Hi all,
We are planning to enable GitHub's merge queue on the Comet main branch to
reduce CI resource usage. This note explains what changes for contributors
and committers, and gives everyone a chance to raise concerns before it is
turned on. The tracking issue is
https://github.com/apache/datafusion-comet/issues/5838.
Why
CI load on pull requests is dominated by suites that do not need to be green
until merge time. Today Spark 3.5 SQL tests run on every push to every PR, and
Spark 3.4/4.0 and the older Iceberg suites run again after every merge. With a
merge queue, GitHub builds the exact tree that will land (up to five PRs
batched together) and runs CI against that once. An iterating PR runs a
smaller set, and nothing runs twice.
What changes
1. Two CI tiers, decided by which event triggered the run.
PR tier (every push to a PR): Linux and macOS build and unit tests,
benchmark check, Spark 4.1 SQL tests, Iceberg 1.11.
Queue tier (when a PR is queued to merge): everything above, plus Spark
3.4, 3.5 and 4.0 SQL tests and Iceberg 1.8, 1.9 and 1.10.
Spark 3.5 is the one contributors will notice: it currently runs on every
PR push and becomes queue-only. If you need it on a PR, add the
run-spark-3.5-tests label (run-spark-3.4-tests, run-spark-4.0-tests and
run-iceberg-tests work the same way).
2. How merging works. Instead of "Squash and merge", committers click "Merge
when ready". GitHub adds the PR to the queue, runs the queue tier against
the merge result, and merges automatically when the single required check
("Required Checks") is green. A PR can be queued before its own PR-tier
checks have finished.
3. No more post-merge test runs. Because the queue already tested the tree
that landed, the heavy jobs no longer run on push to main. Only the docs
site deploy does.
Impact on open PRs
The required check is new. Any open PR whose last CI run happened before
https://github.com/apache/datafusion-comet/pull/5842 merged will not have a
"Required Checks" result and cannot be queued until it gets a fresh CI run.
Pushing a new commit or rebasing is enough.
Flaky tests
Once the older Spark and Iceberg suites gate the queue, a flaky test in any of
them blocks everyone's merges rather than leaving main red for one commit.
That raises the bar on flakiness. Please treat any flake seen in the queue
tier as a bug to fix or quarantine promptly, and file an issue when you hit
one.
Rollout
Step 1, https://github.com/apache/datafusion-comet/pull/5842, adds the
"Required Checks" aggregator job with nothing requiring it yet, so we can
watch it report correctly on real PRs first. Step 2,
https://github.com/apache/datafusion-comet/pull/5843, declares it as the
required status check and turns the queue on via a ruleset in .asf.yaml. The
queue parameters (batch up to 5 PRs, 2 groups in flight, no minimum wait, 5
hour check timeout) are easy to tune afterwards. ASF Infra (apache/root) is a
bypass actor so a wedged queue can always be recovered.
If there are no objections we will merge step 1 shortly and step 2 once it
has reported cleanly on a few PRs.
Thanks,
Andy
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
