Re: [PATCH RFC] rebase: respect --ff-only option

From: Phillip Wood <phillip.wood123@gmail.com>
Date: Jul 5, 2021, 13:54 UTC
Message-ID: <41852b7c-6ac7-3a3b-9f2d-6e537f52cd2a@gmail.com>
In-Reply-To: <xmqq8s2lnkta.fsf@gitster.g>

Hi Junio

On 05/07/2021 10:58, Junio C Hamano wrote:

Show 34 quoted lines

> Phillip Wood <phillip.wood123@gmail.com> writes:
> 
>> As I understand it the motivation for this change is to have 'git -c
>> pull.rebase=true pull --ff-only' actually fast forward. Why cant we
>> just change pull not to rebase in that case?
>> ...
>> Is there a use for this outside of 'git pull --ff-only'? I'm far from
>> convinced we want this new option but if we do end up adding it I
>> think it should error out in combination with '-i' or '-x' as '-i'
>> implies the user wants to change the existing commits and '-x' can end
>> up changing them as well.
>>
>> I think this patch addresses a valid problem but it seems to me that
>> the approach taken pushes complexity into rebase to handle a case when
>> pull does not need to invoke rebase in the first place.
> 
> I share the sentiment, but my conclusion would be different.
> 
> Even though we explain that "pull" is _like_ "fetch" followed by
> "merge" (or "rebase"), at the conceptual level, "pull --ff-only"
> should not have to invoke merge or rebase backend.  If the branch
> fast-fowards to the fetched tip, "pull" should be able to just
> update the branch and check it out, and if it doesn't, "pull" should
> just be able to fail.
> 
> Similarly, "pull --ff" (i.e. fast-forwading allowed) should be able
> to just update the branch and check it out when the current tip of
> the branch can be fast-forwarded to the fetched tip.
> 
> But as you said, pull.rebase=interactive will break such a clean
> separation of concerns.  You can leave "pull" oblivious of what
> "rebase -i" wants to do when seeing a fast-forwardable history by
> teaching "rebase" (and "merge") backend to take "--ff-only", "--ff",
> and "--no-ff" options and having "pull" pass them through.

My main concern with the new rebase option was about a user invoking 'git rebase -i --ff-only' directly.

If a user has pull.rebase=interactive and runs 'git pull --ff-only' then I'm not clear what they expect to happen. Assuming we can fast-forward would they expect pull to run 'rebase -i' which would open their editor with the todo list or would they expect that '--ff-only' means "I just want to fast-forward, I don't want to run 'rebase -i'". If it is the latter then we can just invoke 'git merge --ff-only' (so long as we don't mind running the post-merge hook in this case) and not worry about adding more complexity to 'git rebase'

The relevant section of the pull man page only talks about merging in relation to --ff-only

     --ff, --no-ff, --ff-only
	Specifies how a merge is handled when the merged-in history
         is already a descendant of the current history.  --ff is the
         default unless merging an annotated (and possibly signed)
         tag that is not stored in its natural place in the
         refs/tags/ hierarchy, in which case --no-ff is assumed.

         With --ff, when possible resolve the merge as a fast-forward
         (only update the branch pointer to match the merged branch;
         do not create a merge commit). When not possible (when the
         merged-in history is not a descendant of the current
         history), create a merge commit.

	With --no-ff, create a merge commit in all cases, even when
	the merge could instead be resolved as a fast-forward.

	With --ff-only, resolve the merge as a fast-forward when
	possible. When not possible, refuse to merge and exit with a
	non-zero status.

Show 9 quoted lines

> We can teach "pull" that certain backends (namely "rebase -i" in
> this case) want to handle "ff logic" [*] themselves, and other
> backends (i.e. "rebase" and "merge") do not mind if "pull" handles
> "ff logic" for them, but that will break the abstraction (e.g. do we
> really want to teach "pull" that "rebase -i" wants to rewrite all
> the commits even when the history fast-forwards?) and will become a
> source of duplicated logic.
> 
> Another thing to worry about are hooks.  post-rebase

There is no post-rebase hook. There is the post-rewrite hook, I haven't checked if we invoke it with no input of skip it entirely when we fast-forward.

Show 10 quoted lines

> or post-merge
> operations would want to be carried out even when the history would
> fast-forward, and making "pull" to perform the fast-forwarding and
> know which hooks should be called with what parameter so that we
> could pretend as if the "merge" or "rebase" backend was indeed ran,
> breaks the abstraction.
> 
> So, even though I wish that the world was simpler and we could
> handle "ff logic" inside "pull", I am not sure if it is a realistic
> wish.

I think if we decide that 'pull --ff-only' always implies merging then the world stays fairly simple. On the other hand if we want to somehow combine rebasing with --ff-only it will be more complicated. If we go for the latter then unless someone comes up with a good use for 'rebase --ff-only' in another context I would prefer the new rebase option to be marked with PARSE_OPT_HIDDEN and that we also avoid making incidental changes the existing rebase options.

Best Wishes

Phillip

Show 7 quoted lines

> 
> [Footnote]
> 
> * By "ff logic", I am referring to what should happen in the 3x2=6
>    matrix when one of ("--ff", "--ff-only", or "--no-ff") is given
>    and the history (does or does not) fast-forward.
>