[hauntsaninja]

Authored by Codex, here's its description below.

The UNIQUE constraint on files2.path creates an index on that column. The explicit path_idx indexed the same keys, so every cold-cache insert updated two indexes. Existing databases keep their index and old mypy versions can read the new layout.

This reduced cold torch 2.1.0 import wall time by 1.22% (3.263 s to 3.226 s, 95% interval 1.00% to 1.44%). Self-check fell by 0.43% (6.198 s to 6.171 s, interval 0.20% to 0.68%). These were 16 alternating pairs per input against the parent commit, using compiled mypy on CPython 3.11.8 with a fresh cache and no workers. All diagnostic streams matched. The torch improvement also replicated against the unmodified fork at 1.25% across 14 pairs.

Eight pairs per primer input showed 1.62% on pandas, 1.33% on colour, and 0.88% on jax. The first zulip series was noisy. An independent 16-pair series measured 0.78% (interval 0.50% to 1.02%).

The 16 torch shard files occupied 15,847,424 bytes, down from 16,080,896. Both caches contained 3,061 rows. All 1,020 module data blobs were byte-identical. Each version reused the other's cache and SQLite's integrity checks passed in both directions. The cache diff and fine-grained cache suites passed (552 tests).

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[JukkaL]

Good catch!