Beyond React 16 by Dan Ambramov – Async Rendering and React Suspense | Hacker News Reader

Beyond React 16 by Dan Ambramov – Async Rendering and React Suspense

youtube.com

129 pointstimdorr64 comments

So glad someone captured this talk!

At the 15 minute mark, he introduces a new paradigm for asynchronous IO within render functions; see a screenshot here: https://www.dropbox.com/s/wybqgtftuyqipk8/Screenshot%202018-.... This is groundbreaking, to say the least, as previously this needed to be done by adding higher-order components to inject the result of the fetch as a prop at some higher level; now, it's as simple as changing the src property in a single line of code.

One of Dan's colleagues, Andrew Clark, described how this is implemented on Twitter: https://twitter.com/acdlite/status/969171217356746752

    Here's how suspending works:
    - in the render method, read a value from the cache.
    - if the value is already cached, the render continues like normal
    - if the value is not already cached, the cache *throws a promise*
    - when the promise resolves, React retries where it left off

Digging into the source code released in this companion tweet, imageFetcher.read() indeed throws (instead of an exception) a Promise that is apparently caught by React's new error handling infrastructure and handled somewhere up the chain. https://github.com/facebook/react/blob/master/packages/simpl...

The talk indicated that there will be tons of support for granularly choosing when/how to show spinners, allowing interactivity with the rest of the interface, etc. while this loading is ongoing.

It's a bit magical, but promising to say the least!

The description of suspending sounds exactly like yielding generators and/or async await; why not build this feature on top of existing language constructs? I am quite sure the react core team has put more than the 30 seconds of thought I have into this question, so I'm curious what their reasoning is for using this implementation.

Seb explains the reason here https://github.com/facebook/react/issues/7942#issuecomment-2...

That's a great thread, thanks for the link!

This seems to me like a workaround for the fact that React requires `render` methods to return React elements. And I suppose it also takes advantage of the fact that the `read` method can `throw` its way out of your `render` function, obviating the need to check its result and manually `return`. But not gonna lie, I'm pretty suspicious of overreliance on creative use of `throw`ing as something that is going to cross through application code's control flow. But perhaps it will prove convenient enough to be worth it.

So I am guessing if a prop is `instanceof Promise` or some other constructor then it must trigger this response . . . is that correct?

React isn't changing any logic of the types of props; if you actually pass a Promise as a prop, it will be treated 100% as it was in previous React versions.

What's happening instead is that imageFetcher.read() either returns a string OR throws a promise; it never returns a Promise.

So <Component foo={fooFetcher.read(...)} /> either gets instantiated with the fetched Foo, or throws so React.createElement(Component, ...) never gets called in the first place.

Only if you physically throw a Promise (or "thenable") does the new logic apply: https://github.com/facebook/react/pull/12279/files#diff-1996... is called inside an exception handler.

Wow, I never would have considered throwing anything other than an Error. On the face of it, this seems like an incredibly powerful technique, but is it an abuse of a language feature?

Is a feature of the exceptions to let you handle exceptional cases far away from the caller, up in the stack.

You should keep a distinction between exceptions that means "error" and exceptions that means "some kind of event happened", and handle them accordingly.

Python “abuses” exceptions in a similar way every time you iterate over a generator: https://softwareengineering.stackexchange.com/questions/1124...

At the end of the day it’s just a control flow mechanism. And theoretically, if you lack data needed to render, it’s an exceptional condition.

It's a very interesting technique; I will definitely be keeping an eye out for appropriate use cases.

My guess is that it won't catch on in the long run. But, you never know.

The key phrase is "thenable", which is a term for anything where `typeof variable.then === "function"` is true.

But only when passed to HTML elements I hope. Because say I have my own type of the shape `{ then(): Whatever }` and try to pass that as a prop to one of my components I don't want React to invoke that function.

It's not about passing props - the "thenable" detection comes into play if you throw a promise (or something shaped like a promise) from within `render()`.

Wait, can render functions now throw, legally?

Yup. This builds on error boundaries and componentDidCatch: https://reactjs.org/blog/2017/07/26/error-handling-in-react-...

Yes. (For this use case.)

> throws a promise

> ...

> It's...promising to say the least!

^ this guy :finger_guns:

Is the cache global?

It's up to you. I intentionally kept moving parts to the minimum in my demo, but here's the real thing I was using behind the scenes: https://github.com/facebook/react/tree/master/packages/simpl.... So in that sense it's not global, you can create as many as you like and then isolate them (e.g. per subtree).

The `simple-cache-provider` is just a reference implementation though. It's somewhat naïve, e.g. it doesn't have invalidation strategy and just lets you replace the whole cache with a new one if you want to invalidate.

We'll publish more details on lower level API and how to write your own cache in the future.

