * Mid-Life stage -- where a team possesses some skills for self management and decision making , and can make some of its own decisions without needing a team lead. * Mature stage -- where a team is practically fully self managing and a team leader is mostly a coach rather than a decision maker. Later on in the article he goes on to describe the best approach as a leader of each of these types of team: * The chaotic stage needs a strong, dictator like leadership * The Mid-Life stage needs a coach-dictator * The Mature stage needs a coach My experience over the last few years is that a lot of tech leads seem to start off with the assumption that their team is a chaotic one and that they need to make every decision if a project is going to be successful. Perhaps this is a good place to start but I think it's important to keep checking/iterating your approach to see whether it's still appropriate. Teams tend to improve as time goes on to the point where you don't need to decide as much for the team as you originally did. The dictator approach seems to work more effectively in smaller teams in non political environments where it is possible for the tech lead to be around frequently and have the time available to make the majority of decisions. When we don't have that i.e. it's a bigger team or political situation then it doesn't seem to scale since that person becomes the bottleneck as they don't have the time available to get involved in every decision that gets made. Even if it does work *the team leader can end up discouraging people from taking responsibility by always taking it themselves* instead of coaching people so that they become more effective at decision making. Ironically this isn't what they were trying to achieve but not allowing people to be responsible for decisions means that they'll be less likely to take responsibility in the future i.e. it's a http://en.wikipedia.org/wiki/Self-fulfilling_prophecy[self fulfilling prophecy]. I appreciate that there is a fine balance to strike between allowing people to self manage and potentially make mistakes and wanting to keep things progressing by making a decision which you intuitively know is right but somehow we need to find it! An interesting quote from a http://www.brazencareerist.com/2010/06/11/what-the-celtics-win-can-teach-you-about-leadership[post by Leanne Chase about leadership] sums up my current opinion on where we should aim to drive software teams towards: ____ Finally, I heard the ESPN announcer say at one point last night after a Kobe Bryant 3-pointer that Kobe was essentially telling the team -- no worries, get on my back, I'll carry you. That sort of mentality must be exhausting in basketball, at work and in life. *I don't know about you but I would much rather have a team mentality*. ____ I've never yet had to lead a software team so all the above is from observation of how others have carried out this role. It's interesting picking up ideas from each of them and hopefully I'll be able to make use of these if I'm in that role in the future.

leadership-and-software-teams-some-thoughts

https://markhneedham.com/blog/2010/06/22/leadership-and-software-teams-some-thoughts

2010-06-22 22:27:58

C#: StackTrace

http://twitter.com/dermotkilroy[Dermot] and I were doing a bit of work on a mini testing DSL that we've been writing to try and make some of our interaction tests a bit more explicit and one of the things that we wanted to do was find out which method was being called on one of our collaborators. We have a stub collaborator which gets injected into our system under test. It looks roughly like this: [source,csharp] ---- public class StubCollaborator : IGotForcedToCollaborate { public double Method1() { return CannedValue(); } public double Method2() { return CannedValue(); } private double CannedValue() { return 10; } } ---- We wanted to try and capture which of the methods on that object had been called by our system under test and then assert on that value from our test. While trying to work out how to do this we came across the 'StackTrace' object. We use it like so to work out which public method on that object has been called: [source,csharp] ---- public class StubCollaborator : IGotForcedToCollaborate { private MethodBase methodCalled; .. private double CannedValue() { methodCalled = new StackTrace().GetFrames() .Select(f =>f.GetMethod()) .Where(m => m.DeclaringType.Name == GetType().Name) .Where(m => m.IsPublic) .First() return 10; } public string GetMethodCalled() { return methodBase.Name; } } ---- We needed to only find the public methods because 'CannedValue' was showing up on the stack trace and since it's private this was an easy way to exclude it. I'm sure there are other ways to get this type of information but we were able to solve our problem really quickly with this solution.

c-stacktrace

https://markhneedham.com/blog/2010/06/22/c-stacktrace

2010-06-25 23:32:36

Mercurial: Only pushing some local changes

One problem we've come across a few times over the last couple of months while using Mercurial is the situation where we want to quickly commit a local change without committing other local changes that we've made. The example we came across today was where we wanted to make a change to the build file as we'd made a mistake in the target that runs on our continuous integration server and hadn't noticed for a while during which time we'd accumulated other local changes. The following is a rough diagram of the situation we had: image::{{<siteurl>}}/uploads/2010/06/mercurial1.jpg[mercurial.jpg,441] We had multiple file changes in our working directory which hadn't yet been checked in to the local repository or the central repository. We wanted to push just the change in blue. My initial thought was that I could check in just that one file into our local repository and then push it to the central one. [source,text] ---- hg ci -m "mark: updating build file to fix build" -A /path/to/build.file ---- I then wanted to push that change but when I went to do so I realised that they were other incoming changes which we hadn't yet integrated with. In order to integrate with those changes we need to make sure that we don't have any locally uncommitted changes which of course in this scenario we do since we deliberately chose not to check in some of our local changes. One way around this would be to just force the push and ignore the need to integrate with the remote changes but that doesn't seem the right approach to me but I'm not sure what is. We ended up just checking in everything we had locally, commenting out the bits that we were currently working on, merging with the remote changes and then pushing everything to the remote repository. That's obviously a really poor way of solving the problem so I'd be interested in what a good way to solve this problem would be!

mercurial-only-pushing-some-local-changes

https://markhneedham.com/blog/2010/06/25/mercurial-only-pushing-some-local-changes

2010-12-20 21:23:39

India Cultural Differences: Stretched work day

A couple of months ago I briefly touched on http://www.markhneedham.com/blog/2010/10/17/distributed-agile-cultural-differencesexpectation-disconnect/[the very stretched days I've experienced while working in India]. This is in contrast to what I've experienced in the UK and Australia where the day was much more time boxed and tended to go from 9am to 6pm. At the moment we also have a call with colleagues in Chicago at 9pm for about 30-45 minutes so the day has now stretched out until nearly 10pm. We start with our standup at 10am. We're certainly not working all of that time but you're never really done for the day until the call has finished. With or without the call I've found it a much more difficult environment to work in than anything I've experienced before. I frequently get to the end of the day and can't really determine whether I've actually achieved anything because I've started and stopping working so many times. I find I get distracted really easily working in our Pune office so I need my pair to be quite vocal about keeping focus! One disadvantage of the stretched work day which I've only noticed recently is that you can become less reflective about the code because you don't get as much time where you're awake but not directly working with the code. I frequently came up with possible ways to improve code bases on previous projects while playing around with something else or watching a presentation or similar. I find that I play around with stuff outside the project less than I have previously partly because I've lost the 4 to 5 hour stretch that I used to have each evening. On the positive side you get to know the people you're working with pretty well because you're with them for a lot more hours per day. Not every project in the office has hours as stretched as mine but it certainly seems to be more common in India than elsewhere.

india-cultural-differences-stretched-work-day

https://markhneedham.com/blog/2010/12/20/india-cultural-differences-stretched-work-day

2010-12-20 08:58:35

Distributed Agile: Bringing onshore people offshore

For the last two weeks we've had a ThoughtWorks colleague from the onshore team working with us in Pune and it's been really cool having someone who has been working on 'the other side'. In my time in India there seem to have been many more people going from offshore to onshore than the other way around but based on this experience I don't think that should necessarily be the case. The typical argument in favour of someone from offshore going on site is that they're able to interact with the client and onshore people and pick up any required context which they'll then be able to share with their colleagues offshore. The impact of those trips is perceived to be higher than what would be achieved by having someone do the reverse trip. Having seen the interactions my US colleague was able to have with people has made me realise that perhaps that perception isn't as true as it seems. A big part of the success of distributed projects seems to be about the ability for the two sides to communicate effectively and having someone who interacts daily with people that we only know by phone/email is very helpful. My colleague was able to provide context about our client and what they're trying to do in a way that I don't think someone going onshore for a couple of weeks would be able to do. Sometimes it's felt like we're just building something because we were told to do so and my colleague has been able to explain the underlying thinking and what we're trying to do with the system. I think he's also found it a pretty useful experience in terms of learning about the strengths and weaknesses of the team and what the real pain points are. Everyone offshore now knows one person onshore personally which should be useful for any future interactions.

distributed-agile-bringing-onshore-people-offshore

https://markhneedham.com/blog/2010/12/20/distributed-agile-bringing-onshore-people-offshore

2010-12-27 19:15:51

Vim: Learnings so far

I've been using Vim instead of http://www.jetbrains.com/ruby/index.html[RubyMine] for the last month or so and it's been interesting observing the way that I browse code as I add plugins to make my life easier. == Between files I generally don't know exactly where in the folder structure different files live since I'm used to being able to search by just the name i.e. RubyMine's Ctrl-N http://yehudakatz.com/2010/07/29/everyone-who-tried-to-convince-me-to-use-vim-was-wrong/[Yahuda Katz wrote a blog post] earlier in the year where he listed some of the plugins he's been using - one of which is called https://wincent.com/products/command-t[Command-T] and allows exactly this functionality. I also quite like the ability to quickly access files that I've recently opened i.e. files which are in the Vim buffer or RubyMine's Ctrl-E. The http://www.vim.org/scripts/script.php?script_id=1984[FuzzyFinder] plugin provides that functionality. I've also tagged all my Ruby gems and the source code of the project using http://ctags.sourceforge.net/[Exuberant CTags] which then allows http://blog.bojica.com/2010/06/27/ctags-and-vim-for-ruby-on-rails-development[easy browsing to methods/classes]. == Inside files I've noticed that the way I browse inside files has changed since I started using Vim. I used to just scroll around files using the mouse but now I find myself moving around a file by line numbers instead. A lot of the commands for file editing in Vim are based on moving/changing/deleting to a particular symbol so you become almost a human parser when reading a line of text. I found/am finding the following quite useful for learning Vim shortcuts: * http://www.viemu.com/vi-vim-cheat-sheet.gif[Vim cheat sheet] * http://tnerual.eriogerg.free.fr/vimqrc.pdf[Vim quick reference card] * http://jmcpherson.org/editing.html[Efficient Editing with Vim] When coding in Java/C# I rely quite heavily on auto complete to tell me what methods I have available to me on a certain object. Although I don't use it that frequently the http://www.vim.org/scripts/script.php?script_id=182[SuperTab] plugin works reasonably well when you do need help. http://twitter.com/#!/just3ws[Mike] pointed out https://github.com/spicycode/Vimlander-2-The-Quickening[Vimlander-2-The-Quickening] which has some of the plugins I mentioned and several others ready to use.

vim-learnings-so-far

https://markhneedham.com/blog/2010/12/27/vim-learnings-so-far

2010-12-27 14:16:09

India Cultural Differences: Hierarchy

One of the more interesting differences between Indian culture and my own is that in India there appears to be more adherence to a hierarchy than I've experienced before. ThoughtWorks tries to keep a reasonably flat hierarchy so I think the idea of hierarchy would be much more obvious if I was working at one of the big Indian services organisations. Between peers conversations don't seem to play out any differently but someone in a position of authority is more likely to be able to get their opinion across and accepted with less resistance than they might experience without that authority. On several occasions it's been only me and maybe a couple of others involved in discussions if someone in an authority position expresses an opinion assertively. I asked several colleagues why this was and they pointed out that when they'd previously expressed an opinion in that type of situation it didn't have any impact so they'd stopped doing so. I can't decide whether or not it really matters - the situations where authority has any influence are relatively infrequent - but it would be cool if all opinions were assessed just on merit. Of course these are just my observations so it'd be interesting to hear others' experiences.

india-cultural-differences-hierarchy

https://markhneedham.com/blog/2010/12/27/india-cultural-differences-hierarchy

2010-12-10 03:47:13

Why am I working in India?

A few colleagues have asked me why I chose to work in India so I thought it would be interesting to explore what it is that appealed to me about working here. I've come to the conclusion that there were 2 main drivers for me: == The buzz of the ThoughtWorks office I was in Bangalore in 2006 when I attended http://www.thoughtworks.com/thoughtworks-university[ThoughtWorks University] and one of the things that stood out for me was the atmosphere in the Diamond District office. It was really buzzing with noise and I thought it would be really fun to have the chance to work in that type of environment. I don't think Pune reaches the levels that I remember from Bangalore but it's still quite a fun and relaxed place to work. I haven't noticed as much inter team communication in Pune as I did in Bangalore which is interesting and perhaps means we aren't getting the full benefit of having so many people in the same room. My overall feeling is that the office here acts much more as a central point for people during the week and they'll spend much more time in it than I've seen in other countries. The cool thing about that is that nearly every evening there's people to hang out and go to dinner with! == Curiosity about offshore development I've spent the past four years working predominantly in co-located teams and had always heard that it was much more difficult to work in a distributed team. I have worked in one team which was distributed between Melbourne and Sydney but it's not quite the same as working in a team distributed in two countries with big time zone differences as well. From hearing a couple of talks about distributed agile I was coming to the conclusion that big/complex projects would typically be off-shored since an offshore team could add people much more rapidly. I haven't really noticed that my current project is any more complicated technically or domain wise than others that I've worked on but it certainly has significantly more people than anything else I've worked on before. http://www.markhneedham.com/blog/2010/10/27/distributed-agile-communication/[As I've mentioned before], I do admire the patience of my colleagues in dealing with the slower feedback cycles/lack of face to face client communication which I've found very frustrating.

why-am-i-working-in-india

https://markhneedham.com/blog/2010/12/10/why-am-i-working-in-india

2010-12-26 00:04:54

Theory of Constraints: Blaming the bottleneck

I've been reading http://www.amazon.com/Goal-Process-Ongoing-Improvement/dp/0884271781/ref=sr_1_1?ie=UTF8&qid=1293318612&sr=8-1[The Goal] over the last week or so where Eliyahu Goldratt describes the http://en.wikipedia.org/wiki/Theory_of_Constraints[theory of constraints] as a philosophy for allowing organisations to continually achieve their goal. Goldratt goes on to describe http://en.wikipedia.org/wiki/Bottleneck[bottlenecks] - resources which have a capacity less than the capacity being demanded of the system. The capacity of the system cannot be higher than that of the bottleneck which means that we need to find a way to optimise the bottlenecks in any system. In a software context I think an agile/lean approach to delivery is quite effective for allowing us to quickly see where the bottlenecks are - this will typically be where the cards are piling up on the wall. The problem with this is that we can frequently end up blaming the people working in the role where the bottleneck seems to be. On the projects I've worked on stories tend to get stuck in QA since there are significantly less testers than developers on any team. Since blaming isn't particularly useful Goldratt suggests two ways to optimise bottlenecks: . Don't waste time of the bottleneck ** Processing defective parts ** Sitting idle ** Work on unnecessary parts . Take load off the bottleneck The bottleneck ends up sitting idle when we wait until we've completely finished a story before showcasing what we've done to the testers. It's frequently the case that we can QA test a piece of functionality before the whole thing is complete. We can avoid processing defective parts in this case by talking through a story with a tester before we start and thinking of the bugs we might create in development. Unit testing can also be useful for reducing defects. I'm not sure how working on unnecessary parts would work most of the time - perhaps testing features which aren't required for the next release would be an example of this. We can achieve the second to an extent by ensuring that we have more scenarios automated so that a tester doesn't need to manually cover them each time. We could also get developers to temporarily take on the role of a tester to increase capacity but from my experience that doesn't seem to work out as well in practice as it seems like it should.

theory-of-constraints-blaming-the-bottleneck

https://markhneedham.com/blog/2010/12/26/theory-of-constraints-blaming-the-bottleneck

2010-12-07 05:01:44

Ruby: One method, two parameter types

One interesting thing that I've noticed while coding in Ruby is that due to the dynamicness of the language it's possible to pass values of different types into a given method as parameters. For example, I've recently come across a few examples of methods designed like this: [source,ruby] ---- def calculate_foo_prices(foos) ... [foos].flatten.each do |foo| # do something end end ---- This allows us to use the method like this: [source,ruby] ---- # foos would come in as an array from the UI foos = [Foo.new, Foo.new, Foo.new] calculate_foo_prices(foos) ---- Or like this: [source,ruby] ---- calculate_foo_prices(Foo.new) ---- It becomes quite confusing to understand why what is supposedly already a collection is being put inside an array on line 3 of the first example when you first read it. An alternative would be to pull out a different method for calculating the price of the single Foo: [source,ruby] ---- def calculate_foo_price(foo) calculate_foo_prices([foo]) end ---- And then simplify the original method: [source,ruby] ---- def calculate_foo_prices(foos) ... foos.each do |foo| # do something end end ---- While writing this I was thinking that another way could be to change the original method to look like this by using the splat operator: [source,ruby] ---- def calculate_foo_prices(*foos) ... foos.each do |foo| # do something end end ---- Which means that we can use the same method for both situations: [source,ruby] ---- calculate_foo_prices(Foo.new) ---- [source,ruby] ---- # foos would come in as an array from the UI foos = [Foo.new, Foo.new, Foo.new] calculate_foo_prices(*foos) ---- I'm guessing the latter is more idiomatic Ruby or perhaps there's another way I'm not aware of yet?

ruby-one-method-two-parameter-types

https://markhneedham.com/blog/2010/12/07/ruby-one-method-two-parameter-types

2010-12-01 17:56:51

Ruby: Exiting a 'loop' early

We recently had a problem to solve which at its core required us to iterate through a collection, look up a value for each key and then exit as soon as we'd found a value. The original solution looped through the collection and then explicitly returned once a value had been found: [source,ruby] ---- def iterative_version v = nil [1,2,3,4,5].each do |i| v = long_running_method i return v unless v.nil? end v end ---- [source,ruby] ---- def long_running_method(value) puts "inside the long running method with #{value}" return nil if value > 3 value end ---- Which we run like so: [source,ruby] ---- p "iterative value is #{iterative_version.to_s}" ---- This prints the following when we run it: [source,text] ---- inside the long running method with 1 "iterative value is 1" ---- I figured there must be a more functional way to solve the problem and I eventually came up with this: [source,ruby] ---- def functional_version [1,2,3,4,5].map {|i| long_running_method i }.find { |i| !i.nil? } end ---- Which prints the following when we run it: [source,ruby] ---- inside the long running method with 1 inside the long running method with 2 inside the long running method with 3 inside the long running method with 4 inside the long running method with 5 "functional value is 1" ---- The problem is that collections in Ruby are eager evaluated so we evaluate every single item in the collection before we get the first non nil value. Luckily the http://flori.github.com/lazylist/[lazylist] gem comes to our rescue and allows us to solve the problem in a functional way: [source,ruby] ---- require 'lazylist' def lazy_version lazy_list([1,2,3,4,5]).find { |i| !i.nil? } end def lazy_list(values) list(long_running_method(values.first)) { lazy_list(values - [values.first]) } end ---- Running that gives us this: [source,text] ---- inside the long running method with 1 "lazy value is 1" ---- I've never come across a problem where I needed to use a lazy list but finally I have and I think the version which uses it is pretty neat.

ruby-exiting-a-loop-early

https://markhneedham.com/blog/2010/12/01/ruby-exiting-a-loop-early

2010-12-24 18:12:51

India Cultural Differences: Language

For the majority of the time that I've spent in Pune so far language hasn't been a big deal at all but there are a couple of differences that I didn't initially anticipate. == The local language While the official office language is English my colleagues seem more comfortable talking to each other in Hindi so quite frequently the conversation will move into Hindi if someone isn't directly speaking to me. I initially found it tremendously frustrating that I couldn't understand what was going on in group discussions and while I did point it out people don't realise when they're switching into Hindi so it's a difficult problem to fix. My brain now seems to zone out when people stop speaking in English so I don't even notice that it's happened unless I observe myself becoming much more aware of my surroundings because I've tuned out and started looking around the room. An interesting side effect is that I don't pick up as much information from http://alistair.cockburn.us/Osmotic+communication[osmotic communication] as I have done when working in the UK/Australia because conversations between my colleagues have a 50/50 chance of being in Hindi. I guess the easiest way for me to fix the problem would be to take lessons in Hindi but I've never got around to that. Another interesting thing to note is that it's not only me who doesn't understand Hindi but also people from the Chennai office. == Swearing Swearing in India is a bit of a taboo which is quite strange for me because in the UK/Australia there's not really any stigma attached to the majority of words. There are a few people who are more easily offended so you have to be a bit more careful about word selection when speaking with them. Having said that, with the majority of people that I've worked with the language that I use makes no difference at all.

india-cultural-differences-language

https://markhneedham.com/blog/2010/12/24/india-cultural-differences-language

2010-12-15 19:08:56

India Cultural Differences: Tolerance/Patience

Some colleagues have been asking me recently what cultural differences I've noticed working in India compared to my experiences in the UK and Australia and one of the biggest differences by far is the amount of tolerance and patience people here have compared to me. These attributes seem to show themselves in roughly two situations: == With respect to the environment We've had some building work done in the Pune office recently which has meant that there's been extremely high volume drilling being done to the extent that you can barely hear someone who's sitting a couple of metres away. I was quite amazed to watch colleagues continuing with their conversations as if nothing had happened - I was totally distracted and had to go and sit in a meeting room away from the noise. I've talked to a few colleagues about this and they pointed out that they're pretty used to the environment not being perfect from previous experiences at school or just in general life situations so it's not such a big deal. I find it interesting to see how influential the environment (as far as I can see) can be on some fairly fundamental personality traits. == With people People here seem to give others the benefit of the doubt much more frequently than I've seen elsewhere. I've noticed this especially when communicating with people in remote locations. I worked on a project distributed between two cities a couple of years ago and the distrust between the two locations was fairly evident in all methods of communication. I haven't noticed the same thing here as much and the majority of the time people don't react at all to something which I would consider to be inflammatory. There's also much more tolerance for repetition of information than I've noticed elsewhere. I get very frustrated if I hear the same thing repeated multiple times but the majority of people here don't seem too bothered by that and take it in their stride. == It's a strength not a weakness I think that people here sometimes don't recognise that their tolerance/patience is a strength rather than a weakness and I've frequently heard conversations where what I consider being tolerant is seen to be 'rolling over'. I don't necessarily think the 'western' style of presenting an argument very assertively/confidently is the only way to go and I've seen colleagues successfully communicate while using their own style. Of course these are massive generalisations based on my observations so it'd be interesting to see what other people who've had the chance to work in both cultures/environments think.

india-cultural-differences-tolerancepatience

https://markhneedham.com/blog/2010/12/15/india-cultural-differences-tolerancepatience

2010-12-12 08:11:31

Distributed Agile: Other observations

Some of the difficulties of working in an offshore environment were clear to me before I even came to India but I've come across a few others lately which I either didn't think about before or didn't realise how annoying they were! == Getting data from the client's network For several of the stories that we've been working on lately we needed to make use of huge amounts of reference data residing on the client's network. If we were working in the same building then it would be very quick to download that and use it for whatever we needed. Working offshore we've been downloading those dumps via scp which can take around 5 hours for a 300 MB file. I don't know a lot about networks but I'm told that it would be possible to create a direct VPN between our office and the client's which would allow the transfer of this type of data to be much quicker. == Dealing with internal IP addresses Since we're working on a different network any IP addresses for test environments of external systems aren't directly accessible. We have a box setup in our office connected via a VPN to the client's network which forwards requests we make to local IP addresses to the equivalent IPs on the client's network. There have been several occasions when there have been internal IP addresses hard coded in our configuration files and we've completely forgotten that we need to create an equivalent local address. It's not a big time waster but it's something you wouldn't have to think about if working in the client's office. == Slowness of the feedback cycle It's obvious that if the client is in a different country to you that there will be less face to face communication but I didn't realise how frustrating that would be. If you have an observation about a piece of functionality and the offshore BAs don't have the answer then you'll most likely have to wait 5/6 hours to speak to someone onshore about it or write your thoughts in an email. While both of those are viable options I find my motivation to express my opinion is much lower if I can't get feedback straight away. As a result of this communication lag I think that we spend more time building functionality which doesn't necessarily add a lot of value.

distributed-agile-other-observations

https://markhneedham.com/blog/2010/12/12/distributed-agile-other-observations

2010-12-12 07:59:17

Bugs: Prioritising by bucket

At a lot of organisations that I've worked there is a tendency to prioritise bugs by a priority bucket. We might therefore have priority buckets 1-4 where the bucket number indicates how important the bug is to fix and then any buckets ranked below 4 would not be fixed but would be logged anyway. From what I've noticed this isn't a particularly effective way of managing bugs. To start with there tend to be a lot of discussions around what the priority of each bug should be where a QA will argue that it should be a higher priority while a developer disagrees. I'm not convinced that we actually get any value out of these discussions. Another thing that I've noticed recently is that developers are much less motivated to fix bugs if they've been put in Bucket 4 rather than one of the higher priority buckets. I think a better way to solve this problem is just to *store whether or not we're going to fix the bug before the next release or not*. We can then list them in the order that they should be fixed as well. It's a much simpler binary categorisation and there's no stigma attached to picking up a 'low priority' bug because each bug picked up is just the next one on the list. We did this on one project I worked on and it was interesting to notice the number of bugs the client was completely disinterested in fixing before the next release because they just weren't that important. I'm not convinced the same thing would have happened if we'd categorised by bucket. The added benefit of this approach is that we can avoid discussions where we identify how many bugs of each category have been identified each iteration. In some organisations the ability to detect higher priority bugs is an indication of the quality of QA work but I don't believe this is a useful metric to track.

bugs-prioritising-by-bucket

https://markhneedham.com/blog/2010/12/12/bugs-prioritising-by-bucket

2010-12-13 21:29:02

Technical implementation heavy stories

Earlier this year I wrote about http://www.markhneedham.com/blog/2010/03/26/finding-the-assumptions-in-stories/[some of the problems that we can run into when we have implicit assumptions in stories] and another problematic approach I've seen around this area is where we end up with stories that are very heavily focused on technical implementation. Initially this seems like it will work out pretty well since all the developer then needs to do is follow the steps that have been outlined for them but from my experience it seems to create more problems than it solves. The first problem is that it seems to create a *non thinking mindset* and the more detail there is the less it feels like you need to think. This isn't a great position to end up in since there are frequently multiple ways to solve any problem and we will probably exclude those from our thinking if we're just following instructions. As a developer part of the skill-set is knowing how to implement a solution but an equally important part is being able to *suggest alternative approaches* which may be more appropriate in a given context. Technically heavy stories only encourage the former as the developer may now assume that any other options have somehow already been considered. In addition to this it tends to take a lot of time to explain these types of stories which presumably also means that it takes longer to write them up as well. I've also found that it's more difficult to understand these types of stories since it's not initially obvious why you need to implement the technical approach that has been described. Another thing which I hadn't appreciated until quite recently is that it's very boring/demotivating to implement stories if all the problems have already been 'solved' before you even start. Given that it's somewhat inevitable that we may end up with a story that is shaped like this the key seems to be not to get frustrated by the story but rather to try and find out what the underlying intent of the requirement actually is.

technical-implementation-heavy-stories

https://markhneedham.com/blog/2010/12/13/technical-implementation-heavy-stories

2010-12-14 18:04:18

Ask someone vs work it out yourself

Back in 2007/2008 when I worked on my first couple of projects at ThoughtWorks I always found it strange how frequently my colleagues would try and figure something out themselves rather than asking someone else (who already knew how to do it) how to do it. Fast forward to 2010 and I find myself being the one encouraging people to figure things out themselves. There's still merit in communicating with colleagues when we've tried to work out how to do something and haven't managed to figure it out but it's also useful to not have this as our default mode. This seems to be particularly the case when working offshore if we have a reasonable idea that someone onshore might be able to help us with something we're working on. We can stop and wait for them to start their work day but frequently we'll be able to make some progress without their input and then if we really do need help we can always get it later on. While someone else can often solve a problem quicker than us the disadvantage is that we don't tend to learn by watching others but rather by struggling ourselves. In my first job it was suggested to me by a colleague that I should try and solve any problems alone for an hour before asking someone else for help. Having said that I do think it's still useful to know which people in a team have expertise in certain areas so that we can ask them for help when required.

ask-someone-vs-work-it-out-yourself

https://markhneedham.com/blog/2010/12/14/ask-someone-vs-work-it-out-yourself

2010-12-22 23:32:46

Communication when it's not going your way

I've been reading some of the articles written about the disruption caused by the snow across Europe and I found one quote in http://www.telegraph.co.uk/topics/weather/8214205/UK-snow-flights-cancelled-and-roads-closed-as-Arctic-weather-tightens-grip.html[The Daily Telegraph] by Phillip Hammond particularly interesting ____ "I think whilst people are obviously deeply upset about the inconvenience, particularly at this time of year, of having their travel plans disrupted, most of *what I am hearing is a sense of outrage about the way they were then treated when they were stranded at Heathrow airport.*" ____ I was stuck in Brussels Airport for almost a day and the communication by just about everybody there was as non existent as it sounds like it's been at Heathrow. I think part of the reason for that was that the people in charge didn't know what was happening and part of it was because they didn't want to give people bad news. I've noticed the same type of thing happen in organisations when there there's something bad to be communicated or an unpopular decision has been made and needs to be explained. Often in this situation there will be no further communication because *it's assumed that people won't react well to that communication*. I don't think that's actually an accurate assumption and in addition not communicating is actually quite a dangerous thing to do because it then puts people in the position that they will now guess why certain things are being done. More often than not those guesses will be more damning of people in leadership positions than they deserve. Whenever I've seen someone in a leadership position explain what's actually going on (and the thinking behind it) the response of the people receiving the message has always been much more reasonable than they expect it to be. I think the people in charge of communicating what was going on in the airports would have had similar results if they'd only communicated something!

communication-when-its-not-going-your-way

https://markhneedham.com/blog/2010/12/22/communication-when-its-not-going-your-way

2019-03-27 06:42:00

Neo4j: From Graph Model to Neo4j Import

In this post we're going to learn how to import the https://aminer.org/citation[DBLP citation network^] into Neo4j using the https://neo4j.com/docs/operations-manual/current/tools/import/[Neo4j Import Tool^]. In case you haven't come across this dataset before, Tomaz Bratanic has https://tbgraph.wordpress.com/2018/09/09/article-recommendation-system-on-a-citation-network-using-personalized-pagerank-and-neo4j/[a great blog post^] explaining it. The tl;dr is that we have articles, authors, and venues. Authors can write articles, articles can reference other articles, and articles are presented at a venue. Below is the graph model for this dataset: image::{{<siteurl>}}/uploads/2019/03/citation-model.svg[] Tomaz then goes on to show how to load the data into Neo4j using Cypher and the APOC library. Unfortunately this process is quite slow as there's a lot of data so, since I'm not as patient as Tomaz, I wanted to try and find a faster way to get the data into the graph. At a high level, the diagram below shows what happens when we import data using Cypher: image::{{<siteurl>}}/uploads/2019/03/transaction.png[] When we use the Neo4j import tool we're able to skip that middle bit: image::{{<siteurl>}}/uploads/2019/03/no-transaction.png[] We skip the whole transactional machinery, so it's much faster. We do pay a price for that extra speed: . We can only use this tool to create a brand new database . If there's an error while it runs we need to start again from the beginning . We need to get our data into the format that the tool expects If we're happy with that trade off then Neo4j Import is the way to go. The raw data is in JSON files, so our process will be as follows: image::{{<siteurl>}}/uploads/2019/03/workflow.png[] We don't have to use a Python script to transform the raw data, but that's the scripting language I'm most familiar with so I tend to use that. Before we look at the script, let's explore the JSON files that we need to transform. Below is a sample of one of these files: ++++ <script src="https://gist.github.com/mneedham/fea6c05086637730874c25ec298fc5aa.js"></script> ++++ Each row contains one JSON document and, as we can see, these documents contain the following keys: * `abstract` - the abstract of the article * `authors` - the authors who wrote the article * `n_citation` - we won't use this key * `references` - the other articles that an article cites * `title` - the title of the article * `venue` - the venue where the article was published * `year` - the year the article was published * `id` - the id of the article For each of these JSON documents we're going to create the following graph structure: [source,cypher] ---- (:Article)-[:VENUE]->(:Venue) (:Article)-[:AUTHOR]->(:Author) (:Article)-[:CITED]->(:Article) ---- * where `()` indicates a node i.e. `(:Article)` means that we have a node with the label `Article` * and `[]` indicates a relationship i.e. `[:VENUE]` means that we have a relationship with the type `VENUE` The properties from our JSON document will be assigned as follows: * `id`, `title`, `abstract`, `year` - `Article` node * `authors` - one `Author` node per value in the array, with the value assigned to the `name` property * `references` - one `Article` node per value in the array, with the value assigned to the `id` property * `venue` - one `Venue` node per value in the array, with the value assigned to the `name` property So we know what data we're going to extract from our JSON file, but what should the CSV files that we create look like? The Neo4j Import Tool expects separate CSV files containing the nodes and relationships that we want to create. For each of those files we can either include a header line at the top of the file, or we can store those header lines in a separate line. We're going to take the latter approach in this blog post. The fields in those headers have a very specific format, it's almost a mini DSL. Let's take the example of the `Author` and `Venue` nodes and the corresponding relationship. We'll have three CSV files. The node headers will look like this: articles_header.csv [source, text] ---- id:ID(Article),title:string,abstract:string,year:int ---- venues_header.csv [source, text] ---- name:ID(Venue),name:string ---- For node files one of the fields needs to act as an identifier. We define this by including `:ID` in that field. In our example we're also specifying an optional node group in parentheses e.g. `:ID(Article)`. The node group is used to indicate that the identifier only needs to be unique within that group, rather than across all nodes. And the relationships header will look like this: article_VENUE_venue_header.csv [source, text] ---- :START_ID(Article),:END_ID(Venue) ---- Again our header fields need to contain some custom values: * `:START_ID` indicates that this field contains the identifier for the source node * `:END_ID` indicates that this field contains the identifier for the target node As with the nodes, we can specify a node group in parentheses e.g. `:START_ID(Article)`. So now we know what files we need to create, let's have a look at the script that generates these files: ++++ <script src="https://gist.github.com/mneedham/51bdaaa1d4ec5b8ec9676462817a4b87.js"></script> ++++ There's nothing too clever going here - we're iterating over the JSON files and writing into CSV files. For the relationships we write those directly to the CSV files as soon as we see them. For the nodes we're collecting those in in-memory data structures to remove duplicates, and then we write those to the CSV files at the end. Let's have a look at a subset of the data in those CSV files: ++++ <script src="https://gist.github.com/mneedham/1ace09bd1ef4b48418665ce928025154.js"></script> ++++ This script takes a few minutes to run on my machine. There are certainly some ways we could improve the script, but this is good enough as a first pass. Now it's time to import the data into Neo4j. If we're using the Neo4j Desktop we can access the Neo4j Import Tool via the 'Terminal' tab within a database. image::{{<siteurl>}}/uploads/2019/03/terminal.png[] We can then paste in the script below: ++++ <script src="https://gist.github.com/mneedham/466cb70f3703ec48840d46e043b25406.js"></script> ++++ Let's go through the arguments we're passing to the Neo4j Import Tool. The line below imports our author nodes: ``` --nodes:Author=${DATA_DIR}/authors_header.csv,${DATA_DIR}/authors.csv ``` The syntax of this part of the command is `--nodes:[:Label]=<"file1,file2,...">`. It treats the files that we provide as if they're one big file, and the first line of the first file needs to contain the header line. So in this line we're saying that we want to create one node for each entry in the `authors.csv` file, and each of those nodes should have the label `Author`. The line below creates relationships between our `Author` and `Venue` nodes: ``` --relationships:VENUE=${DATA_DIR}/article_VENUE_venue_header.csv,${DATA_DIR}/article_VENUE_venue.csv ``` The syntax of this part of the command is `--relationships[:RELATIONSHIP_TYPE]=<"file1,file2,...">` Again, it treats the files we provide as if they're one big file, so the first line of the first file must contain the header line. In this line we're saying that we want to create a relationship with the type `VENUE` for each entry in the `article_VENUE_venue.csv` file. Now if we run the script, we'll see the following output: ``` Neo4j version: 3.5.3 Importing the contents of these files into /home/markhneedham/.config/Neo4j Desktop/Application/neo4jDatabases/database-32ed444f-69cd-422c-81aa-0c634210ad50/installation-3.5.3/data/databases/graph.db: Nodes: :Author /home/markhneedham/projects/dblp/data/authors_header.csv /home/markhneedham/projects/dblp/data/authors.csv :Article /home/markhneedham/projects/dblp/data/articles_header.csv /home/markhneedham/projects/dblp/data/articles.csv :Venue /home/markhneedham/projects/dblp/data/venues_header.csv /home/markhneedham/projects/dblp/data/venues.csv Relationships: :REFERENCES /home/markhneedham/projects/dblp/data/article_REFERENCES_article_header.csv /home/markhneedham/projects/dblp/data/article_REFERENCES_article.csv :AUTHOR /home/markhneedham/projects/dblp/data/article_AUTHOR_author_header.csv /home/markhneedham/projects/dblp/data/article_AUTHOR_author.csv :VENUE /home/markhneedham/projects/dblp/data/article_VENUE_venue_header.csv /home/markhneedham/projects/dblp/data/article_VENUE_venue.csv Available resources: Total machine memory: 31.27 GB Free machine memory: 1.93 GB Max heap memory : 910.50 MB Processors: 8 Configured max memory: 27.34 GB High-IO: false WARNING: heap size 910.50 MB may be too small to complete this import. Suggested heap size is 1.00 GBImport starting 2019-03-27 09:43:54.934+0000 Estimated number of nodes: 7.84 M Estimated number of node properties: 24.08 M Estimated number of relationships: 36.99 M Estimated number of relationship properties: 0.00 Estimated disk space usage: 4.78 GB Estimated required memory usage: 1.09 GB IMPORT DONE in 2m 52s 306ms. Imported: 4850632 nodes 37215467 relationships 15328803 properties Peak memory usage: 1.08 GB ``` And now let's start the database and have a look at the graph we've imported: image::{{<siteurl>}}/uploads/2019/03/graph-citation.svg[]

Learn how to import data into Neo4j using the Neo4j Import Tool.

from-graph-model-to-neo4j-import

https://markhneedham.com/blog/2019/03/27/from-graph-model-to-neo4j-import

2019-03-14 06:42:00

Neo4j: Delete/Remove dynamic properties

https://twitter.com/irfannuri[Irfan^] and I were playing with a dataset earlier today, and having run a bunch of https://neo4j.com/docs/graph-algorithms/current/[graph algorithms^], we had a lot of properties that we wanted to clear out. The following Cypher query puts Neo4j into the state that we were dealing with. [source,cypher] ---- CREATE (:Node {name: "Mark", pagerank: 2.302, louvain: 1, lpa: 4 }) CREATE (:Node {name: "Michael", degree: 23, triangles: 12, betweeness: 48.70 }) CREATE (:Node {name: "Ryan", eigenvector: 2.302, scc: 1, unionFind: 4 }) ---- We wanted to delete all the properties except for 'name', but how do we do that? We could do it one at a time. For example: [source,cypher] ---- MATCH (n:Node) REMOVE n.pagerank ---- And then repeat that for all the other properties. That is a bit of a painful process though - it'd be good if we can automate it. First we need to get a list of the properties for each node, excluding the `name` property. The following query does this: [source, cypher] ---- neo4j> MATCH (n:Node) WITH [k in keys(n) where not k in ["name"]] as keys RETURN keys; +---------------------------------------+ | keys | +---------------------------------------+ | ["lpa", "pagerank", "louvain"] | | ["betweeness", "degree", "triangles"] | | ["unionFind", "eigenvector", "scc"] | +---------------------------------------+ ---- Now let's try and remove those properties. This was our first attempt: [source,cypher] ---- neo4j> MATCH (n:Node) WITH n, [k in keys(n) where not k in ["name"]] as keys UNWIND keys AS key REMOVE n[key]; Invalid input '[': expected an identifier character, whitespace, node labels, 'u/U', '{', 'o/O', a property map, a relationship pattern, '.' or '(' (line 4, column 9 (offset: 103)) "REMOVE n[key];" ---- Hmm, that doesn't work so well. We figured that https://neo4j-contrib.github.io/neo4j-apoc-procedures/[APOC^] would come to our rescue, we just didn't know which procedure we needed, so we searched all of them for the term 'remove': [source,cypher] ---- CALL dbms.procedures() YIELD name, signature, description WHERE name starts with "apoc" and description contains "remove" return name, signature, description ---- Here's the output of running that query: image::{{<siteurl>}}/uploads/2019/03/remove_procedures.png[] `apoc.create.removeProperties` looks like it will do the job. Let's give that a try: [source,cypher] ---- neo4j> MATCH (n:Node) WITH n, [k in keys(n) where not k in ["name"]] as keys CALL apoc.create.removeProperties(n, keys) YIELD node RETURN count(*); +----------+ | count(*) | +----------+ | 3 | +----------+ ---- Let's check what keys we have on those nodes now: [source,cypher] ---- neo4j> MATCH (n:Node) RETURN keys(n) AS keys; +----------+ | keys | +----------+ | ["name"] | | ["name"] | | ["name"] | +----------+ ---- Cool, that worked well! I was curious whether we could avoid having to call the `keys` function on each node, and instead just pass a list of all keys except for `name`. We can compute the list of all property keys in the database excluding `name` by executing the following query: [source, cypher] ---- neo4j> CALL db.propertyKeys() YIELD propertyKey WHERE propertyKey <> 'name' RETURN collect(propertyKey); +-----------------------------------------------------------------------------------------+ | collect(propertyKey) | +-----------------------------------------------------------------------------------------+ | ["degree", "pagerank", "louvain", "lpa", "triangles", "betweeness", "scc", "unionFind"] | +-----------------------------------------------------------------------------------------+ ---- If we want to pass in that list to the `apoc.create.removeProperties` procedure, we can do so like this: [source,cypher] ---- CALL db.propertyKeys() YIELD propertyKey WHERE propertyKey <> 'name' WITH collect(propertyKey) AS properties MATCH (n:Node) WITH collect(n) AS nodes, properties CALL apoc.create.removeProperties(nodes, properties) YIELD node RETURN count(*) ---- == Bigger data This approach works well if we want to delete properties from a small number of nodes, but what if we want to do this in bulk? `apoc.periodic.iterate` is our friend. Let's first create 1,000,000 nodes with properties that we want to remove: [source, cypher] ---- CALL apoc.periodic.iterate( "UNWIND range(0, 1000000) AS id RETURN id", "CREATE (:Node {name: 'name-' + id, pagerank: 2.302, louvain: 1, lpa: 4 })", {}) ---- And now we'll adapt our previous query to get rid of all properties except for `name`: [source,cypher] ---- neo4j> CALL db.propertyKeys() YIELD propertyKey WHERE propertyKey <> 'name' WITH collect(propertyKey) AS properties CALL apoc.periodic.iterate( "MATCH (n:Node) RETURN n", "WITH collect(n) AS nodes CALL apoc.create.removeProperties(nodes, $properties) YIELD node RETURN count(*)", {params: {properties: properties}}) YIELD batches RETURN batches; +---------+ | batches | +---------+ | 101 | +---------+ ---- And finally, let's check that all the properties are gone: [source, cypher] ---- neo4j> MATCH (n:Node) RETURN keys(n), count(*); +---------------------+ | keys(n) | count(*) | +---------------------+ | ["name"] | 1000001 | +---------------------+ ---- Sweet, that worked perfectly!

Learn how to remove a list of properties all in one go.

neo4j-delete-dynamic-properties

https://markhneedham.com/blog/2019/03/14/neo4j-delete-dynamic-properties

2019-04-17 09:00:00

pyspark: Py4JJavaError: An error occurred while calling o138.loadClass.: java.lang.ClassNotFoundException: org.graphframes.GraphFramePythonAPI

I've been building a Docker Container that has support for Jupyter, Spark, GraphFrames, and Neo4j, and ran into a problem that had me pulling my (metaphorical) hair out! The https://hub.docker.com/r/jupyter/pyspark-notebook/[pyspark-notebook^] container gets us most of the way there, but it doesn't have GraphFrames or Neo4j support. Adding Neo4j is as simple as pulling in the Python Driver from Conda Forge, which leaves us with GraphFrames. When I'm using GraphFrames with pyspark locally I would pull it in via the `--packages` config parameter, like this: [source, bash] ---- ./bin/pyspark --packages graphframes:graphframes:0.7.0-spark2.4-s_2.11 ---- I thought the same approach would work in the Docker container, so I created a Dockerfile that extends `jupyter/pyspark-notebook`, and added this code into the `SPARK_OPTS` environment variable: [source, docker] ---- ARG BASE_CONTAINER=jupyter/pyspark-notebook FROM $BASE_CONTAINER LABEL maintainer="Mark Needham" USER root USER $NB_UID ENV SPARK_OPTS --driver-java-options=-Xms1024M --driver-java-options=-Xmx4096M --driver-java-options=-Dlog4j.logLevel=info --packages graphframes:graphframes:0.7.0-spark2.4-s_2.11 RUN conda install --quiet --yes 'conda-forge::neo4j-python-driver' && \ pip install graphframes && \ fix-permissions $CONDA_DIR && \ fix-permissions /home/$NB_USER ---- I built the Docker image: [source,bash] ---- docker build . Successfully built fbcc49e923a6 ---- And then ran it locally: [source, bash] ---- docker run -p 8888:8888 fbcc49e923a6 [I 08:12:44.168 NotebookApp] The Jupyter Notebook is running at: [I 08:12:44.168 NotebookApp] http://(1f7d61b2f1de or 127.0.0.1):8888/?token=2f1c9e01326676af1a768b5e573eb9c58049c385a7714e53 [I 08:12:44.168 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation). [C 08:12:44.171 NotebookApp] To access the notebook, open this file in a browser: file:///home/jovyan/.local/share/jupyter/runtime/nbserver-6-open.html Or copy and paste one of these URLs: http://(1f7d61b2f1de or 127.0.0.1):8888/?token=2f1c9e01326676af1a768b5e573eb9c58049c385a7714e53 ---- I navigated to http://localhost:8888/?token=2f1c9e01326676af1a768b5e573eb9c58049c385a7714e53, which is where the Jupyter notebook is hosted. I uploaded a couple of CSV files, created a Jupyter notebook, and ran the following code: [source, python, linenums] ---- from pyspark.sql.types import * from graphframes import * import pandas as pd from pyspark.context import SparkContext from pyspark.sql.session import SparkSession sc = SparkContext.getOrCreate('local') spark = SparkSession(sc) def create_transport_graph(): node_fields = [ StructField("id", StringType(), True), StructField("latitude", FloatType(), True), StructField("longitude", FloatType(), True), StructField("population", IntegerType(), True) ] nodes = spark.read.csv("data/transport-nodes.csv", header=True, schema = StructType(node_fields)) rels = spark.read.csv("data/transport-relationships.csv", header=True) reversed_rels = (rels.withColumn("newSrc", rels.dst) .withColumn("newDst", rels.src) .drop("dst", "src") .withColumnRenamed("newSrc", "src") .withColumnRenamed("newDst", "dst") .select("src", "dst", "relationship", "cost")) relationships = rels.union(reversed_rels) return GraphFrame(nodes, relationships) g = create_transport_graph() ---- Unfortunately it throws the following exception when it tries to read the `data/transport-nodes.csv` file on line 18: [source, text] ---- Py4JJavaError: An error occurred while calling o138.loadClass. : java.lang.ClassNotFoundException: org.graphframes.GraphFramePythonAPI at java.net.URLClassLoader.findClass(URLClassLoader.java:381) at java.lang.ClassLoader.loadClass(ClassLoader.java:424) at java.lang.ClassLoader.loadClass(ClassLoader.java:357) at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method) at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62) at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43) at java.lang.reflect.Method.invoke(Method.java:497) at py4j.reflection.MethodInvoker.invoke(MethodInvoker.java:237) at py4j.reflection.ReflectionEngine.invoke(ReflectionEngine.java:357) at py4j.Gateway.invoke(Gateway.java:280) at py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:128) at py4j.commands.CallCommand.execute(CallCommand.java:79) at py4j.GatewayConnection.run(GatewayConnection.java:211) at java.lang.Thread.run(Thread.java:745) ---- I Googled the error message, and came across https://github.com/graphframes/graphframes/issues/104[this issue^], which has a lot of suggestions for how to fix it. I tried them all! I passed `--packages` to `PYSPARK_SUBMIT_ARGS` as well as `SPARK_OPTS`: [source, docker] ---- ARG BASE_CONTAINER=jupyter/pyspark-notebook FROM $BASE_CONTAINER LABEL maintainer="Mark Needham" USER root USER $NB_UID ENV SPARK_OPTS --driver-java-options=-Xms1024M --driver-java-options=-Xmx4096M --driver-java-options=-Dlog4j.logLevel=info --packages graphframes:graphframes:0.7.0-spark2.4-s_2.11 ENV PYSPARK_SUBMIT_ARGS --master local[*] pyspark-shell --packages graphframes:graphframes:0.7.0-spark2.4-s_2.11 RUN conda install --quiet --yes 'conda-forge::neo4j-python-driver' && \ pip install graphframes && \ fix-permissions $CONDA_DIR && \ fix-permissions /home/$NB_USER ---- I downloaded the GraphFrames JAR, and referenced it directly using the `--jars` argument: [source, docker] ---- ARG BASE_CONTAINER=jupyter/pyspark-notebook FROM $BASE_CONTAINER LABEL maintainer="Mark Needham" USER root USER $NB_UID ENV SPARK_OPTS --driver-java-options=-Xms1024M --driver-java-options=-Xmx4096M --driver-java-options=-Dlog4j.logLevel=info --jars /home/jovyan/graphframes-0.7.0-spark2.4-s_2.11.jar ENV PYSPARK_SUBMIT_ARGS --master local[*] pyspark-shell --jars /home/jovyan/graphframes-0.7.0-spark2.4-s_2.11.jar RUN conda install --quiet --yes 'conda-forge::neo4j-python-driver' && \ pip install graphframes && \ fix-permissions $CONDA_DIR && \ fix-permissions /home/$NB_USER COPY graphframes-0.7.0-spark2.4-s_2.11.jar /home/$NB_USER/graphframes-0.7.0-spark2.4-s_2.11.jar ---- I used the `--py-files` argument as well: [source, docker] ---- ARG BASE_CONTAINER=jupyter/pyspark-notebook FROM $BASE_CONTAINER LABEL maintainer="Mark Needham" USER root USER $NB_UID ENV SPARK_OPTS --driver-java-options=-Xms1024M --driver-java-options=-Xmx4096M --driver-java-options=-Dlog4j.logLevel=info --jars /home/jovyan/graphframes-0.7.0-spark2.4-s_2.11.jar --py-files /home/jovyan/graphframes-0.7.0-spark2.4-s_2.11.jar ENV PYSPARK_SUBMIT_ARGS --master local[*] pyspark-shell --jars /home/jovyan/graphframes-0.7.0-spark2.4-s_2.11.jar --py-files /home/jovyan/graphframes-0.7.0-spark2.4-s_2.11.jar RUN conda install --quiet --yes 'conda-forge::neo4j-python-driver' && \ pip install graphframes && \ fix-permissions $CONDA_DIR && \ fix-permissions /home/$NB_USER COPY graphframes-0.7.0-spark2.4-s_2.11.jar /home/$NB_USER/graphframes-0.7.0-spark2.4-s_2.11.jar ---- But nothing worked and I still had the same error message :( I was pretty stuck at this point, and returned to Google, where I found a https://stackoverflow.com/questions/39261370/unable-to-run-a-basic-graphframes-example[a StackOverflow thread^] that had I hadn't spotted before. https://stackoverflow.com/users/7174028/gilles-essoki[Gilles Essoki^] suggested copying the GraphFrames JAR directly into the `/usr/local/spark/jars` directory, so I updated my Dockerfile to do this: [source, docker] ---- ARG BASE_CONTAINER=jupyter/pyspark-notebook FROM $BASE_CONTAINER LABEL maintainer="Mark Needham" USER root USER $NB_UID ENV SPARK_OPTS --driver-java-options=-Xms1024M --driver-java-options=-Xmx4096M --driver-java-options=-Dlog4j.logLevel=info RUN conda install --quiet --yes 'conda-forge::neo4j-python-driver' && \ pip install graphframes && \ fix-permissions $CONDA_DIR && \ fix-permissions /home/$NB_USER COPY graphframes-0.7.0-spark2.4-s_2.11.jar /usr/local/spark/jars ---- I built it again, and this time my CSV files are happily processed! So thankyou Gilles! If you want to use this Docker container I've put it on GitHub at https://github.com/mneedham/pyspark-graphframes-neo4j-notebook[mneedham/pyspark-graphframes-neo4j-notebook^], or you can pull it directly from Docker using the following command: [source, cypher] ---- docker pull markhneedham/pyspark-graphframes-neo4j-notebook ----

