Expected Behavior

Remote write_to_online_store should round-trip regardless of the feature view's
timestamp column names. Writing works for a local provider (the full dataframe is
converted in-process using the source's configured column names) and should behave the
same when the online store is remote (online_store: { type: remote } / the feature
server /write-to-online-store endpoint).

Current Behavior

Writing to the online store through the remote path fails on the server with:

KeyError: 'Field "created_at" does not exist in schema'

whenever the feature view's batch source sets created_timestamp_column (or
timestamp_field) to anything other than the literals "created" /
"event_timestamp".

Root cause — RemoteOnlineStore.online_write_batch serializes the timestamps into
hardcoded column names when building the request body:

# sdk/python/feast/infra/online_stores/remote.py  (online_write_batch)
columnar_data["event_timestamp"].append(_to_naive_utc(event_ts).isoformat())
columnar_data["created"].append(
    _to_naive_utc(created_ts).isoformat() if created_ts else None
)
req_body = {"feature_view_name": table.name, "df": columnar_data, ...}
post_remote_online_write(config=config, req_body=req_body)

The server rebuilds the dataframe verbatim and runs the standard conversion:

# sdk/python/feast/feature_server.py
df = pd.DataFrame(request.df)     # columns: entities, features, event_timestamp, created
store.write_to_online_store(..., df=df)
# -> _prep_rows_to_write_for_ingestion -> _convert_arrow_to_proto

but _convert_arrow_fv_to_proto (in utils.py) resolves the created-timestamp column
by the feature-view-configured name, not the hardcoded "created":

if feature_view.batch_source.created_timestamp_column:
    created_timestamps = [
        ... table.column(
            feature_view.batch_source.created_timestamp_column   # e.g. "created_at"
        ) ...
    ]

So the client emits the column as created and the server looks it up as created_at
→ KeyError. The timestamp value is not lost — it is on the wire under created;
it is simply labeled with a name the server does not look up. The client write and
the server-side conversion disagree on the timestamp column names.

Server-side traceback:

File ".../feast/feature_server.py", in write_to_online_store
File ".../feast/feature_store.py", in write_to_online_store
File ".../feast/infra/passthrough_provider.py", in _prep_rows_to_write_for_ingestion
File ".../feast/utils.py", in _convert_arrow_fv_to_proto
KeyError: 'Field "created_at" does not exist in schema'

Steps to reproduce

1. Define a feature view whose batch source uses non-default timestamp column names,
   e.g. PostgreSQLSource(timestamp_field="event_timestamp", created_timestamp_column="created_at", ...).
2. Point a client at a remote online store: online_store: { type: remote, path: <feature-server-url> } (or call the feature server /write-to-online-store).
3. store.write_to_online_store(fv_name, df=df) where df includes the created_at
   column. The client-side conversion succeeds; the server returns HTTP 500 with the
   KeyError above. The same write against a local provider succeeds.

Specifications

• Version: 0.64.0 (client and feature server)
• Platform: Linux, Python 3.13
• Subsystem: remote online store (feast/infra/online_stores/remote.py) /
  Python feature server /write-to-online-store (feast/feature_server.py) /
  _convert_arrow_fv_to_proto (feast/utils.py)

Possible Solution

In RemoteOnlineStore.online_write_batch, label the timestamp columns using the
feature view's configured fields instead of hardcoded literals:

event_col   = table.batch_source.timestamp_field or "event_timestamp"
created_col = table.batch_source.created_timestamp_column or "created"
columnar_data[event_col].append(...)
if created_col:
    columnar_data[created_col].append(...)

Alternatively, have the server map the well-known event_timestamp / created
request columns onto the FV's configured names before conversion. Fixing the client to
emit the configured names is the smaller change and keeps the request self-describing.