feat: Validating logged features via Python SDK (#2640)

* simple logged feature validation

Signed-off-by: Oleksii Moskalenko <[email protected]>

* validate with metadata

Signed-off-by: Oleksii Moskalenko <[email protected]>

* typos

Signed-off-by: Oleksii Moskalenko <[email protected]>

* revert entity columns

Signed-off-by: Oleksii Moskalenko <[email protected]>

Author: pyalex

sdk/python/feast/dqm/profilers/ge_profiler.py

@@ -21,6 +21,7 @@
 from feast.protos.feast.core.ValidationProfile_pb2 import (
     GEValidationProfiler as GEValidationProfilerProto,
 )
+from feast.protos.feast.serving.ServingService_pb2 import FieldStatus
 
 
 def _prepare_dataset(dataset: PandasDataset) -> PandasDataset:
@@ -41,6 +42,23 @@ def _prepare_dataset(dataset: PandasDataset) -> PandasDataset:
     return dataset_copy
 
 
+def _add_feature_metadata(dataset: PandasDataset) -> PandasDataset:
+    for column in dataset.columns:
+        if "__" not in column:
+            # not a feature column
+            continue
+
+        if "event_timestamp" in dataset.columns:
+            dataset[f"{column}__timestamp"] = dataset["event_timestamp"]
+
+        dataset[f"{column}__status"] = FieldStatus.PRESENT
+        dataset[f"{column}__status"] = dataset[f"{column}__status"].mask(
+            dataset[column].isna(), FieldStatus.NOT_FOUND
+        )
+
+    return dataset
+
+
 class GEProfile(Profile):
     """
     GEProfile is an implementation of abstract Profile for integration with Great Expectations.
@@ -96,9 +114,12 @@ class GEProfiler(Profiler):
     """
 
     def __init__(
-        self, user_defined_profiler: Callable[[pd.DataFrame], ExpectationSuite]
+        self,
+        user_defined_profiler: Callable[[pd.DataFrame], ExpectationSuite],
+        with_feature_metadata: bool = False,
     ):
         self.user_defined_profiler = user_defined_profiler
+        self.with_feature_metadata = with_feature_metadata
 
     def analyze_dataset(self, df: pd.DataFrame) -> Profile:
         """
@@ -113,6 +134,9 @@ def analyze_dataset(self, df: pd.DataFrame) -> Profile:
 
         dataset = _prepare_dataset(dataset)
 
+        if self.with_feature_metadata:
+            dataset = _add_feature_metadata(dataset)
+
         return GEProfile(expectation_suite=self.user_defined_profiler(dataset))
 
     def to_proto(self):
@@ -158,5 +182,13 @@ def __repr__(self):
         return json.dumps(failed_expectations, indent=2)
 
 
-def ge_profiler(func):
-    return GEProfiler(user_defined_profiler=func)
+def ge_profiler(*args, with_feature_metadata=False):
+    def wrapper(fun):
+        return GEProfiler(
+            user_defined_profiler=fun, with_feature_metadata=with_feature_metadata
+        )
+
+    if args:
+        return wrapper(args[0])
+
+    return wrapper

sdk/python/feast/feature_logging.py

@@ -11,6 +11,7 @@
     FeatureViewNotFoundException,
     OnDemandFeatureViewNotFoundException,
 )
+from feast.feature_view import DUMMY_ENTITY_NAME
 from feast.protos.feast.core.FeatureService_pb2 import (
     LoggingConfig as LoggingConfigProto,
 )
@@ -77,7 +78,11 @@ def get_schema(self, registry: "Registry") -> pa.Schema:
 
             else:
                 for entity_name in feature_view.entities:
+                    if entity_name == DUMMY_ENTITY_NAME:
+                        continue
+
                     entity = registry.get_entity(entity_name, self._project)
+
                     join_key = projection.join_key_map.get(
                         entity.join_key, entity.join_key
                     )

sdk/python/feast/feature_store.py

@@ -45,6 +45,7 @@
 from feast.data_source import DataSource
 from feast.diff.infra_diff import InfraDiff, diff_infra_protos
 from feast.diff.registry_diff import RegistryDiff, apply_diff_to_registry, diff_between
+from feast.dqm.errors import ValidationFailed
 from feast.entity import Entity
 from feast.errors import (
     EntityNotFoundException,
@@ -83,7 +84,7 @@
 from feast.repo_config import RepoConfig, load_repo_config
 from feast.repo_contents import RepoContents
 from feast.request_feature_view import RequestFeatureView
-from feast.saved_dataset import SavedDataset, SavedDatasetStorage
+from feast.saved_dataset import SavedDataset, SavedDatasetStorage, ValidationReference
 from feast.type_map import (
     feast_value_type_to_python_type,
     python_values_to_proto_values,
@@ -2054,6 +2055,58 @@ def write_logged_features(
             registry=self._registry,
         )
 
