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

From: Johannes Sixt <j6t@kdbg.org>
Date: Nov 11, 2020, 07:10 UTC
Message-ID: <ea62f9b7-483b-9187-25ee-b6116a25b757@kdbg.org>
In-Reply-To: <nycvar.QRO.7.76.6.2011102312020.18437@tvgsbejvaqbjf.bet>

Am 10.11.20 um 23:28 schrieb Johannes Schindelin:

Show 38 quoted lines

> 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.

Not to me. In particular, I would prefer to move away from the mental model "two-arg form is shorthand for checkout followed by rebase".

First of all, it does not match the mental model of inexperienced users. You have to have been deep in Git operations long enough to know that the two-arg form is implemented by an initial checkout so that the rebase can proceed as if it were the usual one-arg form.

Second, this initial checkout in two-arg form is not necessary at all to begin the rebase. As a first step, the commits to be rebased must be determined. For this, the traditional way is to ask for the range BASE..HEAD (and in order not to change this query for two-arg form, the checkout was added). But the commits can be determined with BASE..${second_arg:-HEAD} without requiring a checkout. Then the first unavoidable checkout is the one that goes to ONTO (with some further shortcuts in an interactive rebase).

I really don't give a dime for the initial checkout. After a botched two-arg rebase, I usually prefer that --abort brings me back to the branch were I was when I started, and not to the branch that was the second arg of the rebase.

>> In any case, the ship has long sailed, so ...

Well, then order it back. rebase is porcelain, not plumbing.

-- Hannes