frogml-core 0.0.1__py3-none-any.whl

frogml/feature_store/feature_sets/streaming.py

@@ -0,0 +1,635 @@
+import collections
+import functools
+import inspect
+from dataclasses import dataclass, field
+from datetime import datetime
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union
+
+from typeguard import typechecked
+
+from _frogml_proto.qwak.feature_store.features.execution_pb2 import (
+    StreamingExecutionSpec as ProtoStreamingExecutionSpec,
+)
+from _frogml_proto.qwak.feature_store.features.feature_set_pb2 import (
+    FeatureSetSpec as ProtoFeatureSetSpec,
+)
+from _frogml_proto.qwak.feature_store.features.feature_set_types_pb2 import (
+    AggregationSpec as ProtoAggregationSpec,
+)
+from _frogml_proto.qwak.feature_store.features.feature_set_types_pb2 import (
+    FeatureSetType as ProtoFeatureSetType,
+)
+from _frogml_proto.qwak.feature_store.features.feature_set_types_pb2 import (
+    StreamingAggregationFeatureSet as ProtoStreamingAggregationFeatureSet,
+)
+from _frogml_proto.qwak.feature_store.features.feature_set_types_pb2 import (
+    StreamingFeatureSetV1 as ProtoStreamingFeatureSetV1,
+)
+from _frogml_proto.qwak.feature_store.sources.streaming_pb2 import StreamingSource
+from _frogml_proto.qwak.feature_store.sources.streaming_pb2 import (
+    StreamingSource as ProtoStreamingSource,
+)
+from frogml.clients.feature_store import FeatureRegistryClient
+from frogml.exceptions import FrogmlException
+from frogml.feature_store._common.artifact_utils import ArtifactSpec, ArtifactsUploader
+from frogml.feature_store.entities.entity import Entity
+from frogml.feature_store.feature_sets._utils._featureset_utils import FeaturesetUtils
+from frogml.feature_store.feature_sets.base_feature_set import BaseFeatureSet
+from frogml.feature_store.feature_sets.execution_spec import ClusterTemplate
+from frogml.feature_store.feature_sets.metadata import (
+    Metadata,
+    get_metadata_from_function,
+    set_metadata_on_function,
+)
+from frogml.feature_store.feature_sets.streaming_backfill import (
+    BackfillBatchDataSourceSpec,
+    StreamingBackfill,
+)
+from frogml.feature_store.feature_sets.transformations import (
+    BaseTransformation,
+    PySparkTransformation,
+    SparkSqlTransformation,
+    Window,
+)
+from frogml.feature_store.sinks.base import BaseSink
+from frogml.feature_store.sinks.streaming.factory import StreamingSinkFactory
+from frogml.feature_store.validations.validation_options import (
+    FeatureSetValidationOptions,
+)
+
+if TYPE_CHECKING:
+    try:
+        import pandas as pd
+    except ImportError:
+        pass
+
+_OFFLINE_SCHEDULING_ATTRIBUTE = "_qwak_offline_scheduling"
+_OFFLINE_CLUSTER_SPEC = "_qwak_offline_cluster_specification"
+_ONLINE_TRIGGER_INTERVAL = "_qwak_online_trigger_interval"
+_ONLINE_CLUSTER_SPEC = "_qwak_online_cluster_specification"
+_METADATA_ = "_qwak_online_cluster_specification"
+
+
+@typechecked
+def feature_set(
+    *,
+    data_sources: List[str],
+    timestamp_column_name: str,
+    offline_scheduling_policy: Optional[str] = "*/30 * * * *",
+    online_trigger_interval: Optional[int] = 5,
+    name: Optional[str] = None,
+    entity: Optional[str] = None,
+    key: Optional[str] = None,
+    auxiliary_sinks: List[BaseSink] = [],
+):
+    """
+    Creates a streaming feature set for the specified entity using the given streaming data sources.
+
+    A streaming feature set allows for real-time updates of features from live data sources, letting ML models access
+    the most recent values without waiting for batch updates.
