Let only Test save the shared uv cache - #134
Conversation
Sorry, something went wrong.
astral-sh/setup-uv derives its cache key from the runner architecture, the runner image, the Python version and a hash of the dependency files. Documentation, Linter and the ubuntu-latest/3.10 leg of Test therefore all derive the same key, and when one push triggers them together they race to reserve it at save time. Only the first to finish wins; the losers log Failed to save: Unable to reserve cache with key setup-uv-2-..., another job may be creating this cache. The race is save-time only and harmless, since every loser would have written identical content, but it is noise in an otherwise green log. Giving each workflow its own cache-suffix would remove the collision by storing several copies of the same cache, which was rejected as wasteful in audeering/audeer#207. Keep the single shared key instead and make exactly one workflow the writer: Documentation, Linter and Publish set save-cache: false, so they still restore the cache but never save it, and Test remains the sole saver. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Reviewer's guide (collapsed on small PRs)
Reviewer's GuideConfigure GitHub Actions so that only the Test workflow writes the shared uv cache, while Documentation, Linter, and Publish continue to restore but no longer save, eliminating cache reservation races and warnings. File-Level Changes
Tips and commands Interacting with Sourcery
Customizing Your ExperienceAccess your dashboard to:
Getting Help
|
Sorry, something went wrong.
There was a problem hiding this comment.
Hey - I've left some high level feedback:
- Since the design now relies on Test being the sole cache writer, consider making this explicit in test.yml (e.g., a short comment and/or an explicit save-cache: true) so future workflow edits don’t accidentally reintroduce multiple writers or disable saving entirely.
Please address the comments from this code review:
## Overall Comments
- Since the design now relies on Test being the sole cache writer, consider making this explicit in `test.yml` (e.g., a short comment and/or an explicit `save-cache: true`) so future workflow edits don’t accidentally reintroduce multiple writers or disable saving entirely.Sourcery is free for open source - if you like our reviews please consider sharing them ✨ Help me be more useful! Please click 👍 or 👎 on each comment and I'll use the feedback to improve your reviews.
Sorry, something went wrong.
|
@sourcery-ai Thanks — the concern is fair, but I'd rather not act on it in this PR. This is a straight port of audeering/audeer#207, which is merged and reviewer-approved. That PR touches exactly doc.yml, linter.yml and publish.yml, leaves test.yml untouched, and adds no marker comment. audformat and audb are getting the same change, so keeping the four diffs identical is a large part of the value here — a comment in this repo only would make the rollout inconsistent. On the mechanics: save-cache: true is already the default in astral-sh/setup-uv v9.0.0, so writing it out changes no behaviour. It would read as if it were switching something on when it is not. There is also a working guard already in place. The three non-saving workflows each carry an explicit save-cache: false, which is a visible marker at every site where the invariant could be broken; and the failure mode is loud rather than silent — drop the sole saver and the shared key is never written, so every job logs No GitHub Actions cache found for key: … on each run. If @hagenw wants a comment in test.yml I'll add it here, but I'd then add the same one in audeer, audformat and audb so the four stay in step. |
Sorry, something went wrong.
CI verification on bfabeccUnable to reserve cache annotations: 0, across all 20 check-runs on the head SHA. Was 1 on main at a41c0bd. 19 of 20 checks pass. The one failure is Documentation, which fails identically on main at a41c0bd and is unrelated to this PR: FileNotFoundError: [Errno 2] No such file or directory: '/home/runner/audb/emodb/1.1.1/db.pkl' RuntimeError: Cannot find version '1.1.1' for database 'emodb'. Linter — restores, then declines to savesave-cache: false Trying to restore cache from GitHub Actions cache with key: setup-uv-2-x86_64-unknown-linux-gnu-ubuntu-24.04-3.10.20-69a2db3b… Cache restored successfully cache restored from GitHub Actions cache with key: setup-uv-2-x86_64-unknown-linux-gnu-ubuntu-24.04-3.10.20-69a2db3b… save-cache is false. Skipping save cache step. Same shared key as before, restore still works, and the save step is gone — which is exactly the job that used to win the reservation and starve the others. Test — still the writerRun 31098242982, build (ubuntu-latest, 3.10): save-cache: true Cache hit for: setup-uv-2-x86_64-unknown-linux-gnu-ubuntu-24.04-3.10.20-69a2db3b… Cache restored successfully Cache hit occurred on key setup-uv-2-x86_64-unknown-linux-gnu-ubuntu-24.04-3.10.20-69a2db3b…, not saving cache. save-cache is still true here. It does not write on this run because the key already exists — main populated it — and a hit means there is nothing new to store. On the next key rotation it is the one job that will write it. Documentation shows save-cache: false and its restore attempt, but no Skipping save cache step. line, because it aborts at the emodb failure above and setup-uv's post step is post-if: success(). That is also why it produced no warning on main despite deriving the same key: it never reached its save step. Once the emodb failure is fixed it becomes a third contender for that key — and this PR is what keeps that from turning into a second warning. |
Sorry, something went wrong.