It seems from https://github.com/facebook/react/blob/master/packages/simpl... that in the live version you'll be specifying which cache to use; this could be global, provided via React's old or new Context APIs, or passed down through props. And you could always write your own fetcher factory if you want to do something really custom; there's nothing special about the SimpleCacheProvider, and it doesn't use any React internals.

FWIW my demo code looked like

    export let cache;
    function initCache() {
      cache = createCache(initCache);
    }
    initCache();

    export function createFetcher(fetch) {
      const res = createResource(fetch);
      return {
        read(...args) {
          return res(cache, ...args);
        },
      };
    }

typo in the title - its Abramov not Ambramov

It's interesting that Lin Clark basically demoed Time Slicing a -year- ago at ReactConf: https://www.youtube.com/watch?v=ZCuYPiUIONs and in the first 10 mins of Dan's talk you see basically the same thing with user interaction applied.

createFetcher is brand new and here's the metaconversation on Twitter/Github: https://twitter.com/acdlite/status/969168681644179456 and https://github.com/facebook/react/pull/12279/files#diff-50d8...

with a couple implementations here (https://twitter.com/jamiebuilds/status/969169357094842368) and here (https://twitter.com/actualhypercrab/status/94131894239559680...)

// style discussion below

one thing I really appreciate about this talk is hard it is to even communicate asynchronous problems and the solutions to them. Personally the lagging animationFrame clock display (for time slicing) and the loading state devtools (for "react suspense") were actually the bigger "wow" moments. (Yes I know this talk is not about the tools or the charts)

he does another thing I'm a fan of - starting with a question. this helps frame the rest of the talk: explaining why the question is a big deal, breaking the question down, proposing some solutions, showing how the solutions answer the question.

bottom line - there's more to learn from this talk than just React.

Yep, nice link roundup - beat me to it :)

Also see the summary post on the React blog for the key takeaways from the demo:

https://reactjs.org/blog/2018/03/01/sneak-peek-beyond-react-...

Thanks for the link.

But boy, the Apple presentation style is so heavy in that blog post. "We call this...", "We can't wait..."

Sorry if the tone was a bit much! I figured it would be helpful to establish common terminology that people can refer to, and I'm serious about our excitement here – not exaggerating.

I express how excited I am for async rendering.

In my experience React has had 2 major paint points. async anything, and styling. And this fixes the former, and I'd argue that the latter isn't necessarily their job to fix.

Also, to have done this in a way that keeps a LOT of backwards compatibility is amazing.

What do you think is so painful about styling in React? Genuinely curious. The scoping of stuff? The organization?

There are basically 3 options, and they all have really big downsides:

* CSS modules - can use any CSS preprocessors, fast, can be extracted into CSS files, but it's very hard to "theme" and very hard to style/modify 3rd party components.

* Css-in-js - uses Js to render, using preprocessors is much more difficult, can't easily extract into CSS files, somewhat easier to theme and share.

* Inline styles - even slower, harder to preprocess, harder to extract, but very easy to theme and override.

And all 3 don't work well with one another. So if you get a widget made with inline styles, overriding the color is hard with CSS modules...

Isn't the whole point of CSS-in-JS that you don't need a toy language to do preprocessing if you have a real language like JavaScript?

I haven't really used SASS et. al., but my understanding is that they were hacks to get shared constants, functions, and derived values in a CSS that predated custom properties.

But with the downside of basically replicating a lot of what the CSS engine does in Javascript, further polluting the already stressed "main" thread in many JS applications. And doing it in a slower, less battery efficient, and less cacheable way.

And in my opinion, a language like SASS isn't a "hack" but is a very specific and extremely powerful language which is dedicated to one use (styling) and therefore most styling-specific code is very natural with SASS, however with JS it gets more difficult (want to extend another style defined elsewhere? You kind of need to work with JS that wasn't really designed to make that an easy task, in SASS it's `@extend %other-class`)

Thanks for your perspective!

Don't know if you've seen this, I've been toying with this approach: https://www.youtube.com/watch?v=bIK2NwoK9xk , and like the fact that you're writing css, even though it's technically css-in-js.

FYI: A higher quality version of this video is on the React blog at https://reactjs.org/blog/2018/03/01/sneak-peek-beyond-react-....

I certainly understand the bias to host it on your own infrastructure, but I really appreciate that I can download from YouTube for offline viewing on the metro. Couldn't find a way to do that with the official version.

That's really helpful to know for the future, thanks! We didn't have a strong preference so ended up with FB.

So if I understand it correctly, anything after this line inside the render...

    const movie = movieDeatailsFetch.read(props.id);
    ...

... will block rendering and will defer to a parent <Placeholder /> component to render instead?

Interesting magic I'd say, since I wouldn't expect such a line to block my rendering.

