get_historical_features raises TypeError on a zero-row entity_df (Dask/file offline store)

Expected Behavior

get_historical_features with an entity_df that happens to have zero rows should return an empty result with the right columns. A zero-row entity frame is a normal degenerate case in batch scoring — the upstream query simply matched nothing that run.

Current Behavior

It raises an opaque pandas error from deep inside the Dask offline store:

TypeError: Invalid comparison between dtype=datetime64[ns] and DatetimeArray

Traceback tail:

  File "sdk/python/feast/infra/offline_stores/offline_store.py", line 178, in to_arrow
    features_table = self._to_arrow_internal(timeout=timeout)
  File "sdk/python/feast/infra/offline_stores/dask.py", line 103, in _to_arrow_internal
    df = self.evaluation_function().compute()
  File "sdk/python/feast/infra/offline_stores/dask.py", line 317, in evaluate_historical_retrieval
    df_to_join = _filter_ttl(
  File "sdk/python/feast/infra/offline_stores/dask.py", line 1206, in _filter_ttl
    df_to_join = df_to_join.persist()

Steps to reproduce

Identical construction, only the row count differs, against a file offline store:

def mk(n):
    return pd.DataFrame({
        "driver_id": [1001] * n,
        "event_timestamp": pd.to_datetime([datetime(2026, 2, 1)] * n, utc=True),
    })

fs.get_historical_features(entity_df=mk(1), features=["driver_stats:conv_rate"]).to_df()
# OK, shape=(1, 3)

fs.get_historical_features(entity_df=mk(0), features=["driver_stats:conv_rate"]).to_df()
# TypeError: Invalid comparison between dtype=datetime64[ns] and DatetimeArray

Specifications

• Version: master @ 5ad5592390febfca60c9d88edf7daccbdd156fd6
• Platform: Linux x86_64, Python 3.11.15, dask 2026.8.0
• Subsystem: offline store (Dask / file)

Possible Solution

_normalize_timestamp (sdk/python/feast/infra/offline_stores/dask.py:1139) makes timestamp columns tz-aware with a row-wise apply:

df_to_join[timestamp_field] = df_to_join[timestamp_field].apply(
    lambda x: x if x.tzinfo else x.replace(tzinfo=timezone.utc),
    meta=(timestamp_field, "datetime64[ns, UTC]"),
)

meta declares a tz-aware result, but with zero rows the lambda never runs, so the computed partition stays datetime64[ns]. Declared and actual dtypes then diverge, and the tz-naive vs tz-aware comparison in _filter_ttl (sdk/python/feast/infra/offline_stores/dask.py:1183) raises.

Replacing the row-wise apply with a vectorized, empty-safe conversion fixes it and is faster on non-empty frames too: pick dt.tz_localize("UTC") for a tz-naive column and dt.tz_convert("UTC") otherwise, based on the column's declared dtype. That yields the correct dtype even when the partition is empty.

A regression test covering a zero-row entity_df through get_historical_features(...).to_df() on the Dask store would pin this down; the current suite only exercises non-empty entity frames, which is why CI stays green.

Happy to send a PR for this.