Merging: Hg/Git vs. SVN

I often read that Hg (and Git and...) are better at merging than SVN but I have never seen practical examples of where Hg/Git can merge something where SVN fails (or where SVN needs manual intervention). Could you post a few step-by-step lists of branch/modify/commit/...-operations that show where SVN would fail while Hg/Git happily moves on? Practical, not highly exceptional cases please...

Some background: we have a few dozen developers working on projects using SVN, with each project (or group of similar projects) in its own repository. We know how to apply release- and feature-branches so we don't run into problems very often (i.e., we've been there, but we've learned to overcome Joel's problems of "one programmer causing trauma to the whole team" or "needing six developers for two weeks to reintegrate a branch"). We have release-branches that are very stable and only used to apply bugfixes. We have trunks that should be stable enough to be able to create a release within one week. And we have feature-branches that single developers or groups of developers can work on. Yes, they are deleted after reintegration so they don't clutter up the repository. ;)

So I'm still trying to find the advantages of Hg/Git over SVN. I'd love to get some hands-on experience, but there aren't any bigger projects we could move to Hg/Git yet, so I'm stuck with playing with small artificial projects that only contain a few made up files. And I'm looking for a few cases where you can feel the impressive power of Hg/Git, since so far I have often read about them but failed to find them myself.

I too have been looking for a case where, say, Subversion fails to merge a branch and Mercurial (and Git, Bazaar, ...) does the right thing.

The SVN Book describes how renamed files are merged incorrectly. This applies to Subversion 1.5, 1.6, 1.7, and 1.8! I have tried to recreate the situation below:

cd /tmp
rm -rf svn-repo svn-checkout
svnadmin create svn-repo
svn checkout file:///tmp/svn-repo svn-checkout
cd svn-checkout
mkdir trunk branches
echo 'Goodbye, World!' > trunk/hello.txt
svn add trunk branches
svn commit -m 'Initial import.'
svn copy '^/trunk' '^/branches/rename' -m 'Create branch.'
svn switch '^/trunk' .
echo 'Hello, World!' > hello.txt
svn commit -m 'Update on trunk.'
svn switch '^/branches/rename' .
svn rename hello.txt hello.en.txt
svn commit -m 'Rename on branch.'
svn switch '^/trunk' .
svn merge --reintegrate '^/branches/rename'

According to the book, the merge should finish cleanly, but with wrong data in the renamed file since the update on trunk is forgotten. Instead I get a tree conflict (this is with Subversion 1.6.17, the newest version in Debian at the time of writing):

--- Merging differences between repository URLs into '.':
A    hello.en.txt
C hello.txt
Summary of conflicts:
Tree conflicts: 1

There shouldn't be any conflict at all — the update should be merged into the new name of the file. While Subversion fails, Mercurial handles this correctly:

rm -rf /tmp/hg-repo
hg init /tmp/hg-repo
cd /tmp/hg-repo
echo 'Goodbye, World!' > hello.txt
hg add hello.txt
hg commit -m 'Initial import.'
echo 'Hello, World!' > hello.txt
hg commit -m 'Update.'
hg update 0
hg rename hello.txt hello.en.txt
hg commit -m 'Rename.'
hg merge

Before the merge, the repository looks like this (from hg glog):

@  changeset:   2:6502899164cc
|  tag:         tip
|  parent:      0:d08bcebadd9e
|  user:        Martin Geisler <[email protected]>
|  date:        Thu Apr 01 12:29:19 2010 +0200
|  summary:     Rename.
|
| o  changeset:   1:9d06fa155634
|/   user:        Martin Geisler <[email protected]>
|    date:        Thu Apr 01 12:29:18 2010 +0200
|    summary:     Update.
|
o  changeset:   0:d08bcebadd9e
user:        Martin Geisler <[email protected]>
date:        Thu Apr 01 12:29:18 2010 +0200
summary:     Initial import.

The output of the merge is:

merging hello.en.txt and hello.txt to hello.en.txt
0 files updated, 1 files merged, 0 files removed, 0 files unresolved
(branch merge, don't forget to commit)

In other words: Mercurial took the change from revision 1 and merged it into the new file name from revision 2 (hello.en.txt). Handling this case is of course essential in order to support refactoring and refactoring is exactly the kind of thing you will want to do on a branch.

I do not use Subversion myself, but from the release notes for Subversion 1.5: Merge tracking (foundational) it looks like there are the following differences from how merge tracking work in full-DAG version control systems like Git or Mercurial.

• Merging trunk to branch is different from merging branch to trunk: for some reason merging trunk to branch requires --reintegrate option to svn merge.

