add possibility to pushdown clickhouse deduplication to offline store

Signed-off-by: lukas.valatka <lukas.valatka@cast.ai>
astronautas · Oct 15, 2025 · Oct 29, 2025 · Nov 3, 2025 · Nov 3, 2025 · Nov 3, 2025
commit f7b7d3722cfbf1db20b01d748664e9ba69dacd31
@@ -1,4 +1,5 @@
 import contextlib
+import logging
 import re
 from dataclasses import asdict
 from datetime import datetime
@@ -32,6 +33,8 @@
 from feast.infra.utils.clickhouse.connection_utils import get_client
 from feast.saved_dataset import SavedDatasetStorage
 
+logger = logging.getLogger(__name__)
+
 
 class ClickhouseOfflineStoreConfig(ClickhouseConfig):
     type: Literal["clickhouse"] = "clickhouse"
@@ -205,6 +208,29 @@ def pull_all_from_table_or_query(
         assert isinstance(config.offline_store, ClickhouseOfflineStoreConfig)
         assert isinstance(data_source, ClickhouseSource)
 
+        # https://github.com/feast-dev/feast/issues/5707
+        # Not an ideal solution, but least invasion into existing codebase
+        if config.offline_store.deduplicate_pushdown:
+            logger.warning(
+                """
+                deduplicate_pushdown optimization is set to True in ClickhouseOfflineStoreConfig.
+                Calling pull_latest_from_table_or_query instead of pull_all_from_table_or_query.
+                This results in multiple times more efficient materialization jobs for large datasets.
+                However, it comes with a caveat - can't compute historical features (just latest values available) with Feast compute engines
+                Use with caution and ensure your use case aligns with this behavior.
+                """
+            )
+            return ClickhouseOfflineStore.pull_latest_from_table_or_query(
+                config=config,
+                data_source=data_source,
+                join_key_columns=join_key_columns,
+                feature_name_columns=feature_name_columns,
+                timestamp_field=timestamp_field,
+                created_timestamp_column=created_timestamp_column,
+                start_date=start_date,
+                end_date=end_date,
+            )
+
         from_expression = data_source.get_table_query_string()
 
         timestamp_fields = [timestamp_field]

@@ -11,4 +11,9 @@ class ClickhouseConfig(FeastConfigBaseModel):
     password: StrictStr
     use_temporary_tables_for_entity_df: bool = True
 
+    # https://github.com/feast-dev/feast/issues/5707
+    # We observed that for large materialization jobs, it's multiple times more efficient
+    # pushdown deduplication to ClickHouse side rather than doing it in Feast (no matter compute engine used)
+    deduplicate_pushdown: bool = False
+
     model_config = ConfigDict(frozen=True)