Repository navigation
A seed cache per target: warmed wheels-only by the tick, copied into every job - #330
Conversation
…ies it into its own cache evalcache.warm fetches the target's pyproject.toml and uv.lock at its default branch and runs uv sync --frozen --no-install-project --no-build --all-extras into a throwaway environment with the seed as uv's cache, once per lockfile hash; --no-build keeps every build backend of the target's off the tick host. write_eval_job(seed_cache=) adds one line to the job script that copies the seed's contents into the job's scratch cache before uv runs — a copy, never a bind or a hardlink, best-effort. DispatchSettings and DispatchedMeasurer carry the seed path; the launcher and the inline measurer pass it; the tick warms once per cycle. Step 1 of docs/design/eval-cache.md. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
Round 1 — reviewed head 2e6f6f67 — reviewer summarizer:hermes/gpt-5.6-terra over coverage+credentials+deployment+general+lifecycle+prose.
terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.
Verdict: 3 blocking, 1 advisory.
4 findings attached to the lines below.
Merged verdict: two blocking defects and two advisory improvements. [credentials] the warmer leaks the tick environment to uv; [deployment+general+lifecycle+prose] resumed and follow-up dispatches lose the seed cache; [lifecycle+prose] obsolete package entries accumulate after lockfile changes; [coverage] tick wiring lacks an integration test. Rejected findings: none; the deployment, general, lifecycle, and prose observations about missing seed-cache propagation were merged as one finding, and the lifecycle and prose observations about stale cache entries were merged as one finding.
…ile, the seed on wakes and follow-ups, tick integration test - The warmer's uv runs with an allowlisted environment (PATH, HOME, locale, proxies, CA bundles, its own UV_*): no tick credential reaches the download process. - Each warm builds a fresh directory beside the seed and swaps it in whole, so the seed holds exactly the current lockfile's wheels; a failed warm removes its half-built directory. - with_seed fills the seed from the record's target where the CLI gave none: resume_run (wake jobs) and the follow-up's seal-and-measure; a settings object that knows no seed is left as it is. - A tick-level test asserts the scheduled tick warms the target's seed. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
There was a problem hiding this comment.
Round 1 — reviewed head f3469cb7 — reviewer hermes/gpt-5.6-terra.
terra
Advisory findings from outerloop — the code owner decides. Reply to disagree; the outerloop:no-review label opts this PR out.
Verdict: nothing blocking — 1 advisory note.
1 finding attached to the lines below.
Dry-run ticks can still populate the new seed cache.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
renmengye
added
outerloop:review
1 participant
Step 1 of the eval-cache design note (#328): a seed cache per target, so a torch-sized dependency set is downloaded once per lockfile instead of once per job.
evalcache.warm, run byservice_eval_cacheonce per cycle): fetchespyproject.tomlanduv.lockfrom the target's default branch through the GitHub client (never the sources), and runsuv sync --frozen --no-install-project --no-build --all-extrasinto a throwaway environment with<state root>/eval-cache/<target>as uv's cache.--no-buildis the trust line: wheels are downloaded and unpacked, no build backend of the target's ever runs on the tick host; a lockfile that needs a build is reported and its jobs install as today. Re-warms only when the lockfile hash changes; a failure records nothing so the next tick retries.uvruns: one line in the job script (cp -a seed/. "$SCRATCH/cache"/, best-effort, after the scratch dirs exist and before the venv is built). A copy, never a bind or hardlinks; the jail and the private environment are unchanged, and jobs still write nothing shared.DispatchSettings.seed_cache(set from the attempt's--run-rootand--target),DispatchedMeasurer.seed_cache, the launcher and the inline measurer pass it. Local mode is the same code.Not in this PR: the in-image seed and the
environment.containerknob (step 2).🤖 Generated with Claude Code