On Git's Shortcomings
09 Jul 2013
Git receives a lot of positive press. There are countless websites, articles, and blog posts dedicated to the adulation of Git. It’s plenty easy to find a list of reasons to use Git. It’s much harder to find a list of substantial reasons not to use Git.
That’s not surprising. There’s an obvious selection bias at play. Those who spend more time with Git understand it better and are more likely to extol its virtues. Conversely, those who are turned off by Git early are unlikely to make well informed arguments against it.
Unable to find a complete list of Git’s weaknesses, I have attempted to compile a list myself. Don’t get me wrong, I’m a proponent of Git (this site is on GitHub after all), but we must be truthful about Git’s limitations. This list (with a few exceptions) is about Git alone. I will avoid making comparisons to other version control systems unless relevant.
Ease of Use
In brief, Git has a complex information model and it doesn’t really abstract
that from the user. Git’s model is comprised of directed acyclic graphs,
commits, trees, blobs, branches, tags, and remotes. Git has a staging area, a
stash, and a reflog. All in all, Git has 145 commands, but there’s no
undo. Sure, you don’t need to know all of them, but you do need to know at
least 13 to be minimally productive (
Even specifying Git revisions needs its own manpage.
Git is complicated. I welcome the complexity because it brings powerful features along with it. However, if I wasn’t a professional software developer, I might look elsewhere.
Git doesn’t concern itself with access control. It defers that job to the file system or ssh. Because of this, the ability to restrict access to Git repositories is severely limited.
Read Access - You cannot restrict read access to specific files, directories, or branches within a single repository with Git. You can either clone the entire repository or none of it. You need to arrange things such that setting read permissions per-repository is sufficient.
Write Access - Options for restricting write access aren’t quite as limited. While Git doesn’t support write access control out of the box, it does provide hooks that can reject pushes. So, third party tools can add write access control to Git. gitolite, for example, can restrict write access to files, directories, or branches. It can also restrict force pushes.
Unfortunately, this option ties you to a particular third party tool. Want to migrate your gitolite access control to GitHub? Too bad.
Some version control tools have a way of completely removing files from the repository, sometimes called obliterate. Obliterate is more than just delete. It must purge the data from the repository and its history completely, as if it was never there in the first place. There are two reasons you might want to do this:
Confidential Information - If confidential information is ever committed to a repository, deleting it isn’t enough. It will remain recoverable to anyone with read access to the repository; that’s the point of version control. Depending upon the level of confidentiality and the level of exposure of the repository, obliterate may be necessary.
Large (Measured in Bytes) Mistakes - If a file is accidentally committed to a Git repository and then removed by a later commit, a snapshots of that file will forever live in the repository. If that file was both added by mistake and very large, this could be a problem. Obliterate could remove it for good.
I think Git excels in preventing these problems by having multiple steps to
catch and correct a mistake before it ends up in the central repository. You
push (compared to just
commit with SVN) before
your mistake ends up public.
On the other hand, once it’s public, it’s public for good. Git makes cryptographic guarantees that ensure that if someone tries to rewrite history to obliterate a file, every clone of that repository will notice at the next fetch (perhaps to the ire of your fellow developers).
This isn’t to suggest that obliterate is trivial in centralized version control tools. You still have to worry about all the working copies in addition to the central repository. However, with Git, the problem is more complicated thanks to locally attached history in every clone.
Git does not support locking files. How could it? To what central authority
does a distributed version control system go to to obtain a lock?
would be an oxymoron.
That said, sometimes locks are needed. Consider binary files that can’t be merged (video, images, CAD drawings, video game assets, etc.). Fundamentally, each developer needs to take turns working on such files. Locking a file can be the right solution. For some, locking is a critical feature.
Git doesn’t support locking, but veracity, another distributed version control tool, does. If you want to stick with Git, gitolite learned how to lock files in 2012. So, this isn’t an intractable problem, but it’s worth watching out for. At the very least, you’ll need to configure an additional tool to support locking.
