Git workflow and rebase vs merge questions

TL;DR

A git rebase workflow does not protect you from people who are bad at conflict resolution or people who are used to a SVN workflow, like suggested in Avoiding Git Disasters: A Gory Story. It only makes conflict resolution more tedious for them and makes it harder to recover from bad conflict resolution. Instead, use diff3 so that it's not so difficult in the first place.


Rebase workflow is not better for conflict resolution!

I am very pro-rebase for cleaning up history. However if I ever hit a conflict, I immediately abort the rebase and do a merge instead! It really kills me that people are recommending a rebase workflow as a better alternative to a merge workflow for conflict resolution (which is exactly what this question was about).

If it goes "all to hell" during a merge, it will go "all to hell" during a rebase, and potentially a lot more hell too! Here's why:

Reason #1: Resolve conflicts once, instead of once for each commit

When you rebase instead of merge, you will have to perform conflict resolution up to as many times as you have commits to rebase, for the same conflict!

Real scenario

I branch off of master to refactor a complicated method in a branch. My refactoring work is comprised of 15 commits total as I work to refactor it and get code reviews. Part of my refactoring involves fixing the mixed tabs and spaces that were present in master before. This is necessary, but unfortunately it will conflict with any change made afterward to this method in master. Sure enough, while I'm working on this method, someone makes a simple, legitimate change to the same method in the master branch that should be merged in with my changes.

When it's time to merge my branch back with master, I have two options:

git merge: I get a conflict. I see the change they made to master and merge it in with (the final product of) my branch. Done.

git rebase: I get a conflict with my first commit. I resolve the conflict and continue the rebase. I get a conflict with my second commit. I resolve the conflict and continue the rebase. I get a conflict with my third commit. I resolve the conflict and continue the rebase. I get a conflict with my fourth commit. I resolve the conflict and continue the rebase. I get a conflict with my fifth commit. I resolve the conflict and continue the rebase. I get a conflict with my sixth commit. I resolve the conflict and continue the rebase. I get a conflict with my seventh commit. I resolve the conflict and continue the rebase. I get a conflict with my eighth commit. I resolve the conflict and continue the rebase. I get a conflict with my ninth commit. I resolve the conflict and continue the rebase. I get a conflict with my tenth commit. I resolve the conflict and continue the rebase. I get a conflict with my eleventh commit. I resolve the conflict and continue the rebase. I get a conflict with my twelfth commit. I resolve the conflict and continue the rebase. I get a conflict with my thirteenth commit. I resolve the conflict and continue the rebase. I get a conflict with my fourteenth commit. I resolve the conflict and continue the rebase. I get a conflict with my fifteenth commit. I resolve the conflict and continue the rebase.

You have got to be kidding me if this is your preferred workflow. All it takes is a whitespace fix that conflicts with one change made on master, and every commit will conflict and must be resolved. And this is a simple scenario with only a whitespace conflict. Heaven forbid you have a real conflict involving major code changes across files and have to resolve that multiple times.

With all the extra conflict resolution you need to do, it just increases the possibility that you will make a mistake. But mistakes are fine in git since you can undo, right? Except of course...

Reason #2: With rebase, there is no undo!

I think we can all agree that conflict resolution can be difficult, and also that some people are very bad at it. It can be very prone to mistakes, which why it's so great that git makes it easy to undo!

When you merge a branch, git creates a merge commit that can be discarded or amended if the conflict resolution goes poorly. Even if you have already pushed the bad merge commit to the public/authoritative repo, you can use git revert to undo the changes introduced by the merge and redo the merge correctly in a new merge commit.

When you rebase a branch, in the likely event that conflict resolution is done wrong, you're screwed. Every commit now contains the bad merge, and you can't just redo the rebase*. At best, you have to go back and amend each of the affected commits. Not fun.

After a rebase, it's impossible to determine what was originally part of the commits and what was introduced as a result of bad conflict resolution.

*It can be possible to undo a rebase if you can dig the old refs out of git's internal logs, or if you create a third branch that points to the last commit before rebasing.

Take the hell out of conflict resolution: use diff3

Take this conflict for example:

<<<<<<< HEAD
TextMessage.send(:include_timestamp => true)
=======
EmailMessage.send(:include_timestamp => false)
>>>>>>> feature-branch

Looking at the conflict, it's impossible to tell what each branch changed or what its intent was. This is the biggest reason in my opinion why conflict resolution is confusing and hard.

diff3 to the rescue!

git config --global merge.conflictstyle diff3

When you use the diff3, each new conflict will have a 3rd section, the merged common ancestor.

