Really, a concrete example that merging in Git is easier than SVN?

From a practical perspective, merging has traditionally been "hard" because of what I call the "big bang merge" problem. Suppose a developer has been working away on some code for a while and hasn't committed their work yet (maybe the developer is accustomed to working in Subversion against trunk, and doesn't commit unfinished code). When the developer finally commits, there is going to be a lot of changes all rolled up into one commit. For the other developers who want to merge their work with this "big bang" commit, the VCS tool isn't going to have enough information about how the first developer got to the point they committed, so you just get "here's a giant conflict in this whole function, go fix it".

On the other hand, the usual style of working with Git and other DVCS that have cheap local branches, is to commit regularly. Once you've done one bit of work that pretty much makes sense, you commit it. It doesn't have to be perfect but it should be a coherent unit of work. When you come back to merge, you still have that history of smaller commits that shows how you got from the original state to the current state. When the DVCS goes to merge this with the work of others, it has a lot more information about what changes were made when, and you end up getting smaller and fewer conflicts.

The point is that you can still make merging a hard problem with Git by making a single big bang commit only after you've finished something. Git encourages you to make smaller commits (by making them as painless as possible), which makes future merging easier.


I can only tell you of a small experiment were Git was NOT better than Subversion (same problems).

I was wondering about this case: You start with two branches "mytest1" and "mytest2" both based on the same commit. You have got a C file which contains a function blub(). In branch mytest1 you move "blub()" to a different position in the file and commit. In branch mytest2 you modify blub() and commit. On branch mytest2 you try to use "git merge mytest1".

Seems to give a merge conflict. I hoped that Git would recognize that "blub()" was moved in mytest1 and then be able to auto-merge the modification in mytest2 with the move in mytest1. But at least when I tried this did not work automatically...

So while I fully understand that Git is much better at tracking what has been merged and what has not been merged yet, I also wonder if there is "pure" merge case in which Git is better than SVN...

Now because this question has been bugging me for a long time I was really trying to create a concrete example where Git is better, whereas merging in SVN fails.

I found one here https://stackoverflow.com/a/2486662/1917520, but this includes a rename and the question here was for a case without a rename.

So here is an SVN example which basically tries this:

bob        +-----r3----r5---r6---+
          /                /      \
anna     /  +-r2----r4----+--+     \
        /  /                  \     \
trunk  r1-+-------------------r7-- Conflict

The idea here is:

  • Anna and Bob are both developers with their own branches (created in r2,r3).
  • Anna does some modifications (r4),
  • Bob does some modifications (r5).
  • Bob merges the modifications from Anna into his branch; this gives conflicts, which Bob fixes and then commits (r6).
  • Annas modifications are merged back into the trunk (r7).
  • Bob tries to merge his modification back into the trunk and this again gives a conflict.

Here is a Bash script, which produces this conflict (using SVN 1.6.17 and also SVN 1.7.9):

#!/bin/bash
cd /tmp
rm -rf rep2 wk2
svnadmin create rep2
svn co file:///tmp/rep2 wk2
cd wk2
mkdir trunk
mkdir branches
echo -e "A\nA\nB\nB" > trunk/f.txt
svn add trunk branches
svn commit -m "Initial file"
svn copy ^/trunk ^/branches/anna -m "Created branch anna"
svn copy ^/trunk ^/branches/bob  -m "Created branch bob"
svn up 
echo -e "A\nMA\nA\nB\nB" > branches/anna/f.txt
svn commit -m "anna added text"
echo -e "A\nMB\nA\nB\nMB\nB" > branches/bob/f.txt
svn commit -m "bob added text"
svn up
svn merge --accept postpone ^/branches/anna branches/bob
echo -e "A\nMAB\nA\nB\nMB\nB" > branches/bob/f.txt
svn resolved branches/bob/f.txt
svn commit -m "anna merged into bob with conflict"
svn up
svn merge --reintegrate ^/branches/anna trunk
svn commit -m "anna reintegrated into trunk"
svn up
svn merge --reintegrate --dry-run ^/branches/bob trunk

The last "--dry-run" tells you, that there will be a conflict. If you instead first try to merge Anna's reintegration into Bob's branch then you also get a conflict; so if you replace the last svn merge with

svn merge ^/trunk branches/bob

this also shows a conflicts.