+
+    :param entity: The name of the entity for which the feature set is being created. An entity typically represents a
+                   unique object or concept, like 'user', 'product', etc. Entity and key are mutually exclusive.
+    :param key: a column name in the feature set which is the key. Entity and key are mutually exclusive.
+    :param data_sources: A list of references to the data sources from which the feature values will be streamed.
+                                Each data source should be capable of providing data in a streaming manner.
+    :param timestamp_column_name: The name of the column in the data source that contains timestamp information. This
+                                  is used to order the data chronologically and ensure that the feature values are
+                                  updated in the correct order.
+    :param offline_scheduling_policy: Defines the offline ingestion policy - which affects the data freshness of
+                                      the offline store. defaults to */30 * * * * (every 30 minutes)
+    :param online_trigger_interval: Defines the online ingestion policy  - which affects the data freshness of
+                                      the online store. defaults to 5 seconds
+    :param name: An optional name for the feature set. If not provided, the name of the function will be used.
+    :param  auxiliary_sinks: list of auxiliary sinks. Not supported with Aggregations
+    Example:
+
+    ... code-block:: python
+
+        @streaming.feature_set(
+            entity="users",
+            data_sources=["users_registration_stream"],
+            timestamp_column_name="reg_date"
+        )
+        def user_streaming_features():
+            return SparkSqlTransformation("SELECT user_id, reg_country, reg_date FROM data_source")
+    """
+
+    def decorator(function):
+        user_transformation = function()
+        FeaturesetUtils.validate_base_featureset_decorator(
+            user_transformation=user_transformation, entity=entity, key=key
+        )
+
+        FeaturesetUtils.validate_streaming_featureset_decorator(
+            online_trigger_interval=online_trigger_interval,
+            offline_scheduling_policy=offline_scheduling_policy,
+        )
+
+        streaming_backfill: Optional[
+            StreamingBackfill
+        ] = StreamingBackfill.get_streaming_backfill_from_function(function=function)
+
+        fs_name = name or function.__name__
+        streaming_feature_set = StreamingFeatureSet(
+            name=fs_name,
+            entity=entity if entity else None,
+            key=key if key else None,
+            data_sources=data_sources,
+            timestamp_column_name=timestamp_column_name,
+            transformation=user_transformation,
+            metadata=get_metadata_from_function(
+                function, description=fs_name, display_name=fs_name
+            ),
+            online_trigger_interval=(
+                online_trigger_interval if online_trigger_interval else 5
+            ),
+            offline_scheduling_policy=(
+                offline_scheduling_policy
+                if offline_scheduling_policy
+                else "*/30 * * * *"
+            ),
+            offline_cluster_template=getattr(
+                function, _OFFLINE_CLUSTER_SPEC, ClusterTemplate.SMALL
+            ),
+            online_cluster_template=getattr(
+                function, _ONLINE_CLUSTER_SPEC, ClusterTemplate.SMALL
+            ),
+            backfill=streaming_backfill,
+            __instance_module_path__=inspect.stack()[1].filename,
+            auxiliary_sinks=auxiliary_sinks,
+        )
+
+        functools.update_wrapper(streaming_feature_set, user_transformation)
+        return streaming_feature_set
+
+    return decorator
+
+
+@typechecked
+def execution_specification(
+    *,
+    online_cluster_template: Optional[ClusterTemplate] = None,
+    offline_cluster_template: Optional[ClusterTemplate] = None,
+):
+    """
+    Set the execution specification of the cluster running the feature set
+
+    :param online_cluster_template: Predefined template sizes
+    :param offline_cluster_template: Predefined template sizes
+
+    Cluster template example:
+
+    ... code-block:: python
+        @streaming.feature_set(entity="users", data_sources=["streaming_users_source"])
+        @streaming.execution_specification(
+                offline_cluster_template=ClusterTemplate.MEDIUM,
+                online_cluster_template=ClusterTemplate.MEDIUM)
+        def user_streaming_features():
+            return SparkSqlTransformation("SELECT user_id, age, timestamp FROM streaming_users_source"
+    """
+
+    def decorator(user_transformation):
+        setattr(user_transformation, _ONLINE_CLUSTER_SPEC, online_cluster_template)
+
+        setattr(user_transformation, _OFFLINE_CLUSTER_SPEC, offline_cluster_template)
+
+        return user_transformation
+
+    return decorator
+
+
+@typechecked
+def backfill(
+    *,
+    start_date: datetime,
+    end_date: datetime,
+    data_sources: Union[List[str], List[BackfillBatchDataSourceSpec]],
+    backfill_transformation: SparkSqlTransformation,
+    backfill_cluster_template: Optional[ClusterTemplate] = ClusterTemplate.SMALL,
+):
+    """
+    Configures optional backfill specification for a streaming featureset. Currently available for streaming
+    aggregation featuresets only.