Learn how to work around the ClassNotFoundException GraphFramePythonAPI error when using pyspark and GraphFrames.

pyspark-class-not-found-exception-org-graphframes-graphframepythonapi

https://markhneedham.com/blog/2019/04/17/pyspark-class-not-found-exception-org-graphframes-graphframepythonapi

2019-04-07 05:03:00

Python: Getting GitHub download count from the GraphQL API using requests

I was recently trying to use some code I shared just over a year ago to https://markhneedham.com/blog/2018/03/23/github-release-download-count/[compute GitHub Project download numbers from the GraphQL API^], and wanted to automate this in a Python script. It was more fiddly than I expected, so I thought I'd share the code for the benefit of future me more than anything else! == Pre requisites We're going to use the popular requests library to query the API, so we need to import that. We'll also be parsing JSON and reading environment variables, so we'll load libraries to help with those tasks as well. [source,python] ---- import requests import json import os ---- == Creating the script We can use the query from the original post, so let's create a string containing that: [source,python] ---- query = """ query($owner:String!, $name: String!) { repository(owner: $owner, name: $name) { nameWithOwner releases(first: 50, orderBy: {field:CREATED_AT, direction:DESC}) { totalCount nodes { releaseAssets(first: 1) { nodes { name downloadCount createdAt } } } } } } """ ---- That query takes in some variables. We'll put those in a map: [source, python] ---- variables = {"owner": "neo4j-contrib", "name": "neo4j-apoc-procedures"} ---- We're nearly ready to post a request to the GraphQL API, but first we need to create a personal access token. The GitHub docs https://help.github.com/en/articles/creating-a-personal-access-token-for-the-command-line[explain how to do this^]. Let's create a file containing the token. It looks like this: _devenv_ [source, bash] ---- export GITHUB_TOKEN="<github-token-here>" ---- We can now run the following commands so that the token is available as an environment variable: [source,bash] ---- source devenv ---- Now we're ready to construct our POST request: [source, python] ---- token = os.getenv("GITHUB_TOKEN") headers = {"Authorization": f"bearer {token}"} request = {"query": query, "variables": variables} response = requests.post("https://api.github.com/graphql", json=request, headers=headers) ---- The response we get back is extremely nested! We want to print the number of downloads per project, as well as the total download count. The following code does this: [source,python] ---- result = response.json()["data"]["repository"]["releases"]["nodes"] total_downloads = 0 for item in result: release = item["releaseAssets"]["nodes"][0] print( f"{release['name']:<35} {release['createdAt']} {release['downloadCount']}" ) total_downloads += release["downloadCount"] print(f"Total Downloads: {total_downloads}") ---- And if we run that we'll see this output: [source,bash] ---- apoc-3.4.0.5-all.jar 2019-02-15T20:45:21Z 17622 apoc-3.5.0.2-all.jar 2019-02-15T21:28:31Z 7454 apoc-3.5.0.1-all.jar 2018-11-27T14:16:12Z 28113 apoc-3.4.0.4-all.jar 2018-11-16T17:37:27Z 5502 apoc-3.5.0.0-all.jar 2018-10-11T00:17:08Z 414 apoc-3.4.0.3-all.jar 2018-09-20T13:38:29Z 13498 apoc-3.4.0.2-all.jar 2018-08-08T23:51:40Z 10680 apoc-3.3.0.4-all.jar 2018-08-08T23:49:52Z 30670 apoc-3.3.0.3-all.jar 2018-05-16T16:17:20Z 7509 apoc-3.4.0.1-all.jar 2018-05-16T16:13:52Z 29150 apoc-3.1.3.9-all.jar 2018-02-23T19:52:54Z 478 apoc-3.3.0.2-all.jar 2018-02-23T19:26:26Z 16508 apoc-3.2.3.6-all.jar 2018-02-23T19:26:37Z 2596 apoc-3.2.3.5-all.jar 2017-10-23T15:53:42Z 5806 apoc-3.3.0.1-all.jar 2017-10-23T15:54:12Z 29825 apoc-3.2.0.5-beta-all.jar 2017-10-03T18:39:15Z 407 apoc-3.2.0.4-all.jar 2017-08-07T14:29:09Z 6776 apoc-3.1.3.8-all.jar 2017-07-22T09:31:43Z 1275 apoc-3.1.3.7-all.jar 2017-05-15T07:06:33Z 2250 apoc-3.2.0.3-all.jar 2017-05-15T07:28:20Z 6963 apoc-3.2.0.2-all.jar 2017-04-03T08:22:26Z 543 apoc-3.1.3.6-all.jar 2017-04-03T08:19:17Z 1609 apoc-3.0.8.6-all.jar 2017-04-03T02:15:46Z 1014 apoc-3.2.0.1.jar 2017-03-10T01:21:52Z 89 apoc-3.1.2.5.jar 2017-03-10T06:13:03Z 537 apoc-3.1.0.4.jar 2017-03-10T06:12:03Z 594 apoc-3.0.8.5.jar 2017-03-10T01:34:02Z 36 apoc-3.0.8.4-all.jar 2017-01-06T02:15:33Z 862 apoc-3.1.0.3-all.jar 2016-12-14T02:19:30Z 2990 apoc-3.1.0.2-all.jar 2016-11-06T15:09:44Z 201 apoc-3.0.4.2-all.jar 2016-10-29T08:36:49Z 4286 apoc-3.1.0.1-all.jar 2016-10-06T08:54:34Z 199 apoc-3.0.4.1-all.jar 2016-08-22T16:10:57Z 1267 apoc-1.1.0.jar 2016-07-11T05:22:14Z 630 apoc-1.0.0.jar 2016-05-07T22:53:24Z 535 apoc-1.0.0-RC1.jar 2016-04-15T23:26:37Z 53 Total Downloads: 238941 ---- APOC all the things! The full script is available below if you want to do the same for your project: ++++ <script src="https://gist.github.com/mneedham/f73d1d5494586d58d3b8988f54abf1d3.js"></script> ++++

Learn how to write a Python script to get the number of downloads of a GitHub Project from the GraphQL API.

python-github-download-count-graphql-requests

https://markhneedham.com/blog/2019/04/07/python-github-download-count-graphql-requests

2019-04-01 05:03:00

Finding famous MPs based on their Wikipedia Page Views

As part of the https://towardsdatascience.com/graphing-brexit-bbe4314cf70[Graphing^] https://medium.com/neo4j/graphing-brexit-clustering-edition-3b966694e723[Brexit^] series of blog posts, I wanted to work out who were the most important Members of the UK parliament, and after a bit of Googling I realised that views of their Wikipedia pages would do the trick. I initially found my way to https://tools.wmflabs.org/pageviews/?project=en.wikipedia.org&platform=all-access&agent=spider&start=2018-01&end=2019-02&pages=Theresa_May[tools.wmflabs.org^], which is great for exploring the popularity of an individual MP, but not so good if you want to extract the data for 600 of them. I then came to learn that Wikimedia have https://wikimedia.org/api/rest_v1/[a REST API^] and, https://blog.wikimedia.org/2015/12/14/pageview-data-easily-accessible/[hidden at the bottom of a blog post^] from 2015, a Python library called myviews. Yay! It's really easy to use as well. Installation is via PyPi: [source, bash] ---- pip install mwviews ---- And then if we want to find the page views from the last week for our current Prime Minister, Theresa May, we can write the following code: [source, python] ---- from mwviews.api import PageviewsClient p = PageviewsClient("mark-needham") views = p.article_views("en.wikipedia", ["Theresa May"], start="20190324", end="20190331") ---- And now let's iterate over `views` to find the number of pageviews per day: [source, python] ---- for day in views: print(day, views[day]) 2019-03-24 00:00:00 {'Theresa_May': 23461} 2019-03-25 00:00:00 {'Theresa_May': 22272} 2019-03-26 00:00:00 {'Theresa_May': 18661} 2019-03-27 00:00:00 {'Theresa_May': 42541} 2019-03-28 00:00:00 {'Theresa_May': 34310} 2019-03-29 00:00:00 {'Theresa_May': 34514} 2019-03-30 00:00:00 {'Theresa_May': 20604} 2019-03-31 00:00:00 {'Theresa_May': 18137} ---- We can extend our example to compute pageviews for multiple people by adding their names to the array, and we'll also extend the date range back to the EU referendum of 2016: [source, python] ---- people = [ "Boris Johnson", "Theresa May", "Jacob Rees-Mogg", "Jeremy Corbyn" ] views = p.article_views("en.wikipedia", people, start="20160624", end="20190331") ---- That's a lot of days so, rather than printing out each day on its own, let's sum up the pageviews: [source, python] ---- votes = {person: 0 for person in people } for key in views.keys(): for person_key in views[key].keys(): person = person_key.replace("_", " ") if views[key][person_key]: votes[person] += views[key][person_key] ---- And who's the most famous of them all? [source,python] ---- max_width = max([len(key) for key in votes.keys()]) for person in votes: print(f"{person:<{max_width}} {votes[person]:,}") Boris Johnson 5,727,213 Theresa May 12,844,215 Jacob Rees-Mogg 3,631,652 Jeremy Corbyn 5,965,669 ----

Learn how to find out the number of pageviews Wikipedia pages have.

famous-mps-wikipedia-pageviews

https://markhneedham.com/blog/2019/04/01/famous-mps-wikipedia-pageviews

2019-04-14 12:52:00

Neo4j: Delete all nodes

When experimenting with a new database, at some stage we'll probably want to delete all our data and start again. I was trying to do this with Neo4j over the weekend and it didn't work as I expected, so I thought I'd write the lessons I learned. We'll be using Neo4j via the https://neo4j.com/developer/neo4j-desktop/[Neo4j Desktop^] with the default settings. This means that we have a maximum heap size of 1GB. _This blog post assumes that you've got the https://neo4j.com/developer/neo4j-apoc/[Neo4j APOC^] library installed._ _We can find https://neo4j-contrib.github.io/neo4j-apoc-procedures/#_installation_with_neo4j_desktop[instructions for installing that library in the docs^]._ == Cypher Shell In this post we'll be executing Cypher queries using the https://neo4j.com/docs/operations-manual/current/tools/cypher-shell/[Cypher Shell^]. We can launch that from the Neo4j Desktop like this: image::{{<siteurl>}}/uploads/2019/04/cypher-shell.png[] == Creating data Once we've done that we're ready to create 1 million nodes using APOC's https://neo4j-contrib.github.io/neo4j-apoc-procedures/#commit-batching[periodic iterate^] procedure. [source,cypher] ---- neo4j> CALL apoc.periodic.iterate( 'UNWIND range(1, 1000000) as id RETURN id', 'CREATE (:Node {id: id})', {} ) YIELD timeTaken, operations RETURN timeTaken, operations; +-------------------------------------------------------------------------+ | timeTaken | operations | +-------------------------------------------------------------------------+ | 8 | {total: 1000000, committed: 1000000, failed: 0, errors: {}} | +-------------------------------------------------------------------------+ 1 row available after 8249 ms, consumed after another 0 ms ---- Let's check how many nodes our database contains: [source, cypher] ---- neo4j> MATCH () RETURN count(*); +----------+ | count(*) | +----------+ | 1000000 | +----------+ 1 row available after 0 ms, consumed after another 0 ms ---- Great, 1 million nodes, all ready to be deleted! == Deleting nodes My first attempt to delete all this nodes was the following query, which finds all the nodes and then attempts to delete them: [source, cypher] ---- neo4j> MATCH (n) DETACH DELETE n; There is not enough memory to perform the current task. Please try increasing 'dbms.memory.heap.max_size' in the neo4j configuration (normally in 'conf/neo4j.conf' or, if you you are using Neo4j Desktop, found through the user interface) or if you are running an embedded installation increase the heap by using '-Xmx' command line flag, and then restart the database. ---- Hmmm, an OutOfMemory exception. I had some ideas as to why that might be happening, but it's best to get a heap dump to know for sure. We can add the following lines to the Neo4j settings or configuration file to create a heap dump when an OutOfMemory exception is thrown: [source, text] ---- dbms.jvm.additional=-XX:+HeapDumpOnOutOfMemoryError -XX:HeapDumpPath=/tmp/neo4jdump.hprof ---- If you're using the Neo4j Desktop, you can find this file from the 'Settings' tab: image::{{<siteurl>}}/uploads/2019/04/settings.png[] We can visualise the contents of this file using a tool like https://www.yourkit.com/[YourKit^] or https://visualvm.github.io/[VisualVM^]. I have VisualVM installed on my machine, so we'll use that. Below is a print screen showing the heap dump: image::{{<siteurl>}}/uploads/2019/04/hdump.png[] Many of the classes taking up space on the heap are used to create the commands that get stored in the https://neo4j.com/docs/operations-manual/current/configuration/transaction-logs/[transaction log^] whenever we execute a write query. == Batching deletes We need to delete our nodes in batches so that we don't end up with so much in memory state. My go to procedure is `apoc.periodic.iterate`, so let's give that a try: [source,cypher] ---- CALL apoc.periodic.iterate( 'MATCH (n) RETURN n', 'DELETE n', {batchSize: 10000} ) YIELD timeTaken, operations RETURN timeTaken, operations ---- I found that sometimes this works, but sometimes we end up with the whole heap being filled, and a lot of garbage collection pauses: image::{{<siteurl>}}/uploads/2019/04/gc.png[] We can also see all the GC pauses by searching the debug log, which we can access via the 'Terminal' tab: [source, bash] ---- $ grep VmPauseMonitorComponent logs/debug.log | tail -n 10 2019-04-14 16:14:22.377+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=9143, gcTime=4619, gcCount=7} 2019-04-14 16:14:28.845+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=6367, gcTime=6451, gcCount=10} 2019-04-14 16:14:35.730+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=2131, gcTime=6875, gcCount=12} 2019-04-14 16:14:44.455+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=9080, gcTime=4523, gcCount=5} 2019-04-14 16:14:46.721+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=6364, gcTime=6449, gcCount=18} 2019-04-14 16:15:09.106+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=19938, gcTime=22355, gcCount=28} 2019-04-14 16:15:13.288+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=6428, gcTime=4176, gcCount=7} 2019-04-14 16:15:17.807+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=4418, gcTime=4515, gcCount=5} 2019-04-14 16:16:00.108+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=19724, gcTime=42279, gcCount=40} 2019-04-14 16:16:00.209+0000 WARN [o.n.k.i.c.VmPauseMonitorComponent] Detected VM stop-the-world pause: {pauseTime=22476, gcTime=10, gcCount=1} ---- If we take a heap dump using VisualVM, we'll see something like this: image::{{<siteurl>}}/uploads/2019/04/heap-periodic.png[] This time it's not commands that are taking up the space, but rather objects holding all the nodes that we want to delete. To avoid loading all those nodes into memory, we can use the `apoc.periodic.commit` procedure instead. With this procedure we provide one query that must contain a `LIMIT` clause. We also need to include a `RETURN` clause at the end of our query, and as long as a result is returned, it will keep on iterating. [source,cypher] ---- neo4j> CALL apoc.periodic.commit( 'MATCH (n) WITH n LIMIT $limit DELETE n RETURN count(*)', {limit: 10000} ) YIELD updates, executions, runtime, batches RETURN updates, executions, runtime, batches; +------------------------------------------+ | updates | executions | runtime | batches | +------------------------------------------+ | 1000000 | 100 | 7 | 101 | +------------------------------------------+ 1 row available after 7540 ms, consumed after another 0 ms ---- Good times! All the nodes are deleted and we can get on with our day. //// [source,cypher] ---- CREATE INDEX ON :Node(id) ---- [source,cypher] ---- CALL apoc.periodic.iterate( 'WITH ["FOO", "BAR", "BAX"] AS prefixes UNWIND prefixes AS prefix UNWIND range(0, 100000) AS id RETURN prefix, id', 'MERGE (node1:Node {id: id + prefix}) MERGE (node2:Node {id: (id+1) + prefix}) WITH node1, node2, prefix, id CALL apoc.create.relationship(node1, prefix + "_" + toInteger(rand() * 10.0), {}, node2) YIELD rel RETURN count(*)', {}) YIELD operations RETURN operations ---- What if we want to delete all the relationships with the `BAX` prefix? ////

Learn how to delete all the nodes in a Neo4j Database

neo4j-delete-all-nodes

https://markhneedham.com/blog/2019/04/14/neo4j-delete-all-nodes

2019-05-20 11:43:00

KSQL: Create Stream - extraneous input 'properties'

In my continued playing with the https://www.confluent.io/product/ksql/[KSQL streaming engine for Kafka^], I came across another interesting error while trying to put a stream on top of a topic generated by the https://neo4j-contrib.github.io/neo4j-streams/[Neo4j Streams Library^]. We'll simplify the events being posted on the topic for this blog post, so this is what the events on the topic look like: [source, javascript] ---- { "id":"ABCDEFGHI", "properties": { "name":"Mark", "location":"London" } } ---- We then create a stream on that topic: [source, sql] ---- CREATE STREAM users_original( id varchar, properties STRUCT< name varchar, location varchar > ) WITH(KAFKA_TOPIC='users', value_format='json'); ---- But when we execute that query we'll get the following exception: [source, text] ---- line 3:3: extraneous input 'properties' expecting {'INTEGER', 'DATE', 'TIME', 'TIMESTAMP', 'INTERVAL', 'YEAR', 'MONTH', 'DAY', 'HOUR', 'MINUTE', 'SECOND', 'ZONE', 'PARTITION', 'STRUCT', 'EXPLAIN', 'ANALYZE', 'TYPE', 'SHOW', 'TABLES', 'COLUMNS', 'COLUMN', 'PARTITIONS', 'FUNCTIONS', 'FUNCTION', 'ARRAY', 'MAP', 'SET', 'RESET', 'SESSION', 'IF', IDENTIFIER, DIGIT_IDENTIFIER, QUOTED_IDENTIFIER, BACKQUOTED_IDENTIFIER} Statement: CREATE STREAM users_original( ---- I thought I might have a trailing comma or colon - two mistakes I've made before - but on this occassion the problem is that `properties` is a reserved word in KSQL. Luckily it's an https://docs.confluent.io/current/ksql/docs/developer-guide/syntax-reference.html[easy problem to fix^]: [quote, KSQL Syntax Reference] ____ If the name of a column in your source topic is one of the reserved words in KSQL you can use back quotes to define the column. The same applies to the field names in a STRUCT type. For instance, if in the above example we had another field called Properties, which is a reserved word in KSQL, you can use the following statement to declare your stream: ____ The fixed syntax is below: [source, sql] ---- CREATE STREAM users_original( id varchar, `properties` STRUCT< name varchar, location varchar > ) WITH(KAFKA_TOPIC='users', value_format='json'); ---- And if we run that the stream will be created: [source, text] ---- Message ---------------- Stream created ---------------- ----

Learn how to create a KSQL stream from a source topic that contains a column with a reserved word.

kql-create-stream-extraneous-input

https://markhneedham.com/blog/2019/05/20/kql-create-stream-extraneous-input

2019-05-29 06:50:00

Twint: Loading tweets into Kafka and Neo4j

In this post we're going to load tweets via the https://github.com/twintproject/twint[twint library^] into Kafka, and once we've got them in there we'll use the https://www.confluent.io/blog/kafka-connect-neo4j-sink-plugin[Kafka Connect Neo4j Sink Plugin^] to get them into Neo4j. == What is twint? Twitter data has always been some of the most fun to play with, but over the years the official API has become more and more restritive, and it now takes a really long time to download enough data to do anything interesting. I was therefore intrigued when Michael showed me the https://github.com/twintproject/twint[twint^] library, which describes itself as: ____ An advanced Twitter scraping & OSINT tool written in Python that doesn't use Twitter's API, allowing you to scrape a user's followers, following, Tweets and more while evading most API limitations ____ We can install the library via pip using the following command: [source, bash] ---- pip install --upgrade -e git+https://github.com/twintproject/twint.git@origin/master#egg=twint ---- == Retrieving tweets Let's learn how to use the library. The following code will retrieve tweets since 20th May 2019 for the search term `neo4j OR "graph database" OR "graph databases" OR graphdb OR graphconnect OR @neoquestions OR @Neo4jDE OR @Neo4jFr OR neotechnology` and save them into the file `tweets.json`: [source, python] ---- import twint import json c = twint.Config() c.Search = "neo4j OR \"graph database\" OR \"graph databases\" OR graphdb OR graphconnect OR @neoquestions OR @Neo4jDE OR @Neo4jFr OR neotechnology" c.Store_json = True c.Custom["user"] = ["id", "tweet", "user_id", "username", "hashtags", "mentions"] c.User_full = True c.Output = "tweets.json" c.Since = "2019-05-20" c.Hide_output = True twint.run.Search(c) ---- We can have a look at the contents of that file by running the following command: [source, bash] ---- $ head -n5 tweets.json {"id": 1133394643830345728, "conversation_id": "1133334287841992704", "created_at": 1559057315000, "date": "2019-05-28", "time": "16:28:35", "timezone": "BST", "user_id": 900548798, "username": "geolytix", "name": "GEOLYTIX", "place": null, "tweet": "have you benchmarked against OSRM? that is best of the 'standard' approaches. I know others ...incluidng us ;-)... doing interesting r&d with massively parallel approach with 1000's of graph database for crazy speeds.", "mentions": ["murraydata", "rapidsai", "ordnancesurvey", "transportgovuk", "puntofisso"], "urls": [], "photos": [], "replies_count": 1, "retweets_count": 0, "likes_count": 1, "location": "", "hashtags": [], "link": "https://twitter.com/Geolytix/status/1133394643830345728", "retweet": null, "quote_url": "", "video": 0} {"id": 1133393703148687361, "conversation_id": "1133393703148687361", "created_at": 1559057090000, "date": "2019-05-28", "time": "16:24:50", "timezone": "BST", "user_id": 892256485, "username": "neoquestions", "name": "Neo Questions", "place": null, "tweet": "\"neo4j - Return single instance of node - querying by property?\" #neo4jquestions https://stackoverflow.com/questions/56307118/neo4j-return-single-instance-of-node-querying-by-property …", "mentions": "", "urls": ["https://stackoverflow.com/questions/56307118/neo4j-return-single-instance-of-node-querying-by-property"], "photos": [], "replies_count": 0, "retweets_count": 0, "likes_count": 0, "location": "", "hashtags": ["#neo4jquestions"], "link": "https://twitter.com/NeoQuestions/status/1133393703148687361", "retweet": null, "quote_url": "", "video": 0} {"id": 1133393003656167424, "conversation_id": "1133393003656167424", "created_at": 1559056924000, "date": "2019-05-28", "time": "16:22:04", "timezone": "BST", "user_id": 135805905, "username": "phermar", "name": "Pablo Hernández", "place": null, "tweet": "An illuminating story of @emileifrem, the #Entrepreneur who founded @neo4j, the #Startup offering a graph platform for Data Analysis. https://www.forbes.com/sites/alejandrocremades/2019/05/28/this-entrepreneur-went-from-having-2000-left-in-the-bank-to-building-a-billion-dollar-business/#3d0f1b7b3192 …", "mentions": ["emileifrem", "neo4j"], "urls": ["https://www.forbes.com/sites/alejandrocremades/2019/05/28/this-entrepreneur-went-from-having-2000-left-in-the-bank-to-building-a-billion-dollar-business/#3d0f1b7b3192"], "photos": [], "replies_count": 0, "retweets_count": 0, "likes_count": 0, "location": "", "hashtags": ["#entrepreneur", "#startup"], "link": "https://twitter.com/phermar/status/1133393003656167424", "retweet": null, "quote_url": "", "video": 0} {"id": 1133392446279344128, "conversation_id": "1133392446279344128", "created_at": 1559056791000, "date": "2019-05-28", "time": "16:19:51", "timezone": "BST", "user_id": 892256485, "username": "neoquestions", "name": "Neo Questions", "place": null, "tweet": "\"How to efficiently store time-series values for each node in NEO4J?\" #neo4jquestions https://stackoverflow.com/questions/56345345/how-to-efficiently-store-time-series-values-for-each-node-in-neo4j …", "mentions": "", "urls": ["https://stackoverflow.com/questions/56345345/how-to-efficiently-store-time-series-values-for-each-node-in-neo4j"], "photos": [], "replies_count": 0, "retweets_count": 0, "likes_count": 0, "location": "", "hashtags": ["#neo4jquestions"], "link": "https://twitter.com/NeoQuestions/status/1133392446279344128", "retweet": null, "quote_url": "", "video": 0} {"id": 1133392332416606214, "conversation_id": "1132026469675225088", "created_at": 1559056764000, "date": "2019-05-28", "time": "16:19:24", "timezone": "BST", "user_id": 954001, "username": "ryguyrg", "name": "ryan boyd", "place": null, "tweet": "agreed on the spider :-( sounds like a valid comment on the amazon review! https://www.amazon.com/Graph-Algorithms-Practical-Examples-Apache/dp/1492047686/ …", "mentions": ["odbmsorg", "neo4j", "amyhodler", "markhneedham"], "urls": ["https://www.amazon.com/Graph-Algorithms-Practical-Examples-Apache/dp/1492047686/"], "photos": [], "replies_count": 0, "retweets_count": 0, "likes_count": 0, "location": "", "hashtags": [], "link": "https://twitter.com/ryguyrg/status/1133392332416606214", "retweet": null, "quote_url": "", "video": 0} ---- What about if we don't want to write to a JSON file, but want to write those JSON objects to somewhere not supported by the library? The writing of data to the JSON file is done by the `twint.storage.write.Json` function, which we can https://yuji.wordpress.com/2011/01/13/python-globally-override-function-imports/[replace with our own function^] using the following code: [source, python] ---- import twint import sys module = sys.modules["twint.storage.write"] def Json(obj, config): tweet = obj.__dict__ print(tweet) module.Json = Json c = twint.Config() c.Search = "neo4j OR \"graph database\" OR \"graph databases\" OR graphdb OR graphconnect OR @neoquestions OR @Neo4jDE OR @Neo4jFr OR neotechnology" c.Store_json = True c.Custom["user"] = ["id", "tweet", "user_id", "username", "hashtags", "mentions"] c.User_full = True c.Output = "tweets.json" c.Since = "2019-05-20" c.Hide_output = True twint.run.Search(c) ---- If we run that tweets will now be printed to stdout instead of being written to `tweets.json`: [source, bash] ---- $ python print_tweets.py | head -n5 {'id': 1133418564013174784, 'id_str': '1133418564013174784', 'conversation_id': '1133412611347886080', 'datetime': 1559063018000, 'datestamp': '2019-05-28', 'timestamp': '18:03:38', 'user_id': 2481818113, 'user_id_str': '2481818113', 'username': 'onejsninja', 'name': 'ECONNREFUSED', 'profile_image_url': 'https://pbs.twimg.com/profile_images/1122732844231610368/ccIfr_eK.jpg', 'place': None, 'timezone': 'BST', 'mentions': ['pensnaku', 'neo4j'], 'urls': [], 'photos': [], 'video': 0, 'tweet': '🙌', 'location': '', 'hashtags': [], 'replies_count': '0', 'retweets_count': '0', 'likes_count': '0', 'link': 'https://twitter.com/onejsninja/status/1133418564013174784', 'retweet': None, 'quote_url': ''} {'id': 1133417664347594752, 'id_str': '1133417664347594752', 'conversation_id': '1133417664347594752', 'datetime': 1559062803000, 'datestamp': '2019-05-28', 'timestamp': '18:00:03', 'user_id': 2355868690, 'user_id_str': '2355868690', 'username': 'theokraay', 'name': 'Theo van Kraay', 'profile_image_url': 'https://pbs.twimg.com/profile_images/953427963227394050/zMqFsTlX.jpg', 'place': None, 'timezone': 'BST', 'mentions': '', 'urls': ['https://lnkd.in/dC9jYFC'], 'photos': [], 'video': 0, 'tweet': 'Discover how to use the execution profile step for #AzureCosmosDB Gremlin API graph databases. Samples...... https://lnkd.in/dC9jYFC\xa0', 'location': '', 'hashtags': ['#azurecosmosdb'], 'replies_count': '0', 'retweets_count': '0', 'likes_count': '0', 'link': 'https://twitter.com/TheoKraay/status/1133417664347594752', 'retweet': None, 'quote_url': ''} {'id': 1133412611347886080, 'id_str': '1133412611347886080', 'conversation_id': '1133412611347886080', 'datetime': 1559061598000, 'datestamp': '2019-05-28', 'timestamp': '17:39:58', 'user_id': 378668650, 'user_id_str': '378668650', 'username': 'pensnaku', 'name': 'Eedee Naku', 'profile_image_url': 'https://pbs.twimg.com/profile_images/1123867014903287808/1QKtbWAt.jpg', 'place': None, 'timezone': 'BST', 'mentions': ['neo4j'], 'urls': [], 'photos': [], 'video': 0, 'tweet': 'Enjoying my first few days with @neo4j and Cypher. ASCII art. ( ) - [ ] -> ( );', 'location': '', 'hashtags': [], 'replies_count': '1', 'retweets_count': '2', 'likes_count': '7', 'link': 'https://twitter.com/pensnaku/status/1133412611347886080', 'retweet': None, 'quote_url': ''} {'id': 1133410865607319559, 'id_str': '1133410865607319559', 'conversation_id': '1133410865607319559', 'datetime': 1559061182000, 'datestamp': '2019-05-28', 'timestamp': '17:33:02', 'user_id': 2545730773, 'user_id_str': '2545730773', 'username': 'wbsbike', 'name': 'Will Snipes', 'profile_image_url': 'https://pbs.twimg.com/profile_images/474224499441688576/_qqblwJY.jpeg', 'place': None, 'timezone': 'BST', 'mentions': ['youtube'], 'urls': ['https://youtu.be/v6QI3YlYPrE'], 'photos': [], 'video': 0, 'tweet': 'How Graph Technology is Changing AIJake Graham Neo4j,Alicia Frame Neo4j https://youtu.be/v6QI3YlYPrE\xa0 via @YouTube', 'location': '', 'hashtags': [], 'replies_count': '0', 'retweets_count': '0', 'likes_count': '0', 'link': 'https://twitter.com/wbsbike/status/1133410865607319559', 'retweet': None, 'quote_url': ''} {'id': 1133410195537833985, 'id_str': '1133410195537833985', 'conversation_id': '1133410195537833985', 'datetime': 1559061023000, 'datestamp': '2019-05-28', 'timestamp': '17:30:23', 'user_id': 872808354720223236, 'user_id_str': '872808354720223236', 'username': 'oraclecourse', 'name': 'Oracle DBA Courses', 'profile_image_url': 'https://pbs.twimg.com/profile_images/956551490205835264/ODMsVpoX.jpg', 'place': None, 'timezone': 'BST', 'mentions': '', 'urls': ['https://twitter.com/OracleCourse/status/1133409940289277953'], 'photos': [], 'video': 0, 'tweet': '#ECO18 #ERP #ExploreOracle #ethereum #TDD #Essbase25 #EmergingTech #futurecities #frontend #Followback #FakeData #Fintech #fintechgeek #freeads #fourweeks #financecourses #fastest #financialservices #groundbreakerstour #ggupgrade #GraphDB #goals #GDPR #GoOracle #GoldenGate https://twitter.com/OracleCourse/status/1133409940289277953\xa0…', 'location': '', 'hashtags': ['#eco18', '#erp', '#exploreoracle', '#ethereum', '#tdd', '#essbase25', '#emergingtech', '#futurecities', '#frontend', '#followback', '#fakedata', '#fintech', '#fintechgeek', '#freeads', '#fourweeks', '#financecourses', '#fastest', '#financialservices', '#groundbreakerstour', '#ggupgrade', '#graphdb', '#goals', '#gdpr', '#gooracle', '#goldengate'], 'replies_count': '0', 'retweets_count': '0', 'likes_count': '0', 'link': 'https://twitter.com/OracleCourse/status/1133410195537833985', 'retweet': None, 'quote_url': 'https://twitter.com/OracleCourse/status/1133409940289277953'} ---- So far so good. Now that we can get access to the tweet objects we can store them in Kafka instead of printing them to the console. == Storing tweets in Kafka I've created a Docker Compose template that launches the containers we'll use for the rest of the post. It's available in the https://github.com/mneedham/kafka-connect-neo4j[kafka-connect-neo4j^] GitHub repository. [source, bash] ---- git clone git@github.com:mneedham/kafka-connect-neo4j.git && cd kafka-connect-neo4j docker-compose up ---- We should see the following output from running that command: [source, bash] ---- Starting zookeeper-kc ... done Starting broker-kc ... done Starting neo4j-kc ... done Starting schema_registry-kc ... done Starting ksql-server-kc ... done Starting connect-kc ... done Starting control-center-kc ... done Attaching to zookeeper-kc, broker-kc, ksql-server-kc, neo4j-kc, schema_registry-kc, connect-kc, control-center-kc ---- While that's running, we'll install the https://github.com/confluentinc/confluent-kafka-python[confluent-kafka-python^] driver using the following command: [source, bash] ---- pip install confluent-kafka[avro] ---- We can now update the `Json` function that we wrote earlier to store our tweets into the `tweets` topic: [source, python] ---- import twint import sys import json from confluent_kafka import avro from confluent_kafka.avro import AvroProducer # Define Avro schema value_schema_str = """ { "namespace": "my.test", "name": "value", "type": "record", "fields" : [ { "name": "id", "type": "long" }, { "name": "tweet", "type": "string" }, { "name": "datetime", "type": "long" }, { "name": "username", "type": "string" }, { "name": "user_id", "type": "long" }, { "name": "hashtags", "type": {"type": "array", "items": "string"} } ] } """ key_schema_str = """ { "namespace": "my.test", "name": "key", "type": "record", "fields" : [ { "name" : "name", "type" : "string" } ] } """ kafka_broker = 'localhost:9092' schema_registry = 'http://localhost:8081' value_schema = avro.loads(value_schema_str) key_schema = avro.loads(key_schema_str) producer = AvroProducer({ 'bootstrap.servers': kafka_broker, 'schema.registry.url': schema_registry }, default_key_schema=key_schema, default_value_schema=value_schema) module = sys.modules["twint.storage.write"] def Json(obj, config): tweet = obj.__dict__ print(tweet) producer.produce(topic='tweets', value=tweet, key={"name": "Key"}) producer.flush() module.Json = Json c = twint.Config() c.Search = "neo4j OR \"graph database\" OR \"graph databases\" OR graphdb OR graphconnect OR @neoquestions OR @Neo4jDE OR @Neo4jFr OR neotechnology" c.Store_json = True c.Custom["user"] = ["id", "tweet", "user_id", "username", "hashtags", "mentions"] c.User_full = True c.Output = "tweets.json" c.Since = "2019-05-20" c.Hide_output = True twint.run.Search(c) ---- Since the events we stored used an Avro schema, we'll use the `kafka-avro-console-consumer` command to query the topic: [source, bash] ---- $ docker exec schema_registry-kc kafka-avro-console-consumer --topic tweets --bootstrap-server broker:9093 --from-beginning [2019-05-28 17:10:59,610] INFO Registered kafka:type=kafka.Log4jController MBean (kafka.utils.Log4jControllerRegistration$) .... [2019-05-28 17:10:59,918] INFO Kafka version : 2.1.1-cp1 (org.apache.kafka.common.utils.AppInfoParser) [2019-05-28 17:10:59,918] INFO Kafka commitId : 9aa84c2aaa91e392 (org.apache.kafka.common.utils.AppInfoParser) [2019-05-28 17:11:00,035] INFO Cluster ID: Ai8uZd6RS7iUToW3jRwBTQ (org.apache.kafka.clients.Metadata) [2019-05-28 17:11:00,036] INFO [Consumer clientId=consumer-1, groupId=console-consumer-8062] Discovered group coordinator broker:9093 (id: 2147483646 rack: null) (org.apache.kafka.clients.consumer.internals.AbstractCoordinator) [2019-05-28 17:11:00,038] INFO [Consumer clientId=consumer-1, groupId=console-consumer-8062] Revoking previously assigned partitions [] (org.apache.kafka.clients.consumer.internals.ConsumerCoordinator) [2019-05-28 17:11:00,038] INFO [Consumer clientId=consumer-1, groupId=console-consumer-8062] (Re-)joining group (org.apache.kafka.clients.consumer.internals.AbstractCoordinator) [2019-05-28 17:11:00,047] INFO [Consumer clientId=consumer-1, groupId=console-consumer-8062] Successfully joined group with generation 1 (org.apache.kafka.clients.consumer.internals.AbstractCoordinator) [2019-05-28 17:11:00,048] INFO [Consumer clientId=consumer-1, groupId=console-consumer-8062] Setting newly assigned partitions [tweets-0] (org.apache.kafka.clients.consumer.internals.ConsumerCoordinator) [2019-05-28 17:11:00,065] INFO [Consumer clientId=consumer-1, groupId=console-consumer-8062] Resetting offset for partition tweets-0 to offset 0. (org.apache.kafka.clients.consumer.internals.Fetcher) {"id":1133394643830345728,"tweet":"have you benchmarked against OSRM? that is best of the 'standard' approaches. I know others ...incluidng us ;-)... doing interesting r&d with massively parallel approach with 1000's of graph database for crazy speeds.","datetime":1559057315000,"username":"geolytix","user_id":900548798,"hashtags":[]} {"id":1133393703148687361,"tweet":"\"neo4j - Return single instance of node - querying by property?\" #neo4jquestions https://stackoverflow.com/questions/56307118/neo4j-return-single-instance-of-node-querying-by-property …","datetime":1559057090000,"username":"neoquestions","user_id":892256485,"hashtags":["#neo4jquestions"]} {"id":1133393003656167424,"tweet":"An illuminating story of @emileifrem, the #Entrepreneur who founded @neo4j, the #Startup offering a graph platform for Data Analysis. https://www.forbes.com/sites/alejandrocremades/2019/05/28/this-entrepreneur-went-from-having-2000-left-in-the-bank-to-building-a-billion-dollar-business/#3d0f1b7b3192 …","datetime":1559056924000,"username":"phermar","user_id":135805905,"hashtags":["#entrepreneur","#startup"]} {"id":1133392446279344128,"tweet":"\"How to efficiently store time-series values for each node in NEO4J?\" #neo4jquestions https://stackoverflow.com/questions/56345345/how-to-efficiently-store-time-series-values-for-each-node-in-neo4j …","datetime":1559056791000,"username":"neoquestions","user_id":892256485,"hashtags":["#neo4jquestions"]} {"id":1133392332416606214,"tweet":"agreed on the spider :-( sounds like a valid comment on the amazon review! https://www.amazon.com/Graph-Algorithms-Practical-Examples-Apache/dp/1492047686/ …","datetime":1559056764000,"username":"ryguyrg","user_id":954001,"hashtags":[]} ---- Great, all good so far! Now we're ready to get the tweets from Kafka into Neo4j. == Storing tweets in Neo4j As mentioned at the beginning of this post, we're going to use the Kafka Connect Neo4j Sink Plugin to get the data from Kafka into Neo4j. The https://www.confluent.io/blog/kafka-connect-neo4j-sink-plugin[Kafka Connect Neo4j Sink Plugin^] was launched in February, and is a tool that makes it easy to load streaming data from Kafka into Neo4j. You control ingestion by defining Cypher statements per topic that you want to ingest. Those are then executed for batches of events coming in. We can create a new connector by running the following command: [source, bash] ---- curl -i -X POST -H "Accept:application/json" \ -H "Content-Type:application/json" http://localhost:8083/connectors/ \ -d '{ "name": "connect.sink.neo4j.tweets", "config": { "topics": "tweets", "connector.class": "streams.kafka.connect.sink.Neo4jSinkConnector", "neo4j.server.uri": "bolt://neo4j:7687", "neo4j.authentication.basic.username": "neo4j", "neo4j.authentication.basic.password": "neo", "neo4j.topic.cypher.tweets": "WITH event AS data MERGE (t:Tweet {id: data.id}) SET t.text = data.tweet, t.createdAt = datetime({epochmillis:data.datetime}) MERGE (u:User {username: data.username}) SET u.id = data.user_id MERGE (u)-[:POSTED]->(t) FOREACH (ht IN data.hashtags | MERGE (hashtag:HashTag {value: ht}) MERGE (t)-[:HAS_HASHTAG]->(hashtag))" } }' ---- This creates a consumer that takes messages from the `tweets` topic and runs the Cypher query defined by `neo4j.topic.cypher.tweets`. We can then run the following Cypher query to explore the data that's been loaded into Neo4j: [source, cypher] ---- MATCH path = (u:User)-[:POSTED]->(t:Tweet)-[:HAS_HASHTAG]->(ht) RETURN path LIMIT 100 ---- image::{{<siteurl>}}/uploads/2019/05/twitter.png[] == Summary And that's it! Hopefully this post has shown how easy it is to load data from Kafka into Neo4j using the Kafka Connect Sink. Below are useful resources in case you want to reproduce any part of this post: * https://www.confluent.io/blog/kafka-connect-neo4j-sink-plugin[Blog post^] announcing the launch of the Kafka Connect Neo4j Sink Plugin * https://gist.github.com/mneedham/fd0aece612896b8ef7cabd2943d6d90c[GitHub gist^] showing the evolution of our tweet processing code * https://github.com/mneedham/kafka-connect-neo4j[kafka-connect-neo4j^] repository for launching all the infrastructure used

Learn how to load tweets via the twint library into Kafka, and then use the Kafka Connect Sink to load them into Neo4j as well.

loading-tweets-twint-kafka-neo4j

https://markhneedham.com/blog/2019/05/29/loading-tweets-twint-kafka-neo4j

2019-05-16 10:02:00

Kafka: A basic tutorial

In this post we're going to learn how to launch Kafka locally and write to and read from a topic using one of the Python drivers. To make things easy for myself, I've created a Docker Compose template that launches 3 containers: * broker - our Kafka broker * zookeeper - used by Kafka for leader election * jupyter - notebooks for connecting to our Kafka broker This template can be downloaded from the https://github.com/mneedham/basic-kafka-tutorial[mneedham/basic-kafka-tutorial^] repository, and reads as follows: [source, yaml] ---- version: '3' services: zookeeper: image: confluentinc/cp-zookeeper hostname: zookeeper container_name: zookeeper-tutorial ports: - "2181:2181" environment: ZOOKEEPER_CLIENT_PORT: 2181 ZOOKEEPER_TICK_TIME: 2000 broker: image: confluentinc/cp-enterprise-kafka hostname: broker container_name: broker-tutorial depends_on: - zookeeper ports: - "9092:9092" expose: - "9093" environment: KAFKA_ADVERTISED_LISTENERS: PLAINTEXT://broker:9093,OUTSIDE://localhost:9092 KAFKA_LISTENER_SECURITY_PROTOCOL_MAP: PLAINTEXT:PLAINTEXT,OUTSIDE:PLAINTEXT KAFKA_LISTENERS: PLAINTEXT://0.0.0.0:9093,OUTSIDE://0.0.0.0:9092 CONFLUENT_METRICS_REPORTER_BOOTSTRAP_SERVERS: broker:9093 KAFKA_INTER_BROKER_LISTENER_NAME: PLAINTEXT KAFKA_BROKER_ID: 1 KAFKA_ZOOKEEPER_CONNECT: 'zookeeper:2181' KAFKA_METRIC_REPORTERS: io.confluent.metrics.reporter.ConfluentMetricsReporter KAFKA_OFFSETS_TOPIC_REPLICATION_FACTOR: 1 KAFKA_GROUP_INITIAL_REBALANCE_DELAY_MS: 0 jupyter: container_name: jupyter-tutorial image: jupyter/scipy-notebook:${JUPYTER_VERSION:-latest} volumes: - ./notebooks:/home/jovyan/ ports: - "8888:8888" ---- The easiest way to use this template is to first clone the repository locally using the following command: [source,bash] ---- git clone git@github.com:mneedham/basic-kafka-tutorial.git && cd basic-kafka-tutorial ---- And then launch the Docker containers using the following command: [source, bash] ---- docker-compose up ---- This command outputs a lot of stuff, and we'll need to keep a close eye on the first few lines so that we can grab the Jupyter Notebook token. The output we're looking for looks like this: [source, bash] ---- jupyter-tutorial | [I 10:35:27.804 NotebookApp] The Jupyter Notebook is running at: jupyter-tutorial | [I 10:35:27.804 NotebookApp] http://(4a031e4b5326 or 127.0.0.1):8888/?token=7907fef53948168308c829d48d9978b8f9c475b7c621c7d1 jupyter-tutorial | [I 10:35:27.804 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation). jupyter-tutorial | [C 10:35:27.811 NotebookApp] jupyter-tutorial | jupyter-tutorial | To access the notebook, open this file in a browser: jupyter-tutorial | file:///home/jovyan/.local/share/jupyter/runtime/nbserver-6-open.html jupyter-tutorial | Or copy and paste one of these URLs: jupyter-tutorial | http://(4a031e4b5326 or 127.0.0.1):8888/?token=7907fef53948168308c829d48d9978b8f9c475b7c621c7d1 ---- We then need to navigate to `http://localhost:8888?token=<token>` in our web browser. Based on the output above we'd navigate to `http://localhost:8888?token=7907fef53948168308c829d48d9978b8f9c475b7c621c7d1`. Once we've done that we can open https://github.com/mneedham/basic-kafka-tutorial/blob/master/notebooks/Kafka%20Tutorial.ipynb[Kafka Tutorial.ipynb^]. Let's go through that notebook. Once we've installed the `kafka-python` library we write the following functions to create a Kafka Producer and publish a message to a topic: [source, python] ---- from kafka import KafkaConsumer, KafkaProducer import json import uuid def publish_message(producer_instance, topic_name, key, value): try: key_bytes = bytes(key, encoding='utf-8') value_bytes = bytes(value, encoding='utf-8') producer_instance.send(topic_name, key=key_bytes, value=value_bytes) producer_instance.flush() print('Message published successfully.') except Exception as ex: print('Exception in publishing message') print(str(ex)) def connect_kafka_producer(server): _producer = None try: _producer = KafkaProducer(bootstrap_servers=[server], api_version=(0, 10)) except Exception as ex: print('Exception while connecting Kafka') print(str(ex)) finally: return _producer ---- Now we'll create a producer: [source, python] ---- server = 'broker:9093' kafka_producer = connect_kafka_producer(server) ---- And publish a JSON message to the broker: [source, python] ---- topic_name = "intro-to-kafka" message = json.dumps({"name": "Mark"}) publish_message(kafka_producer, topic_name, str(uuid.uuid4()), message) ---- And now we'll create a consumer to process all the message on that topic: [source, python] ---- consumer = KafkaConsumer(topic_name, auto_offset_reset='earliest', bootstrap_servers=[server], api_version=(0, 10), value_deserializer = json.loads, consumer_timeout_ms=1000) for msg in consumer: print(msg.key.decode("utf-8"), msg.value) ---- If we run this code we'll see the following output: [source,text] ---- 716a46df-2b10-4270-b294-b04ced51bd73 {'name': 'Mark'} ---- And now we're ready to go and do some more fun stuff with streams!

