perf(cache): make the result cache pay only where it helps#4
Open
naamanhirschfeld-armis wants to merge 3 commits into
Open
perf(cache): make the result cache pay only where it helps#4naamanhirschfeld-armis wants to merge 3 commits into
naamanhirschfeld-armis wants to merge 3 commits into
Conversation
On a whole-repository run the result cache wrote one tiny file per (file, engine). For the great majority of files the wrapped tools are fast, so the open + write + rename to store a result — and the failed open() to miss it — cost more than simply recomputing. The effect was a cold run several times slower than `--no-cache`, while a warm run barely beat it. Two changes make the cache pay only where it helps: - Duration-gated puts: persist a result only when its engine ran at least MIN_CACHE_DURATION. Cheap results are recomputed each run (fast by construction); genuinely expensive files (large/slow inputs) are still cached and skip the work next time. The gate is on cost, not on emptiness, so an expensive file that produces no diagnostics / no reformat is still cached. - In-memory presence index: ResultCache scans the on-disk keys once at open and serves misses from a hash set, so a miss costs a lookup instead of a failed syscall. A small put-updated layer (guarded by a flag, so it stays lock-free until the first write) keeps get-after-put correct on one handle. Also skip digest/key/serialize entirely when the cache is disabled, and surface slow single-file engine runs at info/warn — noting the file, its size, the backend, and (for a wrapped upstream tool) a hint to report it upstream, since that cost is in the tool, not in polylint. Result on a large mixed corpus: cold ~= --no-cache, and a previously 150s-per-run file is served from cache in well under a second on re-runs.
Add .cargo/config.toml setting rustc-wrapper = "sccache" so clean builds and branch switches reuse cached artifacts (the pinned oxc/ruff git dependencies dominate build time and cache well). This requires sccache on PATH, so every compiling CI job is updated to install it: - ci.yaml: add mozilla-actions/sccache-action to fmt/clippy/test, with SCCACHE_GHA_ENABLED and CARGO_INCREMENTAL=0 (sccache can't cache incremental). - publish.yaml: add sccache to the native release builds; the Alpine/musl container build sets RUSTC_WRAPPER="" to neutralize the wrapper, since rust:alpine has no sccache (the env var overrides the committed config).
…get bench Wire a pprof-based criterion profiler (benches/support/profiler.rs) into all three benches, so `cargo bench -- --profile-time=N` writes a flamegraph.svg per benchmark. It wraps pprof::ProfilerGuard directly rather than enabling pprof's own `criterion` feature, staying decoupled from whichever criterion version that feature pins. pprof is a dev-only dependency (never in the shipped binaries). Add a cache_get bench measuring a miss (served from the in-memory presence index) vs a hit (disk read): the miss is ~14 ns vs ~36 us, so a regression that reintroduces a per-miss syscall shows up here.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Hardening + profiling pass on whole-repository runs. Profiling a large mixed corpus (~70k files) showed the run was spending most of its wall time in filesystem syscalls for the result cache, not in the wrapped tools — a cold run was 2–3× slower than
--no-cache, and a warm run barely beat it.Cache: cost-gate writes + presence index (
perf(cache))The cache wrote one tiny file per
(file, engine). For the majority of files the wrapped tools (oxc, ruff, …) are fast, so the open+write+rename to store — and the failedopen()to miss — cost more than recomputing.MIN_CACHE_DURATION. Cheap results recompute each run (fast by construction); expensive files are still cached. The gate is on cost, not on emptiness, so an expensive file that reformats to unchanged is still cached.Result on the corpus: cold ≈
--no-cache; warm full-corpus run 227s → ~15s; a single pathological file that took ~150s/run is served from cache in <0.1s on re-runs. Lint/format output is byte-for-byte identical across no-cache/cold/warm.Build: sccache (
chore(build)).cargo/config.tomlroutes rustc through sccache; CI (ci.yaml,publish.yaml) provisions it, and the Alpine/musl release container neutralizes the wrapper (no sccache there). Requiressccacheon PATH locally.Benches: flamegraphs + cache_get (
test(bench))A pprof-based criterion profiler emits
flamegraph.svgper bench under--profile-time=N(dev-only dep). Newcache_getbench guards the miss path (~14 ns index miss vs ~36 µs disk hit).Note
The ~150s pathological case is a super-linear blowup in an upstream backend on a large generated input; it is tracked upstream and not addressed here.