This is cool and i keep thinking about CRDTs as a baseline for version control, but CRDTs has some major issues, mainly the fact that most of them are strict and "magic" in the way they actually converge(like the joke: CRDTs always converge, but to what).
i didn't read if he's using some special CRDT that might solve for that, but i think that for agentic work especially this is very interesting
The thing about how merges are presented seems orthogonal to how to represent history. I also hate the default in git, but that is why I just use p4merge as a merge tool and get a proper 4-pane merge tool (left, right, common base, merged result) which shows everything needed to figure out why there is a conflict and how to resolve it. I don't understand why you need to switch out the VCS to fix that issue.
<<<<<<< left
||||||| base
def calculate(x):
a = x * 2
b = a + 1
return b
=======
def calculate(x):
a = x * 2
logger.debug(f"a={a}")
b = a + 1
return b
>>>>>>> right
With this configuration, a developer reading the raw conflict markers could infer the same information provided by Manyana’s conflict markers: that the right side added the logging line.
That still have an issue with the vocabulary. Things like "theirs/our" is still out of touch but it's already better than a loose spatial analogy on some representation of the DAG.
Something like base, that is "common base", looks far more apt to my mind. In the same vein, endogenous/exogenous would be far more precise, or at least aligned with the concern at stake. Maybe "local/alien" might be a less pompous vocabulary to convey the same idea.
I'm on my phone right now so I'm not going to dig too hard for this, but you can also configure a "merge tool" (or something like that) so you can use Meld or Kompare to make the process easier. This has helped me in a pinch to work out some confusing merge conflicts.
Seconding the use of p4merge for easy-to-use three-pane merging. Just like most other issues with Git, if your merges are painful it's probably due to terrible native UX design - not due to anything conceptually wrong with Git.
Did you know that VS Code added support for the same four-pane view as p4merge years ago? I used p4merge as my merge tool for a long time, but I switched to VS Code when I discovered that, as VS Code’s syntax highlighting and text editing features are much better than p4merge’s.
I also use the merge tool of JetBrains IDEs such as IntelliJ IDEA (https://www.jetbrains.com/help/idea/resolve-conflicts.html#r...) when working in those IDEs. It uses a three-pane view, not a four-pane view, but there is a menu that allows you to easily open a comparison between any two of the four versions of the file in a new window, so I find it similarly efficient.
Thirding it except I do it from Emacs. Three side-by-side pane with left / common ancestor / right and then below the merge result. By default it's not like that but then it's Emacs so anything is doable. I hacked some elisp code a great many years ago and I've been using it ever since.
No matter the tool, merges should always be presented like that. It's the only presentation that makes sense.
I think you need to enable 3 way merge by default in git's configuration, and both smerge (minor mode for solving conflicts) and ediff (major mode that encompass diff and patch) will pick it up. In the case of the latter you will have 4 panes, one for version A, another for version B, a third for the result C, and the last is the common ancestor of A and B.
Addendum:
I've since long disabled it. A and B changes are enough for me, especially as I rebase instead of merging.
Is it a good thing to have merges that never fail? Often a merge failure indicates a semantic conflict, not just "two changes in the same place". You want to be aware of and forced to manually deal with such cases.
I assume the proposed system addresses it somehow but I don't see it in my quick read of this.
My understanding of the way this is presented is that merges don't _block_ the workflow. In git, a merge conflict is a failure to merge, but in this idea a merge conflict is still present but the merge still succeeds. You can commit with conflicts unresolved. This allows you to defer conflict resolution to later. I believe jj does this as well?
Technically you could include conflict markers in your commits but I don't think people like that very much
The way I'm reading it (and reading in between the lines) there wouldn't be conflict lines in the file. You'd still have the conflict to resolve eventually, but not tracked by representing it in the file with conflict markup in it.
I could be way off base but that is what it seems the claim is to me.
Should you be counting on confusion of an underpowered text-merge to catch such problems?
It'll fire on merge issues that aren't code problems under a smarter merge, while also missing all the things that merge OK but introduce deeper issues.
Post-merge syntax checks are better for that purpose.