In distributed version control systems like Git or Mercurial there is no technical difference between trunk and branch: all branches are created equal (there might be social difference, though). Merging in either direction is done the same way.

• You need to provide new -g (--use-merge-history) option to svn log and svn blame to take merge tracking into account.

In Git and Mercurial merge tracking is automatically taken into account when displaying history (log) and blame. In Git you can request to follow first parent only with --first-parent (I guess similar option exists also for Mercurial) to "discard" merge tracking info in git log.

• From what I understand svn:mergeinfo property stores per-path information about conflicts (Subversion is changeset-based), while in Git and Mercurial it is simply commit objects that can have more than one parent.

• "Known Issues" subsection for merge tracking in Subversion suggests that repeated / cyclic / reflective merge might not work properly. It means that with the following histories second merge might not do the right thing ('A' can be trunk or branch, and 'B' can be branch or trunk, respectively):
  

  ------x------y------------M2        <-- A
  \       \             /
  ------M1------*---/           <-- B
  

In the case the above ASCII-art gets broken: Branch 'B' is created (forked) from branch 'A' at revision 'x', then later branch 'A' is merged at revision 'y' into branch 'B' as merge 'M1', and finally branch 'B' is merged into branch 'A' as merge 'M2'.

------x--------M1--------M2          <-- A
\       /           /
----y---------/             <-- B

In the case the above ASCII-art gets broken: Branch 'B' is created (forked) from branch 'A' at revision 'x', it is merged into branch 'A' at 'y' as 'M1', and later merged again into branch 'A' as 'M2'.

• Subversion might not support advanced case of criss-cross merge.

---b-----B1--M1-----M3
\     \ /        /
\     X        /
\   / \      /
--B2--M2--*

Git handles this situation just fine in practice using "recursive" merge strategy. I am not sure about Mercurial.

• In "Known Issues" there is warning that merge tracking migh not work with file renames, e.g. when one side renames file (and perhaps modifies it), and second side modifies file without renaming (under old name).

Both Git and Mercurial handle such case just fine in practice: Git using rename detection, Mercurial using rename tracking.

HTH

Without speaking about the usual advantages (offline commits, publication process, ...) here is a "merge" example I like:

The main scenario I keep seeing is a branch on which ... two unrelated tasks are actually developed
(it started from one feature, but it lead to the development of this other feature.
Or it started from a patch, but it lead to the development of another feature).

How to you merge only one of the two feature on the main branch?
Or How do you isolate the two features in their own branches?

You could try to generate some kind of patches, the problem with that is you are not sure anymore of the functional dependencies which could have existed between:

• the commits (or revision for SVN) used in your patches
• the other commits not part of the patch

Git (and Mercurial too I suppose) propose the rebase --onto option to rebase (reset the root of the branch) part of a branch:

From Jefromi's post

- x - x - x (v2) - x - x - x (v2.1)
           \
            x - x - x (v2-only) - x - x - x (wss)

you can untangle this situation where you have patches for the v2 as well as a new wss feature into:

- x - x - x (v2) - x - x - x (v2.1)
          |\
          |  x - x - x (v2-only)
           \
             x - x - x (wss)

, allowing you to:

• test each branch in isolation to check if everything compile/work as intended
• merge only what you want to main.

---

The other feature I like (which influence merges) is the ability to squash commits (in a branch not yet pushed to another repo) in order to present:

• a cleaner history
• commits which are more coherent (instead of commit1 for function1, commit2 for function2, commit3 again for function1...)

That ensure merges which are a lot easier, with less conflicts.

We recently migrated from SVN to GIT, and faced this same uncertainty. There was a lot of anecdotal evidence that GIT was better, but it was hard to come across any examples.

I can tell you though, that GIT is A LOT BETTER at merging than SVN. This is obviously anecdotal, but there is a table to follow.

Here are some of the things we found:

• SVN used to throw up a lot of tree conflicts in situations where it seemed like it shouldn't. We never got to the bottom of this but it doesn't happen in GIT.
• While better, GIT is significantly more complicated. Spend some time on training.
• We were used to Tortoise SVN, which we liked. Tortoise GIT is not as good and this may put you off. However I now use the GIT command line which I much prefer to Tortoise SVN or any of the GIT GUI's.

When we were evaluating GIT we ran the following tests. These show GIT as the winner when it comes to merging, but not by that much. In practice the difference is much larger, but I guess we haven't managed to replicate the situations that SVN handles badly.

Others have covered the more theoretical aspects of this. Maybe I can lend a more practical perspective.

