Re: why "git rebase" searching the duplicate patches in <upstream branch> rather than in <new base branch>?

From: Sergey Organov <sorganov@gmail.com>
Date: Jul 20, 2021, 09:04 UTC
Message-ID: <87a6mhgxv4.fsf@osv.gnss.ru>
In-Reply-To: <xmqqmtqij63t.fsf@gitster.g>

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

Show 18 quoted lines

> Andy Zhang <zhgdrx@gmail.com> writes:
>
>> why "git rebase" searching the duplicate patches in <upstream
>> branch> rather than in <new base branch>?
>>
>> hi, all:
>>
>>  I am reading the help of "git rebase", it says:
>>     "If the upstream branch already contains a change you have made
>> (e.g., because you mailed a patch which was applied upstream), then
>> that commit will be skipped. "
>>
>>  But, because we are applying commits to <new base branch> rather than
>> to <upstream branch>, I really don't understand why we are searching
>> the duplicate patches in <upstream branch> rather than in <new base
>> branch>?
>
> It is either a design bug or a documentation bug, or both ;-)

It's definitely /at least/ a documentation bug, as description of the feature is not precise enough. For example, it's unclear if such a commit will appear in the todo list of --interactive. Will it?

It looks like documentation of "git rebase" should be revised to make clearer distinction between <branch>, <upstream>, and <newbase>.

Show 6 quoted lines

>
> I do think it makes sense to skip commits from the branch we are
> rebasing that have equivalent commits in the upstream, as it is
> expected that upstream might have already applied/cherry-picked some
> of the changes you are rebasing, and you do not want to use the same
> change twice.

To me this only makes sense for the branch we rebase /onto/, and thus it actually makes sense for <newbase>, and for <upstream> it only happens to make sense by default as <newbase>=<upstream> in this case.

If Git currently indeed searches for duplicates in <upstream>, then it looks like implementation bug, or misfeature. I think the <newbase> should rather be used.

Show 15 quoted lines

>
> When we are transplanting a series of commits from an old base to
> totally unrelated base using the --onto option, e.g. when replaying
> the contents of 'topic' relative to 'next' down to 'master' in your
> topology, however,
>
>> Old tree is:
>>
>> o---o---o---o---o  master
>>     \
>>      o---o---o---o---o  next
>>                       \
>>                        o---o---o  topic
>
> it is not necessarily obvious where to stop digging back at.

Similar problem should exist for explicitly specified <upstream> that might happen to have little in common with the current <branch>, right? If so, then it's already somehow being solved, even if simply by ignoring the issue, so adding <newbase> to the picture doesn't actually bring anything significantly new.

Show 9 quoted lines

> In the
> above picture where 'master' and 'next' have ancestry relationship,
> we could try to see if the three commits on 'topic' branch being
> replayed match any of the commits in next..master range, but when
> using the --onto option, there does not have to be any relationship
> between the <upstream> and <new base> (they do not have to share a
> root commit).  So from that point of view, it probably makes sense
> to default to --no-reapply-cherry-picks when --onto is used, while
> defaulting --reapply-cherry-picks when --onto is not used.

I don't actually like this.

First, in general, changing default of another option is not to be taken lightly. For example, defaulting to --fork-point when no <upstream> is specified is already a point of confusion.

Second, changing the default is not backward compatible, so there should be very sound reason to change it.

Finally, if user does specify /both/ --onto and --no-reapply-cherry-picks, where would Git supposedly stop digging for matching cherry-picks? Provided this is to be solved anyway, the rationale to change the default does not sound strong enough.

Overall, it seems that we should take the <newbase> rather than <upstream> (that is still <upstream> when --onto is not specified), and apply the skipping logic from there, to whatever depth the merge-base will give us. If it's already implemented this way, then only the manual page needs to be fixed.

Thanks,

-- 
Sergey Organov