ntkathole · Jun 30, 2026 · Jun 29, 2026 · Jun 28, 2026 · Jun 10, 2026 · Jun 29, 2026
@@ -10,11 +10,13 @@
     ContextManager,
     Dict,
     Iterator,
+    KeysView,
     List,
     Literal,
     Optional,
     Tuple,
     Union,
+    cast,
 )
 
 import numpy as np
@@ -57,7 +59,7 @@
 from feast.on_demand_feature_view import OnDemandFeatureView
 from feast.repo_config import FeastConfigBaseModel, RepoConfig
 from feast.saved_dataset import SavedDatasetStorage
-from feast.utils import _utc_now, get_user_agent
+from feast.utils import _utc_now, compute_non_entity_date_range, get_user_agent
 
 from .bigquery_source import (
     BigQueryLoggingDestination,
@@ -267,10 +269,11 @@ def get_historical_features(
         config: RepoConfig,
         feature_views: List[FeatureView],
         feature_refs: List[str],
-        entity_df: Union[pd.DataFrame, str],
+        entity_df: Optional[Union[pd.DataFrame, str]],
         registry: BaseRegistry,
         project: str,
         full_feature_names: bool = False,
+        **kwargs: Any,
     ) -> RetrievalJob:
         # TODO: Add entity_df validation in order to fail before interacting with BigQuery
         assert isinstance(config.offline_store, BigQueryOfflineStoreConfig)
@@ -297,36 +300,79 @@ def get_historical_features(
             config.offline_store.table_create_disposition,
         )
 
-        entity_schema = _get_entity_schema(
-            client=client,
-            entity_df=entity_df,
-        )
+        # Non-entity mode: create a left temporary table from entity keys - any entity key having an event in the time window
 
-        entity_df_event_timestamp_col = (
-            offline_utils.infer_event_timestamp_from_entity_df(entity_schema)
-        )
+        non_entity_mode = entity_df is None
 
-        entity_df_event_timestamp_range = _get_entity_df_event_timestamp_range(
-            entity_df,
-            entity_df_event_timestamp_col,
-            client,
-        )
+        if non_entity_mode:
+            start_date, end_date = compute_non_entity_date_range(
+                feature_views,
+                start_date=kwargs.get("start_date"),
+                end_date=kwargs.get("end_date"),
+            )
+            entity_df_event_timestamp_range = (start_date, end_date)
 
-        @contextlib.contextmanager
-        def query_generator() -> Iterator[str]:
-            _upload_entity_df(
+            # Pre-compute query contexts to collect entity column names per feature view.
+            fv_query_contexts_pre = offline_utils.get_feature_view_query_context(
+                feature_refs,
+                feature_views,
+                registry,
+                project,
+                entity_df_event_timestamp_range,
+            )
+            all_entities = offline_utils.gather_all_entities(fv_query_contexts_pre)
+            event_timestamp_col = "entity_ts"
+            entity_schema_keys: KeysView[str] = cast(
+                KeysView[str],
+                {k: None for k in (all_entities + [event_timestamp_col])}.keys(),
+            )
+            entity_schema = None
+        else:
+            entity_schema = _get_entity_schema(
                 client=client,
-                table_name=table_reference,
                 entity_df=entity_df,
             )
-
-            expected_join_keys = offline_utils.get_expected_join_keys(
-                project, feature_views, registry
+            event_timestamp_col = offline_utils.infer_event_timestamp_from_entity_df(
+                entity_schema
             )
-
-            offline_utils.assert_expected_columns_in_entity_df(
-                entity_schema, expected_join_keys, entity_df_event_timestamp_col
+            entity_df_event_timestamp_range = _get_entity_df_event_timestamp_range(
+                entity_df,
+                event_timestamp_col,
+                client,
             )
+            entity_schema_keys = entity_schema.keys()
+            all_entities = []
+            fv_query_contexts_pre = None
+            start_date = entity_df_event_timestamp_range[0]
+            end_date = entity_df_event_timestamp_range[1]
+
+        @contextlib.contextmanager
+        def query_generator() -> Iterator[str]:
+            if non_entity_mode:
+                assert fv_query_contexts_pre is not None
+                _bq_create_entity_union_table(
+                    client=client,
+                    table_name=table_reference,
+                    feature_views=feature_views,
+                    fv_query_contexts=fv_query_contexts_pre,
+                    start_date=start_date,
+                    end_date=end_date,
+                    all_entities=all_entities,
+                    event_timestamp_col=event_timestamp_col,
+                )
+            else:
+                _upload_entity_df(
+                    client=client,
+                    table_name=table_reference,
+                    entity_df=entity_df,
+                )
+                expected_join_keys = offline_utils.get_expected_join_keys(
+                    project, feature_views, registry
+                )
+                assert entity_schema is not None
+                offline_utils.assert_expected_columns_in_entity_df(
+                    entity_schema, expected_join_keys, event_timestamp_col
+                )
 