There was a problem hiding this comment.
Pull request overview
This PR updates the GitHub Actions CI workflows to eliminate astral-sh/setup-uv cache save races across concurrently-started workflows by making only the Test workflow the cache writer (this PR disables saving in Documentation, Linter, and Publish while preserving restore behavior).
Changes:
- Disable setup-uv cache saving in .github/workflows/doc.yml.
- Disable setup-uv cache saving in .github/workflows/linter.yml.
- Disable setup-uv cache saving in both build and deploy jobs in .github/workflows/publish.yml.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.
| File | Description |
|---|---|
| .github/workflows/doc.yml | Sets save-cache: false for setup-uv to avoid competing cache saves. |
| .github/workflows/linter.yml | Sets save-cache: false for setup-uv to avoid competing cache saves. |
| .github/workflows/publish.yml | Sets save-cache: false for both setup-uv steps to prevent save collisions (including future release-time collisions). |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Sorry, something went wrong.
Problem
astral-sh/setup-uv derives its cache key from the runner architecture,
the runner image, the Python version and a hash of the dependency files:
Nothing in that key identifies the workflow. Documentation
(ubuntu-latest, 3.10), Linter (ubuntu-latest, 3.10) and the
ubuntu-latest/3.10 leg of Test therefore derive a byte-identical key, and
one push starts all three at once. At save time they race to reserve it;
only the first to finish wins, and the losers log:
It is save-time only and harmless — the loser's content would have been
identical to the winner's — but it is a yellow annotation on an
otherwise green run.
Measured on main: 1 annotation across the last six pushes, all of
it on the one push that has run since setup-uv was bumped to v9.0.0.
Push a41c0bd ("Fix CI caching; bump checkout/setup-python off Node.js
20"),
17 check-runs, 1 warning:
All three logged No GitHub Actions cache found for key: …69a2db3b…
at restore, so all three were candidates to save. Documentation is a
third contender that is only masked today: it fails at Test building documentation with RuntimeError: Cannot find version '1.1.1' for database 'emodb', a pre-existing failure on main unrelated to this PR,
which aborts the job before the save step runs. Fix that and the same
push produces two warnings instead of one.
The five earlier pushes show zero such annotations; they ran
astral-sh/setup-uv@v5 and their logs are past GitHub's retention
window, so the key they derived can no longer be inspected.
Fix
Keep the single shared key and make exactly one workflow the writer.
save-cache: false (an input of setup-uv v9.0.0, default true) turns
off saving while leaving restore intact:
Test's matrix is a pure 3-OS x 5-Python cross product, so every leg
derives a distinct key and there is no race inside it. Everyone still
restores the same shared cache; only Test writes it.
As a side effect this also removes a collision that would have hit
publish.yml on the next release: its build job is a 6-way platform
matrix that all runs on ubuntu-latest with Python 3.10 and the same
dependency hash, so all six legs derive one key. With save-cache: false
none of them tries to save.
Trade-off
After a cache-key rotation — a pyproject.toml or uv.lock change, a
runner image bump, a Python patch bump — the non-saving workflows miss
once, because the new key does not exist until the next Test run writes
it. That is a fresh uv download in Documentation, Linter and Publish
for a single run: a second or two per job, once. From the following run
on, they restore normally.
Context
Ports audeering/audeer#207
(merged, reviewer-approved), where the single-saver design was chosen
over per-workflow cache-suffix values — the latter removes the warning
by storing several identical copies of the same cache, which was rejected
as wasteful.
audformat and audb get the same treatment.
This repo's main never received the fix: its caching PR merged before
the warning was diagnosed, which is why a41c0bd is the push that shows
it.
Test plan
(Cannot find version '1.1.1' for database 'emodb'), which also
fails on main at a41c0bd
on the head SHA