I understand why you were disillusioned with the performance of native GUIs. But in other areas, such as accessibility, nativeness absolutely does help. I'm guessing you know this; I just want to make sure other developers get this message.
Here is a quite comprehensive list of the benefits of a native GUI for accessibility, particularly on macOS:
If by "something like this" you mean a text editor, then there's no question; they can be made accessible with a screen reader, and some complex ones (e.g. Visual Studio Code, Visual Studio, and Xcode) are accessible already.
> it would likely involve specific accessibility work either way.
For the main UI of a full-featured programmer's editor or word processor, that's definitely true.
What I want to avoid is developers using a custom GUI toolkit where the platform's native toolkit would do, because they want the best possible performance (regardless of whether the platform's toolkit is good enough), and forfeiting all of the accessibility benefits that the platform's native toolkit brings more or less automatically.
If you are developing a text editor, you should put the time into your text editor. It's the thing that distinguishes you from your competitors. And it's the part that users interact with the most. Your users are likely to ask for features, like vi keybindings, that you can't easily build with the OS-native text field anyway.
Incidental forms, like your preferences screen? Unless you have a lot of time and money to burn, the OS widgets will probably be better than yours.
You need it to take automated screenshots too.
*Or very close to zero
I remember looking at one of Raph's early presentations and being both convinced by his arguments for and impressed with the design decisions made, the same ones he outlines here. After doing some of my own comparatively rudimentary research, coming to similar conclusions as he did. It left me very excited to watch the project develop over time.
One thing I was wondering about is that the Xi website contains a list of front ends similar to the list in the GitHub repo, but includes a Flutter implementation for Fuchsia that is now a dead link and missing from the GitHub list. He mentions it briefly in this retrospective as well. Any ideas what happened to it?
I've run into that phenomenon a lot with various startup projects over the years. It's soul crushing, and often an immediate reaction is to assume certain projects are direct competition, rendering your project fruitless.
In retrospect, this ends up almost never being the case. In a way, you become your own worst enemy. The correct solution is to simply ignore it and press on, but this is certainly easier said than done—especially when morale is already low, and without the benefit of hindsight.
>Perhaps more important, because these projects are more ambitious than one person could really take on, the community started around xi-editor is evolving into one that can sustain GUI in Rust.
In the end, I still think that counts as a success. Perhaps it will prove to be an even greater success than the original goal.
Thanks for all these write-ups, they're excellent. Like other commenters, I've been watching at a distance for years now. Your work is much appreciated.
There's two other screen filter apps on F-Droid, and at one point I proposed that we [work together], either merging our apps or creating shared libraries. While the other devs initially expressed some interest, they weren't willing/able to put in the effort and nothing came of it. This has killed my motivation to work on Red Moon. It feels like a waste of time, duplicating other people's efforts.
That's not intrinsic to other people working separately to solve the same problem. If Red Moon were an educational project, or if it took a unique approach to screen filtering (like the other project I'm part of and [crowdfunding]), then I'd feel differently. But my main motivation was creating a FLO screen filter app, since previously I depended on the proprietary Twilight. And while there's a handful of valid approaches to screen filtering, the bulk of the effort is on UX, state transitions, and working around the kinks of Android, all of which are shared.
Note: I don't bear any hard feelings towards the other devs, genuinely -- the most frustrating bit was [lack of communication], and even that is totally understandable; it's hard to say no to something you like the idea of, even to yourself (eg, acknowledging you'll never get to that project near the bottom of your list).
[work together]: https://github.com/LibreShift/red-moon/issues/222
[lack of communication]: https://github.com/LibreShift/TransitionScheduler/issues/14
Some other side conversations I've had about this that might be interesting and relevant:
* Native UI is not always the right choice. That might seem surprising coming from someone who pushes native UI extremely strongly in general, and I don't want to be misunderstood as recommending you choose something else for your next project (if you have to ask: don't). But in some very specific contexts, such as performant text editing/rendering, it can be problematic to use a high-level control. In fact, many of the text views I use in are not using the standard platform controls but a custom view doing its own drawing: Terminal.app does its own thing (it doesn't even use NSScrollView!), Xcode has the Pegasus editor written in Swift, iTerm draws text using Metal, and Sublime Text is Skia. Platform controls are meant for 99% of apps that don't need absolute control over performance or text rendering, and sometimes you can push them quite far. But in some cases it's appropriate to abandon them. (I will note that all of the examples handle actual text editing very, very well. If you don't know what the standard shortcuts for editing text are for my platform, you're not going to do it right.)