And imminently: agent-based sanity-checks of preserved intent – operating on a logically-whole result file, without merge-tool cruft. Perhaps at higher intensity when line-overlaps – or even more-meaningful hints of cross-purposes – are present.
> It'll fire on merge issues that aren't code problems under a smarter merge, while also missing all the things that merge OK but introduce deeper issues.
That has not been my experience at all. The changes you introduced is your responsibility. If you synchronizes your working tree to the source of truth, you need to evaluate your patch again whether it introduces conflict or not. In this case a conflict is a nice signal to know where someone has interacted with files you've touched and possibly change their semantics. The pros are substantial, and it's quite easy to resolve conflicts that's only due to syntastic changes (whitespace, formatting, equivalent statement,...)
They address this; it's not that they don't fail, in practice...
the key insight is that changes should be flagged as conflicting when they touch each other, giving you informative conflict presentation on top of a system which never actually fails.
It says that merges that involve overlap get flagged to the user. I don't think that's much more than a defaults difference to git really. You could have a version of git that just warns on conflict and blindly concats the sides.
I realized recently that I've subconsciously routed-around merge conflicts as much as possible. My process has just subtly alerted to make them less likely. To the point of which seeing a 3-way merge feels jarring. It's really only taking on AI tools that bought this to my attention.
I'm surprised to see that some people sync their working tree and does not evaluate their patch again (testing and reviewing the assumptions they have made for their changes).
I agree. Nevertheless I wonder if this approach can help with certain other places where Git sometimes struggles, such as whether or not two commits which have identical diffs but different parents should be considered equivalent.
In the general case, such commits cannot be considered the same — consider a commit which flips a boolean that one branch had flipped in another file. But there are common cases where the commits should be considered equivalent, such as many rebased branches. Can the CRDT approach help with e.g. deciding that `git branch -d BRANCH` should succeed when a rebased version of BRANCH has been merged?
Nobody should have these types of problems in the age of AI agents. This kind of clean up and grunt work is perfect for AI agents. We don’t need new abstractions.
This thing is really short. https://github.com/bramcohen/manyana/blob/main/manyana.py is 473 lines of dependency-free Python (that file only imports difflib, itertools and inspect) and of that ~240 lines are implementation and the rest are tests.
It's really impressive what can be done in a few hundred lines of well-thought-out Python without resorting to brutal hacks. People complain about left-pad incidents etc. in the JS world but I honestly feel like the Python ecosystem could do with more, smaller packages on balance. They just have to be put forward by responsible people who aren't trying to make a point or inflate artificial metrics.
I think something like this needs to be born out of analysis of gradations of scales of teams using version control systems.
- What kind of problems do 1 person, 10 person, 100 person, 1k (etc) teams really run into with managing merge conflicts?
- What do teams of 1, 10, 100, 1k, etc care the most about?
- How does the modern "agent explosion" potentially affect this?
For example, my experience working in the 1-100 regime tells me that, for the most part, the kind of merge conflict being presented here is resolved by assigning subtrees of code to specific teams. For the large part, merge conflicts don't happen, because teams coordinate (in sprints) to make orthogonal changes, and long-running stale branches are discouraged.
However, if we start to mix in agents, a 100 person team could quickly jump into a 1000 person team, esp if each person is using subagents making micro commits.
It's an interesting idea definitely, but without real-world data, it kind of feels like this is just delivering a solution without a clear problem to assign it to. Like, yes merge-conflicts are a bummer, but they happen infrequently enough that it doesn't break your heart.
Team scale doesn't tend to impact this that much, since as teams grow they naturally specialize in parts of the codebase. Shared libs can be hotspots, I've heard horror stories at large orgs about this sort of thing, though usually those shared libs have strong gatekeeping that makes the problem more one of functionality living where it shouldn't to avoid gatekeeping than a shared lib blowing up due to bad change set merges.
This is sort of a revival and elaboration of some of Bram’s ideas from Codeville, an earlier effort that dates back to the early 2000s Cambrian explosion of DVCS.
Codeville also used a weave for storage and merge, a concept that originated with SCCS (and thence into Teamware and BitKeeper).
Codeville predates the introduction of CRDTs by almost a decade, and at least on the face of it the two concepts seem like a natural fit.
