GeezerVison

Addresses item 3 of #6907 — but not the way the issue proposed, because that way does not work.

Why "require vector_length" is off the table

The issue suggested requiring vector_length when vector_index=True. I checked the blast radius first: ~29 call sites declare vector_index=True without vector_length, including

• Feast's own templates: feast/templates/rag/feature_repo/example_repo.py and feast/templates/ray_rag/feature_repo/feature_definitions.py — i.e. feast init -t rag would stop working
• examples/rag/, examples/rag-docling/, examples/agent_feature_store/, the Milvus and pgvector tutorials
• four online stores that deliberately fall back to 512: sqlite, qdrant, elasticsearch, mongodb

Making it mandatory would break the onboarding path. Not worth it.

What this does instead

The underlying complaint is still valid: vector_length defaults to 0, and 0 meant skip, so the common case was unvalidated — a correctness check that is off by default.

So rather than skipping when it is undeclared, the width of the first non-null row becomes the contract for the rest of the table. An ANN index requires a fixed width, so a ragged vector column is a defect whether or not anyone declared a length. Inferring is strictly better than skipping, and it requires nobody to change a single declaration.

The gate also moves from vector_length to vector_index, since vector_index is the attribute that actually says "this field is a vector".

Error messages distinguish the two sources, so an inferred failure is not mistaken for a declared one:

Row 2: Vector length 7 does not match expected 4, inferred from the first row
       for feature 'embedding' in feature view 'fv'.

Testing

11 new tests in test_vector_length_validation.py (32 total), covering both the Arrow and DataFrame paths: ragged rejected, uniform passes, declared still wins over inference, fixed_size_list needs no check, nulls neither infer nor compare, all-null is a no-op, non-sequence still caught, and vector_index=False untouched.

Three of them fail against master and pass here.

The important check for this PR is that inference does not break the 29 undeclared sites. I diffed failure names across the full unit suite rather than comparing counts, since the suite has flaky members:

with this change : 41 failures
master           : 41 failures
only with change : (none)

That is also the evidence that none of those sites produce ragged columns in practice.

ruff check / ruff format --check clean. mypy unchanged — same 2 pre-existing errors at feature_store.py:2123.

Relationship to the other PRs

• fix: Make vector length validation reachable, schema-driven and vectorized #6909 (merged) — items 1, 2, 4 of Vector length validation is unreachable from the batch path, skipped unless the vector is the first field, and off by default #6907
• fix: Include vector attributes in Field equality #6934 — item 5, Field.__eq__
• this PR — item 3

With this, #6907 can be closed.