+
+    :param start_date: backfill start date
+    :param end_date: backfill end date
+    :param data_sources: a list of dict representations connecting each batch source name to its requested time range
+    :param backfill_transformation: a backfill SQL transformation. Only SQL transformations are supported for backfill
+    :param backfill_cluster_template: an optional cluster specification for the backfill job. Defaults to SMALL.
+
+    Example:
+
+    ... code-block:: python
+
+        @streaming.feature_set(
+            entity="users",
+            data_sources=["users_registration_stream"],
+            timestamp_column_name="reg_date"
+        )
+        @streaming.backfill(
+            start_date=start_date=(2022,01,01,0,0,0),
+            end_date=start_date=(2023,09,01,0,0,0),
+            data_sources=[{"batch_backfill_source_name": BackfillSourceRange(start_date=(2023,01,01,0,0,0),
+                                                                             end_date=(2023,08,01,0,0,0))}],
+            backfill_cluster_template=ClusterTemplate.SMALL
+            backfill_transformation=SparkSqlTransformation("SELECT user_id, reg_country, reg_date FROM backfill_data_source")
+        )
+        def user_streaming_features():
+            return SparkSqlTransformation("SELECT user_id, reg_country, reg_date FROM data_source")
+    """
+
+    def decorator(function):
+        _validate_decorator_ordering(function)
+        StreamingBackfill.set_streaming_backfill_on_function(
+            function=function,
+            start_date=start_date,
+            end_date=end_date,
+            data_sources=data_sources,
+            backfill_transformation=backfill_transformation,
+            backfill_cluster_template=backfill_cluster_template,
+        )
+
+        return function
+
+    return decorator
+
+
+@typechecked
+def metadata(
+    *,
+    owner: Optional[str] = None,
+    description: Optional[str] = None,
+    display_name: Optional[str] = None,
+    version_comment: Optional[str] = None,
+):
+    """
+    Sets additional user provided metadata
+
+    :param owner: feature set owner
+    :param description: General description of the feature set
+    :param display_name: Human readable name of the feature set
+    :param version_comment: Comment which describes the version
+
+    Example:
+
+    ... code-block:: python
+
+        @streaming.feature_set(
+            entity="users",
+            data_sources=["users_registration_stream"],
+            timestamp_column_name="reg_date"
+        )
+        @streaming.metadata(
+            owner="datainfra@frogml.com",
+            display_name="User Streaming Features",
+            description="Users feature from the Kafka topic of users registration stream",
+        )
+        def user_streaming_features():
+            return SparkSqlTransformation("SELECT user_id, reg_country, reg_date FROM data_source")
+
+    """
+
+    def decorator(function):
+        _validate_decorator_ordering(function)
+        set_metadata_on_function(
+            function=function,
+            owner=owner,
+            description=description,
+            display_name=display_name,
+            version_comment=version_comment,
+        )
+
+        return function
+
+    return decorator
+
+
+def _validate_decorator_ordering(function):
+    if isinstance(function, StreamingFeatureSet):
+        raise ValueError(
+            "Wrong decorator ordering - @streaming.feature_set should be the top most decorator"
+        )
+
+
+@dataclass
+class StreamingFeatureSet(BaseFeatureSet):
+    timestamp_column_name: str = str()
+    online_trigger_interval: int = int()
+    offline_scheduling_policy: str = str()
+    transformation: Optional[BaseTransformation] = None
+    offline_cluster_template: Optional[ClusterTemplate] = None
+    online_cluster_template: Optional[ClusterTemplate] = None
+    metadata: Optional[Metadata] = None
+    backfill: Optional[StreamingBackfill] = None
+    auxiliary_sinks: List[BaseSink] = field(default_factory=lambda: [])
+
+    def __post_init__(self):
+        self._validate()
+
+    @classmethod
+    def _from_proto(cls, proto: ProtoFeatureSetSpec):
