REST feature server drops feature_view_metadata when include_feature_view_version_metadata is set #6922

Expected Behavior

POST /get-online-features (and /search / /retrieve-online-documents with api_version: 2) accepts include_feature_view_version_metadata: true. With that flag set, the response metadata should contain feature_view_metadata, a list of {name, version} entries, like the Python SDK and gRPC paths return. Three things support this:

• ServingService.proto: repeated FeatureViewMetadata feature_view_metadata = 2; // Only populated when requested.
• The feast/feature_server_utils.py docstring says convert_response_to_dict matches MessageToDict(proto, preserving_proto_field_name=True). The only exception it documents is double_val precision. MessageToDict does emit feature_view_metadata.
• RemoteOnlineStore sends the flag to the server, and its unit test says it does so "so versioned reads work end-to-end".

Current Behavior

The flag reaches store.get_online_features(...), and metadata.feature_view_metadata is filled on the response proto. It is then dropped during JSON serialization. The HTTP response metadata only ever contains feature_names, so over REST the flag has no effect.

The field is lost in three places:

1. feature_server_utils._metadata_to_dict only reads metadata.feature_names.
2. OnlineFeaturesMetadataResponse in feature_server.py has no feature_view_metadata field. Before perf: Replace MessageToDict with optimized custom dict builder #6015, the endpoint returned a dict through response_model=OnlineFeaturesResponse, so this model already filtered the field out. In practice the field has not been returned over REST since the flag was added in feat: Add version tracking to FeatureView #6101.
3. RemoteOnlineStore._build_online_response_from_json (infra/online_stores/remote.py) rebuilds the metadata from feature_names only. Even if the server returned the field, a client using a remote online store would still lose it.

Steps to reproduce

Minimal, no server needed (master @ 810391f):

from google.protobuf.json_format import MessageToDict
from feast import proto_json
from feast.feature_server_utils import convert_response_to_dict
from feast.protos.feast.serving.ServingService_pb2 import GetOnlineFeaturesResponse

proto_json.patch()
r = GetOnlineFeaturesResponse()
r.metadata.feature_names.val.append("trips_today")
r.metadata.feature_view_metadata.add(name="driver_stats", version=2)

print(convert_response_to_dict(r)["metadata"])
# {'feature_names': ['trips_today']}
print(MessageToDict(r, preserving_proto_field_name=True)["metadata"])
# {'feature_names': ['trips_today'], 'feature_view_metadata': [{'name': 'driver_stats', 'version': 2}]}

End to end, with feast serve on any repo. The PR's test_get_online_features_returns_feature_view_version_metadata checks the same thing through the FastAPI test client:

curl -s -X POST localhost:6566/get-online-features -H 'Content-Type: application/json' -d '{
  "features": ["driver_hourly_stats:conv_rate"],
  "entities": {"driver_id": [1001]},
  "include_feature_view_version_metadata": true
}' | jq .metadata
# {"feature_names": ["driver_id", "conv_rate"]}   <- no feature_view_metadata

Specifications

• Version: master (810391f), 0.66.x dev
• Platform: macOS 26.5 arm64, Python 3.11 (not platform specific)
• Subsystem: Python feature server (REST), remote online store

Possible Solution

• In _metadata_to_dict, emit feature_view_metadata when it is non-empty, in the same shape MessageToDict uses (default-valued name/version are omitted).
• Add feature_view_metadata: List[FeatureViewMetadataResponse] to OnlineFeaturesMetadataResponse so the OpenAPI schema documents it.
• In RemoteOnlineStore._build_online_response_from_json, parse the field back into the proto.

I have a small PR with regression tests ready and will link it here.

AI assistance: drafted with an AI coding assistant (Claude). The reproduction above was run locally against the current default branch.