It was always kind of difficult to argue that weaves produced unambiguously better merge results (and more limited conflicts) than the more heuristically driven approaches of git, Mercurial, et al, because the edit histories required to produce test cases were difficult (at least for me) to reason about.
I like that Bram hasn’t let go of the problem, and is still trying out new ideas in the space.
Note that CRDT isn't "a thing". The CRDT paper provides a way to think about and analyze eventually consistent replication mechanisms. So CRDTs weren't "introduced", only the "CRDT way of discussing replication". Every concrete mechanism described in the CRDT paper is very old, widely used for decades beforehand.
This means that everything that implements eventual consistency (including Git) is using "a CRDT".
I have a weird hobby: about once a year I go to the theory page [0] in pijul manual and see if they have fixed the TeX formatting yet.
You would think that if a better, more sound model of storing patches is your whole selling point, you would want to make as easy as possible for people who are interested in the project to actually understand it. It is really weird not to care about the first impression that your manual makes on a curious reader.
Currently, I'm about 6 years into the experiment.
Approximately 2 years in (about 4 years ago), I've actually went to the Pijul Nest and reported [1] the issue. I got an explanation on fixing this issue locally, but weirly enough, the fix still wasn't actually implemented on the public version.
I'll report back in about a year with an update on the experiment.
I'm surprised! Pijul has been discussed here on HN many, many times. My impression is that many people here were hoping that Pijul might eventually become a serious Git contender but these days people seem to be more excited about Jujutsu, likely because migration is much easier.
Interesting idea. While conflicts can be improved, I personally don't see it as a critical challenge with VCS.
What I do think is the critical challenge (particularly with Git) is scalability.
Size of repository & rate of change of repositories are starting to push limits of git, and I think this needs revisited across the server, client & wire protocols.
What exactly, I don't know. :). But I do know that in my current role (mid-size well-known tech company) is hitting these limits today.
This is a really interesting and well thought out idea, especially the way it turns conflicts into something informative instead of blocking. The improved conflict display alone makes it much easier to understand what actually happened. I think using CRDTs to guarantee merges always succeed while still keeping useful history feels like a strong direction for version control. Looks like a solid concept!
doesn't the side by side view in github diff solve this?
conflict free merging sounds cool, but doesn't that just mean that that a human review step is replaced by "changes become intervals rather than collections of lines" and "last set of intervals always wins"? seems like it makes sense when the conflicts are resolved instantaneously during live editing but does it still make sense with one shot code merges over long intervals of time? today's systems are "get the patch right" and then "get the merge right"... can automatic intervalization be trusted?
edit: actually really interesting if you think about it. crdts have been proven with character at a time edits and use of the mouse select tool.... these are inherently intervalized (select) or easy (character at a time). how does it work for larger patches can have loads of small edits?
See vim-mergetool[1]. I use it to manage merge conflicts and it's quite intuitive. I've resolved conflicts that other people didn't even want to touch.
This seems like an excellent idea. I'm sure a lot of us have been idly wondering why CRDTs aren't used for VCS for some time, so it's really cool to see someone take a stab at it! We really do need an improvement over git; the question is how to overcome network effects.
This is actually a very interesting moment to potentially overcome network effects, because more and more code is going to be written by agents. If a crdt approach is measurably better for merging by agent swarms then there is incentive to make the switch. It also much easier to get an agent to change its workflow than a human. The only tricky part is how much git usage is in the training set so some careful thought would need to be given to create a compatibility layer in the tooling to help agents along.
Overcoming network effects cannot be the goal; otherwise, work will never get done.
The goal should be to build a full spec and then build a code forge and ecosystem around this. If it’s truly great, adoption will come. Microsoft doing a terrible job with GitHub is great for new solutions.
Well over half of all people can’t tell you the difference between git and Github. The latter being owned by a corporation that needs the network effect to keep existing.
starts with “based on the fundamentally sound approach of using CRDTs for version control”. How on earth is crdt a sound base for a version control system? This makes no sense fundamentally, you need to reach a consistent state that is what you intended not what some crdt decided and jj shows you can do that also without blocking on merges but with first level conflicts that need to be resolved. ai and language aware merge drivers are helping so much here i really wonder if the world these “replace version control” projects were made for still exists at all.
