git-imerge -- incremental merge and rebase for Git

Perform a merge between two branches incrementally. If conflicts are encountered, figure out exactly which pairs of commits conflict, and present the user with one pairwise conflict at a time for resolution.

git-imerge has two primary design goals:

• Reduce the pain of resolving merge conflicts to its unavoidable minimum, by finding and presenting the smallest possible conflicts: those between the changes introduced by one commit from each branch.

• Allow a merge to be saved, tested, interrupted, published, and collaborated on while it is in progress.

I think that it is easiest to understand the concept of incremental merging visually, and therefore I recommend the video of my git-imerge presentation from the GitMerge 2013 conference (20 min) as a good place to start. The full slides for that talk are available in this repository under doc/presentations/GitMerge-2013. At the same conference, I was interviewed about git-imerge by Thomas Ferris Nicolaisen for his GitMinutes Podcast #12.

To learn how to use the git-imerge tool itself, I suggest the blog article git-imerge: A Practical Introduction and also typing git-imerge --help and git-imerge SUBCOMMAND --help. If you want more information, the theory and benefits of incremental merging are described in minute detail in a series of blog articles [1], as are the benefits of retaining history when doing a rebase [2].

Multiple incremental merges can be in progress at the same time. Each incremental merge has a name, and its progress is recorded in the Git repository as references under refs/imerge/NAME. The current state of an incremental merge can be visualized using the diagram command.

An incremental merge can be interrupted and resumed arbitrarily, or even pushed to a server to allow somebody else to work on it.

git-imerge comes with a Bash completion script, completions/git-imerge, which is installed automatically when installing git-imerge.

Requirements

git-imerge requires:

• A Python interpreter; either

  • Python 3.x, version 3.3 or later.

  • Python 2.x, version 2.6 or later. If you are using Python 2.6.x, then you have to install the argparse module yourself, as it was only added to the standard library in Python 2.7.

• A recent version of Git.

Bash completion requires Git's completion being available.

Installation

git-imerge is available on PyPI, so you can install it with pip:

$ pip install git-imerge

or using setup.py if you have downloaded the source package locally:

$ python setup.py install

If you prefer, you can use Homebrew to install git-imerge:

$ brew install git-imerge

Instructions

To start a merge or rebase operation using git-imerge, you use commands that are similar to the corresponding git commands:

git-imerge drop is also analogous to running git rebase --interactive, then deleting the specified commit(s) from the history. Both the drop and the revert subcommands are included in git-imerge because the equivalent git operations can conflict, so they both can benefit from using the incremental merge approach.)

A few more options are available if you start the incremental merge using git imerge start:

$ git-imerge start --name=NAME --goal=GOAL [--first-parent] BRANCH

where

• NAME is the name for this merge (and also the default name of the branch to which the results will be saved).

• GOAL describes how you want to simplify the results (see next section).

After the incremental merge is started, you will be presented with any conflicts that have to be resolved. The basic procedure is similar to performing an incremental merge using git:

while not done:
    <fix the conflict that is presented to you>
    <"git add" the files that you changed>
    git-imerge continue

When you have resolved all of the conflicts, you finish the incremental merge by typing:

git-imerge finish

That should be enough to get you going. All of these subcommands have additional options; to learn about them type:

git-imerge --help
git-imerge SUBCMD --help

Simplifying results

When the incremental merge is finished, you can simplify its results in various ways before recording it in your project's permanent history by using either the finish or simplify command. The "goal" of the incremental merge can be one of the following:

• merge — keep only a simple merge of the second branch into the first branch, discarding all intermediate merges. The end result is similar to what you would get from

    git checkout BRANCH1
    git merge BRANCH2
  

• rebase — keep the versions of the commits from the second branch rebased onto the first branch. The end result is similar to what you would get from

    git checkout BRANCH2
    git rebase BRANCH1
  

• rebase-with-history — like rebase, except that it retains the old versions of the rebased commits in the history. It is equivalent to merging the commits from BRANCH2 into BRANCH1, one commit at a time. In other words, it transforms this:

    o---o---o---o          BRANCH1
         \
          A---B---C---D    BRANCH2
  

  into this:

    o---o---o---o---A'--B'--C'--D'   NEW_BRANCH
         \         /   /   /   /
          --------A---B---C---D
  

  It is safe to rebase an already-published branch using this approach. See [2] for more information.

• full — don't simplify the incremental merge at all: do all of the intermediate merges and retain them all in the permanent history. In other words, it transforms this:

    o---o---1---2---3      BRANCH1
         \
          A---B---C---D    BRANCH2
  

  into this:

    o---o---1---2---3
         \   \   \   \
          A---A1--A2--A3
           \   \   \   \
            B---B1--B2--B3
             \   \   \   \
              C---C1--C2--C3
               \   \   \   \
                D---D1--D2--D3    NEW_BRANCH
  

  This approach retains the complete history and ancestry information, which gives the maximum flexibility for conducting future merges. On the other hand, it clutters up the permanent Git history considerably.

• border — this experimental goal retains the rebase of BRANCH2 onto BRANCH1 and also the rebase of BRANCH1 onto BRANCH2, plus a merge commit that includes both branches. In other words, it transforms this:

    o---o---1---2---3      BRANCH1
         \
          A---B---C---D    BRANCH2
  

  into this:

    o---o---1---2---3
         \           \
          A           A2
           \           \
            B           B2
             \           \
              C           C2
               \           \
                D---D1--D2--D3    NEW_BRANCH
  

  This approach leaves more history than a simple merge or rebase, possibly making future merges easier.

• border-with-history — this experimental goal retains the rebase-with-history of BRANCH2 onto BRANCH1 and also the rebase (without history) of BRANCH1 onto BRANCH2, plus a merge commit that includes both branches. In other words, it transforms this:

    o---o---1---2---3      BRANCH1
         \
          A---B---C---D    BRANCH2
  

  into this:

    o---o---1---2---3
         \           \
          A-----------A3
           \           \
            B-----------B3
             \           \
              C-----------C3
               \           \
                D---D1--D2--D3    NEW_BRANCH
  

  This approach leaves more history and ancestry information than a simple merge or rebase, possibly making future merges easier.

• border-with-history2 — this experimental goal retains the rebase-with-history of BRANCH1 onto BRANCH2 and also the rebase-with-history of BRANCH2 onto BRANCH1, plus a merge commit that includes both branches. In other words, it transforms this:

    o---o---1---2---3      BRANCH1
         \
          A---B---C---D    BRANCH2
  

  into this:

    o---o---1---2---3
         \   \   \   \
          A--- --- ---A3
           \   \   \   \
            B--- --- ---B3
             \   \   \   \
              C--- --- ---C3
               \   \   \   \
                D---D1--D2--D3    NEW_BRANCH
  

  This approach leaves more history and ancestry information than a simple merge or rebase, possibly making future merges easier.

Technical notes

Suspending/resuming

When git-imerge needs to ask the user to do a merge manually, it creates a temporary branch refs/heads/imerge/NAME to hold the result. If you want to suspend an incremental merge to do something else before continuing, all you need to do is abort any pending merge using git merge --abort and switch to your other branch. When you are ready to resume the incremental merge, just type git imerge continue.

If you need to completely abort an in-progress incremental merge, first remove the temporary branches git-imerge creates using git-imerge remove, then checkout the branch you were in before you started the incremental merge with git checkout ORIGINAL_BRANCH.

Storage

git-imerge records all of the intermediate state about an incremental merge in the Git object database as a bunch of references under refs/imerge/NAME, where NAME is the name of the imerge:

• refs/imerge/NAME/state points to a blob that describes the current state of the imerge in JSON format; for example,

  • The tips of the two branches that are being merged

  • The current "blocker" merges (merges that the user will have to do by hand), if any

  • The simplification goal

  • The name of the branch to which the result will be written.

• refs/imerge/NAME/manual/I-J and refs/imerge/NAME/auto/I-J refer to the manual and automatic merge commits, respectively, that have been done so far as part of the incremental merge. I and J are integers indicating the location (I,J) of the merge in the incremental merge diagram.

Transferring an in-progress imerge between repositories

It might sometimes be convenient to transfer an in-progress incremental merge from one Git repository to another. For example, you might want to make a backup of the current state, or continue an imerge at home that you started at work, or ask a colleague to do a particular pairwise merge for you. Since all of the imerge state is stored in the Git object database, this can be done by pushing/fetching the references named in the previous section. For example,

$ git push --prune origin +refs/imerge/NAME/*:refs/imerge/NAME/*

or

$ git fetch --prune origin +refs/imerge/NAME/*:refs/imerge/NAME/*

Please note that these commands overwrite any state that already existed in the destination repository. There is currently no support for combining the work done by two people in parallel on an incremental merge, so for now you'll just have to take turns.

Interaction with git rerere

git rerere is a nice tool that records how you resolve merge conflicts, and if it sees the same conflict again it tries to automatically reuse the same resolution.

Since git-imerge attempts so many similar test merges, it is easy to imagine rerere getting confused. Moreover, git-imerge relies on a merge resolving (or not resolving) consistently if it is carried out more than once. Having rerere store extra information behind the scenes could therefore confuse git-imerge.

Indeed, in testing it appeared that during incremental merges, the interaction of git-imerge with rerere was sometimes causing merge conflicts to be resolved incorrectly. Therefore, git-imerge explicitly turns rerere off temporarily whenever it runs any git commands.

Log messages for pairwise merge commits

When git imerge continue or git imerge record finds a resolved merge in the working tree, it commits that merge then incorporates it into the incremental merge. Usually it just uses Git's autogenerated commit message for such commits. If you want to be prompted to edit such commit messages, you can either specify --edit on the command line or change the default in your configuration:

$ git config --global imerge.editmergemessages true

Testing

git-imerge uses tox to run tests. To run the test suite with the system's default Python:

$ tox

To run with a specific Python version, such as 3.7, pass the -e argument to tox:

$ tox -e py37

License

git-imerge is released as open-source software under the GNU General Public License (GPL), version 2 or later. See file COPYING for more information.

References

[1]

• https://softwareswirl.blogspot.com/2012/12/the-conflict-frontier-of-nightmare-merge.html
• https://softwareswirl.blogspot.com/2012/12/mapping-merge-conflict-frontier.html
• https://softwareswirl.blogspot.com/2012/12/real-world-conflict-diagrams.html
• https://softwareswirl.blogspot.com/2013/05/git-incremental-merge.html
• https://softwareswirl.blogspot.com/2013/05/one-merge-to-rule-them-all.html
• https://softwareswirl.blogspot.com/2013/05/incremental-merge-vs-direct-merge-vs.html
• https://softwareswirl.blogspot.com/2013/05/git-imerge-practical-introduction.html

[2]

• https://softwareswirl.blogspot.com/2009/04/truce-in-merge-vs-rebase-war.html
• https://softwareswirl.blogspot.com/2009/08/upstream-rebase-just-works-if-history.html
• https://softwareswirl.blogspot.com/2009/08/rebase-with-history-implementation.html