An extremely basic introduction to Kafka, the stream-processing platform.

kafka-basic-tutorial

https://markhneedham.com/blog/2019/05/16/kafka-basic-tutorial

2019-05-10 23:00:00

Jupyter: RuntimeError: This event loop is already running

I've been using the https://github.com/twintproject/twint/wiki/Module[twint^] library to explore the Neo4j twitter community, and ran into an initially confusing error when I moved the code I'd written into a Jupyter notebook. The first three cells of my notebook contain the following code: Cell 1: [source, python] ---- ! pip install twint ---- Cell 2: [source, python] ---- import json import twint ---- Cell 3: [source,python] ---- users = ["vikatakavi11", "tee_mars3"] for username in users[:10]: c = twint.Config() c.Username = username c.User_full = False c.Store_object = True c.Hide_output = True twint.run.Followers(c) followers = twint.output.follow_object if not username in followers: followers[username] = {"followers": []} print(followers) twint.output.follow_object = {} ---- If we execute those three cells we'll get the following error: [source,text] ---- --------------------------------------------------------------------------- RuntimeError Traceback (most recent call last) <ipython-input-5-796d99d0022c> in <module> 6 c.Hide_output = True 7 ----> 8 twint.run.Followers(c) 9 followers = twint.output.follow_object 10 ~/projects/twitter-graph/a/src/twint/twint/run.py in Followers(config) 219 config.Favorites = False 220 config.TwitterSearch = False --> 221 run(config) 222 if config.Pandas_au: 223 storage.panda._autoget("followers") ~/projects/twitter-graph/a/src/twint/twint/run.py in run(config, callback) 198 def run(config, callback=None): 199 logme.debug(__name__+':run') --> 200 get_event_loop().run_until_complete(Twint(config).main(callback)) 201 202 def Favorites(config): /usr/lib/python3.6/asyncio/base_events.py in run_until_complete(self, future) 458 future.add_done_callback(_run_until_complete_cb) 459 try: --> 460 self.run_forever() 461 except: 462 if new_task and future.done() and not future.cancelled(): /usr/lib/python3.6/asyncio/base_events.py in run_forever(self) 412 self._check_closed() 413 if self.is_running(): --> 414 raise RuntimeError('This event loop is already running') 415 if events._get_running_loop() is not None: 416 raise RuntimeError( RuntimeError: This event loop is already running ---- I knew that the twint library was using an event loop, but I didn't realise that this would cause a problem with Jupyter until I read this https://github.com/jupyter/notebook/issues/3397#issuecomment-376803076[GitHub issue^], which explains that the Jupyter kernel is also running an event loop, and these loops can't be nested. This thread (https://github.com/jupyter/notebook/issues/3397#issuecomment-376803076) explains that we have two event loops running - one in Jupyter and one in the twint library. Luckily https://github.com/jupyter/notebook/issues/3397#issuecomment-419386811[Ewald de Wit^] created an awesome library called 'nest_asyncio' that we can use to workaround this error. I added the following cells: Cell 4: [source, python] ---- !pip install nest_asyncio ---- And now we can apply this library: [source,python] ---- import nest_asyncio nest_asyncio.apply() ---- And finally, we'll try to find the followers for our users again: [source, python] ---- for username in users[:10]: c = twint.Config() c.Username = username c.User_full = False c.Store_object = True c.Hide_output = True twint.run.Followers(c) followers = twint.output.follow_object if not username in followers: followers[username] = {"followers": []} print(followers) twint.output.follow_object = {} ---- And this time it works! [source,text] ---- {'vikatakavi11': {'followers': ['ganimohamed7861', 'ganimohamed7861']}} {'tee_mars3': {'followers': ['christophe_cruz', 'leclercq_ub']}} ---- The full code for the notebook is in the gist below: ++++ <script src="https://gist.github.com/mneedham/5599bcc714ea97c1277da0afc786410a.js"></script> ++++

Learn how to use a library that runs in an event loop inside a Jupyter notebook.

jupyter-runtimeerror-this-event-loop-is-already-running

https://markhneedham.com/blog/2019/05/10/jupyter-runtimeerror-this-event-loop-is-already-running

2019-05-19 19:21:00

KSQL: Create Stream - Failed to prepare statement: name is null

I've been playing with https://www.confluent.io/product/ksql/[KSQL^] over the weekend and ran into a basic error message that took me a little while to solve. I was trying to create a stream over a topic `dummy1`, which is the simplest possible thing you can do with KSQL. The events posted to `dummy1` are JSON messages containing only an `id` key. Below is an example of a message posted to the topic: [source, javascript] ---- { "id": "ABCDEFGHI"} ---- I tried to create the stream `dummy1_original` over this topic with the following statement: [source, sql] ---- ksql> CREATE STREAM dummy1_original (id varchar) WITH (kafka_topic='dummy1', value_format="JSON"); Failed to prepare statement: name is null ---- I didn't find any search results to explain this error message, but realised that I'd used double quotes on the `value_format`, which was different to all the examples I'd seen. I've clearly been doing too much Cypher where you can use single or double quotes to represent string values. Anyway...let's update our create statement to use only single quotes: [source, sql] ---- ksql> CREATE STREAM dummy1_original (id varchar) WITH (kafka_topic='dummy1', value_format='JSON'); Message ---------------- Stream created ---------------- ---- Success! We can now query the stream like this: [source, sql] ---- ksql> SELECT * FROM dummy1_original; ----

Learn how to work around a weird KSQL error message.

ksql-create-stream-failed-to-prepare-statement-name-is-null

https://markhneedham.com/blog/2019/05/19/ksql-create-stream-failed-to-prepare-statement-name-is-null

2019-05-24 06:10:00

Docker: Find the network for a container

If we want two Docker containers to communicate with each other they need to belong to the same network. In this post we'll learn how to find out the network of existing containers so that we can attach new containers to that network. All the containers mentioned in this post can be launched locally from Docker compose, using the following command: [source, bash] ---- git clone git@github.com:mneedham/ksql-kafka-neo4j-streams.git && cd ksql-kafka-neo4j-streams docker-compose-up ---- Running this command will create four containers: [source, bash] ---- Starting zookeeper-blog ... Starting broker-blog ... Starting ksql-server-blog ... Starting neo4j-blog ... ---- What if we want to add another container and have it be able to communicate with these containers? We want to know which network these containers are attached to, and a good place to start is the `docker inspect` command, which returns detailed information about containers. We'll process the JSON document that the command returns using the https://stedolan.github.io/jq/[jq library^]: [source, bash] ---- docker inspect neo4j-blog | jq ---- If we run that command, we'll see the following output: image::{{<siteurl>}}/uploads/2019/05/docker_inspect.png[] On the last line of this output we can see `NetworkMode`, which is described as follows in the documentation: [quote, 'https://docs.docker.com/engine/api/v1.24/[Docker Documentation^]'] ____ NetworkMode - Sets the networking mode for the container. Supported standard values are: bridge, host, none, and container:<name|id>. Any other value is taken as a custom network’s name to which this container should connect to. ____ We can write the following command to extract that value: [source, bash] ---- $ docker inspect neo4j-blog --format='{{ .HostConfig.NetworkMode }}' ksql-kafka-neo4j-streams_default ---- If we want to return the network for all running containers, we can just iterate over them and run the command for each one: [source, bash] ---- for container in `docker container ls --format "{{.Names}}"`; do network=`docker inspect $container --format="{{ .HostConfig.NetworkMode }}"` image=`docker inspect $container --format="{{ .Config.Image }}"` printf '%-20s %-35s %-15s\n' $container $image $network done ksql-server-blog confluentinc/cp-ksql-server:5.2.1 ksql-kafka-neo4j-streams_default neo4j-blog neo4j:3.4.10 ksql-kafka-neo4j-streams_default broker-blog confluentinc/cp-enterprise-kafka ksql-kafka-neo4j-streams_default zookeeper-blog confluentinc/cp-zookeeper ksql-kafka-neo4j-streams_default ---- If we want to extract more detailed information about the network, we can find that under https://stackoverflow.com/questions/43904562/docker-how-to-find-the-network-my-container-is-in[`NetworkSettings.Network`^]. The following query returns this information for the `neo4j-blog` container: [source, bash] ---- $ docker inspect neo4j-blog --format='{{ json .NetworkSettings.Networks }}' | jq { "ksql-kafka-neo4j-streams_default": { "IPAMConfig": null, "Links": null, "Aliases": [ "c004bcd09652", "neo4j" ], "NetworkID": "f3bc43280a8a2ae764280e08fa9604097079d7020fac44d7d2012d2b2d353723", "EndpointID": "3ed6ac3c03e83e96fc56b57aedff2606f264237998194ff5215480744ce0b303", "Gateway": "172.27.0.1", "IPAddress": "172.27.0.5", "IPPrefixLen": 16, "IPv6Gateway": "", "GlobalIPv6Address": "", "GlobalIPv6PrefixLen": 0, "MacAddress": "02:42:ac:1b:00:05", "DriverOpts": null } } ---- We can then write the following query to extract the name of the network: [source, bash] ---- $ docker inspect neo4j-blog --format='{{range $k,$v := .NetworkSettings.Networks}} {{$k}} {{end}}' ksql-kafka-neo4j-streams_default ---- And if we want to pull out a couple of properties, we can do that as well: [source, bash] ---- $ docker inspect neo4j-blog --format='{{range $k,$v := .NetworkSettings.Networks}} {{$k}} {{$v.IPAddress}} {{$v.Aliases}} {{end}}' ksql-kafka-neo4j-streams_default 172.27.0.5 [c004bcd09652 neo4j] ----

Learn how to find the network that a Docker container is attached to.

docker-find-network-for-container

https://markhneedham.com/blog/2019/05/24/docker-find-network-for-container

2019-05-23 07:58:00

Deleting Kafka Topics on Docker

In this post we're going to learn how to delete a Kafka Topic when running a Kafka Broker on Docker. [NOTE] ==== **Update: 26th July 2023** While the approach described in this blog post still works, I think I've now got an even easier way using a command line tool called `rpk`. If you're interested in seeing an alternative approach https://www.markhneedham.com/blog/2023/07/26/how-to-delete-kafka-topic/[check out my other blog post]. If not, as you were! ==== We'll spin up a local Kafka environment using the Docker Compose template from the https://markhneedham.com/blog/2019/05/16/kafka-basic-tutorial/[Kafka Basic Tutorial blog post^] that I wrote last week. Let's open a terminal window and run the following commands to set up our environment: [source, bash] ---- git clone git@github.com:mneedham/basic-kafka-tutorial.git && cd basic-kafka-tutorial docker-compose up ---- On another terminal window, run the following command to see the list of Docker containers that are running: [source, bash] ---- $ docker ps CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES 3851185ae46f jupyter/scipy-notebook:latest "tini -g -- start-no…" 7 days ago Up 3 minutes 0.0.0.0:8888->8888/tcp jupyter-tutorial a118a60010ce confluentinc/cp-enterprise-kafka "/etc/confluent/dock…" 7 days ago Up 3 minutes 0.0.0.0:9092->9092/tcp, 9093/tcp broker-tutorial 1606bfe6e93d confluentinc/cp-zookeeper "/etc/confluent/dock…" 7 days ago Up 4 minutes 2888/tcp, 0.0.0.0:2181->2181/tcp, 3888/tcp zookeeper-tutorial ---- We're going to connect to `broker-tutorial`, seen on the 2nd line of this output, which is the container for our Kafka broker. We'll use the https://docs.docker.com/engine/reference/commandline/exec/[`docker exec`^] command to connect to the broker, and use the `kafka-topics` command to create and delete a topic. Jacek Laskowski gives a https://jaceklaskowski.gitbooks.io/apache-kafka/kafka-topic-deletion.html[detailed explanation of this command^] in https://jaceklaskowski.gitbooks.io/apache-kafka/[The Internals of Apache Kafka^] gitbook. Let's start by creating a dummy topic, which we can do with the following command: [source, bash] ---- docker exec broker-tutorial kafka-topics \ --create \ --zookeeper zookeeper:2181 \ --replication-factor 1 \ --partitions 1 \ --topic blog-dummy ---- .Output [source, text] ---- Created topic "blog-dummy". ---- So far, so good. Note that we also to pass in the `--zookeeper` argument to tell the command where our Zookeeper Instance is running. The `zookeeper:2181` value is derived from the `hostname:ZOOKEEPER_CLIENT_PORT` values from the following fragment of our https://github.com/mneedham/basic-kafka-tutorial/blob/master/docker-compose.yml[Docker Compose file^]: [source, yaml] ---- zookeeper: image: confluentinc/cp-zookeeper hostname: zookeeper container_name: zookeeper-tutorial ports: - "2181:2181" environment: ZOOKEEPER_CLIENT_PORT: 2181 ZOOKEEPER_TICK_TIME: 2000 ---- We can list all the topics available on this broker by running the following command: [source, bash] ---- docker exec broker-tutorial kafka-topics \ --list \ --zookeeper zookeeper:2181 ---- .Output [source, text] ---- __confluent.support.metrics _confluent-metrics blog-dummy ---- There are a couple of Confluent metrics topics, but we can see our `blog-dummy` topic as well. Now let's delete it by running the following command: [source, bash] ---- docker exec broker-tutorial kafka-topics \ --delete \ --zookeeper zookeeper:2181 \ --topic blog-dummy ---- .Output [source, text] ---- Topic blog-dummy is marked for deletion. Note: This will have no impact if delete.topic.enable is not set to true. ---- We didn't actually set the `delete.topic.enable` parameter, so maybe our topic isn't actually going to be deleted. If we look back at the other terminal window, however, we'll see the following output: [source, bash] ---- broker-tutorial | [2019-05-23 06:29:23,307] INFO [Topic Deletion Manager 1], Handling deletion for topics blog-dummy (kafka.controller.TopicDeletionManager) broker-tutorial | [2019-05-23 06:29:23,308] INFO [Topic Deletion Manager 1], Deletion of topic blog-dummy (re)started (kafka.controller.TopicDeletionManager) broker-tutorial | [2019-05-23 06:29:23,308] INFO [Topic Deletion Manager 1], Topic deletion callback for blog-dummy (kafka.controller.TopicDeletionManager) broker-tutorial | [2019-05-23 06:29:23,348] INFO [Topic Deletion Manager 1], Deletion of topic blog-dummy successfully completed (kafka.controller.TopicDeletionManager) ---- So it seems like our topic was deleted. And, in fact, scrolling back through the output from running `docker-compose up` we'll find the following line: [source, bash] ---- broker-tutorial | delegation.token.max.lifetime.ms = 604800000 broker-tutorial | delete.records.purgatory.purge.interval.requests = 1 broker-tutorial | delete.topic.enable = true broker-tutorial | fetch.purgatory.purge.interval.requests = 1000 broker-tutorial | group.initial.rebalance.delay.ms = 0 ---- So it seems like this property is set by default in the version that we're using. We can list the topics one more time just to be sure: [source, bash] ---- docker exec broker-tutorial kafka-topics \ --list \ --zookeeper zookeeper:2181 ---- .Output [source, text] ---- __confluent.support.metrics _confluent-metrics ---- The Confluent Metrics topics remain, but `blog-dummy` is nowhere to be seen. Success!

Learning how to delete a Kafka Topic when running a Kafka Broker on Docker.

deleting-kafka-topics-on-docker

https://markhneedham.com/blog/2019/05/23/deleting-kafka-topics-on-docker

2019-05-23 12:46:00

Processing Neo4j Transaction Events with KSQL and Kafka Streams

The https://neo4j-contrib.github.io/neo4j-streams[Neo4j Streams Library^] lets users send transaction events to a Kafka topic, and in this post we're going to learn how to explore these events using the KSQL streaming SQL Engine. All the infrastructure used in this post can be launched locally from Docker compose, using the following command: [source, bash] ---- git clone git@github.com:mneedham/ksql-kafka-neo4j-streams.git && cd ksql-kafka-neo4j-streams docker-compose-up ---- Running this command will create four containers: [source, bash] ---- Starting zookeeper-blog ... Starting broker-blog ... Starting ksql-server-blog ... Starting neo4j-blog ... ---- While that's running, let's learn how to use the Neo4j Streams Library. The library has four parts: * Neo4j Streams Procedure: a procedure to send a payload to a topic * Neo4j Streams Producer: a transaction event handler events that sends data to a Kafka topic * Neo4j Streams Consumer: a Neo4j application that ingest data from Kafka topics into Neo4j via templated Cypher Statements * Kafka-Connect Plugin: a plugin for the Confluent Platform that allows to ingest data into Neo4j, from Kafka topics, via Cypher queries. We're going to use the Neo4j Streams Producer to configure a source that will publish all nodes with the `User` label to the `users_blog` topic. If we were deploying Neo4j in a non Docker environment we'd do this by adding the following line to our Neo4j Configuration file: [source, xml] ---- streams.source.topic.nodes.users_blog= User{*} ---- But in our case we're using Docker, so instead we'll define the following environment variable: [source, yaml] ---- NEO4J_streams_source_topic_nodes_users__blog: User{*} ---- We can also see this configuration in the https://github.com/mneedham/ksql-kafka-neo4j-streams/blob/master/docker-compose.yml#L22[Docker Compose file^]: Now we're going to run a Cypher fragment that creates some `User` nodes in the Neo4j Browser, which is accessible from `http://localhost:7474`: [source, cypher] ---- UNWIND range(0, 100) AS id CREATE (u:User {id: id}) > Added 101 labels, created 101 nodes, set 101 properties, completed after 173 ms. ---- The diagram below shows the workflow of our query going into Neo4j, and then events being published to Kafka: image::{{<siteurl>}}/uploads/2019/05/neo4j_source.png[] Once this query has completed, events capturing the creation of those nodes will be published to the `users_blog` topic. We can then use KSQL to explore that data. But first things first, what is KSQL? [quote, 'https://docs.confluent.io/current/ksql/docs/index.html[KSQL Documentation^]'] ____ KSQL is the streaming SQL engine for Apache Kafka®. It provides an easy-to-use yet powerful interactive SQL interface for stream processing on Kafka, without the need to write code in a programming language such as Java or Python. KSQL is scalable, elastic, fault-tolerant, and real-time. It supports a wide range of streaming operations, including data filtering, transformations, aggregations, joins, windowing, and sessionization. ____ We'll execute KSQL queries using the KSQL CLI Docker container. Before we launch that container we need to know the name of the network on which our other containers are running. We can run the following code to determine the network: [source, bash] ---- for container in `docker container ls --format "{{.Names}}"`; do network=`docker inspect $container --format='{{ .HostConfig.NetworkMode }}'` printf '%-20s %-15s\n' $container $network done ksql-server-blog ksql-kafka-neo4j-streams_default neo4j-blog ksql-kafka-neo4j-streams_default broker-blog ksql-kafka-neo4j-streams_default zookeeper-blog ksql-kafka-neo4j-streams_default ---- All of our containers are using the network called `ksql-kafka-neo4j-streams_default`, a value that we'll pass to the `--network` parameter when we launch the KSQL CLI container: [source, bash] ---- $ docker run --network ksql-kafka-neo4j-streams_default --rm --interactive --tty confluentinc/cp-ksql-cli:5.2.1 http://ksql-server:8088 ---- Note that the `http://ksql-server:8088` value at the end of the command refers to the container name and listener port of our KSQL Server, which we can see in the following fragment of our Docker Compose file: [source, yaml] ---- ksql-server: container_name: ksql-server image: "confluentinc/cp-ksql-server:5.2.1" depends_on: - broker environment: KSQL_BOOTSTRAP_SERVERS: broker:9093 KSQL_LISTENERS: http://0.0.0.0:8088 ---- When we run that command we should see the following output: [source, bash] ---- =========================================== = _ __ _____ ____ _ = = | |/ // ____|/ __ \| | = = | ' /| (___ | | | | | = = | < \___ \| | | | | = = | . \ ____) | |__| | |____ = = |_|\_\_____/ \___\_\______| = = = = Streaming SQL Engine for Apache Kafka® = =========================================== Copyright 2017-2018 Confluent Inc. CLI v5.2.1, Server v5.2.1 located at http://ksql-server:8088 Having trouble? Type 'help' (case-insensitive) for a rundown of how things work! ksql> ---- If we see this prompt we're ready to roll. We'll create a stream over the `users_blog` topic by running the following query: [source, sql] ---- CREATE STREAM users_blog( payload STRUCT< id varchar, type varchar, before STRUCT< labels ARRAY<varchar>, `properties` MAP<varchar,varchar> >, after STRUCT< labels ARRAY<varchar>, `properties` MAP<varchar,varchar> > >, meta STRUCT < timestamp bigint, operation varchar, username varchar > ) WITH(KAFKA_TOPIC='users_blog', value_format='json'); ---- The fields and field types defined in this stream are based on the event definitions from the https://neo4j-contrib.github.io/neo4j-streams/#_transaction_event_handler[Transaction Event Handler section^] of the Neo4j Streams documentation. Next we'll run the following statement to tell KSQL to read from the beginning of the topic: [source, sql] ---- ksql> SET 'auto.offset.reset' = 'earliest'; Successfully changed local property 'auto.offset.reset' to 'earliest'. Use the UNSET command to revert your change. ---- And now let's query the stream: [source, sql] ---- ksql> SELECT * FROM users_blog LIMIT 10; 1558621108556 | 4-0 | {ID=0, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=0}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108828 | 5-1 | {ID=1, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=1}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108830 | 6-2 | {ID=2, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=2}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108831 | 7-3 | {ID=3, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=3}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108832 | 8-4 | {ID=4, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=4}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108833 | 9-5 | {ID=5, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=5}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108834 | 10-6 | {ID=6, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=6}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108835 | 11-7 | {ID=7, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=7}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108835 | 12-8 | {ID=8, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=8}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108837 | 13-9 | {ID=9, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=9}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} Limit Reached Query terminated ---- We can see the first ten nodes that were created by the Cypher query that we ran earlier. Notice that the operation is `created`. What if we take each of those nodes and add a `name` property by running the following Cypher query in the Neo4j Browser? [source, cypher] ---- MATCH (u:User) SET u.name = "Name-" + u.id > Set 101 properties, completed after 57 ms. ---- We can now run the following query to see the different types of events on our stream: [source, sql] ---- ksql> SELECT meta->operation, count(*) FROM users_blog GROUP BY meta->operation; created | 101 updated | 101 ---- We can view those events by running the following command and waiting until it gets to the last few entries: [source, sql] ---- ksql> SELECT * FROM users_blog; ... 1558622181521 | 101-95 | {ID=90, TYPE=node, BEFORE={LABELS=[User], properties={id=90}}, AFTER={LABELS=[User], properties={name=Name-90, id=90}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} 1558622181522 | 102-96 | {ID=75, TYPE=node, BEFORE={LABELS=[User], properties={id=75}}, AFTER={LABELS=[User], properties={name=Name-75, id=75}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} 1558622181522 | 103-97 | {ID=60, TYPE=node, BEFORE={LABELS=[User], properties={id=60}}, AFTER={LABELS=[User], properties={name=Name-60, id=60}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} 1558622181523 | 104-98 | {ID=45, TYPE=node, BEFORE={LABELS=[User], properties={id=45}}, AFTER={LABELS=[User], properties={name=Name-45, id=45}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} 1558622181523 | 105-99 | {ID=30, TYPE=node, BEFORE={LABELS=[User], properties={id=30}}, AFTER={LABELS=[User], properties={name=Name-30, id=30}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} 1558622181523 | 106-100 | {ID=15, TYPE=node, BEFORE={LABELS=[User], properties={id=15}}, AFTER={LABELS=[User], properties={name=Name-15, id=15}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} ---- Note that the properties now include both the `id` from when we created the nodes, as well as the `name` that we just added. We can now create separate streams to process the created and updated records: __users_blog_created__ [source, sql] ---- CREATE STREAM users_blog_created AS SELECT * FROM users_blog WHERE meta->operation = 'created'; ---------------------------- Stream created and running ---------------------------- ---- __users_blog_updated__ [source, sql] ---- CREATE STREAM users_blog_updated AS SELECT * FROM users_blog WHERE meta->operation = 'updated'; ---------------------------- Stream created and running ---------------------------- ---- And now we can query these streams individually. __users_blog_created__ [source, sql] ---- ksql> SELECT * FROM users_blog_created LIMIT 5; 1558621108556 | 4-0 | {ID=0, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=0}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108841 | 16-12 | {ID=12, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=12}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108835 | 11-7 | {ID=7, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=7}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108844 | 17-13 | {ID=13, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=13}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} 1558621108835 | 12-8 | {ID=8, TYPE=node, BEFORE=null, AFTER={LABELS=[User], properties={id=8}}} | {TIMESTAMP=1558621108484, OPERATION=created, USERNAME=neo4j} Limit Reached Query terminated ---- __users_blog_updated__ [source, sql] ---- ksql> SELECT * FROM users_blog_updated LIMIT 5; 1558622181456 | 7-1 | {ID=17, TYPE=node, BEFORE={LABELS=[User], properties={id=17}}, AFTER={LABELS=[User], properties={name=Name-17, id=17}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} 1558622181460 | 12-6 | {ID=8, TYPE=node, BEFORE={LABELS=[User], properties={id=8}}, AFTER={LABELS=[User], properties={name=Name-8, id=8}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} 1558622181454 | 6-0 | {ID=0, TYPE=node, BEFORE={LABELS=[User], properties={id=0}}, AFTER={LABELS=[User], properties={name=Name-0, id=0}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} 1558622181457 | 8-2 | {ID=34, TYPE=node, BEFORE={LABELS=[User], properties={id=34}}, AFTER={LABELS=[User], properties={name=Name-34, id=34}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} 1558622181459 | 11-5 | {ID=85, TYPE=node, BEFORE={LABELS=[User], properties={id=85}}, AFTER={LABELS=[User], properties={name=Name-85, id=85}}} | {TIMESTAMP=1558622181434, OPERATION=updated, USERNAME=neo4j} Limit Reached Query terminated ---- We could then create consumers that subscribe to these streams and process the events published, but that's for another blog post!

Learn how to process Neo4j Transaction events with KSQL and Kafka Streams.

processing-neo4j-transaction-events-ksql-kafka-streams

https://markhneedham.com/blog/2019/05/23/processing-neo4j-transaction-events-ksql-kafka-streams

2019-05-12 17:58:00

Neo4j: keep/filter keys in a map using APOC

In this post we'll learn how to write a Cypher query to create a node in Neo4j containing some of the keys from a map. This post assumes that the https://neo4j.com/developer/neo4j-apoc/[APOC library^] is installed. We'll start by creating a map that contains data from my twitter profile: [source, cypher] ---- :param document => { id: 14707949, name: "Mark Needham", username: "markhneedham", bio: "Developer Relations @neo4j", location: "London, United Kingdom", url: "http://www.markhneedham.com", join_date: "8 May 2008", join_time: "5:58 PM", tweets: 24710, following: 2479, followers: 5054, likes: 1014 }; ---- We want to create a `User` node based on this data, but we don't want to use all of the keys in the map. If we want to remove some keys we can use the `apoc.map.clean` function. The following function call removes the `join_date` and `join_time` keys: [source, cypher] ---- WITH ["join_date", "join_time"] AS keysToRemove RETURN apoc.map.clean($document, keysToRemove, []) AS result ---- image::{{<siteurl>}}/uploads/2019/05/keysToRemove.png[] But what if we want to keep keys rather than remove them? I ran `apoc.help('apoc.map')` to see if there were any functions to do this, and as far as I can tell there aren't: image::{{<siteurl>}}/uploads/2019/05/apochelp.png[] We can, however, combine two APOC functions to get the desired result. First up, `apoc.map.values` extracts the values for a list of keys: [source, cypher] ---- WITH ["name", "username", "bio", "following", "followers"] AS keysToKeep RETURN apoc.map.values($document, keysToKeep) ---- image::{{<siteurl>}}/uploads/2019/05/values.png[] And we can then use `apoc.map.fromLists` to reconstruct our map: [source,cypher] ---- WITH ["name", "username", "bio", "following", "followers"] AS keysToKeep RETURN apoc.map.fromLists(keysToKeep, apoc.map.values($document, keysToKeep)) AS result ---- image::{{<siteurl>}}/uploads/2019/05/keysToKeep.png[] This code forms part of the following query that creates a `User` node keyed on the `id` property, and then sets the other properties afterwards: [source, cypher] ---- WITH ["name", "username", "bio", "following", "followers"] AS keysToKeep MERGE (u:User {id: $document.id }) SET u += apoc.map.fromLists($keysToKeep, apoc.map.values($document, $keysToKeep)) ---- All done!

Learn how to create a node based on a subset of keys in a map.

neo4j-keep-filter-keys-map-apoc

https://markhneedham.com/blog/2019/05/12/neo4j-keep-filter-keys-map-apoc

2019-11-24 00:21:00

Elasticsearch: Importing data into App Search

For a side project that I'm working on I wanted to create a small React application that can query data stored in https://www.elastic.co/[Elasticsearch^], and https://www.elastic.co/blog/how-to-build-application-search-with-react-and-elastic-app-search[most of the tutorials^] I found suggested using a tool called Elastic App Search. I'd not heard of App Search before, and it took me a while to figure out that it's the mid level product in between Elasticsearch Service and Elastic Site Search Service, as described on https://www.elastic.co/cloud[elastic.co/cloud] image::{{<siteurl>}}/uploads/2019/11/appsearch.png[] == Launching Elastic App Search locally Now that we've figured that out we're going to setup a local running App Search server and import some data into it. I found a https://okode.com/demystified-elasticsearch-with-elastic-app-search/[Docker compose file^] on the Okode blog that I adapted to the following: _docker-compose.yml_ [source,yaml] ---- version: '3.7' services: elasticsearch: image: docker.elastic.co/elasticsearch/elasticsearch:7.4.2 environment: - "node.name=es-node" - "discovery.type=single-node" - "cluster.name=app-search-docker-cluster" - "bootstrap.memory_lock=true" - "ES_JAVA_OPTS=-Xms512m -Xmx2048m" ulimits: memlock: soft: -1 hard: -1 ports: - 9200:9200 - 9300:9300 appsearch: image: docker.elastic.co/app-search/app-search:7.4.2 depends_on: - elasticsearch environment: - "elasticsearch.host=http://elasticsearch:9200" - "allow_es_settings_modification=true" - "JAVA_OPTS=-Xmx2048m" ports: - 3002:3002 ---- We can run the following command to launch AppSearch: [source,bash] ---- docker-compose up ---- Once that command has run App Search should be running at http://localhost:3002/. If we navigate to that URL in our web browser, we'll see the following screen: image::{{<siteurl>}}/uploads/2019/11/create-engine.png[] We need to create an `engine`, which is App Search's name for an `index`. Let's create one called `meals`, as in the Okode tutorial mentioned earlier. Once we've done that we'll see the following screen, which has instructions for importing data into our engine: image::{{<siteurl>}}/uploads/2019/11/import-data.png[] But we're not going to use any of these approaches! == Installing the Python elastic-app-search library Instead we'll use the Python https://pypi.org/project/elastic-app-search/[elastic-app-search^] library to import data into AppSearch. We'll install the library using https://pipenv-fork.readthedocs.io/en/latest/index.html[Pipenv^] via the following Pipfile: _Pipfile_ [source,yaml] ---- [[source]] name = "pypi" url = "https://pypi.org/simple" verify_ssl = true [dev-packages] [packages] elastic-app-search = "*" requests = "*" stringcase = "*" [requires] python_version = "3.7" ---- We can set everything up by running the following commands: [source,bash] ---- pipenv shell pipenv install ---- Once we've run those commands, we can check that the library is installed by executing the following command: [source,bash] ---- pipenv graph ---- If we run that we'll see the following output: [source,bash] ---- elastic-app-search==7.4.0 - PyJWT [required: Any, installed: 1.7.1] - requests [required: Any, installed: 2.22.0] - certifi [required: >=2017.4.17, installed: 2019.9.11] - chardet [required: >=3.0.2,<3.1.0, installed: 3.0.4] - idna [required: >=2.5,<2.9, installed: 2.8] - urllib3 [required: >=1.21.1,<1.26,!=1.25.1,!=1.25.0, installed: 1.25.7] stringcase==1.2.0 ---- == Importing data We can now write a Python script to import some of the documents from https://themealdb.com/[themealdb.com^]: [source,python] ---- from elastic_app_search import Client import requests as r engine_name = 'meals' api_key = "private-kwicp7mhwssdxv54as9buzen" client = Client( api_key=api_key, base_endpoint='localhost:3002/api/as/v1', use_https=False ) response = r.get("https://www.themealdb.com/api/json/v1/1/search.php?f=a").json() documents = [] for entry in response["meals"]: documents.append(entry) if len(documents) % 50 == 0: res = client.index_documents(engine_name, documents) print(res) documents = [] res = client.index_documents(engine_name, documents) print(res) ---- We get the `api_key` via the `Credentials` menu item: image::{{<siteurl>}}/uploads/2019/11/credentials.png[] If we execute this script we'll see the following output: [source,text] ---- [{'id': None, 'errors': ['Fields can only contain lowercase letters, numbers, and underscores: idMeal.', 'Fields can only contain lowercase letters, numbers, and underscores: strMeal.', 'Fields can only contain lowercase letters, numbers, and underscores: strDrinkAlternate.', 'Fields can only contain lowercase letters, numbers, and underscores: strCategory.', 'Fields can only contain lowercase letters, numbers, and underscores: strArea.', 'Fields can only contain lowercase letters, numbers, and underscores: strInstructions.', 'Fields can only contain lowercase letters, numbers, and underscores: strMealThumb.', ... ]}] ---- We're not allowed to have fields that contain uppercase letters, so we'll need to fix that. We can use the https://pypi.org/project/stringcase/[stringcase^] library to fix this. The following script does this: [source,python] ---- from elastic_app_search import Client import requests as r import stringcase engine_name = 'meals' api_key = "private-kwicp7mhwssdxv54as9buzen" client = Client( api_key=api_key, base_endpoint='localhost:3002/api/as/v1', use_https=False ) response = r.get("https://www.themealdb.com/api/json/v1/1/search.php?f=a").json() documents = [] for entry in response["meals"]: new_entry = {stringcase.snakecase(key):entry[key] for key in entry} new_entry["id"] = new_entry["id_meal"] documents.append(new_entry) if len(documents) % 50 == 0: res = client.index_documents(engine_name, documents) print(res) documents = [] res = client.index_documents(engine_name, documents) print(res) ---- If we execute that query, we'll see the following output: [source,text] ---- [{'id': '52768', 'errors': []}, {'id': '52893', 'errors': []}] ---- And now let's navigate to http://localhost:3002/as#/engines/meals/documents to have a look at what we've imported: image::{{<siteurl>}}/uploads/2019/11/data-imported.png[] Success!

Learn how to import data into Elastic App Search

elasticsearch-import-data-appsearch

https://markhneedham.com/blog/2019/11/24/elasticsearch-import-data-appsearch

2019-07-29 11:08:00

Python: Click - Handling Date Parameter

I've been building a little CLI application using the https://click.palletsprojects.com/en/7.x/[Python Click Library^], and I wanted to pass in a Date as a parameter. There's more than one way to do this. Let's first install the Click library: [source,bash] ---- pip install click ---- And now we'll import our required libraries: [source, python] ---- from datetime import date import click ---- Now we'll create a sub command that takes two parameters: `date-start` and `date-end`. These parameters have the type `DateTime`, and we can pass a string in the format `yyyy-mm-dd` from the command line: [source,python] ---- @click.group() def cli(): pass @click.command() @click.option('--date-start', type=click.DateTime(formats=["%Y-%m-%d"]), default=str(date.today())) @click.option('--date-end', type=click.DateTime(formats=["%Y-%m-%d"]), default=str(date.today())) def dummy(date_start, date_end): click.echo(f"Start: {date_start}, End: {date_end} ") cli.add_command(dummy) if __name__ == '__main__': cli() ---- We can execute this command by running the following: [source, bash] ---- $ python blog.py dummy --date-start 2018-01-01 Start: 2018-01-01 00:00:00, End: 2019-07-29 00:00:00 ---- We only passed in `date-start`, which means `date-end` has defaulted to today's date. The dates are correct, but we still have time information that we don't want. We can get rid of that by calling the `date` function on those `DateTime` objects: [source,python] ---- @click.command() @click.option('--date-start', type=click.DateTime(formats=["%Y-%m-%d"]), default=str(date.today())) @click.option('--date-end', type=click.DateTime(formats=["%Y-%m-%d"]), default=str(date.today())) def dummy(date_start, date_end): date_start = date_start.date() date_end = date_end.date() click.echo(f"Start: {date_start}, End: {date_end} ") ---- Now let's run the script again: [source, bash] ---- $ python blog.py dummy --date-start 2018-01-01 Start: 2018-01-01, End: 2019-07-29 ---- Alternatively we can create our own Date type based on the DateTime type: [source, python] ---- from datetime import datetime class Date(click.ParamType): name = 'date' def __init__(self, formats=None): self.formats = formats or [ '%Y-%m-%d', '%Y-%m-%dT%H:%M:%S', '%Y-%m-%d %H:%M:%S' ] def get_metavar(self, param): return '[{}]'.format('|'.join(self.formats)) def _try_to_convert_date(self, value, format): try: return datetime.strptime(value, format).date() except ValueError: return None def convert(self, value, param, ctx): for format in self.formats: date = self._try_to_convert_date(value, format) if date: return date self.fail( 'invalid date format: {}. (choose from {})'.format( value, ', '.join(self.formats))) def __repr__(self): return 'Date' ---- And then we'll add a new command that uses this parameter type: [source, python] ---- @click.command() @click.option('--date-start', type=Date(formats=["%Y-%m-%d"]), default=str(date.today())) @click.option('--date-end', type=Date(formats=["%Y-%m-%d"]), default=str(date.today())) def dummy2(date_start, date_end): click.echo(f"Start: {date_start}, End: {date_end} ") cli.add_command(dummy2) ---- And now let's call this sub command: [source, python] ---- $ python blog.py dummy2 --date-start 2018-01-02 --date-end 2019-04-05 Start: 2018-01-02, End: 2019-04-05 ----

Learn how to process date parameters with the Python Click Library.

python-click-date-parameter-type

https://markhneedham.com/blog/2019/07/29/python-click-date-parameter-type

2019-07-31 11:08:00

Neo4j: Conditional WHERE clause with APOC

Sometimes we want to be able to vary our Cypher queries based on the value of a parameter. I came across such a situation today, and thought I'd share how I solved it using the APOC library. Let's first setup some sample data: [source, cypher] ---- UNWIND range(0, 5) AS id CREATE (:Person {name: "person-" + id}) ---- Now, if we want to get all pairs of people, we could write the following query: [source,cypher] ---- MATCH (p1:Person), (p2:Person) WHERE id(p1) < id(p2) RETURN p1.name, p2.name ---- If we run that query we'll see the following output: .Results [opts="header",cols="1,1"] |=== | p1.name | p2.name |person-0|person-1 |person-0|person-2 |person-0|person-3 |person-0|person-4 |person-0|person-5 |person-1|person-2 |person-1|person-3 |person-1|person-4 |person-1|person-5 |person-2|person-3 |person-2|person-4 |person-2|person-5 |person-3|person-4 |person-3|person-5 |person-4|person-5 |=== But we also want to search for a Person, based on a query parameter, and return the pairs from that Person to everybody else. Let's see how to do that. We'll first create a parameter to find a subset of people: [source, cypher] ---- :param personSubstring => "person-3" ---- And then the following query returns pairs from those people to everybody else: [source,cypher] ---- MATCH (p1:Person), (p2:Person) WHERE p1.name CONTAINS $personSubstring AND p2 <> p1 RETURN p1.name, p2.name ---- If we run that query we'll see the following output: .Results [opts="header",cols="1,1"] |=== | p1.name | p2.name |person-3|person-0 |person-3|person-1 |person-3|person-2 |person-3|person-4 |person-3|person-5 |=== What if we want to combine those queries? If `personSubstring` contains a value we'll use that, and if not we'll return all pairs of people. We can use the https://neo4j.com/developer/neo4j-apoc/[APOC Library^]'s `apoc.when` procedure to solve this problem. This procedure has the following parameters: * `condition` - a statement that must return a boolean value * `ifQuery` - the query to execute if the condition is true * `elseQuery` - the query to execute if the condition is false * `params` - parameters to pass to the queries And we can call it like this: [source,cypher] ---- CALL apoc.when( $personSubstring = '', 'MATCH (p1:Person), (p2:Person) WHERE id(p1) < id(p2) RETURN p1, p2', 'MATCH (p1:Person), (p2:Person) WHERE p1.name CONTAINS $personSubstring AND p2 <> p1 RETURN p1, p2', {personSubstring: $personSubstring}) YIELD value RETURN value.p1.name, value.p2.name ---- And now we can handle both use cases in one Cypher query.

Learn how to write conditional queries using the APOC library.

neo4j-conditional-where-query-apoc

https://markhneedham.com/blog/2019/07/31/neo4j-conditional-where-query-apoc

2019-09-05 00:47:00

Neo4j: apoc.load.csv - Neo.ClientError.Statement.SyntaxError: Type mismatch: expected Float, Integer, Number or String but was Any

The https://neo4j.com/docs/labs/apoc/3.5/[Neo4j APOC library^]'s https://neo4j.com/docs/labs/apoc/current/import/load-csv/[Load CSV procedure^] is very useful if you want more control over the import process than the `LOAD CSV` clause allows. I found myself using it last week to import a CSV file of embeddings, because I wanted to know the line number of the row in the CSV file while importing the data. I had a file that looked like this, which I put into the `import` directory: [source,text] ---- $ cat import/data.csv 0.034 0.765 0.452 0.312 0.413 0.789 ---- And before I imported it, I added the following entries to my Neo4j Settings file so that I could read locally: [source,text] ---- apoc.import.file.enabled=true apoc.import.file.use_neo4j_config=true ---- The following query processes the file, with optional config that indicates that the file doesn't have a header and uses a space as a separator: [source,cypher] ---- CALL apoc.load.csv("file:///data.csv", {header: false, sep: " "}) YIELD lineNo, map, list RETURN lineNo, list ---- If we run that query we'll get this result: .Results [opts="header",cols="1,2"] |=== | lineNo | list |0 | ["0.034","0.765","0.452"] |1 | ["0.312","0.413","0.789"] |=== But we want to have each item of the list be a float value rather than a string. I tried to coerce each of the values using the `toFloat` function: [source,cypher] ---- CALL apoc.load.csv("file:///data.csv", {header: false, sep: " "}) YIELD lineNo, map, list RETURN lineNo, [item in list | toFloat(item)] AS list ---- Unfortunately this doesn't quite work, as the following error indicates: [source,text] ---- Neo.ClientError.Statement.SyntaxError: Type mismatch: expected Float, Integer, Number or String but was Any (line 3, column 40 (offset: 129)) "RETURN lineNo, [item in list | toFloat(item)] AS list" ^ ---- Instead we need to use APOC's `apoc.convert.toFloat` function to do the type coercion. The following query does the trick: [source,cypher] ---- CALL apoc.load.csv("file:///data.csv", {header: false, sep: " "}) YIELD lineNo, map, list RETURN lineNo, [item in list | apoc.convert.toFloat(item)] AS list ---- .Results [opts="header",cols="1,2"] |=== | lineNo | list |0 | [0.034,0.765,0.452] |1 | [0.312,0.413,0.789] |===

Learn how to coerce the type of values loaded using APOC's apoc.load.csv procedure.

neo4j-apoc-load-csv-type-mismatch-expected-float-integer-number-string

https://markhneedham.com/blog/2019/09/05/neo4j-apoc-load-csv-type-mismatch-expected-float-integer-number-string

2019-09-20 00:47:00

Graphing Brexit: Plotting how the parties voted

Over the last week I've revisited the https://towardsdatascience.com/graphing-brexit-bbe4314cf70[Brexit Graph^] that I created in March 2019, this time looking at how the parties voted on average on each of the indicative votes. To recap, we have a graph that has the following schema: image::{{<siteurl>}}/uploads/2019/09/schema.png[] Since the initial post I've slightly changed how the `MEMBER_OF` relationship works. As several MPs have switched MPs in the intervening months, we're now storing a `start` property to indicate when they started representing a party and an `end` property to indicate when they stopped representing a party. This change to the model means that we need to do some filtering to work out which party they were representing when a vote took place. We want to work out how the parties voted on average across each of the different Brexit indicative votes. We want to come up with a score for each party for each vote, which we'll do by aggregating scores by party using the following scoring system for each MP's position on a particular motion: * Voted For = 1.0 * Did Not Vote = 0.5 * Voted Against = 0 Let's run through a couple of examples: * If everybody in a party voted `For` a motion, that party would have a score of 1.0 for that motion. * If everybody in a party voted `Against` a motion, that party would have a score of 0.0 for that motion. * If a party has 5 MPs, with 3 voting `For` and 2 `Did Not Vote`, they'd have a score of `(3*1.0 + 2*0.5)/5 = 4/5 = 0.8` We can compute the average vote by a party for a motion by writing the following query: [source,cypher] ---- MATCH (m:Motion)<-[vote]-(person:Person) MATCH (person)-[memberOf:MEMBER_OF]->(party) WHERE memberOf.start <= m.date AND (not(exists(memberOf.end)) OR m.date <= memberOf.end) WITH m, party, CASE WHEN type(vote) = "FOR" THEN 1 WHEN type(vote) = "DID_NOT_VOTE" THEN 0.5 ELSE 0 END AS score RETURN m.division, m.name, party.name, avg(score) AS score, count(*) AS count ORDER BY m.name, party.name LIMIT 10 ---- .Results [opts="header",cols="1,3,1,1,1"] |=== | m.division | m.name | party.name | score | count | "388" | "George Eustice's motion H (EFTA and EEA)" | "Conservative" | 0.2731629392971245 | 313 | "388" | "George Eustice's motion H (EFTA and EEA)" | "DUP" | 0.5 | 10 | "388" | "George Eustice's motion H (EFTA and EEA)" | "Green" | 0.0 | 1 | "388" | "George Eustice's motion H (EFTA and EEA)" | "Independent" | 0.15 | 20 | "388" | "George Eustice's motion H (EFTA and EEA)" | "Labour" | 0.25352112676056326 | 213 | "388" | "George Eustice's motion H (EFTA and EEA)" | "Labour/Co-operative" | 0.265625 | 32 | "388" | "George Eustice's motion H (EFTA and EEA)" | "Liberal Democrat" | 0.4090909090909091 | 11 | "388" | "George Eustice's motion H (EFTA and EEA)" | "Plaid Cymru" | 0.5 | 4 | "388" | "George Eustice's motion H (EFTA and EEA)" | "Scottish National Party" | 0.014285714285714292 | 35 | "388" | "George Eustice's motion H (EFTA and EEA)" | "Sinn Féin" | 0.5 | 7 |=== None of the parties much liked this motion. The best average score it received was by the DUP, Plaid Cymru, and Sinn Fein, whose members on average didn't vote. For the other parties the average score indicates that at least some of their members were against it. We'll store this information in our graph, by creating a relationship between the party and the motion, storing the average rating as a property on the relationship: [source,cypher] ---- MATCH (m:Motion)<-[vote]-(person:Person) MATCH (person)-[memberOf:MEMBER_OF]->(party) WHERE memberOf.start <= m.date AND (not(exists(memberOf.end)) OR m.date <= memberOf.end) WITH m, party, CASE WHEN type(vote) = "FOR" THEN 1 WHEN type(vote) = "DID_NOT_VOTE" THEN 0.5 ELSE 0 END AS score WITH m, party, avg(score) AS score MERGE (party)-[averageVote:AVERAGE_VOTE]->(m) SET averageVote.score = score; ---- The image below shows the relationships that have been created: image::{{<siteurl>}}/uploads/2019/09/average-votes.svg[] Now that we've stored the average votes, we can write a bit of Python code to create bar charts showing the average votes across all the motions. The following code does this: [source,python] ---- import matplotlib.pyplot as plt import numpy as np from neo4j import GraphDatabase driver = GraphDatabase.driver("bolt://localhost") fig = plt.figure() fig.set_size_inches(16.5, 14.5) fig.subplots_adjust(hspace=0.4, wspace=0.4) with driver.session() as session: for index, division in enumerate(["386", "387", "388", "389", "390", "391", "392", "393"]): rows = session.run(""" MATCH (m:Motion {division: $division}) RETURN m.name AS motion """, {"division": division}) motion = rows.peek()["motion"] rows = session.run(""" MATCH (p:Party)-[vote:AVERAGE_VOTE]->(m:Motion {division: $division}) RETURN p.name AS party, vote.score AS score, m.name AS motion ORDER BY party """, {"division": division}) result = [{"party": row["party"], "score": row["score"]} for row in rows ] plt.rcdefaults() ax = fig.add_subplot(4, 2, index+1) parties = [item["party"] for item in result] y_pos = np.arange(len(parties)) scores = [item["score"] for item in result] ax.barh(y_pos, scores, align='center') ax.set_yticks(y_pos) ax.set_yticklabels(parties) ax.invert_yaxis() ax.set_title(motion) ax.set_xlim([0,1]) plt.xticks([0, 0.5, 1], ["Against", "Did Not Vote", "For"]) plt.tight_layout() plt.savefig("images/votes.svg") plt.close() ---- If we execute this script, we'll see the following charts: image::{{<siteurl>}}/uploads/2019/09/votes.svg[] From this diagram we can clearly see that the only indicative vote that Conservative Party members in general voted in favour of was the Mr Baron's Motion B (no deal). I guess this chart actually foresaw the current composition of the cabinet. All the parties except for the DUP, Sinn Fein, and the Conservatives on average were in favour of Joanna Cherry's motion L (Revocation to avoid no deal) and Margaret Beckett's motion M (Confirmatory public vote). Only Labour MPs voted on average in favour of Jeremy Corbyn's motion K (Labour's alternative plan) and Mr Clarke's motion J (Customs union). Now that we've computed the average vote of each party we can see whether there are people whose average votes correlate more closely with another party than their own. We'll have a look at that in our next post in the Graphing Brexit series.