+        streaming_def: ProtoStreamingFeatureSetV1 = (
+            proto.feature_set_type.streaming_feature_set_v1
+        )
+
+        return cls(
+            name=proto.name,
+            entity=Entity._from_proto(proto.entity).name,
+            data_sources=[ds.name for ds in streaming_def.data_sources],
+            timestamp_column_name=streaming_def.timestamp_column_name,
+            online_trigger_interval=streaming_def.online_trigger_interval,
+            offline_scheduling_policy=streaming_def.offline_scheduling_policy,
+            transformation=BaseTransformation._from_proto(streaming_def.transformation),
+            offline_cluster_template=ClusterTemplate.from_cluster_template_number(
+                streaming_def.execution_spec.offline_cluster_template
+            ),
+            online_cluster_template=ClusterTemplate.from_cluster_template_number(
+                streaming_def.execution_spec.online_cluster_template
+            ),
+            metadata=Metadata.from_proto(proto.metadata),
+            auxiliary_sinks=[
+                StreamingSinkFactory.get_streaming_sink(proto)
+                for proto in streaming_def.auxiliary_sinks
+            ],
+        )
+
+    def _get_data_sources(
+        self, feature_registry: FeatureRegistryClient
+    ) -> List[ProtoStreamingSource]:
+        sources: List[ProtoStreamingSource] = list()
+
+        for name in self.data_sources:
+            ds = feature_registry.get_data_source_by_name(name)
+            if not ds:
+                raise FrogmlException(
+                    f"Trying to register a featureset with a non existing data source {name}"
+                )
+            else:
+                sources.append(
+                    ds.data_source.data_source_definition.data_source_spec.stream_source
+                )
+        return sources
+
+    def _to_proto(
+        self,
+        git_commit,
+        features,
+        feature_registry: FeatureRegistryClient,
+        artifact_url: Optional[str] = None,
+        **kwargs,
+    ) -> Tuple[ProtoFeatureSetSpec, Optional[str]]:
+        maybe_initial_tile_size: Optional[int] = self._validate_streaming_aggregation()
+
+        data_sources: List[StreamingSource] = self._get_data_sources(feature_registry)
+
+        if not artifact_url:
+            artifact: Optional[ArtifactSpec] = ArtifactsUploader.get_artifact_spec(
+                transformation=self.transformation,
+                featureset_name=self.name,
+                __instance_module_path__=self.__instance_module_path__,
+            )
+            if artifact:
+                artifact_url = ArtifactsUploader.upload(artifact)
+
+        proto_featureset_type: ProtoFeatureSetType
+        if maybe_initial_tile_size is None:
+            # row-level streaming
+            proto_featureset_type = self._get_streaming_featureset_proto(
+                artifact_url=artifact_url, streaming_sources=data_sources
+            )
+        else:
+            # streaming aggregation
+            proto_featureset_type = self._get_streaming_aggregation_featureset_proto(
+                artifact_url=artifact_url,
+                streaming_sources=data_sources,
+                feature_registry=feature_registry,
+                initial_tile_size=maybe_initial_tile_size,
+            )
+
+        return (
+            ProtoFeatureSetSpec(
+                name=self.name,
+                metadata=self.metadata.to_proto(),
+                git_commit=git_commit,
+                features=features,
+                entity=self._get_entity_definition(feature_registry),
+                feature_set_type=proto_featureset_type,
+            ),
+            artifact_url,
+        )
+
+    def _validate(self):
+        import croniter
+
+        super()._validate()
+
+        # verify the cron expression is valid
+        if not self.offline_scheduling_policy or not croniter.croniter.is_valid(
+            self.offline_scheduling_policy
+        ):
+            raise FrogmlException(
+                f"offline scheduling policy "
+                f"'{self.offline_scheduling_policy}'"
+                f" is not a valid cron expression"
+            )
+
