Jujutsu (jj), a Git compatible VCS
tonyfinn.com
tonyfinn.com
My ideal git workflow is pretty simple:
1. `git switch -C new-branch` 2. Make changes, time passes 3. `git add .` 4. `git commit -m "Description of changes made"` 5. `git fetch && git rebase origin/main` 6. `git push` 7. Make pr to main
If main is out of date I rebase again to update.
This to me feels pretty light weight and is not a hinderence in any way on my day to day tasks.
What exactly makes `jj` better? Is there a specific usecase i'm not understanding?
1. If it works for you, no need to stop
I remember plenty of people who used SVN saying why did they need to use git. In fact, I remember companies complaining about distributed version control generally, saying they preferred centralized version control.
If it works for you, you can keep using it. No one will stop you, but you may find that just as SVN users have largely shrunk down, maybe the same will happen with git.
2. jj today
I use jj today and the biggest advantages for me are not needing to worry as much about the order of operations.
I write some code and then I come across a bug. I can pull out the bug and fix it while still working on my existing code. I don't have to go back in time to fix it.
I also appreciate not needing to stop just because of a conflict. This is something I miss from SVN, which was similar-ish in that conflicts didn't need to be addressed immediately, only jj's implementation is better.
Generally I find the jj workflow to be something that takes a few days of adjustment, then it just makes more sense.
3. The way things will be
jj is still relatively early. The commands to interact with git aren't all very intuitive or easy to use. I had to make some aliases to make this easier. More importantly, there's no jj backend yet. Once that happens and it's adopted by forgjo or another forge (Gitlab, Github), I think we'll see a big shift towards it.
I think git is going to be around a long time, but the day to day use of jj is so much nicer than git that I hope it will gain fast adoption.
I remember everyone saying “SVN is a piece of shit, I hate it, why can’t I easily create branches.” Also mixed revisions which was the worst part of it.
The jj people should really write a guide to show how it’s better than git because, so far, I’m still very confused and we’ve had that conversation on HN at least 10 times without anyone giving concrete examples.
Take the index. I love(d) git’s index. It is a powerful way to build up focused commits. Jj does away with the index, yet gives you equivalent power. How? Because instead of a separate index concept, jj has a special name for a commit representing the index: @. Files are (by default) automatically tracked and included in @. Just like how you’d git add -p, you create a new commit for your work, then create a commit on top of it. As you work, @ gets built up with your code, and when you’re ready, jj squash (with various flags or paths to control exactly what gets added) to move them into your parent, “real” work commit.
So if it’s the same, why is it better? Because we’ve reduced the number of states: instead of dirty, staged, and committed, you just have committed. And with that, commands are simpler: no distinction between —hard or —soft for something like git reset, everything just operates on committed state. This means you can bring the full set of tools to manipulate commits to bear on your index!
The same goes with stashing: because heads are allowed to be detached, to use the git terminology, we don’t need a special stash concept: create a new commit that’s forward of where you’re working to “stash” your changes in there until you need them.
These are two simple examples, but it’s really about how everything fits together into an extremely powerful and cohesive tool.
I’ll give you some concrete advantages:
1. Every time you run a `jj` command a snapshot of your local changes is made. I don’t know about you but often while I’m working on something experimental I have an idea for a slightly different approach, try it out, it doesn’t pan out, and I’m like fuck I wish I’d saved where I was an hour ago. git won’t do that for you but jj will.
2. The rebase workflow is dramatically better. Rebasing multiple commits with conflicts is excruciating. You have to fix it right here and right now but you’re in a weird repo state where you can’t easily to another commit and tweak it or look at the state of things at different times. It’s a pain in the ass and easy to get wrong. With jj there are no weird intermediate rebase states. Rebasing always succeeds and if there are conflicts you’re able to address them just by editing your changes as you would in any other situation.
3. I don’t have to make WIP commits or stash when switching between lines of work and I don’t have to remember to undo them when I come back. Your work is always committed. Stashing doesn’t need to exist any more and as a result I never have to worry about conflicts when unstashing. I never have to realize I’ve forgotten work that was stashed.