* I love that LSP exists and that it makes my "dumb" editor "smart". But having worked with it a bit, I think it might suffer from some of the same problems: it uses a client-server architecture and JSON to communicate, which means it is not all that performant and needs to keep track of asynchronous state in a coherent way. And not just the state of one document, but the states of many documents, all in different contexts and workspaces, all with their own compiler flags and completion contexts and cursor locations–it is really hard to do this correctly and quickly, although many language servers do an admirably good job. Perhaps it's just "good enough" that it's here to stay, or maybe it has enough support that it can have some of its issues wrinkled out with time.
I agree with you that getting shortcuts (and other similar aspects) right is more important than which code is used to implement text editing. We're very much going to try to take that to heart as we implement more text capabilities in Druid.
Regarding LSP. Yes, while the functionality is there, it is very difficult to implement a really polished user experience with it. I think a lot of that is in the design of the protocol itself; it's hard to avoid annotating the wrong region when there are concurrent edits. It's interesting to think about how to do better, but I fear you'd be bringing in a lot of complexity, especially if you started doing some form of Operational Transformation. I would highly recommend that if someone were to take this on, they study xi in detail, both for what it got right and what was difficult.
I think LSP will end up being the "good enough" solution for niche languages, while popular languages get dedicated IDEs/mega-plugins, a la the various language flavours of IDEA, XCode, etc.
JSON. The protocol for front-end / back-end communication, as well as between the back-end
and plug-ins, is based on simple JSON messages. I considered binary formats, but the actual
improvement in performance would be completely in the noise. Using JSON considerably lowers
friction for developing plug-ins, as it’s available out of the box for most modern languages,
and there are plenty of the libraries available for the other ones.
The choice of JSON was controversial from the start. It did end up being a source of friction, but for surprising reasons.
For one, JSON in Swift is shockingly slow.
Also, Swift is marketed as a fast language (also based on LLVM), yet in my measurements it's 20x to 50x slower than Rust for JSON processing. I found that surprising. Would you not?
JSON's parsing performance is difficult, whether or not the language doing the parsing is fast or not. There's a reason that there are almost always libraries claiming faster JSON performance, regardless of the language.
Go is marketed as a fast language, and they're still trying to build a high performance JSON parser .
Parsing JSON will always be slower than most of the alternatives.
My exasperation is you were surprised by JSON becoming an issue, when this was the most controversial design decision about the project. To me, it's completely unsurprising.
It is possible to write a blazing fast JSON serializer/deserializer on pretty much every programming language, and most languages have many of them.
Swift does not, and it does not seem that it will have one in the near future. This is quite disappointing, taking into account that Swift is one of the 3 major-platform native languages (C# windows, swift macos, C on Linux).
So yeah, it is quite surprising that the major platform language in one of the main platforms in use has extremely poor JSON support. Even more surprising is that there is no path forward to fix that.
There are many pros/cons of using json, but this was a "non issue" (if it isn't fast enough, we can make it fast - turns out we cannot because the platform is controlled by Apple and they don't want to).
What's a rationale for having syntax highlighting server-side as opposed to client-side? I'm working on side project that uses ProseMirror and CRDTs through yjs, and the idea of having a server for syntax highlighting for editable text never occurred to me.
> Even so, xray was a bit of a wake-up call for me. It was evidence that the vision I had set out for xi was not quite compelling enough that people would want to join forces.
I think in hindsight this was a virtue. If you look at the xray project, it died after the Microsoft acquisition of Github: https://github.com/atom-archive/xray/issues/177. By not combining projects, it allowed for redundancy.