Wouldn't it be a more "React way" to have HOC component or a render prop that would work exactly like this .read() API?

   <MovieDetailsFetcher>
     {movie => <MovieDetails movie={movie} /> }
   ...

Then this HOC component doesn't need to have the IO-logic in the render, it can put it in the `componentWillMount` or inside another lifecycle method of React.

What's the advantage of putting the IO operation in the render instead of inside a lifecycle method?

Also, how's race-conditioning done? Does the fetcher library have the logic to only render the last fetcher, or is it more complicated than that?

There's lots of magic, especially in the IO part, and would've loved if he went more into detail on how it's actually done - it's a JS-conf afterall.

Also, a fetch()-promise-based API is not the only type of IO-interface we have. In fact it's only one of many.

HTTP2 for instance is on the horizon offering more complex interfaces with open connections possibly returning several results like websockets - on the other hand a Promise only returns one result and then dies. Observables might be a more interesting abstraction for IO-kind of things.

Also, testing this seems rather complex since the render is effectively doing a side-effect.

In any case, cool lateral thinking and I love hearing these type of talks about cool new futuristic stuff rather than some boring old thing everybody knows about.

>will block rendering and will defer to a parent <Placeholder /> component to render instead?

Yes, but only after the delay period passes. Before that, nothing will happen on the screen (at least not until I add an inline spinner as one of the last steps). Preventing a spinner from appearing and instantly disappearing on fast connections is exactly what this feature enables.

>Interesting magic I'd say, since I wouldn't expect such a line to block my rendering.

It works by throwing a Promise (and React catches it and waits for it before retrying rendering). So it’s not really magic, just an unfamiliar pattern. Conceptually this is similar to algebraic effects in languages that have them.

>Wouldn't it be a more "React way" to have HOC component or a render prop that would work exactly like this .read() API?

No. If you watch the demo closely you’ll notice we don’t show the spinner if the load is fast enough. React prevents updating the whole tree. This is impossible to express with a HOC without reimplementing the whole React in it.

>There's lots of magic, especially in the IO part, and would've loved if he went more into detail on how it's actually done - it's a JS-conf afterall.

As I said earlier the actual mechanism is quite straightforward. We throw and later React retries.

We’ll be posting an RFC with details. But for now we felt it was important to motivate this work, and even the motivation barely fit into 30 minutes.

>Observables might be a more interesting abstraction for IO-kind of things.

The core purpose of this feature is to suspend rendering until data is ready to avoid flicker and things jumping around. You can use observables in your code but I’d argue that, save for a few use cases, having components jump around as new items “come in” is a janky experience. You can always build a Promise interface that coalesces those observable updates into Promises of a final list before rendering.

>Also, testing this seems rather complex since the render is effectively doing a side-effect.

I actually think testing will get easier because now you can wrap your tree with a fake cache in tests. Then even components that do data fetching won’t attempt to reach the real network.

>Also, how's race-conditioning done? Does the fetcher library have the logic to only render the last fetcher, or is it more complicated than that?

The fetcher library doesn’t have much logic. It’s basically a cache backed by two Maps (for pending and resolved data).

React just retries rendering after Promise has resolved (or re-render happened due to another reason, like prop change). So there is no need for explicit logic handling race conditions. Instead of keeping continuations “in flight” we just abort rendering and retry when data is ready. So there’s nothing to race for.

Thank you! I appreciate all your replies, but want clarify this one:

> This is impossible to express with a HOC without reimplementing the whole React in it.

What is the problem here? We return `null` from the beginning, then spinner, if reached timeout and finally, the provided component, if data has arrived.

Please watch my demo. It doesn’t just unmount the children when I start fetching. It keeps the previous screen in place (and fully interactive) while the new screen is loaded. If the new screen is ready fast enough, then it fully replaces the old screen without any intermediate state (empty or loading).

I know it’s a bit hard to believe because none of JS libraries I’m aware of can do this yet. But that’s what it does thanks to cooperative scheduling and a double buffering-like system we implemented. I totally understand that it doesn’t quite “click” from the first explanation because you really need to try it to get it.

But to explain it again: React doesn’t remove the child tree, it suspends the whole update from committing before the tree is ready. So there’s no “holes” or “spinners” if it loads fast enough.

Intuitively my last example with Img component may help. There’s no a way a leaf Img component could delay the whole page from rendering before. With suspense it is possible.

Thanks for the answer. Since you're also the creator of Redux, how do you see this suspense feature work together with something like Redux?

I imagine that Redux can continue to be used as before, but instead of using middlewares (redux-thunk) for async things, we'd use suspense instead?

