ci: parallelize workspace checks to cut PR build time - #353

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

CI timing — measured results (baseline)

Measured on real Actions runs. "Old CI" = a recent main run (build + build-in-minimal, warm actions/cache). "New" = this PR's runs (fmt/clippy/test/dogfood in parallel).

Wall-clock (PR critical path)

Old jobs for reference: build = 32.3 min (serial cargo build 11.8m → test 3.5m → clippy 8.9m), build-in-minimal = 30.1 min.

Takeaways

• ~32 → ~14 min (~57%), almost entirely from parallelizing the jobs and dropping the redundant cargo build (cargo test already builds every target).
• The sccache layer under-delivered. The warm clippy job hit a 90% cache rate yet only improved 13.1 → 10.7 min, because (from the sccache stats in the logs):
  • ~138s spent downloading 875 cached objects from the GHA cache (per-object round-trips: cache_read_hit_duration: 137.96s);
  • 184 compilations are non-cacheable every run (proc-macros, crate-type=bin, and build scripts — this repo is protobuf-heavy via prost-build/tonic-prost-build, and sccache never caches build-script output or linking);
  • 239 cache write errors on the cold run (parallel clippy+test write to one shared sccache namespace and collide) → 97 forced misses on the warm run.
• dogfood builds -p minimal through the sandboxed minimal tool and barely warms (15.1 → 13.8 min); it is the warm long pole. (Tracked separately; not addressed in this round.)

Next: replacing sccache with Swatinem/rust-cache target caching + trimming Free Disk Space

Incoming in this PR. Expectation: one tarball restore instead of 875 object reads, build-script/dep artifacts cached directly, no cross-job write contention. Re-measured numbers to follow in a reply.

Generated by Claude Code

CI timing — after replacing sccache with Swatinem/rust-cache (target caching) + trimming Free Disk Space

Measured on real runs of identical config (5905f31 cold → 22a2515 warm), compared against the earlier sccache setup (d659ce6 cold → 663f553 warm). Per-job wall-clock; the Run column is the actual compile/lint/test step.

clippy

test

Free Disk Space step (trimmed: dropped the slow apt package purge)

~1.8m → 44s–1.3m.

Why rust-cache wins

• One tarball restore (~13–22s) instead of sccache fetching 875 objects one-by-one (~138s of cache_read_hit_duration).
• Caches dependency .rlibs and build-script output directly → cargo skips deps entirely. This repo's protobuf build scripts (prost-build/tonic-prost-build) were in sccache's 184 "non-cacheable" compilations; rust-cache restores their output.
• Per-job cache keys → no cross-job write contention (sccache had 239 write errors → 97 forced warm misses). rust-cache warm runs hit clean.

Net for the cached jobs, warm: clippy 10.7m → 2.5m, test 11.1m → 4.4m.

⚠️ Honest caveat: end-to-end PR wall-clock is unchanged

The PR's critical path is now gated by dogfood (~15–16m), which runs minimal run build-smoke inside minimal's sandbox using minimal's own state cache — it does not use rust-cache and barely warms. So clippy/test finishing in 2.5/4.4m now just means they idle while dogfood runs.

• Compute/cost drops meaningfully (warm clippy+test went from ~22 job-min to ~7).
• Wall-clock won't drop below ~15m until dogfood's sandboxed build is cached or slimmed — tracked separately; not in scope here.

Generated by Claude Code