Repository navigation
Feature views lack required data for offline_utils.get_feature_view_query_context #3188
Description
Expected Behavior
Context: running get_historical_features from feast.infra.offline_stores.bigquery
Version: v0.24
With arguments unchanged, lines 210-216 should still work to populate a FeatureViewQueryContext:
from feast.infra.offline_stores import offline_utils
query_context = offline_utils.get_feature_view_query_context(
feature_refs,
feature_views,
registry,
project,
entity_df_event_timestamp_range,
)This context should include an entities field for feature views with entities, that can be used to populate the query in lines 552-556. This entities field should be a list of entity join keys.
Current Behavior
Currently, even if the feature views have entities, the context's entities field is always empty.
The key problem: offline_utils.get_feature_view_query_context requires the feature view to have two attributes to populate this entities field: feature_view.entity_columns and feature_view.projection.join_key_map. However, neither of these fields are populated. These feature views have been pulled from our feature store using the function feature_store.get('feature_view_name').
Here is the piece of the code that fails to populate (join_keys is set to the entities field here):
join_keys: List[str] = []
entity_selections: List[str] = []
for entity_column in feature_view.entity_columns:
join_key = feature_view.projection.join_key_map.get(
entity_column.name, entity_column.name
)
join_keys.append(join_key)Steps to reproduce
Here is an example based on my feature store. To reproduce, use a feature view with at least one entity:
from datetime import datetime, timezone
from feast.infra.offline_stores import offline_utils
feature_refs = ['shop_stats:feature']
feature_views = [feature_store.get_feature_view('shop_stats')]
registry = feature_store.registry
project = feature_store.project
entity_df_timestamp_range = (datetime(2022, 9, 7, 12, 12, 19, tzinfo=timezone.utc),
datetime(2022, 9, 7, 12, 12, 19, tzinfo=timezone.utc))
print(example_feature_views[0].entities)
print(example_feature_views[0].entity_columns)
print(example_feature_views[0].projection.join_key_map)
query_context = offline_utils.get_feature_view_query_context(
example_feature_refs,
example_feature_views,
registry,
project,
entity_df_event_timestamp_range,
)
print(query_context)Output:
>>> ['shop']
>>> []
>>> {}
>>> [FeatureViewQueryContext(name='shop_stats', ttl=0, entities=[], features=['feature'], field_mapping={}, timestamp_field='observed_at', created_timestamp_column='', table_subquery='`<TABLE_NAME>`', entity_selections=[], min_event_timestamp=None, max_event_timestamp='2022-09-07T12:12:19', date_partition_column='')]Specifications
- Version: 0.24
- Platform:
- Subsystem:
Activity
chhabrakadabra commentedon Sep 7, 2022
Upon further examination, this is potentially not an issue. I think we were connecting to a registry that had not been applied to in version 0.24, leading to the entity_columns fields being empty.
Potentially a useful thing to consider is that perhaps in the
FeatureView.from_proto()method, iffeature_view.entitiesandfeature_view.entity_columnslook substantially different (different lengths), a warning to the user would make a huge difference.