Daniel R. MCloy
Git submodules

11 Jun 2015

Recently I set up the PHOIBLE development repos to be a bit more flexible and extensible for future needs. This required setting up submodules in git. Here are some notes on how that works:

  1. Both repos must already exist. If either the parent or the child module doesn’t exist yet, create it first. Here the two repos are called “dev” and “data”, and we’ll be nesting “data” inside of “dev”.

  2. Create the submodule. In the local clone of “dev”: git submodule add git@github.com:phoible/data.git. This creates the submodule in detached head state, meaning it is not actually “on” a branch, and changes made to files in dev/data won’t be tracked.

    • Note: if we had wanted to control the specific path of the submodule, we could add it to the end of the command, like: git submodule add git@github.com:phoible/data.git nest/in/folder/called/datums.
  3. Change some settings. In the parent repo:

    • git config status.submodulesummary 1 makes it so that git status will include changes to submodules.
    • git config diff.submodule log makes it so that git diff will show a list of submodule commit messages instead of just the previous and current commit hashes.

    Either one of these can take the --global flag.

  4. Get the submodule onto a branch. This is as simple as cd data; git checkout master. Now the submodule is set up to track local changes, not just sit there as a snapshot of the upstream version of data.

  5. Decide what to do when local submodule changes conflict with remote changes. Normally, submodules are the kind of thing that you might change while working on your project, but other people might change upstream too. In setups like that, you can run cd data; git fetch; git pull origin/master to pull in upstream changes to the data submodule. Alternatively, you can run git submodule update --remote from the parent repo to pull in upstream changes (defaults to pulling master, but you can change that in the .gitmodules config file).

    If you’ve made local changes to data, you need to tell git what to do when you try to pull in upstream changes. Do this by adding a flag to the git submodule update --remote command: either --merge (to merge remote changes on top of your local ones) or --rebase (to rewind your changes, apply remote changes, then replay your changes afterward). If you have local changes and don’t specify either --merge or --rebase, your changes will live on in whatever branch of the submodule you made them in, but the submodule will revert to detached head state before applying the remote changes.

    In the case of PHOIBLE, it is expected that all changes to data will occur via dev, so pulling in upstream changes to data should never need to happen. But it would if data became a submodule for other repos as well.

  6. Push. A normal git push command in the parent repo will not push submodule changes. A safer version is git push --recurse-submodules=check which prevents pushing the parent repo if the submodule(s) are not pushed first. Another option is git push --recurse-submodules=on-demand which will try to push the submodules automatically (if necessary) before pushing the parent repo.

Here’s a worked example. First, make some changes to a script in the parent module (not shown), and re-run the script (whose effect is to re-generate files files within the submodule). What does git status show us?

On branch master
Changes not staged for commit:
  modified:   aggregate.R
  modified:   data (modified content)

The important thing to remember is that the changes in the submodule need to be committed twice: once within the submodule, and once within the parent. One way to think about this is that the parent repo doesn’t track changes to files in the submodule, but instead tracks commits in the submodule tree. The words modified content in the status message above mean that things have changed in the submodule, but they aren’t committed to the submodule tree and therefore cannot be committed in the parent tree yet either.

For these reasons, I find it easier to commit the changes in the submodule repo first, so that I can use the -a flag when committing the changes in the parent repo and it will pick up the submodule commits too:

cd data
git commit -am 'regenerated the data files after bugfix'
cd ..

At this point, git status would say something a little different:

Changes not staged for commit:
  modified:   aggregate.R
  modified:   data (new commits)

Submodules changed but not updated:
* data bd58f1d...0c9c816 (1):
  > regenerated the data files after bugfix

The new commits means the submodule changes are now available to be committed within the parent repo:

git commit -am 'bugfix and associated regenerated data'

Now what does git status tell us?

On branch master
Your branch is ahead of 'origin/master' by 1 commit.

That 1 commit is the parent repo commit, which includes changes to files within the parent repo and changes (i.e., commits) within the submodule. Now we can push, using a special submodule flag: git push origin master --recurse-submodules=on-demand

Pushing submodule 'data'
To git@github.com:phoible/data.git
bd58f1d..0c9c816  master -> master

To git@github.com:phoible/dev.git
92b3cd5..a11f3f0  master -> master

Note that git pushes to the submodule repo first (phoible/data) before pushing the the parent repo (phoible/dev). This would also have worked:

cd data
git push origin master
cd ..
git push origin master --recurse-submodules=check

The last line simply checks to see if we had already pushed the submodule changes, and fails to push if we have not yet done so. This is safer than a plain git push when dealing with submodules.