perf(onset): lands never wait on Upstream, cheaper lands and git view, small-step compaction

Radin committed …

⧉115913154a49

Author

Radin · radin@coline.appOct 29, 2026 at 05:36

Committer

Radin · radin@coline.appOct 29, 2026 at 05:36

Parents

07149df2c388

perf(onset): lands never wait on Upstream, cheaper lands and git view, small-step compaction

- Sign-in, repository and rules answers are used while a newer one is
  fetched in the background (up to 60 s / 2 min old); a request only waits
  on Upstream when nothing is known. The data center's copy keeps when
  Upstream gave it, so new instances after a deploy start from it too.
- The Worker's land batching sends the next call as soon as the last one
  ends (it polled every 25 ms).
- An edits land writes its change once, already landed, with no extra
  roots; objects go in one statement; main's latest version is kept in
  memory: 18 statements a land down to 7.
- Tree diffs walk both folders' sorted entries side by side instead of
  building maps; objects the core made aren't hashed again. A land on a
  900-file folder costs half the CPU.
- The git view writes a folder's tree as a delta on its tree earlier in
  the same pack (25x fewer bytes pushed on radin/vcs-check), sends each
  unchanged entry's git ID once a pack instead of once a version, and
  loads a batch's folders a level at a time in one read: a 256-version
  push of vcs-check builds in 12 s instead of 85-98 s.
- Compaction reads only the rows it moves (it read 20,000 rows' bytes to
  keep 8 MB) and moves them 256 KB at a time, so lands aren't held for
  seconds behind one big write.
- bench/land.mjs: --flat (a big top folder) and a no-rules Upstream for --git.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>