The downside is if you select & delete two functions, and I concurrently insert a new function in the middle of that range, then the merged result will delete my function. Thats a really bad outcome - but its pretty rare in practice.
A much more common case is, I deleted 2 lines of code and you modify a parameter in one of those function calls. In this case, using range deletes your edits will also be deleted (which is what we want). Using individual deletes, the result will end up with some random junk left behind, in the form of whichever characters I typed.
When we're editing code as plain text, I don't think there's any such thing as perfect here. For asyncronous collaborative editing (like git), I think generating conflict markers & human supervision for merging might be the best answer.
I'd love to see someone make a prototype editor which can do AST-level concurrent editing. That would be super interesting in a bunch of ways - like, you could use the AST to do syntax highlighting in any editor too. But I'm not sure how it'd merge code when the parse tree is invalid - like, if I remove a closing brace (}) early in the function, what happens?
There's a lot of cool ideas in this space. I'm excited to see people explore them as CRDTs get more popular and more mature.
Seph, thanks for showing up. Author here. This write-up is just a generalization I've landed on, while thinking about the nature of another problem: handling block elements in rich text in Peritext - a project you've collaborated with I guess.
Here's the link - https://github.com/inkandswitch/peritext/issues/27
Do throw your thoughts too. I'd be happy pick your brain on this :)
If the document has a syntax tree format, you would get the right thing if you require that range operations are broken into subranges correspond to subtrees.
Let me work this out where I imagine range operations occur as the insertion of a pair of special characters [range-delete-start] and [range-delete-end], so then we can understand results in terms of merging insertions.
Following your example, if the original is:
f(x) { ... }
g(x) { ... }
[range-delete-start]
f(x) { ... }
[range-delete-end]
[range-delete-start]
g(x) { ... }
[range-delete-end]
f(x) { ... }
[insert-start]
h(x) { ... }
[insert-end]
g(x) { ... }
[range-delete-start]
f(x) { ... }
[range-delete-end]
[insert-start]
h(x) { ... }
[insert-end]
[range-delete-start]
g(x) { ... }
[range-delete-end]
h(x) { ... }I would love to work on a structural editor for a programming language. Maybe I will start some experiments after my current project. Keep up the good work!
[0]: https://dominicpm.github.io/publications/kleppmann-highly-av...
Conflict-free means that there is no conflict in the semantics, not just that the operations on a syntax tree shape result in a valid syntax tree shape.
If we pass a syntax tree node with 12 children to two different servers which each add a node, we can easily resolve the structural conflict by adding both nodes, so now we have 14, and that may be a valid syntax tree.
However, that 12-child node could be, say, a call to a function that takes 12 arguments. (So 11 are being passed, and one is defaulted). Adding one argument is semantically valid. Adding two means the call is bad.
If you start with an incorrect premise you will wind up in a strange place. Conflict free has a specific meaning, and "semantics" can have multiple layers. If your threshold is perfect conflict free semantics at all levels, then you are indeed doomed. But a CRDT can provide a distributed, conflict-free canonical source substrate, with a specific semantic meaning at a minimally useful lowest semantic level. The challenge is to find transformations that are "good enough" to minimize conflicts at the higher semantic levels that you care about.
I'm not sure these issues can't be overcome, but it's admittedly still an open problem.
The real problem is that we're trying to solve this problem at the wrong level of abstraction.
I think CRDTs can be reformulated as a subset of a larger replicated data type model that explicitly includes conflicts as part of the model, that any chosen set of transformations can be programmatically projected into a corresponding set of conflicts, that the choices of resolutions to those conflicts can be pre-packaged and composable, that some of those resolutions may need to be 'add a new transformation that choses between conflicts created by two other conflicting transformations' (basically exposing the problem up into the model of the data type), and that a composition of chosen resolutions can be proved to be fully conflict-free.
The biggest benefits of this approach are: 1. It enables modeling domains where conflicts are an unavoidable part of the domain itself. 2. It allows tuning an initially pessimistic model where all conflicts are exposed to the user by gradually adjusting the conflict resolution choices so that it becomes more conflict-free over time; especially by composing known conflict-free subsets of the model into larger conflict-free subsets.
Can you talk a bit more on this stuff? Like with a basic example maybe? Thanks!
One nice example is from a prior thread where I aim at the same idea ( https://news.ycombinator.com/item?id=28717848#28724977 ):
> edit A sets an invoice status to paid with a payment of 100, edit B changes the invoice amount from 100 to 120
Here, the merged result should be a conflict state that needs additional attention to resolve. Maybe this happens once and the resolution is to call the customer to work it out, or maybe this happens a lot and there's a whole conflict resolution tree that applies based on the details. This kind of conflict is not possible to "eliminate by model"; its presence is integral to the problem domain itself.
But even if there is a conflict we should still be able to represent it in our model, and be able to apply one of a number of pre-packaged resolutions. For example: if the invoice amount changed due to adding a new item to the order, then the resolution could branch with a choice: 1. split the item into a separate order, 2. remove the item and notify the person that made the change to the order, 3. set the invoice status to partially-paid and notify the customer of the remaining balance and reason why it changed. If instead the price changed then the resolutions could branch differently: 1. backdate the order to before the price change, 2. give the customer a discount, 3. set the invoice to partially-paid and notify the customer of the remaining balance, etc.
Note that these choices are concerned with business processes and customer relationship which have no "right" answer and could change business to business, customer to customer, or even order to order. We can still get desirable CRDT-like high-level guarantees out of this system (like "every invoice is paid in full or billed for the remaining balance") while modeling conflicts in it. Another benefit: two customers could request different default conflict resolutions, maybe one wants to always split into a new order and the other always wants to bill the added balance; with the conflict included as part of the model interactions with both of these customers could be represented in the same system with different custom resolution strategies.
I'm not sure how to continue advancing this idea, or if I've missed something important that invalidates it. I have a feeling that this could be layered on top of existing CRDT work with a rich enough data model. My email is in my profile if you want to talk more. :)
I also think a range move or range delete is missing from many of these algorithms. A submission here on HN did implement a range delete on rich text (but I forgot what the name was).
[0]: https://www.youtube.com/watch?v=x7drE24geUw
EDIT: Ah actually Peritext (which you linked to) was the article I was thinking of. I think the range delete would be useful for regular lists as well.