How and/or why is merging in Git better than in SVN?

The claim of why merging is better in a DVCS than in Subversion was largely based on how branching and merge worked in Subversion a while ago. Subversion prior to 1.5.0 didn't store any information about when branches were merged, thus when you wanted to merge you had to specify which range of revisions that had to be merged.

So why did Subversion merges suck?

Ponder this example:

      1   2   4     6     8
trunk o-->o-->o---->o---->o
       \
        \   3     5     7
b1       +->o---->o---->o

When we want to merge b1's changes into the trunk we'd issue the following command, while standing on a folder that has trunk checked out:

svn merge -r 2:7 {link to branch b1}

… which will attempt to merge the changes from b1 into your local working directory. And then you commit the changes after you resolve any conflicts and tested the result. When you commit the revision tree would look like this:

      1   2   4     6     8   9
trunk o-->o-->o---->o---->o-->o      "the merge commit is at r9"
       \
        \   3     5     7
b1       +->o---->o---->o

However this way of specifying ranges of revisions gets quickly out of hand when the version tree grows as subversion didn't have any meta data on when and what revisions got merged together. Ponder on what happens later:

           12        14
trunk  …-->o-------->o
                                     "Okay, so when did we merge last time?"
              13        15
b1     …----->o-------->o

This is largely an issue by the repository design that Subversion has, in order to create a branch you need to create a new virtual directory in the repository which will house a copy of the trunk but it doesn't store any information regarding when and what things got merged back in. That will lead to nasty merge conflicts at times. What was even worse is that Subversion used two-way merging by default, which has some crippling limitations in automatic merging when two branch heads are not compared with their common ancestor.

To mitigate this Subversion now stores meta data for branch and merge. That would solve all problems right?

And oh, by the way, Subversion still sucks…

On a centralized system, like subversion, virtual directories suck. Why? Because everyone has access to view them… even the garbage experimental ones. Branching is good if you want to experiment but you don't want to see everyones' and their aunts experimentation. This is serious cognitive noise. The more branches you add, the more crap you'll get to see.

The more public branches you have in a repository the harder it will be to keep track of all the different branches. So the question you'll have is if the branch is still in development or if it is really dead which is hard to tell in any centralized version control system.

Most of the time, from what I've seen, an organization will default to use one big branch anyway. Which is a shame because that in turn will be difficult to keep track of testing and release versions, and whatever else good comes from branching.

So why are DVCS, such as Git, Mercurial and Bazaar, better than Subversion at branching and merging?

There is a very simple reason why: branching is a first-class concept. There are no virtual directories by design and branches are hard objects in DVCS which it needs to be such in order to work simply with synchronization of repositories (i.e. push and pull).

The first thing you do when you work with a DVCS is to clone repositories (git's clone, hg's clone and bzr's branch). Cloning is conceptually the same thing as creating a branch in version control. Some call this forking or branching (although the latter is often also used to refer to co-located branches), but it's just the same thing. Every user runs their own repository which means you have a per-user branching going on.

The version structure is not a tree, but rather a graph instead. More specifically a directed acyclic graph (DAG, meaning a graph that doesn't have any cycles). You really don't need to dwell into the specifics of a DAG other than each commit has one or more parent references (which what the commit was based on). So the following graphs will show the arrows between revisions in reverse because of this.

A very simple example of merging would be this; imagine a central repository called origin and a user, Alice, cloning the repository to her machine.

         a…   b…   c…
origin   o<---o<---o
                   ^master
         |
         | clone
         v

         a…   b…   c…
alice    o<---o<---o
                   ^master
                   ^origin/master

What happens during a clone is that every revision is copied to Alice exactly as they were (which is validated by the uniquely identifiable hash-id's), and marks where the origin's branches are at.

Alice then works on her repo, committing in her own repository and decides to push her changes:

         a…   b…   c…
origin   o<---o<---o
                   ^ master

              "what'll happen after a push?"


         a…   b…   c…   d…   e…
alice    o<---o<---o<---o<---o
                             ^master
                   ^origin/master

The solution is rather simple, the only thing that the origin repository needs to do is to take in all the new revisions and move it's branch to the newest revision (which git calls "fast-forward"):

         a…   b…   c…   d…   e…
origin   o<---o<---o<---o<---o
                             ^ master

         a…   b…   c…   d…   e…
alice    o<---o<---o<---o<---o
                             ^master
                             ^origin/master

The use case, which I illustrated above, doesn't even need to merge anything. So the issue really isn't with merging algorithms since three-way merge algorithm is pretty much the same between all version control systems. The issue is more about structure than anything.

So how about you show me an example that has a real merge?

Admittedly the above example is a very simple use case, so lets do a much more twisted one albeit a more common one. Remember that origin started out with three revisions? Well, the guy who did them, lets call him Bob, has been working on his own and made a commit on his own repository:

         a…   b…   c…   f…
bob      o<---o<---o<---o
                        ^ master
                   ^ origin/master

                   "can Bob push his changes?" 

         a…   b…   c…   d…   e…
origin   o<---o<---o<---o<---o
                             ^ master

Now Bob can't push his changes directly to the origin repository. How the system detects this is by checking if Bob's revisions directly descents from origin's, which in this case doesn't. Any attempt to push will result into the system saying something akin to "Uh... I'm afraid can't let you do that Bob."