In this post we create bar charts of the Brexit indicative votes using Python's matplotlib library.

graphing-brexit-charting-how-the-parties-voted

https://markhneedham.com/blog/2019/09/20/graphing-brexit-charting-how-the-parties-voted

2019-09-18 00:47:00

Neo4j: Approximate string matching/similarity

I've been playing with the https://towardsdatascience.com/graphing-brexit-clustering-edition-3b966694e723[Brexit Graph^] over the last few days, and wanted to map the MPs that I got from https://commonsvotes.digiminster.com/?page=1[CommonsVotes^] with data from the https://www.theyworkforyou.com/api[TheyWorkForYou API^]. I already had voting records loaded into Neo4j, but to recap, this is how I did that: [source,cypher] ---- UNWIND [655,656,657,658,659,660,661,662,711, 669, 668, 667, 666, 664] AS division LOAD CSV FROM "https://github.com/mneedham/graphing-brexit/raw/master/data/commonsvotes/Division" + division + ".csv" AS row // Create motion nodes WITH division, collect(row) AS rows MERGE (motion:Motion {division: trim(split(rows[0][0], ":")[1]) }) SET motion.name = rows[2][0], motion.date = date(datetime({epochmillis: apoc.date.parse(trim(split(rows[1][0], ":")[1]), "ms", "dd/MM/yyyy")})) // Skip the first 6 rows as they have metadata we don't need WITH motion, rows UNWIND rows[7..] AS row // Create person, party, constituency, and corresponding rels MERGE (person:Person {name: row[0]}) MERGE (constituency:Constituency {name: row[2]}) MERGE (person)-[:REPRESENTS]->(constituency) WITH person, motion, CASE WHEN row[3] = "Aye" THEN "FOR" WHEN row[3] = "No" THEN "AGAINST" ELSE "DID_NOT_VOTE" END AS vote CALL apoc.merge.relationship(person, vote, {}, {}, motion) YIELD rel RETURN count(*); ---- I extracted MP details from the TheyWorkForYou API and stored them in a JSON file so that I could load them into Neo4j. == Exact Match We'll start by finding exact matches between MPs in both datasets. The following query does this: [source,cypher] ---- WITH "https://github.com/mneedham/graphing-brexit/raw/master/data/mps_formatted.json" AS uri CALL apoc.load.json(uri) YIELD value MATCH (other:Person) WHERE other.name = value.name SET other.id = value.person_id; ---- This sorts out the mapping for 612 MPs, which is the vast majority of them. We only have 40 MPs where we'll need to be a bit clever. == String Similarity Given that there's not an exact string match, another approach is to compare the string similarity of names. But before we do that we'll reduce the number of string similarity computations we have to do by building a list of potential matches for each person based on having the same last name. The following query does this: [source,cypher] ---- WITH "https://github.com/mneedham/graphing-brexit/raw/master/data/mps_formatted.json" AS uri CALL apoc.load.json(uri) YIELD value WITH collect(value) AS values MATCH (other:Person) WHERE not(exists(other.id)) WITH other, [value in values WHERE other.name ENDS WITH split(value.name, " ")[-1] | value.name] AS potentialValues RETURN other.name, potentialValues LIMIT 10 ---- If we execute this query, we'll get the following results: .Results [opts="header",cols="1,2"] |=== | other.name | potentialValues | "David Amess" | ["Sir David Amess"] | "Graham P Jones" | ["Marcus Jones", "Andrew Jones", "Susan Elan Jones", "Kevan Jones", "Graham Jones", "Darren Jones", "David Jones", "Gerald Jones", "Helen Jones", "Sarah Jones"] | "Steve Baker" | ["Steven Baker"] | "Joseph Johnson" | ["Diana R. Johnson", "Boris Johnson", "Gareth Johnson", "Jo Johnson", "Dr Caroline Johnson"] | "Nick Boles" | ["Nicholas Boles"] | "Nicholas Brown" | ["Nick Brown", "Alan Brown", "Lyn Brown"] | "Vince Cable" | ["Vincent Cable"] | "William Cash" | ["Bill Cash"] | "Th?r?se Coffey" | ["Therese Coffey", "Ann Coffey"] | "Nic Dakin" | ["Nicholas Dakin"] |=== For some MPs there's only one potential match, which makes our life easy, but for others we have a few to choose from. Now we can apply string similarity algorithms to work out which of these names is the best match. The APOC library supports several https://neo4j.com/docs/labs/apoc/3.5/misc/text-functions/#text-functions-text-similarity[text similarity functions^], including Sorensen Dice Similarity and Jaro Winkler Distance. I decided to use both of them, and pick the candidate that had the highest average score, because each of them excelled on different names. Let's give it a try for one of our MPs to see how it fares. The following query finds the most similar people to Joseph Johnson: [source,cypher] ---- WITH "https://github.com/mneedham/graphing-brexit/raw/master/data/mps_formatted.json" AS uri CALL apoc.load.json(uri) YIELD value WITH collect(value) AS values match (other:Person {name: "Joseph Johnson"}) WHERE not(exists(other.id)) with other, [value in values WHERE other.name ends with split(value.name, " ")[-1] | value.name] AS potentialValues WHERE size(potentialValues) > 1 WITH other, apoc.coll.sortMaps([value in potentialValues | { value: value, sorensen: apoc.text.sorensenDiceSimilarity(other.name, value), jaro: apoc.text.jaroWinklerDistance(other.name, value), average: apoc.coll.avg([apoc.text.sorensenDiceSimilarity(other.name, value), apoc.text.jaroWinklerDistance(other.name, value)]) }], "average") AS similarities UNWIND similarities AS similarity RETURN similarity.value, similarity.sorensen, similarity.jaro ---- If we execute this query, we'll get the following results: .Results [opts="header",cols="1,1,1"] |=== | similarity.value | similarity.sorensen | similarity.jaro | "Jo Johnson" | 0.7777777777777778 | 0.8326530612244899 | "Gareth Johnson" | 0.5454545454545454 | 0.8095238095238096 | "Boris Johnson" | 0.5714285714285714 | 0.7278388278388279 | "Diana R. Johnson" | 0.5454545454545454 | 0.6071428571428571 | "Dr Caroline Johnson" | 0.48 | 0.5802005012531328 |=== Pretty good, it found the right person for this one. What about if we match Stewart McDonald, where there's actually another person with a different spelling of the same first name? [source, cypher] ---- WITH "https://github.com/mneedham/graphing-brexit/raw/master/data/mps_formatted.json" AS uri CALL apoc.load.json(uri) YIELD value WITH collect(value) AS values match (other:Person {name: "Stewart Malcolm McDonald"}) WHERE not(exists(other.id)) with other, [value in values WHERE other.name ends with split(value.name, " ")[-1] | value.name] AS potentialValues WHERE size(potentialValues) > 1 WITH other, apoc.coll.sortMaps([value in potentialValues | { value: value, sorensen: apoc.text.sorensenDiceSimilarity(other.name, value), jaro: apoc.text.jaroWinklerDistance(other.name, value), average: apoc.coll.avg([apoc.text.sorensenDiceSimilarity(other.name, value), apoc.text.jaroWinklerDistance(other.name, value)]) }], "average") AS similarities UNWIND similarities AS similarity RETURN similarity.value, similarity.sorensen, similarity.jaro ---- If we execute this query, we'll get the following results: .Results [opts="header",cols="1,1,1"] |=== | similarity.value | similarity.sorensen | similarity.jaro | "Stewart McDonald" | 0.8125 | 0.8914930555555556 | "Stuart McDonald" | 0.6451612903225806 | 0.7868386243386243 | "Andy McDonald" | 0.4827586206896552 | 0.540954415954416 |=== I tried this across all the MPs that didn't have an exact name match and it found the correct person every time, based on manual inspection. So for this data set it seems to be a decent approach. We'll finish the mapping by storing the `person_id` of the best matching person: [source,cypher] ---- WITH "https://github.com/mneedham/graphing-brexit/raw/master/data/mps_formatted.json" AS uri CALL apoc.load.json(uri) YIELD value WITH collect(value) AS values match (other:Person) WHERE not(exists(other.id)) with other, [value in values WHERE other.name ends with split(value.name, " ")[-1] | value] AS potentialValues WITH other, apoc.coll.sortMaps([value in potentialValues | { value: value, sorensen: apoc.text.sorensenDiceSimilarity(other.name, value.name), jaro: apoc.text.jaroWinklerDistance(other.name, value.name), average: apoc.coll.avg([apoc.text.sorensenDiceSimilarity(other.name, value), apoc.text.jaroWinklerDistance(other.name, value)]) }], "average") AS similarities SET other.id = similarities[0].value.person_id ---- Let's check if we have any MPs that haven't been mapped yet. We can do this by executing the following query: [source, cypher] ---- MATCH (p:Person) WHERE not(exists(p.id)) RETURN p.name ---- .Results [opts="header",cols="1"] |=== | p.name | "Ian Paisley" | "Liz Saville Roberts" |=== Let's filter the KnowYourMP data to figure out what's going on: [source,cypher] ---- call apoc.load.json("https://github.com/mneedham/graphing-brexit/raw/master/data/mps_formatted.json") YIELD value WHERE value.name contains "Saville" OR value.name contains "Paisley" RETURN value.name ---- .Results [opts="header",cols="1"] |=== | value.name | "Liz Saville-Roberts" | "Ian Paisley Jnr" |=== Both these people are in the dataset, but their different surnames meant that they got filtered out before we applied the text similarity algorithms. [source,cypher] ---- WITH "https://github.com/mneedham/graphing-brexit/raw/master/data/mps_formatted.json" AS uri CALL apoc.load.json(uri) YIELD value WITH collect(value) AS potentialValues match (other:Person) WHERE not(exists(other.id)) WITH other, apoc.coll.sortMaps([value in potentialValues | { value: value, sorensen: apoc.text.sorensenDiceSimilarity(other.name, value.name), jaro: apoc.text.jaroWinklerDistance(other.name, value.name), average: apoc.coll.avg([apoc.text.sorensenDiceSimilarity(other.name, value.name), apoc.text.jaroWinklerDistance(other.name, value.name)]) }], "average") AS similarities SET other.id = similarities[0].value.person_id ---- If we remove that filtering criteria we'll be able to match these people: [source, cypher] ---- WITH "https://github.com/mneedham/graphing-brexit/raw/master/data/mps_formatted.json" AS uri CALL apoc.load.json(uri) YIELD value WITH collect(value) AS potentialValues match (other:Person) WHERE not(exists(other.id)) WITH other, apoc.coll.sortMaps([value in potentialValues | { value: value, sorensen: apoc.text.sorensenDiceSimilarity(other.name, value.name), jaro: apoc.text.jaroWinklerDistance(other.name, value.name), average: apoc.coll.avg([apoc.text.sorensenDiceSimilarity(other.name, value.name), apoc.text.jaroWinklerDistance(other.name, value.name)]) }], "average") AS similarities RETURN other.name, similarities[0].value.name AS name, similarities[0].value.person_id AS personId ---- .Results [opts="header", cols="1,1,1"] |=== | other.name |name | personId | "Ian Paisley" | "Ian Paisley Jnr" | "13852" | "Liz Saville Roberts" | "Liz Saville-Roberts" | "25302" |=== And then we can tweak the query to store the personId: [source, cypher] ---- WITH "https://github.com/mneedham/graphing-brexit/raw/master/data/mps_formatted.json" AS uri CALL apoc.load.json(uri) YIELD value WITH collect(value) AS potentialValues match (other:Person) WHERE not(exists(other.id)) WITH other, apoc.coll.sortMaps([value in potentialValues | { value: value, sorensen: apoc.text.sorensenDiceSimilarity(other.name, value.name), jaro: apoc.text.jaroWinklerDistance(other.name, value.name), average: apoc.coll.avg([apoc.text.sorensenDiceSimilarity(other.name, value.name), apoc.text.jaroWinklerDistance(other.name, value.name)]) }], "average") AS similarities SET other.id = similarities[0].value.person_id ---- And now all of our MPs are mapped!

Learn how to use the text matching algorithms in the APOC Library.

neo4j-string-matching-similarity

https://markhneedham.com/blog/2019/09/18/neo4j-string-matching-similarity

2019-09-27 00:47:00

Graphing Brexit: Did the threat work?

Following on from the blog post where https://markhneedham.com/blog/2019/09/23/graphing-brexit-mps-vs-parties/[we compared how MPs and parties voted on Brexit indicative measures^], in this post we're going to explore how Conservative MPs have voted with respect to a no deal exit from the European Union. In particular we'd like to know whether the threat to have the party whip removed had an impact on how they voted in the recent motion to request an extension to work out a deal. We know that there were 21 people who voted against the government and had the Conservative party whip removed. Presumably the other Conservative MPs voted against the motion, but let's check by executing the following query: [source,cypher] ---- MATCH (m:Motion {division: "439"})<-[vote]-(person:Person), (person)-[memberOf:MEMBER_OF]->(:Party {name: "Conservative"}) WHERE memberOf.start <= m.date AND (not(exists(memberOf.end)) OR m.date <= memberOf.end) RETURN type(vote), count(*) ORDER BY count(*) DESC ---- .Results [opts="header",cols="1,1"] |=== | type(vote) | count(*) | "AGAINST" | 286 | "FOR" | 21 | "DID_NOT_VOTE" | 3 |=== As expected, most MPs voted against this motion, although there are also 3 MPs who didn't vote. Those 3 who didn't vote might be interesting to explore further, but for now let's go back to our 21 rebel MPs and see how they voted on Mr Baron's motion B (No deal) back in March. We'd assume that they mostly voted against this motion, but let's check by executing the following query: [source,cypher] ---- MATCH (person:Person)-[vote]->(:Motion {division: "386"}) WHERE (person)-[:MEMBER_OF {end: date({year: 2019, month: 9, day: 3})}] ->(:Party {name: "Conservative"}) WITH person, type(vote) AS vote ORDER BY person.pageviews DESC RETURN person.name AS person, vote ---- .Results [opts="header",cols="1,1"] |=== | person | vote | "Philip Hammond" | "DID_NOT_VOTE" | "Kenneth Clarke" | "AGAINST" | "Justine Greening" | "AGAINST" | "Rory Stewart" | "AGAINST" | "Nicholas Soames" | "AGAINST" | "Dominic Grieve" | "AGAINST" | "Oliver Letwin" | "AGAINST" | "Sam Gyimah" | "AGAINST" | "David Gauke" | "DID_NOT_VOTE" | "Greg Clark" | "DID_NOT_VOTE" | "Caroline Nokes" | "DID_NOT_VOTE" | "Margot James" | "AGAINST" | "Richard Benyon" | "AGAINST" | "Anne Milton" | "AGAINST" | "Guto Bebb" | "AGAINST" | "Alistair Burt" | "AGAINST" | "Antoinette Sandbach" | "AGAINST" | "Stephen Hammond" | "AGAINST" | "Steve Brine" | "AGAINST" | "Richard Harrington" | "AGAINST" | "Edward Vaizey" | "AGAINST" |=== Unsurprisingly most of these people voted against the motion, 17 in total, although the 4 cabinet members didn't vote. Were there any other Conservative MPs who also voted against Mr Baron's motion B (No deal)? We can find out by executing the following query: [source,cypher] ---- MATCH (person:Person)-[vote]->(m:Motion {division: "386"}), (person)-[memberOf:MEMBER_OF]->(:Party {name: "Conservative"}) WHERE memberOf.start <= m.date AND (not(exists(memberOf.end)) OR m.date <= memberOf.end) RETURN type(vote), count(*) ORDER BY count(*) DESC ---- .Results [opts="header",cols="1,1"] |=== | type(vote) | count(*) | "FOR" | 156 | "AGAINST" | 94 | "DID_NOT_VOTE" | 63 |=== There were 94 people who voted against this motion. We know who 17 of them are, but that still leaves us with 77 others. [source,cypher] ---- MATCH (person:Person)-[:AGAINST]->(noDealMotion:Motion {division: "386"}), (person)-[memberOf:MEMBER_OF]->(:Party {name: "Conservative"}) WHERE memberOf.start <= noDealMotion.date AND (not(exists(memberOf.end)) OR noDealMotion.date <= memberOf.end) MATCH (person)-[vote]->(:Motion {division: "439"}) RETURN type(vote) AS vote, collect(person.name) AS people, count(*) AS count ---- .Results [opts="header",cols="1,3,1"] |=== | vote | people | count | "AGAINST" | ["James Cartlidge", "John Glen", "David Morris", "John Howell", "Luke Hall", "Mark Pawsey", "Chris Skidmore", "Alister Jack", "Gary Streeter", "Luke Graham", "Mel Stride", "Kevin Hollinrake", "Jeremy Lefroy", "Jeremy Quin", "Bim Afolami", "Victoria Atkins", "Michael Fallon", "Robin Walker", "Damien Moore", "Oliver Heald", "Nick Herbert", "Nicky Morgan", "Alok Sharma", "Keith Simpson", "Stephen Kerr", "Paul Masterton", "Alberto Costa", "Nick Hurd", "Colin Clark", "Lucy Frazer", "Stephen Crabb", "Bill Grant", "Andrew Bowie", "Robert Buckland", "Richard Graham", "Damian Collins", "Roger Gale", "David Duguid", "Jesse Norman", "Guy Opperman", "Charles Walker", "Paul Beresford", "Greg Hands", "Hugo Swire", "Patrick McLoughlin", "Mark Prisk", "Vicky Ford", "Damian Green", "Joseph Johnson", "Victoria Prentis", "Mike Freer", "Oliver Dowden", "David Tredinnick", "Andrew Jones", "Helen Grant", "George Hollingbery", "Robert Halfon", "Cheryl Gillan", "Rebecca Pow", "Simon Hoare", "Gillian Keegan", "Alex Chalk", "Graham Brady", "Caroline Spelman", "Peter Bottomley", "Alan Duncan", "Sarah Newton", "Robert Neill", "Philip Dunne", "Peter Aldous", "Jo Churchill", "Mark Field", "Seema Kennedy", "Jonathan Djanogly", "Mary Robinson"] | 75 | "FOR" | ["Stephen Hammond", "Justine Greening", "Steve Brine", "Rory Stewart", "Margot James", "Oliver Letwin", "Dominic Grieve", "Sam Gyimah", "Nick Boles", "Richard Benyon", "Alistair Burt", "Guto Bebb", "Kenneth Clarke", "Edward Vaizey", "Anne Milton", "Richard Harrington", "Nicholas Soames", "Antoinette Sandbach", "Phillip Lee"] | 19 |=== We can see that there are 19 people who voted against both motions. If we add our 4 ex-Cabinet members, that gives us a total of 23 former Conservative MPs who voted against the motion to ask for an extension. 21 of those had the party whip removed on the 3rd September 2019, but that still gives us two that are not accounted for. We can write the following query to find out what happened with them: [source, cypher] ---- MATCH (person:Person)-[:AGAINST]->(noDealMotion:Motion {division: "386"}), (person)-[memberOf:MEMBER_OF]->(:Party {name: "Conservative"}) WHERE memberOf.start <= noDealMotion.date AND (not(exists(memberOf.end)) OR noDealMotion.date <= memberOf.end) MATCH (person)-[vote]->(:Motion {division: "439"}) WITH person, type(vote) AS vote, memberOf WHERE memberOf.end <> date({year: 2019, month: 9, day: 3}) RETURN person.name AS person, vote, memberOf.end AS endDate ---- .Results [opts="header",cols="1,3,1"] |=== | person | vote | endDate | "Nick Boles" | "FOR" | 2019-04-01 | "Phillip Lee" | "FOR" | 2019-09-02 |=== Phillip Lee famously walked across the floor of the Houses of Common to join the Liberal Democrats, and I'd forgotten that Nick Boles had resigned the party whip back in April. That still leave us with 75 people who changed their position from not supporting no deal in March 2019, but wanting to leave it as an option in September 2019. We'll give them the benefit of the doubt, and make the assumption that if they represent a constituency that had a majority vote to leave the EU their constituents may be happy with them taking this position. But if their constituency had a majority vote to remain in the EU we can't really make that argument. The following query find MPs who voted against no deal in the indicative votes, voted against asking for an extension (under threat of having the party whip removed), and represent constituencies that want to remain in the EU: [source, cypher] ---- MATCH (person:Person)-[:AGAINST]->(noDealMotion:Motion {division: "386"}), (person)-[memberOf:MEMBER_OF]->(:Party {name: "Conservative"}) WHERE memberOf.start <= noDealMotion.date AND (not(exists(memberOf.end)) OR noDealMotion.date <= memberOf.end) MATCH (person)-[:AGAINST]->(:Motion {division: "439"}), (person)-[:REPRESENTS]->(constituency:Constituency) WHERE constituency.leave < 50 RETURN person.name AS person, constituency.name AS constituency, constituency.leave AS leavePercentage ORDER BY constituency.leave ---- .Results [opts="header",cols="1,1,1"] |=== | person | constituency | leavePercentage | "Paul Masterton" | "East Renfrewshire" | 25.7 | "Mark Field" | "Cities of London and Westminster" | 28.1 | "Greg Hands" | "Chelsea and Fulham" | 29.1 | "Mike Freer" | "Finchley and Golders Green" | 31.1 | "Stephen Kerr" | "Stirling" | 32.3 | "Graham Brady" | "Altrincham and Sale West" | 38.6 | "Andrew Bowie" | "West Aberdeenshire and Kincardine" | 39.2 | "Luke Graham" | "Ochil and South Perthshire" | 39.5 | "Bim Afolami" | "Hitchin and Harpenden" | 39.8 | "Mary Robinson" | "Cheadle" | 42.7 | "Alex Chalk" | "Cheltenham" | 42.9 | "John Howell" | "Henley" | 43.1 | "Bill Grant" | "Ayr, Carrick and Cumnock" | 44.1 | "Colin Clark" | "Gordon" | 44.3 | "Cheryl Gillan" | "Chesham and Amersham" | 45.0 | "Alister Jack" | "Dumfries and Galloway" | 45.1 | "Guy Opperman" | "Hexham" | 45.3 | "Lucy Frazer" | "South East Cambridgeshire" | 45.3 | "Sarah Newton" | "Truro and Falmouth" | 45.9 | "Damien Moore" | "Southport" | 46.3 | "Andrew Jones" | "Harrogate and Knaresborough" | 47.2 | "Paul Beresford" | "Mole Valley" | 47.3 | "Mark Prisk" | "Hertford and Stortford" | 49.2 | "Jeremy Quin" | "Horsham" | 49.5 | "Nick Hurd" | "Ruislip, Northwood and Pinner" | 49.5 | "Nick Herbert" | "Arundel and South Downs" | 49.7 | "Robert Neill" | "Bromley and Chislehurst" | 49.8 | "John Glen" | "Salisbury" | 49.9 |=== Several of these constituencies are in London or Scotland. They would therefore seem like seats that remain favouring parties, such as the Liberal Democrats or SNP would try to win in the next election. And indeed the Liberal Democrats have already moved MPs into https://inews.co.uk/news/politics/chuka-umunna-lib-dem-candidate-constituency-cities-of-london-and-westminster-general-election-2019-496539[Cities of London and Westminster^] and https://www.bbc.co.uk/news/uk-politics-49831648[Finchley and Golders Green^]. It'll be fascinating to see what will happen with these seats if and when there is actually an election. If you have any other ideas for how we can explore the data, let me know in the comments.

In this post we look at how Conservative MPs feel about a no deal exit from the European Union.

graphing-brexit-did-the-threat-work

https://markhneedham.com/blog/2019/09/27/graphing-brexit-did-the-threat-work

2019-09-23 00:47:00

Graphing Brexit: MPs vs Parties

In the previous post of the Graphing Brexit series we computed https://markhneedham.com/blog/2019/09/20/graphing-brexit-charting-how-the-parties-voted/[the average vote by party^]. In this post we're going to take those average party scores and compare them against the votes placed by individual MPs. The goal is to determine whether, Brexit wise, MPs are representing the right party! It won't be perfect since we know that https://towardsdatascience.com/graphing-brexit-clustering-edition-3b966694e723[not everyone in a party voted the same way^], but it should still give us some fun results. Let's start with the https://www.theguardian.com/politics/2019/sep/03/what-have-tory-rebels-voted-for-and-will-an-election-now-happen[Conservative MPs who lost the party whip^] after voting to stop the government forcing a no-deal departure from the EU on 31st October. We can find those MPs by executing the following query: [source,cypher] ---- MATCH (person:Person)-[vote]->(m:Motion {date: date({year: 2019, month: 3, day: 27})}) WHERE (person)-[:MEMBER_OF {end: date({year: 2019, month: 9, day: 3})}] ->(:Party {name: "Conservative"}) RETURN person ---- image::{{<siteurl>}}/uploads/2019/09/conservative-rebels.svg[] Now let's compare their vote on each of the indicative motions with the average vote of each party on those motions. To recap from our last post, we're going to represent an MP's vote on a motion using the following scoring system: * `0.0` means an MP voted against a motion * `0.5` means an MP didn't vote on a motion * `1.0` means an MP voted for a motion We'll then build an lists of all their votes, and compare it against a list of the average votes of each of the parties using the https://neo4j.com/docs/graph-algorithms/current/experimental-algorithms/cosine/[Cosine Similarity algorithm^] in the Graph Algorithms Library. [quote, Neo4j Graph Algorithms Documentation] _____ Cosine similarity is the cosine of the angle between two n-dimensional vectors in an n-dimensional space. It is the dot product of the two vectors divided by the product of the two vectors' lengths (or magnitudes). _____ The query below shows how we'd build those lists of votes for one person, and then shows the result of running the Cosine Similarity function on those lists: [source,cypher] ---- MATCH (person:Person) WHERE (person)-[:MEMBER_OF {end: date({year: 2019, month: 9, day: 3})}] ->(:Party {name: "Conservative"}) WITH person LIMIT 1 MATCH (person)-[vote]->(m:Motion {date: date({year: 2019, month: 3, day: 27})}) MATCH (party:Party)-[ave:AVERAGE_VOTE]->(m) WHERE party.name <> "Speaker" WITH person, party, collect(CASE WHEN type(vote) = "FOR" THEN 1 WHEN type(vote) = "DID_NOT_VOTE" THEN 0.5 ELSE 0 END) AS personVotes, collect(ave.score) AS partyVotes RETURN person.name AS person, party.name AS party, personVotes, partyVotes, algo.similarity.cosine(personVotes, partyVotes) AS score ---- If we run that query we'll see the following results: .Results [opts="header",cols="1,1,2,2,1"] |=== | person | party | personVotes | partyVotes | score | "Guto Bebb" | "Conservative" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.5990415335463259, 0.20127795527156556, 0.2731629392971245, 0.1821086261980831, 0.062300319488817986, 0.10383386581469652, 0.10862619808306717, 0.5047923322683707] | 0.3760038974248517 | "Guto Bebb" | "Labour" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.02582159624413147, 0.7042253521126761, 0.25352112676056326, 0.927230046948357, 0.9600938967136151, 0.6643192488262909, 0.8403755868544599, 0.03521126760563382] | 0.7186196179929221 | "Guto Bebb" | "Green" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 1.0, 0.0] | 0.7071067811865476 | "Guto Bebb" | "Plaid Cymru" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.0, 1.0, 0.5, 0.5, 0.5, 1.0, 1.0, 0.0] | 0.7745966692414834 | "Guto Bebb" | "DUP" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.45, 0.5, 0.5, 0.0, 0.0, 0.0, 0.0, 1.0] | 0.19160041630983682 | "Guto Bebb" | "Scottish National Party" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.014285714285714292, 0.5, 0.014285714285714292, 0.5, 0.5, 0.9857142857142855, 0.9571428571428571, 0.028571428571428584] | 0.7562796166478085 | "Guto Bebb" | "Independent" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.075, 0.225, 0.15, 0.3, 0.25, 0.7999999999999998, 0.7250000000000001, 0.075] | 0.8338456535684794 | "Guto Bebb" | "Labour/Co-operative" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.0, 0.7187499999999999, 0.265625, 1.0, 1.0, 0.796875, 0.90625, 0.0] | 0.7381883979420002 | "Guto Bebb" | "Sinn Féin" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5] | 0.7071067811865476 | "Guto Bebb" | "Liberal Democrat" | [0, 0, 1, 1, 0, 1, 1, 0] | [0.0, 0.5, 0.4090909090909091, 0.5, 0.4090909090909091, 0.9545454545454545, 1.0, 0.0] | 0.8640682817262345 |=== We can see from the results here that Guto Bebb voted most similarly to the Liberal Democrats, and least similarly to the DUP and Conservatives. Let's run this query for all the rebel MPs: [source,cypher] ---- MATCH (person:Person)-[vote]->(m:Motion {date: date({year: 2019, month: 3, day: 27})}) WHERE (person)-[:MEMBER_OF {end: date({year: 2019, month: 9, day: 3})}] ->(:Party {name: "Conservative"}) MATCH (party:Party)-[ave:AVERAGE_VOTE]->(m) WHERE party.name <> "Speaker" WITH person, party, collect(CASE WHEN type(vote) = "FOR" THEN 1 WHEN type(vote) = "DID_NOT_VOTE" THEN 0.5 ELSE 0 END) AS personVotes, collect(ave.score) AS partyVotes WITH person, party, algo.similarity.cosine(personVotes, partyVotes) AS similarity ORDER BY similarity DESC WITH person, collect({party: party.name, score: similarity}) AS parties RETURN person.name AS person, parties[0].party AS party, parties[0].score AS score ---- If we run that query we'll see the results below: .Results [opts="header",cols="1,1,1"] |=== | person | party | score | "David Gauke" | "Sinn Féin" | 1.0 | "Caroline Nokes" | "Sinn Féin" | 1.0 | "Philip Hammond" | "Sinn Féin" | 1.0 | "Greg Clark" | "Sinn Féin" | 1.0 | "Justine Greening" | "Liberal Democrat" | 0.9585976908510777 | "Kenneth Clarke" | "Labour/Co-operative" | 0.9573767170592801 | "Dominic Grieve" | "Independent" | 0.954606992329263 | "Sam Gyimah" | "Liberal Democrat" | 0.931030158349453 | "Richard Harrington" | "Plaid Cymru" | 0.8970852271450604 | "Stephen Hammond" | "Plaid Cymru" | 0.8767140075192094 | "Antoinette Sandbach" | "Plaid Cymru" | 0.8767140075192094 | "Guto Bebb" | "Liberal Democrat" | 0.8640682817262345 | "Margot James" | "Plaid Cymru" | 0.8563488385776752 | "Steve Brine" | "Plaid Cymru" | 0.828078671210825 | "Alistair Burt" | "Plaid Cymru" | 0.828078671210825 | "Edward Vaizey" | "Plaid Cymru" | 0.828078671210825 | "Anne Milton" | "Liberal Democrat" | 0.8243496193719115 | "Richard Benyon" | "Labour" | 0.6174088452074383 | "Nicholas Soames" | "Labour" | 0.6174088452074383 | "Rory Stewart" | "Labour" | 0.6144467241017605 | "Oliver Letwin" | "Sinn Féin" | 0.6123724356957945 |=== The top 4 on the list were Cabinet members, which meant that they didn't vote on any of the motions, just like Sinn Féin representatives. Just below them we have Justine Greening, who used to be part of the cabinet until January 2018. She voted most similarly to the Liberal Democrats, and we can see how she voted on each issue by executing the following query: [source, cypher] ---- MATCH (person:Person {name: "Justine Greening"})-[vote]->(m:Motion {date: date({year: 2019, month: 3, day: 27})}) MATCH path2 = (party:Party {name: "Liberal Democrat"})-[ave:AVERAGE_VOTE]->(m) WITH person, CASE WHEN type(vote) = "FOR" THEN 1 WHEN type(vote) = "DID_NOT_VOTE" THEN 0.5 ELSE 0 END AS score, m, path2 CALL apoc.create.vRelationship(person, toString(score), {}, m) YIELD rel RETURN path2, rel, person, m ---- image::{{<siteurl>}}/uploads/2019/09/greening-indicative.svg[] She differs to the average position of her Conservative colleagues in a couple of ways: * She's not in favour of No Deal (Joanna Cherry's motion L) * She'd like there to be a confirmatory public vote (Margaret Beckett's motion M) I wonder if she'll be the next person to join the Liberal Democrats? One person who did recently do that is Phillip Lee. ++++ <iframe width="560" height="315" src="https://www.youtube.com/embed/27ngBYn6Y-E" frameborder="0" allow="accelerometer; autoplay; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> ++++ Let's see how he voted in the indicative votes: [source,cypher] ---- MATCH (person:Person {name: "Phillip Lee"})-[vote]->(m:Motion {date: date({year: 2019, month: 3, day: 27})}) MATCH (party:Party)-[ave:AVERAGE_VOTE]->(m) WHERE party.name <> "Speaker" RETURN party.name, algo.similarity.cosine( collect(CASE WHEN type(vote) = "FOR" THEN 1 WHEN type(vote) = "DID_NOT_VOTE" THEN 0.5 ELSE 0 END), collect(ave.score)) AS similarity ORDER BY similarity DESC ---- If we run that query we'll see the results below: .Results [opts="header",cols="1,1"] |=== | party | score | "Green" | 1.0 | "Independent" | 0.9105491868904616 | "Scottish National Party" | 0.8456834950587977 | "Liberal Democrat" | 0.8340478501880517 | "Plaid Cymru" | 0.7302967433402214 | "Labour/Co-operative" | 0.5989010989010989 | "Labour" | 0.5694375104718962 | "Sinn Féin" | 0.5 | "Conservative" | 0.1691931217592533 | "DUP" | 0.0 |=== He voted reasonably similarly to his Liberal Democrat colleagues, but voted identically to the Green party. Let's have a look at a graph of those votes: image::{{<siteurl>}}/uploads/2019/09/lee-green-indicative.svg[] One thing to keep in mind is that Caroline Lucas is the only person representing the Green Party, so he only voted identically to her rather than to a larger group of people. We can see that, like Justine Greening, he's only in favour of a confirmatory public vote and is not in favour of no deal. He voted against all the other motions. Let's see a graph of his votes compared to those of the Liberal Democrats: image::{{<siteurl>}}/uploads/2019/09/lee-liberal-indicative.svg[] Brexit wise he looks like a good fit for the party. He only really differs because he voted on every motion and many of his colleagues didn't vote on half of them. Of course to know if he's really a good fit for the party in general we'd need to compare his voting record across more issues than just the Brexit motions. We can tweak our query slightly to run it over all MPs and see which of them voted more similarly to another party than their own: [source, cypher] ---- MATCH (person:Person)-[vote]->(m:Motion {date: date({year: 2019, month: 3, day: 27})}) MATCH (person)-[memberOf:MEMBER_OF]->(actualParty) WHERE memberOf.start <= m.date AND (not(exists(memberOf.end)) OR m.date <= memberOf.end) MATCH (party:Party)-[ave:AVERAGE_VOTE]->(m) WHERE party.name <> "Speaker" WITH person, actualParty, party, collect(CASE WHEN type(vote) = "FOR" THEN 1 WHEN type(vote) = "DID_NOT_VOTE" THEN 0.5 ELSE 0 END) AS personVotes, collect(ave.score) AS partyVotes WITH person, actualParty, party, algo.similarity.cosine(personVotes, partyVotes) AS similarity ORDER BY similarity DESC, party = actualParty DESC WITH person, actualParty, collect({party: party, score: similarity}) AS parties WHERE actualParty <> parties[0].party WITH person, actualParty, parties[0].party.name AS mostSimilarParty, parties[0].score AS score ORDER BY person.pageviews DESC RETURN person.name AS person, actualParty.name AS actualParty, mostSimilarParty, score LIMIT 20 ---- If we run that query we'll see the results below: .Results [opts="header",cols="1,1,1,1"] |=== | person | actualParty | mostSimilarParty | score | "Theresa May" | "Conservative" | "Sinn Féin" | 1.0 | "Amber Rudd" | "Conservative" | "Sinn Féin" | 1.0 | "John Bercow" | "Speaker" | "Sinn Féin" | 1.0 | "Michael Gove" | "Conservative" | "Sinn Féin" | 1.0 | "Andrea Leadsom" | "Conservative" | "Sinn Féin" | 1.0 | "Sajid Javid" | "Conservative" | "Sinn Féin" | 1.0 | "Philip Hammond" | "Conservative" | "Sinn Féin" | 1.0 | "Anna Soubry" | "Independent" | "Green" | 1.0 | "Jim McMahon" | "Labour/Co-operative" | "Labour" | 0.9591990396603212 | "Jeremy Hunt" | "Conservative" | "Sinn Féin" | 1.0 | "Liam Fox" | "Conservative" | "Sinn Féin" | 1.0 | "Helen Hayes" | "Labour" | "Labour/Co-operative" | 0.9799919151000505 | "Kenneth Clarke" | "Conservative" | "Labour/Co-operative" | 0.9573767170592801 | "Justine Greening" | "Conservative" | "Liberal Democrat" | 0.9585976908510777 | "Chuka Umunna" | "Independent" | "Green" | 1.0 | "Dennis Skinner" | "Labour" | "Sinn Féin" | 0.6123724356957945 | "Vince Cable" | "Liberal Democrat" | "Independent" | 0.954606992329263 | "Angela Eagle" | "Labour" | "Labour/Co-operative" | 0.9835164835164835 | "Elizabeth Truss" | "Conservative" | "Sinn Féin" | 1.0 | "Harriet Harman" | "Labour" | "Labour/Co-operative" | 0.9834336020084081 |=== We can mostly ignore the first few names on here since they were all cabinet members who didn't vote on any of the motions. In a future iteration of the Brexit Graph we might want to store information about members of the government so that we could exclude them from this type of analysis. Labour/Co-operative and Labour tend to vote in similar ways to each other, so I don't think it's interesting to see a difference in the votes by Jim McMahon, Harriet Harman, or Angela Eagle. Justine Greening and Ken Clarke show up again - Justine voting in a similar way to Liberal Democrats and Ken Clarke in a similar way to Labour. Dennis Skinner is an interesting one. He's a long serving Labour MP, so it's surprising to see that he didn't vote in line with his party. We can write the following query to explore his votes: [source, cypher] ---- MATCH (person:Person {name: "Dennis Skinner"})-[rep:REPRESENTS]->(const) CALL apoc.create.vNode(["Constituency"], { caption: const.name + " (Leave" + const.leave + ")"}) YIELD node AS constituency CALL apoc.create.vRelationship(person, type(rep), {}, constituency) YIELD rel as representing MATCH (person)-[vote]->(m:Motion {date: date({year: 2019, month: 3, day: 27})}) MATCH path2 = (party:Party {name: "Labour"})-[ave:AVERAGE_VOTE]->(m) WITH person, representing, constituency, CASE WHEN type(vote) = "FOR" THEN 1 WHEN type(vote) = "DID_NOT_VOTE" THEN 0.5 ELSE 0 END AS score, m, path2 CALL apoc.create.vRelationship(person, toString(score), {}, m) YIELD rel RETURN path2, rel, person, m, representing, constituency ---- image::{{<siteurl>}}/uploads/2019/09/skinner-labour-indicative.svg[] He voted in favour of Jeremy Corbyn's alternative deal, but also in favour of the No Deal motion. We can see why he voted in favour of the latter by looking at the constituency he represents. Bolsover voted 70.4% in favour of leaving the EU, so he's in a tricky situation where he can't vote in favour of motions that would not respect the leave vote. This post has gone on a lot longer than I intended, but hopefully the exploration was interesting. If you have any ideas for further analysis that I should do, let me know in the comments.

In this post we look at how MPs voted in the Brexit indicative votes and find the party that they voted most in line with.

graphing-brexit-mps-vs-parties

https://markhneedham.com/blog/2019/09/23/graphing-brexit-mps-vs-parties

2019-08-23 00:47:00

Neo4j: Cypher - Nested Path Comprehensions vs OPTIONAL MATCH

While writing my previous post about https://markhneedham.com/blog/2019/08/22/neo4j-cypher-nested-pattern-comprehensions/[Cypher nested path comprehensions^], I realised that for this particular problem, the https://neo4j.com/docs/cypher-manual/current/clauses/optional-match/[OPTIONAL MATCH^] clause is a better choice. To recap, we have the following graph: [source,cypher] ---- MERGE (club:Club {name: "Man Utd"}) MERGE (league:League {name: "Premier League"}) MERGE (country:Country {name: "England"}) MERGE (club)-[:IN_LEAGUE]->(league) MERGE (league)-[:IN_COUNTRY]->(country) MERGE (club2:Club {name: "Juventus"}) MERGE (league2:League {name: "Serie A"}) MERGE (club2)-[:IN_LEAGUE]->(league2) ---- [.graph-model] image::{{<siteurl>}}/uploads/2019/08/nested-projection.svg[] We started the post with the following query that returns `(club)-[:IN_LEAGUE]->(league)-[:IN_COUNTRY]->(country)` paths: [source,cypher] ---- MATCH (club:Club) RETURN club.name, [path = (club)-[:IN_LEAGUE]->(league)-[:IN_COUNTRY]->(country) | {path: path}] AS path ---- .Results [opts="header",cols="1,2"] |=== | club.name | path | Man Utd | [{path: (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"})}] | Juventus | [] |=== The equivalent query using the `OPTIONAL MATCH` clause looks like this: [source,cypher] ---- MATCH (club:Club) OPTIONAL MATCH path = (club)-[:IN_LEAGUE]->(league)-[:IN_COUNTRY]->(country) RETURN club.name, path ---- If we run that query we get the following result: .Results [opts="header",cols="1,2"] |=== | club.name | path | Man Utd | (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"}) | Juventus | NULL |=== We concluded the post with the following query that returns as much of the `(club)-[:IN_LEAGUE]->(league)-[:IN_COUNTRY]->(country)` as exists: [source,cypher] ---- MATCH (club:Club) WITH club, [path1 = (club)-[:IN_LEAGUE]->(league) | {path1: path1, path2: [path2 = (league)-[:IN_COUNTRY]->(country) | path2]}] AS path RETURN club.name, [item in path | {path: apoc.path.combine(item.path1, item.path2[0])}] ---- .Results [opts="header",cols="1,2"] |=== | club.name | path | Man Utd | [{path: (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"})}] | Juventus | [{path: (:Club {name: "Juventus"})-[:IN_LEAGUE]->(:League {name: "Serie A"})}] |=== The equivalent query using the `OPTIONAL MATCH` clause looks like this: [source,cypher] ---- MATCH (club:Club) OPTIONAL MATCH path1 = (club)-[:IN_LEAGUE]->(league) OPTIONAL MATCH path2 = (league)-[:IN_COUNTRY]->(country) RETURN club.name, path1, path2 ---- .Results [opts="header",cols="1,2,2"] |=== | club.name | path1 | path2 | Man Utd | (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"}) | (:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"}) | Juventus | (:Club {name: "Juventus"})-[:IN_LEAGUE]->(:League {name: "Serie A"}) | NULL |=== As in the other post, we can then use https://neo4j.com/developer/neo4j-apoc/[APOC^]'s `apoc.path.combine` function on these paths: [source,cypher] ---- MATCH (club:Club) OPTIONAL MATCH path1 = (club)-[:IN_LEAGUE]->(league) OPTIONAL MATCH path2 = (league)-[:IN_COUNTRY]->(country) RETURN club.name, apoc.path.combine(path1, path2) ---- .Results [opts="header",cols="1,2"] |=== | club.name | path | Man Utd | (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"}) | Juventus | (:Club {name: "Juventus"})-[:IN_LEAGUE]->(:League {name: "Serie A"}) |=== That works well when we only have two paths to combine, but what if we have more than that? Let's run the following query to create an `IN_CONFEDERATION` relationship from England: [source,cypher] ---- MATCH (country:Country {name:"England"}) MERGE (confederation:Confederation {name: "UEFA"}) MERGE (country)-[:IN_CONFEDERATION]->(confederation) ---- [.graph-model] image::{{<siteurl>}}/uploads/2019/08/optional-match.svg[] Now we can write the following query to optionally find each of our three relationships, `IN_LEAGUE`, `IN_COUNTRY`, and `IN_CONFEDERATION`: [source,cypher] ---- MATCH (club:Club) OPTIONAL MATCH path1 = (club)-[:IN_LEAGUE]->(league) OPTIONAL MATCH path2 = (league)-[:IN_COUNTRY]->(country) OPTIONAL MATCH path3 = (country)-[:IN_CONFEDERATION]->(confederation) WITH club, [path1, path2, path3] AS paths RETURN club.name, reduce(acc=null, path in paths | apoc.path.combine(acc, path)) AS path ---- Note that we're using the `reduce` function to iterate over our paths and join them together. If we run this query we'll see the following result: .Results [opts="header",cols="1,2"] |=== | club.name | path | Man Utd | (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"})-[:IN_CONFEDERATION]->(:Confederation {name: "UEFA"}) | Juventus | (:Club {name: "Juventus"})-[:IN_LEAGUE]->(:League {name: "Serie A"}) |===

Learn how to write nested path comprehensions with the Cypher query language

neo4j-cypher-path-comprehensions-optional-match

https://markhneedham.com/blog/2019/08/23/neo4j-cypher-path-comprehensions-optional-match

2019-08-22 11:08:00

Neo4j: Cypher - Nested Path Comprehensions

I've recently been building an application using the https://grandstack.io/[GRANDstack^], which uses nested https://neo4j.com/blog/cypher-graphql-neo4j-3-1-preview/[Cypher path comprehensions^] to translate GraphQL queries to Cypher ones. I'd not done this before, so I was quite curious how this feature worked. We'll explore it using the following dataset: [source,cypher] ---- MERGE (club:Club {name: "Man Utd"}) MERGE (league:League {name: "Premier League"}) MERGE (country:Country {name: "England"}) MERGE (club)-[:IN_LEAGUE]->(league) MERGE (league)-[:IN_COUNTRY]->(country) MERGE (club2:Club {name: "Juventus"}) MERGE (league2:League {name: "Serie A"}) MERGE (club2)-[:IN_LEAGUE]->(league2) ---- [.graph-model] image::{{<siteurl>}}/uploads/2019/08/nested-projection.svg[] If we want to return a path containing a club, the league they play in, and the country that the league belongs to, we could write the following query: [source,cypher] ---- MATCH (club:Club) RETURN club.name, [path = (club)-[:IN_LEAGUE]->(league)-[:IN_COUNTRY]->(country) | {path: path}] AS path ---- If we execute that query, we'll get the following result: .Results [opts="header",cols="1,2"] |=== | club.name | path | Man Utd | [{path: (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"})}] | Juventus | [] |=== We only get a path back for Man Utd because there isn't a path from Juventus that has both the `IN_LEAGUE` and `IN_COUNTRY` relationships. What about if we do one path comprehension to get the `IN_LEAGUE` part of the path, and a second one to get the `IN_COUNTRY` part of the path? [source,cypher] ---- MATCH (club:Club) RETURN club.name, [path1 = (club)-[:IN_LEAGUE]->(league) | [path2 = (league)-[:IN_COUNTRY]->(country) | {path1: path1, path2: path2}]] AS path ---- If we execute that query, we'll get the following result: .Results [opts="header",cols="1,2"] |=== | club.name | path | Man Utd | [[{path1: (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"}), path2: (:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"})}]] | Juventus | [[]] |=== Hmmm, we expected not to get a value for `path2` for Juventus, but we should have got a value for `path1`. How do we do that? We need to pull the returning the map earlier in the path comprehension so that we return `path1` regardless of whether `path2` exists. The following query does this: [source,cypher] ---- MATCH (club:Club) RETURN club.name, [path1 = (club)-[:IN_LEAGUE]->(league) | {path1: path1, path2: [path2 = (league)-[:IN_COUNTRY]->(country) | path2]}] ---- If we execute that query, we'll get the following result: .Results [opts="header",cols="1,2"] |=== | club.name | path | Man Utd | [{path1: (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"}), path2: [(:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"})]}] | Juventus | [{path1: (:Club {name: "Juventus"})-[:IN_LEAGUE]->(:League {name: "Serie A"}), path2: []}] |=== That's more like it! Man Utd have paths for `path1` and `path2`, whereas Juventus only have a path for `path2`. Now, we'd like to join those paths together, which we can do using the `apoc.path.combine` function from the https://neo4j.com/developer/neo4j-apoc/[APOC library^]: [source,cypher] ---- MATCH (club:Club) WITH club, [path1 = (club)-[:IN_LEAGUE]->(league) | {path1: path1, path2: [path2 = (league)-[:IN_COUNTRY]->(country) | path2]}] AS path RETURN club.name, [item in path | {path: apoc.path.combine(item.path1, item.path2[0])}] ---- Since we know there can only be one `IN_COUNTRY` relationship between `league` and `country` we select the first value in the `path2` array on the last line of the query. If we execute that query, we'll get the following result: .Results [opts="header",cols="1,2"] |=== | club.name | path | Man Utd | [{path: (:Club {name: "Man Utd"})-[:IN_LEAGUE]->(:League {name: "Premier League"})-[:IN_COUNTRY]->(:Country {name: "England"})}] | Juventus | [{path: (:Club {name: "Juventus"})-[:IN_LEAGUE]->(:League {name: "Serie A"})}] |=== This example shows how to write a nested path comprehension at just one level of nesting. GraphQL queries can have an indefinite number of levels, so if you want to see extreme usage of this feature of the Cypher query language you'll need to create a GRANDstack app and then inspect the Cypher queries that get generated. ++++ <style> .graph-model img { margin-top: 10px; } </style> ++++

Learn how to write nested path comprehensions with the Cypher query language

neo4j-cypher-nested-pattern-comprehensions

https://markhneedham.com/blog/2019/08/22/neo4j-cypher-nested-pattern-comprehensions

2019-01-11 09:42:00

Python: Add query parameters to a URL

I was recently trying to automate adding a query parameter to a bunch of URLS and came across a neat approach a long way down this https://stackoverflow.com/questions/2506379/add-params-to-given-url-in-python[StackOverflow answer^], that uses the http://docs.python-requests.org/en/master/api/#requests.PreparedRequest[PreparedRequest^] class from the http://docs.python-requests.org/en/master/[requests library^]. Let's first get the class imported: [source,cypher] ---- from requests.models import PreparedRequest req = PreparedRequest() ---- And now let's use use this class to add a query parameter to a URL. We can do this with the following code: [source, cypher] ---- url = "http://www.neo4j.com" params = {'ref':"mark-blog"} req.prepare_url(url, params) ---- And then we need to access the `url` attribute to see our new URL: [source, cypher] ---- >>> req.url 'http://www.neo4j.com/?ref=mark-blog' ---- Neat! We can also use this approach to add parameters to URLs that already have existing ones. For example, we could update this YouTube URL: [source, cypher] ---- url = "https://www.youtube.com/watch?v=Y-Wqna-hC2Y&list=RDhlznpxNGFGQ&index=2" params = {'ref':"mark-blog"} req.prepare_url(url, params) ---- And let's see what the updated URL looks like: [source, cypher] ---- >>> req.url 'https://www.youtube.com/watch?v=Y-Wqna-hC2Y&list=RDhlznpxNGFGQ&index=2&ref=mark-blog' ---- I'm sure I'll be using this code snippet in future!

Learn how to add query parameters to a URL in Python.

python-add-query-parameters-url

https://markhneedham.com/blog/2019/01/11/python-add-query-parameters-url

2019-01-12 19:05:00

Neo4j: APOC - Caused by: java.io.RuntimeException: Can't read url or key file (No such file or directory)

I've been using Neo4j's https://neo4j-contrib.github.io/neo4j-apoc-procedures/[APOC^] library to load some local JSON files this week, and ran into an interesting problem. The `LOAD CSV` tool assumes that any files you load locally are in the `import` directory, so I've got into the habit of putting my data there. Let's check what I'm trying to import by opening the import directory: image::{{<siteurl>}}/uploads/2019/01/import-directory.png[] What's in there? image::{{<siteurl>}}/uploads/2019/01/import-directory-contents.png[] Just the one JSON file needs processing. If we want to import local files we need to add the following property to our Neo4j configuration file: [source, text] ---- apoc.import.file.enabled=true ---- If you're using the Neo4j Desktop, you can add this property via the 'Settings' tab: image::{{<siteurl>}}/uploads/2019/01/Selection_142.png[] Once we've done that we'll need to restart the database so our new settings will be picked up: image::{{<siteurl>}}/uploads/2019/01/restart-after-config.png[] Now let's try to process our JSON file: [source, cypher] ---- neo4j> CALL apoc.load.json("file:///dummy.json"); Failed to invoke procedure `apoc.load.json`: Caused by: java.lang.RuntimeException: Can't read url or key file:/dummy.json as json: /dummy.json (No such file or directory) ---- Hmm, that didn't work as we expected - it seems to be trying to read the file from the root of the machine rather than from the `import` directory. It turns out I hadn't RTFM: image::{{<siteurl>}}/uploads/2019/01/local-files.png[] Let's update our Neo4j configuration file to add the following property: [source, text] ---- apoc.import.file.use_neo4j_config=true ---- If we re-run our query we'll see that it now finds and processes the file: image::{{<siteurl>}}/uploads/2019/01/load-json.png[]

Learn how to import files from the import folder when using the APOC library.

neo4j-apoc-file-not-found-exception-no-such-file-directory

https://markhneedham.com/blog/2019/01/12/neo4j-apoc-file-not-found-exception-no-such-file-directory

2019-01-12 04:32:00

Neo4j: Cypher - Remove consecutive duplicates from a list

I was playing with a dataset this week and wanted to share how I removes duplicate consecutive elements from a list using the Cypher query language. For simplicity's sake, imagine that we have this list: [source, cypher] ---- neo4j> return [1,2,3,3,4,4,4,5,3] AS values; +-----------------------------+ | values | +-----------------------------+ | [1, 2, 3, 3, 4, 4, 4, 5, 3] | +-----------------------------+ ---- We want to remove the duplicate 3's and 4's, such that our end result should be: [source, cypher] ---- [1,2,3,4,5,3] ---- https://github.com/neo4j-contrib/neo4j-apoc-procedures[APOC]'s `apoc.coll.toSet` doesn't quite do the trick because it removes duplicates regardless of where they appear in the collection: [source, cypher] ---- neo4j> return apoc.coll.toSet([1,2,3,3,4,4,4,5,3]) AS values; +-----------------+ | values | +-----------------+ | [1, 2, 3, 4, 5] | +-----------------+ ---- Luckily it's quite easy to translate https://stackoverflow.com/a/46977206[Ulf Aslak's Python one liner] to do what we want. This is the Python version: [source,python] ---- values = [1,2,3,3,4,4,4,5,3] >>> [v for i, v in enumerate(values) if i == 0 or v != values[i-1]] [1, 2, 3, 4, 5, 3] ---- We'll use the https://neo4j.com/docs/cypher-manual/current/functions/list/#functions-range[`range`^] function to iterate over our list and list comprehensions to do the rest. The following code does the trick: [source, cypher] ---- neo4j> WITH [1,2,3,3,4,4,4,5,3] AS values RETURN [i in range(0, size(values)-1) WHERE i=0 OR values[i] <> values[i-1] | values[i] ] AS values; +--------------------+ | values | +--------------------+ | [1, 2, 3, 4, 5, 3] | +--------------------+ ---- You can use this on collections containing nodes, strings, or anything else - I've just used numbers in the example to keep the example simple.

