[python] Vectorize batch raw vector scoring

[TheR1sing3un]

Purpose

Batch raw-vector fallback already streams each Arrow batch once, but it expands every vector into a Python list and scores every row/query pair through the scalar dimension loop. Convert each bounded regular FLOAT vector block to one owned float64 matrix, reuse that block across all queries, and keep only one active scratch matrix while updating the existing per-query Top-K heaps.

The vectorized reductions preserve the existing scalar accumulation semantics and Top-K tie breaking. Null, irregular, unsupported, non-finite, and dimension-mismatched inputs retain the scalar validation path.

Tests

• python -m pytest -q pypaimon/tests/vector_scoring_test.py pypaimon/tests/batch_vector_raw_scan_test.py: 16 passed.
• Exact score checks cover 2,051 vectors x 128 dimensions and five queries for L2, cosine, and inner product, including binary score equality with the scalar path.
• Existing tests cover Arrow list layouts, nulls, non-finite values, dimension failures, filters, snapshots, split parallelism, bounded batch consumption, and reader cleanup.
• Flake8 and git diff --check passed.

Benchmark

macOS arm64, Python 3.9.6, NumPy 2.0.2, PyArrow 19.0.1. 8,192 FLOAT vectors x 128 dimensions, 16 queries, Top-K=10. Values are medians of three runs.

The per-query-vectorized ablation converts and scores the Arrow block separately for every query. The shared-block variant is this change: it converts the Arrow block once and reuses it across queries.

Metric	Scalar kernel (s)	Per-query vectorized (s)	Shared block (s)	Kernel speedup	Top-K scoring speedup
L2	1.177249	0.043614	0.040448	29.10x	18.73x
Cosine	1.783921	0.088984	0.085943	20.76x	16.45x
Inner product	1.045036	0.166346	0.171815	6.08x	5.46x

All variants produced identical scores and Top-K heaps. The Top-K measurement includes Arrow-to-score conversion, scoring, and heap maintenance; it excludes table I/O, planning, and indexed search. The temporary benchmark driver is not included in the change.

[JingsongLi]

Requirement fit: SUPPORTED. Implementation: CLEAN.

Reviewed c22880574ba9. Raw batch fallback scores every row against each query, making repeated scalar dimension loops a real CPU cost. Reusing one owned float64 block and one active query scratch buffer improves that path while preserving scalar fallback, exact accumulation and deterministic Top-K behavior. I checked Arrow offsets/layouts, null/irregular vectors and non-finite/dimension validation.

Validation: 16 tests and 51 subtests passed, including exact scalar score/byte equality and bounded batch consumption. Current head CI is green. Kernel and end-to-end throughput measurements were not rerun locally.

No actionable implementation regression found in this review.