Skip to content

perf(compiler): bound cold analysis and tighten regression budgets - #772

Merged
kvlonge merged 7 commits into
mainfrom
perf/cold-compile-guarded-release
Sep 22, 2026
Merged

kvlonge merged 7 commits into
mainfrom
perf/cold-compile-guarded-release

Conversation

@kvlonge

@kvlonge kvlonge commented Sep 22, 2026 •

Copy link
Copy Markdown
Contributor

Why

Cold compilation repeatedly copies query trees and reconstructs compiler facts. Preserve the implemented improvements with representative, bounded-memory regression gates before further architectural work.

Changes

  • Reuse bounded native query facts with conservative fallback, compact analysis results, and reduced validation allocations.
  • Batch static interpolation/reference work and overlap independent rule preparation while preserving diagnostics and cache identities.
  • Add dense 1k/3k/5k/10k all-rules guards, exact semantic/artifact fingerprints, and zero-cache-hit assertions.
  • Enforce 14s/34s/55s/115s regression ceilings and a 4GiB no-swap CI limit; include the matrix in the required Verify gate.
  • Keep experimental project-partitioning and prototype shortcuts out of this release.

Verification

  • All four dense guards passed before and after tightening budgets.
  • Hosted CI measured 11.92s at 1k and 30.18s at 3k; calibrated their ceilings to 14s/34s. Larger dense guards and other required suites passed. Log successful measurements for future calibration.
  • uv sync --all-extras and make check-ci passed.
  • Focused compiler, native-analysis, cache and real DuckDB integration checks passed at the implementation checkpoints.
  • Full cross-platform and integration validation is gated by repository CI.

Back up static preparation, bounded native analysis, compatibility recovery, and completed-cache reuse work. Pin the matching Polyglot checkpoint for reproducible checkouts. Full semantic equivalence and final cold-time/RSS acceptance remain outstanding; this is an unmerged work-in-progress checkpoint.
Combine native analysis and binding, reuse SQL expansion evidence, batch lint work, and preserve input-column lineage precedence. Keep replacement fact/type evaluation behind experimental gates and bypass analysis caches during experiments. Pin the shared parser checkpoint for reproducible builds. This backup is not release-ready; semantic parity and cold performance work continue.
@kvlonge
kvlonge enabled auto-merge (squash) September 22, 2026 10:05
@kvlonge
kvlonge merged commit 9f248e6 into main Sep 22, 2026
29 checks passed
@kvlonge
kvlonge deleted the perf/cold-compile-guarded-release branch September 22, 2026 10:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant