[webkit-dev] Proposal: Mandatory Commit and Merge Queue

Mon Jun 6 11:22:53 PDT 2022

Thanks for gathering this data!

>> What are some notable cases of recent regressions that have landed because of non-use of commit queue and caused serious problems?
> 
> Some recent examples of regressions that would have been prevented by mandatory commit/merge-queue as proposed:
> 
> https://commits.webkit.org/250940@main <https://commits.webkit.org/250940@main> (PR did use Merge-Queue, but we’re missing the feature that would have caught the problem)

What feature specifically?

> https://commits.webkit.org/250343@main <https://commits.webkit.org/250343@main> (Commit did use Commit-Queue, but skipped layout tests because of a previous iteration passed EWS)

Are you also proposing that merge-queue should re-run layout tests unconditionally? Do we know how long that takes on average? Do we have the hardware to support that?

> https://commits.webkit.org/250791@main <https://commits.webkit.org/250791@main> (Should have failed EWS, but flakey tests had different names, cross referencing trunk results would have made that obvious)

How would merge-queue cross reference with trunk results or otherwise resolve this confusion?

> https://commits.webkit.org/248624@main <https://commits.webkit.org/248624@main> (Landed manually, broke the build)

Seems like this would have been avoided by Merge-Queue alone. That’s nice.

For the other three cases, whether the proposal as-is would have resolved them or not seems muddy. Might need revision.

Thanks,
Geoff

>> Do you have any data on how frequent such regressions are, compared to the base rate of regressions that have landed despite use of commit queue?
> 
> Commit and Merge queue have basically prevented us from breaking the Mac build, the only example of a broken Mac build I can find come from by-passing Commit and Merge Queue. Layout test breakages are quite a bit more difficult to reason about because breakages tend to not just be platform specific, but configuration specific as well.
> 
>> 
>> Do you have any data on how frequently regressions are resolved by patches that land outside commit queue?
> 
> In the last week, we had 200 commits. 50 of those were made through the Unsafe-Merge-Queue. Break down is here:
> 
> 15 were feature work that should have gone through Merge Queue
> 12 were test gardening
> 9 were build or test fixes
> 3 were 3rd party imports
> 3 were reverts
> 3 were tooling changes
> 2 were buildbot changes
> 2 were contributors.json changes
> 1 was a documentation change
> 
> My proposal would have sent the feature work, gardening work, imports, tooling change and documentation change to Merge-Queue rather than Unsafe-Merge-Queue. The build and test fixes would have needed modification to their commit messages, but were the kind of changes I would want landed via Unsafe-Merge-Queue.
> 
>> 
>>> and reduce demands of post-commit test gardening.
>> 
>> Is this goal distinct from preventing regressions from landing? If so, how?
> 
> I suppose I should have said:
> "The goal is to increase the stability of the build, speed up EWS and reduce demands of post-commit test gardening by preventing regressions from landing”, since build instability, slow EWS and post-commit test gardening are all consequences of regressions landing.
> 
>> 
>>>> What problem are you trying to solve, and with what level of urgency?
>>> 
>>> Urgency is not high. I started this with the expectation it would be a somewhat long and contentious discussion. The motivating change is that the GitHub transition makes this proposal possible, from a technical perspective, in a way it is not while the project is still backed by Subversion.
>> 
>> I don’t understand the premise here. There are lots of ways to enforce commit policy on a Subversion repository.
> 
> Kind of. The problem here is that we want to provide enough escape hatches so that any committer can quickly repair a broken build, so we have to check not just the committer, but the commit itself. This is more analogous to ensuring that commit has “Reviewed by” in it’s commit message (something that Subversion does not enforce in our repository, despite it being policy) than it is to ensuring that the committer has certain privileges. We’ve always implemented these kind of checks in buildbot, not the Subversion server, and it’s much easier to implement complicated logic that takes into consideration the content of the commit in buildbot than it is Subversion hooks.
> 
>> On the meta level, while we are still dealing with serious regressions in our workflow caused by git and GitHub, I recommend that we do not push forward with more unrelated sweeping changes to the project and its workflow. Just like in software development, where we need to fix regressions before we can move forward with major feature work, so too in tooling we need to do the same. Otherwise we just pile chaos on top of chaos, and there is no way to know if things are improving or getting worse, and no way to hold ourselves accountable for improvement. 
>> 
>> Thanks,
>> Geoff
>> 
>>> 
>>> Jonathan
>>> 
>>>> 
>>>> Thanks,
>>>> Geoff
>>>> 
>>>> 
>>>>> On Jun 2, 2022, at 2:35 PM, Jonathan Bedard via webkit-dev <webkit-dev at lists.webkit.org <mailto:webkit-dev at lists.webkit.org>> wrote:
>>>>> 
>>>>> As we move to GitHub, I would like to propose we strengthen our protections on `main` by making MergeQueue and CommitQueue mandatory. This would mean that with a few exceptions, all changes would need to be built and run layout tests before they are landed. To spell out what the exceptions I had in mind are:
>>>>> 
>>>>> - Revert commits, identified by a commit message that starts with “Unreviewed, revering…” would be exempt
>>>>> - Changes which only modify files that do not effect building or testing WebKit would be exempt. These files specifically are:
>>>>>          .github/
>>>>>          JSTests/
>>>>>          ManualTests/
>>>>>          metadata/
>>>>>          PerformanceTests/
>>>>>          Tools/
>>>>> 	         CISuport/
>>>>> 	         EWSTools/
>>>>> 	         WebKitBot
>>>>>         Websites/
>>>>> - Emergency build and infrastructure fixes, identified by a commit message that starts with “Emergency build fix” or “Emergency infrastructure fix” would be exempt
>>>>> - A reviewer who is not the commit author can overwrite this protection by adding `unsafe-merge-queue` instead of the commit author
>>>>> - Changes which passed an EWS layout test queue within the last 7 days would skip the layout test check
>>>>> 
>>>>> These exceptions are designed to provide contributors for a way to by-pass potentially slow checks if extraordinary situations, or in ones where CI has already validated the change. I think we should keep the ability for any committer to deploy an emergency fix, because our project has many contributors in different timezones and with different holiday schedules.
>>>>> 
>>>>> We know that this policy change would potentially slow down development, so I think these 3 improvements block making MergeQueue and CommitQueue mandatory:
>>>>> 
>>>>> - run-webkit-tests consulting results.webkit.org <http://results.webkit.org/> to avoid retrying known flakey or failing tests
>>>>> - Another MergeQueue bot
>>>>> - Xcode workspace builds to speed up incremental builds
>>>>> 
>>>>> Jonathan Bedard
>>>>> WebKit Continuous Integration
>>>>> 
>>>>> _______________________________________________
>>>>> webkit-dev mailing list
>>>>> webkit-dev at lists.webkit.org <mailto:webkit-dev at lists.webkit.org>
>>>>> https://lists.webkit.org/mailman/listinfo/webkit-dev <https://lists.webkit.org/mailman/listinfo/webkit-dev>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.webkit.org/pipermail/webkit-dev/attachments/20220606/260eb79c/attachment.htm>