Repository navigation
Conversation
The Rocky, Debug and Debian build workflows ran on every push to main, so each merged pull request was built and tested a second time. That second run tests nothing new: the main branch protection in .asf.yaml uses strict status checks, so a pull request must be up to date with main before it is merged, and only squash and rebase merges are allowed. The tree that lands on main is therefore the tree the pull request runs were made on. For example, PR apache#1942 was rebased onto the then-current main, and its merged head 4730e4c has the same tree as the pull request head 6c9a5fa, yet all of these workflows ran on it again for up to almost four hours. Such reruns only add load on the runners and delay the queue, and a flaky test in them turns main red although the change was already green. Drop main from the push trigger of these three workflows. Keep it for REL_2_STABLE, which has no required status checks, so pushes there are still tested. Leave the other workflows alone: Behave only runs on pull requests touching management code, so its push run is the only one for other changes, SQLancer fuzzes with random input, and the Rat check is cheap. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
🟢 Approval recommended
There are no active review issues with the CI workflow trigger changes.
No open issues 🎉
What changed in this PR
This PR reduces duplicate CI work by stopping the Rocky Linux, Debug, and Debian build workflows from running again after merges to main.
Changes:
- Remove
mainfrom the three workflows’ push triggers while retaining pushes toREL_2_STABLE, pull requests to both branches, and manual runs. - Update workflow comments and README trigger documentation to reflect the change and explain the branch-protection rationale.
| File | Description |
|---|---|
.github/workflows/README.md |
Clarifies push triggers for the main and debug builds. |
.github/workflows/build-deb-cloudberry.yml |
Removes Debian builds on pushes to main and updates trigger comments. |
.github/workflows/build-dbg-cloudberry.yml |
Removes debug builds on pushes to main and updates trigger comments. |
.github/workflows/build-cloudberry.yml |
Removes Rocky Linux builds on pushes to main and updates trigger comments. |
Review effort: Balanced
Last week we used 520,300 minutes (https://github.com/apache/cloudberry/actions/metrics/usage?dateRangeType=DATE_RANGE_TYPE_CURRENT_WEEK), double our quota. This is mainly because adding more tests to the test matrix is a strong quality gate but also increases runner minutes. Previously, I didn’t want to cancel the push running in main after merging; since we are exceeding our runner quota, I now agree with canceling it. |
The Rocky, Debug and Debian build workflows ran on every push to main, so each merged pull request was built and tested a second time. That second run tests nothing new: the main branch protection in .asf.yaml uses strict status checks, so a pull request must be up to date with main before it is merged, and only squash and rebase merges are allowed. The tree that lands on main is therefore the tree the pull request runs were made on. For example, PR #1942 was rebased onto the then-current main, and its merged head 4730e4c has the same tree as the pull request head 6c9a5fa, yet all of these workflows ran on it again for up to almost four hours.
Such reruns only add load on the runners and delay the queue, and a flaky test in them turns main red although the change was already green.
Drop main from the push trigger of these three workflows. Keep it for REL_2_STABLE, which has no required status checks, so pushes there are still tested. Leave the other workflows alone: Behave only runs on pull requests touching management code, so its push run is the only one for other changes, SQLancer fuzzes with random input, and the Rat check is cheap.
One could see the issue in https://github.com/apache/cloudberry/actions?query=is%3Ain_progress page. After merge here we run tests once again. Just spend CPU in a vain )