git-subtrac: all your git submodules in one place
git-subtrac is a helper tool that makes it easier to keep track of your git submodule contents. It collects the entire contents of the entire history of all your submodules (recursively) into a separate git branch, which can be pushed, pulled, forked, and merged however you want.
git-subtrac is a git extension written in the Go language using the lovely go-git library. If you have Go 1.12 or higher and a project that already uses go modules (contains a go.mod file), you can install the tool like this:
go install github.com/apenwarr/git-subtrac
If you have an older version of go or don't care about go modules, you can install it like this instead:
go get -v github.com/apenwarr/git-subtrac
This will drop the compiled program into Go's $GOPATH/bin directory, often
$HOME/go/bin. As long as that's in your
$PATH, you will then be able to run commands like
git subtrac <whatever>.
To collect the submodule history for all your branches into new tracking branches, just do this (from inside the git repository you want to affect):
git subtrac update
If your repository is incomplete (ie. you have old submodule links that no longer exist, because the submodules have since been rebased or something), this might give an error about missing commits. You can work around it with the
git subtrac --auto-exclude update
git subtrac update will generate, for each branch, a new branch with the
.trac extension, referencing the complete history of all its submodule references. The easiest way to use this branch is as follows:
Push the tracking branch (eg.
master.trac) into the same upstream repository (github or whatever) as the branch (eg.
master) that it tracks.
.gitmodulesfile to change all the
url = <whatever>lines to just
url = .(a single dot). This tells git to always fetch the submodule contents from the same repo as your parent module.
End users downloading your project don't need to have git-subtrac installed, unless they make their own changes to submodules. All the
git submodule commands continue to work exactly the same way as before.
What problem are we solving?
git submodules have been very complicated and hard to use since the earliest days of git, and have not improved much. The complications stem from some major design issues:
The contents of submodules come from other git repositories, which can move around or be rebased/deleted/etc without warning, causing your own repository to no longer be usable. At the very least, using multiple upstream repositories makes it hard to fork a project and make wide-ranging changes: you have to fork every sub-repo that you change, then update .gitmodules links, tell everyone how to obtain your changes, etc. As a result, submodules end up forcing people to centralize on a single upstream repository.
Submodule links inside your git repository are links to commits, not trees. This is incredibly powerful ("this directory is provided by the contents of commit X of project P") because it acts like you copied the contents of one project into a subdir of another, but you have the complete history of the subproject still intact. Unfortunately, this power turns git's already-complicated history management (which few people understand) into something exponentially more complex. The subproject has (we hope) a history that always moves forward (commits always being added on top). And your main project has a history that moves forward. But your link to the subproject can move backward! You can make a new commit to your parent project that moves the subproject from commit X to an earlier commit X-2 instead of a later commit X+2. As a result, all of git's normal fork, merge, revert, stash, etc algorithms are useless for submodules. And sure enough, nobody has ever updated them to work well with submodules.
Submodules are optional. An early design goal was to make it so that some people accessing a project might not have access to all the commits in all the subprojects. To make that work, all of git was designed to not mess with submodule links when you do regular operations, and to not abort when submodule links go missing. So it's super easy to make a mistake like forgetting to push your changes to a submodule when you push a change to the parent project, thus making it so nobody else can check out your parent code anymore.
git-subtrac can't solve all the problems created by these complexities, but it tries to narrow the set of constraints to make some of them easier. In particular, we put your submodule contents in the same repository as your parent project, eliminating all the problems from #1. And we make it easy to be sure you've pushed all your submodules correctly (just regenerate and push the trac branch), which greatly reduces #3.
We'll have to look somewhere else for a solution to #2. My earlier tool, git-subtree takes a different approach, avoiding submodule links entirely and merging submodule content directly into the parent repo. This solves all three of the problems above, but it makes submitting your changes back upstream more complex, because you have to split them out again and regenerate the original submodule history. It also gets confused when you move the linked subtrees around in your repository, because they just look like files being moved around, not the history of those files. Still, you might like git-subtree too. There are some discussions around about the pros and cons of submodules vs subtrees, such as this pretty good one from Atlassian.
Design: How does it actually work?
git-subtrac borrows some tricks from my earlier git-subtree project, but uses them with a more submodule-centric design.
The main thing you need to know is that, unlike every other kind of git object, a "submodule" reference from a git tree does not cause the referenced object (a commit) to be included in git object packs, or pushed along with your branch. In other words, when you add a file (blob) or subdirectory (tree) to your project, and then commit it to a branch, and then push that branch, the commit is never sent without also including a copy of the trees and blobs it references. But if your tree links to a commit - which is all a submodule is, a tree linking to a commit instead of to another tree or blob - then git push does not package up the subcommit or anything underneath it. It assumes you will push that commit yourself. And that's where all the problems come from.