Perhaps an even better motivation is indentation. When I use editors with regex-based indentation (the norm), I'm regularly annoyed when it gets it wrong. Being able to budget a few milliseconds or tends of milliseconds to do indentation that exactly matches what rustfmt or gofmt would recommend would be worth it, imho.
as a side note, i think more and more people are starting to realize that the async paradigm is not a panacea. that paradigm or at least the way we implement it might even mean more trouble than a synchronous counterpart.
also, other editors that i feel will be sunset or not worked on at some point are atom and brackets.
This is why libvim (from the oni2 project) is based on vim rather than neovim. Even aside from performance, it is a huge simplification if you can interact with an editor engine in the same process, synchronously. At some point we replaced simple function calls with baroque APIs accessed over localhost...
It's mostly due to their build system.
> This was a smaller consideration - but more fundamentally, the model Neovim uses for input - queuing it on an event loop and handling it asynchronously - is at odds with what we required - to be able to process the input -> handle updates synchronously.
(And it and that something better is now both owned/maintained by the same organization, making it even more of a no-brainer to slowly phase the inferior option out.)
Why not put the input into the window, put a version number with the buffers and keyboard input and use shared memory? Of course something isn't going to work if you don't confront the problems you have at a fundamental level.
What are your thoughts on this topic @raphlinus ?
If you have any asynchronous background task it means the state of your application can change without the current thread having made any change. And your code needs to account for that possibility, which makes it a lot more complicated.
I couldn't find this point discussed elsewhere in the comments. I think this is one of the most important sentences in the whole article. I know this is essentially just another perspective on Conway's law, but I think it is important to identify factors like this early in a project if at all possible so that human factors around development can be factored out of engineering decisions.
I think that this is especially critical for open projects that want to achieve long term sustainability. Open communities don't have the resources to maintain all the packages, even if they use the same build system and follow the same pattern, even if most of the maintenance can be automated, some human being still has to deal each of those systems separately because they are part of the environment surrounding that individual part of the project.
The tree-sitter framework provides a pretty good engine for syntax highlight:
Once upon a time (a decade ago) it was OK for apps to directly draw to the front window, compositing wasn't a super established thing. This saves on precious memory bandwidth, which CPUs didn't really get all that much more of over the last decade.
However now that GPU composition has to happen, GPUs actually don't super like the linear formats that CPUs write to. They want swizzled textures, and they keep those formats private. And on not-heterogeneous memory systems, the buffer the CPU writes into needs to be sent over to the GPU as well.
Meanwhile screen resolutions & pixel counts skyrocket, as did UI visual & animation expectations. Being responsive is a lot easier than being fluid. Especially if you're trying to be fluid on a high-resolution and/or high-refresh rate display.
And this is all while ignoring things like Apple's "high DPI" handling, which is to just say fuck it and downscale instead. Which means you're pushing resolutions far higher than the display's actual resolution quite commonly.
For complex reasons, that's all been changing. For example, Metal on macOS doesn't even have a way to specify partial screen updates. Those optimizations are still valid, though, and in my ideal world we have both good support for partial invalidation and fast GPU rendering. Among other things, that would be really good for power usage. But to get there requires some attention to detail that seems to be mostly gone from the desktop UI space.
Even sending the buffer to the gpu for compositing makes sense as a problem, but still feels that should be faster than you would care about in a text editor.
I'm also on a mac that, if I do something that is "gpu accelerated", I'm likely to get a frozen session. Such that most applications don't seem to need gpu help. Are they just using different parts of it?
Linear is just your normal buffer where you index into a pixel at '(y * stride) + x', where stride is probably just the width * bytes per pixel.
But this isn't how GPUs store texture data. They swizzle it so that locality can be maintained well enough regardless of how the texture is rotated. httsp://fgiesen.wordpress.com/2011/01/17/texture-tiling-and-swizzling/ is a decent introduction. https://en.wikipedia.org/wiki/Z-order_curve has more of the general side of things.
There's then also framebuffer compression in addition to all this.
A good intro to layout transitions is: https://www.gamasutra.com/blogs/EgorYusov/20181211/332596/Ta...
The number of copies required to get a pixel from CPU space into photons today is ridiculous. It used to be you'd just write into a memory-mapped buffer, and your graphics card would scan out directly from that where it would get to the electron beam modulator in microseconds.