+        # verify the online scheduling policy is valid
+        if self.online_trigger_interval < 0:
+            raise FrogmlException(
+                f"Value '{self.online_trigger_interval}'"
+                f" is not a legal online scheduling policy, "
+                f"only non-negative integers are allowed"
+            )
+
+        # verify timestamp_col_name was set
+        if not self.timestamp_column_name:
+            raise FrogmlException("'timestamp_col_name' field must be set")
+
+        is_streaming_agg = bool(self._validate_streaming_aggregation())
+
+        # if sinks were configured, make sure it's not streaming-agg
+        if len(self.auxiliary_sinks) > 0 and is_streaming_agg:
+            raise FrogmlException(
+                "Auxiliary Sinks Are not supported in Streaming Aggregation Feature Sets"
+            )
+
+        # streaming backfill not yet supported for row-level streaming
+        if self.backfill and not is_streaming_agg:
+            raise FrogmlException(
+                "Streaming backfill is only supported for streaming aggregation feature sets at the moment"
+            )
+
+        # Validate transformation is PySpark when multiple data sources are used
+        if len(self.data_sources) > 1 and not isinstance(
+            self.transformation, PySparkTransformation
+        ):
+            raise FrogmlException(
+                "When using multiple data sources, only `PySparkTransformation` is allowed."
+            )
+
+    def _validate_streaming_aggregation(self) -> Optional[int]:
+        if not (self.transformation.windows or self.transformation.aggregations):
+            # definitely not streaming aggregation
+            return None
+
+        # at least 1 window and/or at least 1 aggregate - so we need to verify
+        # it's a valid streaming aggregation definition
+        if not self.transformation.windows:
+            raise FrogmlException(
+                "When specifying aggregations, at least one time window must be defined"
+            )
+
+        if not self.transformation.aggregations:
+            raise FrogmlException(
+                "When specifying time windows, at least one aggregation must be defined"
+            )
+
+        (
+            initial_tile_size,
+            min_window,
+        ) = StreamingFeatureSet._get_default_slide_period(self.transformation.windows)
+
+        if initial_tile_size > max(10.0, (min_window / 10)):
+            raise FrogmlException(
+                "Windows with such different cardinality (For example years vs days) can't be in "
+                "the same feature set. Please separate to different feature sets."
+            )
+
+        # Validate no feature duplications exist
+
+        # Get feature names, with or without aliases
+        final_feature_names = self.transformation.get_features_names()
+
+        # Look for duplicated feature names
+        group_by_feature_name_to_count = collections.Counter(final_feature_names)
+        features_with_duplicates = [
+            feature
+            for feature, count in group_by_feature_name_to_count.items()
+            if count > 1
+        ]
+        if features_with_duplicates:
+            error_message_str = ""
+            for dup in features_with_duplicates:
+                error_message_str += (
+                    f"{repr(dup)} feature, appears more than once in Aggregations.\n"
+                )
+            raise FrogmlException(error_message_str)
+
+        # Validate the backfill, if defined
+        if self.backfill:
+            self.backfill._validate_tile_size(initial_tile_size)
+
+        return initial_tile_size
+
+    def _validate_streaming_aggregation_backfill(self):
+        initial_tile_size, _ = StreamingFeatureSet._get_default_slide_period(
+            self.transformation.windows
+        )
+
+        self.backfill._validate_tile_size(initial_tile_size)
+
+    @staticmethod
+    def _get_default_slide_period(
+        windows: List[Window], minimum_time_window_seconds: int = 10
+    ):
+        """
+        Tile size is the max between the `minimum_time_window_seconds`,
+        and the round down of the max window size divided by 2048.
