Re: git rebase/git rebase --abort cause inconsistent state

[Johannes Schindelin <Johannes.Schindelin@gmx.de>]

Hi Junio,

On Mon, 9 Nov 2020, Junio C Hamano wrote:

> Eugen Konkov <kes-kes@yandex.ru> writes:
>
> >> You start at branch dev. Then you use the two argument form
> >
> >>      git rebase dev local/dev
> >
> >> and when you later
> >
> >>      git rebase --abort
> >
> >> then you are not warped back to dev, but to local/dev:
> >
> > I suppose `git rebase --abort` should return me back to `dev`, because
> > this is the state I was before the command. hmm... suppose it will not
> > return to original branch when [branch] parameter is specified for git
> > rebase
>
> Yes, "git rebase [--onto C] A B" has always been a short-hand for
>
> 	git checkout B
> 	git rebase [--onto C] A
>
> which means that if the second rebase step aborts, rebase wants to
> go back to the state before the rebase started, i.e. immediately
> after "checkout B" was done.
>
> I think the root cause of the problem is that addition of the
> "--autostash" feature (which came much later than the two-arg form)
> was designed poorly.  If it wanted to keep the "two-arg form is a
> mere short-hand for checkout followed by rebase" semantics to avoid
> confusing existing users (which is probably a good thing and that
> seems to be what the code does), then the auto-stash should have
> been added _after_ we switch to the branch we rebase, i.e. B.  That
> way, the stash would be applicable if the rebase gets aborted and
> goes back to the original B, where the stash was taken from.

That makes a ton of sense to me.

> Of course, that would also mean that the original modification in
> the working tree and the index may not allow you to move to branch B
> (i.e. starting from your original branch O, and editing files in the
> working tree, "git checkout B" may notice that you edited files that
> are different between O and B and refuse to check out branch B to
> prevent you from losing your local modifications), but that probably
> is a good thing, if "two-arg form is a mere short-hand" paradigm is
> to be kept.  So, "use autostash and you can always rebase in a clean
> state" would no longer hold.

I agree with that, too.

> Another thing we could have done when adding "--autostash", was to
> redefine the meaning of the two-arg form.  Then it starts to make
> sense to take a stash _before_ switching to the branch to be rebased
> (i.e.  B), to go back to the original branch before switching to B,
> and then to unstash on the working tree of the original branch that
> is checked out after aborting.
>
> Note that such an alternative design would have had its own issues.
> With such a different semantics of two-arg form, if a rebase cleanly
> finishes, instead of staying on the rebased branch B, we MUST go
> back to the original branch to unstash what was autostashed.
> Usually people expect after a rebase to play with the rebased state
> (e.g. test build), so staying on branch B that was just rebased
> would be far more usable than going back to unrelated original
> branch (and possibly unstashing).
>
> In any case, the ship has long sailed, so ...

Right. I think by now, the sanest way out of this fix is to do as you say,
stash only _after_ switching to the branch (if that was asked for).

Unfortunately, it is not that trivial to change `git rebase` to autostash
_after_ switching branches: we actually do skip the actual _checkout_
part, for performance reasons, as of 767a9c417eb (rebase -i: stop checking
out the tip of the branch to rebase, 2020-01-24).

My wishful thinking part wants Elijah's merge-ort work to be complete
already so that we can implement a purely in-memory, throw-away
cherry-pick of the autostashed changes on top of the branch to switch to,
and add that as a mandatory check _right after_ autostashing in `git
rebase`. I guess at some stage that will happen.

In the meantime, we might need to disallow `--autostash` with implicit
branch switching.

Ciao,
Dscho