Learn how to remove consecutive duplicate elements/duplicate neighbours from a list.

neo4j-cypher-remove-consecutive-duplicates

https://markhneedham.com/blog/2019/01/12/neo4j-cypher-remove-consecutive-duplicates

2019-01-13 06:42:00

Neo4j: Cypher - Date ranges

As part of a dataset I've been working with this week, I wanted to generate a collection of a range of dates using the Cypher query language. I've previously used the https://neo4j.com/docs/cypher-manual/current/syntax/temporal/#cypher-temporal-durations[`duration`^] function, which lets you add (or subtract) from a specific date, so I thought I'd start from there. If we want to find the day after 1st January 2019, we could write the following query: [source,cypher] ---- neo4j> WITH date("2019-01-01") AS startDate RETURN startDate + duration({days: 1}) AS date; +------------+ | date | +------------+ | 2019-01-02 | +------------+ ---- We can extend this code sample to find the next 5 dates from 1st January 2019 by using the `range` function: [source,cypher] ---- neo4j> WITH date("2019-01-01") AS startDate RETURN [day in range(0, 5) | startDate + duration({days: day})] AS dates; +--------------------------------------------------------------------------+ | dates | +--------------------------------------------------------------------------+ | [2019-01-01, 2019-01-02, 2019-01-03, 2019-01-04, 2019-01-05, 2019-01-06] | +--------------------------------------------------------------------------+ ---- That works fine if we know how many days we want to find the range for, but what if we only know the start and end dates for which we want to extract a range? We can use the `duration.inDays` function to compute a duration between two dates: [source, cypher] ---- neo4j> RETURN duration.inDays(date("2019-01-01"), date("2019-01-06")) AS difference; +------------+ | difference | +------------+ | P0M5DT0S | +------------+ ---- We can then get a count of the number of days from the `days`attribute: [source,cypher] ---- neo4j> RETURN duration.inDays(date("2019-01-01"), date("2019-01-06")).days AS days; +------+ | days | +------+ | 5 | +------+ ---- We can now update our query where we hard coded the number of days to use this computed value instead: [source,cypher] ---- neo4j> WITH date("2019-01-01") AS startDate, date("2019-01-06") AS endDate WITH startDate, duration.inDays(startDate, endDate).days AS days RETURN [day in range(0, days) | startDate + duration({days: day})] AS dates; +--------------------------------------------------------------------------+ | dates | +--------------------------------------------------------------------------+ | [2019-01-01, 2019-01-02, 2019-01-03, 2019-01-04, 2019-01-05, 2019-01-06] | +--------------------------------------------------------------------------+ ----

Learn how to remove consecutive duplicate elements/duplicate neighbours from a list.

neo4j-cypher-date-ranges

https://markhneedham.com/blog/2019/01/13/neo4j-cypher-date-ranges

2019-06-03 11:08:00

Kafka: Python Consumer - No messages with group id/consumer group

When I'm learning a new technology, I often come across things that are incredibly confusing when I first come across them, but make complete sense afterwards. In this post I'll explain my experience writing a Kafka consumer that wasn't finding any messages when using https://kafka.apache.org/documentation/#intro_consumers[consumer groups^] . == Setting up Kafka infrastructure We'll set up the Kafka infrastructure locally using the Docker Compose Template that I describe in my https://markhneedham.com/blog/2019/05/16/kafka-basic-tutorial/[Kafka: A Basic Tutorial blog post^]. We can run the following command to do this: [source,bash] ---- git clone git@github.com:mneedham/basic-kafka-tutorial.git && cd basic-kafka-tutorial ---- And then launch the Docker containers using the following command: [source, bash] ---- docker-compose up ---- While that's running let's install the https://github.com/dpkp/kafka-python[kafka-python^] library, which we'll use to put messages onto a Kafka topic, as well as consume messages from that topic. We can install this library using the following command: [source, bash] ---- pip install kafka-python ---- == Creating a Kafka topic Now let's create a topic named `foobar`, which we can do using the `kafka-topics` tool [source, bash] ---- docker exec broker-tutorial kafka-topics \ --zookeeper zookeeper:2181 \ --create \ --topic foobar \ --partitions 2 \ --replication-factor 1 ---- .Output [source, text] ---- Created topic "foobar". ---- Cool! Now we're ready to write some messages to the topic. == Producing and consuming messages The following code adds 10 JSON messages to the `foobar` topic: [source, python] ---- from kafka import KafkaProducer import uuid import json producer = KafkaProducer( bootstrap_servers=['localhost:9092'], value_serializer=lambda v: json.dumps(v).encode('utf-8')) for _ in range(10): producer.send('foobar', {"id": str(uuid.uuid4())}) producer.flush() ---- Let's read the messages from the topic. The following code does this: [source, python] ---- from kafka import KafkaConsumer import json consumer = KafkaConsumer('foobar', bootstrap_servers='localhost:9092', auto_offset_reset='earliest', consumer_timeout_ms=1000, value_deserializer = json.loads) for msg in consumer: print(msg.value) ---- Note that we set `auto_offset_reset` to `earliest` so that our consumer will read all the messages from the beginning. If we don't add this config our consumer will only see new messages. The output of running the consumer is below: [source, bash] ---- python dummy_consumer.py ---- .Output [source, json] ---- {'id': '939cd4fe-79e9-4050-a9bc-3f94b31f62e3'} {'id': 'a90c71b3-4516-4df9-871d-047264f1d6b6'} {'id': '06d45529-6888-4d2d-a4df-fec15b4b1d87'} {'id': '25642eee-b51c-432a-89d1-d7a17c7ef30a'} {'id': '03229045-94c9-4825-9cb6-ce495a68e7a9'} {'id': '3b6875f4-0a64-443f-9206-3bf1a3d31dc8'} {'id': '9f45a12b-97d2-4585-b296-a80ec5c0223c'} {'id': '485b4946-18a9-47d1-a849-87a6fc60365a'} {'id': 'b1c9d75a-d56e-4dd8-9e08-89b3a818fbb1'} {'id': '14486560-b5cb-41fa-bd0e-845a424c8ed4'} ---- If we run that code again, we'll see the same list of 10 messages. What about if we provide a https://kafka.apache.org/documentation/#intro_consumers[consumer group^]? (using the `group_id` config) The following consumer reads from the `foobar` topic using a group id named `blog_group`: [source, python] ---- from kafka import KafkaConsumer import json consumer = KafkaConsumer('foobar', bootstrap_servers='localhost:9092', group_id='blog_group', auto_offset_reset='earliest', consumer_timeout_ms=10000, value_deserializer = json.loads) for msg in consumer: print(msg.value) ---- The first time we run this script we'll see those 10 messages, but if we run it again we won't get any messages. The reason for this is that when we provide a group id, the broker https://github.com/confluentinc/confluent-kafka-python/issues/275[keeps track of the current offset^] so that messages aren't consumed twice. We can run the following command to see this: [source, bash] ---- $ docker exec broker-tutorial kafka-consumer-groups \ --bootstrap-server broker:9093 \ --group blog_group \ --describe ---- .Output [source,text] ---- TOPIC PARTITION CURRENT-OFFSET LOG-END-OFFSET LAG CONSUMER-ID HOST CLIENT-ID foobar 0 8 8 0 kafka-python-1.4.6-74f3f185-b563-40c8-958a-ad6c0c4815c0 /172.26.0.1 kafka-python-1.4.6 foobar 1 2 2 0 kafka-python-1.4.6-74f3f185-b563-40c8-958a-ad6c0c4815c0 /172.26.0.1 kafka-python-1.4.6 ---- From this output we need to look at two columns: * `CURRENT_OFFSET`, which indicates the offset that our consumer has read up to * `LOG-END-OFFSET`, which indicates the maximum offset for that partition If we want to consume all the messages on the `foobar` topic again, we'll need to reset `CURRENT_OFFSET` back to 0. == Resetting offsets We can do this by passing the `--reset-offsets` argument to `kafka-consumer-groups.` The following code shows what a dry run of this command will do: [source, bash] ---- docker exec broker-tutorial kafka-consumer-groups \ --bootstrap-server broker:9093 \ --group blog_group \ --topic foobar \ --reset-offsets \ --to-earliest \ --dry-run ---- .Output [source, text] ---- TOPIC PARTITION NEW-OFFSET foobar 0 0 foobar 1 0 ---- And if we want to execute it for real, we need to change `--dry-run` to `--execute`: [source, bash] ---- docker exec broker-tutorial kafka-consumer-groups \ --bootstrap-server broker:9093 \ --group blog_group \ --topic foobar \ --reset-offsets \ --to-earliest --execute ---- .Output [source, text] ---- TOPIC PARTITION NEW-OFFSET foobar 0 0 foobar 1 0 ---- Once we've done this we can re-run our group id consumer and we'll be able to read all the messages again.

Learn how to work around a confusing error message when using the Kafka Python Consumer.

kafka-python-consumer-no-messages-group-id-consumer-group

https://markhneedham.com/blog/2019/06/03/kafka-python-consumer-no-messages-group-id-consumer-group

2019-12-19 00:21:00

React React Router: Setting parent component state based on route change event

I've been working on a https://reach.tech/router[React Reach Router^] based application that has several routes and wanted to show a search box in the header unless the user was on the search page. After a lot of trial and error I learnt that I could use a route change event listener to do this. The CodeSandbox below shows all the code to do this: ++++ <iframe src="https://codesandbox.io/embed/github/mneedham/hugo-blog/tree/master/blog/code/my-app?fontsize=14&hidenavigation=1&module=%2Fsrc%2FApp.js&theme=dark" style="width:100%; height:500px; border:0; border-radius: 4px; overflow:hidden;" title="my-app" allow="geolocation; microphone; camera; midi; vr; accelerometer; gyroscope; payment; ambient-light-sensor; encrypted-media; usb" sandbox="allow-modals allow-forms allow-popups allow-scripts allow-same-origin" ></iframe> ++++ Let's walk through the code. We have a top level component called `App` that has the state `showSearchBox`, which defaults to `true`: [source,javascript] ---- import React, {Component} from 'react'; import {globalHistory, Link, Router} from "@reach/router"; class App extends Component { constructor(props) { super(props) this.state = { showSearchBox: true }; } } ---- In our render function we display an input box only if this value is set to `true`: [source,javascript] ---- class App extends Component { render() { return ( <div className="App"> <header className="App-header"> <p> React Reach Router Demo </p> <div> {this.state.showSearchBox && <input type={"text"} style={{margin: "10px"}}/>} </div> </header> <nav style={{padding: "5px"}}> <Link to="/">Home</Link> <Link to="dashboard">Dashboard</Link> <Link to="search">Search</Link> </nav> <div style={{paddingTop: "10px"}}> <Router> <Home path="/"/> <Dashboard path="/dashboard"/> <Search path="/search"/> </Router> </div> </div> ); } } ---- In this code sample we can also see that we have 3 paths: * `/` * `/dashboard` and * `/search` We want to set `showSearchBox` to `false` if we're on the `/search` route, but set to `true` on any of the other routes. I initially tried to control this state from the child components, without much success. The closest I got was a maximum update depth exceeded error, which wasn't great. By chance I came across a https://github.com/reach/router/issues/262[GitHub issue^] created by Marvin Heilemann, in which he asked whether there was a hook to capture route change events. And indeed there is. We can update our `App` component to listen to these events and update the `showSearchBox` state by adding the following functions: [source, javascript] ---- class App extends Component { componentDidMount() { this.toggleSearchBox(globalHistory.location) this.historyUnsubscribe = globalHistory.listen(({action, location}) => { if (action === 'PUSH') { this.toggleSearchBox(location) } }); } componentWillUnmount() { this.historyUnsubscribe(); } toggleSearchBox(location) { if (location.pathname === "/search") { this.setState({ showSearchBox: false }) } else { this.setState({ showSearchBox: true }) } } } ---- In `componentDidMount` we also make sure that we toggle the search box based on the URL that we start on as well, otherwise it would default to `true`. If we try the CodeSandbox provided at the top of the post we'll see the following if we click the `Dashboard` link in the header: image::{{<siteurl>}}/uploads/2019/12/dashboard.png[] As expected, the search box is still showing. But if we click the `Search` link, we'll see the following screen: image::{{<siteurl>}}/uploads/2019/12/search.png[] After I'd got this working I came across https://github.com/reach/router/issues/203[another GitHub issue^], where Martin Mende showed how to achieve the same thing https://reactjs.org/docs/hooks-intro.html[using state and effect hooks^]. The following code does the same thing as the `App` component that we defined above: [source, javascript] ---- import React, {useEffect, useState} from 'react'; import {globalHistory, Link, Router} from "@reach/router"; function App() { const initialState = true; const [showSearchBox, setShowSearchBox] = useState(initialState); useEffect(() => { const removeListener = globalHistory.listen(params => { const { location } = params; const newState = location.pathname !== "/search"; setShowSearchBox(newState); }); return () => { removeListener(); }; }, []); return ( <div className="App"> <header className="App-header"> <p>React Reach Router Demo</p> <div> {showSearchBox && <input type={"text"} style={{ margin: "10px" }} />} </div> </header> <nav style={{ padding: "5px" }}> <Link to="/">Home</Link> <Link to="dashboard">Dashboard</Link>{" "} <Link to="search">Search</Link> </nav> <div style={{ paddingTop: "10px" }}> <Router> <Home path="/" /> <Dashboard path="/dashboard" /> <Search path="/search" /> </Router> </div> </div> ); } ----

Learn how to listen to route change events when using React Reach Router.

react-reach-parent-event-route-change

https://markhneedham.com/blog/2019/12/19/react-reach-parent-event-route-change

2019-12-23 00:21:00

QuickGraph #3: Itsu Allergens

As someone who's allergic to lots of different things, the introduction of allergen charts in restaurants over the last few years has been very helpful. These charts are often hidden away in PDF files, but the Asian inspired Itsu restaurant have all this information available on their online menus. This therefore seemed like a great opportunity for another https://markhneedham.com/blog/tag/quickgraph/[QuickGraph^]. image::{{<siteurl>}}/uploads/2019/12/Itsu_logo.svg[] == Scraping the Itsu website I wrote a couple of Python scripts to download each of the menu items and then extract the product name, description, and allergens. Instructions for https://github.com/mneedham/itsu-neo4j#downloading-and-scraping-product-data[how to download and scrape this data^] are described in the https://github.com/mneedham/itsu-neo4j[itsu-neo4j^] repository. The scraped data for each of the products is in the https://github.com/mneedham/itsu-neo4j/blob/master/import/items.json[import/items.json^] file. == Exploring the data Let's have a quick look at the first few lines of this file using command line tools: [source,bash] ---- $ head -n3 import/items.json | jq '.' { "title": "aluminium refill flask", "allergens": [], "description": "for only £1.99, our aluminium refill flask is eco-friendly and can be used again and again. Proudly supporting Blue Marine Foundation.", "url": "https://www.itsu.com/menu/sides-snacks/aluminium-refill-flask/" } { "title": "itsu still water", "allergens": [], "description": "", "url": "https://www.itsu.com/menu/sides-snacks/itsu-still-water/" } { "title": "salmon sashimi", "allergens": [ "fish", "soya", "mustard", "gluten" ], "description": "omega 3 rich sashimi grade salmon sashimi on magnesium packed wakame & pickled ginger", "url": "https://www.itsu.com/menu/sushi/salmon-sashimi-2/" } ---- We can also use the https://neo4j.com/docs/labs/apoc/3.5/[APOC^] Library's https://neo4j.com/docs/labs/apoc/current/import/load-json/[`apoc.load.json`^] procedure to explore the data. The following query returns the 1st product: [source,cypher] ---- CALL apoc.load.json("https://github.com/mneedham/itsu-neo4j/raw/master/import/items.json") YIELD value RETURN value LIMIT 1 ---- If we run that query we'll see the following output: .Exploring the data [opts="header"] |=== | value | {description: "for only £1.99, our aluminium refill flask is eco-friendly and can be used again and again. Proudly supporting Blue Marine Foundation.", title: "aluminium refill flask", url: "https://www.itsu.com/menu/sides-snacks/aluminium-refill-flask/", allergens: []} |=== This isn't a particularly interesting product as it doesn't have any allergens! Next we're going to learn how we can use `LOAD CSV` to import the products into Neo4j. == Importing the data We're going to import the data into the following graph model: image::{{<siteurl>}}/uploads/2019/12/itsu-graph-model.png[] The following statement imports the products and their allergens: [source,cypher] ---- CALL apoc.load.json("https://github.com/mneedham/itsu-neo4j/raw/master/import/items.json") YIELD value MERGE (product:Product {url: value.url}) SET product.name = value.title, product.description = value.description WITH product, value UNWIND value.allergens AS a MERGE (allergen:Allergen {name: a}) MERGE (product)-[:CONTAINS_ALLERGEN]->(allergen); ---- We can see a sample of the imported graph in the Neo4j Browser visualisation below: image::{{<siteurl>}}/uploads/2019/12/itsu-sample.svg[] == Querying the graph Now let's explore the data using Neo4j's https://neo4j.com/developer/cypher-basics-i/[Cypher^] query language. == How many allergens are there? [source,cypher] ---- MATCH (allergen:Allergen) RETURN count(*), collect(allergen.name) AS allergens; ---- .How many allergens are there? [opts="header",cols="1,3"] |=== | count(*) | allergens | 14 | ["fish", "soya", "mustard", "gluten", "celery", "sesame", "sulphur dioxide", "dairy", "egg", "crustaceans", "wheat", "mullusc", "nuts", "peanuts"] |=== == Which allergen is most prevalent? [source,cypher] ---- MATCH (:Product) WITH count(*) AS productCount MATCH (allergen)<-[:CONTAINS_ALLERGEN]-() WITH allergen, count(*) AS count, productCount RETURN allergen.name AS allergen, count, apoc.math.round(count*1.0/productCount,2) AS percentageOfProducts ORDER BY count DESC LIMIT 5; ---- .Which allergen is most prevalent? [opts="header"] |=== | allergen | count | percentageOfProducts | "soya" | 78 | 0.75 | "gluten" | 55 | 0.53 | "sesame" | 54 | 0.52 | "mustard" | 42 | 0.4 | "celery" | 24 | 0.23 |=== Itsu is clearly not a good place to go if you're allergic to soya, gluten, or sesame. == Which allergens appear together most frequently? [source,cypher] ---- MATCH (allergen1:Allergen)<-[:CONTAINS_ALLERGEN]-()-[:CONTAINS_ALLERGEN]->(allergen2) WHERE allergen1.name < allergen2.name RETURN allergen1.name AS allergen1, allergen2.name AS allergen2 , count(*) AS count ORDER BY count DESC LIMIT 5 ---- .Which allergens appear together most frequently? [opts="header"] |=== | allergen1 | allergen2 | count | "gluten" | "soya" | 55 | "sesame" | "soya" | 53 | "gluten" | "sesame" | 47 | "mustard" | "soya" | 40 | "mustard" | "sesame" | 36 |=== It looks like Itsu recipes often use soya alongside gluten and sesame. And now let's finally see what I can eat the next time that I go to Itsu. == How many things can I eat in Itsu? [source,cypher] ---- MATCH (:Product) WITH count(*) AS productCount WITH ["crustaceans", "nuts", "peanuts", "egg", "dairy", "fish"] AS allergens, productCount MATCH (product:Product) WHERE all(allergen in allergens WHERE not((product)-[:CONTAINS_ALLERGEN]->(:Allergen {name: allergen}))) WITH count(*) AS count, collect(product.name) AS products, productCount RETURN count, apoc.math.round(count*1.0/productCount, 2) AS percentageOfProducts, products ---- .How many things can I eat in Itsu? [opts="header", cols="1,1,3"] |=== | count | percentageOfProducts | products | 61 | 0.59 | ["aluminium refill flask", "itsu still water", "edamame", "dark chocolate rice cakes", "veggie rice’bowl & quinoa burgers", "winter wonderland", "veggie dragon roll", "christmas cracker gyoza", "detox miso’noodle soup", "little choc pot", "porridge’power", "itsu sparkling water", "hoisin duck tokyo wrap", "vegetable fusion gyoza", "chicken noodle soup", "ginger & lemon kombucha", "ginger’low", "veg press", "lean chicken machine", "miso soup", "original kombucha", "coconut chicken soup", "thai chicken rice’bowl", "blueberry’boost porridge", "beef twerky", "crispy seaweed thins sweet soy & sea salt", "passionfruit kombucha", "raw fruitfix beauty’smoothie", "cucumber & mint zen’water", "Hawaii 5.0 fruit cup", "crispy seaweed thins sea salt", "bacon bao’bun", "ginger detox zinger", "pork & truffle gyoza", "crispy seaweed thins wasabi", "veggie club rolls", "veggie’gyoza udon ", "spicy korean chicken rice’bowl", "veggie thai soup", "the sesame chicken salad", "i’thai udon [stir-fry style]", "lean satay chicken tokyo wrap", "quinoa burgers tokyo wrap", "little salted caramel pot", "orange press", "hoisin duck bao buns", "peach & lychee zen’water", "veggie sushi collection", "raw veg cleanse beauty’smoothie", "chilli’chicken udon", "super’seeds porridge", "no meat mondays", "avo baby rolls", "lemon’low", "tenderstem broccoli with sesame dressing", "crushed coconut & chocolate oishi bar", "teriyaki chicken rice’bowl", "vegetable dumplings", "wasabi peas [healthy snack]", "chargrilled chicken udon", "goji, mandarin & lime"] |=== A lot more than I expected! Let's simplify this query a bit by creating a function that returns the product count: [source,cypher] ---- CALL apoc.custom.asFunction("productCount", "MATCH (:Product) RETURN count(*) AS count", "LONG", null, true) ---- And now we'll update our previous query to use this function: [source,cypher] ---- WITH ["crustaceans", "nuts", "peanuts", "egg", "dairy", "fish"] AS allergens MATCH (product:Product) WHERE all(allergen in allergens WHERE not((product)-[:CONTAINS_ALLERGEN]->(:Allergen {name: allergen}))) WITH count(*) AS count, collect(product.name) AS products RETURN count, apoc.math.round(count*1.0/custom.productCount(), 2) AS percentage, products ---- == How many hot things can I eat in Itsu? Let's say I want to eat something from the `hot` category. We haven't modelled that in our graph, but it is embedded in the `url` property stored on each product. The following query will find the products that I can eat in this category: [source,cypher] ---- WITH ["crustaceans", "nuts", "peanuts", "egg", "dairy", "fish"] AS allergens MATCH (product:Product) WHERE all(allergen in allergens WHERE not((product)-[:CONTAINS_ALLERGEN]->(:Allergen {name: allergen}))) WITH product WHERE split(product.url, "/")[-3] = "hot" RETURN product.name AS product, product.url AS url, [(product)-[:CONTAINS_ALLERGEN]->(allergen) | allergen.name] AS allergens ---- .Which hot things can I eat in Itsu? [opts="header", cols="1,2,3"] |=== | product | url | allergens | "veggie rice’bowl & quinoa burgers" | "https://www.itsu.com/menu/hot/quinoa-falafel-veg-ricebowl/" | ["sulphur dioxide", "sesame", "mustard", "celery", "soya", "gluten"] | "winter wonderland" | "https://www.itsu.com/menu/hot/winter-wonderland/" | ["sesame", "mustard", "celery", "soya", "gluten"] | "christmas cracker gyoza" | "https://www.itsu.com/menu/hot/christmas-cracker-gyoza/" | ["sesame", "soya", "gluten"] | "detox miso’noodle soup" | "https://www.itsu.com/menu/hot/detox-noodles/" | ["soya", "sesame", "gluten"] | "vegetable fusion gyoza" | "https://www.itsu.com/menu/hot/vegetable-fusion-gyoza/" | ["soya", "sesame", "gluten"] | "chicken noodle soup" | "https://www.itsu.com/menu/hot/the-chicken-noodle-soup/" | ["mustard", "soya", "celery", "sulphur dioxide", "gluten", "sesame"] | "miso soup" | "https://www.itsu.com/menu/hot/miso-soup/" | ["soya"] | "coconut chicken soup" | "https://www.itsu.com/menu/hot/coconutchicken-greens/" | ["sulphur dioxide", "sesame", "soya", "mustard", "celery"] | "thai chicken rice’bowl" | "https://www.itsu.com/menu/hot/chicken-thai-ricebowl/" | ["celery", "sulphur dioxide", "sesame", "mustard", "soya", "gluten"] | "veggie’gyoza udon " | "https://www.itsu.com/menu/hot/veggie-gyoza-noodles/" | ["celery", "sesame", "sulphur dioxide", "mustard", "soya", "gluten"] | "spicy korean chicken rice’bowl" | "https://www.itsu.com/menu/hot/korean-bbq-chicken-ricebowl/" | ["gluten", "sesame", "soya", "mustard", "celery", "sulphur dioxide"] | "veggie thai soup" | "https://www.itsu.com/menu/hot/thai-coconut-veggierice/" | ["celery", "soya", "mustard", "sulphur dioxide", "sesame"] | "i’thai udon [stir-fry style]" | "https://www.itsu.com/menu/hot/ithai-udon-noodles-yaki-udon/" | ["gluten", "sesame", "mustard", "celery", "soya", "sulphur dioxide"] | "hoisin duck bao buns" | "https://www.itsu.com/menu/hot/hoisin-duck-bao-buns/" | ["sesame", "soya", "gluten", "mullusc"] | "chilli’chicken udon" | "https://www.itsu.com/menu/hot/chilli-chicken-udon-2/" | ["sulphur dioxide", "celery", "soya", "sesame", "gluten", "mustard"] | "teriyaki chicken rice’bowl" | "https://www.itsu.com/menu/hot/chicken-teriyaki-ricebowl/" | ["mustard", "gluten", "celery", "sesame", "soya", "sulphur dioxide"] | "chargrilled chicken udon" | "https://www.itsu.com/menu/hot/chargrilled-chicken-noodles/" | ["sulphur dioxide", "celery", "sesame", "mustard", "soya", "gluten"] |=== It occurs to me after writing this post that this dataset would be much easier to explore via a web app, so perhaps a GRANDstack allergen application is the next thing in my future.

Learn how to build a graph of Itsu allergens.

uploads/2019/12/GUIDE_LOGO_itsu.png

quick-graph-itsu-allergens

https://markhneedham.com/blog/2019/12/23/quick-graph-itsu-allergens

2019-12-22 00:21:00

QuickGraph #2: Guardian Top 100 Male Footballers

Over the last week the Guardian have been counting down their https://www.theguardian.com/global/ng-interactive/2019/dec/17/the-100-best-male-footballers-in-the-world-2019[top 100 male footballers of 2019^], and on Friday they also published a https://docs.google.com/spreadsheets/d/1f4nc8vehOiZhEB2_3L2bvXTpag29VcEKJ0PJeo5Runc/edit#gid=3[Google sheet containing all the votes^], which seemed like a perfect candidate for a https://markhneedham.com/blog/tag/quickgraph/[QuickGraph^]. image::{{<siteurl>}}/uploads/2019/12/top-100-header.png[] We can see a preview of the Google sheet in the printscreen below: image::{{<siteurl>}}/uploads/2019/12/spreadsheet.png[] We can also download Google Sheets in CSV format based on the following URI template: [source, text] ---- https://docs.google.com/spreadsheets/d/KEY/export?format=csv&id=KEY&gid=SHEET_ID ---- where: * `KEY` is the spreadsheet id, in our case `1f4nc8vehOiZhEB2_3L2bvXTpag29VcEKJ0PJeo5Runc` * `SHEET_ID` is the sheet, in our case `3` == Exploring the data We can use Neo4j's `LOAD CSV` tool to explore the data. The following query returns the 1st player on the list: [source,cypher] ---- WITH "1f4nc8vehOiZhEB2_3L2bvXTpag29VcEKJ0PJeo5Runc" AS id WITH "https://docs.google.com/spreadsheets/d/" + id + "/export?format=csv&id=" + id + "&gid=3" AS uri LOAD CSV FROM uri AS row WITH row SKIP 4 LIMIT 1 RETURN row ---- If we run that query, we'll see the following output: .Results [opts="header",cols="1"] |=== | row | ["1", "2", "3", "1", "2", "1", "2", "1", "Lionel Messi", "Argentina", "Barcelona", "Spain", "32", "31", "9198", "40", "9158", "239", "108", "40", "36", "38", "38", "32", "40", "13", "39", "40", "40", "39", "30", "40", "40", "39", "37", "40", "38", "39", "39", "40", "39", "38", "38", "40", "38", "40", "40", "40", "39", "39", "38", "38", "39", "40", "40", "40", "40", "38", "40", "40", "40", "39", "40", "40", "39", "40", "39", "37", "39", "40", "40", "39", "40", "39", "37", "36", "40", "39", "38", "40", "39", "39", "40", "40", "34", "40", "40", "40", "40", "39", "38", "40", "40", "39", "40", "38", "40", "40", "39", "39", "35", "40", "40", "39", "39", "39", "38", "40", "39", "39", "36", "39", "40", "24", "40", "39", "30", "39", "39", "40", "40", "39", "39", "39", "37", "40", "35", "40", "35", "39", "39", "39", "37", "40", "39", "40", "38", "40", "40", "39", "40", "39", "39", "40", "37", "40", "36", "39", "40", "39", "38", "39", "40", "40", "40", "38", "22", "38", "39", "40", "39", "37", "40", "40", "40", "40", "40", "38", "40", "38", "40", "39", "40", "40", "39", "40", "40", "22", "39", "40", "40", "40", "40", "38", "39", "39", "38", "39", "39", "19", "40", "32", "40", "40", "40", "30", "39", "40", "40", "40", "38", "40", "40", "40", "39", "40", "40", "37", "40", "36", "37", "38", "39", "40", "39", "40", "40", "40", "38", "39", "40", "40", "40", "38", "39", "36", "39", "30", "38", "40", "40", "39", "39", "40", "39", "40", "40", "37", "40", "40", "39", "40", "39", "38", "40", "38", "39", "39", "38", "38", "40", "40", "39", "40", "40", "39", "38", "40"] |=== We're not interested in the first 8 columns, but we do want to capture the next 4 columns, which contain the player's name, nationality, club, and the country where that club competes. We'll then skip the next 7 columns until we get to the votes given by each of the judges. == Importing the data We're going to import the data into the following graph model: image::{{<siteurl>}}/uploads/2019/12/guardian-graph-model.svg[] Before we run any import statements, we'll make sure we don't accidentally import any duplicates by creating unique constraints on each of the node labels: [source,cypher] ---- CALL apoc.schema.assert(null,{Judge:['id'], Player:["name"], Country:["name"], Club:["name"]}); ---- We can run the following query to check that they've been created: .Results [opts="header"] |=== | description | indexName | tokenNames | properties | state | type | progress | provider | id | failureMessage | "INDEX ON :Club(name)" | "index_7" | ["Club"] | ["name"] | "ONLINE" | "node_unique_property" | 100.0 | {version: "1.0", key: "native-btree"} | 7 | "" | "INDEX ON :Country(name)" | "index_1" | ["Country"] | ["name"] | "ONLINE" | "node_unique_property" | 100.0 | {version: "1.0", key: "native-btree"} | 1 | "" | "INDEX ON :Judge(id)" | "index_10" | ["Judge"] | ["id"] | "ONLINE" | "node_unique_property" | 100.0 | {version: "1.0", key: "native-btree"} | 10 | "" | "INDEX ON :Player(name)" | "index_4" | ["Player"] | ["name"] | "ONLINE" | "node_unique_property" | 100.0 | {version: "1.0", key: "native-btree"} | 4 | "" |=== All good so far. Now let's import the votes. We're going to use the same `LOAD CSV` tool that we used to explore the data. Importing the data is a bit fiddly because rows 105 and 186 don't contain players, so we need to skip those rows. The following `LOAD CSV` statement imports the top 100 players, along with their club, country, and votes: [source,cypher] ---- WITH "1f4nc8vehOiZhEB2_3L2bvXTpag29VcEKJ0PJeo5Runc" AS id WITH "https://docs.google.com/spreadsheets/d/" + id + "/export?format=csv&id=" + id + "&gid=3" AS uri LOAD CSV FROM uri AS row WITH row SKIP 4 LIMIT 100 MERGE (player:Player {name: row[8]}) SET player.rawScore = toInteger(row[14]), player.score = toInteger(row[16]) MERGE (country:Country {name: row[9]}) MERGE (club:Club {name: row[10]}) MERGE (clubCountry:Country {name: row[11]}) MERGE (player)-[:NATIONALITY]->(country) MERGE (player)-[:PLAYS_FOR]->(club) MERGE (club)-[:PLAY_IN]->(clubCountry) FOREACH(index in range(19, size(row)-1) | MERGE (judge:Judge {id: index-18}) FOREACH(ignoreMe IN CASE WHEN row[index] is null THEN [] ELSE [1] END | MERGE (judge)-[voted:VOTED]->(player) SET voted.score = toInteger(row[index]) )); ---- To import the rest of the players we'll have to vary the `SKIP` and `LIMIT` values on the 4th line of the query The GitHub gist below includes statements that import the votes for all players: ++++ <script src="https://gist.github.com/mneedham/1ad3733452f90730fd3de7595121e637.js"></script> ++++ We can see a sample of the imported graph in the Neo4j Browser visualisation below: image::{{<siteurl>}}/uploads/2019/12/sample-graph.svg[] == Querying the graph Now that we've imported the data, it's time to start querying it. The Google sheet already contains the answers to the following questions: * How many judges included each player in their top 40? * How many judges voted for a player as their number 1? * What's the top ranking that a player received? Let's see what else we can learn. The queries that follow use Neo4j's https://neo4j.com/developer/cypher-basics-i/[Cypher^] query language. == How many judges included the top 5 in their top 5? [source,cypher] ---- MATCH (player:Player) WHERE exists(player.score) WITH player ORDER BY player.score DESC LIMIT 5 MATCH (player)<-[voted:VOTED]-() WHERE voted.score >= 36 WITH player, 41-voted.score AS rank, count(*) AS count RETURN player.name AS player, apoc.map.fromPairs(collect([rank, count])) AS votes, player.score AS score; ---- .How many judges included the top 5 in their top 5? [opts="header", cols="1,3,1"] |=== | player | votes | score | "Lionel Messi" | {`1`: 108, `2`: 69, `3`: 31, `4`: 10, `5`: 6} | 9158 | "Virgil van Dijk" | {`1`: 77, `2`: 59, `3`: 30, `4`: 30, `5`: 8} | 8849 | "Sadio Mané" | {`1`: 22, `2`: 34, `3`: 46, `4`: 36, `5`: 28} | 8383 | "Cristiano Ronaldo" | {`1`: 11, `2`: 29, `3`: 53, `4`: 31, `5`: 28} | 8172 | "Mohamed Salah" | {`1`: 5, `2`: 8, `3`: 16, `4`: 24, `5`: 29} | 7421 |=== We can see that the judges overwhelmingly have Messi and van Dijk in their top 2. This is not that surprising given that they finished 1st and 2nd in the https://en.wikipedia.org/wiki/The_Best_FIFA_Football_Awards_2019[The Best FIFA Football Awards 2019^] and https://en.wikipedia.org/wiki/2019_Ballon_d%27Or[2019 Ballon d'Or^]. Did any judges get the top 3 exactly right? == Which judges got the top 3 exactly correct? [source,cypher] ---- MATCH (player1:Player {name: "Lionel Messi"})<-[:VOTED {score: 40}]-(judge) MATCH (player2:Player {name: "Virgil van Dijk"})<-[:VOTED {score: 39}]-(judge) MATCH (player3:Player {name: "Sadio Mané"})<-[:VOTED {score: 38}]-(judge) WITH collect(judge.id) AS judges RETURN size(judges) AS count, judges ---- .Which judges got the top 3 exactly correct? [opts="header", cols="1,3"] |=== | count | judges | 14 | [179, 145, 102, 239, 187, 204, 220, 215, 144, 51, 202, 42, 79, 218] |=== Only 14 out of 239 (or 6% of the) judges got the top 3 exactly right. Football is clearly a game of different opinions! Next let's see which clubs have fared best. == Which club has the most players on the list? [source,cypher] ---- MATCH (club:Club)<-[:PLAYS_FOR]-(:Player) RETURN club.name AS club, count(*) AS count ORDER BY count DESC LIMIT 10 ---- .Which club has the most players on the list? [opts="header"] |=== | club | count | "Real Madrid" | 19 | "Manchester City" | 17 | "Liverpool" | 15 | "Juventus" | 14 | "Bayern Munich" | 14 | "Barcelona" | 14 | "Paris St-Germain" | 12 | "Chelsea" | 11 | "Atlético Madrid" | 10 | "Tottenham Hotspur" | 9 |=== Manchester City and Liverpool had very strong years, so it's not surprising to see that they have a lot of players on the list. Real Madrid having the most players is a strange one given their relatively poor year. Perhaps the players that they have on the list didn't receive many votes. Let's tweak the query to sum the voting scores given to the players that play for those clubs == Which clubs have the highest ranking players on the list? [source,cypher] ---- MATCH (club:Club)<-[:PLAYS_FOR]-(player:Player) OPTIONAL MATCH (player)<-[voted:VOTED]-() RETURN club.name, count(DISTINCT player) AS count, sum(voted.score) AS totalScore ORDER BY totalScore DESC LIMIT 10 ---- .Which clubs have the highest ranking players on the list? [opts="header"] |=== | club | count | totalScore | "Liverpool" | 15 | 46803 | "Manchester City" | 17 | 25915 | "Barcelona" | 14 | 19154 | "Juventus" | 14 | 15265 | "Real Madrid" | 19 | 13078 | "Paris St-Germain" | 12 | 12500 | "Ajax" | 9 | 11608 | "Tottenham Hotspur" | 9 | 10956 | "Bayern Munich" | 14 | 10255 | "Atlético Madrid" | 10 | 4950 |=== That's more like it. Liverpool are way clear at the top, which makes sense given they won the Champions League and have a big lead in this year's Premier League season. After them come the 2018/2019 winners of the Premier League, La Liga, and Serie A, before we find Real Madrid in 5th place. The top 10 contains 3 clubs that play in England, 3 clubs that play in Spain, and 1 each from the Netherlands, Germany, France, and Italy. Let's next see which countries have the highest ranking players on the whole list. == Which countries have the highest ranking players on the list? [source,cypher] ---- MATCH (club:Club)<-[:PLAYS_FOR]-(player:Player) OPTIONAL MATCH (player)<-[voted:VOTED]-() WITH club, count(DISTINCT player) AS count, sum(voted.score) AS totalScore ORDER BY club, totalScore DESC MATCH (club)-[:PLAY_IN]->(country) RETURN country.name, sum(count) AS count, sum(totalScore) AS totalScore, collect([club.name, totalScore])[..5] AS clubs ORDER BY totalScore DESC LIMIT 5 ---- .Which clubs have the highest ranking players on the list? [opts="header", cols="1,1,1,6"] |=== | club | count | totalScore | clubs | "England" | 85 | 94221 | [["Liverpool", 46803], ["Manchester City", 25915], ["Tottenham Hotspur", 10956], ["Chelsea", 4273], ["Arsenal", 2635]] | "Spain" | 62 | 37926 | [["Barcelona", 19154], ["Real Madrid", 13078], ["Atlético Madrid", 4950], ["Villarreal", 206], ["Valencia", 210]] | "Italy" | 49 | 21167 | [["Juventus", 15269], ["Napoli", 2638], ["Internazionale", 1174], ["Lazio", 491], ["Atalanta", 504]] | "Germany" | 33 | 13882 | [["Bayern Munich", 10255], ["Borussia Dortmund", 2655], ["RB Leipzig", 576], ["Bayer Leverkusen", 279], ["Borussia Mönchengladbach", 72]] | "France" | 26 | 13170 | [["Paris St-Germain", 12500], ["Lyon", 368], ["Lille", 108], ["Monaco", 52], ["Nice", 71]] |=== Although both England and Spain had 3 players each in the top 10, the total score of those playing in England is almost twice as high as for those in Spain. This is mostly due to the high ranking of Manchester City and Liverpool players. Looking at the rest of the list, we can see that the top 5 countries are the big 5 European leagues, which is also where the majority of football's money lives. That's all I can think of for now, but if you can think of any other analysis we can do on this data, let me know in the comments!

Learn how to build a graph of the Guardian's Top 100 Male Footballers

uploads/2019/12/top-100-header.png

quick-graph-guardian-top-100-male-footballers

https://markhneedham.com/blog/2019/12/22/quick-graph-guardian-top-100-male-footballers

2021-03-28 00:44:37

Pandas: Filter column value in array/list - ValueError: The truth value of a Series is ambiguous

The UK government publishes Coronavirus vaccinations data on https://coronavirus.data.gov.uk/details/vaccinations[coronavirus.data.gov.uk^], but I wanted to create some different visualisations so I downloaded the data and have been playing with it in the https://github.com/mneedham/covid-vaccines[mneedham/covid-vaccines^] GitHub repository. I massaged the data so that I have rows in a Pandas DataFrame representing the numbers of first doses, second doses, and total doses done each day. I then wanted to filter this DataFrame based on the type of dose, but initially got a bit stuck. In this blog post we'll explore how I got around this problem. The DataFrame that we're working with looks like the one below: [source, python] ---- df = pd.DataFrame({ "date": ["2021-03-25", "2021-03-25", "2021-03-25"], "dose": ["firstDose", "secondDose", "totalDoses"], "vaccinations": [324942, 234382, 559324] }) ---- .Results [opts="header", cols="1,30,30,30"] |=== | | date |dose | vaccinations |0 |2021-03-25 |firstDose | 324942 |1 |2021-03-25 |secondDose | 234382 |2 | 2021-03-25 |totalDoses | 559324 |=== I wanted to filter the DataFrame to return only the rows with a `dose` value of `firstDose` or `secondDose`. One way of doing this would be to combine filters for `firstDose` or `secondDose`. I initially tried to do that using the Python `or` clause: [source, python] ---- df[(df.dose == "firstDose") or (df.dose == "secondDose")] ---- But that throws the following error: .Error [source, text] ---- Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/generic.py", line 1442, in __nonzero__ raise ValueError( ValueError: The truth value of a Series is ambiguous. Use a.empty, a.bool(), a.item(), a.any() or a.all(). ---- Instead we have to use the `|` operator, as shown below: [source, python] ---- df[(df.dose == "firstDose") | (df.dose == "secondDose")] ---- .Results [opts="header", cols="1,30,30,30"] |=== | | date |dose | vaccinations |0 |2021-03-25 |firstDose | 324942 |1 |2021-03-25 |secondDose | 234382 |=== That works fine if we're filtering on two values, but if we wanted to add a 3rd or 4th filter, our expression would get more complicated. Instead it would be cool if we could filter on a list of values. I initially tried to do this using the following code: [source, python] ---- df[df.dose in ["firstDose", "secondDose"]] ---- But that takes us back to the same error that we had before: .Error [source, text] ---- Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/generic.py", line 1442, in __nonzero__ raise ValueError( ValueError: The truth value of a Series is ambiguous. Use a.empty, a.bool(), a.item(), a.any() or a.all(). ---- This time we can use Pandas' https://pandas.pydata.org/docs/reference/api/pandas.Series.isin.html[`isin`^] function to solve the problem, as shown below: [source, python] ---- df[df.dose.isin(["firstDose", "secondDose"])] ---- .Results [opts="header", cols="1,30,30,30"] |=== | | date |dose | vaccinations |0 |2021-03-25 |firstDose | 324942 |1 |2021-03-25 |secondDose | 234382 |===

In this post we'll learn how to filter a Pandas DataFrame based on a column value existing in an array/list.

pandas-column-value-in-array-list-truth-value-ambiguous

https://markhneedham.com/blog/2021/03/28/pandas-column-value-in-array-list-truth-value-ambiguous

2021-04-02 00:44:37

Altair - Remove margin/padding on discrete X axis

One of the Altair charts on https://share.streamlit.io/mneedham/covid-vaccines/main/app.py[my Covid Vaccine Dashboards Streamlit app^] shows the % of first doses, but when I first created it there was some padding on the X axis that I wanted to remove. In this blog post we'll learn how to do that. == Pre requisites Let's start by installing the following libraries: [source, bash] ---- pip install pandas altair altair_viewer ---- Next let's import them, as shown below: [source, python] ---- import pandas as pd import altair as alt ---- == Visualising % of first doses Now we're going to create a DataFrame that contains two columns - one contains the year and week number, the other the percentage of 1st doses administered. [source, python] ---- df = pd.DataFrame({ 'dateWeek': ['2021-02', '2021-03', '2021-04', '2021-05', '2021-06', '2021-07', '2021-08', '2021-09', '2021-10', '2021-11', '2021-12', '2021-13'], 'percentageFirstDose': [95.53223579198118, 99.05357715009595, 99.29015227195728, 99.3040526396809, 99.17822125167659, 97.17701207004448, 93.13782375333588, 86.52577108509273, 80.75997640077365, 84.62332165884469, 67.53684465759456, 46.83433617577248] }) df ---- .Results [opts="header", cols="1,20,20"] |=== | |dateWeek |percentageFirstDose | 0 | 2021-02 | 95.532236 | 1 | 2021-03 | 99.053577 | 2 | 2021-04 | 99.290152 | 3 | 2021-05 | 99.304053 | 4 | 2021-06 | 99.178221 | 5 | 2021-07 | 97.177012 | 6 | 2021-08 | 93.137824 | 7 | 2021-09 | 86.525771 | 8 | 2021-10 | 80.759976 | 9 | 2021-11 | 84.623322 | 10 | 2021-12 | 67.536845 | 11 | 2021-13 | 46.834336 |=== Next we'll create a line chart using the Altair visualisation libray: [source, python] ---- chart = (alt.Chart(df, padding={"left": 10, "top": 10, "right": 10, "bottom": 10}) .mark_line(point=True) .encode( x=alt.X("dateWeek"), y=alt.Y('percentageFirstDose', axis=alt.Axis(title='% first dose'))) .properties(title="% of first doses by week")) chart.show() ---- .Chart showing % of first doses by week image::{{<siteurl>}}/uploads/2021/04/altair-padding.svg[] As we can see, this diagram has a padding/margin on either side of the X axis that I wanted to remove. After trying out a lot of things that didn't solve the problem, I came across https://github.com/altair-viz/altair/issues/558#issuecomment-371262131[a suggestion^] by Eitan Lees to set the `Scale` of the X axis. We can remove all padding by setting that value to 0, as shown below: [source, python] ---- chart = (alt.Chart(df, padding={"left": 10, "top": 10, "right": 10, "bottom": 10}) .mark_line(point=True) .encode( x=alt.X("dateWeek", scale=alt.Scale(padding=0)), y=alt.Y('percentageFirstDose', axis=alt.Axis(title='% first dose'))) .properties(title="% of first doses by week")) chart.show() ---- .Chart showing % of first doses by week image::{{<siteurl>}}/uploads/2021/04/altair-no-padding.svg[] Job done!

