What this does

pa_to_feast_value_type had no handling for fixed_size_list, so decoding an Arrow vector column raised a KeyError:

>>> pa_to_feast_value_type("fixed_size_list<item: float>[8]")
KeyError: 'fixed_size_list<item: float>[8]'

>>> pa_to_feast_value_type("list<item: float>")   # variable-width worked
ValueType.FLOAT_LIST

Only the fixed-width spelling was affected. "fixed_size_list" not in inspect.getsource(feast.type_map) held before this change.

Why it matters

fixed_size_list is how Arrow-native stores spell an embedding column, and it is the type _validate_vector_field_lengths reads a width from — the validator added in #6909. So schema inference failed on exactly the columns that vector support is meant to describe.

Any source whose schema inference routes through pa_to_feast_value_type is affected, not one particular store. It surfaces through get_pyarrow_schema_from_batch_source, so offline_write_batch and feast apply schema inference both hit it.

Approach

A Feast value type carries no width — Field.vector_length carries it — so the two spellings must agree. The fixed-width spelling is rewritten into the variable-width one and then flows through the existing list handling unchanged, which keeps that logic single-sourced rather than duplicating the element-type table.

Nesting is preserved: fixed_size_list<item: list<item: float>>[8] still decodes to VALUE_LIST, matching list<item: list<item: float>>.

Verification

Each is asserted equal to its variable-width counterpart, and one test pins the parser against the exact string PyArrow emits (str(pyarrow.list_(pyarrow.float32(), 8))) so the format cannot drift away unnoticed.

5 tests added. sdk/python/tests/unit/test_type_map.py: 24 failures before the change and the same 24 after, by name — all pre-existing TestSparkNativeTypeValidation cases that need pyspark. 247 passed.

ruff check and ruff format --check clean.