Jujutsu Megamerges and jj Absorb
v5.chriskrycho.com
v5.chriskrycho.com
But jj is more than that, and you couldn't implement it as a git wrapper. I think the core innovation is that rebases are so fast they are essentially free, which opens up amazing potential.
But so far I haven't found a way to really convey the full weight of this without a long explanation, because people are so used to thinking in git, it takes a long time to go beyond "so rebase -i will be faster, then? But it's not slow, is it?"
It's funny, that git is now in the position that subversion was in almost two decades ago: People were so used to svn and felt productive with it, it was hard to convey why free branches and a distributed data model are a gamechanger, and how much svn limited their thinking. But as a git user you had certainty that you were right, and that your tool is superior.
Now, people are so used to git, in an eerily similar way. But ask most git users, "how would you move a specific change from your branch to another?" I have yet to receive a confident answer and workflow to this (they do exist ofc), the common reaction I get is that this isn't something they'd ever need anyways.
I'm curious how this will look in a few years :)
Of course, no tool is truly universal, and git frontends and alternatives leverage different parts of the underlying data structures with different trade-offs.
I personally never use rebase, and am a merge guy. I don't use a gazillion branches either, not because I'm afraid of them, but my mental model works like that. Some people squash commits, others rebase, etc. Everybody is different in their ways.
svn to git was a fundamental change and allowed new things to happen. jj on the other hand makes some workflows easier while others take the back seat. So I don't see jj as a git disruptor, but an opinionated alternative tool which uses git data structures as the backend.
In my mind this demonstrates convincingly that gits conceptual model carries unnecessary incidental complexity. Within a fixed complexity budget jj absolutely does allow new things to happen.
I also don't really see any workflow examples that truly suffer under jj, so I don't think it's just another trade-off. It's a real substantial improvement. If it wasn't it wouldn't be making headway against gits network effects.
In my experience software developers come in two flavors. a) Developers who want to understand all the pieces they work with and have the desire to deep-dive into them b) Developers who write and commit code. They think that former part is their job and care about it. They don't care after it leaves their machine (in the sense of "I give box to machine it magically integrates").
IMHO, jj is very enticing for the latter camp, and that's not a bad thing. Not all of us are passion programmers/developers, and some are don't want to think about beyond their immediate realm. I respect these people. But, again, from my perspective, jj doesn't make sense for me. It's useful alright, but I actively use index, commit and stash very naturally. I might be good at understanding it because Eclipse has a great git integration which makes everything super workable and understandable, and I'd never change a power tool like git to something "push button, magic happens" class of tool.
jj might be very conductive to more complex workflows by removing some complexities, but unifying my stash, staging area and index is not something I want to be dome to my workflow for example.
Some trade-offs make great headways into networks because a large part of that network is silently suffering without being aware, and the trade-offs are worth it. See Rust for example. People accept 5x mental load and 10x compile times for memory safety. Same for jj. People accept a much simpler interface because it makes their painful workflows "push button, magic happens" levels of simplicity. Or, the people who like jj work under time pressure and want to remove a time-sink from their workflows, again understandable and respectable.
Having something like "jj" is not a bad thing. What I'm squarely against is "git is dead, new king is jj. Now move to your new kingdom and worship jj" mentality.
Unifying concepts is not taking away any expressive power. It just makes the system as a whole simpler to reason about.
I think it's a similar step up in power as moving from Windows to Linux. On Windows, to toy with the system at all you need to set up a C++ project and dig through win32 API docs and deal with opening/closing handles and void* pointers just to change anything, so I rarely bothered. On Linux everything is right there in the file system so there's almost no friction to changing things. You can do anything imaginable in a minute or two with just a shell script.
In Linux, Everything Is A File. In Jujutsu, Everything Is A Commit.
I actually like git’s CLI. I never understood why people wanted a simplified git. I’m never going back to git from jj. It’s just better.
