[raphlinus]

I think that's overstating it a bit. What I would say is that the CRDT added a lot of complexity, and we weren't able to encapsulate that complexity into an "engine" as was originally hoped.

Previous HN discussion on the mini-retrospective might be interesting: https://news.ycombinator.com/item?id=19886883

[AriaMinaei]

Thanks for the correction! Is it more accurate to say that in Xi, CRDTs ended up leaking too much complexity for the purpose of synchronizing state between async components of the editor (like plugins) that run on the same machine, but not necessarily for the purpose of collaborative editing between users where there is significant latency?

[raphlinus]

I think that's fair. We had an experimental collaborative editing prototype running on Fuchsia, but that was never officially part of the scope of the project. So this is all about whether technology developed for multiuser collaboration is helpful for the single user case.

A good example is inserting indentation in response to, say, the "Enter" key. In a synchronous model, this is pretty straightforward. You calculate the indentation as essentially a function of the contents of the buffer and the delta (the insert of the newline), then any other typing gets processed after that. In a CRDT model, the additional typing can be interleaved in any order, and the merge operation has to converge to the correct results. You can think about how bad that is: the user could type a hundred lines of code, complete with complex editing operations including cut and paste, deletions, undo, and then the auto-indent plugin wakes up, calculates the diff, and sends it to the core. Meanwhile, the user has typed another hundred lines, including more cut'n'pasting of the affected lines. This is perhaps unrealistic, but the CRDT model requires it. And I think it is possible to solve it (I wrote some "rope science" articles, then there was more discussion in several issues), but it's a hugely intricate puzzle compared with just applying it synchronously, and the benefit is not worth it. It's quite practical to compute indentation for any realistic source file in under a millisecond, if you've got a nice fast language.