Here is the same with Git 1.7.9.5:

#!/bin/bash
cd /tmp
rm -rf rep2
mkdir rep2
cd rep2
git init .
echo -e "A\nA\nB\nB" > f.txt
git add f.txt
git commit -m "Initial file"
git branch anna
git branch bob
git checkout anna
echo -e "A\nMA\nA\nB\nB" > f.txt
git commit -a -m "anna added text"
git checkout bob
echo -e "A\nMB\nA\nB\nMB\nB" > f.txt
git commit -a -m "bob added text"
git merge anna
echo -e "A\nMAB\nA\nB\nMB\nB" > f.txt
git commit -a -m "anna merged into bob with conflict"
git checkout master
git merge anna
git merge bob

The contents of f.txt change like this.

Initial version

A
A
B
B

Anna's modifications

A
MA
A
B
B

Bob's modifications

A
MB
A
B
MB
B

After Anna's branch is merged into Bob's branch

A
MAB
A
B
MB
B

As so many people already pointed out: The problem is, that subversion cannot remember that Bob already resolved a conflict. So when you try to now merge Bob's branch into the trunk, then you have to re-resolve the conflict.

Git works completely different. Here some graphical representation what git is doing

bob         +--s1----s3------s4---+
           /                /      \
anna      /  +-s1----s2----+--+     \
         /  /                  \     \
master  s1-+-------------------s2----s4

s1/s2/s3/s4 are the snapshots of the working directory git takes.

Notes:

  • When anna and bob create their development branches, this will NOT create any commits under git. git will just remember that both branches initially refer to the same commit object as the master branch. (This commit in turn will refer to the s1 snapshot).
  • When anna implements her modification, this will create a new snapshot "s2" + a commit object. A commit object includes:
    • A reference to the snapshot (s2 here)
    • A commit message
    • Information about ancestors (other commit objects)
  • When bob implements his modification, this will create another snapshot s3 + a commit object
  • When bob merges annas modifications into his development branch this will create yet another snapshot s4 (containing a merge of his changes and anna's changes) + yet another commit object
  • When anna merges her changes back into the master branch, this will be a "fast-forward" merge in the shown example, because the master has not changed in the meantime. What "fast-forward" here means is, that the master will simply point to the s2 snapshot from anna without merging anything. With such a "fast-forward" there will not even be another commit object. The "master" branch will just directly now refer to the last commit from the "anna" branch
  • When bob now merges his changes into the trunk, the following will happen:
    • git will find out that the commit from anna which created the s2 snapshot is a (direct) ancestor for bobs commit, which created the s4 snapshot.
    • because of this git will again "fast-forward" the master branch to the last commit of the "bob" branch.
    • again this will not even create a new commit object. The "master" branch will simply be pointed to the last commit of the "bob" branch.

Here is the output of "git ref-log" which shows all of this:

88807ab HEAD@{0}: merge bob: Fast-forward
346ce9f HEAD@{1}: merge anna: Fast-forward
15e91e2 HEAD@{2}: checkout: moving from bob to master
88807ab HEAD@{3}: commit (merge): anna merged into bob with conflict
83db5d7 HEAD@{4}: commit: bob added text
15e91e2 HEAD@{5}: checkout: moving from anna to bob
346ce9f HEAD@{6}: commit: anna added text
15e91e2 HEAD@{7}: checkout: moving from master to anna
15e91e2 HEAD@{8}: commit (initial): Initial file

As you can see from this:

  • when we go to anna's development branch (HEAD@{7}) we do not change to a different commit, we keep the commit; git just remembers that we are now on a different branch
  • At HEAD@{5} we move to bob's initial branch; this will move the working copy to the same state as the master branch, because bob has not changed anything yet
  • At HEAD@{2} we move back to the master branch, so to the same commit object everything started from.
  • Head@{1},HEAD@{0} show the "fast-forward" merges, which do not create new commit objects.

