[RFD] make rebase abort to original branch, not rebased branch

When git rebase is aborted, it will return the HEAD to the rebased branch (which has not moved). If the current branch is "topic" and a run of "git rebase master" is aborted, "topic" is therefore checked out, which is both what I expect and what I want. When the two-argument version of git rebase is used, as in "git rebase master topic", and then aborted, "topic" will be checked out. This is not what I would expect, although I'm aware of it by now, and usually not what I want. I would expect to be returned to the branch/commit that was checked out when I started "git rebase master topic".

In most cases, this is just a small annoyance, since it's usually quick and easy to manually switch back to the original branch. However, I have run into at least two cases where it has been a bit more annoying:

 1. When on a detached HEAD and running "git rebase HEAD topic", if
    you abort the rebase, you will have to look up the old commit in
    the reflog.

 2. I once was rebasing many old branches, mostly to see which ones
    could be rebased without conflicts. The project I was working on
    was pretty big and a lot had changed since those old branches were
    created. If a "new" branch is currently checked out and "git
    rebase master old-branch" is run and causes conflicts, the old
    branch would be checked out. In my case, this would take 5-10
    seconds. Since Junio improved rebase in 0cb0664 (rebase [--onto O]
    A B: omit needless checkout, 2008-03-15), there will be no
    checkouts of any "old" commit in the successful case, so it's a
    shame that it has to happen in the unsuccessful case when the
    rebase is aborted.

Are there valid cases where the current behavior is bettter?

I suppose we can not just change the current behavior due to backward compatibility constraints. We could introduce a configuration variable, of we could perhaps change it in 1.8.0. Maybe the problem is not big enough to do either...

/Martin

Martin von Zweigbergk <martin.von.zweigbergk@gmail.com> writes:

Doesn't this merely show a bad discipline? What were you envisioning to do to your detached HEAD state if the rebase were to succeed? IOW, if the state was so precious, why did you decide to switch to topic and rebase it onto that state, without marking?

I don't particularly like the "when aborted it returns to the original location" behaviour even for a single argument "git rebase A" case (I do deeply care about the tip of the branch that you attempted to rebase _is_ set back to the original state, but I don't care deeply on which branch you would end up on myself), but because "git rebase A B" is a shorthand for "git checkout B; git rebase A" (at least that is how I view it myself), I would imagine that it would be more surprising to switch back to the branch you were on which may not have anything to do with A nor B.

At least going back to B conceptually makes more sense in one use case I have, which was the original reason I invented rebase with the "checkout B and rebase it ono A" shorthand in the first place (see 59e6b23), back when I was an active contributor throwing patches at Linus (note that back then I didn't have "abort then go back" in the code--and that is why I don't care too deeply about this "which branch should I be after aborting?" myself).

The reason I tried to rebase B on A with the short-hand form was because I wanted to clean up what is on B; I may say "abort" when my first attempt to rebase failed because it was a bit too much to bite at once (e.g. the history diverged a bit too much since B forked from A's ancestor).

But then, the next thing I would want to do in such a case after aborting is not to give up and forget about what I needed to do, which is to clean up B into a shape easier to merge with the updated codebase that leads to A. I would want to stay on B and examine the situation a bit deeper, and try to figuire out a different base (e.g. a bit older commit in the history leading to A) to rebase to, so that I can keep up with the other branch incrementally without lagging too far behind. Switching away from the original B would be majorly annoying in such a case.

Junio C Hamano <gitster@pobox.com> writes:

Having said all that, my opinion would be very different if the definition of "git rebase A B" were different. I can imagine that some people may be tempted to (mis)understand the form of the command as a way to "rebase B on top of A, even though I am doing something unrelated to B (my work may or may not be related to A), because I need to help somebody who wants to have an access to an updated B _now_, and I'll continue what I am doing (which is not related to B) after I am done rebasing".

Actually, I would not say that such a workflow is unreasonable.

And for people who want such a workflow out of "rebase A B", it would be a bug that the command does not to come back to your original branch when it finishes rebasing B on top of A, and it would equally be a bug when it stops with a conflict and you tell it to abort. In _both_ cases, you would want to come back to the branch you started from.

But "git rebase A B" after successful completion does not come back to your original branch (it stays on B), because the current semantics of the command is not about supporting such a workflow.

On Sat, 12 Mar 2011, Junio C Hamano wrote:

This usually happens when I see something that looks a bit suspicious in one of my own commits on my topic branch. I then check out that commit to have a look and perhaps run some test. If I find something broken, I would fix it and either amend the commit or create a new commit on top. I would then run "git rebase --onto HEAD HEAD~1 topic" or "git rebase HEAD topic", respectively. If the merge conflicts turn out to be bigger than I expected, I may decide to abort and to instead create a new branch for the commit(s) until I find time/energy.

This has sometimes annoyed me as well. Maybe a --stop/discard would be a nice alternative to "rf -r .git/(rebase-apply|rebase-merge)"?

Certainly a valid use case. Maybe the best solution would be to introduce a new kind of --abort (say --abort-to-branch-before-rebase, but with a better name)?

/Martin

This can be easily fixed for your own repository by installing a pre-rebase hook that force-sets a temporary branch (named e.g. "pre-rebase") to the current commit before starting the rebase. I can highly recommend having such a hook, it's damn useful.