cloudops_tsf.py

# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass, field
from typing import Iterator, Optional
from functools import cached_property

import datasets
import pandas as pd
import pyarrow as pa
import pyarrow.parquet as pq
from gluonts.dataset.field_names import FieldName

_CITATION = """\
@article{woo2023pushing,
  title={Pushing the Limits of Pre-training for Time Series Forecasting in the CloudOps Domain},
  author={Woo, Gerald and Liu, Chenghao and Kumar, Akshat and Sahoo, Doyen},
  journal={arXiv preprint arXiv:2310.05063},
  year={2023}
}
"""


_CONFIGS = {
    "azure_vm_traces_2017": {
        "optional_fields": (
            FieldName.FEAT_STATIC_CAT,
            FieldName.FEAT_STATIC_REAL,
            FieldName.PAST_FEAT_DYNAMIC_REAL,
        ),
        "prediction_length": 48,
        "freq": "5T",
        "stride": 48,
        "univariate": True,
        "multivariate": False,
        "rolling_evaluations": 12,
        "test_split_date": pd.Period(
            year=2016, month=12, day=13, hour=15, minute=55, freq="5T"
        ),
        "_feat_static_cat_cardinalities": {
            "train_test": (
                ("vm_id", 17568),
                ("subscription_id", 2713),
                ("deployment_id", 3255),
                ("vm_category", 3),
            ),
            "pretrain": (
                ("vm_id", 177040),
                ("subscription_id", 5514),
                ("deployment_id", 15208),
                ("vm_category", 3),
            ),
        },
        "target_dim": 1,
        "feat_static_real_dim": 3,
        "past_feat_dynamic_real_dim": 2,
    },
    "borg_cluster_data_2011": {
        "optional_fields": (
            FieldName.FEAT_STATIC_CAT,
            FieldName.PAST_FEAT_DYNAMIC_REAL,
        ),
        "prediction_length": 48,
        "freq": "5T",
        "stride": 48,
        "univariate": False,
        "multivariate": True,
        "rolling_evaluations": 12,
        "test_split_date": pd.Period(
            year=2011, month=5, day=28, hour=18, minute=55, freq="5T"
        ),
        "_feat_static_cat_cardinalities": {
            "train_test": (
                ("job_id", 850),
                ("task_id", 11117),
                ("user", 282),
                ("scheduling_class", 4),
                ("logical_job_name", 718),
            ),
            "pretrain": (
                ("job_id", 6072),
                ("task_id", 154503),
                ("user", 518),
                ("scheduling_class", 4),
                ("logical_job_name", 3899),
            ),
        },
        "target_dim": 2,
        "past_feat_dynamic_real_dim": 5,
    },
    "alibaba_cluster_trace_2018": {
        "optional_fields": (
            FieldName.FEAT_STATIC_CAT,
            FieldName.PAST_FEAT_DYNAMIC_REAL,
        ),
        "prediction_length": 48,
        "freq": "5T",
        "stride": 48,
        "univariate": False,
        "multivariate": True,
        "rolling_evaluations": 12,
        "test_split_date": pd.Period(
            year=2018, month=1, day=8, hour=11, minute=55, freq="5T"
        ),
        "_feat_static_cat_cardinalities": {
            "train_test": (
                ("container_id", 6048),
                ("app_du", 1292),
            ),
            "pretrain": (
                ("container_id", 64457),
                ("app_du", 9484),
            ),
        },
        "target_dim": 2,
        "past_feat_dynamic_real_dim": 6,
    },
}

PRETRAIN = datasets.splits.NamedSplit("pretrain")
TRAIN_TEST = datasets.splits.NamedSplit("train_test")

Cardinalities = tuple[tuple[str, int], ...]


@dataclass
class CloudOpsTSFConfig(datasets.BuilderConfig):
    """BuilderConfig for CloudOpsTSF."""

    # load_dataset kwargs
    train_test: bool = field(default=True, init=False)
    pretrain: bool = field(default=False, init=False)
    _include_metadata: tuple[str, ...] = field(default_factory=tuple, init=False)

    # builder kwargs
    prediction_length: int = field(default=None)
    freq: str = field(default=None)
    stride: int = field(default=None)
    univariate: bool = field(default=None)
    multivariate: bool = field(default=None)
    optional_fields: tuple[str, ...] = field(default=None)
    rolling_evaluations: int = field(default=None)
    test_split_date: pd.Period = field(default=None)
    _feat_static_cat_cardinalities: dict[str, Cardinalities] = field(
        default_factory=dict
    )
    target_dim: int = field(default=1)
    feat_static_real_dim: int = field(default=0)
    past_feat_dynamic_real_dim: int = field(default=0)

    METADATA = [
        "freq",
        "prediction_length",
        "stride",
        "rolling_evaluations",
    ]

    @property
    def include_metadata(self) -> tuple[str, ...]:
        return self._include_metadata