<<<<<<< HEAD
TextMessage.send(:include_timestamp => true)
||||||| merged common ancestor
EmailMessage.send(:include_timestamp => true)
=======
EmailMessage.send(:include_timestamp => false)
>>>>>>> feature-branch

First examine the merged common ancestor. Then compare each side to determine each branch's intent. You can see that HEAD changed EmailMessage to TextMessage. Its intent is to change the class used to TextMessage, passing the same parameters. You can also see that feature-branch's intent is to pass false instead of true for the :include_timestamp option. To merge these changes, combine the intent of both:

TextMessage.send(:include_timestamp => false)

In general:

  1. Compare the common ancestor with each branch, and determine which branch has the simplest change
  2. Apply that simple change to the other branch's version of the code, so that it contains both the simpler and the more complex change
  3. Remove all the sections of conflict code other than the one that you just merged the changes together into

Alternate: Resolve by manually applying the branch's changes

Finally, some conflicts are terrible to understand even with diff3. This happens especially when diff finds lines in common that are not semantically common (eg. both branches happened to have a blank line at the same place!). For example, one branch changes the indentation of the body of a class or reorders similar methods. In these cases, a better resolution strategy can be to examine the change from either side of the merge and manually apply the diff to the other file.

Let's look at how we might resolve a conflict in a scenario where merging origin/feature1 where lib/message.rb conflicts.

  1. Decide whether our currently checked out branch (HEAD, or --ours) or the branch we're merging (origin/feature1, or --theirs) is a simpler change to apply. Using diff with triple dot (git diff a...b) shows the changes that happened on b since its last divergence from a, or in other words, compare the common ancestor of a and b with b.

    git diff HEAD...origin/feature1 -- lib/message.rb # show the change in feature1
    git diff origin/feature1...HEAD -- lib/message.rb # show the change in our branch
    
  2. Check out the more complicated version of the file. This will remove all conflict markers and use the side you choose.

    git checkout --ours -- lib/message.rb   # if our branch's change is more complicated
    git checkout --theirs -- lib/message.rb # if origin/feature1's change is more complicated
    
  3. With the complicated change checked out, pull up the diff of the simpler change (see step 1). Apply each change from this diff to the conflicting file.


"Conflicts" mean "parallel evolutions of a same content". So if it goes "all to hell" during a merge, it means you have massive evolutions on the same set of files.

The reason why a rebase is then better than a merge is that:

  • you rewrite your local commit history with the one of the master (and then reapply your work, resolving any conflict then)
  • the final merge will certainly be a "fast forward" one, because it will have all the commit history of the master, plus only your changes to reapply.

I confirm that the correct workflow in that case (evolutions on common set of files) is rebase first, then merge.

However, that means that, if you push your local branch (for backup reason), that branch should not be pulled (or at least used) by anyone else (since the commit history will be rewritten by the successive rebase).


On that topic (rebase then merge workflow), barraponto mentions in the comments two interesting posts, both from randyfay.com:

  • A Rebase Workflow for Git: reminds us to fetch first, rebase:

Using this technique, your work always goes on top of the public branch like a patch that is up-to-date with current HEAD.

(a similar technique exists for bazaar)

  • Avoiding Git Disasters: A Gory Story: about the dangers of git push --force (instead of a git pull --rebase for instance)

In my workflow, I rebase as much as possible (and I try to do it often. Not letting the discrepancies accumulate drastically reduces the amount and the severity of collisions between branches).

However, even in a mostly rebase-based workflow, there is a place for merges.

Recall that merge actually creates a node that has two parents. Now consider the following situation: I have two independent feature brances A and B, and now want to develop stuff on feature branch C which depends on both A and B, while A and B are getting reviewed.

What I do then, is the following:

  1. Create (and checkout) branch C on top of A.
  2. Merge it with B

Now branch C includes changes from both A and B, and I can continue developing on it. If I do any change to A, then I reconstruct the graph of branches in the following way:

  1. create branch T on the new top of A
  2. merge T with B
  3. rebase C onto T
  4. delete branch T

This way I can actually maintain arbitrary graphs of branches, but doing something more complex than the situation described above is already too complex, given that there is no automatic tool to do the rebasing when the parent changes.