Such that I agree it would be better to do dirty tracking and sending just a small update to the screen. But the tricks to do that should be a lot easier than they used to be.
The gpu thing in my Mac is amusing just because if I run intellij, it will cause my machine to crash. If I run emacs? Not so much. Even if I am stressing the machine with several compiles or some silly pandas data frames. If I enable gpu accelerations in my browser? Expect instability. My video chat program just boasted that they use gpu for video. And for the past two weeks, it is common for the entire video system of my machine to hang during a video chat... Literally get a crash screen.
Not really, no. Over the last 15 years CPU memory bandwidth has increased around ~10x. Meanwhile monitor resolutions have also increased around ~10x. Meaning per-pixel CPU bandwidth has been flat over the years or even regressing a little bit. Laptops have become more common yet they also lag on memory speeds, or sometimes are even just single channel. Yet they also tend to have the highest resolution displays, not a great mix if you're trying to make CPU rendering still be viable.
As for your Mac issues I think you just have a broken computer. Software shouldn't be designed with the expectation that it'll be used on broken systems, that's not a realistic design constraint.
Define Hella fast.
If you want 16ms text rendering (~60 Hz) you pretty much need a GPU today. If you want 8ms (120Hz) text rendering, there is just no CPU that can handle it today with an OS running next to your text editor doing other stuff.
The whole point of Xi was being able to edit huge files (Tb size) on 8k displays at over 120 Hz, and being able to resize the screen and scroll, resize text, with great fonts, without any delay or sluggishness.
Xi delivered in some of those things, and druid delivers on some of the others.
But essentially, there is no system today where you press a key and that key appears on your screen in less than 8ms. Xi achieves ~16ms on the right platforms - not the fastest but much faster than most text editors.
The idea behind Xi was to preserve that as the editor got more features, like syntax highlightning, and it did.
I get that rendering a full display from scratch could be slow. But updating a single key press should not be. Scrolling? Sure. I guess. But even that feels like we over complicated something in the process.
I note that I do not disbelieve you. My asserting that it is ridiculous is that it sounds ridiculous.
> I have a multi core machine running at speeds well above my comprehension and you are saying they can't render simple 2d primitives at speed? That just blows my mind. What are they spending the time on?
Moving memory from RAM to the CPU, doing stuff in the CPU, moving the results back to RAM, moving those results from RAM to the GPU, etc.
The CPU is fast, but moving memory around is not, so the CPU and the GPU just end up doing nothing most of the time, waiting for memory most of the time.
Raph's blog (linked at the top), has a bunch of articles about rendering latency. But just google about typing latency, there are a couple of projects and tools that measure the time between a physical keystroke is pressed, and the letter being rendered on the screen (end-to-end). Beyond the memory copies, what usually happens is also: an interrupt is triggered on the CPU, the kernel might take some time to context switch to catch it, context switching requires saving all registers, and restoring them, and registers have exploded in size over time (e.g. with AVX-512 you need to save quite a bit of memory), then the kernel notifies the application, which needs to do something with the key press, like scheduling a render into a frame buffer, etc.
So there are just quite a bit of bounces from here to there in the system, most of which deal with memory latencies, and memory latencies is the part of the system that hasn't been getting much faster over the last 30 years.
How does one measure this? Say for the text editor I'm using (Emacs) or for Chrome/Google Docs?
Resolution could also matter, but we did have higher resolution displays back in the day, as well. Feels like 2d acceleration somehow regressed.
// Mac only. Valid values are true, false, and "auto". Auto will enable
// the setting when running on a screen 2560 pixels or wider (i.e., a
// Retina display). When this setting is enabled, OpenGL is used to
// accelerate drawing. Sublime Text must be restarted for changes to take
I'd looked at Xi about a year ago when I was finally running out of patience with Atom's terrible performance, but the lack of a workable Windows solution pushed me to VSCode in the end.
The same thing I experienced. Text rendering performance is terrible with Cocoa, and even worse with Win32. Calculating text width is an order of magnitude slower than with FreeType.
That's why for my UI framework I also went from using "native" APIs to custom GPU rendering.
I was always wondering if CRDT would really achieve encapsulation, but this kind of things are very hard to know without trying. Thank you for you works!