4. Named branches don’t follow you automatically. This is huge if you ever have to do some linear work that involves multiple PRs that have to be merged and deployed one by one by one. With git you have to make branch after branch after branch. And god help you if one of the earlier lines of work needs to be changed, now you have to rebase all the descendants. With jj all this work is linear. If you have to edit an earlier commit or insert a new commit inbetween two, it’s trivial.
5. And the real killer feature? I get to have all this and nobody on my team needed to change a thing for themselves.
I was a git expert. I’ve used it since before GitHub. I’ve written a barebones implementation of git from scratch. I was the guy everyone came to when they got stuck. I gave talks on git to my local language user meetup. And I will never need to touch it again.
I think that's going to be pretty divisive. For projects where your input is here, your output is there, and that's that I imagine it's fine. Getting people to implement one of the workarounds to avoid junking up the repo feels like a huge ask.
The auto add feature is configurable these days though, so if you don’t like it, you can turn it off.
I was skeptical until I tried it.
I guess I should ask... is that final step made super easy? That is does jj make it easy to completely remove all traces from the repo of those auto-added files (like filter-branch)?
If git auto-added them from the get go, you would have used a pattern in your personal or project gitignore to skip them, or you would have put them into an ignored tmpdir, or you would keep them on a separate branch if you intend to make them part of the repo. None of these steps would have felt painful or out of place. You see files you don’t intend to include in `jj status` and you’d do whatever is appropriate to exclude them.
I’m with Steve here. I too was skeptical, and it’s turned out to be a complete non-issue.
> and I'd have to something akin to git's filter-branch to clean them out again to avoid that bloat
It’s even more trivial than you might imagine. If you have a commit that added files you don’t want to include you `jj edit` that commit, remove them however is appropriate (rm, mv, add to gitignore, whatever), and… there is no step 3. Later commits are instantaneously and automatically rebased so as to not include those files. Tools like filter-branch are completely unnecessary.
As someone who also kept a lot of files in their repo that went untracked, it was a small adjustment. Add to my .gitignore, untrack the files, and done.
That's one of the workarounds I had in mind. :)
This week I've had to work with CSVs, SQLite databases, and video in... six... different formats. Some goes in the repo, some doesn't.
It seems like my choices are:
- Disable it in the repo's .jj/repo/config.toml.
- Get people that don't like it to turn it off.
- Get people to do their exploratory coding in a gitignored directory.
It feels like any choice I make is going to be met with resistance.
For me, writing temporary output to a gitignored tmp directory has been trivial. And I do exploratory coding in a separate branch. Just because it’s part of the repo doesn’t mean it needs to get pushed.
Currently, sure. I was more thinking about it as a possible Git successor.
However, why would the contributors to jj not just try to make git better by addressing these weaknesses?
I’m not being glib, just that it puzzzles me when a new oss comes out that does the same thing as another tool but a a bit different. I would have though that there is a way to have a git plug-in or even a way to contribute to git to enhance the issues outlined.
Yes, I know that some code bases are too far gone for major enhancements…
Because the git workflow causes a lot of the difficulty and fixing it involves making a new concept: mutable changesets with a unique id which are built on top of immutable commits.
jj isn’t just some bugfixes or a few porcelain improvements, it deeply changes the way you approach things. And there’s a pretty decent amount of unlearning of subtly-broken concepts you need to do.
None of that could ever make its way into git.
Also, some of the UI improvements or even protocol improvements might make it into git eventually.
I’m skeptical about it being worthwhile to switch to a new backend. Maybe it would work if it’s just local, but a lot of IDE plugins and other tools would need to be written, and git has lots of inertia.
Because jj isn't being built to improve git, but to allow using git at Google to work on Piper-backed code instead of hg/fig, and adding support for the abstractions that are needed to support git and piper was probably seen as free complexity on git's side.
I'm a bit surprised then, that's one of the things that has kept me from trying it as it feels weird to not have that much control over what gets into the commit. Using magit doesn't help as it's just way too good once you get used to it.
