Joachim Breitner
Hi,
as far as I understand, the expected workflow for MRs is that when they are ready, the developer manually squashes the chronological commit history of the MR into a logical one with polished commit messages, typically consisting of a single commit, but could be multiple ones, and then assigns the MR to margebot, which will rebase that sequence of commits onto the staging branch and eventually merges that into master.
One downside of this approach is that it requires destructive changes to work-in-progres branches: I might think the MR is ready, squash the commit sequence into a single commit, but then more work is ready. Now it’s hard to revert individual patches, or collaborate with others, because the git history was disrupted.
Another is that the commit message itself isn’t very easily visible to reviewers.
In other similarly sized projects (e.g. mathlib) I often see a mode where the actual commits of the MR are ignored (so they can represent the true git history of the branch, including merges and all that grit, which is good for collaboration and for reviewers to understand what has happened, without requiring developers to spend cosmetics effort on them), and upon merging by margebot/bors/mergify/whatever, the MR is merged as a single commit with the description taken from the MR description (which encourages developers to keep the MR description up to date as the MR develops, reviewers can easily see that).
A downside of this that you’ll always get one commit on master per MR. If you like to submit a curated list of logical commits within one MR, then this would not work, and you’d have to use multiple MRs.
We would certainly want this to be optional. I, for one, do try when possible to maintain fine-grained histories. Requiring one MR per commit would make this significantly more labor-intensive.
Has this been considered?
(I don’t want to cause unnecessary disruption with a presumptious call for action here; take it as a comment to weigh in in if and when this part of our infrastructure is about to change anyways, or maybe a careful probe if my sentiment may be shared widely.)
Indeed we have considered this and, if we didn't need to use marge-bot, GitLab itself has quite good support for optionally squashing. Sadly though, we do need to use Marge for the reasons described in #19046. In the past we have been rather conservative about what features of Marge we have used since experience has shown it lacking in robustness. In general I'd very much like to move away from Marge but this will require help from GitLab upstream. I have made our need of this feature clear to upstream (and thankfully several other FOSS projects have made similar requests) but progress has been quite slow. Thankfully there appears [1] to be a recent uptick in activity; here's to hoping that it will happen in the next few releases. Cheers, - Ben [1] https://gitlab.com/groups/gitlab-org/-/epics/4911