> "git is dead, new king is jj. Now move to your new kingdom and worship jj"
No one is saying this. If you feel like people getting excited about Jujutsu is them telling you that you must love it, that's on you. The fact that it is built on top of Git means that even if others are using Jujutsu, you are not forced to when working in the same repos.
I have converted to jj and will never go back. It is simpler and more powerful.
Probably a lot of that is because it is painful. When it becomes a no-op, suddenly a lot of things become easy to do.
Do you ever want to go back and edit a previous unpublished commit? Yes fixups exist, but that’s a band aid over the fact that editing a previous one is annoying and hard.
Do you ever want to maintain a linear chain of branches? Branch A needs to be merged before Branch B needs to be merged before Branch C. It’s a massive pain with git, especially if you need to change Branch A. It’s a no-op in jj.
There are a ton of straightforward and useful workflows in git that are just completely impractical.
Also unnecessary, "git rebase -i" has an "edit" command that pauses the rebase at that commit to let you do whatever.
The rebase next level part is not about using git rebase in the right way, it's about every command including status sometimes doing 15 rebases without even mentioning it.
In-memory, always succeeding rebases are just really nice.
It’s really a completely different way of thinking, but shockingly, one that essentially directly maps to what your actual intentions are.
> Do you ever want to go back and edit a previous unpublished commit?
Honestly, no? I plan what I'm going to change, do it, commit it, and if I messed up, I don't hide it, just fix it and write "I messed up X in $HASH. This fixes that".
> Do you ever want to maintain a linear chain of branches?
No, I always keep a single thing moving at one time. If the project is big, I use a branch per feature, if it's a smaller branch, there's a development branch. That's all. I have a habit of designing before touching code, so I always move in well planned moves.
It's not like that because I'm afraid of rebase. It's just how I design/implement software. Yes, it's generally not a group work, but I do what I do, and it works.
> There are a ton of straightforward and useful workflows in git that are just completely impractical.
That's fine. I don't claim otherwise. I just don't do to prefer source code / branch acrobatics. I'd love to share a seven year development tree to show how I work, but unfortunately that repo is not meant to be open, at least for now.
For example, the solution to almost all problems is oddly kubernetes shaped even though that was not the case 15 years ago. And Google themselves would be the 1st to tell you that kubernetes is not capable of doing all jobs.
Monocultures are insidiously bad, they poison solutions and reasoning, but they have their upshots too- if you never need to learn a new mental model; us tech workers become more interchangeable and the cost to onboard goes down.
Oh I do that quite often, either because I didn’t realise I was committing on the wrong branch, or I decided to split my work in two branches, or i need some other unmerited change for my current branch…
It’s usually a rebase —onto or cherry-pick, I’d say I agree it’s not simple.
jj seems interesting, I’m very comfortable with git but I think its UX is terrible, but everyone use git and it is not an area where I have capacity or willingness to innovate. I’d love to see git being replaced with something better designed though
I don't understand how absorb works in your example. Why would the commit "z" be affiliated with some random documentation changes you come across and fix in the wip commit.
All that I see is it makes it easier to make small refinements to preciously altered lines ?
Super excited for jj.... Almost want to find a work opportunity to promote it. Been a refreshing tool I've adopted in 2024 I'm happy about.
It's not? z is the root pseudo-commit.
> All that I see is it makes it easier to make small refinements to preciously altered lines ?
Well yes, the point is to untangle working copies with various unrelated changes (usually fixups) which should really go into older commits.
There's a whole operation log (`jj op log`), as another sequence of actions, and you can undo them. It gets crazier from there, but I've also been enjoying jujutsu lately and I had to RTFM a good couple of times to get comfortable with it.
There's the usual cognitive shift, and in my case a bit of a best-practice shift I had to go through to get comfortable with `jj` -- specifically, keeping the repo directory pristine (and putting my temp output in a git-ignored dir, or outside the repo) because /everything/ is tracked. Flip side being, I've forgotten to `git add` new files at least a few times a year, and now that won't be an issue.