The second thing you need to know is that if a commit (say, X) is referenced as the parent of another commit (say, Y), then when you push commit Y somewhere, it will always make sure X is there too (either by checking it already exists, or packaging it up along with Y). This is true recursively, so if you push the head of your commit history somewhere, then all the commits it was based on - and all the trees and blobs referred to by those commits - are pushed along with it. This is normal, of course; you'd be pretty surprised if that didn't work.
Also relevant is that a commit can have more than one parent commit. When you make a merge, your "merge commit" has at least two parents. It's not too commonly used, but a single commit can have as many parents as you want. You can merge a bunch of branches together in one shot.
So here's what we do: git-subtrac looks through all the trees of all the commits in your main project, and finds all the submodule links (ie. links to commit objects). Then it creates a new, parallel history, where for each commit, the "parent commits" of that commit include not just the "real" parent(s), but also the commits referred to inside the trees of that commit. In other words, if you have a commit Y that is based on X, and Y's filesystem contains submodule links to A, B, and C, then we produce a new commit Y+, which has parents X+, A, B, and C. X+ is generated the same way, by appending the submodule links referred to by X.
This way, when you push Y+ to a git server somewhere, it will include A, B, and C... and all their trees, files, and parents, exactly like you might have expected in the first place.
Features of the subtrac history branch
There are a few subtleties here that make the results extra nice. First of all, the way we generate the synthetic commits is completely stable: anyone seeing a commit Y (and a copy of its entire history, including all its submodule links and their trees, blobs, and histories) can reproduce exactly the same Y+, including the modification dates, parents, and even the commit hashes.
But it goes even further. Someone who doesn't have commit Y, but does have commit X, can generate exactly the same X+ as the person who generates Y+ from Y. (Remember that Y+ depends on X+.) If you have a subset of the history of Y, you can generate a perfect subset of the history of Y+.
This is really important: it means that two different people, given the same input, can always regenerate the same git-subtrac output. As long as you have all the objects, there's no danger in deleting and regenerating the subtrac branch from scratch.
Secondly, git-subtrac trims redundancy out of the
.trac branch as far as possible. It only generates a new Y+ for commit Y (based on X) if the set of submodule links is different from X. If the submodules haven't changed, then commit Y+ is X+ (not a new commit). This minimizes the changes to your
.trac branch across new commits, rebases, etc.
As a result, we get the following features for free:
If two people differently extend commit X, for example by producing commit Y and commit Z on top of X, then their subtrac trees Y+ and Z+ will both have X+ as a parent. When you then merge Y and Z together (into, say, commit YZ), the resulting YZ+ commit will also look like a git merge of Y+ and Z+. That merge could be produced by git just as easily as by git-subtrac. The main outcome is that you should always be able to push a newly generated
.tracbranch to upstream without using a "force push", because it should always look like a "fast forward" of the previous branch, even though it's actually been regenerated from scratch every time.
Imagine that commit X of our project depends on commit N of a submodule. For some reason commit N doesn't work out very well, and in commit Y we rewind the submodule to commit M (one commit before N) while we wait for the submodule people to fix it. In our fork of the submodule, we then realize we want to apply a cherry pick of some other patch Q, so the 'master' branch of our forked submodule looks like M+Q. If we push that new master branch to our real submodule repo, it would create a surprising problem: our subproject's commit Y works fine, because it links to M+Q. But if someone then wants to rewind to an earlier version of our superproject (eg. using
git bisect) and try commit X again, it won't work! The submodule repo no longer contains N at all, because we rewound the submodule's master branch.
In this case, git-subtrac helps a lot. Since X+ includes N, and Y+ includes M+Q, both sets of submodule links are included in the
.tracbranch. We end up successfully tracking "the history of history" of the submodules.
Similarly, a common use of submodules is to maintain patch queues for sending upstream. That is, we pull in some version of a submodule project (a library), and start working on improving it, testing our patches with our superproject (an app) until we're sure they really work. When we're feeling confident, we want to rebase the subproject to the latest version and apply our patches on top before sending them to the subproject's maintainer. In a traditional submodule setup, we'd have to be very careful not to lose the old history when we rebase the master branch of the submodule. With git-subtrac, this all works exactly like you want it to: both the old and the new patch histories are available when you need them.
Forking a superproject (eg. on github) now carries all its submodule history along with it, making pull requests very easy. (You can submit a pull request for the .trac branch if you want to share your submodule patches that way, I guess.)
Please try it out! git-subtrac is fairly new, so I apologize for any bugs you might run into.
You can email me at email@example.com. I don't have as much free time as I used to, but I'll try to respond. If you fix a bug, please send pull requests on github.