The author's entire section on "Real Reactivity" hit quite close to home.
I recently played around with using https://reactpixi.org/ to build a simple ant farm simulation (https://meomix.github.io/antfarm/) and was left disappointed regarding rendering performance, but not surprised. It's a real challenge to get strong numbers when stacking a declarative wrapper on an imperative base. I found myself fighting React's reconciler for performance almost immediately.
I was talking to my coworkers about the issue and one of them suggested trying to use react-three-fiber and just force it to render 2D, but, as the author notes, the problem feels intractable with competing layers of abstraction.
I'm really excited to learn about this library. I feel that I was between a rock and a place with my web-first, declarative graphics tooling. I was pushing myself to learn Rust, to use Bevvy, to have a well-supported, declarative framework, but I felt I would prototype quicker if I stuck with JavaScript. I considered A-Frame, but it's really not about 2D rendering at all with its VR-first approach.
There's really quite a desert of modern, active, declarative, web-first graphics rendering frameworks. Stoked for UseGPU to hit 1.0!
React acknowledges that nothing can ever be purely declarative, it allows you to declare a view while giving you an outlet for imperative computations and side-effects (hooks). There is no contradiction in that.
In Fiber to animate is to use useFrame, which happens to be outside of React and bears no extra cost or overhead, nor does it conflict with reactive prop updates. useFrame allows any individual component to tie its mutations into the render loop.
The reactivity debate is next to irrelevant in WebGL because that is a loop driven system, it is not event-driven like the DOM. If you doubt that observe Svelte/Threlte, which can update the view directly, yet relies on useFrame modeled after Fibers. That is because you need deltas (to be refresh-rate independent), flags (needsUpdate), imperative functions (updateProjectionMatrix etc), it is never just a prop update.
That said, i don't know much about React-Pixi/Konva and so on, and if these do not have a loop outlet then yes i agree. But the whole premise of that article just falls flat.
The Live reconciler[1] included in this project looks quite interesting as a reconciler for building a canvas-based image editing app.
Since WebGPU is still quite experimental and not usable without install canary browser builds, I'm really hoping that Live can be used separately. Like the author, I'm pretty sick of doing the data retained/declarative dance with libraries like pixijs many times now.
At least from this article, it seems that the author is using their @use-gpu/state[2] package independently. Would love to know if anyone might have any more details or experience?
> …certain properties have an invisible setter, which immediately triggers a "change" event when you assign a new value to it.
I’ve hit this with SwiftUI. Isn’t this a general problem of the entire “reactive” style of programming? If you have anything that requires a computation step after model changes, you need to gate when you take into account a set of model changes before you kick off your calculations or you end up with this problem of constant updates that kill performance. I don’t do a lot of front end work, but that’s been my experience with this method of programming in general.
You can have reactive programming with batch state updates. For the longest time, most React code used setState to update the entirety of a component's state at once rather than set each property individually and hope the framework batches up changes fast (or you use manual state update limitation logic like debounces). These days, I see a lot of React code having switched to functional components that suffer from rapid state updates, possibly without the authors even knowing.
In theory, these state updates shouldn't be a problem as your rendering logic and your state logic should be separate, at most rendering one or two frames wrong if the state updates are inconsistent. Video games have done this for over a decade and every "modern" GUI framework uses a GPU surface to paint its own controls rather than use existing system windowing, so you may as well take a page out of games' book.
Sadly, I don't think many frameworks do this stuff right and a lot of time is spend doing unnecessary state recalculations. It's certainly easier to write bug free GUI code through reactive programming, as state conflict bugs now get gracefully handled by the frameworks, but I think because these issues are now no longer clearly visible they also don't get recognised as bugs anymore.
I think this is just one of the many ways development has shifted towards developer comfort over user experience; changing a number in a region in memory and redrawing a string doesn't need to go through 50 chained method calls to update application state, but they make life for the devs easier so they're often tolerated. You get more features in return, but I'm not happy with the price we paid for those (mostly useless) features.
>redrawing a string doesn't need to go through 50 chained method calls to update application state
a) If you use memoization properly, and don't put all your state at the tree root, it doesn't.
b) In my experience, devs dramatically overestimate how useful such a hand-optimized fast path is, because in a real application, users expect the UI to always be consistent. It's not just about responding to a string change but also potentially the reason you're displaying a string in the first place.
And FWIW, Live is actually a lot leaner than React.
For Graphistry's GPU UI, we solve via react code for HTML chrome & interactions, and the visual analytics in webgl controlled by rxjs. Beyond rxjs being an overall coordination language, it lets you do thinks like custom schedulers, and so we are overall synced to requestanimationframe. And we still need to be careful on over updating, but at least we now have abstractions for reasoning about it and guiding it.