In this post we'll learn how to remove padding on the X axis of an Altair visualisation.

altair-discrete-x-axis-margin-padding

https://markhneedham.com/blog/2021/04/02/altair-discrete-x-axis-margin-padding

2021-04-11 00:44:37

Pandas - Format DataFrame numbers with commas and control decimal places

I'm still playing around with the UK's COVID-19 vaccination data and in this blog post we'll learn how to format a DataFrame that contains a mix of string and numeric values. [NOTE] ==== On 10th November 2022 I created https://www.youtube.com/embed/m1FEHPz90oI[a video^] that covers the same content as this blog post. Let me know if it's helpful 😊 ==== ++++ <iframe width="560" height="315" src="https://www.youtube.com/embed/m1FEHPz90oI" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> ++++ We'll be using Pandas' styling functionality, which generates CSS and HTML, so if you want to follow along you'll need to install Pandas and Jupyter: [source, bash] ---- pip install pandas jupyter ---- Next, launch Jupyter and create a notebook: [source, bash] ---- jupyter notebook ---- Let's start by importing Pandas: [source, python] ---- import pandas as pd ---- And now we'll create a DataFrame containing the data that we want to format: [source, python] ---- df = pd.DataFrame({ "LTLA Name": ["Amber Valley", "Ashfield", "Bassetlaw"], "Population": [72179, 77988, 70832], "PercentageVaccinated": [0.228571, 0.223342, 0.184493] }) ---- .Results [opts="header", cols="1,20,20,20"] |=== || LTLA Name | Population | PercentageVaccinated |0 |Amber Valley | 72179 | 0.228571 |1 | Ashfield | 77988 | 0.223342 |2 | Bassetlaw | 70832 | 0.184493 |=== One way to do this is to format the values in place, as shown below: [source, python] ---- df.loc[:, "Population"] = df["Population"].map('{:,d}'.format) df.loc[:, "PercentageVaccinated"] = df["PercentageVaccinated"].map('{:.2f}'.format) ---- After this transformation, the DataFrame looks like this: .Results [opts="header", cols="1,20,20,20"] |=== || LTLA Name | Population | PercentageVaccinated |0 |Amber Valley | 72,179 | 0.23 |1 | Ashfield | 77,988 | 0.22 |2 | Bassetlaw | 70,832 | 0.18 |=== This works, but it changes the underlying values in the DataFrame to be objects, which we can see by calling the `dtypes` function: [source, python] ---- df.dtypes ---- .Output [source, text] ---- LTLA Name object Population object PercentageVaccinated object dtype: object ---- This means that we can't do any number based computations anymore. For example, if we try to multiple the `Population` and `PercentageVaccinated` columns: [source, python] ---- df.Population * df.PercentageVaccinated ---- .Output [source, text] ---- Traceback (most recent call last): File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/array_ops.py", line 142, in _na_arithmetic_op result = expressions.evaluate(op, left, right) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/computation/expressions.py", line 235, in evaluate return _evaluate(op, op_str, a, b) # type: ignore[misc] File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/computation/expressions.py", line 69, in _evaluate_standard return op(a, b) TypeError: can't multiply sequence by non-int of type 'str' During handling of the above exception, another exception occurred: Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/common.py", line 65, in new_method return method(self, other) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/arraylike.py", line 105, in __mul__ return self._arith_method(other, operator.mul) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/series.py", line 4998, in _arith_method result = ops.arithmetic_op(lvalues, rvalues, op) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/array_ops.py", line 189, in arithmetic_op res_values = _na_arithmetic_op(lvalues, rvalues, op) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/array_ops.py", line 149, in _na_arithmetic_op result = _masked_arith_op(left, right, op) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/array_ops.py", line 91, in _masked_arith_op result[mask] = op(xrav[mask], yrav[mask]) TypeError: can't multiply sequence by non-int of type 'str' ---- So if we want to use the underlying data for anything else, formatting like this isn't a good solution. Instead we can use Pandas styling functionality. The documentation is a bit overwhelming, but Chris Moffitt https://pbpython.com/styling-pandas.html[has a great introductory article on the styling API^]. If we want to apply the same formatting to every column, we can pass a style to `style.format`. e.g. we could restrict every column to 2 decimal places, as shown below: [source, python] ---- df.style.format("{.2f") ---- [NOTE] ==== For a description of valid format values, see the https://docs.python.org/3/library/string.html#format-specification-mini-language[Format Specification Mini-Language^] documentation or https://mkaz.blog/code/python-string-format-cookbook/[Python String Format Cookbook^]. ==== [source, output] ---- --------------------------------------------------------------------------- ValueError Traceback (most recent call last) ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/IPython/core/formatters.py in __call__(self, obj) 343 method = get_real_method(obj, self.print_method) 344 if method is not None: --> 345 return method() 346 return None 347 else: ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/io/formats/style.py in _repr_html_(self) 203 Hooks into Jupyter notebook rich display system. 204 """ --> 205 return self.render() 206 207 @doc( ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/io/formats/style.py in render(self, **kwargs) 619 self._compute() 620 # TODO: namespace all the pandas keys --> 621 d = self._translate() 622 # filter out empty styles, every cell will have a class 623 # but the list of props may just be [['', '']]. ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/io/formats/style.py in _translate(self) 403 "value": value, 404 "class": " ".join(cs), --> 405 "display_value": formatter(value), 406 "is_visible": (c not in hidden_columns), 407 } ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/io/formats/style.py in <lambda>(x) 1715 ) -> Callable: 1716 if isinstance(formatter, str): -> 1717 formatter_func = lambda x: formatter.format(x) 1718 elif callable(formatter): 1719 formatter_func = formatter ValueError: Unknown format code 'f' for object of type 'str' ---- That doesn't work because the `LTLA Name` column contains string values, which can't be formatted as a number. We can work around that problem by dropping the `LTLA Name` column: [source, python] ---- df.drop(["LTLA Name"], axis=1).style.format("{:.2f}") ---- .Results [opts="header", cols="1,20,20"] |=== | | Population | PercentageVaccinated |0| 72179.00| 0.23 |1 |77988.00| 0.22 |2| 70832.00| 0.18 |=== This works, but we've lost the `LTLA Name` column and the `Population` column isn't formatted how we'd like. Instead of passing a single style to `style.format`, we can instead pass a dictionary of `{"column: "style"}`. So to style `Population` with a comma as thousands separator and `PercentageVaccinated` with two decimal places, we can do the following: [source, python] ---- df.style.format({ "Population": "{:,d}", "PercentageVaccinated": "{:.2f}" }) ---- .Results [opts="header", cols="1,20,20,20"] |=== || LTLA Name | Population | PercentageVaccinated |0 |Amber Valley | 72,179 | 0.23 |1 | Ashfield | 77,988 | 0.22 |2 | Bassetlaw | 70,832 | 0.18 |=== And if we go one step further, we can also use the `hide_index` function to get rid of the index column: [source, python] ---- df.style.format({ "Population": "{:,d}", "PercentageVaccinated": "{:.2f}" }).hide_index() ---- .Results [opts="header", cols="20,20,20"] |=== | LTLA Name | Population | PercentageVaccinated |Amber Valley | 72,179 | 0.23 | Ashfield | 77,988 | 0.22 | Bassetlaw | 70,832 | 0.18 |===

In this post we'll learn how to format numbers in Pandas DataFrames.

uploads/2021/04/pandas-format-banner.png

pandas-format-dataframe-numbers-commas-decimals

https://markhneedham.com/blog/2021/04/11/pandas-format-dataframe-numbers-commas-decimals

2021-04-28 00:44:37

Altair/Pandas: TypeError: Cannot interpret 'Float64Dtype()' as a data type

I ran into an interesting problem when trying to use Altair to visualise a Pandas DataFrame containing vaccination rates of different parts of England. In this blog post we'll look at how to work around this issue. First, let's install Pandas, numpy, and altair: [source, bash] ---- pip install pandas altair numpy ---- And now we'll import those modules into a Python script or Jupyter notebook: [source,python] ---- import pandas as pd import altair as alt import numpy as np ---- Next, we'll create a DataFrame containing the vaccinations rates of a couple of regions: [source,python] ---- vaccination_rates_by_region = pd.DataFrame([ {"Region": "East Midlands", "Overall": 48.877331}, {"Region": "London", "Overall": 32.58} ]) vaccination_rates_by_region = vaccination_rates_by_region.convert_dtypes() ---- .vaccination_rates_by_region DataFrame [format="csv", options="header"] |=== include::content/2021/04/28/data/vaccination_rates_by_region.csv[] |=== Let's now try to create a bar chart based on this data: [source,python] ---- (alt.Chart(vaccination_rates_by_region).mark_bar().encode( x=alt.X('Region'), y=alt.Y('Overall', axis=alt.Axis(title='Vaccinations')), tooltip=[alt.Tooltip('Overall', format=",")]) .properties(width=600)) ---- If we run this code, we'll see the following error: .Output [source, text] ---- TypeError Traceback (most recent call last) ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/altair/vegalite/v4/api.py in to_dict(self, *args, **kwargs) 361 copy = self.copy(deep=False) 362 original_data = getattr(copy, "data", Undefined) --> 363 copy.data = _prepare_data(original_data, context) 364 365 if original_data is not Undefined: ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/altair/vegalite/v4/api.py in _prepare_data(data, context) 82 # convert dataframes or objects with __geo_interface__ to dict 83 if isinstance(data, pd.DataFrame) or hasattr(data, "__geo_interface__"): ---> 84 data = _pipe(data, data_transformers.get()) 85 86 # convert string input to a URLData ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/toolz/functoolz.py in pipe(data, *funcs) 625 """ 626 for func in funcs: --> 627 data = func(data) 628 return data 629 ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/toolz/functoolz.py in __call__(self, *args, **kwargs) 301 def __call__(self, *args, **kwargs): 302 try: --> 303 return self._partial(*args, **kwargs) 304 except TypeError as exc: 305 if self._should_curry(args, kwargs, exc): ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/altair/vegalite/data.py in default_data_transformer(data, max_rows) 17 @curried.curry 18 def default_data_transformer(data, max_rows=5000): ---> 19 return curried.pipe(data, limit_rows(max_rows=max_rows), to_values) 20 21 ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/toolz/functoolz.py in pipe(data, *funcs) 625 """ 626 for func in funcs: --> 627 data = func(data) 628 return data 629 ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/toolz/functoolz.py in __call__(self, *args, **kwargs) 301 def __call__(self, *args, **kwargs): 302 try: --> 303 return self._partial(*args, **kwargs) 304 except TypeError as exc: 305 if self._should_curry(args, kwargs, exc): ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/altair/utils/data.py in to_values(data) 147 return {"values": data} 148 elif isinstance(data, pd.DataFrame): --> 149 data = sanitize_dataframe(data) 150 return {"values": data.to_dict(orient="records")} 151 elif isinstance(data, dict): ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/altair/utils/core.py in sanitize_dataframe(df) 334 col = df[col_name].astype(object) 335 df[col_name] = col.where(col.notnull(), None) --> 336 elif np.issubdtype(dtype, np.integer): 337 # convert integers to objects; np.int is not JSON serializable 338 df[col_name] = df[col_name].astype(object) ~/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/numpy/core/numerictypes.py in issubdtype(arg1, arg2) 417 """ 418 if not issubclass_(arg1, generic): --> 419 arg1 = dtype(arg1).type 420 if not issubclass_(arg2, generic): 421 arg2 = dtype(arg2).type TypeError: Cannot interpret 'Float64Dtype()' as a data type ---- We can check the types used in our DataFrame by running the following code: [source, python] ---- vaccination_rates_by_region.dtypes ---- .Output [source, text] ---- Region string Overall Float64 dtype: object ---- The problem is that altair https://github.com/altair-viz/altair/issues/2398[doesn't yet support the Float64Dtype^] type. We can work around this problem by coercing the type of that column to `float32`: [source, python] ---- vaccination_rates_by_region= vaccination_rates_by_region.astype({ column: np.float32 for column in vaccination_rates_by_region.drop(["Region"], axis=1).columns }) ---- And now if we create a chart: [source, python] ---- chart = (alt.Chart(vaccination_rates_by_region).mark_bar().encode( x=alt.X('Region'), y=alt.Y('Overall', axis=alt.Axis(title='Vaccinations')), tooltip=[alt.Tooltip('Overall', format=",")]) .properties(width=600)) ---- image::{{<siteurl>}}/uploads/2021/04/vaccinations-by-region.png[title="Vaccinations by Region"] Happy days!

In this post we'll learn how to work around an issue with Pandas new Float64Dtype when using the Altair visualisation library.

altair-pandas-cannot-interpret-Float64Dtype-as-data-type

https://markhneedham.com/blog/2021/04/28/altair-pandas-cannot-interpret-Float64Dtype-as-data-type

2021-04-17 00:44:37

Vaccinating England: The Data (cleanup)

Over the last 13 months I've spent countless hours looking at dashboards that showed Coronavirus infection rates, death rates, and numbers of people vaccinated. The UK government host a dashboard at https://coronavirus.data.gov.uk[coronavirus.data.gov.uk^], which contains charts and tables showing all of the above. One thing I haven't been able to find, however, is a drill down of vaccinations by local area and age group. So I'm going to try to build my own! image::{{<siteurl>}}/uploads/2021/04/covid-vaccines.png[] This is the first in a series of blog posts showing how to do that, in this post we'll find the data and get it into a format that's easier to process. == The Data First, we need to find the data! I haven't found anywhere that contains the data for the whole of the United Kingdom, but Public Health England do https://www.england.nhs.uk/statistics/statistical-work-areas/covid-19-vaccinations/[post the data^] about vaccinations every Thursday. image::{{<siteurl>}}/uploads/2021/04/phe.png[title="Public Health England Data"] This page contains various PDFs and spreadsheets and the 'Weekly data' spreadsheet is the one that we want. This spreadsheet contains data on vaccinations by local area and age group, as well as the populations of those areas and age groups. == Reading Excel Sheets with Pandas We're going to use Python's Pandas library to read the data from the spreadsheet, but we'll also need to install some other libraries that deal with reading `xlsx` files: [source,bash] ---- pip install numpy pandas xlrd openpyxl ---- Once we've done that, let's import pandas and numpy: [source, python] ---- import pandas as pd import numpy as np ---- == Vaccinations We're going to start by extracting vaccination data. This is stored in the `LTLA` sheet: image::{{<siteurl>}}/uploads/2021/04/ltla-spreadsheet-vaccinations.png[title="Vaccinations - LTLA Spreadsheet"] From visually inspecting the spreadsheet, we can see that columns B to M contain the data we're interested in. We can extract just those columns by using the `usecols` parameter, as shown below: [source, python] ---- spreadsheet = "data/COVID-19-weekly-announced-vaccinations-15-April-2021.xlsx" vaccinations = pd.read_excel(spreadsheet, "LTLA", usecols="B:M") vaccinations.head() ---- .vaccinations DataFrame [opts="header"] |=== | Unnamed: 1 | Unnamed: 2 | Unnamed: 3 | Unnamed: 4 | Unnamed: 5 | Unnamed: 6 | Unnamed: 7 | Unnamed: 8 | Unnamed: 9 | Unnamed: 10 | Unnamed: 11 | Unnamed: 12 | Title: | COVID-19 Vaccinations By Lower Tier Local Authority (LTLA) of Residence and Age Group | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | Summary: | The number of people who have been vaccinated for COVID-19, split by LTLA of residence and age group. All figures are presented by date of vaccination as recorded on the National Immunisation Management Service (NIMS) database. NIMS vaccination counts and populations by LTLA are provided by Public Health England. | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | Period: | 8th December 2020 to 11th April 2021 | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | Source: | National Immunisation Management Service (NIMS), Public Health England | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan | Basis: | England | nan | nan | nan | nan | nan | nan | nan | nan | nan | nan |=== We clearly still have some cleanup work to do. We already filtered on columns when loading the sheet, so let's now select just the rows that we need: [source,python] ---- spreadsheet = "data/COVID-19-weekly-announced-vaccinations-15-April-2021.xlsx" vaccinations = pd.read_excel(spreadsheet, "LTLA", usecols="B:M") vaccinations = vaccinations.loc[14:327,] ---- .vaccinations DataFrame [opts="header"] |=== | Unnamed: 1 | Unnamed: 2 | Unnamed: 3 | Unnamed: 4 | Unnamed: 5 | Unnamed: 6 | Unnamed: 7 | Unnamed: 8 | Unnamed: 9 | Unnamed: 10 | Unnamed: 11 | Unnamed: 12 | E12000004 | East Midlands | E07000032 | Amber Valley | 20018 | 9225 | 9212 | 8078 | 7294 | 7968 | 5798 | 6808 | E12000004 | East Midlands | E07000170 | Ashfield | 18535 | 8581 | 8581 | 7189 | 6358 | 6465 | 5163 | 6070 | E12000004 | East Midlands | E07000171 | Bassetlaw | 17562 | 8208 | 8629 | 7450 | 6589 | 7098 | 5247 | 6428 | E12000004 | East Midlands | E07000129 | Blaby | 16281 | 6770 | 6647 | 6004 | 5157 | 5601 | 4166 | 5559 | E12000004 | East Midlands | E07000033 | Bolsover | 11594 | 5819 | 5644 | 4967 | 4213 | 4249 | 3411 | 3825 |=== That's looking better already, but those column names aren't great. Let's extract those and set them on our DataFrame: [source,python] ---- spreadsheet = "data/COVID-19-weekly-announced-vaccinations-15-April-2021.xlsx" vaccinations = pd.read_excel(spreadsheet, "LTLA", usecols="B:M") columns = np.concatenate( (vaccinations.loc[10,:][:4].values, vaccinations.loc[11,:][4:].values), axis=None) vaccinations = vaccinations.loc[14:327,] vaccinations.columns = columns ---- .vaccinations DataFrame [opts="header"] |=== | Region Code (Administrative) | Region Name (administrative) | LTLA Code | LTLA Name | Under 50 | 50-54 | 55-59 | 60-64 | 65-69 | 70-74 | 75-79 | 80+ | E12000004 | East Midlands | E07000032 | Amber Valley | 20018 | 9225 | 9212 | 8078 | 7294 | 7968 | 5798 | 6808 | E12000004 | East Midlands | E07000170 | Ashfield | 18535 | 8581 | 8581 | 7189 | 6358 | 6465 | 5163 | 6070 | E12000004 | East Midlands | E07000171 | Bassetlaw | 17562 | 8208 | 8629 | 7450 | 6589 | 7098 | 5247 | 6428 | E12000004 | East Midlands | E07000129 | Blaby | 16281 | 6770 | 6647 | 6004 | 5157 | 5601 | 4166 | 5559 | E12000004 | East Midlands | E07000033 | Bolsover | 11594 | 5819 | 5644 | 4967 | 4213 | 4249 | 3411 | 3825 |=== This is all looking good, but there's one final bit of cleanup that we need to do. The `dtypes` function returns the data type of each column. Let's check what it returns for our DataFrame: [source,python] ---- vaccinations.dtypes ---- .Output [source,text] ---- Region Code (Administrative) object Region Name (administrative) object LTLA Code object LTLA Name object Under 50 object 50-54 object 55-59 object 60-64 object 65-69 object 70-74 object 75-79 object 80+ object dtype: object ---- Hmmm, every column is an object, when we know that some columns only have numeric values. We can fix column types by calling the `convert_dtypes` function: [source, python] ---- vaccinations.convert_dtypes().dtypes ---- .Output [source,text] ---- Region Code (Administrative) string Region Name (administrative) string LTLA Code string LTLA Name string Under 50 Int64 50-54 Int64 55-59 Int64 60-64 Int64 65-69 Int64 70-74 Int64 75-79 Int64 80+ Int64 dtype: object ---- Looks good. Let's now wrap all of that code into a function: [source, python] ---- def vaccinations_dataframe(spreadsheet): vaccinations = pd.read_excel(spreadsheet, "LTLA", usecols="B:M") columns = np.concatenate( (vaccinations.loc[10,:][:4].values, vaccinations.loc[11,:][4:].values), axis=None) vaccinations = vaccinations.loc[14:327,] vaccinations.columns = columns vaccinations = vaccinations.convert_dtypes() return vaccinations ---- [source, python] ---- spreadsheet = "data/COVID-19-weekly-announced-vaccinations-15-April-2021.xlsx" vaccinations = vaccinations_dataframe(spreadsheet) ---- == Population Now we'll do the same for each Local Tier Local Authority's population, which we can find in the `Population estimates (NIMS)` sheet. image::{{<siteurl>}}/uploads/2021/04/ltla-spreadsheet-population.png[title="Population - LTLA Spreadsheet"] We'll skip straight to the final function for creating the DataFrame: [source, python] ---- def population_dataframe(spreadsheet): population = pd.read_excel(spreadsheet, "Population estimates (NIMS)", usecols="B:L") population_columns = np.concatenate( (population.loc[10,:][:2], population.loc[11, :][2:]), axis=None) population = population.loc[14:327,] population.columns = population_columns population.insert( loc=2, column="Under 50", value=population["Under 16"] + population["16-49"]) population = population.drop(["Under 16", "16-49"], axis=1) population = population.convert_dtypes() return population ---- One extra thing we had to do was add the `Under 50` column, since that didn't exist in the spreadsheet. We need that column to exist so that we can directly map the population with the number of people vaccinated per age group per area. Now let's use that function to extract the population for each area: [source, python] ---- spreadsheet = "data/COVID-19-weekly-announced-vaccinations-15-April-2021.xlsx" population = population_dataframe(spreadsheet) ---- .population DataFrame [opts="header"] |=== | LTLA Code | LTLA Name | Under 50 | 50-54 | 55-59 | 60-64 | 65-69 | 70-74 | 75-79 | 80+ | E07000032 | Amber Valley | 72286 | 10198 | 10139 | 8672 | 7633 | 8242 | 5928 | 7010 | E07000170 | Ashfield | 78053 | 9798 | 9503 | 7886 | 6752 | 6769 | 5357 | 6316 | E07000171 | Bassetlaw | 70918 | 9376 | 9572 | 8099 | 6997 | 7367 | 5414 | 6636 | E07000129 | Blaby | 62265 | 7595 | 7395 | 6528 | 5465 | 5801 | 4290 | 5781 | E07000033 | Bolsover | 48519 | 6660 | 6280 | 5376 | 4498 | 4445 | 3501 | 3947 |=== We've now got DataFrames that contain vaccination and population data. Next we're going to compute the vaccination rate per age group per area. == Combining Vaccinations and Populations We can compute the vaccination rate by dividing the non string columns in `vaccinations` by the non string columns in `population` and then multiplying by 100. We can see how to do this in the following code: [source,python] ---- vaccination_rates = (vaccinations.select_dtypes(exclude='string') .div(population.select_dtypes(exclude='string')) * 100) ---- .Vaccination Rates [opts="header"] |=== | Under 50 | 50-54 | 55-59 | 60-64 | 65-69 | 70-74 | 75-79 | 80+ | 27.6928 | 90.4589 | 90.8571 | 93.1504 | 95.5588 | 96.6756 | 97.807 | 97.1184 | 23.7467 | 87.5791 | 90.2978 | 91.1616 | 94.1647 | 95.5089 | 96.3786 | 96.1051 | 24.7638 | 87.5427 | 90.1483 | 91.9867 | 94.1689 | 96.3486 | 96.9154 | 96.8656 | 26.1479 | 89.1376 | 89.8851 | 91.973 | 94.3641 | 96.5523 | 97.1096 | 96.1598 | 23.8958 | 87.3724 | 89.8726 | 92.3921 | 93.6639 | 95.5906 | 97.4293 | 96.909 |=== That looks good, but we've lost the area names, which we can get back by calling the https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.combine_first.html[`combine_first`^] function: [source,python] ---- vaccination_rates = ( (vaccinations.select_dtypes(exclude='string') .div(population.select_dtypes(exclude='string')) * 100) .combine_first(vaccinations)[vaccinations.columns]) vaccination_rates = vaccination_rates.convert_dtypes() ---- .Vaccination Rates [opts="header"] |=== | Region Code (Administrative) | Region Name (administrative) | LTLA Code | LTLA Name | Under 50 | 50-54 | 55-59 | 60-64 | 65-69 | 70-74 | 75-79 | 80+ | E12000004 | East Midlands | E07000032 | Amber Valley | 27.6928 | 90.4589 | 90.8571 | 93.1504 | 95.5588 | 96.6756 | 97.807 | 97.1184 | E12000004 | East Midlands | E07000170 | Ashfield | 23.7467 | 87.5791 | 90.2978 | 91.1616 | 94.1647 | 95.5089 | 96.3786 | 96.1051 | E12000004 | East Midlands | E07000171 | Bassetlaw | 24.7638 | 87.5427 | 90.1483 | 91.9867 | 94.1689 | 96.3486 | 96.9154 | 96.8656 | E12000004 | East Midlands | E07000129 | Blaby | 26.1479 | 89.1376 | 89.8851 | 91.973 | 94.3641 | 96.5523 | 97.1096 | 96.1598 | E12000004 | East Midlands | E07000033 | Bolsover | 23.8958 | 87.3724 | 89.8726 | 92.3921 | 93.6639 | 95.5906 | 97.4293 | 96.909 |=== == Exploring vaccination rates Now that we've compute vaccination rates, let's do a bit of exploration of the data. We can find the maximum rates in each age group using the `max` function: [source,python] ---- vaccination_rates.max() ---- .Vaccination Rates Maximums [source, text] ---- Region Code (Administrative) E12000009 Region Name (administrative) Yorkshire and The Humber LTLA Code E09000033 LTLA Name York Under 50 39.916551 50-54 91.650635 55-59 93.316424 60-64 94.840627 65-69 96.086819 70-74 99.305556 75-79 98.061389 80+ 99.386503 dtype: object ---- We can ignore the values for the string columns, they don't really mean anything. But if we look at the age categories, we can see that in one area 99.38% of over 80s have received one doses of the vaccine and the best area for under 50s has vaccinated almost 40% of that cohort. We can find out which areas those are by sorting on the appropriate column. For example, these are the top areas for vaccinating Under 50s: [source, python] ---- vaccination_rates.sort_values(["Under 50"], ascending=False).head() ---- .Vaccination Rates - Best for Under 50s [opts="header"] |=== | Region Code (Administrative) | Region Name (administrative) | LTLA Code | LTLA Name | Under 50 | 50-54 | 55-59 | 60-64 | 65-69 | 70-74 | 75-79 | 80+ | E12000009 | South West | E06000053 | Isles of Scilly | 39.9166 | 89.8734 | 89.4444 | 91.1765 | 95.2381 | 99.3056 | 95.4545 | 99.3865 | E12000008 | South East | E07000090 | Havant | 28.4815 | 89.6867 | 91.423 | 93.2137 | 94.7948 | 95.7627 | 96.7189 | 96.7105 | E12000004 | East Midlands | E07000034 | Chesterfield | 28.4194 | 87.8095 | 89.4942 | 92.7704 | 94.8408 | 95.9601 | 96.4349 | 96.7665 | E12000004 | East Midlands | E07000032 | Amber Valley | 27.6928 | 90.4589 | 90.8571 | 93.1504 | 95.5588 | 96.6756 | 97.807 | 97.1184 | E12000002 | North West | E07000128 | Wyre | 27.56 | 87.7091 | 89.6141 | 92.7322 | 94.5789 | 95.7072 | 96.8809 | 96.7518 |=== We could also group by `Region Name` and compute the average vaccination rate per age group: [source, python] ---- average_rates = vaccination_rates.groupby(["Region Name (administrative)"]).mean() average_rates.insert(0, "Region", list(average_rates.index)) ---- .Vaccination Rates - By Region [opts="header",cols="2,1,1,1,1,1,1,1,1"] |=== | Region | Under 50 | 50-54 | 55-59 | 60-64 | 65-69 | 70-74 | 75-79 | 80+ | East Midlands | 21.3713 | 86.2392 | 88.4962 | 90.6407 | 93.2752 | 95.3722 | 96.3283 | 96.1347 | East of England | 18.4956 | 85.1457 | 87.6499 | 89.4512 | 91.7769 | 94.2033 | 95.3887 | 95.6347 | London | 14.4234 | 69.3903 | 73.0272 | 76.6469 | 80.2523 | 83.9104 | 85.1132 | 84.9085 | North East | 18.7766 | 84.5177 | 88.4894 | 91.2096 | 93.7479 | 95.6958 | 96.8194 | 96.4659 | North West | 19.751 | 83.3976 | 86.5955 | 89.3941 | 91.9802 | 94.3731 | 95.5474 | 95.4938 | South East | 18.0919 | 85.6189 | 87.8544 | 89.6232 | 91.963 | 94.2038 | 95.2525 | 95.4839 | South West | 19.7912 | 87.5552 | 89.477 | 90.7873 | 93.1417 | 95.1491 | 96.0111 | 96.4188 | West Midlands | 20.426 | 85.6018 | 87.7443 | 90.4331 | 92.825 | 94.8981 | 95.8844 | 95.4323 | Yorkshire and The Humber | 19.7412 | 85.6536 | 88.9223 | 90.8691 | 93.3894 | 95.4119 | 96.4073 | 96.2381 |=== Most of the country is doing really well, but it's interesting to see that London has much lower vaccination rates across every category. I assume this is due to a combination of vaccine hesitancy, difficulty getting to the vaccination centres, or perhaps there's an error with the data. This isn't a perfect calculation though because we're assigning equal weight to each local area, which will vary in size. That means that we might be giving too much weight to smaller places with especially high or low vaccine uptake. Let's instead sum together the vaccinations given and populations per region and compute the percentages again. First we'll create DataFrames for each region: [source, python] ---- population.loc[:, "Region Name (administrative)"] = vaccinations["Region Name (administrative)"] population_by_region = population.groupby(["Region Name (administrative)"]).sum() population_by_region.insert(0, "Region", list(population_by_region.index)) vaccinations_by_region = vaccinations.groupby(["Region Name (administrative)"]).sum() vaccinations_by_region.insert(0, "Region", list(vaccinations_by_region.index)) ---- .population_by_region DataFrame [opts="header",cols="2,1,1,1,1,1,1,1,1"] |=== | Region | Under 50 | 50-54 | 55-59 | 60-64 | 65-69 | 70-74 | 75-79 | 80+ | East Midlands | 3170332 | 364235 | 356411 | 304211 | 261878 | 265843 | 193933 | 249586 | East of England | 4165578 | 470144 | 456846 | 387478 | 331616 | 344449 | 252007 | 346692 | London | 7592384 | 664924 | 587034 | 458373 | 342329 | 293428 | 204762 | 297433 | North East | 1687882 | 193270 | 200330 | 180677 | 154636 | 150338 | 102373 | 140390 | North West | 4953735 | 549622 | 544180 | 465484 | 389203 | 388096 | 274374 | 364876 | South East | 6112065 | 695721 | 676804 | 570183 | 480072 | 493718 | 365478 | 510052 | South West | 3496476 | 415359 | 424070 | 374972 | 333106 | 349326 | 260864 | 355024 | West Midlands | 4019092 | 439809 | 418626 | 356127 | 304688 | 300890 | 230075 | 308908 | Yorkshire and The Humber | 3670427 | 404075 | 392613 | 341298 | 289223 | 289865 | 204261 | 277539 |=== .vaccinations_by_region DataFrame [opts="header",cols="2,1,1,1,1,1,1,1,1"] |=== | Region | Under 50 | 50-54 | 55-59 | 60-64 | 65-69 | 70-74 | 75-79 | 80+ | East Midlands | 658569 | 310156 | 312653 | 274052 | 243321 | 253103 | 186500 | 239377 | East of England | 755452 | 399663 | 400596 | 347169 | 305122 | 325649 | 241126 | 332207 | London | 1097594 | 459513 | 428659 | 352909 | 276040 | 248388 | 176083 | 255922 | North East | 313030 | 164486 | 178221 | 165482 | 145496 | 144315 | 99336 | 135749 | North West | 937404 | 451446 | 466463 | 412713 | 356303 | 365507 | 261756 | 347927 | South East | 1078781 | 594115 | 594208 | 511143 | 442087 | 466018 | 348965 | 487877 | South West | 668237 | 361313 | 377659 | 339592 | 309999 | 332055 | 250762 | 342276 | West Midlands | 763887 | 361823 | 357134 | 315110 | 278568 | 283073 | 218910 | 291583 | Yorkshire and The Humber | 694645 | 339544 | 344640 | 307351 | 268164 | 275646 | 196338 | 266059 |=== And now we can compute the vaccination rate per region: [source, python] ---- vaccination_rates_by_region = ((vaccinations_by_region .select_dtypes(exclude='object') .div(population_by_region.select_dtypes(exclude='object')) * 100) .combine_first(vaccinations_by_region)[vaccinations_by_region.columns]) vaccination_rates_by_region = vaccination_rates_by_region.convert_dtypes() ---- .vaccinations_by_region DataFrame [opts="header",cols="2,1,1,1,1,1,1,1,1"] |=== | Region | Under 50 | 50-54 | 55-59 | 60-64 | 65-69 | 70-74 | 75-79 | 80+ | East Midlands | 20.7729 | 85.1527 | 87.7226 | 90.0862 | 92.9139 | 95.2077 | 96.1672 | 95.9096 | East of England | 18.1356 | 85.0086 | 87.6873 | 89.5971 | 92.0106 | 94.542 | 95.6823 | 95.8219 | London | 14.4565 | 69.1076 | 73.0212 | 76.9917 | 80.6359 | 84.6504 | 85.994 | 86.0436 | North East | 18.5457 | 85.1068 | 88.9637 | 91.59 | 94.0893 | 95.9937 | 97.0334 | 96.6942 | North West | 18.9232 | 82.1375 | 85.7185 | 88.6632 | 91.5468 | 94.1795 | 95.4012 | 95.3549 | South East | 17.65 | 85.3956 | 87.7962 | 89.6454 | 92.0876 | 94.3895 | 95.4818 | 95.6524 | South West | 19.1117 | 86.9881 | 89.0558 | 90.5646 | 93.0632 | 95.0559 | 96.1275 | 96.4093 | West Midlands | 19.0065 | 82.2682 | 85.311 | 88.4825 | 91.4273 | 94.0786 | 95.1472 | 94.3915 | Yorkshire and The Humber | 18.9255 | 84.0299 | 87.7811 | 90.0536 | 92.7188 | 95.0946 | 96.1211 | 95.8636 |=== London is doing slightly better on some of the age groups measured like this, but still seems to be behind the other regions. We could do some further analysis to figure out if there are specific areas in London that have lower uptake, but I think that can wait for another post. //// print(tabulate(vaccinations.head().values,vaccinations.columns, tablefmt="pipe")) ////

In this post we'll learn how to extract data about England's Covid vaccinations from a Public Health England spreadsheet.

uploads/2021/04/covid-vaccines.png

england-covid-vaccination-rates-the-data

https://markhneedham.com/blog/2021/04/17/england-covid-vaccination-rates-the-data

2021-04-21 00:44:37

Pandas: Compare values in DataFrame to previous days

I'm still playing around with Covid vaccine data, this time exploring how the number of doses varies week by week. I want to know how many more (or less) vaccines have been done on a given day compared to that same day last week. We'll be using Pandas in this blog post, so let's first install that library and import it: .Install Pandas [source, bash] ---- pip install pandas ---- .Import module [source,python] ---- import pandas as pd ---- And now let's create a DataFrame containing a subset of the data that I'm working with: [source,python] ---- df = pd.DataFrame([ {'date': '2021-01-11', 'vaccinations': 165844.0}, {'date': '2021-01-18', 'vaccinations': 208641.0}, {'date': '2021-01-25', 'vaccinations': 281725.0}, {'date': '2021-02-01', 'vaccinations': 352935.0}, {'date': '2021-02-08', 'vaccinations': 356291.0} ]) ---- .Output [format="csv", options="header"] |=== include::content/2021/04/21/data/initial.csv[] |=== I've filtered the data to include the data for 5 Mondays start from the beginning of January. In the real data set we have the data for every day from January until today. I want to add a column that shows the number of vaccinations done on the previous week. So: * for `2021-01-18`, we'd have the value for `2021-01-11` * for `2021-01-25`, we'd have the value for `2021-02-18` and so on. It took me a while to find the function to do this, but it turns out that https://pandas.pydata.org/docs/reference/api/pandas.Series.shift.html[`pandas.Series.shift`] is what we want. This function shifts the index by the desired number of periods. We'll pass in `periods=1` because we want to shift every row down by 1: [source, python] ---- df["vaccinations"].shift(periods=1) ---- [NOTE] ==== I should note that for the real DataFrame I had data for every day and therefore wanted to compare the data from 7 rows earlier in the DataFrame,which meant I had to use `periods=7`. If we wanted to compare the data from 3 days earlier, we could use `periods=3`, and so on. ==== .Output [format="csv", options="header"] |=== include::content/2021/04/21/data/shift_standalone.csv[] |=== We can then add a new column to our DataFrame using the following code: [source, python] ---- df.loc[:, "lastWeekVaccinations"] = df["vaccinations"].shift(periods=1) ---- .Output [format="csv", options="header"] |=== include::content/2021/04/21/data/shift.csv[] |=== If we then wanted to compute the difference between this week and last week, we could compute this by writing the following code: [source, python] ---- df["vaccinations"] - df["lastWeekVaccinations"] ---- Or we could use the https://pandas.pydata.org/docs/reference/api/pandas.Series.diff.html[`pandas.Series.diff`] function, which achieves the same thing: [source, python] ---- df.loc[:, "lastWeekVaccinationsDiff"] = df["vaccinations"].diff(periods=1) ---- .Output [format="csv", options="header"] |=== include::content/2021/04/21/data/diff.csv[] |=== We can also compute the percentage change between the weeks using https://pandas.pydata.org/docs/reference/api/pandas.Series.pct_change.html[`pandas.Series.pct_change`]: [source, python] ---- df.loc[:, "lastWeekVaccinationsChange"] = df["vaccinations"].pct_change(periods=1) ---- .Output [format="csv", options="header"] |=== include::content/2021/04/21/data/change.csv[] |=== With these functions I've been able to do exactly what I wanted and could then create a chart based on this data.

In this post we'll learn how to compare values from earlier rows in a Pandas DataFrame.

pandas-compare-dataframe-to-previous-days

https://markhneedham.com/blog/2021/04/21/pandas-compare-dataframe-to-previous-days

2021-04-08 00:44:37

Pandas - Dividing two DataFrames (TypeError: unsupported operand type(s) for /: 'str' and 'str')

I've been doing some more exploration of the UK Coronavirus vaccine data, this time looking at the https://www.england.nhs.uk/statistics/statistical-work-areas/covid-19-vaccinations/[number of people vaccinated by Local Tier Local Authority^]. The government publish data showing the number of people vaccinated in each authority by age group, as well as population estimates for each cohort. Having loaded that data into two Pandas DataFrames, I wanted to work out the % of people vaccinated per age group per local area. In this blog post we'll see how to do that, including my missteps along the way. Let's start by importing Pandas: [source, python] ---- import pandas as pd ---- And now we're going to create DataFrames with a subset of the data from the Public Health England spreadsheet as of 1st April 2021. First for the number of vaccinations given: [source, python] ---- vaccinations = pd.DataFrame({ "LTLA Code": ["E07000032", "E07000170", "E07000171"], "LTLA Name": ["Amber Valley", "Ashfield", "Bassetlaw"], "Under 50": [16498, 17418, 13068], "50-54": [9002, 8399, 7858] }) vaccinations ---- .vaccinations DataFrame [opts="header", cols="1,10,20,20,20"] |=== | |LTLA Code | LTLA Name | Under 50 |50-54 |0 | E07000032 | Amber Valley | 16498 | 9002 |1 |E07000170 | Ashfield | 17418 | 8399 |2 |E07000171 | Bassetlaw | 13068 | 7858 |=== And now the population estimates: [source,python] ---- population = pd.DataFrame({ "LTLA Code": ["E07000032", "E07000170", "E07000171"], "LTLA Name": ["Amber Valley", "Ashfield", "Bassetlaw"], "Under 50": [72179, 77988, 70832], "50-54": [10194, 9795, 9354] }) population ---- .population DataFrame [opts="header", cols="1,10,20,20,20"] |=== | |LTLA Code | LTLA Name | Under 50 |50-54 |0 | E07000032 | Amber Valley | 72179 | 10194 |1 |E07000170 | Ashfield | 77988 | 9795 |2 |E07000171 | Bassetlaw | 70832 | 9354 |=== For my first attempt at working out the % of people vaccinated per age group, I optimistically tried dividing the DataFrames, as shown below: [source, python] ---- vaccinations / population ---- .Output [source, text] ---- Traceback (most recent call last): File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/array_ops.py", line 142, in _na_arithmetic_op result = expressions.evaluate(op, left, right) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/computation/expressions.py", line 235, in evaluate return _evaluate(op, op_str, a, b) # type: ignore[misc] File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/computation/expressions.py", line 69, in _evaluate_standard return op(a, b) TypeError: unsupported operand type(s) for /: 'str' and 'str' During handling of the above exception, another exception occurred: Traceback (most recent call last): File "<stdin>", line 1, in <module> File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/common.py", line 65, in new_method return method(self, other) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/arraylike.py", line 113, in __truediv__ return self._arith_method(other, operator.truediv) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/frame.py", line 5982, in _arith_method new_data = self._dispatch_frame_op(other, op, axis=axis) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/frame.py", line 6018, in _dispatch_frame_op bm = self._mgr.operate_blockwise(right._mgr, array_op) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/internals/managers.py", line 374, in operate_blockwise return operate_blockwise(self, other, array_op) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/internals/ops.py", line 54, in operate_blockwise res_values = array_op(lvals, rvals) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/array_ops.py", line 189, in arithmetic_op res_values = _na_arithmetic_op(lvalues, rvalues, op) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/array_ops.py", line 149, in _na_arithmetic_op result = _masked_arith_op(left, right, op) File "/home/markhneedham/.local/share/virtualenvs/covid-vaccines-xEbcGJTy/lib/python3.8/site-packages/pandas/core/ops/array_ops.py", line 91, in _masked_arith_op result[mask] = op(xrav[mask], yrav[mask]) TypeError: unsupported operand type(s) for /: 'str' and 'str' ---- That doesn't work so well because we still have `LTLA Code` and `LTLA Name`, which aren't numeric values and therefore can't be divided. What we can do instead is use the `divide` function, which lets us pass in a mask that will exclude non-numeric columns. I learnt about this thanks to https://stackoverflow.com/a/49412743/1093511[a StackOverflow answer^] by cs95. Before we try that, we need to check the types of our columns so that we can exclude the right ones. We can do this using the following code: [source,python] ---- vaccinations.dtypes ---- .Output [source,text] ---- LTLA Code object LTLA Name object Under 50 int64 50-54 int64 dtype: object ---- [source,python] ---- population.dtypes ---- .Output [source,text] ---- LTLA Code object LTLA Name object Under 50 int64 50-54 int64 dtype: object ---- So it looks like 'object' is the type that we need to exclude. [NOTE] ==== While playing around with this I found that sometimes columns containing strings might have the type `string`, in which case we'd need to exclude that type instead. ==== Let's now try to divide the DataFrames, excluding any columns of type `object`: [source, python] ---- vaccinations.select_dtypes(exclude='object').div(population.select_dtypes(exclude='object')) ---- .Results [opts="header", cols="1,20,20"] |=== | |Under 50 |50-54 |0 | 0.228571 | 0.883068 |1 |0.223342 |0.857478 |2 |0.184493 | 0.840068 |=== Cool! That's what we hoped to see, but it would be good if we could also have the `LTLA Code` and `LTLA Name` columns back as well. We can do this using the `combine_first` function: [source, python] ---- vaccinations.select_dtypes(exclude='object').div(population.select_dtypes(exclude='object')).combine_first(population) ---- .Results [opts="header", cols="1,15,15,15,15"] |=== | | 50-54 | LTLA Code | LTLA Name |Under 50 |0 | 0.883068 | E07000032 |Amber Valley |0.228571 |1 |0.857478 |E07000170 | Ashfield |0.223342 |2 |0.840068 |E07000171 | Bassetlaw |0.184493 |=== That's got the columns back, but the order is a bit messed up. Luckily it's not too difficult to retain the column order: [source, python] ---- vaccinations.select_dtypes(exclude='object').div(population.select_dtypes(exclude='object')).combine_first(population)[vaccinations.columns] ---- .Results [opts="header", cols="1,15,15,20,20"] |=== || LTLA Code | LTLA Name |Under 50 | 50-54 |0 |E07000032 |Amber Valley |0.228571 |0.883068 |1 |E07000170 | Ashfield |0.223342 | 0.857478 |2 | E07000171 | Bassetlaw |0.184493 |0.840068 |===

In this post we'll learn how to divide only the numeric columns in two Pandas DataFrames.

pandas-divide-dataframes-unsupported-operand-type-str

https://markhneedham.com/blog/2021/04/08/pandas-divide-dataframes-unsupported-operand-type-str

2021-05-19 00:44:37

jq: Select multiple keys

I recently started a new job, working for a FinTech company called Finbourne, who build a data platform for investment data. It's an API first product that publishes a Swagger API JSON file that I've been trying to parse to get a list of the end points and their operation ids. In this blog post I'll show how I've been parsing that file using https://stedolan.github.io/jq/manual/[jq^], my favourite tool for parsing JSON files. Every LUSID user gets their own tenant/subdomain that has its own Swagger file. We'll download my one by running the following command: [source, bash] ---- wget https://markn.lusid.com/api/swagger/v0/swagger.json ---- This file has lots of stuff that we're not interested in, so let's clean it up. We'll remove the parts of the file that we aren't interested in and write the rest to `swagger-clean.json`: [source,bash] ---- cat swagger.json | \ jq 'del(.paths[] [] | .requestBody,.responses,.parameters,.security,.tags) | del (.info,.servers)' \ > swagger-clean.json ---- And now let's have a look at what we've got left: [source,bash] ---- cat swagger-clean.json | jq -r '.paths' ---- .Results [source, json] ---- { "/api/aggregation/$valuation": { "post": { "summary": "[BETA] Perform valuation for a list of portfolios and/or portfolio groups", "description": "Perform valuation on specified list of portfolio and/or portfolio groups for a set of dates.", "operationId": "GetValuation", "x-fbn-apistatus": "Beta" } }, "/api/allocations": { "get": { "summary": "[EXPERIMENTAL] List Allocations", "description": "Fetch the last pre-AsAt date version of each allocation in scope (does not fetch the entire history).", "operationId": "ListAllocations", "x-fbn-apistatus": "Experimental" }, "post": { "summary": "[EXPERIMENTAL] Upsert Allocations", "description": "Upsert; update existing allocations with given ids, or create new allocations otherwise.", "operationId": "UpsertAllocations", "x-fbn-apistatus": "Experimental" } }, "/api/allocations/{scope}/{code}": { "get": { "summary": "[EXPERIMENTAL] Get Allocation", "description": "Fetch an Allocation matching the provided identifier", "operationId": "GetAllocation", "x-fbn-apistatus": "Experimental" }, "delete": { "summary": "[EXPERIMENTAL] Delete allocation", "description": "Delete an allocation. Deletion will be valid from the allocation's creation datetime.\r\nThis means that the allocation will no longer exist at any effective datetime from the asAt datetime of deletion.", "operationId": "DeleteAllocation", "x-fbn-apistatus": "Experimental" } }, ... } ---- It's a nested structure, keyed by endpoint URIs, which can then support multiple HTTP verbs. So given this subset of the file, we want to extract the URI, `summary`, and `operationId` for each endpoint. The first slightly tricky thing is getting hold of the URIs. If we use the `paths[]` selector to get an array of paths, we lose the URI, as shown below: [source, bash] ---- cat swagger-clean.json | jq -r '.paths[]' ---- .Results [source, json] ---- { "post": { "summary": "[BETA] Perform valuation for a list of portfolios and/or portfolio groups", "description": "Perform valuation on specified list of portfolio and/or portfolio groups for a set of dates.", "operationId": "GetValuation", "x-fbn-apistatus": "Beta" } } { "get": { "summary": "[EXPERIMENTAL] List Allocations", "description": "Fetch the last pre-AsAt date version of each allocation in scope (does not fetch the entire history).", "operationId": "ListAllocations", "x-fbn-apistatus": "Experimental" }, "post": { "summary": "[EXPERIMENTAL] Upsert Allocations", "description": "Upsert; update existing allocations with given ids, or create new allocations otherwise.", "operationId": "UpsertAllocations", "x-fbn-apistatus": "Experimental" } } { "get": { "summary": "[EXPERIMENTAL] Get Allocation", "description": "Fetch an Allocation matching the provided identifier", "operationId": "GetAllocation", "x-fbn-apistatus": "Experimental" }, "delete": { "summary": "[EXPERIMENTAL] Delete allocation", "description": "Delete an allocation. Deletion will be valid from the allocation's creation datetime.\r\nThis means that the allocation will no longer exist at any effective datetime from the asAt datetime of deletion.", "operationId": "DeleteAllocation", "x-fbn-apistatus": "Experimental" } } ---- Luckily we can access the URI using the `keys` function in combination with the `.paths` selector (instead of `.paths[]`): [source, bash] ---- cat swagger-clean.json | jq -r '.paths | keys[]' ---- .Results [source, json] ---- "/api/aggregation/$valuation", "/api/allocations", "/api/allocations/{scope}/{code}", ---- We can then alias the keys and drill down into each one. So if we want to extract the `operationId` and `summary` for `post` requests, we can do it like this: [source, bash] ---- cat swagger-clean.json | jq -r '.paths | keys[] as $k | [$k, (.[$k] .post.operationId), (.[$k] .post.summary)]' ---- .Results [source, json] ---- [ "/api/aggregation/$valuation", "GetValuation", "[BETA] Perform valuation for a list of portfolios and/or portfolio groups" ] [ "/api/allocations", "UpsertAllocations", "[EXPERIMENTAL] Upsert Allocations" ] [ "/api/calendars/businessday/{scope}/{code}", null, null ] ---- That works to some extent, but it doesn't work if the key is `get` or `delete`. So it turns out that we want to get all of the keys, which we can do using the `keys` function again! This leaves us with the following nested query: [source, bash] ---- cat swagger-clean.json | jq -r '.paths | keys[] as $k | [ (.[$k] | keys[] as $k1 | [$k, $k1, .[$k1].operationId, .[$k1].summary] ) ]' ---- .Results [source, json] ---- [ [ "/api/aggregation/$valuation", "post", "GetValuation", "[BETA] Perform valuation for a list of portfolios and/or portfolio groups" ] ] [ [ "/api/allocations", "get", "ListAllocations", "[EXPERIMENTAL] List Allocations" ], [ "/api/allocations", "post", "UpsertAllocations", "[EXPERIMENTAL] Upsert Allocations" ] ] [ [ "/api/allocations/{scope}/{code}", "delete", "DeleteAllocation", "[EXPERIMENTAL] Delete allocation" ], [ "/api/allocations/{scope}/{code}", "get", "GetAllocation", "[EXPERIMENTAL] Get Allocation" ] ] ---- And then to flatten it out into single arrays instead of nested ones, we can pipe the result through the `.[]` selector: [source, bash] ---- cat swagger-clean.json | jq -r '.paths | keys[] as $k | [ (.[$k] | keys[] as $k1 | [$k, $k1, .[$k1].operationId, .[$k1].summary] ) ] | .[]' ---- .Results [source, json] ---- [ "/api/aggregation/$valuation", "post", "GetValuation", "[BETA] Perform valuation for a list of portfolios and/or portfolio groups" ] [ "/api/allocations", "get", "ListAllocations", "[EXPERIMENTAL] List Allocations" ] [ "/api/allocations", "post", "UpsertAllocations", "[EXPERIMENTAL] Upsert Allocations" ] [ "/api/allocations/{scope}/{code}", "delete", "DeleteAllocation", "[EXPERIMENTAL] Delete allocation" ] [ "/api/allocations/{scope}/{code}", "get", "GetAllocation", "[EXPERIMENTAL] Get Allocation" ] ---- And then if we want to go one step further, we could even convert that all into a CSV file using the `@csv` operator: [source, bash] ---- cat swagger-clean.json | jq -r '.paths | keys[] as $k | [ (.[$k] | keys[] as $k1 | [$k, $k1, .[$k1].operationId, .[$k1].summary] ) ] | .[] | @csv' ---- .Results [format="csv", cols="45,5,20,30"] |=== "/api/aggregation/$valuation","post","GetValuation","[BETA] Perform valuation for a list of portfolios and/or portfolio groups" "/api/aggregation/$valuationinlined","post","GetValuationOfWeightedInstruments","[BETA] Perform valuation for an inlined portfolio" "/api/aggregation/{scope}/{code}/$generateconfigurationrecipe","post","GenerateConfigurationRecipe","[EXPERIMENTAL] Generates a recipe sufficient to perform valuations for the given portfolio." "/api/allocations","get","ListAllocations","[EXPERIMENTAL] List Allocations" "/api/allocations","post","UpsertAllocations","[EXPERIMENTAL] Upsert Allocations" "/api/allocations/{scope}/{code}","delete","DeleteAllocation","[EXPERIMENTAL] Delete allocation" "/api/allocations/{scope}/{code}","get","GetAllocation","[EXPERIMENTAL] Get Allocation" "/api/calendars/businessday/{scope}/{code}","get","IsBusinessDateTime","[EXPERIMENTAL] Check whether a DateTime is a ""Business DateTime""" "/api/calendars/generic","get","ListCalendars","[EXPERIMENTAL] List Calenders" "/api/calendars/generic","post","CreateCalendar","[EXPERIMENTAL] Create a calendar in its generic form" |=== Job done and jq to the rescue again!

