Sharing merge conflict resolution between multiple developers

Hi List,

At $dayjob we maintain internal forks of the a number of upstream repositories.

Unsurprisingly updating these forks can be extremely problematic, especially when it's only one person doing the merge. Fortunately most of us are in the same physical location so it is possible to drag in someone who knows more about the code than the person merging but I can't see that scaling with remote developers.

Is there any way where we could share the conflict resolution around but still end up with a single merge commit. I'm thinking of something like the following workflow

developer A:
  git merge $upstream
  <conflicts>
  git mergetool ...
  <resolves some of the conflicts>
  git commit -m "WIP: Merge upstream" --something-like--all-but-not

developer B:
  git pull developer_A
  git reset HEAD^
  <at this point it's be nice if developer B's work tree and index
were in the same state as developer A's before they ran 'git commit'>
  git mergetool ...
  <resolves the conflicts that they know about>
  git commit -m "WIP: Merge upstream" --something-like--all-but-not

developer A:
  git pull developer_B
  git reset HEAD^
  git mergetool ....
  <finishes off anything left>
  git commit
  <It'd be really nice if MERGE_MSG was retained through all of this>
  git push

Any thoughts on if something like this is currently possible? Is this something other git users would find useful?

Thanks, Chris

Chris Packham <judge.packham@gmail.com> writes:

One idea that immediately comes to me is to use something like "rerere" (not its implementation and storage, but the underlying idea) enhanced with the trick I use to fix-up merges in my daily integration cycle (look for "merge-fix" in howto/maintain-git.txt in Documentation/).

And then commit this immediately, together with conflict markers (i.e. "commit -a"), and discard it with "reset --hard HEAD^" *after* storing it somewhere safe. And then redo the same merge, resolve the conflicts and commit the usual way.

The difference between the final conflict resolution and the original conflicted state can be used as a reference for others to redo the same conflict resolution later elsewhere. That can most easily be done by creating a commit that records the final state whose parent is the one you recorded the initial conflicted state.

So, the "recording" phase may go something like this:

    git checkout $this
    git merge $that
    git commit -a -m 'merge-fix/$this-$that preimage'
    git branch merge-fix/$this-$that
    git reset --hard HEAD^
    git merge $that
    edit
    git commit -a -m 'merge $that to $this'
    git checkout merge-fix/$this-$that
    git read-tree -m -u HEAD $this
    git commit -a -m 'merge-fix/$this-$that postimage'

The rough idea is "git show merge-fix/$this-$that" will show the "patch" you can apply on top of the conflicted state other people would get by running "git merge $that" while on "$this" branch.

"rerere" essentially does the above recording (and replaying) per-path and it comes with a clever indexing scheme to identify which previous conflict resolution would apply to the conflicts you see in your working tree.

On Tue, Aug 12, 2014 at 6:44 AM, Junio C Hamano <gitster@pobox.com> wrote:

So how would someone else pickup that postimage and use it?

  git checkout $this
  git merge $that
  git fetch $remote ':/merge-fix/$this-$that postimage'
  git show ':/merge-fix/$this-$that postimage' | git apply (or patch -p1)
  edit

I feel a toy patch coming on.

On Mon, Aug 11, 2014 at 4:29 PM, Chris Packham <judge.packham@gmail.com> wrote:

For a simpler case that would work, but because we are not saving just a patch but two full trees to compare (i.e. merge-fix/$this-$that is the postimage, its ^1 is the preimage), you should be able to use the three-way merge in a similar way cherry-pick works. In fact, that is how rerere replays the recorded resolution, not with a "patch -p1".

IIUC, this might help,

http://www.mail-archive.com/git@vger.kernel.org/msg56418.html http://www.mail-archive.com/git@vger.kernel.org/msg56468.html

On Mon, Aug 11, 2014 at 04:59:15PM +1200, Chris Packham wrote:

This came up once a while back. Here's the discussion:

  http://thread.gmane.org/gmane.comp.version-control.git/169187

I proposed a solution there where developers push their partial resolutions, and the integrator (or just "next person" if it is a chain of developers) uses "git checkout" to pull the fixes. But note that I never actually _used_ that in practice. It was a thought experiment, so there may be gotchas.

-Peff

I am working on an open source project right now where we are looking to enforce a N of M audit approval process. It turns out that git supports verifying multiple signatures because gpg supports signature merging.

My question is how can this workflow best be added into git and if not added atleast supported.

Here are the manual procedures (scripts are in the bundle too):

To do this most properly I feel like there needs to be a way to "share" the repository state and intterrupt the commit process.

Comments?

1: 
$ cat merge-multisigs.sh
#!/bin/bash
(
 for i in "$@"
 do
  gpg --dearmor < "$i"
 done
) | gpg --enarmor

2: $ cat write-commit.ruby #!/usr/bin/irb require 'fileutils' file = File.open(ARGV[0], "rb") content = file.read header = "commit #{content.length}\0" store = header + content require 'digest/sha1' sha1 = Digest::SHA1.hexdigest(store) require 'zlib' zlib_content = Zlib::Deflate.deflate(store) path = '.git/objects/' + sha1[0,2] + '/' + sha1[2,38] FileUtils.mkdir_p(File.dirname(path)) File.open(path, 'w') { |f| f.write zlib_content }

P.S. This was inspired by actual events and the parent thread.

--
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
-                                                               -
- Jason Pyeron                      PD Inc. http://www.pdinc.us -
- Principal Consultant              10 West 24th Street #100    -
- +1 (443) 269-1555 x333            Baltimore, Maryland 21218   -
-                                                               -
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
This message is copyright PD Inc, subject to license 20080407P00.

On Sun, Aug 17, 2014 at 09:30:47AM -0400, Jason Pyeron wrote:

In the scheme you propose, the commit object is actually rewritten. So whoever made and signed it first will then need to rebase on top of the rewritten multi-signed version.

Is there a reason not to use detached signatures, and let each person add them after the fact? You can store them in git-notes and then push them along with the other commits (you can even check in a pre-receive hook that the commits meet your N of M criteria, as long as everybody has pushed up their signature notes).

I think this is just "git hash-object -w -t commit <file>", isn't it?

-Peff

Not exactly. A known and shared commit is used as the parent of an empty, no changes commit. The "no changes" commit object is taken and passed around before being added into the repository. There is no need for a rebase.

But my scheme uses out-of-band process to accomplish this. The idea of using git to "distribute" the conflict resolution seemed like a valid use case of sharing a working copy state for a distributed commit, just like this. [1][2]

Yes. The embeded signatures provides the best user experience (UX) for verification.

Let me find the most complicated way of saying this, yes. I feel silly for that.

-Jason

[1]: http://git.661346.n2.nabble.com/Sharing-a-massive-distributed-merge-td6178696.html [2]: http://git.661346.n2.nabble.com/Sharing-merge-conflict-resolution-between-multiple-developers-td7616700.html

Le 11 août 2014 07:50, "Christian Couder" <christian.couder@gmail.com> a écrit :

Sorry I sent the above using the gmail app on my mobile phone and unfortunately I can't make it send plain text emails. (Emails which are not plain text are rejected by vger.kernel.org.)

Best, Christian.