BoringCache benchmarks

Build real projects with BoringCache and other cache providers. Each result identifies the source revision, provider, runner, machine, phase, and GitHub Actions run.

Results · JSON · Actions · Configuration

Schedule

The rolling schedule checks the projects in rolling.toml every 15 minutes. It dispatches an idle project when its upstream branch has advanced by the case's rolling_stride (one commit by default). Running or queued projects wait for a later tick. Each tick's Actions summary lists dispatched, active, current, and failed checks.

Every lane of a case builds the same first-parent commit. Step 0 builds start_sha with empty rolling caches. Later steps restore and save each lane's fixed cache scope. Recorded build failures advance the series; setup failures without records do not. Fresh runs use separate caches.

The report updates hourly. Stale BuildKit cache cleanup runs every two hours; ordinary Actions caches are pruned after rolling results are published. GitHub may delay scheduled runs.

Run

Use the Ruby version in .tool-versions:

bundle install
bundle exec bin/bench check
bundle exec bin/bench matrix grpc --tool bazel
gh workflow run project.yml --ref main -f tool=Bazel -f project=grpc -f mode=fresh

• fresh: cold and warm builds of start_sha on separate runners.
• rolling: the next upstream commit due at the configured stride.
• preflight: setup and the lane's declared probes, without a build.

Add -f lane=<lane> to select lanes or -f cli=vcli-canary-<sha> to test a published CLI canary in fresh mode. Canary runs are manual. A cold phase cannot be retried under the same run ID; dispatch another fresh run.

Rolling dispatches use the tool's display title, such as Rust/Cargo, so scheduled and manual runs share a concurrency group.

Preview the rolling scheduler without starting builds:

gh workflow run schedule.yml --ref main -f dry_run=true

For local builds and report generation:

bundle exec bin/bench run bazel/grpc --lane boringcache-bazel
bundle exec bin/bench report --results results --out tmp/report

Local builds need the lane's tools and credentials. See configuration for container runs and AGENTS.md for validation commands.

Results

results/ contains measurements and provider evidence. data/report.md and data/results.json report the current lanes and runners, excluding experimental lanes. Times cover cache restore, build, and cache save; checkout, preparation, and output checks are excluded. Measurement details describe matching, failures, cache misses, and post-job saves.

data/latest/ is the historical feed still read by the website. It is retained until that consumer migrates. PLAN.md lists the remaining migration work and records removed lanes and restarted series.