    @include_metadata.setter
    def include_metadata(self, value: tuple[str, ...]):
        assert all([v in self.METADATA for v in value]), (
            f"Metadata: {value} is not supported, each item should be one of"
            f" {self.METADATA}"
        )
        self._include_metadata = value

    @cached_property
    def feat_static_cat_cardinalities(self) -> Optional[list[int]]:
        if FieldName.FEAT_STATIC_CAT not in self.optional_fields:
            return None

        if self.pretrain:
            split = "pretrain"
        elif self.train_test:
            split = "train_test"
        else:
            raise ValueError(
                "At least one of `train_test` and `pretrain` should be True"
            )
        return [c[1] for c in self._feat_static_cat_cardinalities[split]]


class CloudOpsTSF(datasets.ArrowBasedBuilder):
    VERSION = datasets.Version("1.0.0")

    BUILDER_CONFIGS = []
    for dataset, config in _CONFIGS.items():
        BUILDER_CONFIGS.append(
            CloudOpsTSFConfig(name=dataset, version=VERSION, description="", **config)
        )

    def _info(self) -> datasets.DatasetInfo:
        def sequence_feature(dtype: str, univar: bool) -> datasets.Sequence:
            if univar:
                return datasets.Sequence(datasets.Value(dtype))
            return datasets.Sequence(datasets.Sequence(datasets.Value(dtype)))

        features = {
            FieldName.START: datasets.Value("timestamp[s]"),
            FieldName.TARGET: sequence_feature("float32", self.config.univariate),
            FieldName.ITEM_ID: datasets.Value("string"),
        }

        CAT_FEATS = (
            FieldName.FEAT_STATIC_CAT,
            FieldName.FEAT_DYNAMIC_CAT,
            FieldName.PAST_FEAT_DYNAMIC,
        )
        REAL_FEATS = (
            FieldName.FEAT_STATIC_REAL,
            FieldName.FEAT_DYNAMIC_REAL,
            FieldName.PAST_FEAT_DYNAMIC_REAL,
        )
        STATIC_FEATS = (FieldName.FEAT_STATIC_CAT, FieldName.FEAT_STATIC_REAL)
        DYNAMIC_FEATS = (
            FieldName.FEAT_DYNAMIC_CAT,
            FieldName.FEAT_DYNAMIC_REAL,
            FieldName.PAST_FEAT_DYNAMIC,
            FieldName.PAST_FEAT_DYNAMIC_REAL,
        )

        for ts_field in self.config.optional_fields:
            # Determine field dtype
            if ts_field in CAT_FEATS:
                dtype = "int32"
            elif ts_field in REAL_FEATS:
                dtype = "float32"
            else:
                raise ValueError(f"Invalid field: {ts_field}")

            # Determine field shape
            if ts_field in STATIC_FEATS:
                univar = True
            elif ts_field in DYNAMIC_FEATS:
                univar = False
            else:
                raise ValueError(f"Invalid field: {ts_field}")

            features[ts_field] = sequence_feature(dtype, univar)

        for metadata in self.config.include_metadata:
            if metadata == "freq":
                features[metadata] = datasets.Value("string")
            elif metadata in ("prediction_length", "stride", "rolling_evaluations"):
                features[metadata] = datasets.Value("int32")
            else:
                raise ValueError(f"Invalid metadata: {metadata}")

        features = datasets.Features(features)

        return datasets.DatasetInfo(
            features=features,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager) -> list[datasets.SplitGenerator]:
        generators = []
        if self.config.train_test:
            downloaded_files = dl_manager.download_and_extract(f"{self.config.name}/train_test.zip")
            generators.append(
                datasets.SplitGenerator(
                    name=TRAIN_TEST if self.config.train_test else PRETRAIN,
                    gen_kwargs={"filepath": downloaded_files}
                )
            )
        if self.config.pretrain:
            downloaded_files = dl_manager.download_and_extract(f"{self.config.name}/pretrain.zip")
            generators.append(
                datasets.SplitGenerator(
                    name=PRETRAIN,
                    gen_kwargs={"filepath": downloaded_files}
                )
            )
        return generators

    def _generate_tables(self, filepath: str) -> Iterator[pa.Table]:
        table = pq.read_table(filepath)

        for batch in table.to_batches():
            columns = batch.columns
            schema = batch.schema
            if self.config.include_metadata:
                freq = pa.array([self.config.freq] * len(batch))
                prediction_length = pa.array([self.config.prediction_length] * len(batch))
                rolling_evaluations = pa.array([self.config.rolling_evaluations] * len(batch))
                stride = pa.array([self.config.stride] * len(batch))
                columns += [freq, prediction_length, rolling_evaluations, stride]
                for pa_field in [pa.field('freq', pa.string()),
                                 pa.field('prediction_length', pa.int32()),
                                 pa.field('rolling_evaluations', pa.int32()),
                                 pa.field('stride', pa.int32())]:
                    schema = schema.append(pa_field)
            yield batch[FieldName.ITEM_ID].to_pylist(), pa.Table.from_arrays(columns, schema=schema)