Re: rebase invoking pre-commit

From: Elijah Newren <newren@gmail.com>
Date: Dec 23, 2023, 21:37 UTC
Message-ID: <CABPp-BFbvRDCbMp9Gs9PuV7WfgoVNwyOOn1rB7fe_8UvEEdehA@mail.gmail.com>
In-Reply-To: <m0sf3vi86g.fsf@epic96565.epic.com>

On Thu, Dec 21, 2023 at 12:59 PM Sean Allred <allred.sean@gmail.com> wrote:

Show 9 quoted lines

>
> Is there a current reason why pre-commit shouldn't be invoked during
> rebase, or is this just waiting for a reviewable patch?
>
> This was brought up before at [1] in 2015, but that thread so old at
> this point that it seemed prudent to double-check before investing time
> in a developing and testing a patch.
>
> [1]: https://lore.kernel.org/git/1m55i3m.1fum4zo1fpnhncM%25lists@haller-berlin.de/

I'm very opinionated here. I'm just one person, so definitely take this with a grain of salt, but in my view...

Personally, I think implementing any per-commit hook in rebase by default is a mistake. It enforces a must-be-in-a-worktree-and-the-worktree-must-be-updated-with-every-replayed-commit mindset, which I find highly problematic[2], even if that's "what we always used to do". Because of that, I would prefer to see this at most be a command line flag. However, we've already got a command line flag that while not identical, is essentially equivalent: "--exec $MY_SCRIPT" (it's not the same because it's a post-commit check, but you get notification of any problematic commits, and an immediate stop of the rebase for you to fix up the problematic commit; fixing up the commit shouldn't be problematic since you are, after all, already rebasing).

I see Phillip already responded and suggested not running the pre-commit hook with every commit, but only upon the first commit after a "git rebase --continue". That seems far more reasonable to me than running on every commit...though even that worries me in regards to what assumptions that entails about what is present in the working tree. (For example, what about folks with large repositories, so large that a branch switch or full checkout is extremely costly, and which could benefit from resolving conflicts in a separate sparse-checkout worktree, potentially much more sparse than their main checkout? And what if people like that really fast rebase resolution (namely, done in a separate very sparse checkout which also has the advantage of not polluting your current working tree) so much that they use it on smaller repositories as well? Can I not even experiment with this idea because of the historical per-commit-at-least-as-full-as-main-worktree-checkout assumptions we've baked into rebase?)

While at it, I should also mention that I'm not a fan of the broken pre-rebase hook; its design ties us to doing a single branch at a time. Maybe that hook is not quite as bad, though, since we already broke that hook and no one seemed to care (in particular, when --update-refs was implemented). But if no one seems to care about broken hooks, I think the proper answer is to either get rid of the hook or fix it.

Anyway, as I mentioned, I'm quite opinionated here. To the point that I deemed git-rebase in its current form to possibly be unfixable (after first putting a lot of work into improving it over the past years) and eventually introduced a new "git-replay" command which I hope to make into a competent replacement for it. Given that I do have another command to do my experiments, others on the list may think it's fine to send rebase further down this route, but I was hoping to avoid further drift so that there might be some way of re-implementing rebase on top of my "git-replay" ideas/design.

Just my $0.02, Elijah

[2] https://lore.kernel.org/git/20231124111044.3426007-1-christian.couder@gmail.com/