+    def validate_logged_features(
+        self,
+        source: Union[FeatureService],
+        start: datetime,
+        end: datetime,
+        reference: ValidationReference,
+        throw_exception: bool = True,
+    ) -> Optional[ValidationFailed]:
+        """
+        Load logged features from an offline store and validate them against provided validation reference.
+
+        Args:
+            source: Logs source object (currently only feature services are supported)
+            start: lower bound for loading logged features
+            end:  upper bound for loading logged features
+            reference: validation reference
+            throw_exception: throw exception or return it as a result
+
+        Returns:
+            Throw or return (depends on parameter) ValidationFailed exception if validation was not successful
+            or None if successful.
+
+        """
+        warnings.warn(
+            "Logged features validation is an experimental feature. "
+            "This API is unstable and it could and most probably will be changed in the future. "
+            "We do not guarantee that future changes will maintain backward compatibility.",
+            RuntimeWarning,
+        )
+
+        if not isinstance(source, FeatureService):
+            raise ValueError("Only feature service is currently supported as a source")
+
+        j = self._get_provider().retrieve_feature_service_logs(
+            feature_service=source,
+            start_date=start,
+            end_date=end,
+            config=self.config,
+            registry=self.registry,
+        )
+
+        # read and run validation
+        try:
+            j.to_arrow(validation_reference=reference)
+        except ValidationFailed as exc:
+            if throw_exception:
+                raise
+
+            return exc
+
+        return None
+
 
 def _validate_entity_values(join_key_values: Dict[str, List[Value]]):
     set_of_row_lengths = {len(v) for v in join_key_values.values()}

sdk/python/feast/feature_view.py

@@ -285,6 +285,7 @@ def __copy__(self):
             online=self.online,
         )
         fv.projection = copy.copy(self.projection)
+        fv.entities = self.entities
         return fv
 
     def __eq__(self, other):

sdk/python/feast/infra/offline_stores/file.py

@@ -73,6 +73,7 @@ def on_demand_feature_views(self) -> Optional[List[OnDemandFeatureView]]:
     def _to_df_internal(self) -> pd.DataFrame:
         # Only execute the evaluation function to build the final historical retrieval dataframe at the last moment.
         df = self.evaluation_function().compute()
+        df = df.reset_index(drop=True)
         return df
 
     @log_exceptions_and_usage
@@ -555,11 +556,18 @@ def _filter_ttl(
     # Filter rows by defined timestamp tolerance
     if feature_view.ttl and feature_view.ttl.total_seconds() != 0:
         df_to_join = df_to_join[
-            (
-                df_to_join[timestamp_field]
-                >= df_to_join[entity_df_event_timestamp_col] - feature_view.ttl
+            # do not drop entity rows if one of the sources returns NaNs
+            df_to_join[timestamp_field].isna()
+            | (
+                (
+                    df_to_join[timestamp_field]
+                    >= df_to_join[entity_df_event_timestamp_col] - feature_view.ttl
+                )
+                & (
+                    df_to_join[timestamp_field]
+                    <= df_to_join[entity_df_event_timestamp_col]
+                )
             )
-            & (df_to_join[timestamp_field] <= df_to_join[entity_df_event_timestamp_col])
         ]
 
         df_to_join = df_to_join.persist()

sdk/python/tests/integration/e2e/test_validation.py

@@ -1,10 +1,21 @@
+import datetime
+
 import pandas as pd
+import pyarrow as pa
 import pytest
 from great_expectations.core import ExpectationSuite
 from great_expectations.dataset import PandasDataset
 
+from feast import FeatureService
 from feast.dqm.errors import ValidationFailed
 from feast.dqm.profilers.ge_profiler import ge_profiler
+from feast.feature_logging import (
+    LOG_TIMESTAMP_FIELD,
+    FeatureServiceLoggingSource,
+    LoggingConfig,
+)
+from feast.protos.feast.serving.ServingService_pb2 import FieldStatus
+from feast.wait import wait_retry_backoff
 from tests.integration.feature_repos.repo_configuration import (
     construct_universal_feature_views,
 )
@@ -13,6 +24,7 @@
     driver,
     location,
 )
+from tests.utils.logged_features import prepare_logs
 
