Traduction incomplète

Cette page n'est pas encore entièrement traduite : les passages non traduits sont affichés en anglais.

0040. Merge queue

  • Status: accepted

  • Date: 2026-11-08

Context

The main ruleset requires pull requests to be up to date with main before they merge (ADR 0014). After every merge, each open pull request has to receive a merge commit from main (« Update branch ») and run the whole CI again, one after the other, before the next one can merge.

GitHub’s merge queue does that work itself. It is available in any public repository owned by an organization. The maintainer queues pull requests whose checks pass; for each one, GitHub builds a temporary branch gh-readonly-queue/main/pr-<number>-<sha> that holds main, the pull requests ahead in the queue and this one, runs the required checks on it through the merge_group event, and merges the entries in order. A required check that does not report on that branch blocks the queue; a job that is skipped counts as passing.

Decision

main is merged through the merge queue. Every workflow that reports a required check also runs on merge_group (types: [checks_requested]), and each check keeps the same name, so the queue waits for the same thirteen checks as a pull request. A test (tests/test_ci_scripts.py) fails when one of these workflows loses the trigger.

What each check does on a queue entry:

Check

On the queue entry

Lint and types, Tests (SQLite), Tests (PostgreSQL), Django and repository checks

Run as on a pull request, on the code of the entry: what lands on main is what was tested.

zizmor

Runs on the workflows of the entry. The SARIF upload job stays limited to main and to pull requests.

Dependency review

The action reads merge_group.base_sha and head_sha and reviews the dependencies of the entry against the advisories of that moment.

Container image

Builds, smoke-tests and scans the image of the entry. A vulnerability published while the pull request waited is found here. Its SARIF upload stays limited to main and to pull requests.

Container image (multi-arch build)

Builds amd64 and arm64 again. It takes about three minutes, like Container image, so the queue keeps it; skipping it would also turn it into a passing check without a build.

Documentation

Builds and checks the site of the entry. Documentation (external links), not required, does not run on the queue: an outage elsewhere must not hold it up.

Conventional Commits title

Reads the title of the queued pull request again from the API and checks it. A title edited after queuing, which becomes the squash commit message, is still checked before it reaches main.

Signed-off-by

Reads the queued pull request from the API, fetches its head (refs/pull/<n>/head, which also covers forks), checks that it is the head the API returned, and checks the sign-off of its commits as on the pull request. The commits of the queue branch itself are written by GitHub, not by a contributor, so they are not what the sign-off certifies.

Documentation updated

Same as Signed-off-by, with the diff of the pull request and its labels of that moment.

CLA

A separate job of cla.yml, CLA queue, checks the signatures of the queued pull request (its commit authors and its opener) again and sets the CLA status on the head of the entry. It only reads and sets statuses: no signature is recorded and no comment is posted from the queue.

No check passes without checking. The four checks that read the pull request find it from the name of the queue branch (scripts/merge_queue.py, and the same pattern in cla.js), and fail when the branch is not a queue branch of main or when the pull request is not open against main. Each queue entry names one pull request and runs its own checks, so with the grouping that requires every entry to pass (« only merge non-failing pull requests »), each pull request of a group is checked on its own entry.

Code scanning. CodeQL runs in default setup, which analyzes pull requests, pushes to main and a weekly schedule; it cannot run on merge_group. The ruleset’s code_scanning rule is not a status check, and GitHub documents that it does not apply to merge queue entries. A pull request is still analyzed and gated before it can be queued, and main after each merge. What is lost is the analysis of the exact combination before the merge; restoring it means an advanced CodeQL workflow with a merge_group trigger, which replaces default setup.

Settings of the queue (ruleset main, applied by the maintainer): squash, as today; only non-failing pull requests are merged (ALLGREEN); up to five entries built and merged at a time; no minimum group size; one hour for the checks to report. The rule « branches must be up to date » is removed: the queue tests every entry on the latest main.

Consequences

  • The maintainer no longer updates pull request branches after each merge. « Merge when ready » queues a pull request once its checks pass; GitHub merges it once the checks of its entry pass, or removes it from the queue and says why.

  • Every merge runs the required checks twice, once on the pull request and once on its entry: about three more minutes of waiting per entry (the longest jobs, the two image builds), on the free runners of a public repository.

  • Push nothing to a queued pull request: remove it from the queue first. The checks that read the pull request read its head when they run, and a push while it is queued is not a case this design relies on.

  • The commit on main is the squash commit created by the queue, as it was the squash commit created by the merge button.

  • A new required check needs its workflow to run on merge_group before it is added to the ruleset, or the queue waits for it until the timeout.

Modifier cette page sur GitHub Signaler un problème sur cette page