In this post we'll learn how to extract multiple keys from a Swagger JSON file using the jq JSON processing library.

jq-select-multiple-keys

https://markhneedham.com/blog/2021/05/19/jq-select-multiple-keys

2021-05-13 00:44:37

Pandas: Add row to DataFrame

Usually when I'm working with Pandas DataFrames I want to add new columns of data, but I recently wanted to add a row to an existing DataFrame. It turns out there are more than one ways to do that, which we'll explore in this blog post. Let's start by importing Pandas into our Python script: [source, bash] ---- import pandas as pd ---- We'll start from a DataFrame that has two rows and the columns `name` and `age`: [source, python] ---- df = pd.DataFrame(data=[{"name": "Mark", "age": 37}, {"name": "David", "age": 36}]) ---- .DataFrame with no explicit index [format="csv", options="header"] |=== include::content/2021/05/13/data/no_index.csv[] |=== One way to add rows to this DataFrame is by creating a new DataFrame and joining it with our initial DataFrame using the `append` function: [source, python] ---- to_append = pd.DataFrame([ {"name": "Ryan", "age": 42}, {"name": "John", "age": 25} ]) append_df = df.append(to_append) ---- .DataFrame with new row [format="csv", options="header"] |=== include::content/2021/05/13/data/no_index_append.csv[] |=== The index for the new rows start again from 0, so we now have multiple rows with the index 0 and 1. Another way that we can append a new row is using the `loc` function: [source, python] ---- df.loc[2] = ["Ryan", 42] df.loc[3] = ["John", 25] ---- If we take this approach we need to explicitly specify the index of the row. We can put whatever value we want, but let's create indexes that increment the existing values: .DataFrame with new row [format="csv", options="header"] |=== include::content/2021/05/13/data/no_index_individual.csv[] |=== What if we have an explicit index set on the DataFrame? We can convert the `name` column into an index by running the following code: [source, python] ---- df_with_index = df.set_index("name") ---- .DataFrame with `name` index [format="csv", options="header"] |=== include::content/2021/05/13/data/index.csv[] |=== And now if we want to add a new row the index identifier should be a name instead of a numeric value: [source, python] ---- df_with_index.loc["Ryan"] = [42] df_with_index.loc["John"] = [25] ---- .DataFrame with `name` index with new row [format="csv", options="header"] |=== include::content/2021/05/13/data/index_with_row.csv[] |===

In this post we'll add a new row to a Pandas DataFrame.

pandas-add-row-to-dataframe-with-index

https://markhneedham.com/blog/2021/05/13/pandas-add-row-to-dataframe-with-index

2021-02-03 00:44:37

Neo4j Graph Data Science 1.5: Exploring the HITS Algorithm

The https://neo4j.com/product/graph-data-science-library/[Neo4j Graph Data Science Library^] provides efficiently implemented, parallel versions of common graph algorithms for Neo4j, exposed as Cypher procedures. It recently published https://github.com/neo4j/graph-data-science/releases/tag/1.5.0[version 1.5^], which has lots of goodies to play with. image::{{<siteurl>}}/uploads/2021/02/gds-v1.5-hits.png[] In this blog post, we're going to explore the newly added https://neo4j.com/docs/graph-data-science/1.5/algorithms/hits/[HITS algorithm^] with the help of a citations dataset. == Launching Neo4j We're going to run Neo4j with the Graph Data Science Library using the following Docker Compose configuration: .docker-compose.yml [source,yaml] ---- version: '3.7' services: neo4j: image: neo4j:4.2.3-enterprise container_name: "neo4j4.2-gds1.5-exploration" volumes: - ./plugins-4.2:/plugins - ./data-4.2:/data ports: - "7474:7474" - "7687:7687" environment: - NEO4J_ACCEPT_LICENSE_AGREEMENT=yes - NEO4J_AUTH=neo4j/neo - NEO4JLABS_PLUGINS=["apoc", "graph-data-science"] ---- If you want to follow along with the examples used in the blog post, you can copy the configuration above to a file titled `docker-compose.yml` We can now launch the Neo4j server by running the following command: [source, bash] ---- docker-compose up ---- After we've launched that command, we need to wait until we see the following output: .Output [source,text] ---- neo4j4.2-gds1.5-exploration | 2021-02-03 21:39:15.346+0000 INFO Bolt enabled on 0.0.0.0:7687. neo4j4.2-gds1.5-exploration | 2021-02-03 21:39:16.053+0000 INFO Remote interface available at http://localhost:7474/ neo4j4.2-gds1.5-exploration | 2021-02-03 21:39:16.053+0000 INFO Started. ---- == Importing the Citation Graph We're going to import the Citations Graph from the https://neo4j.com/graphacademy/training-gdsds-40/enrollment/[Using a Machine Learning Workflow for Link Prediction^] online training. Let's first connect to Neo4j using the Cypher Shell: [source,bash] ---- docker exec -it neo4j4.2-gds1.5-exploration bin/cypher-shell -u neo4j -p neo ---- .Output [source,text] ---- Connected to Neo4j using Bolt protocol version 4.2 at neo4j://localhost:7687 as user neo4j. Type :help for a list of available commands or :exit to exit the shell. Note that Cypher queries must end with a semicolon. neo4j@neo4j> ---- The data for the citation graph is available as a set of JSON lines files, which we can find at https://github.com/mneedham/link-prediction/tree/master/data. Below is an example of one line of one of the files: [source,json] ---- {"authors": ["Tegegne Marew", "Doo-Hwan Bae"], "n_citation": 1, "references": ["2134bf3b-fd89-4724-90ce-5993b4fa3218", "906c17e0-db09-407b-b760-41df5a3f0293", "94f4382e-cfa6-4aec-92b8-3711fc55da54", "9f172585-8d42-4fce-b6ae-aede321f3fd4", "a3aee287-efd0-4b9d-9cda-d47dd192c9f4", "a9a7fd07-ef71-4b3c-8fcf-d7fe114d2148", "d63dd4ae-4b30-484b-8ffc-88d21839ddad"], "title": "Using Classpects for Integrating Non-Functional and Functional Requirements.", "venue": "international conference on software engineering", "year": 2006, "id": "01f1d231-80ae-4cce-b56c-9d821e0924d0"} ---- We're going to use some https://neo4j.com/labs/apoc/4.2/overview/[APOC procedures^] to convert these JSON files into the following graph structure: .Graph Model image::https://neo4j.com/graphacademy/training-gdsds-40/_images/graph.png[] So we'll have the following node labels: * `Article` - a paper on a topic * `Venue` - where the paper was presented * `Author` - the person/people that worked on the `Article` And these relationship types: * `CITED` - which articles were cited by an article * `AUTHOR` - which articles were written by a person/people * `VENUE` - where an article was presented Let's first create a new database and setup some constraints to make sure we don't end up with duplicates: .Create new database and setup constraints [source,cypher] ---- CREATE OR REPLACE DATABASE citationsblogpost; :use citationsblogpost; CREATE CONSTRAINT ON (a:Article) ASSERT a.index IS UNIQUE; CREATE CONSTRAINT ON (a:Author) ASSERT a.name IS UNIQUE; CREATE CONSTRAINT ON (v:Venue) ASSERT v.name IS UNIQUE; ---- Now we'll import the data: .Create citations graph [source,cypher] ---- CALL apoc.periodic.iterate( 'UNWIND ["dblp-ref-0.json", "dblp-ref-1.json", "dblp-ref-2.json", "dblp-ref-3.json"] AS file CALL apoc.load.json("https://github.com/mneedham/link-prediction/raw/master/data/" + file) YIELD value WITH value return value', 'MERGE (a:Article {index:value.id}) SET a += apoc.map.clean(value,["id","authors","references", "venue"],[0]) WITH a, value.authors as authors, value.references AS citations, value.venue AS venue MERGE (v:Venue {name: venue}) MERGE (a)-[:VENUE]->(v) FOREACH(author in authors | MERGE (b:Author{name:author}) MERGE (a)-[:AUTHOR]->(b)) FOREACH(citation in citations | MERGE (cited:Article {index:citation}) MERGE (a)-[:CITED]->(cited))', {batchSize: 1000, iterateList: true}) YIELD batches, total, timeTaken, committedOperations; ---- .Results [opts="header"] |=== | batches | total | timeTaken | committedOperations | 52 | 51956 | 21 | 51956 |=== And finally, a bit of cleanup to remove articles that don't have a title: .Remove articles that don't have a title [source,cypher] ---- MATCH (a:Article) WHERE not(exists(a.title)) DETACH DELETE a; ---- == HITS Algorithm The https://neo4j.com/docs/graph-data-science/1.5/algorithms/hits/[HITs algorithm^], like many other graph algorithms, was invented to do link analysis on web pages. It is a centrality algorithm, which means that it indicates node importance based on some metric. We can learn more about it from the https://en.wikipedia.org/wiki/HITS_algorithm[HITS Wikipedia page^]: [quote] _____ The idea behind Hubs and Authorities stemmed from a particular insight into the creation of web pages when the Internet was originally forming; that is, certain web pages, known as hubs, served as large directories that were not actually authoritative in the information that they held, but were used as compilations of a broad catalog of information that led users direct to other authoritative pages. The scheme therefore assigns two scores for each page: its authority, which estimates the value of the content of the page, and its hub value, which estimates the value of its links to other pages. _____ So a page with a high authority score has high value content, whereas a page with a high hub score links out to important pages. We're going to use this algorithm to analyse the citations between articles in our graph, so what does those different scores mean for us? * An article with a high authority score will likely have a lot of citations, perhaps some of those by other important articles * An article with a high hub score can help direct us (via its citations) to the important articles. It's not clear to me that the hub score makes so much sense in this graph because there aren't really articles written with the intention of pointing people towards a bunch of other articles! Let's give the algorithm a try and see what we find. We can return a list of the available procedures by running the following query: .List the HITS procedures [source,cypher] ---- CALL gds.list("hits") YIELD name, description RETURN name, description; ---- .Results [opts="header", cols="1,3"] |=== | name | description | "gds.alpha.hits.mutate" | "Hyperlink-Induced Topic Search (HITS) is a link analysis algorithm that rates nodes" | "gds.alpha.hits.mutate.estimate" | "Returns an estimation of the memory consumption for that procedure." | "gds.alpha.hits.stats" | "Hyperlink-Induced Topic Search (HITS) is a link analysis algorithm that rates nodes" | "gds.alpha.hits.stats.estimate" | "Returns an estimation of the memory consumption for that procedure." | "gds.alpha.hits.stream" | "Hyperlink-Induced Topic Search (HITS) is a link analysis algorithm that rates nodes" | "gds.alpha.hits.stream.estimate" | "Returns an estimation of the memory consumption for that procedure." | "gds.alpha.hits.write" | "Hyperlink-Induced Topic Search (HITS) is a link analysis algorithm that rates nodes" | "gds.alpha.hits.write.estimate" | "Returns an estimation of the memory consumption for that procedure." |=== Before we run the algorithm, we'll create a projected graph called `citation_graph`, by running the following: .Create projected graph [source,cypher] ---- CALL gds.graph.create("citation_graph", "Article", "CITED"); ---- .Results [opts="header"] |=== | nodeProjection | relationshipProjection | graphName | nodeCount | relationshipCount | createMillis | {Article: {properties: {}, label: "Article"}} | {CITED: {orientation: "NATURAL", aggregation: "DEFAULT", type: "CITED", properties: {}}} | "citation_graph" | 51956 | 28706 | 149 |=== And now we'll run the write version of the algorithm against the projected graph: .Run HITS algorithm [source,cypher] ---- CALL gds.alpha.hits.write("citation_graph", { hitsIterations: 20 }) YIELD writeMillis, nodePropertiesWritten, ranIterations, postProcessingMillis, createMillis, computeMillis; ---- .Results [opts="header"] |=== | writeMillis | nodePropertiesWritten | ranIterations | postProcessingMillis | createMillis | computeMillis | 174 | 103912 | 81 | 0 | 3 | 390 |=== By default, this procedure will create `pregel_auth` and `pregel_hub` properties on each of the `Article` nodes storing the computed scores. == Analysing authority scores Let's see which articles rank highest, starting with authority: [source,cypher] ---- MATCH (a:Article) RETURN a.title, a.year, substring(a.abstract, 0, 300) AS abstract, [(a)-[:AUTHOR]->(auth) | auth.name] AS authors, round(a.pregel_auth, 3) AS auth ORDER BY auth DESC LIMIT 10; ---- .Results [opts="header", cols="30,10,25,25,10"] |=== | a.title | a.year | abstract | authors | auth | "Rough sets" | 1995 | "Rough set theory, introduced by Zdzislaw Pawlak in the early 1980s [11, 12], is a new mathematical tool to deal with vagueness and uncertainty. This approach seems to be of fundamental importance to artificial intelligence (AI) and cognitive sciences, especially in the areas of machine learning, kno" | ["Jerzy W. Grzymala-Busse", "Wojciech Ziarko", "Zdzisław Pawlak", "Roman Słowiński"] | 0.99 | "Fuzzy Similarity Relation as a Basis for Rough Approximations" | 1998 | "The rough sets theory proposed by Pawlak was originally founded on the idea of approximating a given set by means of indiscernibility binary relation, which was assumed to be an equivalence relation (reflexive, symmetric and transitive). With respect to this basic idea, two main theoretical developm" | ["Roman Słowiński", "Salvatore Greco", "Benedetto Matarazzo"] | 0.042 | "Toward Intelligent Systems: Calculi of Information Granules" | 2001 | "We present an approach based on calculi of information granules as a basis for approximate reasoning in intelligent systems. Approximate reasoning schemes are defined by means of information granule construction schemes satisfying some robustness constraints. In distributed environments such schemes" | ["Andrzej Skowron"] | 0.042 | "Approximation spaces and information granulation" | 2005 | "In this paper, we discuss approximation spaces in a granular computing framework. Such approximation spaces generalise the approaches to concept approximation existing in rough set theory. Approximation spaces are constructed as higher level information granules and are obtained as the result of com" | ["Andrzej Skowron", "Piotr Synak", "Roman Świniarski"] | 0.038 | "Layered learning for concept synthesis" | 2004 | "We present a hierarchical scheme for synthesis of concept approximations based on given data and domain knowledge. We also propose a solution, founded on rough set theory, to the problem of con- structing the approximation of higher level concepts by composing the approximation of lower level concep" | ["Andrzej Skowron", "Jan G. Bazan", "Hung Son Nguyen", "Sinh Hoa Nguyen"] | 0.037 | "A Comparison of Several Approaches to Missing Attribute Values in Data Mining" | 2000 | "In the paper nine different approaches to missing attribute values are presented and compared. Ten input data files were used to investigate the performance of the nine methods to deal with missing attribute values. For testing both naive classification and new classification techniques of LERS (Lea" | ["Jerzy W. Grzymala-Busse", "Ming Hu"] | 0.036 | "Variable Consistency Model of Dominance-Based Rough Sets Approach" | 2000 | "Consideration of preference-orders requires the use of an extended rough set model called Dominance-based Rough Set Approach (DRSA). The rough approximations defined within DRSA are based on consistency in the sense of dominance principle. It requires that objects having not-worse evaluation with re" | ["Benedetto Matarazzo", "Salvatore Greco", "Roman Słowiński", "Jerzy Stefanowski"] | 0.029 | "RSES and RSESlib - A Collection of Tools for Rough Set Computations" | 2000 | "Rough Set Exploration System - a set of software tools featuring a library of methods and a graphical user interface is presented. Methods, features and abilities of the implemented software are discussed and illustrated with a case study in data analysis." | ["Marcin S. Szczuka", "Jan G. Bazan"] | 0.026 | "A New Version of Rough Set Exploration System" | 2002 | "We introduce a new version of the Rough Set Exploration System - a software tool featuring a library of methods and a graphical user interface supporting variety of rough-set-based computations. Methods, features and abilities of the implemented software are discussed and illustrated with a case stu" | ["Marcin S. Szczuka", "Jakub Wróblewski", "Jan G. Bazan"] | 0.026 | "Rough sets and information granulation" | 2003 | "In this paper, the study of the evolution of approximation space theory and its applications is considered in the context of rough sets introduced by Zdzislaw Pawlak and information granulation as well as computing with words formulated by Lotfi Zadeh. Central to this evolution is the rough-mereolog" | ["Piotr Synak", "James F. Peters", "Andrzej Skowron", "Sheela Ramanna"] | 0.026 |=== The top article by some distance on this metric is https://dl.acm.org/doi/10.1145/219717.219791["Rough sets"^], which was written more than 25 years ago. I found it interesting that the abstract talks about it being an approach that is fundamental to AI and machine learning, which are important fields in 2021. We can have a look at the hub nodes that point to these articles by running the following query: [source,cypher] ---- MATCH (a:Article) WITH a, [(a)<-[:CITED]-(other) | other] AS citations WITH a, apoc.coll.sortNodes(citations, "pregel_hub")[..5] AS topHubs RETURN a.title, a.year, round(a.pregel_auth, 3) AS auth, [c in topHubs | {article: c.title, score: round(c.pregel_hub, 3)}] AS topHubs ORDER BY auth DESC LIMIT 10; ---- .Results [opts="header", cols="25,10,10,55"] |=== | a.title | a.year | auth | topHubs | "Rough sets" | 1995 | 0.99 | [{score: 0.083, article: "Rough ethology: towards a biologically-inspired study of collective behavior in intelligent systems with approximation spaces"}, {score: 0.082, article: "Some Issues on Rough Sets"}, {score: 0.079, article: "A treatise on rough sets"}, {score: 0.079, article: "Approximate boolean reasoning: foundations and applications in data mining"}, {score: 0.075, article: "Multimodal classification: case studies"}] | "Fuzzy Similarity Relation as a Basis for Rough Approximations" | 1998 | 0.042 | [{score: 0.082, article: "Some Issues on Rough Sets"}, {score: 0.079, article: "A treatise on rough sets"}, {score: 0.079, article: "Approximate boolean reasoning: foundations and applications in data mining"}, {score: 0.075, article: "On generalized rough fuzzy approximation operators"}, {score: 0.074, article: "Lattices with Interior and Closure Operators and Abstract Approximation Spaces"}] | "Toward Intelligent Systems: Calculi of Information Granules" | 2001 | 0.042 | [{score: 0.083, article: "Rough ethology: towards a biologically-inspired study of collective behavior in intelligent systems with approximation spaces"}, {score: 0.082, article: "Some Issues on Rough Sets"}, {score: 0.072, article: "Rough sets and information granulation"}, {score: 0.071, article: "A Note on Ziarko's Variable Precision Rough Set Model and Nonmonotonic Reasoning"}, {score: 0.071, article: "A Partition Model of Granular Computing"}] | "Approximation spaces and information granulation" | 2005 | 0.038 | [{score: 0.083, article: "Rough ethology: towards a biologically-inspired study of collective behavior in intelligent systems with approximation spaces"}, {score: 0.082, article: "Some Issues on Rough Sets"}, {score: 0.079, article: "A treatise on rough sets"}, {score: 0.075, article: "On generalized rough fuzzy approximation operators"}, {score: 0.074, article: "Matching 2d image segments with genetic algorithms and approximation spaces"}] | "Layered learning for concept synthesis" | 2004 | 0.037 | [{score: 0.083, article: "Rough ethology: towards a biologically-inspired study of collective behavior in intelligent systems with approximation spaces"}, {score: 0.079, article: "A treatise on rough sets"}, {score: 0.079, article: "Approximate boolean reasoning: foundations and applications in data mining"}, {score: 0.075, article: "Multimodal classification: case studies"}, {score: 0.072, article: "P300 wave detection based on rough sets"}] | "A Comparison of Several Approaches to Missing Attribute Values in Data Mining" | 2000 | 0.036 | [{score: 0.082, article: "Some Issues on Rough Sets"}, {score: 0.075, article: "The rough set exploration system"}, {score: 0.071, article: "Missing template decomposition method and its implementation in rough set exploration system"}, {score: 0.07, article: "Data with Missing Attribute Values: Generalization of Indiscernibility Relation and Rule Induction"}, {score: 0.07, article: "Characteristic relations for incomplete data: a generalization of the indiscernibility relation"}] | "Variable Consistency Model of Dominance-Based Rough Sets Approach" | 2000 | 0.029 | [{score: 0.072, article: "Rough Set Analysis of Preference-Ordered Data"}, {score: 0.072, article: "Variable-precision dominance-based rough set approach"}, {score: 0.071, article: "On variable consistency dominance-based rough set approaches"}, {score: 0.071, article: "Multicriteria choice and ranking using decision rules induced from rough approximation of graded preference relations"}, {score: 0.07, article: "Rough set approach to customer satisfaction analysis"}] | "RSES and RSESlib - A Collection of Tools for Rough Set Computations" | 2000 | 0.026 | [{score: 0.079, article: "Approximate boolean reasoning: foundations and applications in data mining"}, {score: 0.073, article: "Hybridization of rough sets and statistical learning theory"}, {score: 0.072, article: "Ontology driven concept approximation"}, {score: 0.072, article: "Processing of musical data employing rough sets and artificial neural networks"}, {score: 0.069, article: "A statistical method for determining importance of variables in an information system"}] | "A New Version of Rough Set Exploration System" | 2002 | 0.026 | [{score: 0.075, article: "Multimodal classification: case studies"}, {score: 0.072, article: "Processing of musical data employing rough sets and artificial neural networks"}, {score: 0.069, article: "Introducing a rule importance measure"}, {score: 0.069, article: "NetTRS induction and postprocessing of decision rules"}, {score: 0.069, article: "Classification of Swallowing Sound Signals: A Rough Set Approach"}] | "Rough sets and information granulation" | 2003 | 0.026 | [{score: 0.083, article: "Rough ethology: towards a biologically-inspired study of collective behavior in intelligent systems with approximation spaces"}, {score: 0.079, article: "A treatise on rough sets"}, {score: 0.075, article: "On generalized rough fuzzy approximation operators"}, {score: 0.074, article: "Matching 2d image segments with genetic algorithms and approximation spaces"}, {score: 0.071, article: "Time complexity of decision trees"}] |=== Based on the top hubs, it's not really obvious why the authority score for "Rough sets" is so much higher than the other articles. Perhaps if we return the max, min, and average hub scores we'll be able to figure it out? [source,cypher] ---- MATCH (a:Article) WITH a, [(a)<-[:CITED]-(other) | other] AS citations RETURN a.title, a.year, round(a.pregel_auth, 3) AS auth, round(apoc.coll.max([c in citations | c.pregel_hub]), 3) AS maxHub, round(apoc.coll.min([c in citations | c.pregel_hub]), 3) AS minHub, round(apoc.coll.avg([c in citations | c.pregel_hub]), 3) AS averageHub, size(citations) AS citations ORDER BY auth DESC LIMIT 10; ---- .Results [opts="header", cols="4,1,1,1,1,1,1"] |=== | a.title | a.year | auth | maxHub | minHub | averageHub | citations | "Rough sets" | 1995 | 0.99 | 0.083 | 0.068 | 0.069 | 211 | "Toward Intelligent Systems: Calculi of Information Granules" | 2001 | 0.042 | 0.083 | 0.003 | 0.036 | 17 | "Fuzzy Similarity Relation as a Basis for Rough Approximations" | 1998 | 0.042 | 0.082 | 0.003 | 0.061 | 10 | "Approximation spaces and information granulation" | 2005 | 0.038 | 0.083 | 0.005 | 0.055 | 10 | "Layered learning for concept synthesis" | 2004 | 0.037 | 0.083 | 0.003 | 0.05 | 11 | "A Comparison of Several Approaches to Missing Attribute Values in Data Mining" | 2000 | 0.036 | 0.082 | 0.002 | 0.048 | 11 | "Variable Consistency Model of Dominance-Based Rough Sets Approach" | 2000 | 0.029 | 0.072 | 0.004 | 0.061 | 7 | "RSES and RSESlib - A Collection of Tools for Rough Set Computations" | 2000 | 0.026 | 0.079 | 0.002 | 0.032 | 12 | "Rough sets and information granulation" | 2003 | 0.026 | 0.083 | 0.005 | 0.065 | 6 | "A New Version of Rough Set Exploration System" | 2002 | 0.026 | 0.075 | 0.002 | 0.038 | 10 |=== From this output we learn that "Rough sets" is being cited by a lot of articles with a good hub score. The other articles have a similar `maxHub` score and some even have a similar `averageHub`, but their `minHub` is significantly less. It also has 10x as many citations as any of the other articles in the top 10, so that would contribute to the higher score as well. == HITS Authority vs PageRank The HITS Authority score and the PageRank algorithm both compute scores that indicate the importance of a node in a graph, so I was curious whether there was any correlation between the scores. i.e. do the nodes with the highest HITS authority score also have a high PageRank score? To recap, https://neo4j.com/docs/graph-data-science/current/algorithms/page-rank/[this is what PageRank measures^]: [quote] _____ The PageRank algorithm measures the importance of each node within the graph, based on the number incoming relationships and the importance of the corresponding source nodes. _____ We can compute the PageRank score for articles, by running the following query: [source,cypher] ---- CALL gds.pageRank.write("citation_graph", { maxIterations: 20, writeProperty: "pagerank" }) YIELD writeMillis, nodePropertiesWritten, ranIterations, postProcessingMillis, createMillis, computeMillis; ---- .Results [opts="header"] |=== | writeMillis | nodePropertiesWritten | ranIterations | postProcessingMillis | createMillis | computeMillis | 29 | 51956 | 20 | 0 | 0 | 112 |=== And now let's put the PageRank scores alongside the HITS Authority scores: [source,cypher] ---- MATCH (a:Article) RETURN a.title, a.year, round(a.pregel_auth, 3) AS auth, round(a.pagerank, 3) AS pagerank, size([(a)<-[:CITED]-(other) | other]) AS citations ORDER BY auth DESC LIMIT 10; ---- .Results [opts="header",cols="60,10,10,10,10"] |=== | a.title | a.year | auth | pagerank | citations | "Rough sets" | 1995 | 0.99 | 25.609 | 211 | "Fuzzy Similarity Relation as a Basis for Rough Approximations" | 1998 | 0.042 | 0.738 | 10 | "Toward Intelligent Systems: Calculi of Information Granules" | 2001 | 0.042 | 1.862 | 17 | "Approximation spaces and information granulation" | 2005 | 0.038 | 0.418 | 10 | "Layered learning for concept synthesis" | 2004 | 0.037 | 0.505 | 11 | "A Comparison of Several Approaches to Missing Attribute Values in Data Mining" | 2000 | 0.036 | 0.896 | 11 | "Variable Consistency Model of Dominance-Based Rough Sets Approach" | 2000 | 0.029 | 0.471 | 7 | "RSES and RSESlib - A Collection of Tools for Rough Set Computations" | 2000 | 0.026 | 1.296 | 12 | "A New Version of Rough Set Exploration System" | 2002 | 0.026 | 0.682 | 10 | "Rough sets and information granulation" | 2003 | 0.026 | 0.375 | 6 |=== Rough Sets is the only one with a high PageRank score as well. In fact, its PageRank score is the 3rd highest in the graph, which we can see by running the following query: [source,cypher] ---- MATCH (a:Article) RETURN a.title, a.year, round(a.pregel_auth, 5) AS auth, round(a.pagerank, 5) AS pagerank, size([(a)<-[:CITED]-(other) | other]) AS citations ORDER BY pagerank DESC LIMIT 10; ---- .Results [opts="header",cols="60,10,10,10,10"] |=== | a.title | a.year | auth | pagerank | citations | "A method for obtaining digital signatures and public-key cryptosystems" | 1978 | 5.0E-5 | 93.94313 | 125 | "Secure communications over insecure channels" | 1978 | 0.0 | 79.86924 | 7 | "Rough sets" | 1995 | 0.9902 | 25.60911 | 211 | "An axiomatic basis for computer programming" | 1969 | 4.4E-4 | 23.02937 | 93 | "Pastry: Scalable, Decentralized Object Location, and Routing for Large-Scale Peer-to-Peer Systems" | 2001 | 0.0 | 21.46956 | 108 | "SCRIBE: The Design of a Large-Scale Event Notification Infrastructure" | 2001 | 0.0 | 19.4863 | 14 | "A field study of the software design process for large systems" | 1988 | 0.0 | 19.02815 | 53 | "Productivity factors and programming environments" | 1984 | 0.0 | 18.49935 | 5 | "Analyzing medium-scale software development" | 1978 | 0.0 | 16.45275 | 5 | "A Calculus of Communicating Systems" | 1982 | 0.0 | 15.43059 | 55 |=== I find it kinda interesting that while these articles have very high transitive importance, their HITS Authority score is very low. Many of them have a lot of citations as well, but presumably most of those citations aren't from hub nodes. == Analysing hub scores Speaking of hubs, let's explore those in a bit more detail. We can find the articles with the highest hub score, by running the following query: [source,cypher] ---- MATCH (a:Article) WITH a, [(a)-[:CITED]->(other) | other] AS cited RETURN a.title, a.year, round(a.pregel_hub, 3) AS hub, round(apoc.coll.max([c in cited | c.pregel_auth]), 3) AS maxAuth, round(apoc.coll.min([c in cited | c.pregel_auth]), 3) AS minAuth, round(apoc.coll.avg([c in cited | c.pregel_auth]), 3) AS averageAuth, size(cited) AS cited ORDER BY a.pregel_hub DESC LIMIT 10; ---- .Results [opts="header",cols="40,10,10,10,10,10,10"] |=== | a.title | a.year | hub | maxAuth | minAuth | averageAuth | cited | "Rough ethology: towards a biologically-inspired study of collective behavior in intelligent systems with approximation spaces" | 2005 | 0.083 | 0.99 | 0.006 | 0.102 | 12 | "Some Issues on Rough Sets" | 2004 | 0.082 | 0.99 | 0.006 | 0.134 | 9 | "A treatise on rough sets" | 2005 | 0.079 | 0.99 | 0.005 | 0.145 | 8 | "Approximate boolean reasoning: foundations and applications in data mining" | 2006 | 0.079 | 0.99 | 0.005 | 0.115 | 10 | "Multimodal classification: case studies" | 2006 | 0.075 | 0.99 | 0.005 | 0.122 | 9 | "The rough set exploration system" | 2005 | 0.075 | 0.99 | 0.005 | 0.157 | 7 | "On generalized rough fuzzy approximation operators" | 2006 | 0.075 | 0.99 | 0.026 | 0.274 | 4 | "Lattices with Interior and Closure Operators and Abstract Approximation Spaces" | 2009 | 0.074 | 0.99 | 0.005 | 0.136 | 8 | "Matching 2d image segments with genetic algorithms and approximation spaces" | 2006 | 0.074 | 0.99 | 0.005 | 0.154 | 7 | "Hybridization of rough sets and statistical learning theory" | 2011 | 0.073 | 0.99 | 0.005 | 0.214 | 5 |=== The `maxAuth` scores tell us that all of these articles cite the "Rough sets" article that we came across in the previous section. There aren't really any other articles with a high authority score, so we can assume that nearly all of the hub score is coming from citing "Rough sets". In any case, let's have a look at the other authorities that these articles have cited: [source,cypher] ---- MATCH (a:Article) WITH a, [(a)-[:CITED]->(other) | other] AS cited WITH a, apoc.coll.sortNodes(cited, "pregel_auth")[..5] AS topAuthorities RETURN a.title, a.year, round(a.pregel_hub, 3) AS hub, [c in topAuthorities | {article: c.title, score: round(c.pregel_auth, 3)}] AS topAuthorities ORDER BY hub DESC LIMIT 10; ---- .Results [opts="header", cols="30,10,10,50"] |=== | a.title | a.year | hub | topAuthorities | "Rough ethology: towards a biologically-inspired study of collective behavior in intelligent systems with approximation spaces" | 2005 | 0.083 | [{score: 0.99, article: "Rough sets"}, {score: 0.042, article: "Toward Intelligent Systems: Calculi of Information Granules"}, {score: 0.038, article: "Approximation spaces and information granulation"}, {score: 0.037, article: "Layered learning for concept synthesis"}, {score: 0.026, article: "Rough sets and information granulation"}] | "Some Issues on Rough Sets" | 2004 | 0.082 | [{score: 0.99, article: "Rough sets"}, {score: 0.042, article: "Toward Intelligent Systems: Calculi of Information Granules"}, {score: 0.042, article: "Fuzzy Similarity Relation as a Basis for Rough Approximations"}, {score: 0.038, article: "Approximation spaces and information granulation"}, {score: 0.036, article: "A Comparison of Several Approaches to Missing Attribute Values in Data Mining"}] | "A treatise on rough sets" | 2005 | 0.079 | [{score: 0.99, article: "Rough sets"}, {score: 0.042, article: "Fuzzy Similarity Relation as a Basis for Rough Approximations"}, {score: 0.038, article: "Approximation spaces and information granulation"}, {score: 0.037, article: "Layered learning for concept synthesis"}, {score: 0.026, article: "Rough sets and information granulation"}] | "Approximate boolean reasoning: foundations and applications in data mining" | 2006 | 0.079 | [{score: 0.99, article: "Rough sets"}, {score: 0.042, article: "Fuzzy Similarity Relation as a Basis for Rough Approximations"}, {score: 0.037, article: "Layered learning for concept synthesis"}, {score: 0.026, article: "RSES and RSESlib - A Collection of Tools for Rough Set Computations"}, {score: 0.021, article: "Some Issues on Rough Sets"}] | "Multimodal classification: case studies" | 2006 | 0.075 | [{score: 0.99, article: "Rough sets"}, {score: 0.037, article: "Layered learning for concept synthesis"}, {score: 0.026, article: "A New Version of Rough Set Exploration System"}, {score: 0.015, article: "The rough set exploration system"}, {score: 0.01, article: "Rough Set Methods in Approximation of Hierarchical Concepts"}] | "The rough set exploration system" | 2005 | 0.075 | [{score: 0.99, article: "Rough sets"}, {score: 0.036, article: "A Comparison of Several Approaches to Missing Attribute Values in Data Mining"}, {score: 0.021, article: "Rough Sets and Decision Algorithms"}, {score: 0.021, article: "In Pursuit of Patterns in Data Reasoning from Data The Rough Set Way"}, {score: 0.015, article: "Classification of Swallowing Sound Signals: A Rough Set Approach"}] | "On generalized rough fuzzy approximation operators" | 2006 | 0.075 | [{score: 0.99, article: "Rough sets"}, {score: 0.042, article: "Fuzzy Similarity Relation as a Basis for Rough Approximations"}, {score: 0.038, article: "Approximation spaces and information granulation"}, {score: 0.026, article: "Rough sets and information granulation"}] | "Lattices with Interior and Closure Operators and Abstract Approximation Spaces" | 2009 | 0.074 | [{score: 0.99, article: "Rough sets"}, {score: 0.042, article: "Fuzzy Similarity Relation as a Basis for Rough Approximations"}, {score: 0.024, article: "Approximation Operators in Qualitative Data Analysis"}, {score: 0.015, article: "Data with Missing Attribute Values: Generalization of Indiscernibility Relation and Rule Induction"}, {score: 0.005, article: "Algebraic structures for rough sets"}] | "Matching 2d image segments with genetic algorithms and approximation spaces" | 2006 | 0.074 | [{score: 0.99, article: "Rough sets"}, {score: 0.038, article: "Approximation spaces and information granulation"}, {score: 0.026, article: "Rough sets and information granulation"}, {score: 0.01, article: "K-means Indiscernibility Relation over Pixels"}, {score: 0.006, article: "Rough ethology: towards a biologically-inspired study of collective behavior in intelligent systems with approximation spaces"}] | "Hybridization of rough sets and statistical learning theory" | 2011 | 0.073 | [{score: 0.99, article: "Rough sets"}, {score: 0.038, article: "Approximation spaces and information granulation"}, {score: 0.026, article: "RSES and RSESlib - A Collection of Tools for Rough Set Computations"}, {score: 0.01, article: "Accuracy and Coverage in Rough Set Rule Induction"}, {score: 0.005, article: "Generalized indiscernibility relations: applications for missing values and analysis of structural objects"}] |=== The top 2 articles both cited "Toward Intelligent Systems: Calculi of Information Granules", which gives them a marginally higher score than the other 8. But I don't think these hub scores are telling us all that much about these articles. == In Summary While I'm not sure that this is the greatest data set to show off this algorithm, I think the algorithm itself is an interesting addition to the library. I'm curious to see how well it would fare on a Twitter graph - perhaps the HITS Hub score would help to identify those accounts that primarily tweet out links to interesting content? I guess that exploration will have to wait for another post!

In this post we'll explore the HITS link analysis algorithm that was added to the Neo4j Graph Data Science Library in version 1.5.0.

uploads/2021/02/gds-v1.5-hits.png

neo4j-gdsl-hits-algorithm

https://markhneedham.com/blog/2021/02/03/neo4j-gdsl-hits-algorithm

2021-02-08 00:44:37

Neo4j Graph Data Science 1.5: Exploring the Speaker-Listener LPA Overlapping Community Detection Algorithm

The https://neo4j.com/product/graph-data-science-library/[Neo4j Graph Data Science Library^] provides efficiently implemented, parallel versions of common graph algorithms for Neo4j, exposed as Cypher procedures. It recently published https://github.com/neo4j/graph-data-science/releases/tag/1.5.0[version 1.5^], which introduces some fun new algorithms. image::{{<siteurl>}}/uploads/2021/02/sllpa-banner.png[] In this blog post, we're going to explore the newly added https://neo4j.com/docs/graph-data-science/1.5/algorithms/sllpa/[Speaker-Listener Label Propagation algorithm^] with the help of a twitter dataset. == Launching Neo4j We're going to run Neo4j with the Graph Data Science Library using the following Docker Compose configuration: .docker-compose.yml [source,yaml] ---- version: '3.7' services: neo4j: image: neo4j:4.2.3-enterprise container_name: "neo4j4.2-gds1.5-exploration" volumes: - ./plugins-4.2:/plugins - ./data-4.2:/data ports: - "7474:7474" - "7687:7687" environment: - NEO4J_ACCEPT_LICENSE_AGREEMENT=yes - NEO4J_AUTH=neo4j/neo - NEO4JLABS_PLUGINS=["apoc", "graph-data-science"] ---- If you want to follow along with the examples used in the blog post, you can copy the configuration above to a file titled `docker-compose.yml` We can now launch the Neo4j server by running the following command: [source, bash] ---- docker-compose up ---- After we've launched that command, we need to wait until we see the following output: .Output [source,text] ---- neo4j4.2-gds1.5-exploration | 2021-02-03 21:39:15.346+0000 INFO Bolt enabled on 0.0.0.0:7687. neo4j4.2-gds1.5-exploration | 2021-02-03 21:39:16.053+0000 INFO Remote interface available at http://localhost:7474/ neo4j4.2-gds1.5-exploration | 2021-02-03 21:39:16.053+0000 INFO Started. ---- == Importing a Twitter Social Graph We're going to import the Twitter Social Graph that I used in the https://medium.com/neo4j/finding-influencers-and-communities-in-the-graph-community-e3d691296325[Finding influencers and communities in the Graph Community^] blog post that I wrote in 2019. Let's first connect to Neo4j using the Cypher Shell: [source,bash] ---- docker exec -it neo4j4.2-gds1.5-exploration bin/cypher-shell -u neo4j -p neo ---- .Output [source,text] ---- Connected to Neo4j using Bolt protocol version 4.2 at neo4j://localhost:7687 as user neo4j. Type :help for a list of available commands or :exit to exit the shell. Note that Cypher queries must end with a semicolon. neo4j@neo4j> ---- The data for the Twitter Social Graph graph is available as a set of JSON lines files, which we can find at https://github.com/neo4j-devtools/neuler/blob/master/sample-data/twitter/users.json. Below is an example of one line from this file: [source,json] ---- {"user":{"followers":2793,"following":1121,"name":"Synechron","bio":"Synechron is a leading Digital IT Consulting firm Accelerating Digital initiatives for banks, asset managers & insurance companies around the world.","id":625428620,"username":"Synechron"},"following":[],"followers":[50230435,117780256,856240505826496513,31112812,999543859398037504,873919885096693761,61085452,75043311,268191768]} ---- We're going to use some https://neo4j.com/labs/apoc/4.2/overview/[APOC procedures^] to convert these JSON files into the following graph structure: .Graph Model image::{{<siteurl>}}/uploads/2021/02/twitter-social-graph.svg[] Let's first create a new database and setup some constraints to make sure we don't end up with duplicates: .Create new database and setup constraints [source,cypher] ---- CREATE OR REPLACE DATABASE twitterblogpost; :use twitterblogpost; CREATE CONSTRAINT ON(u:User) ASSERT u.id IS unique; ---- Now we'll import the data: .Create citations graph [source,cypher] ---- WITH "https://github.com/neo4j-apps/neuler/raw/master/sample-data/twitter/users.json" AS url CALL apoc.load.json(url) YIELD value MERGE (u:User {id: value.user.id }) SET u += value.user WITH u, value CALL { WITH u, value UNWIND value.following AS following MERGE (f1:User {id: following}) MERGE (u)-[:FOLLOWS]->(f1) RETURN count(*) AS followingCount } CALL { WITH u, value UNWIND value.followers AS follower MERGE (f2:User {id: follower}) MERGE (u)<-[:FOLLOWS]-(f2) RETURN count(*) AS followersCount } RETURN count(*) AS users; ---- .Results [opts="header"] |=== | users | 6526 |=== We can see a Neo4j Browser visualisation of the import graph in the diagram below: .Twitter Social Graph image::{{<siteurl>}}/uploads/2021/02/twitter-graph-imported.svg[] == Speaker Listener Label Propagation Algorithm The https://neo4j.com/docs/graph-data-science/1.5/algorithms/sllpa/[Speaker Listener Label Propagation (SLLPA)^] algorithm is a variation of the Label Propagation Algorithm, where instead of returning one community per node, it can return multiple communities per node. We can alternatively say that the SLLPA algorithm detects overlapping communities. We can find out the name of the SLLPA procedures in the GDS library by running the following query: [source,cypher] ---- CALL gds.list("sllpa") YIELD name, description; ---- .Results [opts="header", cols="20,80"] |=== | name | description | "gds.alpha.sllpa.mutate" | "The Speaker Listener Label Propagation algorithm is a fast algorithm for finding overlapping communities in a graph." | "gds.alpha.sllpa.mutate.estimate" | "Returns an estimation of the memory consumption for that procedure." | "gds.alpha.sllpa.stats" | "The Speaker Listener Label Propagation algorithm is a fast algorithm for finding overlapping communities in a graph." | "gds.alpha.sllpa.stats.estimate" | "Returns an estimation of the memory consumption for that procedure." | "gds.alpha.sllpa.stream" | "The Speaker Listener Label Propagation algorithm is a fast algorithm for finding overlapping communities in a graph." | "gds.alpha.sllpa.stream.estimate" | "Returns an estimation of the memory consumption for that procedure." | "gds.alpha.sllpa.write" | "The Speaker Listener Label Propagation algorithm is a fast algorithm for finding overlapping communities in a graph." | "gds.alpha.sllpa.write.estimate" | "Returns an estimation of the memory consumption for that procedure." |=== We're going to use the `gds.alpha.sllpa.write` procedure. This algorithm runs the algorithm against an in-memory projected graph and stores the results as node properties. The in-memory graph that we're going to use will consist of `USER` nodes and the `FOLLOWS` relationships between them. We'll ignore the direction of the `FOLLOWS` relationship when computing communities. We can do this by running the following query: [source,cypher] ---- CALL gds.alpha.sllpa.write({ nodeProjection: "User", relationshipProjection: { FOLLOWS: { orientation: 'UNDIRECTED' } }, maxIterations: 20 }) YIELD nodePropertiesWritten, ranIterations, writeMillis, createMillis, computeMillis; ---- .Results [opts="header"] |=== | nodePropertiesWritten | ranIterations | writeMillis | createMillis | computeMillis | 6526 | 20 | 20 | 14 | 478 |=== By default, results are written to the `pregel_communityids` property. We can have a look at some of these values by running the following query: [source,cypher] ---- MATCH (user:User) RETURN user.username, user.pregel_communityIds LIMIT 5; ---- .Results [opts="header"] |=== | user.username | user.pregel_communityIds | "webmink" | [119] | "Synechron" | [119] | "flablog" | [185, 119] | "didierdaglinckx" | [3] | "DailyPythonInfo" | [119] |=== webmink, Synechron, didierdaglinckx, and DailyPythonInfo all belong to only one community, whereas flablog belongs to a couple of communities. == How many users belong to multiple communities? I wonder how many other users belong to multiple communities? We can find out by running the following query: [source,cypher] ---- MATCH (u:User) WITH count(*) AS totalCount MATCH (u:User) WITH totalCount, size(u.pregel_communityIds) as communities, count(*) AS count RETURN communities, count, round(count * 100.0 / totalCount, 2) AS percentage ORDER BY communities; ---- .Results [opts="header"] |=== | communities | count | percentage | 1 | 5703 | 87.39 | 2 | 801 | 12.27 | 3 | 22 | 0.34 |=== We can also use https://nielsdejong.nl/neodash/[Niels de Jong's NeoDash^] to create a quick bar chart of these results: .Community Size Frequency image::{{<siteurl>}}/uploads/2021/02/community-size-frequency.png[] A massive majority of nodes have only one community, just over 10% have two communities, and only 0.3% fit into three communities. == Create community graph Let's have a look at the communities in more detail. We're going to add the graph structure coloured in red in the diagram below: .Graph Model with communities image::{{<siteurl>}}/uploads/2021/02/twitter-social-graph-communities.svg[] We'll first create a unique constraint on the `id` property for the `Community` label: [source,cypher] ---- CREATE CONSTRAINT ON (c:Community) ASSERT c.id IS UNIQUE; ---- And now we can create a `Community` node for each community and an `IN_COMMUNITY` relationship from each user to the communities that they belong to: [source,cypher] ---- MATCH (u:User) UNWIND u.pregel_communityIds AS communityId MERGE (c:Community {id: communityId}) MERGE (u)-[:IN_COMMUNITY]->(c); ---- .Output [source,text] ---- 0 rows available after 299 ms, consumed after another 0 ms Added 1028 nodes, Created 7371 relationships, Set 1028 properties, Added 1028 labels ---- We can see a Neo4j Browser visualisation of the new graph structure in the diagram below: image::{{<siteurl>}}/uploads/2021/02/multi-communities.svg[] == Find important users I find that the best way to do community analysis is to look at the important nodes that belong to each one. We can compute the important nodes by running the PageRank algorithm, as shown below: [source,cypher] ---- CALL gds.pageRank.write({ nodeProjection: "User", relationshipProjection: "FOLLOWS", maxIterations: 20, writeProperty: "pagerank" }) YIELD writeMillis, nodePropertiesWritten, ranIterations, postProcessingMillis, createMillis, computeMillis; ---- .Results [opts="header"] |=== | writeMillis | nodePropertiesWritten | ranIterations | postProcessingMillis | createMillis | computeMillis | 35 | 6526 | 20 | 0 | 9 | 111 |=== Each of the `User` nodes now has a `pagerank` property. == Individual communities We can use this property as part of a query to find the top 5 users per community, as shown below: [source,cypher] ---- MATCH (:User) WITH count(*) AS allUsers MATCH (c:Community)<-[:IN_COMMUNITY]-(u:User) WITH allUsers, c, u ORDER BY c, u.pagerank DESC WITH allUsers, c, collect(u) AS users RETURN c.id, size(users) AS users, round(size(users)*100.0 / allUsers, 3) AS percentage, [u IN users | u {.username, score: round(u.pagerank, 3)}][..5] AS topUsers ORDER BY size(users) DESC LIMIT 10; ---- .Results [opts="header", cols="10,10,10,70"] |=== | c.id | users | percentage | topUsers | 119 | 3960 | 60.68 | [{score: 46.822, username: "TechCrunch"}, {score: 33.246, username: "awscloud"}, {score: 22.015, username: "hmason"}, {score: 20.332, username: "kellabyte"}, {score: 17.816, username: "KirkDBorne"}] | 157 | 1152 | 17.652 | [{score: 46.105, username: "neo4j"}, {score: 25.847, username: "emileifrem"}, {score: 23.085, username: "mesirii"}, {score: 15.83, username: "GraphConnect"}, {score: 14.879, username: "jimwebber"}] | 34 | 636 | 9.746 | [{score: 14.666, username: "jessitron"}, {score: 10.754, username: "WardCunningham"}, {score: 9.464, username: "springrod"}, {score: 8.98, username: "starbuxman"}, {score: 8.528, username: "garybernhardt"}] | 144 | 212 | 3.249 | [{score: 12.834, username: "danbri"}, {score: 6.759, username: "ontotext"}, {score: 6.198, username: "StardogHQ"}, {score: 6.113, username: "kidehen"}, {score: 5.223, username: "juansequeda"}] | 22 | 96 | 1.471 | [{score: 3.431, username: "_wald0"}, {score: 2.753, username: "Cyb3rWard0g"}, {score: 2.582, username: "ItsReallyNick"}, {score: 2.302, username: "TheColonial"}, {score: 1.86, username: "DanielGallagher"}] | 91 | 46 | 0.705 | [{score: 0.358, username: "margueritegg"}, {score: 0.3, username: "budejicky"}, {score: 0.26, username: "politicalHEDGE"}, {score: 0.246, username: "420Cyber"}, {score: 0.237, username: "CryptoTrackerBt"}] | 81 | 33 | 0.506 | [{score: 2.447, username: "vaaaaanquish"}, {score: 0.903, username: "__snow_rabbit__"}, {score: 0.841, username: "Moririn47273285"}, {score: 0.604, username: "OSS_News"}, {score: 0.591, username: "shiget84"}] | 1747 | 28 | 0.429 | [{score: 0.278, username: "tech__lib"}, {score: 0.15, username: "niturkan"}, {score: 0.15, username: "djoman_fidele"}, {score: 0.15, username: "scholarsuniv"}, {score: 0.15, username: "seantabatabai"}] | 250 | 27 | 0.414 | [{score: 1.666, username: "anwagnerdreas"}, {score: 1.243, username: "patrick_sahle"}, {score: 1.097, username: "Mareike2405"}, {score: 1.088, username: "AndreasKuczera"}, {score: 1.04, username: "fthierygeo"}] | 1160 | 22 | 0.337 | [{score: 2.357, username: "Silkjaer"}, {score: 1.681, username: "Hodor"}, {score: 1.414, username: "HammerToe"}, {score: 1.262, username: "XrpCenter"}, {score: 1.036, username: "DevNullProd"}] |=== * Community 119 looks like it could be a Data Science cluster, but it also has very popular tech accounts. * Community 157 is full of Neo4j folks * Community 34 has people that are big in the Spring world * Community 22 looks like it has people doing security work * Community 144 is all about the semantic web and triple stores == Overlapping communities Let's now have a look at community overlap. We can compute the biggest overlaps between communities, by running the following query: [source,cypher] ---- MATCH (c1:Community)<-[:IN_COMMUNITY]-(u)-[:IN_COMMUNITY]->(c2:Community) WHERE id(c1) < id(c2) WITH c1, c2, count(*) AS count ORDER BY count DESC LIMIT 20 CALL apoc.create.vRelationship(c1, "COMMON", {count: count}, c2) YIELD rel RETURN c1, rel, c2; ---- We can see the results of running this query in the Neo4j Browser below: .Overlapping communities image::{{<siteurl>}}/uploads/2021/02/overlapping-communities.svg[] The biggest overlap is between community 119 (Data Science/Popular Tech) and communities 34 (Spring) and 157 (Neo4j). The absolute overlap numbers between communities are much lower after that. We can compute similarity scores between the communities using the https://neo4j.com/docs/graph-data-science/1.5/alpha-algorithms/jaccard/[Jaccard Similarity algorithm^], as shown below: [source,cypher] ---- MATCH (c1:Community), (c2:Community) WHERE id(c1) < id(c2) WITH c1, c2, gds.alpha.similarity.jaccard( [(c1)<-[:IN_COMMUNITY]-(u) | id(u)], [(c2)<-[:IN_COMMUNITY]-(u) | id(u)]) AS score WHERE score > 0 MERGE (c1)-[similar:SIMILAR]-(c2) SET similar.score = score; ---- This query computes the similarity betweeen `Community` nodes based on the users that they have in common and then creates a `SIMILAR` relationship between those `Community` nodes. We can see the results of running this query in the Neo4j Browser below: .Overlapping communities - Jaccard Similarity image::{{<siteurl>}}/uploads/2021/02/node-similarity.svg[] From this visualisation we can see that the overlap is tiny between community 119 and the others. The biggest similarity is 0.05 between community 119 and community 34, but the other similarities are much smaller. We can create a table of similarities by running the following query: [source,cypher] ---- MATCH (c1:Community)<-[:IN_COMMUNITY]-(u)-[:IN_COMMUNITY]->(c2:Community) WHERE id(c1) < id(c2) WITH c1, c2, count(*) AS count ORDER BY count DESC LIMIT 20 MATCH (c1)-[similar:SIMILAR]-(c2) RETURN c1.id, c2.id, count, round(similar.score, 3) AS score ORDER BY score DESC LIMIT 10; ---- .Results [opts="header"] |=== | c1.id | c2.id | count | score | 498 | 107 | 3 | 1.0 | 119 | 34 | 228 | 0.052 | 119 | 157 | 187 | 0.038 | 119 | 91 | 46 | 0.012 | 157 | 34 | 20 | 0.011 | 119 | 22 | 36 | 0.009 | 119 | 144 | 36 | 0.009 | 119 | 1747 | 28 | 0.007 | 157 | 144 | 8 | 0.006 | 119 | 81 | 21 | 0.005 |=== There's 100% overlap in the users in community 498 and community 107, but those communities only have 3 nodes! The general amount of overlap is very small. Next we're going to have a look at the highest-ranking users that belong to two communities. We can compute this by running the following query: [source,cypher] ---- MATCH (c1:Community)<-[:IN_COMMUNITY]-(u)-[:IN_COMMUNITY]->(c2:Community) WHERE id(c1) < id(c2) WITH c1, c2, count(*) AS count, apoc.coll.sortNodes(collect(u), "pagerank")[..5] AS topNodes RETURN c1.id, c2.id, count, [node in topNodes | node {.username, score: round(node.pagerank, 3)}] AS topNodes ORDER BY count DESC LIMIT 10; ---- .Results [opts="header", cols="10,10,10,70"] |=== | c1.id | c2.id | count | topNodes | 119 | 34 | 228 | [{score: 14.666, username: "jessitron"}, {score: 10.754, username: "WardCunningham"}, {score: 8.528, username: "garybernhardt"}, {score: 6.882, username: "InfoQ"}, {score: 4.005, username: "skillsmatter"}] | 119 | 157 | 187 | [{score: 6.681, username: "arcadeanalytics"}, {score: 4.694, username: "arangodb"}, {score: 4.55, username: "irregularbi"}, {score: 3.915, username: "CamSemantics"}, {score: 3.7, username: "CluedInHQ"}] | 119 | 91 | 46 | [{score: 0.358, username: "margueritegg"}, {score: 0.3, username: "budejicky"}, {score: 0.26, username: "politicalHEDGE"}, {score: 0.246, username: "420Cyber"}, {score: 0.237, username: "CryptoTrackerBt"}] | 119 | 144 | 36 | [{score: 3.195, username: "Synaptica"}, {score: 1.208, username: "mrgunn"}, {score: 1.126, username: "agentGav"}, {score: 0.684, username: "richdijkstra"}, {score: 0.544, username: "h2cm"}] | 119 | 22 | 36 | [{score: 1.437, username: "s7ephen"}, {score: 0.983, username: "d4rkm4tter"}, {score: 0.978, username: "virusbtn"}, {score: 0.919, username: "CVEnew"}, {score: 0.795, username: "InsanityBit"}] | 119 | 1747 | 28 | [{score: 0.278, username: "tech__lib"}, {score: 0.15, username: "niturkan"}, {score: 0.15, username: "djoman_fidele"}, {score: 0.15, username: "scholarsuniv"}, {score: 0.15, username: "seantabatabai"}] | 119 | 81 | 21 | [{score: 2.447, username: "vaaaaanquish"}, {score: 0.604, username: "OSS_News"}, {score: 0.591, username: "shiget84"}, {score: 0.584, username: "kabukawa"}, {score: 0.424, username: "insomnyan"}] | 157 | 34 | 20 | [{score: 2.631, username: "delitescere"}, {score: 2.424, username: "j_palka"}, {score: 2.398, username: "patbaumgartner"}, {score: 0.844, username: "hackcommitpush"}, {score: 0.673, username: "svzdvd"}] | 119 | 3361 | 12 | [{score: 0.159, username: "hnbot"}, {score: 0.159, username: "Hn150"}, {score: 0.159, username: "SpacedadUNI"}, {score: 0.159, username: "TradeFlo"}, {score: 0.159, username: "nplexROC"}] | 119 | 209 | 10 | [{score: 0.815, username: "matt_zeus"}, {score: 0.698, username: "stereocat"}, {score: 0.675, username: "hi86074659"}, {score: 0.552, username: "tetoran6"}, {score: 0.419, username: "qb0C80aE"}] |=== https://twitter.com/jessitron[@jessitron^] and https://twitter.com/wardcunningham[@wardcunningham^] belong to communities 119 (Data Science/Popular Tech) and 34 (Spring). https://twitter.com/arcadeanalytics[@arcadeanalytics^], https://twitter.com/arangodb[@arangodb^], and https://twitter.com/CluedInHQ[@CluedInHQ] all do work around graph databases, so they end up overlapping communities 119 (Data Science/Popular Tech) and 57 (Neo4j). It would be interesting to see how strongly they belong in each community. The algorithm does compute an association strength score per node per community, but that score isn't returned (at least at the moment!). == How do we use overlapping community detection? So how do we use the results that we get from this algorithm? One way is to use this algorithm to work out which users to follow to learn about other communities. If we pick users who overlap with our community, those users may be better advocates for that community (from our perspective) than the top-ranked users in that community. For example, if we're a member of the Neo4j community, we could find fellow Neo4j users that interact with other communities, by running the following query: [source,cypher] ---- MATCH (c1:Community {id: 157})<-[:IN_COMMUNITY]-(u)-[:IN_COMMUNITY]->(c2:Community) WITH c1, c2, count(*) AS count, apoc.coll.sortNodes(collect(u), "pagerank")[..10] AS topNodes RETURN c2.id, count, [node in topNodes | node {.username, score: round(node.pagerank, 3)}] AS topNodes ORDER BY count DESC LIMIT 10; ---- .Results [opts="header", cols="10,10,80"] |=== | c2.id | count | topNodes | 119 | 187 | [{score: 6.681, username: "arcadeanalytics"}, {score: 4.694, username: "arangodb"}, {score: 4.55, username: "irregularbi"}, {score: 3.915, username: "CamSemantics"}, {score: 3.7, username: "CluedInHQ"}, {score: 2.89, username: "agm1984"}, {score: 2.882, username: "micahstubbs"}, {score: 2.208, username: "newsyc50"}, {score: 2.077, username: "gijn"}, {score: 2.013, username: "jwyg"}] | 34 | 20 | [{score: 2.631, username: "delitescere"}, {score: 2.424, username: "j_palka"}, {score: 2.398, username: "patbaumgartner"}, {score: 0.844, username: "hackcommitpush"}, {score: 0.673, username: "svzdvd"}, {score: 0.643, username: "tonyennis"}, {score: 0.434, username: "jgerity"}, {score: 0.345, username: "gasi"}, {score: 0.298, username: "softvisresearch"}, {score: 0.279, username: "rafalkoko"}] | 144 | 8 | [{score: 0.743, username: "PMissier"}, {score: 0.544, username: "intermineorg"}, {score: 0.343, username: "ConTechLive"}, {score: 0.338, username: "justin_littman"}, {score: 0.226, username: "yooj0907"}, {score: 0.213, username: "imo_weg"}, {score: 0.184, username: "webdevOp"}, {score: 0.15, username: "MahekHanfi"}] | 1558 | 3 | [{score: 0.366, username: "liayeaaah"}, {score: 0.254, username: "King_Sloth95"}, {score: 0.254, username: "yasabdulkadir"}] | 1097 | 2 | [{score: 0.311, username: "BiintAbib"}, {score: 0.242, username: "Fall__yaaram"}] | 266 | 2 | [{score: 0.211, username: "i18nsolutions"}, {score: 0.166, username: "HarishMinions20"}] | 22 | 2 | [{score: 3.431, username: "_wald0"}, {score: 0.15, username: "atatrdp"} | 3846 | 2 | [{score: 0.15, username: "iamarvil"}, {score: 0.15, username: "JETZT_PRde"} | 3281 | 1 | [{score: 0.15, username: "AlanPowiatowy"} | 1226 | 1 | [{score: 0.15, username: "LearnPHPOnline_"}] |=== The top users in community 119 are generally popular tech accounts, but the overlap with the Neo4j community returns other folks doing graph database work. If we want to learn about Spring, https://twitter.com/delitescere[@delitescere^] and https://twitter.com/patbaumgartner[@patbaumgartner^] would be good people to follow. And if we're interested in semantic web, https://twitter.com/pmissier[@Pmissier^] or https://twitter.com/intermineorg[@intermineorg^] would be good bets. == In Summary I first came across this algorithm about 18 months ago and thought it looked awesome, so I'm happy to see it added to the Graph Data Science Library. It seems to find only slight overlap between communities, so I'm not sure how well it'd do if there were communities with very similar members. We've only looked at one use case for it in this blog post, but I'm sure there are others as well. Let me know if you have any questions/ideas in the comments.