The real problem with the idea is that the user expectations of editor performance are high and the computer can only just barely do it. People expect to be able to open a 5GB file, have it drawn really nicely on a display with 30 million pixels, insert 1 character at the beginning of the file while the syntax highlighting updates instantly and then save and exit instantly. It's actually a lot of work for the machine, no room for RPC overhead.
The syntect library mentioned uses fancy-regex, which is explicitly different than Rust's default regex crate, in that it supports regex constructs that require exponential backtracking:
In particular, it uses backtracking to implement "fancy" features such as look-around and backtracking, which are not supported in purely NFA-based implementations (exemplified by RE2, and implemented in Rust in the regex crate).
I think "regexes" have gotten bad name so I call them "regular languages" now. That is, RE2, rust/regex, and libc are "regular language" engines while Perl/Python/PCRE/Java/etc. are "regex" engines.
In my experience, lexing with regular languages can beat hand written code. In the case of Oil vs. zsh, it's beating hand-written C code by an order of magnitude (9-10x).
First, TextMate / Sublime style syntax highlighting is not really all that great. It is quite slow, largely because it grinds through a lot of regular expressions with captures
pedantic: I think the issue is probably more lookahead than captures ? You can do captures quickly in a "regular language" engine.
It may be surprising just how much slower regex-based highlighting is than fast parsers. The library that xi uses, syntect, is probably the fastest open source implementation in existence (the one in Sublime is faster but not open source). Even so, it is approximately 2500 times slower for parsing Markdown than pulldown-cmark.
The main reason is that the parsing model is applying a whole bunch of unanchored regexes to the unparsed remainder of the line one after another until one matches, then starting again for the next token. This means each token parsed can require dozens of regex matches over the same characters. I implemented a bunch of caching schemes to cut down on this number but it still tends to need many matches per token. It sounds like Oil's lexer probably does about one regex run per token, with probably a somewhat faster regex engine, and sure enough it's something like 40x faster than syntect.
Oniguruma is actually pretty fast and anyhow most of the regexes in Sublime syntax definitions are written to not require backtracking, because Sublime has two regex engines and only uses the fast one when no backtracking is required. In fact fancy-regex should delegate directly to Rust's regex crate on these regexes but is somewhat slower than Oniguruma, for reasons I haven't yet looked into (edit: see Raph's comment, it's probably heavy use of captures, another thing Oil's lexer doesn't need).
Also note that byte-serial table-driven lexers have a speed limit of one byte per l1 cache round trip (~5 cycles), whereas faster lexers can take advantage of multi-byte instructions (even SIMD) and out of order execution to go much faster, hence why pulldown-cmark is 2500x faster than syntect rather than just 40.
Hm but why can't you just OR them together? That is perfectly fine with a regular language engine. For example, I OR together about 50 different regexes here (including a whole bunch of constant strings) for the ShCommand mode:
Despite this big mess, everything "just works", i.e. the lexer reads every byte of input just once. re2c picks the alternative with longest match, and it picks the first match to break ties.
I suspect the reason that you can't do this is (1) Sublime was originally implemented with a Perl-style regex engine and (2) the order of | clauses matters more when using a backtracking engine. It doesn't have the simple rule that an automata-based engine has.
My claim is that you avoid performance problems by using a "regular language" engine. So I think what you point out supports this, even it might be a slightly different issue than "more backtracking".
I think you are saying that Sublime's parsing model prevents composition by |, which makes lexing slow, because it forces you to read each byte many times. (in addition to the captures issue, let me think about that)
ETA: Heh, I'm amused to find the latter link to be another point in what seems to be an extended conversation between Andy Chu and the Rust text-processing community :)
If you scroll way down you will see a benchmark I did for the "constant string" problem.
So you can see that re2c does scale better from 1000-6000 fixed strings than either RE2 or rust/regex. But I uncovered a whole bunch of other problems, like re2c segfaulting, the output being slow to compile, egrep blowing up, the non-regular heuristics of "grep" playing a role, etc.
