The "Build darwin binaries" jobs are the critical path for Publish npm in publish.yaml (~10.5 min per release) and rebuild ~388 crates from scratch every run.
Current state
docker/build/darwin.Dockerfile uses FROM ghcr.io/rivet-dev/rivet/builder-base-osxcross:0e33ceb98 — the same prebaked base rivet uses (osxcross + MacOSX 11.3 SDK + both apple Rust targets + sccache preinstalled) — but uses none of its caching affordances:
- re-installs the rust toolchain/targets the base already has
- no
RUSTC_WRAPPER=sccache, no cache backend, no BuildKit cache mounts
- plain
docker build with no --cache-from/--cache-to
- Measured on the 0.2.5 stable run (28801821648): base image pull ~1 min, rustup ~1 min,
cargo build --release of 388 crates = 7m52s, every run, forever.
- By contrast the linux sidecar/plugin jobs use Swatinem/rust-cache and finish in 1–5 min warm.
Fix: port the rivet consumer pattern
Rivet already solved this for the identical build in rivet-dev/rivet:
docker/build/darwin-{arm64,x64}.Dockerfile: RUSTC_WRAPPER=sccache + SCCACHE_WEBDAV_ENDPOINT=https://cache.depot.dev, --mount=type=secret,id=DEPOT_TOKEN,env=SCCACHE_WEBDAV_TOKEN with a graceful health-check fallback that disables sccache when the token/backend is unavailable, plus --mount=type=cache for cargo registry/git/target.
- Base image build/push lives in
docker/builder-base/osxcross.Dockerfile + scripts/docker-builder-base/build-push.sh.
Port notes for agentos:
- agentos builds on GitHub-hosted runners (not Depot builders), so BuildKit cache mounts alone will not persist across runs — the cross-run win is the sccache remote backend. Either:
- (a) add a
DEPOT_TOKEN secret and reuse rivet's Depot cache config verbatim (preferred: byte-for-byte port of a production pattern), or
- (b) use sccache's native GHA backend (
SCCACHE_GHA_ENABLED=on).
- Drop the redundant rustup toolchain/target install from
darwin.Dockerfile (already in the base).
- Pass the secret through in the
build-sidecar-darwin job in .github/workflows/publish.yaml (docker buildx build --secret id=DEPOT_TOKEN).
Expected: darwin job drops from ~10 min to ~2–3 min warm, cutting release npm-publish latency roughly in half.
Related smaller wins (same investigation, can be separate PRs)
Publish crates.io job: give it its own concurrency group (or workflow) so a slow crates tail (crates.io new-crate rate-limit backoff can sleep for many minutes) doesn't queue the next release's npm path behind it (concurrency: publish-${{ github.ref }}, cancel-in-progress: false caused the 0.2.6 dispatch to sit pending behind 0.2.5's crates job). Also surface the "rate limited, retrying at " sleep as a ::notice:: so the job doesn't look hung.
- Release rust-cache eviction: linux jobs key caches by
<target>-<trigger>, so -release caches are only warmed by prior release runs and get LRU-evicted during quiet stretches while -preview stays warm; consider a shared key or a scheduled warm job.
- WASM Commands (~8.6 min): cache
packages/runtime-core/commands keyed on hashFiles('registry/native/**') and skip make on hit.
The "Build darwin binaries" jobs are the critical path for
Publish npmin publish.yaml (~10.5 min per release) and rebuild ~388 crates from scratch every run.Current state
docker/build/darwin.DockerfileusesFROM ghcr.io/rivet-dev/rivet/builder-base-osxcross:0e33ceb98— the same prebaked base rivet uses (osxcross + MacOSX 11.3 SDK + both apple Rust targets + sccache preinstalled) — but uses none of its caching affordances:RUSTC_WRAPPER=sccache, no cache backend, no BuildKit cache mountsdocker buildwith no--cache-from/--cache-tocargo build --releaseof 388 crates = 7m52s, every run, forever.Fix: port the rivet consumer pattern
Rivet already solved this for the identical build in
rivet-dev/rivet:docker/build/darwin-{arm64,x64}.Dockerfile:RUSTC_WRAPPER=sccache+SCCACHE_WEBDAV_ENDPOINT=https://cache.depot.dev,--mount=type=secret,id=DEPOT_TOKEN,env=SCCACHE_WEBDAV_TOKENwith a graceful health-check fallback that disables sccache when the token/backend is unavailable, plus--mount=type=cachefor cargo registry/git/target.docker/builder-base/osxcross.Dockerfile+scripts/docker-builder-base/build-push.sh.Port notes for agentos:
DEPOT_TOKENsecret and reuse rivet's Depot cache config verbatim (preferred: byte-for-byte port of a production pattern), orSCCACHE_GHA_ENABLED=on).darwin.Dockerfile(already in the base).build-sidecar-darwinjob in.github/workflows/publish.yaml(docker buildx build --secret id=DEPOT_TOKEN).Expected: darwin job drops from ~10 min to ~2–3 min warm, cutting release npm-publish latency roughly in half.
Related smaller wins (same investigation, can be separate PRs)
Publish crates.iojob: give it its own concurrency group (or workflow) so a slow crates tail (crates.io new-crate rate-limit backoff can sleep for many minutes) doesn't queue the next release's npm path behind it (concurrency: publish-${{ github.ref }},cancel-in-progress: falsecaused the 0.2.6 dispatch to sit pending behind 0.2.5's crates job). Also surface the "rate limited, retrying at " sleep as a::notice::so the job doesn't look hung.<target>-<trigger>, so-releasecaches are only warmed by prior release runs and get LRU-evicted during quiet stretches while-previewstays warm; consider a shared key or a scheduled warm job.packages/runtime-core/commandskeyed onhashFiles('registry/native/**')and skipmakeon hit.