Skip to content

README.md

Latest commit

 

History

History
103 lines (83 loc) · 5.25 KB

README.md

File metadata and controls

103 lines (83 loc) · 5.25 KB
Raw

Vendor-commit build cost

What it measures: how much a vendor commit actually changes, and what that costs to rebuild — the churn per commit, the ccache hit rate on adjacent commits, and the size of a cached object archive.

When and on what: a local session, March–May 2026, against real v1.5-variegata-dev history (~2 weeks+ old at the time), R 4.3.3, ccache 4.9.1, -j4, unity build linked with LTO. Sample for A.1: 163 vendor commits, March–May 2026.

What it supports: the cost model in operations/ci/per-commit/planning/ — its shard weighting, its bimodal timing, and the archive trade-off.

Provenance: run as Appendix A of the agentic-loop design, now plan/superseded/vendoring-loop.md, and moved here unchanged when that design was superseded but its measurements were not. No script survives; the method is described with each measurement.

A.1 Churn / path-filter validation (163 vendor commits, Mar–May 2026)

  • Path filter holds. Every vendor commit touches only src/duckdb/ plus two generated files: R/version.R (163/163) and occasionally src/include/sources.mk. The only non-mechanical touches in the whole sample were 3 genuine folded fixes (1 test, 2 snapshots) — i.e. exactly what the review surface (§3.4) is meant to flag.
  • Churn is tiny and header-light. Per vendor commit: median 2 .cpp changed (106/163 change exactly 2); 66% (107/163) change zero headers; header changes are a small tail (mostly 1, rarely up to 14).

A.2 ccache behaviour on adjacent commits (8 consecutive v1.5 commits)

Full in-place rebuild at each commit, shared ccache, --preclean so every object is offered to ccache (measures the cache hit rate, not incremental make). Build is a unity build (~340 objects) linked with LTO.

step Δ cpp Δ hdr wall hits misses hit %
1 (cold) — — 841 s 0 351 0%
2 1 1 (narrow) 173 s 325 26 92%
3 2 0 91 s 346 5 98%
4 2 0 92 s 346 5 98%
5 5 3 (wide) 738 s 161 190 45%
6 6 0 89 s 347 5 98%
7 1 1 (narrow) 89 s 347 5 98%
8 1 1 (narrow) 163 s 326 26 92%

Takeaways:

  • Typical adjacent commit: ~98% cached, ~90 s (only ~5/351 objects rebuilt).
  • Mean across all incremental steps ≈ 89% / ~205 s, dragged down by the one wide-header commit; median ≈ 98% / ~92 s.
  • Header reach, not count, is the cost driver: steps 2, 7, 8 are all "1 cpp + 1 hdr" yet span 5–26 misses (98%↔92%); step 5's 3 wide headers invalidated 190/351 (45%, near-cold). ⇒ the sharding cost-estimator must weight headers by reverse-include reach (§4.3).
  • The ~70–90 s floor on cheap commits is LTO link + install + smoke test, not compilation ⇒ drop LTO for the smoke build (§4.4).
  • A.1 (66% zero-header) + A.2 (zero-header ⇒ 98%) ⇒ most adjacent rebuilds are near-free. The bulk path realises this with incremental make within a shard (no shared cache, §4.2); the normal per-commit path with the existing CI caches. The --preclean here only forces a full recompile to measure the hit rate; production never cleans between consecutive commits.

A.3 duckdb.tar archive size — debug info dominates

Scope: this measurement informed an earlier cached bulk design. The bulk path is now cache-free (§4.2), so the archive-size question applies only if the normal per-commit cached path's duckdb.tar is ever shrunk; it is retained here as a measured reference (and the -g/strip trade-off is real for that path).

The cached object archive ($(SOURCES) = 341 unity .o, one v1.5 tree):

variant total .o tar (raw) gzip -6 zstd -3 zstd -19
unstripped (-g, current CI) 2.5 GB 2.5 GB 489 MB 457 MB —
stripped (--strip-debug) 97 MB 97 MB 20 MB 19 MB 14 MB

(Context: the shipped duckdb.so, stripped via _R_SHLIB_STRIP_, is ~46 MB.)

  • Debug info is ~96% of the archive — stripping is a 26× reduction.
  • At ~457 MB zstd (current -g), only ~20 trees fit the 10 GB actions/cache budget ⇒ per-commit archiving of a bulk replay is infeasible. At ~19 MB zstd (stripped), ~500 trees fit ⇒ per-commit archive caching is cheap (§4.2).

End-to-end validation of "build -g, strip only the archive":

  1. Live build (-g, archive absent → to-tar.mk): rc=0, both the .o and the installed .so carry .debug_info ⇒ real traces on a failing test.
  2. Strip the extracted .o with --strip-debug, re-tar → 97 MB / 19 MB zstd.
  3. Cache-hit build (stripped archive present → from-tar.mk): rc=0 in 6 s, engine objects extracted from the archive, only the ~15 glue .cpp recompiled; the .so linked cleanly from the stripped objects and the connect/insert/SELECT sum smoke test passed (loaded strictly from the cache-hit library).

Caveats confirmed: use --strip-debug (a full strip removes the symbol table and breaks linking); on a cache hit the engine frames are debug-less (only glue has -g), which is acceptable since novel failures occur on the -g miss path.