I'm currently working for a company that uses SVN in a "feature branch" development model. That is:

• No work can be done on trunk
• Each developer can have create their own branches
• Branches should last for the duration of the task undertaken
• Each task should have it's own branch
• Merges back to trunk need to be authorized (normally via bugzilla)
• At times when high levels of control are needed, merges may be done by a gatekeeper

In general, it works. SVN can be used for a flow like this, but it's not perfect. There are some aspects of SVN which get in the way and shape human behaviour. That gives it some negative aspects.

• We've had quite a few problems with people branching from points lower than ^/trunk. This litters merge information records throughout the tree, and eventually breaks the merge tracking. False conflicts start appearing, and confusion reigns.
• Picking up changes from trunk into a branch is relatively straight forward. svn merge does what you want. Merging your changes back requires (we're told) --reintegrate on the merge command. I've never truly understood this switch, but it means that the branch can't be merged into trunk again. This means it's a dead branch and you have to create a new one to continue work. (See note)
• The whole business of doing operations on the server via URLs when creating and deleting branches really confuses and scares people. So they avoid it.
• Switching between branches is easy to get wrong, leaving part of a tree looking at branch A, whilst leaving another part looking at branch B. So people prefer to do all their work in one branch.

What tends to happen is that an engineer creates a branch on day 1. He starts his work and forgets about it. Some time later a boss comes along and asks if he can release his work to trunk. The engineer has been dreading this day because reintegrating means:

• Merging his long lived branch back into trunk and solving all conflicts, and releasing unrelated code that should have been in a separate branch, but wasn't.
• Deleting his branch
• Creating a new branch
• Switching his working copy to the new branch

...and because the engineer does this as little as they can, they can't remember the "magic incantation" to do each step. Wrong switches and URLs happen, and suddenly they're in a mess and they go get the "expert".

Eventually it all settles down, and people learn how to deal with the shortcomings, but each new starter goes through the same problems. The eventual reality (as opposed to what I set out at he start) is:

• No work is done on trunk
• Each developer has one major branch
• Branches last until work needs to be released
• Ticketed bug fixes tend to get their own branch
• Merges back to trunk are done when authorized

...but...

• Sometimes work makes it to trunk when it shouldn't because it's in the same branch as something else.
• People avoid all merging (even easy stuff), so people often work in their own little bubbles
• Big merges tend to occur, and cause a limited amount of chaos.

Thankfully the team is small enough to cope, but it wouldn't scale. Thing is, none of this is a problem with CVCS, but more that because merges aren't as important as in DVCS they're not as slick. That "merge friction" causes behaviour which means that a "Feature Branch" model starts to break down. Good merges need to be a feature of all VCS, not just DVCS.

---

According to this there's now a --record-only switch that could be used to solve the --reintegrate problem, and apparently v1.8 chooses when to do a reintegrate automatically, and it doesn't cause the branch to be dead afterwards

Prior to subversion 1.5 (if I'm not mistaken), subversion had a significant dissadvantage in that it would not remember merge history.

Let's look at the case outlined by VonC:

- x - x - x (v2) - x - x - x (v2.1)
          |\
          |  x - A - x (v2-only)
           \
             x - B - x (wss)

Notice revisions A and B. Say you merged changes from revision A on the "wss" branch to the "v2-only" branch at revision B (for whatever reason), but continued using both branches. If you tried to merge the two branches again using mercurial, it would only merge changes after revisions A and B. With subversion, you'd have to merge everything, as if you didn't do a merge before.

This is an example from my own experience, where merging from B to A took several hours due to the volume of code: that would have been a real pain to go through again, which would have been the case with subversion pre-1.5.

Another, probably more relevant difference in merge behaviour from Hginit: Subversion Re-education:

> Imagine that you and I are working on > some code, and we branch that code, > and we each go off into our separate > workspaces and make lots and lots of > changes to that code separately, so > they have diverged quite a bit. > > When we have to merge, Subversion > tries to look at both revisions—my > modified code, and your modified > code—and it tries to guess how to > smash them together in one big unholy > mess. It usually fails, producing > pages and pages of “merge conflicts” > that aren’t really conflicts, simply > places where Subversion failed to > figure out what we did. > > By contrast, while we were working > separately in Mercurial, Mercurial was > busy keeping a series of changesets. > And so, when we want to merge our code > together, Mercurial actually has a > whole lot more information: it knows > what each of us changed and can > reapply those changes, rather than > just looking at the final product and > trying to guess how to put it > together.

In short, Mercurial's way of analyzing differences is (was?) superior to subversion's.