On Git History – or, “The Case for Merge Commits”

(I keep having to argue the case for merge commits in git, so this is a brief summary of my position I can point people to.)

tl;dr: NeverThink hard before you remove information! AlwaysYou should usually merge branches with git merge --no-ff to get a merge commit, and keep the “Conflicts” lines added by git. Only fast forward when pulling in remote changes to a branch with no local commits.

I don’t think this is very controversial anymore, but I’ll start by presenting the case for merge commits anyway.

A new feature is born

So, you want to implement a new feature and you branch off master at the point b, write the code, and end up with commits x, y and z (possibly after first having cleaned up your feature branch). Meanwhile, a new commit, c, has been pushed to the master branch:

[master]  a-b-c
            \
[feature]    x-y-z

You now have two main options: You can rebase your feature branch on top of the current master:

[master]  a-b-c-x-y-z

or merge your feature branch into master with a merge commit m:

[master]  a-b-c----m
            \     /
[feature]    x-y-z

The first option – to rebase and get a linear history – is for some reason very appealing to some people that consider this to be a “cleaner” history, but in reality you’re just losing history. I would even consider this to be a slight lie, since the environment where you tested commits x, y and z is not the same environment where the commits ended up, with no traces of a merge having taken place.

Now, ideally, real tests should uncover bugs anyway, but you probably did informal testing all the time when writing the code (“Does this work? Ah, of course not, that argument needs to be in milliseconds!”), and if anything changed on the master branch in the meantime, assumptions you made while developing may no longer hold. In those cases, when digging through the git history, seeing a merge commit is very valuable information and will probably lead to closer scrutiny if it seems at all relevant to the issue under investigation.

When git creates a merge commit it will also by default append a list of files that had conflicts to the commit message:1

Conflicts:
    src/foo-service.c
    src/bar-client.c

Do not remove these lines! This is also a piece of useful information for our when poring over the history to find out wtf is wrong. Even if your merge conflict was trivial, there is always a non-zero chance of introducing a bug when resolving a conflict, and seeing those lines in the merge commit message could be valuable information. They are basically a hint saying “Still confused? Maybe you should be extra careful when reviewing the changes in these files”.

An alternate history

Okay, fine, you say; merge commits make sense when you actually merge two or more commits, but what if the history was completely linear? Without commit c our history would be:

[master]  a-b
            \
[feature]    x-y-z

If you now merge your feature branch into master, git will by default do a fast-forward, simply changing master to point to z:

$ git merge feature
Updating 4fe28cd..68a8d69
Fast-forward

We then still have the same git history as above with no new commits:

[master,feature]  a-b-x-y-z

In this case it’s not as easy as above to argue for introducing a merge commit (as mentioned, git’s default behaviour is even to just fast-forward), but except for cases where the feature branch ends up just having one commit, you should still create a merge commit, by using git merge --no-ff feature, giving us this history:

[master]  a-b------m
            \     /
[feature]    x-y-z

This introduction of a merge commit allows you to write a summary of the changes in the branch you’re merging, and allows people reading the history in the future to choose to view the merge as just one commit, or – if they choose to – to dive into the commits that compromise the feature that was merged. Again, this is an example of having more information available.2 Which is almost always a good thing.

Howto

Ok, so how do you get git to behave like this? Well, there is a config option merge.ff3 that allows you to specify that merges should never be a fast-forward, but this is probably not what you want since this means that every time you pull new commits from a remote, you will introduce a merge commit.4 Instead, you should stop and think for a second when merging a branch. If you want to introduce a merge commit (you probably do), you should pass the --no-ff flag to merge.

Postscript

In reality, a long lived feature branch will often look more like the following, having merged in the master branch at various points to keep up with changes:

[master]  a-b-c-d-e-f-g-h----m
            \     \     \   /
[feature]    u-v-w-m-x-y-m-z

  1. Actually, since git 2.3.0, these lines are commented out by default, and you have to explicitly un-comment them to keep them.

  2. As a bonus, it also makes it easy to revert the introduction of a feature without the need to carefully study the history around the point where it was introduced.

  3. $ git config --global merge.ff false

  4. Remember that pull = fetch + merge

Comments