             # Build a query context containing all information required to template the BigQuery SQL query
             query_context = offline_utils.get_feature_view_query_context(
@@ -341,16 +387,16 @@ def query_generator() -> Iterator[str]:
             query = offline_utils.build_point_in_time_query(
                 query_context,
                 left_table_query_string=table_reference,
-                entity_df_event_timestamp_col=entity_df_event_timestamp_col,
-                entity_df_columns=entity_schema.keys(),
+                entity_df_event_timestamp_col=event_timestamp_col,
+                entity_df_columns=entity_schema_keys,
                 query_template=MULTIPLE_FEATURE_VIEW_POINT_IN_TIME_JOIN,
                 full_feature_names=full_feature_names,
             )
 
             try:
                 yield query
             finally:
-                # Asynchronously clean up the uploaded Bigquery table, which will expire
+                # Asynchronously clean up the uploaded BigQuery table, which will expire
                 # if cleanup fails
                 client.delete_table(table=table_reference, not_found_ok=True)
 
@@ -364,7 +410,7 @@ def query_generator() -> Iterator[str]:
             ),
             metadata=RetrievalMetadata(
                 features=feature_refs,
-                keys=list(entity_schema.keys() - {entity_df_event_timestamp_col}),
+                keys=list(set(entity_schema_keys) - {event_timestamp_col}),
                 min_event_timestamp=entity_df_event_timestamp_range[0],
                 max_event_timestamp=entity_df_event_timestamp_range[1],
             ),
@@ -567,6 +613,62 @@ def clear_monitoring_baseline(
         )
 
 
+def _bq_create_entity_union_table(
+    client: "Client",
+    table_name: str,
+    feature_views: List[FeatureView],
+    fv_query_contexts: List[offline_utils.FeatureViewQueryContext],
+    start_date: datetime,
+    end_date: datetime,
+    all_entities: List[str],
+    event_timestamp_col: str,
+) -> None:
+    """
+    Creates a BigQuery temp table containing the UNION DISTINCT of entity keys observed
+    across all feature views in [start_date, end_date], plus a stable as-of timestamp
+    column set to end_date. Used as the left table for PIT joins in non-entity mode.
+    """
+    start_str = start_date.strftime("%Y-%m-%dT%H:%M:%S")
+    end_str = end_date.strftime("%Y-%m-%dT%H:%M:%S")
+
+    per_view_selects: List[str] = []
+    for fv, ctx in zip(feature_views, fv_query_contexts):
+        assert isinstance(fv.batch_source, BigQuerySource)
+        from_expression = fv.batch_source.get_table_query_string()
+        timestamp_field = ctx.timestamp_field
+
+        ctx_entities_set = set(ctx.entities)
+        select_entities: List[str] = []
+        for col in all_entities:
+            if col in ctx_entities_set:
+                select_entities.append(f"`{col}`")
+            else:
+                select_entities.append(f"NULL AS `{col}`")
+
+        per_view_selects.append(
+            f"SELECT DISTINCT {', '.join(select_entities)} "
+            f"FROM {from_expression} "
+            f"WHERE `{timestamp_field}` BETWEEN TIMESTAMP('{start_str}') AND TIMESTAMP('{end_str}')"
+        )
+
+    union_query = "\nUNION DISTINCT\n".join(per_view_selects)
+    entity_cols = (
+        ", ".join(f"`{e}`" for e in all_entities) if all_entities else "TRUE AS _dummy"
+    )
+
+    create_sql = (
+        f"CREATE TABLE `{table_name}` AS "
+        f"SELECT {entity_cols}, TIMESTAMP('{end_str}') AS `{event_timestamp_col}` "
+        f"FROM ({union_query}) AS _entity_union"
+    )
+
+    block_until_done(client, client.query(create_sql))
+
+    table = client.get_table(table_name)
+    table.expires = _utc_now() + timedelta(minutes=30)
+    client.update_table(table, ["expires"])
+
+
 # ------------------------------------------------------------------ #
 #  BigQuery monitoring metrics (native)
 # ------------------------------------------------------------------ #

@@ -429,3 +429,12 @@ def format_date(val: Union[str, datetime]) -> str:
             filters.append(f"{dp_field} <= '{format_date(end_date)}'")
 
     return " AND ".join(filters) if filters else ""
+
+
+def gather_all_entities(fv_query_contexts: List[FeatureViewQueryContext]):
+    all_entities: List[str] = []
+    for ctx in fv_query_contexts:
+        for e in ctx.entities:
+            if e not in all_entities:
+                all_entities.append(e)
+    return all_entities