Very Large Repositories
Git stores snapshots of the entire history of a repository locally when you clone. Disk space is cheap and still getting cheaper. More likely, the bottleneck is network speed. If you need to clone a large repository over the network, it’s going to take a while. Megabytes, no problem. Gigabytes, manageable. Terabytes, don’t bother (in 2013). This is made more troublesome by the lack of partial clones.
Partial Clone by Path - Partial clone by path is not possible in Git.
git-archive(1) lets you download part of a repository by path as a tar or zip. But that’s all you get, some files, not a working repository.
Git learned how to do sparse checkout in version 1.7. This could definitely help in some repositories, but it only impacts
clone. So, it will save some disk space, but no bandwidth. The documentation for sparse checkout is burried in git-read-tree(1).
Partial Clone by Branch -
git clone --single-branchmakes it possible to clone only a selected branch. However, this will not save much bandwidth unless the branches in the remote repository are very far diverged. This can be useful in situations like GitHub Pages where you have an orphaned branch (
gh-pages) with different content than the rest of your branches.
Shallow Clone - Git allows shallow clones, which only clone recent history, with
git clone --depth <depth>. However, they come with significant limitations. From git-clone(1):
A shallow repository has a number of limitations (you cannot clone or fetch from it, nor push from nor into it), but is adequate if you are only interested in the recent history of a large project with a long history, and would want to send in fixes as patches.
This can drastically reduce the time to clone a repository with long history, particularly for content that doesn’t delta-compress well.
That said, Git’s delta-compression can help quite a bit for most projects and
the results may surprise you. After a recent SVN to Git migration, one of our
repositories went from a 5 minute 8 second
svn checkout to a 4 minute 17
git clone. I did not expect that.
Aside from file size concerns, Git commands can start to perform poorly in very large repositories (many commits, many files). In short, Git becomes unusable around 10^6 of each. That’s about 2 orders of magnitude larger than the Linux Kernel which the Git community tends to think of as a large repository. The general advice here is to split up your repository before things get that bad, but that advice can seem awfully hollow to an organization that operates successfully with such a super-repository.
If you’re concerned about performance, some of these client side or server side benchmarks might interest you. If you’re considering migrating a repository larger than the Linux kernel to Git, you should probably run your own benchmarks.
Large Number of Contributors to One Branch
This point takes a little explaining since it is counter-intuitive. Git works great for large projects with many contributors. Thanks to its distributed nature, you don’t have to give everyone write access to a central repository. Because of this, Git is fantastic for open source projects. However, if you want to give many people write access to a single branch on a single repository, you might run into trouble.
Consider what happens when two people (Alice and Bob) try to push to the same branch on the same remote at the same time. Let’s say Alice gets in first and successfully updates the branch. Bob’s push gets rejected because the remote branch is no longer reachable from the snapshot he’s trying to push. This ensures that Alice’s commits don’t get abandoned when Bob pushes. This is a good thing, but it can become a bottleneck.
Bob now has to fetch, redo his merge, and then try the push again. This process will take a minute or two. It will take longer in larger repositories and longer still if you sign commits. This limits the rate of pushes possible to a single branch. A branch receiving a dozen or more pushes per hour peak is high traffic, but not unfathomable. 100 pushes per hour is infeasible.
As a counter point, Git makes it painless to branch and merge, so it’s probably worth asking why you need to push that often to one branch.
Finally, here are a couple of minor points that deserve mention.
Empty Directories - Git doesn’t support empty directories. The standard work-around is to add an empty
.gitignoreto the directory and commit that instead.
Revision Numbers - Git doesn’t have revision numbers, it has 40 character hashes. Because of this, a lexicographic sort of revisions no longer makes sense. If you are used to appending a revision number to a build artifact and having a meaningful sort, you may need to rethink your strategy.
SVN Externals Equivalent - If I had written this post much sooner, this would have been an issue. Submodules used to be unable to track branches; they had to integrate with a specific commit. This left them less flexible than SVN externals. Git 1.8.2 recently added this feature. Problem solved.
These shortcomings are a non-issue for many developers. And most of these weaknesses can be avoided without too much trouble. Still, if you are looking to use Git on a new project or migrate an existing project to Git, consider these limitations. Decide for yourself if any of them will be problematic for your particular use case.