> ai and language aware merge drivers are helping so much here i really wonder if the world these “replace version control” projects were made for still exists at all.
I really wonder what kinds of magical AI you're using, because in my experience, Claude Code chokes and chokes hard on complex rebases/merge conflicts to the point that I couldn't trust it anymore.
But the presentation does not show how it resolves conflicts. For the first example, Git has the 3 way-merge that shows the same kind of info. And a conflict is not only to show that two people have worked on a file. More often than not, it highlight a semantic changes that happened differently in two instances and it's a nice signal to pay attention to this area. But a lot of people takes merge conflicts as some kind of nuisance that prevents them from doing their job (more often due to the opinion that their version is the only good one).
There are many ways to instantiate a CRDT, and a trivial one would be "last write wins". LWW is obviously not what you'd want for source version control. It is "correct" per its own definition, but it is not useful.
Anyone saying "CRDTs solve this" without elaborating on the specifics of their CRDT is not saying very much at all.
A suggestion : is there any info to provide in diffs that is faster to parse than "left" and "right" ? Can the system have enough data to print "bob@foo.bar changed this" ?
My issue with git is handling non-text files, which is a common issue with game development. git-lfs is okay but it has some tricky quirks, and you end up with lots of bloat, and you can't merge. I don't really have an answer to how to improve it, but it would be nice if there was some innovation in that area too.
Totally agree. After trying to flesh out Unreal's git plugin, it really shows how far from ideal git really is.
Partial checkouts are awkward at best, LFS locks are somehow still buggy and the CLI doesn't support batched updates. Checking the status of a remote branch vs your local (to prevent conflicts) is at best a naive polling.
Better rebase would be a nice to have but there's still so much left to improve for trunk based dev.
What strategies would you like to use to diff the binaries? Or else how are you going to avoid bloat?
Is it actually okay to try to merge changes to binaries? If two people modify, say, different regions of an image file (even in PNG or another lossless compression format), the sum of the visual changes isn't necessarily equal to the sum of the byte-level changes.
Has there ever been a consideration for the git file format to allow storage of binary blobs uncompressed?
When I was screwing around with the Git file format, tricks I would use to save space like hard-linking or memory-mapping couldn't work, because data is always stored compressed after a header.
A general copy-on-write approach to save checkout space is presumably impossible, but I wonder what other people have traveled down similar paths have concluded.
I really think something like Xet is a better idea to augment Git than LFS, though it seems to pretty much only be used by HuggingFace for ML model storage, and I think their git plugin was deprecated? Too bad if it ends up only serving the HuggingFace niche.
In case the name doesn't jump out at you, this is Bram Cohen, inventory of Bittorrent. And Chia proof-of-storage (probably better descriptions available) cryptocurrency. https://en.wikipedia.org/wiki/Bram_Cohen
It's not the same as capturing it, but I would also note that there are a wide wide variety of ways to get 3-way merges / 3 way diffs from git too. One semi-recent submission (2022 discussing a 2017) discussed diff3 and has some excellent comments (https://news.ycombinator.com/item?id=31075608), including a fantastic incredibly wide ranging round up of merge tools (https://www.eseth.org/2020/mergetools.html).
It sounds interesting but the main selling point doesn’t really reasonate:
If you haven’t resolved conflicts then it probably doesn’t compile and of course tests won’t pass, so I don’t see any point in publishing that change? Maybe the commit is useful as a temporary state locally, but that seems of limited use?
Nowadays I’d ask a coding agent to figure out how to rebase a local branch to the latest published version before sending a pull request.
Jujutsu honestly is the future IMO, it already does what you have outlined but solved in a different way with merges, it'll let you merge but outline you have conflicts that need to be resolved for instance.
It's been amazing watching it grow over the last few years.
Something like base, that is "common base", looks far more apt to my mind. In the same vein, endogenous/exogenous would be far more precise, or at least aligned with the concern at stake. Maybe "local/alien" might be a less pompous vocabulary to convey the same idea.