# grep is faster than both fgrep and the "optimal" DFA in native code
# (generated by re2c). I think grep is benefitting from SKIPPING bytes.
# All times are 'user' time, which is most of the 'real' time.
# re2c compile | re2c code size | re2c match time | ripgrep time | RE2
# n= 100 7 ms 11 KiB 1,566 ms 687 ms 1,398 ms
# n=1000 66 ms 57 KiB 2,311 ms 1,803 ms 1,874 ms
# n=2000 120 ms 93 KiB 2,499 ms 3,591 ms 2,681 ms
# n=3000 204 ms 125 KiB 2,574 ms 5,801 ms 3,471 ms
# n=4000 266 ms 159 KiB 2,563 ms 8,083 ms 4,323 ms
# n=5000 363 ms 186 KiB 2,638 ms 10,431 ms 5,294 ms
# n=6000 366 ms 213 KiB 2,659 ms 13,182 ms 6,397 ms
# n=47,000 2,814 ms
# - egrep blows up around 400 strings!
# - RE2 says "DFA out of memory" at 2000 strings, because it exhausts its 8 MB
# budget. We simply bump it up.
# - at 48,000 words, re2c segfaults!
# - At 10,000 words, GCC takes 36 seconds to compile re2c's output! It's 74K
# lines in 1.2 MB of source.
As mentioned, I think you would uncover similarly interesting things by benchmarking Sublime-like workloads with re2c's capture algorithm. They use some fundamentally different automata-based implementation techniques.
Here are hundreds of regexes OR'd together so the lexer reads the input exactly once, not 100 times:
And in the lobste.rs thread linked below, I was basically saying for all practical purposes you can ignore Aho-Corasick and use the more general regex version. Since the "fgrep problem" (fixed strings) problem doesn't involve captures, you should get a DFA that runs at the same speed either way. (I don't recall if the compile time was longer but I don't think so, it is buried in the thread probably :) )
Second, the design of fancy-regex is precisely to delegate to the regex crate when features like lookahead aren't needed. I had high hopes for the performance, but it turned out to be lackluster, a highly tuned backtracking engine beats it in most cases. And much of the problem is indeed that RE2-style approaches are a lot faster when captures aren't required. And, by nature, Sublime-style syntax definitions rely extremely heavily on captures. A bit of discussion (including comments from burntsushi) here:
Third, the question of whether "regular languages" would support faster parsing is somewhat academic, though perhaps you could build something. Though the languages being parsed often have some regular features, as a rule there are exceptions. Markdown is an extreme example, and cannot be parsed well with either regular languages or Perl-style regexes, though that hasn't stopped people from trying.
I don't have experience with captures in re2c, but it's entirely automata-based, and they apparently implement something "fast" for captures and published it in this 2017 paper, so it's pretty new:
I'm sort of interested in benchmarking some Sublime-ish workloads with it, but I'm not sure I'll have time.
And the other claim from that comment is that while regular languages aren't powerful enough by themselves for syntax highlighting, you can add a trivial state machine on top of them (perhaps Vim-like), and get something more powerful and faster than Sublime's model.
I'd have to look at the captures vs. lookahead issue more closely, but the general claim is that I think Sublime has a bad parsing model based on a sloppy application of Perl-style regexes (e.g. if you are really forced to read every byte multiple times, that seems dumb, and can be avoided with regular languages)
> On the plus side, there is a large and well-curated open source collection of syntax definitions
Whenever I tried to get Java highlighting updated in GitHub for “var” my issue is closed minutes later by people who don’t work at GitHub telling me to open an issue with some random unpaid TextMate / Sublime repo.
Really the overall problem is not owning the front end. It means you can’t do things like fix syntax highlighting, at least as important to programming as high performance word wrap.
I wish that after druid becomes production ready you'll pick it up again.
The coolest thing about Xi is that some have attempted to make it a vim replacement, others an emacs replacement, etc. I wish it becomes a framework that can be used to build better text editors in general.
It is always better to start with simple, dumb and reliable code, with minimal architecture, and then grow organically.
With experience, I tend to write the less smart code possible, I felt in love with the power of brute-forcing everything.
I'm always amused by the hn threads patting ourselves on the backs for in depth technical discussions, when so often the top comments are basically non-specific thought leader tweets.