 _features = [
     "customer_profile:current_balance",
@@ -32,6 +44,39 @@ def configurable_profiler(dataset: PandasDataset) -> ExpectationSuite:
 
     return UserConfigurableProfiler(
         profile_dataset=dataset,
+        ignored_columns=["event_timestamp"],
+        excluded_expectations=[
+            "expect_table_columns_to_match_ordered_list",
+            "expect_table_row_count_to_be_between",
+        ],
+        value_set_threshold="few",
+    ).build_suite()
+
+
+@ge_profiler(with_feature_metadata=True)
+def profiler_with_feature_metadata(dataset: PandasDataset) -> ExpectationSuite:
+    from great_expectations.profile.user_configurable_profiler import (
+        UserConfigurableProfiler,
+    )
+
+    # always present
+    dataset.expect_column_values_to_be_in_set(
+        "global_stats__avg_ride_length__status", {FieldStatus.PRESENT}
+    )
+
+    # present at least in 70% of rows
+    dataset.expect_column_values_to_be_in_set(
+        "customer_profile__current_balance__status", {FieldStatus.PRESENT}, mostly=0.7
+    )
+
+    return UserConfigurableProfiler(
+        profile_dataset=dataset,
+        ignored_columns=["event_timestamp"]
+        + [
+            c
+            for c in dataset.columns
+            if c.endswith("__timestamp") or c.endswith("__status")
+        ],
         excluded_expectations=[
             "expect_table_columns_to_match_ordered_list",
             "expect_table_row_count_to_be_between",
@@ -127,3 +172,88 @@ def test_historical_retrieval_fails_on_validation(environment, universal_data_so
 
     assert failed_expectations[1].check_name == "expect_column_values_to_be_in_set"
     assert failed_expectations[1].column_name == "avg_passenger_count"
+
+
+@pytest.mark.integration
+def test_logged_features_validation(environment, universal_data_sources):
+    store = environment.feature_store
+
+    (_, datasets, data_sources) = universal_data_sources
+    feature_views = construct_universal_feature_views(data_sources)
+    feature_service = FeatureService(
+        name="test_service",
+        features=[
+            feature_views.customer[
+                ["current_balance", "avg_passenger_count", "lifetime_trip_count"]
+            ],
+            feature_views.order[["order_is_success"]],
+            feature_views.global_fv[["num_rides", "avg_ride_length"]],
+        ],
+        logging_config=LoggingConfig(
+            destination=environment.data_source_creator.create_logged_features_destination()
+        ),
+    )
+
+    store.apply(
+        [driver(), customer(), location(), feature_service, *feature_views.values()]
+    )
+
+    entity_df = datasets.entity_df.drop(
+        columns=["order_id", "origin_id", "destination_id"]
+    )
+
+    # add some non-existing entities to check NotFound feature handling
+    for i in range(5):
+        entity_df = entity_df.append(
+            {
+                "customer_id": 2000 + i,
+                "driver_id": 6000 + i,
+                "event_timestamp": datetime.datetime.now(),
+            },
+            ignore_index=True,
+        )
+
+    reference_dataset = store.create_saved_dataset(
+        from_=store.get_historical_features(
+            entity_df=entity_df, features=feature_service, full_feature_names=True
+        ),
+        name="reference_for_validating_logged_features",
+        storage=environment.data_source_creator.create_saved_dataset_destination(),
+    )
+
+    log_source_df = store.get_historical_features(
+        entity_df=entity_df, features=feature_service, full_feature_names=False
+    ).to_df()
+    logs_df = prepare_logs(log_source_df, feature_service, store)
+
+    schema = FeatureServiceLoggingSource(
+        feature_service=feature_service, project=store.project
+    ).get_schema(store._registry)
+    store.write_logged_features(
+        pa.Table.from_pandas(logs_df, schema=schema), source=feature_service
+    )
+
+    def validate():
+        """
+        Return Tuple[succeed, completed]
+        Succeed will be True if no ValidateFailed exception was raised
+        """
+        try:
+            store.validate_logged_features(
+                feature_service,
+                start=logs_df[LOG_TIMESTAMP_FIELD].min(),
+                end=logs_df[LOG_TIMESTAMP_FIELD].max() + datetime.timedelta(seconds=1),
+                reference=reference_dataset.as_reference(
+                    profiler=profiler_with_feature_metadata
+                ),
+            )
+        except ValidationFailed:
+            return False, True
+        except Exception:
+            # log table is still being created
+            return False, False
+
+        return True, True
+
+    success = wait_retry_backoff(validate, timeout_secs=30)
+    assert success, "Validation failed (unexpectedly)"