Expected Behavior

Materializing a feature view that contains one or more UnixTimestamp feature columns into a remote online store should succeed.

The remote online-store client should serialize timestamp-valued feature payloads into a JSON-safe representation before calling the feature server.

Current Behavior

When online_store.type: remote is used and a feature value itself is typed as UnixTimestamp, materialization fails in the remote client before the request is sent.

The failure path is:

• RemoteOnlineStore.online_write_batch() builds req_body
• feature values are converted via _proto_value_to_transport_value()
• unix_timestamp_val falls through to feast_value_type_to_python_type(...)
• Feast returns a timezone-aware Python datetime
• requests.post(..., json=req_body) calls json.dumps(...)
• JSON encoding fails with:

TypeError: Object of type datetime is not JSON serializable

A representative feature view looks like:

schema=[
    Field(name="first_checkin_timestamp", dtype=UnixTimestamp),
    Field(name="last_checkin_timestamp", dtype=UnixTimestamp),
    Field(name="total_checkins", dtype=Int64),
    Field(name="item_type", dtype=String),
]

The source event_timestamp / created_timestamp_column do not have this problem because online_write_batch() already serializes those fields explicitly with .isoformat(). The issue only affects timestamp-valued feature columns.

This was observed with a Trino-backed feature view, but the bug appears to be backend-independent because the exception is raised in the remote online-store HTTP transport before the request reaches the server.

Steps to reproduce

1. Configure Feast with a remote online store (online_store.type: remote) pointing at a Feast feature server.
2. Define a feature view with at least one feature column of dtype UnixTimestamp.
3. Ensure the feature view can be materialized normally from the offline store.
4. Run materialization, for example:

from datetime import datetime, timezone

start = datetime(2023, 1, 1, 0, 0, tzinfo=timezone.utc)
end = datetime(2027, 1, 1, 0, 0, tzinfo=timezone.utc)

fs.materialize(
    start_date=start,
    end_date=end,
    disable_event_timestamp=True,
    feature_views=["historical_item_usage_view_v2"],
)

5. Observe the remote online-store write failure:

File ".../feast/infra/online_stores/remote.py", line ..., in online_write_batch
  response = post_remote_online_write(config=config, req_body=req_body)
...
File ".../json/encoder.py", line 180, in default
  raise TypeError(f'Object of type {o.__class__.__name__} '
TypeError: Object of type datetime is not JSON serializable

Specifications

• Version: v0.63.0
• Platform: Linux / Python 3.12
• Subsystem: remote online store / materialization / HTTP transport

Possible Solution

Handle UnixTimestamp feature values explicitly in the remote transport path instead of letting them fall through to feast_value_type_to_python_type(...).

For example, in RemoteOnlineStore._proto_value_to_transport_value() / online_write_batch():

• serialize unix_timestamp_val into a JSON-safe transport format
  • preferably epoch integer(s), since UNIX_TIMESTAMP already round-trips cleanly from numeric values
  • alternatively ISO-8601 strings, if the server-side write path also parses them back using the feature schema
• keep the existing explicit handling for event_timestamp / created

At the moment the code special-cases JSON/UUID/Map/Struct transport, but not timestamp-valued feature payloads, which makes remote materialization fail for UnixTimestamp features.