But yeah, can highly recommend, and I'm excited to start to jump between multiple open branches^Wbookmarks at will and learn more about the intricacies of conflict management (and the original link is a good glimpse at that!)
If you do somehow add a terabyte of small files by accident, it’s still just git under the hood so you can make sure nothing active points at them and GC them.
Also, this behavior (automatically tracking) is configurable. I thought I would hate it but I actually really like it.
And I didn't have to learn anything like the hell of git refs. The UI is self-explanatory as long as you know the feature exists.
Another very nice thing is that conflicts aren't roadblocks. They're legit changes, first-class jj citizens. So, when I do something that ends in a conflict, I can just move over to a different change, do the work I'm already trying to do, and then go back and resolve the conflict when I want.
I would honestly use jj for just this even if that was all it offered.
Two hours later you realize your experiment is a total failure and you wish you had the half-working changes from earlier. You don’t need to have actively commited anything, it’s all in the obsolte log and you can retrieve it.
I want to repeat that last bit. Unlike git, there was no point where you needed to finalize your changes with a commit in order for them to be persisted and archived for later retrieval.
I was live demoing something and performing some manual repo surgery in order to demonstrate some thing that can happen with git but is nicely solved by jj. In doing the surgery I absolutely fucked everything.
I had never used the undo feature or the op log before. In less than five minutes—during the live demo—I was able to figure out how to reset things back to a known good state. It was stupidly easy.
1. Work on smaller projects with simple histories. Git is perfectly satisfactory.
2. Work on huge projects with very complex git histories and weird legacy workflows (primarily Linux kernel).
jj is only interesting for part 2, but then it's only useful if it's easy to bridge into the Git world that everyone else uses.
But this workflow is exactly the kind of thing that seems handy. There's so much stuff that Git can totally _do_ but could just do a lot _better_.
I converted 6mo ago after finally pulling together the motivation to give it a shot. My biggest hesitation was in how long I expected to struggle with it before feeling comfortable, followed closely by skepticism about how good the compatibility story actually is.
Both, it turns out, were non issues. It took all of a day to feel perfectly comfortable using it. I spent the rest of the week gradually plugging most of the remaining holes in my workflow. And git compatibility is as advertised. A coworker switched shortly after I did and zero people have noticed or cared.
Though at this point I don’t even bother colocating the .git repo alongside .jj. Meaning I haven’t found a need to fall back to git commands in maybe four or five months.
That interests me -- is there a native jj protocol for push/fetch? Even if just ssh? Or do you just work in local repos?
Microsoft has a custom git client internally which includes a filesystem shim (think: the Windows equivalent of FUSE) since stat-ing monorepo amounts of files doesn't work. Also GitHub's running custom git servers, though their custom database backend is behind a compatibility layer.
Finally, sapling, Meta's git replacement which maintains supports for git servers also supports sapling servers (which was not released). Unsure whether to count that as a git backend though.
(I agree I probably shouldn't focus on that first, and could just try jj first on smaller rpos... But git is already slow enough in there that it's an honest question, I don't need to keep using git there as long as I can keep pulling from stable kernels git trees for regular merges)
jj git clone/fetch/push are using the corresponding git commands under the hood so they won't improve on git performance but also it doesn't have much overheard of its own.
If you're using the -T revset syntax, you can specify a revset that requires looking at every commit, which is slow, but that's equivalent to asking for `git log -n 1000000`
I think the nixpkgs workflow is much less prone to rebase/merges (some cherry-picks to stable branches I don't do much), but it's a very good data point, thank you. I guess I'll give it a try over christmas break..
Large file stuff is also a “want to do something better but no progress yet” kind of thing.
Fingers crossed. Glad we're finally getting some viable Git competition with jj, Sapling and Pijul.
Working with submodules natively (see steveklabnik’s sibling comment for the “non-native” bit) will definitely be a big win.