🔴 Silent data loss for empty-collection values (e.g., []) now treated as null instead of erroring

Replacing the TypeError with return [ProtoValue() for _ in values] silently converts legitimate empty-collection values (like []) into null ProtoValues. The _non_empty_value function at sdk/python/feast/type_map.py:1351-1360 treats empty lists/sets as "empty" (returns False), so when the only values are empty collections, sample is None, value_type stays UNKNOWN, and the new code returns null ProtoValues.

This causes silent data corruption in at least two call paths:

1. _python_dict_to_map_proto (line 983): A struct like {"scores": []} will have its empty-list value silently converted to a null ProtoValue, losing the semantic distinction between "empty list" and "null".

2. remote.py:205-213: When the remote online store returns a feature with status "PRESENT" and value [] (a valid empty list), the function returns ProtoValue() (null). The caller then stores this as the feature value (feature_values_dict[feature_name] = message[0]), silently dropping the actual empty-list value.

A safer fix would be to only return empty ProtoValues when all values are actually None, not when they are empty collections.

In sdk/python/feast/type_map.py at line 1017-1018, the fix for all-NULL columns is too broad — it also silently converts empty-collection values (like []) to null ProtoValues. Instead of unconditionally returning empty ProtoValues when value_type is UNKNOWN, check whether ALL values are actually None. If so, return empty ProtoValues (the intended fix for all-NULL columns). If some values are non-None but just empty collections (like []), either raise the original TypeError or attempt more aggressive type inference. For example:

if value_type == ValueType.UNKNOWN:
    if all(v is None for v in values):
        return [ProtoValue() for _ in values]
    raise TypeError("Couldn't infer value type from empty value")

Was this helpful? React with 👍 or 👎 to provide feedback.

@yuan1j Looks like a perfect issue identification here. Could you please change the code to match the devins expectations here.