Re: interactive rebase results across shared histories

From: Eric Sunshine <sunshine@sunshineco.com>
Date: Feb 21, 2016, 19:08 UTC
Message-ID: <CAPig+cSxmWc_Guab0UoQbRMEkVLr-qhF=LCiVk10G5AdnTqnGA@mail.gmail.com>
In-Reply-To: <8737sm6kmk.fsf@gmail.com>

On Sun, Feb 21, 2016 at 12:25 PM, Seb <spluque@gmail.com> wrote:

Show 15 quoted lines

> The scenario is much simpler; imagine master has a longer history behind
> the point where the topic branch started:
>
>                 A---B---C topic
>                /
>   *---D---E---F---G master
>
> And we want to keep both branches separate (no desire to merge them for
> now), but we realize that, say, commits D and E should be
> squashed/fixup, so we do an interactive rebase.  Now, the problem is
> that if I do that from the topic branch, the results are not reflected
> in the master branch, even though these commits are certainly shared
> with master.  It seems counterintuitive that a part of history that is
> shared among branches can be independently manipulated/rewritten with
> rebase.  I must be missing something...

What you're probably missing is that you can't actually edit commits in Git. Instead, what you think of as "editing" actually creates a new commit with its own commit-ID, and the original commit still exists with its own commit-ID. Since Git commits are chained together by their commit-ID's, any commits pointing at the original commit-ID continue to point to that commit, and only commits rebased atop the new commit-ID of the "edited" commit point at it.

In your example, you're "editing" D and E, which creates new commits D' and E', so your resulting graph looks like this:

    D'---E'---A---B---C topic
   /
  *---D---E---F---G master

So, "master" and "topic" really are not sharing D and E (or D' and E'). You could "fix" this to match your intuition by rebasing F...G onto E' (see git-rebase --onto, for instance), which would give you this:

                  A---B---C topic
                 /
  *---D'---E'---F---G master

and then "master" and "topic" would really be sharing D' and E' as common history. (Of course, rebasing "master" or any branch may not be desirable if you've published it, so applicable warnings about rebasing apply.)

By the way, the problem isn't restricted to when you rebase "topic" (as your problem description implies). You'd see the same behavior if you'd rebased D and E in "master" to become D' and E'. "topic" would still have old D and E in its history, and not D' and E'.