I think my only concern for throwing a Promise to do a non-local goto and cause the scheduler to delay rendering that component is whether the thrown Promise can be caught by mis-behaving code that makes it difficult to debug what's going on. It would be nice to have a more line-of-sight API as well vs only an exception based model.

Is there plan to implement the react-core in C or other native language ?

I can see that being compiled to webassembly for browsers and react-native can have multiple language support ?

No, there's no plans to rewrite the React core in C, Rust, Reason, WASM, or anything else right now.

Lin Clark, who works for Mozilla, gave a talk on "What WebAssembly Means for React" last year ( https://www.youtube.com/watch?v=3GHJ4cbxsVQ ). In a recent Twitter thread, she said that some more useful WASM pieces are in the works, but still not likely to be a big improvement or worth the effort to rewrite React ( https://twitter.com/linclark/status/967860936550879232 ).

Maybe Reason one day?

Then it could compile directly to native.

It would be really interesting to one day have React be a polyfill for an in-browser async DOM renderer.

Is that in (very, very, long-term) consideration?

I think my talk paints a relatively convincing picture why DOM nodes aren’t a sufficient primitive if you care about the features I described, as you’d need something like React to orchestrate those updates.

Hello Dan,

Great talk ! My thinking behind the top statement was with a native implementation of react-core, the react-native story becomes more compelling with JSX(ish) UI templates / CSS Styles and multi language business logic support.

Once you have this, the browser becomes another target. I see Flexbox implementation is already native with Yoga [0].

[0] https://yogalayout.com/

Oh sorry, I didn't do a good job explaining myself.

Currently:

    Application(JS) + React(JS) | Browser(C/C++/Rust)

Possible Future:

    Application(JS) + React(Polyfill - JS) | React(C/C++/Rust) + Browser(C/C++/Rust)

Something akin to the Browser building a "higher level" React-like API/scheduler for the DOM.

I saw somewhere that someone was working on something similar in Swift, but can't seem to find the link.

I remember a while ago someone shared a SSR C implementation of React.

I’m pretty excited to see how the api could interop with some new native features of js, like for ex async class functions are now possible and I bet a setup like this could remove a lot of IO overhead in the future

  class Test {
    async render() {}
  }

Kind of hard to come up with a concrete use-case with how little we know about the new features. Still super interesting though!

Interop with language feature is definitely a thing. It helps people to understand a tool without the need of digging into implementation details, since those details are language features that developers already familiar with.

There is a way to do this kind of things with current React version (16.2): https://github.com/alexeyraspopov/react-coroutine

You can use generators, async functions, and even more, async generators. Worth checking examples in the repo.

Nice, awesome work!

I did a little demo of working with the new Suspense API.

https://github.com/birkir/react-suspense-demo/blob/master/sr...

This is pretty amazing, kudos to the React team! Just hope that SSR isn't an afterthought with this feature. That is probably been my only little grouse with the React team that SSR compatibility is always a 'Not sure, someone should try it' response.

We aren't sure if we'll have time to build it ourselves (vs finding external contributors to build support and upstream it) but this design should handily address the long-standing requests to wait for async data during SSR. We thought about it.

Also curious how this would work with SSR? Will the render be considered incomplete until all subtree promises have resolved?

If so, that would solve one of the biggest pain points working with SSR + React today. Today, the solution is to have you route level component load data which is far from ideal.

I guess another question I have is how does the React or the Redux team see this playing with Redux. Dispatch returns promises but then the cache is the redux store itself - so would the fetcher read from the store possibly?

I'm a Redux maintainer (as is Tim Dorr, who submitted the post).

Use of promises with dispatching is, as far as I can see, not relevant to this particular discussion. It seems that the caching aspect is up to you - the important thing is that if a component decides it needs to wait for data before it resumes rendering, it needs to throw a promise.

What _is_ relevant is figuring out how to synchronize dispatched actions that cause updates to the Redux store with React possibly pausing a given component tree re-render cycle, applying a different update, and then resuming the rest of the original update. We've had some initial discussions, and are still trying to grasp what the impacts are here.

Ultimately, it's possible there will be internal changes to the React-Redux library - we just don't know what those changes will be yet. But, if there are, we'll do our best to keep the public API stable.

Maybe a dumb question or wrong forum. But how do you start using latest React in a project?

None of the demoed capabilities are released yet.

The current version of React is 16.2 (see release notes: https://reactjs.org/blog/2017/11/28/react-v16.2.0-fragment-s... ). 16.3 is getting closer to release and will include several new aspects, such as revisions to the lifecycle methods and a simpler `createRef` API. The async stuff that was demoed today is likely to be in a 16.4 release later this year.

Npm has alpha versions react@16.4.0-alpha.0911da3 and react-dom@16.4.0-alpha.0911da3 you can experiment with.

there be bugs

Reply on news.ycombinator.com