What this PR does / why we need it:

An on-demand feature view cannot have schema fields with a description or tags.

feast apply fails on such a view with SpecifiedFeaturesNotPresentError. _feature_exists_in_inferred checks specified_feature in inferred_features, which uses Field.__eq__ and compares the description, tags and vector settings. The inferred fields carry only a name and a type, and a scalar field with a description never matches. The check now compares the name and the type.

With only that change, feast apply succeeds, but the view read back from the registry has an empty description. _parse_features_from_proto builds each Field from the name, the value type and the vector settings instead of calling Field.from_proto, which FeatureView and StreamFeatureView use. Besides the description and tags, it loses the inner types of nested outputs:

_parse_features_from_proto now returns Field.from_proto(feature) for each feature.

To reproduce, add description="val plus one" to a Field in the schema of an @on_demand_feature_view with a RequestSource, in a repo with the local provider and the SQLite online store, and run feast apply. On master it stops after "Applying changes for project repro" with SpecifiedFeaturesNotPresentError. With this change it prints "Created on demand feature view plus_one", and store.get_on_demand_feature_view("plus_one").features[0].description is 'val plus one'.

Which issue(s) this PR fixes:

Fixes #3617. The stale bot closed it in 2025 without a fix. It reports the first half; the registry round-trip has no issue.

Checks

• I've made sure the tests are passing.
• My commits are signed off (git commit -s)
• My PR title follows conventional commits format

Testing Strategy

• Unit tests
• Integration tests
• Manual tests
• Testing is not required for this change

Misc

test_schema_field_metadata_survives_inference_and_proto_round_trip in tests/unit/test_on_demand_feature_view.py checks that the description and tags survive infer_features, that a proto round-trip returns the original fields, including Struct and nested Array types, and that a wrong type still raises SpecifiedFeaturesNotPresentError. It fails on master. test_on_demand_feature_view.py and test_on_demand_python_transformation.py pass, and ruff is clean.

ODFVs that are already registered with Struct or nested outputs now load with their real inner types instead of _value: String.