Skip to content

ci(pstack): test macOS with its stock bash only, since its runners are scarce - #25

Merged
mstephenholl merged 1 commit into
mainfrom
ci-drop-macos-bash
Oct 20, 2026
Merged

mstephenholl merged 1 commit into
mainfrom
ci-drop-macos-bash

Conversation

@mstephenholl

Copy link
Copy Markdown
Owner

Why

GitHub's macOS arm64 runners are short on capacity. The Install (macos-latest, bash) job was cancelled five times in a row today, with the annotation "The job was not acquired by Runner of type hosted even after multiple attempts". That failed CI gate on main, so #23 never released, and it blocks #22.

What changes

The install matrix drops the macOS bash entry. Ubuntu still tests a current bash. The macOS /bin/bash job still tests bash 3.2 and macOS's stat and date, and it got a runner on every attempt today. Each run now needs one macOS runner instead of two.

Risk

A bug that shows up only with a current bash on macOS's userland would go uncaught. Neither of the other two jobs would catch it.

Proof

test_github_settings.py and check.py pass. Merging this releases #23's change too, since the release covers everything since v0.2.14.

…e scarce

GitHub's macOS arm64 runners are short on capacity, and the macos-latest bash job was cancelled five times in a row without getting a runner, which blocked CI gate on main and on #22. Ubuntu already covers a current bash, and the /bin/bash job still covers bash 3.2 and macOS's stat and date.
@mstephenholl
mstephenholl merged commit 1b0b579 into main Oct 20, 2026
10 checks passed
@mstephenholl
mstephenholl deleted the ci-drop-macos-bash branch October 20, 2026 14:38
mstephenholl added a commit that referenced this pull request Oct 20, 2026
#27)

## Why

#25 cut the macOS install job to one, `/bin/bash`, because the other one
kept failing to get a runner. Today the `/bin/bash` job failed the same
way three times in a row ("The job was not acquired by Runner of type
hosted even after multiple attempts"): once on `main` after #25 merged,
and twice on #22. GitHub cancels a job that waits about 15 minutes for a
macOS runner. `CI gate` treats a cancelled job as a failure, and the
`main` ruleset requires the gate with no bypass, so each shortage blocks
every merge and release until a retry gets lucky.

## What changes

- **`pstack-ci.yml`**: the install job runs on Ubuntu only. Every job in
this file still feeds `CI gate`.
- **New `pstack-macos.yml`**: the same install steps on `macos-latest`
with `/bin/bash`, on every push to `main`, daily at 06:23 UTC, and on
demand. It has no `pull_request` trigger, so it never adds a check to a
pull request. A newer commit cancels a run that is still waiting for a
runner, so a burst of merges doesn't queue several runs for scarce
runners.
- **Tests**: `CiWiringTest` now fails if any gated job runs on macOS.
The new `MacosWorkflowTest` fixes the triggers, the read-only
permissions, and the cancel-on-newer concurrency group, and checks that
the macOS steps match the gated Ubuntu install job's steps except for
`/bin/bash`. Without that check, the two copies could drift apart now
that they live in separate files.
- **`port/README.md`**: the CI section describes the new workflow, its
trade-off, and how to test a branch on macOS before merging (`gh
workflow run pstack-macos.yml --ref <branch>`).

## Risk

A bug that shows up only on macOS (bash 3.2, BSD `stat` or `date`) can
now merge and ship in a release before its macOS run reports. The run on
`main` still catches it within hours, but the fix lands after the
release.

## Proof

- Each new test fails on the drift it guards against: dropping a step
from the macOS job, adding a `pull_request` trigger, and `main`'s
current `pstack-ci.yml` with macOS in the gate.
- All 248 unit tests pass, `install_test.sh` passes, and `check.py`
reports 0 problems. Both workflows parse, with the triggers and runners
listed above.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant