Re: git merge banch w/ different submodule revision

From: Elijah Newren <newren@gmail.com>
Date: Apr 27, 2018, 00:19 UTC
Message-ID: <CABPp-BE5jRG8JdDfH1XG-Btz9jJxfwf_oyNni8Ci1j+J3icbVQ@mail.gmail.com>
In-Reply-To: <1524739599.20251.17.camel@klsmartin.com>

On Thu, Apr 26, 2018 at 3:49 AM, Middelschulte, Leif <Leif.Middelschulte@klsmartin.com> wrote:

Show 26 quoted lines

> Hi,
>
> we're using git-flow as a basic development workflow. However, doing so revealed unexpected merge-behavior by git.
>
> Assume the following setup:
>
> - Repository `S` is sourced by repository `p` as submodule `s`
> - Repository `p` has two branches: `feature_x` and `develop`
> - The revisions sourced via the submodule have a linear history
>
>
> * 1c1d38f (feature_x) update submodule revision to b17e9d9
> | * 3290e69 (HEAD -> develop) update submodule revision to 0598394
> |/
> * cd5e1a5 initial submodule revision
>
>
> Problem case: Merge either branch into the other
>
> Expected behavior: Merge conflict.
>
> Actual behavior: Auto merge without conflicts.
>
> Note 1: A merge conflict does occur, if the sourced revisions do *not* have a linear history
>
> Did I get something wrong about how git resolves merges? Shouldn't git be like: "hey, you're trying to merge two different contents for the same line" (the submodule's revision)

Hard to say without saying what commit was referenced for the submodule in the merge-bases for the two repositories you have. In the basic case..

If branch A and branch B have different commits checked out in the
submodule, say:
   A: deadbeef
   B: ba5eba11

then it's not clear whether there's a conflict or not.  The merge-base
(the common point of history) matters.  So, for example if the
original version (which I'll refer to as 'O") had:
  O: deadbeef

then you would say, "Oh, branch A made no change to this submodule but B did. So let's go with what B has." Conversely, of O had ba5eba11, then you'd go the other way.

But, there is some further smarts in that if either A or B point at commits that contain the other in their history and both contain the commit that O points at, then you can just do a fast-forward update to the newest.

You didn't tell us how the merge-base (cd5e1a5 from the diagram you gave) differed in your example here between the two repositories. In fact, the non-linear case could have several merge-bases, in which case they all become potentially relevant (as does their merge-bases since at that point you'll trigger the recursive portion of merge-recursive). Giving us that info might help us point out what happened, though if either the fast-forward logic comes into play or the recursive logic gets in the mix, then we may need you to provide a testcase (or access to the repo in question) in order to explain it and/or determine if you've found a bug.

Does that help?

Elijah