Repository navigation
fix(docs): never queue the GPU pass for a busy runner (#174) - #175
Merged
Merged
Conversation
The concurrency group of a documentation caller is held by the whole run. A build-gpu queued for a busy runner made the runs of the next pushes to main wait, CPU build included (OptimalControl.jl, 2026-11-14: three merges not published until the run was force-cancelled). - select-gpu-runner.yml: input when_busy, 'wait' (default, unchanged for ci-gpu.yml) or 'skip' (only an online and idle runner counts, else the new output skip is 'true'); input ctactions_ref. - The decision moves to scripts/select-gpu/pick.sh, tested by scripts/select-gpu/test.sh (19 cases) in select-gpu-tests.yml; an unknown state still means the fallback, never a skip. - documentation.yml: select-gpu with when_busy 'skip'; build-gpu runs only if skip is not 'true', so the site of the CPU build stays. - Docs: select-gpu-runner, documentation (choosing the runner, ordering, manual force-cancel), design (choice, set-aside options, known limits, timeline), guide, principles, README. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Written by Claude (Claude Code) at the request of ocots.
Closes #174.
Why
A documentation caller with a GPU pass (OptimalControl.jl, CTSolvers.jl) has a concurrency group per ref that queues runs on
main(ordering of the deployments, no cancellation of a Julia process on a self-hosted runner). The group is held by the whole run: abuild-gpuqueued for a busy runner made the next runs wait, CPUbuildincluded, with no time limit of its own. On 2026-11-14, OptimalControl.jl'sdevwas not updated for three merges until the stuck run was force-cancelled (details in #174).What
select-gpu-runner.yml: new inputwhen_busy:wait(default): today's behaviour, unchanged forci-gpu.yml: only offline counts, a busy runner is waited for;skip: a runner counts only if it is online and idle;kkt, elseoccidata, else the new outputskipis'true'andrunneris empty.New input
ctactions_ref(defaultmain), as inguard.yml, for the script checkout. An unknown state (noCT_RUNNERS_*secrets, API error) still gives the fallback, never a skip (principle 7).The decision moves to a script (principles 2 and 3):
scripts/select-gpu/pick.sh, tested byscripts/select-gpu/test.sh(19 cases: both modes, several runners per label, another GPU runner ignored, unknown state, a badWHEN_BUSY), run by the newselect-gpu-tests.yml.select-gpu-runner.ymlleaves the "only linted" exceptions.documentation.yml:select-gpuwithwhen_busy: skip;build-gpuruns only ifskip != 'true'. When skipped, the site published bybuildstays, as when the pass fails.Callers: no change. Their concurrency group stays; it no longer waits on a pass queued for a busy runner.
Docs:
workflows/select-gpu-runner.md(inputs, outputs, rules),workflows/documentation.md(choosing the runner, ordering, how to force-cancel a stuck queued run),design/documentation.md(the choice, the options set aside, two known limits, timeline),guides/documentation.md,principles.md,README.md.Options set aside
cancel-in-progress: trueonmainin the callers: cancels a running GPU pass, the orphan-process case.buildandbuild-gpuwith a "still the head ofmain?" check before the GPU deployment: the deployment happens inside each package'sdocs/make.jl, in the same step as the build, so CTActions cannot place the check.Known limits
select-gpureads it can be taken beforebuild-gpustarts; the pass then waits for that job, bounded by its timeout. The manual way out is documented.Checked
bash scripts/select-gpu/test.sh: 19/19;shellcheck scripts/select-gpu/*.sh: clean;actionlinton every workflow: clean.gh api … | jq -sc .as the workflow):kktoffline,occidataonline and busy →wait:runner=occidata,skip=false(unchanged for the tests);skip:runner=,skip=true(no pass queued, today's case);runner=occidata,skip=false.select-gpu-tests.ymlandlint-workflows.ymlrun on this PR. The reusable workflows call each other at@main(uses:takes no expression), sodocumentation.ymlwithwhen_busy: skipis first exercised by the next docs run of OptimalControl.jl or CTSolvers.jl onmainafter the merge.🤖 Generated with Claude Code