uv: Deduplicate all files in the wheel cache(github.com)
230 points by tosh 10 days ago | 118 comments
tl;dr: uv's content-addressed cache (preview) now deduplicates identical files within and across cached wheels via hardlinks, reducing disk usage for overlapping dependencies. Follow-up optimizations reuse a single 64 KiB hashing buffer during extraction (yielding ~8-10% faster cold installs on large wheels like PyTorch and NumPy) and speed up cache cleanup on macOS by batching hardlink-count reads with `getattrlistbulk` (~4x faster scans).
HN Discussion:
  • ~Pip maintainer acknowledges uv's cache advantages while noting outstanding limitations
  • General enthusiasm for uv as essential Python tooling worth improving
  • Appreciation for BLAKE3 as a great choice for content-addressed deduplication
  • Skepticism that the size/speed tradeoff and added complexity are worthwhile
  • Objection to supporting Astral/uv on ideological grounds regarding AI