You would have thought? Plugins? Git doesn’t have plugins to my knowledge. Hooks are not plugins for security reasons. Meaning you can’t distribute hooks, have people install them and call it an extension.
Then, assuming that they could just make Git better. Are they interested in supporting all of Git plus their new work? Why would they be?
But when an alternative comes along and long-term git users try to tell people how much better it is? Everyone immediately forgets all their issues and frustrations. It’s honestly kind of bizarre.
My thought would be that most people are happy and it’s the <5% of people who complain the loudest that are heard.
Your org doesn’t have to adopt the tool. Nobody on my team needs to know or care that I use jj instead of git. The only people who do know or care have themselves switched when I showed it off to them.
I’ve worked too many places where I’ve helped fix too many coworkers’ broken git repos to believe in a simple git workflow. Basically everyone uses the same fetch/branch/commit/merge-to-main approach and people still constantly run into problems. None of the people with some claimed simple workflow are doing anything meaningfully different than what everyone else is doing.
It’s just astonishingly easy to internalize all the fixes and band aids we’ve adopted to smooth the sharp edges and forget how often we have to work around them.
I will admit there is occasional call where I have to get someone out of a twisted pretzel, but that is few and far between.
All of these things (and others) can be worked around with varying levels of annoyance ranging from just living with it to new commands that help out to changing your workflow to completely avoiding some workflows. But in my experience nearly everyone deals with those annoyances on a semi-regular basis.
And before you say well this doesn’t work with a bigger team, my team is 8 and the org is 50 on the same codebase. At “google scale”, I understand this might not be the same case.
If someone has to go out of that workflow to fix a previous commit on main, they submit a pr on top of latest.
Again, I don’t see the major complications here. It seems to me to be fixing a communication issue in an org more than anything.
jj has no evidence today to support the idea that it will be a main tool in the future. An even heavier burden of proof comes from the idea that it could be adopted by mainstream forges, so it's certainly not an argument on how jj is better than git.
The others are also listed in "the good" section of the article, but that's the biggest one. How much of a benefit that is for you will really depend on your usage patterns - maybe it's solving a problem you don't have. But the nice thing is since it's git compatible, people for whom this solves a problem can use jj, without having to migrate their whole team/company from git, unlike other solutions like fossil, darcs or pijul.
Also in my experience everyone says “my default workflow is easy” when you talk about an alternative VCS but in every other context they complain about how often that workflow starts getting exceptional: rebase conflicts, stash conflicts, WIP commits, uncommitted work accidentally blown away, etc.
On the absolute happy path what you do above is the same I do, where jj shines is when you het out of the happy path.
Theres no overhead at all in thinking these, you can never lose work by mistake, you can never mess it up. It just works
Tools like the stash are band aids that come with their own set of “fun” failure modes.
The easiest and by far most common thing to do of course is to just complete the commit you're working on before switching branches.
Wouldn’t it be nice if there was a tool that didn’t need you to internalize all these workarounds?
I don't consider any of this a problem, but why don't you tell me how your preferred tool does it.
jj needs neither because everything in your working copy is already part of a commit. You can’t accidentally push unfinished changes because jj will complain if it hasn’t been described (given a commit message). If you need to move changes between commits you don’t need to stash because you can easily move the contents of a commit around.
I can see that being nice I guess.
If jj's stashes are just regular commits, don't you have the problem where you accidentally push a stash just like you push a WIP commit?
Because jj doesn’t require commits to be on a branch, and the equivalent feature (bookmarks) don’t automatically advance, if you created a new commit to stash some stuff, the name wouldn’t move to it, and so when you push nothing would happen.
Then checkout the new branch and do my work.
When I come back, `git reset --soft HEAD^`, and continue
In jj, resolutions are first-class objects—treated the same as any normal commit, pushed/pulled like normal, etc.
Linus doesn’t want people who he pulls from to back-merge, i.e. to merge at “random points” from him into their trees to keep up to date. But what about merge conflicts? Then he has to deal with them. So then he might begrudgingly ask them to back-merge right before making a pull: a real zig-zag pattern.