+        By dividing by 2048 we promise that the max aggregation time contains
+        up to 2048 different tiles, which limit the amount of keys that exists in the Online Store.
+        :param windows:
+        :param minimum_time_window_seconds:
+        :return: tuple of tile size and the minimum window size seconds
+        """
+        time_windows_in_seconds = [w.length * w.seconds_in_time_unit for w in windows]
+        max_window = max(time_windows_in_seconds)
+        min_window = min(time_windows_in_seconds)
+
+        if min_window < minimum_time_window_seconds:
+            raise FrogmlException(
+                f"The minimum time window is {minimum_time_window_seconds} seconds"
+            )
+
+        def round(k):
+            return k - (k % minimum_time_window_seconds)
+
+        initial_tile_size = round(max((max_window / 2048), minimum_time_window_seconds))
+
+        return int(initial_tile_size), min_window
+
+    def get_sample(
+        self,
+        number_of_rows: int = 10,
+        validation_options: Optional[FeatureSetValidationOptions] = None,
+    ) -> "pd.DataFrame":
+        return super().get_sample(
+            number_of_rows=number_of_rows, validation_options=validation_options
+        )
+
+    def _get_streaming_featureset_proto(
+        self, artifact_url: Optional[str], streaming_sources: List[StreamingSource]
+    ) -> ProtoFeatureSetType:
+        return ProtoFeatureSetType(
+            streaming_feature_set_v1=ProtoStreamingFeatureSetV1(
+                transformation=self.transformation._to_proto(
+                    artifact_path=artifact_url
+                ),
+                data_sources=streaming_sources,
+                execution_spec=ProtoStreamingExecutionSpec(
+                    online_cluster_template=ClusterTemplate.to_proto(
+                        self.online_cluster_template
+                    ),
+                    offline_cluster_template=ClusterTemplate.to_proto(
+                        self.offline_cluster_template
+                    ),
+                ),
+                timestamp_column_name=self.timestamp_column_name,
+                online_trigger_interval=self.online_trigger_interval,
+                offline_scheduling_policy=self.offline_scheduling_policy,
+                auxiliary_sinks=[
+                    s.to_proto_streaming_sink() for s in self.auxiliary_sinks
+                ],
+            )
+        )
+
+    def _get_streaming_aggregation_featureset_proto(
+        self,
+        artifact_url: Optional[str],
+        streaming_sources: List[StreamingSource],
+        feature_registry: FeatureRegistryClient,
+        initial_tile_size: int,
+    ) -> ProtoFeatureSetType:
+        return ProtoFeatureSetType(
+            streaming_aggregation_feature_set=ProtoStreamingAggregationFeatureSet(
+                transformation=self.transformation._to_proto(
+                    artifact_path=artifact_url
+                ),
+                data_sources=streaming_sources,
+                execution_spec=ProtoStreamingExecutionSpec(
+                    online_cluster_template=ClusterTemplate.to_proto(
+                        self.online_cluster_template
+                    ),
+                    offline_cluster_template=ClusterTemplate.to_proto(
+                        self.offline_cluster_template
+                    ),
+                ),
+                timestamp_column_name=self.timestamp_column_name,
+                online_trigger_interval=self.online_trigger_interval,
+                compaction_scheduling_policy=self.offline_scheduling_policy,
+                aggregation_spec=ProtoAggregationSpec(
+                    slide_seconds=initial_tile_size,
+                    allowed_late_arrival_seconds=60 * 10,
+                    aggregations=self.transformation._get_aggregations_proto(),
+                ),
+                backfill_spec=(
+                    self.backfill._to_proto(
+                        feature_registry=feature_registry,
+                        original_instance_module_path=self.__instance_module_path__,
+                        featureset_name=self.name,
+                    )
+                    if self.backfill
+                    else None
+                ),
+            )
+        )