[webkit-dev] Proposal: Immediate Deprecation of ChangeLogs

[Jonathan Bedard]

I’m measuring from “engineer tells GitHub to land a change”, not “Merge-Queue checks out the change”. The basic steps of the process are:

- GitHub sends hook to buildbot
- buildbot validates hook and PR
- buildbot checks out PR
- buildbot inserts reviewer to commit message and ChangeLog
- buildbot runs basic pre-commit checks
- buildbot adds identifier
- buildbot commits
- buildbot updates PR with landed commit
- buidbot comments on PR and bug with information on landed commit
- buildbot closes associated bug

The “commit” step usually happens at 2 minute mark currently, I gave the time for the whole process (including bug comments, bug closing, etc). It will be a bit faster once we can drop the `git-svn` pieces of this, it’s the “checkout” and “add identifier” step that can take longer than you would expect. 10 minutes is a worst-case scenario when there are already other changes in the unsafe queue and some bots are busy with Merge-Queue and Commit-Queue requests, so there is only a single available bot serving the unsafe queue.

Jonathan

> On May 11, 2022, at 4:31 PM, Chris Dumez <cdumez at apple.com> wrote:
> 
> 
>> On May 11, 2022, at 11:56 AM, Jonathan Bedard via webkit-dev <webkit-dev at lists.webkit.org> wrote:
>> 
>> Trying to embed previous replies is going to get messy, will be referencing those replies but not embedding them.
>> 
>> Unsafe-Merge-Queue should be very fast, I haven’t seen anything take longer than 10 minutes from label application to landing or rejection. The average case is 3-4 minutes. We’re aware of what the architectural problem with Commit-Queue is that slows down the fast path, Unsafe-Merge-Queue has fixed that. The solution we used isn’t transferable to Commit-Queue.
> 
> I personally don’t think that delaying a critical build fix / gardening by 10 minutes is OK. It’s called “unsafe-merge-queue”, why does it take this long to commit? Why isn't it as fast as "generate a unique identifier and git push”?
> Sure, I have no idea how the unsafe-merge-queue does but it is hard for me to imagine it should take more than 1 minute given what it needs to do and given that it should need to do extremely little testing given that it is “unsafe”.
> 
> 4 to 10 minutes is nowhere near as fast as `git svn dcommit` and is therefore a pretty large regression.