With "git cat-file HEAD@{8} -p" you can inspect the complete details of the initial commit object. For the example above, I got:

tree b634f7c9c819bb524524bcada067a22d1c33737f
author Ingo <***> 1475066831 +0200
committer Ingo <***> 1475066831 +0200

Initial file

The "tree" line identifies the snapshot s1 (==b634f7c9c819bb524524bcada067a22d1c33737f) to which this commit refers.

If I do "git cat-file HEAD@{3} -p" I get:

tree f8e16dfd2deb7b99e6c8c12d9fe39eda5fe677a3
parent 83db5d741678908d76dabb5fbb0100fb81484302
parent 346ce9fe2b613c8a41c47117b6f4e5a791555710
author Ingo <***> 1475066831 +0200
committer Ingo <***> 1475066831 +0200

anna merged into bob with conflict

This above shows the commit object, bob created when merging anna's development branch. Again the "tree" line refers to the created snapshot (s3 here). Additionally note the "parent" lines. The second one which starts with "parent 346ce9f" later tells git, when you try to merge back bob's development branch into the master branch, that this last commit of bob has anna's last commit as an ancestor. This is why git knows that the merge of bob's development branch into the master branch is a "fast-forward".


I don't have concrete examples, but any kind of repeated merge is difficult, in particular so called criss-cross merge.

   a
  / \
 b1  c1
 |\ /|
 | X |
 |/ \|
 b2  c2

merging b2 and c2


The wiki page on Subversion Wiki describing differences between mergeinfo based assymetric Subversion merge (with 'sync' and 'reintegrate' directions) and merge tracking based symmetric merge in DVCS has a section "Symmetric Merge with Criss-Cross Merge"


The most concrete example I can think of is the simplest merge that does not result in merge conflicts. However (TL;DR) with that example Git is still inherently a simpler procedure than with Subversion. Lets review why:

Subversion

Consider the following scenario in subversion; the trunk and the feature branch:

   1  2  3
…--o--o--o              trunk
          \4  5
           o--o         branches/feature_1

To merge you can use the following command in subversion:

# thank goodness for the addition of the --reintegrate flag in SVN 1.5, eh?
svn merge --reintegrate central/repo/path/to/branches/feature_1

# build, test, and then... commit the merge
svn commit -m "Merged feature_1 into trunk!"

In subversion merging the changes requires another commit. This is to publish the changes that the merge did with applying the changes on the feature branch virtual directory back into the trunk. That way everyone working with the trunk can now use it and the revision graph looks like sort of like this:

   1  2  3      6
…--o--o--o------o       /trunk
          \4  5/
           o--o         /branches/feature_1

Lets see how this is done in git.

Git

In Git this merge commit is really not necessary as branches are glorified bookmarks on the revision graph. So with the same kind of revision graph structure it sort of looks like this:

         v-- master, HEAD

   1  2  3
…--o--o--o
          \4  5
           o--o

              ^-- feature_branch

With the head currently on the master branch we can perform a simple merge with the feature branch:

# Attempt a merge
git merge feature_branch

# build, test, and then... I am done with the merge

... and it will fast-forward the branch over to the commit where feature branch is pointing at. This is made possible because Git knows that the goal of the merge is a direct descendant and the current branch only needs to take in all the changes that happened. The revision graph will end up looking like this:

   1  2  3  4  5
…--o--o--o--o--o
               ^-- feature_branch, master, HEAD

The changes does not need a new commit as all git has done is to move the branch references further up to the front. All that is left is to publish this to the public repository if you have any:

# build, test, and then... just publish it
git push

Conclusion

Given this simple scenario you can assert two things in the difference between Subversion and Git:

  • SVN requires at least a couple of commands to finalize the merge and forces you to publish your merge as a commit.
  • Git only requires one command and does not force you to publish your merge.

Given this to be the most simplest merge scenario it is difficult to argue that subversion is easier than git.

In more difficult merge scenarios, git also provides you the ability to rebase the branch that produces a simpler revision graph at the cost of history rewriting. Once you get the hang of it though, and avoid publishing history rewrites of things already published, it isn't really that bad of a thing to do.

Interactive rebases is outside the scope of the question but to be honest; it enables you the ability to rearrange, squish and remove commits. I wouldn't willingly want to switch back to Subversion as history rewriting is not possible by design.