I also use the merge tool of JetBrains IDEs such as IntelliJ IDEA (https://www.jetbrains.com/help/idea/resolve-conflicts.html#r...) when working in those IDEs. It uses a three-pane view, not a four-pane view, but there is a menu that allows you to easily open a comparison between any two of the four versions of the file in a new window, so I find it similarly efficient.
No matter the tool, merges should always be presented like that. It's the only presentation that makes sense.
Addendum: I've since long disabled it. A and B changes are enough for me, especially as I rebase instead of merging.
I assume the proposed system addresses it somehow but I don't see it in my quick read of this.
Technically you could include conflict markers in your commits but I don't think people like that very much
True but it is not valid syntax. Like, you mean with the conflict lines?
I could be way off base but that is what it seems the claim is to me.
It'll fire on merge issues that aren't code problems under a smarter merge, while also missing all the things that merge OK but introduce deeper issues.
Post-merge syntax checks are better for that purpose.
And imminently: agent-based sanity-checks of preserved intent – operating on a logically-whole result file, without merge-tool cruft. Perhaps at higher intensity when line-overlaps – or even more-meaningful hints of cross-purposes – are present.
That has not been my experience at all. The changes you introduced is your responsibility. If you synchronizes your working tree to the source of truth, you need to evaluate your patch again whether it introduces conflict or not. In this case a conflict is a nice signal to know where someone has interacted with files you've touched and possibly change their semantics. The pros are substantial, and it's quite easy to resolve conflicts that's only due to syntastic changes (whitespace, formatting, equivalent statement,...)
the key insight is that changes should be flagged as conflicting when they touch each other, giving you informative conflict presentation on top of a system which never actually fails.
FWIW I've struggled to get AI tools to handle merge conflicts well (especially rebase) for the same underlying reason.
I realized recently that I've subconsciously routed-around merge conflicts as much as possible. My process has just subtly alerted to make them less likely. To the point of which seeing a 3-way merge feels jarring. It's really only taking on AI tools that bought this to my attention.
In the general case, such commits cannot be considered the same — consider a commit which flips a boolean that one branch had flipped in another file. But there are common cases where the commits should be considered equivalent, such as many rebased branches. Can the CRDT approach help with e.g. deciding that `git branch -d BRANCH` should succeed when a rebased version of BRANCH has been merged?
- What kind of problems do 1 person, 10 person, 100 person, 1k (etc) teams really run into with managing merge conflicts?
- What do teams of 1, 10, 100, 1k, etc care the most about?
- How does the modern "agent explosion" potentially affect this?
For example, my experience working in the 1-100 regime tells me that, for the most part, the kind of merge conflict being presented here is resolved by assigning subtrees of code to specific teams. For the large part, merge conflicts don't happen, because teams coordinate (in sprints) to make orthogonal changes, and long-running stale branches are discouraged.
However, if we start to mix in agents, a 100 person team could quickly jump into a 1000 person team, esp if each person is using subagents making micro commits.
It's an interesting idea definitely, but without real-world data, it kind of feels like this is just delivering a solution without a clear problem to assign it to. Like, yes merge-conflicts are a bummer, but they happen infrequently enough that it doesn't break your heart.
Codeville also used a weave for storage and merge, a concept that originated with SCCS (and thence into Teamware and BitKeeper).
Codeville predates the introduction of CRDTs by almost a decade, and at least on the face of it the two concepts seem like a natural fit.
It was always kind of difficult to argue that weaves produced unambiguously better merge results (and more limited conflicts) than the more heuristically driven approaches of git, Mercurial, et al, because the edit histories required to produce test cases were difficult (at least for me) to reason about.
I like that Bram hasn’t let go of the problem, and is still trying out new ideas in the space.
This means that everything that implements eventual consistency (including Git) is using "a CRDT".
> ... CRDTs for version control, which is long overdue but hasn’t happened yet
Pijul happened and it has hundreds - perhaps thousands - of hours of real expert developer's toil put in it.
Not that Bram is not one of those, but the post reads like you all know what.
You would think that if a better, more sound model of storing patches is your whole selling point, you would want to make as easy as possible for people who are interested in the project to actually understand it. It is really weird not to care about the first impression that your manual makes on a curious reader.
Currently, I'm about 6 years into the experiment.