So Bob has to pull-in and then merge the changes (with git's pull; or hg's pull and merge; or bzr's merge). This is a two-step process. First Bob has to fetch the new revisions, which will copy them as they are from the origin repository. We can now see that the graph diverges:

                        v master
         a…   b…   c…   f…
bob      o<---o<---o<---o
                   ^
                   |    d…   e…
                   +----o<---o
                             ^ origin/master

         a…   b…   c…   d…   e…
origin   o<---o<---o<---o<---o
                             ^ master

The second step of the pull process is to merge the diverging tips and make a commit of the result:

                                 v master
         a…   b…   c…   f…       1…
bob      o<---o<---o<---o<-------o
                   ^             |
                   |    d…   e…  |
                   +----o<---o<--+
                             ^ origin/master

Hopefully the merge won't run into conflicts (if you anticipate them you can do the two steps manually in git with fetch and merge). What later needs to be done is to push in those changes again to origin, which will result into a fast-forward merge since the merge commit is a direct descendant of the latest in the origin repository:

                                 v origin/master
                                 v master
         a…   b…   c…   f…       1…
bob      o<---o<---o<---o<-------o
                   ^             |
                   |    d…   e…  |
                   +----o<---o<--+

                                 v master
         a…   b…   c…   f…       1…
origin   o<---o<---o<---o<-------o
                   ^             |
                   |    d…   e…  |
                   +----o<---o<--+

There is another option to merge in git and hg, called rebase, which'll move Bob's changes to after the newest changes. Since I don't want this answer to be any more verbose I'll let you read the git, mercurial or bazaar docs about that instead.

As an exercise for the reader, try drawing out how it'll work out with another user involved. It is similarly done as the example above with Bob. Merging between repositories is easier than what you'd think because all the revisions/commits are uniquely identifiable.

There is also the issue of sending patches between each developer, that was a huge problem in Subversion which is mitigated in git, hg and bzr by uniquely identifiable revisions. Once someone has merged his changes (i.e. made a merge commit) and sends it for everyone else in the team to consume by either pushing to a central repository or sending patches then they don't have to worry about the merge, because it already happened. Martin Fowler calls this way of working promiscuous integration.

Because the structure is different from Subversion, by instead employing a DAG, it enables branching and merging to be done in an easier manner not only for the system but for the user as well.

Historically, Subversion has only been able to perform a straight two-way merge because it's didn't store any merge information. This involves taking a set of changes and applying them to a tree. Even with merge information, this is still the most commonly-used merge strategy.

Git uses a 3-way merge algorithm by default, which involves finding a common ancestor to the heads being merged and making use of the knowledge that exists on both sides of the merge. This allows Git to be more intelligent in avoiding conflicts.

Git also has some sophisticated rename finding code, which also helps. It doesn't store changesets or store any tracking information -- it just stores the state of the files at each commit and uses heuristics to locate renames and code movements as required (the on-disk storage is more complicated than this, but the interface it presents to the logic layer exposes no tracking).

Put simply, the merge implementation is done better in Git than in SVN. Before 1.5 SVN did not record a merge action, so it was incapable to do future merges without help by the user which needed to provide information that SVN did not record. With 1.5 it got better, and indeed the SVN storage model is slightly more capable that Git's DAG. But SVN stored the merge information in a rather convoluted form that lets merges take massively more time than in Git - I've observed factors of 300 in execution time.

Also, SVN claims to track renames to aid merges of moved files. But actually it still stores them as a copy and a separate delete action, and the merge algorithm still stumbles over them in modify/rename situations, that is, where a file is modified on one branch and rename on the other, and those branches are to be merged. Such situations will still produce spurious merge conflicts, and in the case of directory renames it even leads to silent loss of modifications. (The SVN people then tend to point out that the modifications are still in the history, but that doesn't help much when they aren't in a merge result where they should appear.

Git, on the other hand, does not even track renames but figures them out after the fact (at merge time), and does so pretty magically.

The SVN merge representation also has issues; in 1.5/1.6 you could merge from trunk to branch as often as just liked, automatically, but a merge in the other direction needed to be announced (--reintegrate), and left the branch in an unusable state. Much later they found out that this actually isn't the case, and that a) the --reintegrate can be figured out automatically, and b) repeated merges in both directions are possible.

But after all this (which IMHO shows a lack of understanding of what they are doing), I'd be (OK, I am) very cautions to use SVN in any nontrivial branching scenario, and would ideally try to see what Git thinks of the merge result.

Other points made in the answers, as the forced global visibility of branches in SVN, aren't relevant to merge capabilities (but for usability). Also, the 'Git stores changes while SVN stores (something different)' are mostly off the point. Git conceptually stores each commit as a separate tree (like a tar file), and then uses quite some heuristics to store that efficiently. Computing the changes between two commits is separate from the storage implementation. What is true is that Git stores the history DAG in a much more straightforward form that SVN does its mergeinfo. Anyone trying to understand the latter will know what I mean.

In a nutshell: Git uses a much simpler data model to store revisions than SVN, and thus it could put a lot of energy into the actual merge algorithms rather than trying to cope with the representation => practically better merging.