If I’m someone who just wants to get an app out the door, then I’d argue that not being able to gracefully bridge to other paradigms when there is a gap in what is available under a reactive pattern is by itself a problem (an ecosystem problem). That’s why I use React. Despite the issues with bridging to non-reactive abstractions, it remains a useful pattern.
Others may be more concerned with the applicability of the pattern itself.
I remember seeing it years ago, still amazed by it. What's even cooler is that it was created almost 10 years ago and I believe it hasn't changed much since then. You can read about it [here](https://acko.net/blog/zero-to-sixty-in-one-second/)
> What reactivity really does is take cache invalidation, said to be the hardest problem, and turn the problem itself into the solution. You never invalidate a cache without immediately refreshing it, and you make that the sole way to cause anything to happen at all. Crazy, and yet it works.
> When I tell people this, they often say "well, it might work well for your domain, but it couldn't possibly work for mine." And then I show them how to do it.
Signals-based reactivity is garbage collection on steroids. It does solve cache invalidation, and extends it to effects as well. It really should be a language feature at this point.
I wish the model could be extended to distributed systems, but adding network unreliability to the mix is non-trivial.
This post is great. Having helped move a large codebase over to React, I love React's emphasis of one-way data flow which I've seen greatly simplify many things and help avoid whole classes of bugs. I've been looking into graphics libraries (including three.js) lately, and the retained-mode style APIs in many that end up encouraging two-way data flow feel like a large step back. I've been wanting for graphics libraries that emphasize a React-style one-way data flow. I had been looking into react-three-fiber, but I've suspected that a wrapper like it doesn't take the fullest advantage of one-way data flow. This post addresses what I've been hoping for while being a good introduction to the issue for others.
>Three.js might seem like a great choice for the job: it has a 3D scene, editing controls and so on. But, my scene is not the source of truth, it's the output of a process. The actual source of truth being live-edited is another tree that sits before it. So I need to solve a two-way synchronization problem between both. This requires careful reasoning about state changes.
i dislike reactjs's way of building applications. it left a bad taste in my mouth and seeing it applied to 3D rendering makes me sad. but i do appreciate author for his previous work. especially the real time math in the browser stuff. wishing him luck with this new project.
Can't speak for the grandparent, but for me, my beef basically boils down to one thing. When you use a stateful hook, there has to be some concept of identity of the calling component. You don't want to get someone else's state back when you call useState(). However, I've never seen it documented what actual process is used to determine the identity of a calling component. In fact, it is something that I've had problem with before.
Related, I don't feel comfortable giving up control of when a real DOM element is replace by a new instance. For some things, like canvas, file inputs, audio tags, this replacement needs to be carefully controlled or prevented. This seems to violate the react philosophy.
This biggest issue I have with React is that I always feel like I have to re-write my model to use it. I've already got a model, now I want to add a UI. I try, and at least for me, React doesn't like it.
A simple example might be a tree structure. Think of folders and files on your computer. So first I go implemented a file system and I have folders that are a collection of files and folders. This data structure is hierarchical.
Now I want to make a React based UI where the user can create and delete folders and files.
5 folders deep the user want's to add a file. According to all the react docs I'm not allowed to just add a file to that folder as that would be mutating data. Instead I need to make a new folder, copy all the old entries to it, put the new file in. But, since it's 5 level deep subfolder, by the same rule, I need to do the same with its parent (make a new parent, copy all the old entries except the old folder, put in the new folder that's holding the new file). Oh but we're 5 levels deep so repeat all the way up the tree.
React people will often say "flatten your data" but that's the entire point, I shouldn't have to change my model to satisfy the UI library! I've already got working code without a UI. I just want to add a UI, not re-write all the other non-UI code.
I'm sure I just don't get it. Maybe a react expert will tell me how I use react without having to change my model.
Readit News
I recently played around with using https://reactpixi.org/ to build a simple ant farm simulation (https://meomix.github.io/antfarm/) and was left disappointed regarding rendering performance, but not surprised. It's a real challenge to get strong numbers when stacking a declarative wrapper on an imperative base. I found myself fighting React's reconciler for performance almost immediately.
I was talking to my coworkers about the issue and one of them suggested trying to use react-three-fiber and just force it to render 2D, but, as the author notes, the problem feels intractable with competing layers of abstraction.
I'm really excited to learn about this library. I feel that I was between a rock and a place with my web-first, declarative graphics tooling. I was pushing myself to learn Rust, to use Bevvy, to have a well-supported, declarative framework, but I felt I would prototype quicker if I stuck with JavaScript. I considered A-Frame, but it's really not about 2D rendering at all with its VR-first approach.
There's really quite a desert of modern, active, declarative, web-first graphics rendering frameworks. Stoked for UseGPU to hit 1.0!