Approximately 2 years in (about 4 years ago), I've actually went to the Pijul Nest and reported [1] the issue. I got an explanation on fixing this issue locally, but weirly enough, the fix still wasn't actually implemented on the public version.
I'll report back in about a year with an update on the experiment.
[0] https://pijul.org/manual/theory.html
[1] https://nest.pijul.com/pijul/manual/discussions/46
... and of course it is, because Pijul uses Pijul for development, not Git and GitHub!
I'm surprised! Pijul has been discussed here on HN many, many times. My impression is that many people here were hoping that Pijul might eventually become a serious Git contender but these days people seem to be more excited about Jujutsu, likely because migration is much easier.
Funny, there was just a post a couple of days ago how this is false.
https://news.ycombinator.com/item?id=47359712
What I do think is the critical challenge (particularly with Git) is scalability.
Size of repository & rate of change of repositories are starting to push limits of git, and I think this needs revisited across the server, client & wire protocols.
What exactly, I don't know. :). But I do know that in my current role (mid-size well-known tech company) is hitting these limits today.
conflict free merging sounds cool, but doesn't that just mean that that a human review step is replaced by "changes become intervals rather than collections of lines" and "last set of intervals always wins"? seems like it makes sense when the conflicts are resolved instantaneously during live editing but does it still make sense with one shot code merges over long intervals of time? today's systems are "get the patch right" and then "get the merge right"... can automatic intervalization be trusted?
edit: actually really interesting if you think about it. crdts have been proven with character at a time edits and use of the mouse select tool.... these are inherently intervalized (select) or easy (character at a time). how does it work for larger patches can have loads of small edits?
[1]: https://github.com/samoshkin/vim-mergetool
The goal should be to build a full spec and then build a code forge and ecosystem around this. If it’s truly great, adoption will come. Microsoft doing a terrible job with GitHub is great for new solutions.
I really wonder what kinds of magical AI you're using, because in my experience, Claude Code chokes and chokes hard on complex rebases/merge conflicts to the point that I couldn't trust it anymore.
Well, isn't that what the CRDT does in its own data structure ?
Also keep in mind that syntactic correctness doesn't mean functional correctness.
There are many ways to instantiate a CRDT, and a trivial one would be "last write wins". LWW is obviously not what you'd want for source version control. It is "correct" per its own definition, but it is not useful.
Anyone saying "CRDTs solve this" without elaborating on the specifics of their CRDT is not saying very much at all.
Partial checkouts are awkward at best, LFS locks are somehow still buggy and the CLI doesn't support batched updates. Checking the status of a remote branch vs your local (to prevent conflicts) is at best a naive polling.
Better rebase would be a nice to have but there's still so much left to improve for trunk based dev.
Is it actually okay to try to merge changes to binaries? If two people modify, say, different regions of an image file (even in PNG or another lossless compression format), the sum of the visual changes isn't necessarily equal to the sum of the byte-level changes.
When I was screwing around with the Git file format, tricks I would use to save space like hard-linking or memory-mapping couldn't work, because data is always stored compressed after a header.
A general copy-on-write approach to save checkout space is presumably impossible, but I wonder what other people have traveled down similar paths have concluded.
It's not the same as capturing it, but I would also note that there are a wide wide variety of ways to get 3-way merges / 3 way diffs from git too. One semi-recent submission (2022 discussing a 2017) discussed diff3 and has some excellent comments (https://news.ycombinator.com/item?id=31075608), including a fantastic incredibly wide ranging round up of merge tools (https://www.eseth.org/2020/mergetools.html).
However/alas git 2.35's (2022) fabulous zdiff3 doesn't seems to have any big discussions. Other links welcome but perhaps https://neg4n.dev/blog/understanding-zealous-diff3-style-git...? It works excellently for me; enthusiastically recommended!
If you haven’t resolved conflicts then it probably doesn’t compile and of course tests won’t pass, so I don’t see any point in publishing that change? Maybe the commit is useful as a temporary state locally, but that seems of limited use?
Nowadays I’d ask a coding agent to figure out how to rebase a local branch to the latest published version before sending a pull request.
It's been amazing watching it grow over the last few years.