For those that got this far, the article is well worth reading about a design space with difficult tradeoffs, and there are comments actually engaging with the content below this.
 today brings us multiple instances in https://news.ycombinator.com/item?id=23664067
I guess my point is he was trying to "correctly solve" a hard problem, which does require engineering. He didn't seem to set out to solve a dumb problem that didn't need solving (Yet Another Text Editor).
I love Emacs, and it continues to be my editor of choice, but it doesn't seem the design is amenable to adding good threading primitives. NeoVIM seems like a success story in this regard, though VIMScript is a much less pleasant extension language than Elisp, which is not perfect, but is at least a real programming language.
There may be a way forward for Emacs, building a sandbox that looks like a full Emacs instance to existing Elisp, and slowly factoring out the whole-editor blocking issues; but that might be almost as complicated as starting fresh, with fewer benefits.
No, but if that's the only criteria, we already have plenty of those.
I think the pagers - less - is somewhat ok with long lines? Not much else that deals gracefully with multi megabyte lines.
I've rarely seen a worse example of absurdly over-engineered code than the Android codebase...
For me, picking Rust and the JSON-based IPC were huge red flags and this post-mortem confirms that. But what I find odd is that many are still not willing to accept the conclusion that using Rust for a UI-intensive app is a bad idea. And using multiple languages in a project and splitting in back-end vs. front-end or lib + UI is also quite popular around here, unjustifiably in my opinion when one considers the extra complexity involved.
However, I fail to see how the choice of Rust is relevant in this particular project, as the UI part was meant to be programmed in any language, making Rust an irrelevant part of the equation.
On another note, I'm baffled why you are downvoted. I find your comment negative, but not unreasonable.
This postmortem says Rust was, and still is, a great fit for the problem domain for which it was used.
Where do you reach the conclusion that Rust is bad for UI-intensive apps? That claim appears to have no relation to the post-mortem as Xi didn't use Rust for the UI in the first place? Things might have been better if Rust was used for the UI if anything, as attempting to be native instead was one of the problem areas.
However if we want to remain positive, we could say like Edison that this project did not fail and is merely attempt 1 of 10000 at learning how not to build UI apps with Rust. :)
> Where do you reach the conclusion that Rust is bad for UI-intensive apps?
I've reached that conclusion by applying logic and looking at the state of Rust and its capabilities as a language and the scarcity of UI apps built in Rust.
Although I personally think Rust is fine. It's just that I thought the frontend and backend split is too complex. I am looking for something simpler than neovim, personally, and xi didn't tick that box for me.
Not mentioning Rust as a cause for failure at all is mildly disappointing but at the same time mostly expected given the lack of self-reflection in the community. I would have also liked to see some thoughts on how this debacle could have been avoided. What were the mental processes that made the author(s) ignore some pretty serious alarm signs about the design? How can these be better controlled?
> CRDT as a mechanism for concurrent modification.
i am genuinely wondering if concurrent modification is really, truly needed ?
Perhaps that changed, but you’d think that would take higher feature precedence over something like collaborative editing right?
Note how none of the above mentions "async", "ropes" or "crdt".
The goal was to make an editor based on sound technology principles, and that included investigating said technology principles. It turned out some of those principles were bad ideas, and that's that.
: My involvement with xi was limited to being part of conversations he had about rendering performance on Windows in the winapi crate's IRC channel.
Also, thanks for those early winapi discussions. The attitude of Rust towards winapi is one of the early reasons I started seeing Rust as being viable for actually building GUI.
I think what got people excited about Xi with Ropes and CRDT and Rust was that it was going to be a modern codebase with performance. Now that Moore's Law is buried next to Denard scaling, we need less software written with time-to-MVP in mind and more software written with performance in mind. Software is used a lot more than it's created. If I looked at editors like emacs and vi/vim/neovim, and tried to calculate the person-decades spent making them vs. the person-millennia spent using it, it's quite clear that optimizing for the single-core performance we have today pays off tomorrow when singe-core performance is 10% faster.
* Using native technologies
* Modular and extensible
* Open source
It is always amazing to discover niches where the conceptual and practical power of logging has not yet penetrated.