In Fiber to animate is to use useFrame, which happens to be outside of React and bears no extra cost or overhead, nor does it conflict with reactive prop updates. useFrame allows any individual component to tie its mutations into the render loop.
The reactivity debate is next to irrelevant in WebGL because that is a loop driven system, it is not event-driven like the DOM. If you doubt that observe Svelte/Threlte, which can update the view directly, yet relies on useFrame modeled after Fibers. That is because you need deltas (to be refresh-rate independent), flags (needsUpdate), imperative functions (updateProjectionMatrix etc), it is never just a prop update.
That said, i don't know much about React-Pixi/Konva and so on, and if these do not have a loop outlet then yes i agree. But the whole premise of that article just falls flat.
Since WebGPU is still quite experimental and not usable without install canary browser builds, I'm really hoping that Live can be used separately. Like the author, I'm pretty sick of doing the data retained/declarative dance with libraries like pixijs many times now.
At least from this article, it seems that the author is using their @use-gpu/state[2] package independently. Would love to know if anyone might have any more details or experience?
[1] https://usegpu.live/docs/guides-live-vs-react
[2] https://usegpu.live/docs/reference-live-@use-gpu-state
I’ve hit this with SwiftUI. Isn’t this a general problem of the entire “reactive” style of programming? If you have anything that requires a computation step after model changes, you need to gate when you take into account a set of model changes before you kick off your calculations or you end up with this problem of constant updates that kill performance. I don’t do a lot of front end work, but that’s been my experience with this method of programming in general.
In theory, these state updates shouldn't be a problem as your rendering logic and your state logic should be separate, at most rendering one or two frames wrong if the state updates are inconsistent. Video games have done this for over a decade and every "modern" GUI framework uses a GPU surface to paint its own controls rather than use existing system windowing, so you may as well take a page out of games' book.
Sadly, I don't think many frameworks do this stuff right and a lot of time is spend doing unnecessary state recalculations. It's certainly easier to write bug free GUI code through reactive programming, as state conflict bugs now get gracefully handled by the frameworks, but I think because these issues are now no longer clearly visible they also don't get recognised as bugs anymore.
I think this is just one of the many ways development has shifted towards developer comfort over user experience; changing a number in a region in memory and redrawing a string doesn't need to go through 50 chained method calls to update application state, but they make life for the devs easier so they're often tolerated. You get more features in return, but I'm not happy with the price we paid for those (mostly useless) features.
a) If you use memoization properly, and don't put all your state at the tree root, it doesn't.
b) In my experience, devs dramatically overestimate how useful such a hand-optimized fast path is, because in a real application, users expect the UI to always be consistent. It's not just about responding to a string change but also potentially the reason you're displaying a string in the first place.
And FWIW, Live is actually a lot leaner than React.
You never need to preemptively invalidate a cache, you just reevaluate in tree order.
Others may be more concerned with the applicability of the pattern itself.
And you can scroll up to see it in full 3d!
> When I tell people this, they often say "well, it might work well for your domain, but it couldn't possibly work for mine." And then I show them how to do it.
Signals-based reactivity is garbage collection on steroids. It does solve cache invalidation, and extends it to effects as well. It really should be a language feature at this point.
I wish the model could be extended to distributed systems, but adding network unreliability to the mix is non-trivial.
>Three.js might seem like a great choice for the job: it has a 3D scene, editing controls and so on. But, my scene is not the source of truth, it's the output of a process. The actual source of truth being live-edited is another tree that sits before it. So I need to solve a two-way synchronization problem between both. This requires careful reasoning about state changes.
Related, I don't feel comfortable giving up control of when a real DOM element is replace by a new instance. For some things, like canvas, file inputs, audio tags, this replacement needs to be carefully controlled or prevented. This seems to violate the react philosophy.
A simple example might be a tree structure. Think of folders and files on your computer. So first I go implemented a file system and I have folders that are a collection of files and folders. This data structure is hierarchical.
Now I want to make a React based UI where the user can create and delete folders and files.
5 folders deep the user want's to add a file. According to all the react docs I'm not allowed to just add a file to that folder as that would be mutating data. Instead I need to make a new folder, copy all the old entries to it, put the new file in. But, since it's 5 level deep subfolder, by the same rule, I need to do the same with its parent (make a new parent, copy all the old entries except the old folder, put in the new folder that's holding the new file). Oh but we're 5 levels deep so repeat all the way up the tree.
React people will often say "flatten your data" but that's the entire point, I shouldn't have to change my model to satisfy the UI library! I've already got working code without a UI. I just want to add a UI, not re-write all the other non-UI code.
I'm sure I just don't get it. Maybe a react expert will tell me how I use react without having to change my model.