In this post we'll explore the Speaker Listener LPA overlapping community algorithm that was added to the Neo4j Graph Data Science Library in version 1.5.0.

uploads/2021/02/sllpa-banner.png

neo4j-gdsl-overlapping-community-detection-sllpa

https://markhneedham.com/blog/2021/02/08/neo4j-gdsl-overlapping-community-detection-sllpa

2021-11-30 00:44:37

Apache Pinot: Exploring indexing techniques on Chicago Crimes

In Neha Pawar's recent blog post, https://www.startree.ai/blogs/what-makes-apache-pinot-fast-chapter-1/[What Makes Apache Pinot fast?], she summarises it with the following sentence: [quote, What makes Apache Pinot fast?, 'https://www.startree.ai/blogs/what-makes-apache-pinot-fast-chapter-1/''] ____ At the heart of the system, Pinot is a columnar store with several smart optimizations that can be applied at various stages of the query by the different Pinot components. Some of the most commonly used and impactful optimizations are data partitioning strategies, segment assignment strategies, smart query routing techniques, a rich set of indexes for filter optimizations, and aggregation optimization techniques. ____ In this blog post we're going to explore one of these techniques, the indexes used for filter optimizations. We'll do this with the help of one of my favourite datasets, the https://data.cityofchicago.org/Public-Safety/Crimes-2001-to-Present/ijzp-q8t2/data[Chicago Crimes data set^], which contains just over 7 million reported incidents of crime since 2001. image::{{<siteurl>}}/uploads/2021/11/indexing-techniques.png[] == How will we assess the impact of indexes? Before we look at this dataset, I want to recommend a video that Apache Pinot co-author Kishore Gopalakrishna recorded in December 2020, in which he explains how to know whether indexes that you've added are working. ++++ <iframe width="560" height="315" src="https://www.youtube.com/embed/VdwVDiXOOVo?start=999" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe> ++++ Kishore goes through different query meta data properties that get returned, as shown in the screenshot below: .Apache Pinot Query Optimization image::{{<siteurl>}}/uploads/2021/11/query-optimization.png[Apache Pinot Query Optimization, role='medium-zoom-image'] Each of the properties is impacted by the smart optimisations that Neha mentioned in her blog post, but here we'll be focusing on `numEntriesScannedInFilter`. Let's get to it! == Setup We're going to spin up a local instance of Pinot using the following Docker compose config: .docker-compose.yml [source, yaml] ---- version: '3.7' services: zookeeper: image: zookeeper:3.5.6 hostname: zookeeper container_name: manual-zookeeper ports: - "2181:2181" environment: ZOOKEEPER_CLIENT_PORT: 2181 ZOOKEEPER_TICK_TIME: 2000 pinot-controller: image: apachepinot/pinot:0.9.0 command: "StartController -zkAddress manual-zookeeper:2181" container_name: "manual-pinot-controller" volumes: - ./config:/config - ./data:/data restart: unless-stopped ports: - "9000:9000" depends_on: - zookeeper pinot-broker: image: apachepinot/pinot:0.9.0 command: "StartBroker -zkAddress manual-zookeeper:2181" restart: unless-stopped container_name: "manual-pinot-broker" volumes: - ./config:/config - ./data:/data ports: - "8099:8099" depends_on: - pinot-controller pinot-server: image: apachepinot/pinot:0.9.0 command: "StartServer -zkAddress manual-zookeeper:2181" restart: unless-stopped container_name: "manual-pinot-server" volumes: - ./config:/config - ./data:/data depends_on: - pinot-broker ---- == Data The Chicago Crimes dataset is available as https://data.cityofchicago.org/Public-Safety/Crimes-2001-to-Present/ijzp-q8t2/data[a CSV file that contains just over 7 million crimes committed in Chicago^] from 2001 until today. A subset of the data is shown below: .Chicago Crimes Dataset image::{{<siteurl>}}/uploads/2021/11/chicago-crimes.png[Chicago Crimes Dataset, role='medium-zoom-image'] == Add Table We're going to import the data into a https://docs.pinot.apache.org/basics/components/table[Pinot table^]. First we'll create a https://docs.pinot.apache.org/basics/components/schema[schema^] that defines the columns, types, and data types: ./config/schema.json [source, json] ---- { "schemaName": "crimes", "dimensionFieldSpecs": [ { "name": "ID", "dataType": "INT" }, { "name": "CaseNumber", "dataType": "STRING" }, { "name": "Block", "dataType": "STRING" }, { "name": "IUCR", "dataType": "STRING" }, { "name": "PrimaryType", "dataType": "STRING" }, { "name": "Arrest", "dataType": "BOOLEAN" }, { "name": "Domestic", "dataType": "BOOLEAN" }, { "name": "Beat", "dataType": "STRING" }, { "name": "District", "dataType": "STRING" }, { "name": "Ward", "dataType": "STRING" }, { "name": "CommunityArea", "dataType": "STRING" }, { "name": "FBICode", "dataType": "STRING" }, { "name": "Latitude", "dataType": "DOUBLE" }, { "name": "Longitude", "dataType": "DOUBLE" } ], "dateTimeFieldSpecs": [{ "name": "Date", "dataType": "STRING", "format" : "1:SECONDS:SIMPLE_DATE_FORMAT:MM/dd/yyyy HH:mm:ss a", "granularity": "1:HOURS" }] } ---- And now a https://docs.pinot.apache.org/configuration-reference/table[table config^]: ./config/table-basic.json [source, json] ---- { "tableName": "crimes", "tableType": "OFFLINE", "segmentsConfig": { "replication": 1 }, "tenants": { "broker":"DefaultTenant", "server":"DefaultTenant" }, "tableIndexConfig": { "loadMode": "MMAP" }, "nullHandlingEnabled": true, "ingestionConfig": { "batchIngestionConfig": { "segmentIngestionType": "APPEND", "segmentIngestionFrequency": "DAILY" }, "transformConfigs": [ {"columnName": "CaseNumber", "transformFunction": "\"Case Number\"" }, {"columnName": "PrimaryType", "transformFunction": "\"Primary Type\"" }, {"columnName": "CommunityArea", "transformFunction": "\"Community Area\"" }, {"columnName": "FBICode", "transformFunction": "\"FBI Code\"" } ] }, "metadata": {} } ---- We're mostly using the defaults here, the only unusual thing that we're doing is specifying some `transformConfigs` to take care of column names containing spaces. If you want to learn more about these transformations, see my blog post on https://www.markhneedham.com/blog/2021/11/25/apache-pinot-csv-columns-spaces/[Importing CSV files with columns containing spaces]. We can create the table by running the following command: [source, bash] ---- docker exec -it manual-pinot-controller bin/pinot-admin.sh AddTable \ -tableConfigFile /config/table-basic.json \ -schemaFile /config/schema.json -exec ---- == Import CSV file Next we're going to import the CSV file. To do this we'll define the following ingestion job spec: ./config/job-spec.yml [source, yaml] ---- executionFrameworkSpec: name: 'standalone' segmentGenerationJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentGenerationJobRunner' segmentTarPushJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentTarPushJobRunner' jobType: SegmentCreationAndTarPush inputDirURI: '/data' includeFileNamePattern: 'glob:**/Crimes_-_2001_to_Present.csv' outputDirURI: '/opt/pinot/data/crimes' overwriteOutput: true pinotFSSpecs: - scheme: file className: org.apache.pinot.spi.filesystem.LocalPinotFS recordReaderSpec: dataFormat: 'csv' className: 'org.apache.pinot.plugin.inputformat.csv.CSVRecordReader' configClassName: 'org.apache.pinot.plugin.inputformat.csv.CSVRecordReaderConfig' tableSpec: tableName: 'crimes' pinotClusterSpecs: - controllerURI: 'http://localhost:9000' ---- [source, bash] ---- docker exec \ -it manual-pinot-controller bin/pinot-admin.sh LaunchDataIngestionJob \ -jobSpecFile /config/job-spec.yml ---- == Querying Pinot We're going to run queries against Pinot using the Query Console of the https://docs.pinot.apache.org/basics/components/exploring-pinot[Pinot Data Explorer^]. You can access this at http://localhost:9000/#/query. We can write SQL queries in the SQL Editor and then run them by pressing 'Cmd + Enter'. We'll then toggle "Show JSON Format" so that we can see the meta data of our query. You can see a screenshot below: .JSON Format image::{{<siteurl>}}/uploads/2021/11/show-json-format.png[JSON Format] .Query metadata [source, json] ---- { "numServersQueried": 1, "numServersResponded": 1, "numSegmentsQueried": 1, "numSegmentsProcessed": 1, "numSegmentsMatched": 1, "numConsumingSegmentsQueried": 0, "numDocsScanned": 10, "numEntriesScannedInFilter": 0, "numEntriesScannedPostFilter": 150, "numGroupsLimitReached": false, "totalDocs": 7434990, "timeUsedMs": 5 } ---- From this meta data, the main thing that we learn is that there are 7,434,990 documents/rows in this table In the rest of this post we're only going to focus on the following properties: [source, json] ---- { "numDocsScanned": 10, "numEntriesScannedInFilter": 0, "numEntriesScannedPostFilter": 150, "timeUsedMs": 5 } ---- We're going to analyse a query that checks the equality of one column. == Forward index Let's start with a query that counts the number of crimes committed where an arrest has happened: [source, sql] ---- select count(*) from crimes WHERE Beat = '1434' ---- .Results [source, json] ---- { "numDocsScanned": 27973, "numEntriesScannedInFilter": 7434990, "numEntriesScannedPostFilter": 0, "timeUsedMs": 71 } ---- From these values we can see that the SQL engine has had to scan every document to check its value for the `Arrest` column and that there were 27,973 documents that matched this predicate. == Forward index + Inverted Index on Beat column One optimisation that we can do is to add the `Arrest` column as an https://docs.pinot.apache.org/basics/indexing/inverted-index[inverted index^]. With an inverted index, Pinot keeps a map from each unique value to a bitmap of rows, meaning that we'll no longer have to scan all the values in these column. We can add an inverted index as `tableIndexConfig.invertedIndexColumns`, as shown in the following table config: ./config/table-inverted-index.json [source, json] ---- { "tableName": "crimes_inverted", "tableType": "OFFLINE", "segmentsConfig": { "replication": 1 }, "tenants": { "broker":"DefaultTenant", "server":"DefaultTenant" }, "tableIndexConfig": { "loadMode": "MMAP", "invertedIndexColumns": [ "Beat" ] }, "ingestionConfig": { "batchIngestionConfig": { "segmentIngestionType": "APPEND", "segmentIngestionFrequency": "DAILY" }, "transformConfigs": [ {"columnName": "CaseNumber", "transformFunction": "\"Case Number\"" }, {"columnName": "PrimaryType", "transformFunction": "\"Primary Type\"" }, {"columnName": "CommunityArea", "transformFunction": "\"Community Area\"" }, {"columnName": "FBICode", "transformFunction": "\"FBI Code\"" } ] }, "metadata": {} } ---- We could apply that to our existing `crimes` table, but to make it easier to compare the different techniques we're going to create a new table for each technique. Run the following command to add a new table `crimes_inverted` based on this table config: .Add table with inverted index [source, bash] ---- docker exec -it manual-pinot-controller bin/pinot-admin.sh AddTable \ -tableConfigFile /config/table-basic.json \ -schemaFile /config/schema.json -exec ---- Now we're going to import the same CSV into this table, using the following ingestion job spec: ./config/job-spec-inverted.yml [source, yaml] ---- executionFrameworkSpec: name: 'standalone' segmentGenerationJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentGenerationJobRunner' segmentTarPushJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentTarPushJobRunner' segmentUriPushJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentUriPushJobRunner' jobType: SegmentCreationAndTarPush inputDirURI: '/data' includeFileNamePattern: 'glob:**/Crimes_beat_sorted.csv' outputDirURI: '/opt/pinot/data/crimes_inverted' overwriteOutput: true pinotFSSpecs: - scheme: file className: org.apache.pinot.spi.filesystem.LocalPinotFS recordReaderSpec: dataFormat: 'csv' className: 'org.apache.pinot.plugin.inputformat.csv.CSVRecordReader' configClassName: 'org.apache.pinot.plugin.inputformat.csv.CSVRecordReaderConfig' tableSpec: tableName: 'crimes_inverted' schemaURI: 'http://localhost:9000/tables/crimes/schema' pinotClusterSpecs: - controllerURI: 'http://localhost:9000' ---- Because our table name and schema name are different we need to explicitly specify the `schemaURI`, otherwise it will try to look for a non existent schema at http://localhost:9000/tables/crimes_inverted/schema. We're also using a different `outputDirURI` than for the previous job spec. We need to do this so that the segments from the `crimes` table don't get included in the `crimes_inverted` table. Run the ingestion job: [source, bash] ---- docker exec \ -it manual-pinot-controller bin/pinot-admin.sh LaunchDataIngestionJob \ -jobSpecFile /config/job-spec-inverted.yml ---- Once the job has finished, we can run our query again: [source, sql] ---- select count(*) from crimes_inverted WHERE Beat = '1434' ---- We should see the following output: .Results [source, json] ---- { "numDocsScanned": 1992434, "numEntriesScannedInFilter": 0, "numEntriesScannedPostFilter": 0, "timeUsedMs": 3, } ---- Our query is almost 20x faster than it was before and the `numEntriesScannedInFilter` is down to 0. === Sorted Forward Index on Beat column We could instead create a https://docs.pinot.apache.org/basics/indexing/forward-index#sorted-forward-index-with-run-length-encoding[sorted index^]. With a sorted index, Pinot keeps a mapping from unique values to start and end document/row ids. [NOTE] ==== The sorted (forward) index for the `Beat` column will replace the default forward index used in the 'No Indexes' example. ==== A table can only have one sorted column and, for offline data ingestion the data in that column must be sorted before we ingest it into Pinot. Since the CSV file isn't too big, we can sort it using Pandas and write the sorted data to a new CSV file, using the following script: [source,python] ---- import pandas as pd df = pd.read_csv("data/Crimes_-_2001_to_Present.csv", dtype=object) df.sort_values(by=["Beat"]).to_csv("data/Crimes_beat_sorted.csv", index=False) ---- Now let's create a new table, which we'll call `crimes_sorted`: ./config/table-sorted-index.json [source, json] ---- { "tableName": "crimes_sorted", "tableType": "OFFLINE", "segmentsConfig": { "replication": 1 }, "tenants": { "broker":"DefaultTenant", "server":"DefaultTenant" }, "tableIndexConfig": { "loadMode": "MMAP", "sortedColumn": [ "Beat" ] }, "ingestionConfig": { "batchIngestionConfig": { "segmentIngestionType": "APPEND", "segmentIngestionFrequency": "DAILY" }, "transformConfigs": [ {"columnName": "CaseNumber", "transformFunction": "\"Case Number\"" }, {"columnName": "PrimaryType", "transformFunction": "\"Primary Type\"" }, {"columnName": "CommunityArea", "transformFunction": "\"Community Area\"" }, {"columnName": "FBICode", "transformFunction": "\"FBI Code\"" } ] }, "metadata": {} } ---- Run the following command to add a new table `crimes_sorted` based on this table config: .Add table with sorted index [source, bash] ---- docker exec -it manual-pinot-controller bin/pinot-admin.sh AddTable \ -tableConfigFile /config/table-sorted-index.json \ -schemaFile /config/schema.json -exec ---- And finally let's create an ingestion job spec to import the CSV file: ./config/job-spec-sorted.yml [source, yaml] ---- executionFrameworkSpec: name: 'standalone' segmentGenerationJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentGenerationJobRunner' segmentTarPushJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentTarPushJobRunner' segmentUriPushJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentUriPushJobRunner' jobType: SegmentCreationAndTarPush inputDirURI: '/data' includeFileNamePattern: 'glob:**/Crimes_beat_sorted.csv' outputDirURI: '/opt/pinot/data/crimes-sorted' overwriteOutput: true pinotFSSpecs: - scheme: file className: org.apache.pinot.spi.filesystem.LocalPinotFS recordReaderSpec: dataFormat: 'csv' className: 'org.apache.pinot.plugin.inputformat.csv.CSVRecordReader' configClassName: 'org.apache.pinot.plugin.inputformat.csv.CSVRecordReaderConfig' tableSpec: tableName: 'crimes_sorted' schemaURI: 'http://localhost:9000/tables/crimes/schema' pinotClusterSpecs: - controllerURI: 'http://localhost:9000' ---- Again we need to specify the `schemaURI` since our table name and schema name differ. We can ingest the data by running the following command: [source, bash] ---- docker exec \ -it manual-pinot-controller bin/pinot-admin.sh LaunchDataIngestionJob \ -jobSpecFile /config/job-spec-sorted.yml ---- Once the job has finished, we can run our query again: [source, sql] ---- select count(*) from crimes_sorted WHERE Beat = '1434' ---- We should see the following output: .Results [source, json] ---- { "numDocsScanned": 27973, "numEntriesScannedInFilter": 0, "numEntriesScannedPostFilter": 0, "timeUsedMs": 3, } ---- Again we don't have any `numEntriesScannedInFilter`, but the query time isn't all that different to when we used the inverted index. == Space vs Time Trade-off The disadvantage of adding extra indexes is that they take up more space on disk. We can check how much space each of our tables consumes by running the following: .Compute table size [source, bash] ---- { printf "%-20s%-12s\n" "Table" "Bytes" for table in 'crimes' 'crimes_sorted' 'crimes_inverted'; do size=`curl -X GET "http://localhost:9000/tables/${table}/size?detailed=true" -H "accept: application/json" 2>/dev/null | jq '.reportedSizeInBytes'`; printf "%-20s%-12s\n" ${table} ${size}; done } ---- .Results [source, text] ---- Table Bytes crimes 340797740 crimes_sorted 332435811 crimes_inverted 355940552 ---- Looking at these numbers, we can see that: * The `crimes_inverted` table takes up the most space, but we shouldn't be too surprised because this is the only table that has two indexes for the `Beat` column: a default forward index and a inverted index. * The sorted forward index used by the `crimes_sorted` table for the `Beat` column actually takes up less space than the unsorted forward index that's used by default. If we want to see a break down of the space usage, we can do that by executing the following command: .Table space breakdown [source, bash] ---- for table in 'crimes' 'crimes_sorted' 'crimes_inverted'; do printf "Table: $table\n" docker exec -it manual-pinot-server ls -l /tmp/data/pinotServerData/${table}_OFFLINE/${table}_OFFLINE_0/v3/ printf "\n" done ---- .Results [source, text] ---- Table: crimes total 332824 -rw-r--r-- 1 root root 340783919 Nov 30 12:11 columns.psf -rw-r--r-- 1 root root 16 Nov 30 12:11 creation.meta -rw-r--r-- 1 root root 2335 Nov 30 12:11 index_map -rw-r--r-- 1 root root 11470 Nov 30 12:11 metadata.properties Table: crimes_sorted total 324656 -rw-r--r-- 1 root root 332421987 Nov 30 14:02 columns.psf -rw-r--r-- 1 root root 16 Nov 30 14:02 creation.meta -rw-r--r-- 1 root root 2331 Nov 30 14:02 index_map -rw-r--r-- 1 root root 11477 Nov 30 14:02 metadata.properties Table: crimes_inverted total 347612 -rw-r--r-- 1 root root 355926631 Nov 30 13:51 columns.psf -rw-r--r-- 1 root root 16 Nov 30 13:16 creation.meta -rw-r--r-- 1 root root 2417 Nov 30 13:51 index_map -rw-r--r-- 1 root root 11488 Nov 30 13:16 metadata.properties ---- == Conclusion In this post we've taken a brief look at two of Apache Pinot's indexing techniques and applied them to a query that counted the number of crimes on a specific beat. Our query runs 20x faster and we only had a small space increase when using an inverted index and actually saved space with the sorted forward index. In our next post we'll look at other indexes and apply them to queries that filter on more than one field. // Now let's start with a query that aggregates the types of crime where an arrest has happened: // [source, sql] // ---- // select PrimaryType, count(*) // from crimes // WHERE Arrest = true // GROUP BY PrimaryType // ORDER BY count(*) DESC // limit 10 // ---- // [source, json] // ---- // { // "numServersQueried": 1, // "numServersResponded": 1, // "numSegmentsQueried": 1, // "numSegmentsProcessed": 1, // "numSegmentsMatched": 1, // "numConsumingSegmentsQueried": 0, // "numDocsScanned": 1992434, // "numEntriesScannedInFilter": 7434990, // "numEntriesScannedPostFilter": 1992434, // "numGroupsLimitReached": false, // "totalDocs": 7434990, // "timeUsedMs": 167, // } // ----

In this post we'll learn how to use Apache Pinot indexes on a Chicago Crime dataset.

uploads/2021/11/indexing-techniques.png

apache-pinot-exploring-index-chicago-crimes

https://markhneedham.com/blog/2021/11/30/apache-pinot-exploring-index-chicago-crimes

2021-11-23 00:44:37

Apache Pinot: org.apache.helix.HelixException: Cluster structure is not set up for cluster: PinotCluster

In my continued exploration of Apache Pinot, I wanted to spin up all the components individually rather than relying on one of the QuickStarts that takes care of that for me. In doing so I came across an interesting error that we'll explore in this post. == Setup We're going to spin up a local instance of Pinot using the following Docker compose config: [source, yaml] ---- version: '3.7' services: zookeeper: image: zookeeper:3.5.6 hostname: zookeeper container_name: manual-zookeeper ports: - "2181:2181" environment: ZOOKEEPER_CLIENT_PORT: 2181 ZOOKEEPER_TICK_TIME: 2000 pinot-controller: image: apachepinot/pinot:0.9.0 command: "StartController -zkAddress manual-zookeeper:2181" container_name: "manual-pinot-controller" ports: - "9000:9000" depends_on: - zookeeper pinot-broker: image: apachepinot/pinot:0.9.0 command: "StartBroker -zkAddress manual-zookeeper:2181" container_name: "manual-pinot-broker" ports: - "8099:8099" depends_on: - pinot-controller pinot-server: image: apachepinot/pinot:0.9.0 command: "StartServer -zkAddress manual-zookeeper:2181" container_name: "manual-pinot-server" depends_on: - pinot-broker ---- This will spin up a ZooKeeper instance that will store all of Pinot's metadata. We'll also spin up a Pinot Controller to manage the metadata, a Pinot Server to host the data and serve queries, and a Pinot Broker that receives queries from clients and sends them to the right server. We can launch the containers by running the following command: [source, bash] ---- docker-compose up ---- If we wait a few seconds we'll eventually see the following exception: [source, text] ---- manual-pinot-broker | 2021/11/23 13:20:02.792 ERROR [ZKHelixManager] [Start a Pinot [BROKER]] fail to createClient. manual-pinot-broker | org.apache.helix.HelixException: Cluster structure is not set up for cluster: PinotCluster manual-pinot-broker | at org.apache.helix.manager.zk.ZKHelixManager.handleNewSession(ZKHelixManager.java:1124) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.helix.manager.zk.ZKHelixManager.createClient(ZKHelixManager.java:701) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.helix.manager.zk.ZKHelixManager.connect(ZKHelixManager.java:738) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.broker.broker.helix.BaseBrokerStarter.start(BaseBrokerStarter.java:195) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.service.PinotServiceManager.startBroker(PinotServiceManager.java:147) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.service.PinotServiceManager.startRole(PinotServiceManager.java:96) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.admin.command.StartServiceManagerCommand$1.lambda$run$0(StartServiceManagerCommand.java:276) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.admin.command.StartServiceManagerCommand.startPinotService(StartServiceManagerCommand.java:302) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.admin.command.StartServiceManagerCommand$1.run(StartServiceManagerCommand.java:276) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] ---- The problem here is that the broker is unable to start because the cluster structure hasn't been set up yet. I found that this only happened the first time that I ran `docker-compose up`. If I killed everything and ran the command again it would work fine. This made me suspicious that some data was being written to ZooKeeper the first time that everything started up on my machine and that it was being kept around for subsequent runs. I was able to confirm that by running `docker-compose rm` to delete everything before I ran `docker-compose up` again. After spending a bit of time reading the Pinot code, I could see that the Broker was trying to read metadata from ZooKeeper before the Controller had written it: .Race condition - ZooKeeper metadata image::{{<siteurl>}}/uploads/2021/11/pinot-zookeeper.png[] One way to work around this would be to stop the Broker from starting until the Controller was ready, but there isn't really a clean way to do this in Docker. Instead what we can do is use the https://docs.docker.com/config/containers/start-containers-automatically/[`restart: unless-stopped`^] config parameter, which was shown to me by Diogo Baeder on the Pinot user Slack. .Docker restart policy image::{{<siteurl>}}/uploads/2021/11/restart-policy.png[] Our Docker Compose file therefore looks like this: [source, yaml] ---- version: '3.7' services: zookeeper: image: zookeeper:3.5.6 hostname: zookeeper container_name: manual-zookeeper ports: - "2181:2181" environment: ZOOKEEPER_CLIENT_PORT: 2181 ZOOKEEPER_TICK_TIME: 2000 pinot-controller: image: apachepinot/pinot:0.9.0 command: "StartController -zkAddress manual-zookeeper:2181" container_name: "manual-pinot-controller" restart: unless-stopped ports: - "9000:9000" depends_on: - zookeeper pinot-broker: image: apachepinot/pinot:0.9.0 command: "StartBroker -zkAddress manual-zookeeper:2181" restart: unless-stopped container_name: "manual-pinot-broker" ports: - "8099:8099" depends_on: - pinot-controller pinot-server: image: apachepinot/pinot:0.9.0 command: "StartServer -zkAddress manual-zookeeper:2181" restart: unless-stopped container_name: "manual-pinot-server" depends_on: - pinot-broker ---- And if we run `docker-compose up` now, we'll see the following: [source, text] ---- manual-pinot-broker | 2021/11/23 13:53:19.419 ERROR [PinotServiceManager] [Start a Pinot [BROKER]] Failed to start Pinot Broker manual-pinot-broker | org.apache.helix.HelixException: Cluster structure is not set up for cluster: PinotCluster manual-pinot-broker | at org.apache.helix.manager.zk.ZKHelixManager.handleNewSession(ZKHelixManager.java:1124) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.helix.manager.zk.ZKHelixManager.createClient(ZKHelixManager.java:701) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.helix.manager.zk.ZKHelixManager.connect(ZKHelixManager.java:738) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.broker.broker.helix.BaseBrokerStarter.start(BaseBrokerStarter.java:195) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.service.PinotServiceManager.startBroker(PinotServiceManager.java:147) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.service.PinotServiceManager.startRole(PinotServiceManager.java:96) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.admin.command.StartServiceManagerCommand$1.lambda$run$0(StartServiceManagerCommand.java:276) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.admin.command.StartServiceManagerCommand.startPinotService(StartServiceManagerCommand.java:302) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.admin.command.StartServiceManagerCommand$1.run(StartServiceManagerCommand.java:276) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | 2021/11/23 13:53:19.420 ERROR [StartServiceManagerCommand] [Start a Pinot [BROKER]] Failed to start a Pinot [BROKER] at 0.927 since launch manual-pinot-broker | org.apache.helix.HelixException: Cluster structure is not set up for cluster: PinotCluster manual-pinot-broker | at org.apache.helix.manager.zk.ZKHelixManager.handleNewSession(ZKHelixManager.java:1124) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.helix.manager.zk.ZKHelixManager.createClient(ZKHelixManager.java:701) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.helix.manager.zk.ZKHelixManager.connect(ZKHelixManager.java:738) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.broker.broker.helix.BaseBrokerStarter.start(BaseBrokerStarter.java:195) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.service.PinotServiceManager.startBroker(PinotServiceManager.java:147) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.service.PinotServiceManager.startRole(PinotServiceManager.java:96) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.admin.command.StartServiceManagerCommand$1.lambda$run$0(StartServiceManagerCommand.java:276) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.admin.command.StartServiceManagerCommand.startPinotService(StartServiceManagerCommand.java:302) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] manual-pinot-broker | at org.apache.pinot.tools.admin.command.StartServiceManagerCommand$1.run(StartServiceManagerCommand.java:276) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] .... manual-pinot-broker | 2021/11/23 13:53:21.834 INFO [StartBrokerCommand] [main] Executing command: StartBroker -brokerHost null -brokerPort 8099 -zkAddress manual-zookeeper:2181 manual-pinot-broker | 2021/11/23 13:53:21.850 INFO [StartServiceManagerCommand] [main] Executing command: StartServiceManager -clusterName PinotCluster -zkAddress manual-zookeeper:2181 -port -1 -bootstrapServices [] manual-pinot-broker | 2021/11/23 13:53:21.850 INFO [StartServiceManagerCommand] [main] Starting a Pinot [SERVICE_MANAGER] at 0.603s since launch manual-pinot-broker | 2021/11/23 13:53:21.853 INFO [StartServiceManagerCommand] [main] Started Pinot [SERVICE_MANAGER] instance [ServiceManager_f165640a2780_-1] at 0.606s since launch manual-pinot-broker | 2021/11/23 13:53:21.860 INFO [StartServiceManagerCommand] [Start a Pinot [BROKER]] Starting a Pinot [BROKER] at 0.612s since launch manual-pinot-broker | Nov 23, 2021 1:53:27 PM org.glassfish.grizzly.http.server.NetworkListener start manual-pinot-broker | INFO: Started listener bound to [0.0.0.0:8099] manual-pinot-broker | Nov 23, 2021 1:53:27 PM org.glassfish.grizzly.http.server.HttpServer start manual-pinot-broker | INFO: [HttpServer] Started. manual-pinot-server | 2021/11/23 13:53:30.322 INFO [StartServiceManagerCommand] [Start a Pinot [SERVER]] Started Pinot [SERVER] instance [Server_192.168.144.5_8098] at 11.279s since launch manual-pinot-broker | 2021/11/23 13:53:32.578 INFO [StartServiceManagerCommand] [Start a Pinot [BROKER]] Started Pinot [BROKER] instance [Broker_192.168.144.4_8099] at 11.33s since launch ---- The Broker still has the same problem the first time that it starts, but when it restarts and tries again the cluster structure is ready to go.

In this post we'll learn how to work around an error message when starting up an Apache Pinot cluster.

apache-pinot-helix-exception-cluster-structure-not-set-up

https://markhneedham.com/blog/2021/11/23/apache-pinot-helix-exception-cluster-structure-not-set-up

2021-11-25 00:44:37

Apache Pinot: Importing CSV files with columns containing spaces

I've been playing around with one of my favourite datasets from the https://data.cityofchicago.org/Public-Safety/Crimes-2001-to-Present/ijzp-q8t2[Chicago Data Portal^] and spent a while figuring out how to import columns that contain spaces into Apache Pinot. In this blog post we'll learn how to do that using a subset of the data. == Setup We're going to spin up a local instance of Pinot using the following Docker compose config: .docker-compose.yml [source, yaml] ---- version: '3.7' services: zookeeper: image: zookeeper:3.5.6 hostname: zookeeper container_name: manual-zookeeper ports: - "2181:2181" environment: ZOOKEEPER_CLIENT_PORT: 2181 ZOOKEEPER_TICK_TIME: 2000 pinot-controller: image: apachepinot/pinot:0.9.0 command: "StartController -zkAddress manual-zookeeper:2181" container_name: "manual-pinot-controller" volumes: - ./config:/config - ./data:/data restart: unless-stopped ports: - "9000:9000" depends_on: - zookeeper pinot-broker: image: apachepinot/pinot:0.9.0 command: "StartBroker -zkAddress manual-zookeeper:2181" restart: unless-stopped container_name: "manual-pinot-broker" volumes: - ./config:/config - ./data:/data ports: - "8099:8099" depends_on: - pinot-controller pinot-server: image: apachepinot/pinot:0.9.0 command: "StartServer -zkAddress manual-zookeeper:2181" restart: unless-stopped container_name: "manual-pinot-server" volumes: - ./config:/config - ./data:/data depends_on: - pinot-broker ---- We can launch the container by running the following command: [source, bash] ---- docker-compose up ---- == Data Let's say we have the following CSV file that contains two columns, one that contains spaces and one that doesn't: .data/import.csv [options="header"] |=== | ID | Case Number | 10224738 | HY411648 | 10224739 | HY411615 | 11646166 | JC213529 | 10224740 |HY411595 |=== This CSV contains a subset of the https://data.cityofchicago.org/Public-Safety/Crimes-2001-to-Present/ijzp-q8t2/data[crimes data from the Chicago Data Portal^]. == Create Table We're going to create a `crimes` Pinot table and associated schema based on this CSV file: .config/schema.json [source, json] ---- { "schemaName": "crimes", "dimensionFieldSpecs": [ { "name": "ID", "dataType": "INT" }, { "name": "Case Number", "dataType": "STRING" } ] } ---- .config/table.json [source, json] ---- { "tableName": "crimes", "tableType": "OFFLINE", "segmentsConfig": { "replication": 1 }, "tenants": { "broker":"DefaultTenant", "server":"DefaultTenant" }, "tableIndexConfig": { "loadMode": "MMAP" }, "ingestionConfig": { "batchIngestionConfig": { "segmentIngestionType": "APPEND", "segmentIngestionFrequency": "DAILY" } }, "metadata": {} } ---- Let's try to create a table based on this config: [source, bash] ---- docker exec -it manual-pinot-controller bin/pinot-admin.sh AddTable -tableConfigFile /config/table.json -schemaFile /config/schema.json -exec ---- [source, text] ---- 2021/11/25 11:57:27.088 ERROR [AddTableCommand] [main] Got Exception to upload Pinot Schema: crimes org.apache.pinot.common.exception.HttpErrorStatusException: Got error status code: 400 (Bad Request) with reason: "Cannot add invalid schema: crimes. Reason: The column name "Case Number" should not contain blank space." while sending request: http://192.168.144.3:9000/schemas to controller: d15b07933b22, version: Unknown at org.apache.pinot.common.utils.FileUploadDownloadClient.sendRequest(FileUploadDownloadClient.java:510) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at org.apache.pinot.common.utils.FileUploadDownloadClient.addSchema(FileUploadDownloadClient.java:616) ~[pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at org.apache.pinot.tools.admin.command.AddTableCommand.uploadSchema(AddTableCommand.java:166) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at org.apache.pinot.tools.admin.command.AddTableCommand.execute(AddTableCommand.java:203) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at org.apache.pinot.tools.Command.call(Command.java:33) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at org.apache.pinot.tools.Command.call(Command.java:29) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at picocli.CommandLine.executeUserObject(CommandLine.java:1953) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at picocli.CommandLine.access$1300(CommandLine.java:145) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at picocli.CommandLine$RunLast.executeUserObjectOfLastSubcommandWithSameParent(CommandLine.java:2352) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at picocli.CommandLine$RunLast.handle(CommandLine.java:2346) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at picocli.CommandLine$RunLast.handle(CommandLine.java:2311) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at picocli.CommandLine$AbstractParseResultHandler.execute(CommandLine.java:2179) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at picocli.CommandLine.execute(CommandLine.java:2078) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at org.apache.pinot.tools.admin.PinotAdministrator.execute(PinotAdministrator.java:161) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] at org.apache.pinot.tools.admin.PinotAdministrator.main(PinotAdministrator.java:192) [pinot-all-0.9.0-jar-with-dependencies.jar:0.9.0-cf8b84e8b0d6ab62374048de586ce7da21132906] ---- Pinot schemas don't allow column names that contain spaces, so we'll have to get rid of the space in `Case Number`. We can update the schema to look like this: .config/schema.json [source, json] ---- { "schemaName": "crimes", "dimensionFieldSpecs": [ { "name": "ID", "dataType": "INT" }, { "name": "CaseNumber", "dataType": "STRING" } ] } ---- If we re-rerun the `AddTable` command, we'll see the following output: [source, text] ---- 2021/11/25 12:02:04.606 INFO [AddTableCommand] [main] Executing command: AddTable -tableConfigFile /config/table.json -schemaFile /config/schema.json -controllerProtocol http -controllerHost 192.168.144.3 -controllerPort 9000 -user null -password [hidden] -exec 2021/11/25 12:02:05.084 INFO [AddTableCommand] [main] {"status":"Table crimes_OFFLINE succesfully added"} ---- == Ingest CSV file Now we're going to imort the CSV file that we saw at the beginning of the post. To do this, we'll create the following data ingestion job spec: .config/job-spec.yml [source, yaml] ---- executionFrameworkSpec: name: 'standalone' segmentGenerationJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentGenerationJobRunner' segmentTarPushJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentTarPushJobRunner' segmentUriPushJobRunnerClassName: 'org.apache.pinot.plugin.ingestion.batch.standalone.SegmentUriPushJobRunner' jobType: SegmentCreationAndTarPush inputDirURI: '/data' includeFileNamePattern: 'glob:**/import.csv' outputDirURI: '/opt/pinot/data/crimes/segments/' overwriteOutput: true pinotFSSpecs: - scheme: file className: org.apache.pinot.spi.filesystem.LocalPinotFS recordReaderSpec: dataFormat: 'csv' className: 'org.apache.pinot.plugin.inputformat.csv.CSVRecordReader' configClassName: 'org.apache.pinot.plugin.inputformat.csv.CSVRecordReaderConfig' tableSpec: tableName: 'crimes' pinotClusterSpecs: - controllerURI: 'http://localhost:9000' ---- We can import the CSV file by running the following command: [source, bash] ---- docker exec \ -it manual-pinot-controller bin/pinot-admin.sh LaunchDataIngestionJob \ -jobSpecFile /config/job-spec.yml ---- Once we've run this we can navigate to the Query Console at http://localhost:9000/#/query and run the following query: [source, sql] ---- select * from crimes limit 10 ---- .Result [options="header"] |=== | CaseNumber | ID | null | 10224738 | null| 10224739 | null| 11646166 | null | 10224740 |=== Hmmm, the `CaseNumber` is always null, which isn't what we want. To deal with this problem we'll need to add an https://docs.pinot.apache.org/developers/advanced/ingestion-level-transformations#extract-value-from-a-column-containing-space[ingestion transformation config^] to our table config. == Update table config and reingest Let's update our table config to add the transform config: .config/table.json [source, json] ---- { "tableName": "crimes", "tableType": "OFFLINE", "segmentsConfig": { "replication": 1 }, "tenants": { "broker":"DefaultTenant", "server":"DefaultTenant" }, "tableIndexConfig": { "loadMode": "MMAP" }, "ingestionConfig": { "batchIngestionConfig": { "segmentIngestionType": "APPEND", "segmentIngestionFrequency": "DAILY" }, "transformConfigs": [ {"columnName": "CaseNumber", "transformFunction": "\"Case Number\"" } ] }, "metadata": {} } ---- The `columnName` refers to the column name in the schema and the `tranformFunction` describes a function for processing a value from the source data. In this case we're specifying the name of the property/column name from our CSV file and it will extract the values from that column. Before we update the table config, let's first delete the segment that we ingested in the previous section: .Drop crimes table segments [source, bash] ---- curl -X DELETE "http://localhost:9000/segments/crimes?type=OFFLINE" -H "accept: application/json" ---- Now we can update the table config: .Update table config [source, bash] ---- curl 'http://localhost:9000/tables/crimes_OFFLINE' \ -X 'PUT' \ -H 'Content-Type: application/json' \ --data-binary "@config/table.json" ---- And finally, we can run the data ingestion job again: [source, bash] ---- docker exec \ -it manual-pinot-controller bin/pinot-admin.sh LaunchDataIngestionJob \ -jobSpecFile /config/job-spec.yml ---- Now if we run a query against this table we'll see populated values for the `CaseNumber`: [source, sql] ---- select * from crimes limit 10 ---- .Result [options="header"] |=== | CaseNumber | ID |HY411648 |10224738 |HY411615 |10224739 |JC213529 |11646166 |HY411595 |10224740 |===

In this post we'll learn how to import CSV files with columns that contain spaces into Apache Pinot.

apache-pinot-csv-columns-spaces

https://markhneedham.com/blog/2021/11/25/apache-pinot-csv-columns-spaces

2021-07-16 00:44:37

Apache Pinot: BadQueryRequestException - Cannot convert value to type: LONG

In my continued exploration of Apache Pinot I've been trying out the https://docs.pinot.apache.org/basics/recipes/github-events-stream[GitHub events recipe ^], which imports data from the GitHub events stream into Pinot. In this blog post I want to show how I worked around an exception I was getting when trying to filter the data by one of the timestamp's column. == Setup We're going to spin up a local instance of Pinot using the following Docker compose config: .docker-compose.yml [source, yaml] ---- version: '3' services: pinot: image: apachepinot/pinot:0.7.1 command: "GitHubEventsQuickStart -personalAccessToken ${GITHUB_TOKEN}" container_name: "pinot-github-events-quick-start" ports: - "9000:9000" - "8000:8000" ---- I then created a personal access token and made it available via the `GITHUB_TOKEN` environment variable. We can launch the container by running the following command: [source, bash] ---- docker-compose up ---- == Querying Pinot Once that's ready, we'll navigate to http://localhost:9000/#/query, where we'll see the following screen: image::{{<siteurl>}}/uploads/2021/07/pinot-controller.png[] You can see we have a `pullRequestMergedEvents` table, which contains the most recent events from GitHub. I wanted to count how many events happened in the last 60 minutes, which I did by writing the following query: [source, sql] ---- select count(*) from pullRequestMergedEvents where createdTimeMillis < (now() - 1000*60*60) limit 10 ---- If we run that query we'll get the following output: [source, text] ---- [ { "errorCode": 200, "message": "QueryExecutionError:\norg.apache.pinot.core.query.exception.BadQueryRequestException: Cannot convert value: '1.626415601834E12' to type: LONG\n\tat org.apache.pinot.core.query.pruner.ColumnValueSegmentPruner.convertValue(ColumnValueSegmentPruner.java:261)\n\tat org.apache.pinot.core.query.pruner.ColumnValueSegmentPruner.pruneRangePredicate(ColumnValueSegmentPruner.java:191)\n\tat org.apache.pinot.core.query.pruner.ColumnValueSegmentPruner.pruneSegment(ColumnValueSegmentPruner.java:105)\n\tat org.apache.pinot.core.query.pruner.ColumnValueSegmentPruner.prune(ColumnValueSegmentPruner.java:76)\n\tat org.apache.pinot.core.query.pruner.SegmentPruner.prune(SegmentPruner.java:45)\n\tat org.apache.pinot.core.query.pruner.SegmentPrunerService.prune(SegmentPrunerService.java:63)\n\tat org.apache.pinot.core.query.executor.ServerQueryExecutorV1Impl.processQuery(ServerQueryExecutorV1Impl.java:271)\n\tat org.apache.pinot.core.query.executor.ServerQueryExecutorV1Impl.processQuery(ServerQueryExecutorV1Impl.java:215)\n\tat org.apache.pinot.core.query.executor.QueryExecutor.processQuery(QueryExecutor.java:60)\n\tat org.apache.pinot.core.query.scheduler.QueryScheduler.processQueryAndSerialize(QueryScheduler.java:157)\n\tat org.apache.pinot.core.query.scheduler.QueryScheduler.lambda$createQueryFutureTask$0(QueryScheduler.java:141)\n\tat java.util.concurrent.FutureTask.run(FutureTask.java:266)\n\tat java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511)\n\tat shaded.com.google.common.util.concurrent.TrustedListenableFutureTask$TrustedFutureInterruptibleTask.runInterruptibly(TrustedListenableFutureTask.java:111)" } ] ---- If the WHERE clause read `where createdTimeMillis < now()` it was working fine, so the issue was with my attempts to subtract from that timestamp. https://medium.com/@xiangfu0[Xiang Fu ^] suggested that `(now() - 1000*60*60)` was resulting in a double value, which we can work around by casting it to a long, as shown below: [source, sql] ---- select count(*) from pullRequestMergedEvents where createdTimeMillis > now() - CAST(1000*60*60 AS long) limit 10 ---- If we run that query we'll now get a result: .Result [options="header"] [source, csv] ---- count(*) 178 ---- I found that I could also put the `CAST` function around the whole `(now() - 1000*60*60)` expression, instead of just the `1000*60*60` part: [source, sql] ---- select count(*) from pullRequestMergedEvents where createdTimeMillis > CAST(now() - 1000*60*60 AS long) limit 10 ----

In this post we'll learn how to work around an error message in Apache Pinot when filtering by timestamp.

pinot-bad-query-request-exception-cannot-convert-value-long

https://markhneedham.com/blog/2021/07/16/pinot-bad-query-request-exception-cannot-convert-value-long