First-class conflicts would solve all that I think.
edit: everyone is so focused on the "single use case" part and not the fact that it looks more complicated and verbose than regular git. Unless everyone using it is some git god that runs into edge cases left and right... maybe my workflow is just not that complicated to need a solution to problems I don't have?
Using jj is less complicated in—as far as I can tell with six months of conversion—actually 100% of the time. I spent one day adapting and maybe two week of having to look something up every other day. And now all of git’s infamous pain points are gone.
There is no index. There is no stash. Rebase conflicts are (and rebasing in general is) no longer painful. Jumping between work is a no-op. Fixing up earlier work is a no-op. There are no modal “interim” states like during conflicts where you have to abort or continue to get back to normal.
You don’t have to be a git god to run into the many sharp edges of git, and all of the “but my workflow is simple” comments break down upon even cursory inspection. Everyone’s workflow is simple until a rebase goes wrong, or they unstash in the wrong branch, or they need to unfuck an earlier commit, and that list goes on and on. And when you’re in these states you have no practical way to save or persist incomplete work.
Jujutsu is both more powerful and simpler, and I say that as someone who knows git inside and out and who’s used it pre-GitHub.
Is Git not verbose and complicated?
It is like the poster child of it among ALL THE OTHER options in his space!
---
I was a heavy user of mercurial, I learned it in a day or two, then maybe only had one or 2 major complications for years.
I use git now for 'social' pressure reasons, and despite I use it more heavily than hg, i need to check stuff every week.
And have some bash scripts around to automate how to back out of rebase hell.
I predict I will have enough of it in the next mess of trying to stack PRs or resolve another rebase of 30 steps.
Another thing you can do is work on multiple branches at the same time. See: https://ofcr.se/jujutsu-merge-workflow
But as I’ve read (according to Martin) doing it this way simplifies the implementation a lot. You track everything directly in the DB. In turn you don’t have to care about all sorts of states like being in conflicted git-am state, conflicted git-merge, etc.—these states cannot happen, it’s all in the DB (not some dirty working tree state where some flags or files somewhere in `.git` or `.jj` tell you “there are some conflict markers thereabouts, enjoy”).
And that frames it in terms of being easier for the people implementing jj. But I think it will also benefit us users by making it easier for them to create streamlined tools.
git-rebase(1) and all that is great and all. But definitely messy. And we only recently got `--update-refs` to update dependent branches, which is kind of late considering how important branches are in Git.
Every single engineer has to start over in their mastery experience by learning all the new quirks and issues, and all the tools, libraries, pipelines need to be rewritten.
IMO better to support the dominant tooling than invest in fracturing one of the few areas where engineers can agree on something for once (to everyone's benefit).
But jj supports git as a backend, so not clear what fracturing you mean, do people using magit front-end fracture?
This allows the people who want the 15% gain to have it, while not forcing anyone to do a costly migration.
Having distinct local working copies of same repo per git branch still works great, not stash needed, no conflicts or pain with local wip and drafts changes overwritten by pull. No blocker for switching branches. No reset hard/soft whatever. And makes files comparisons easier and makes my git workflows easier.
I still feel vcs could be simpler and solved problem if more engineers care about the topic of vcs and generally speaking about the ux of their main tools instead of creating new TUI and plugins to workaround broken ones.
I was a git power user. I became comfortable enough with jj to replace git entirely in one day.
If Linus had this mentality we’d all still be using Subversion.
Anyway Jujutsu stil feels great, I just hope now to not need to suffer like I did with git.
For people who work on repos by themselves or with a small number of others, the advantages of `jj` would be smaller in most cases, I would think.
I did it in half a day.
Git is bad.
It's written in Rust.
To me, the real value I saw when playing with jj was the drag-and-drop change history manipulation in the `gg` GUI app -- and as far as I can tell, that could just as well live directly on top of git.
I don't see why I'd use this over fossil for my personal projects though.
I learned jj in one morning and will never go back to git.