I think the gist of it is that JavaScript doesn't have threads, or isolated state. So your code can theoretically do anything it wants, but the React concurrency stuff is riding on your code only using a specific set of APIs, and does things like calling your functions on cloned instances with different props. If you aren't writing your code "in the React way" it very quickly leads to very confusing inconsistent behaviors.
Maybe somebody else has the energy to summarize them. :-)
The concern was because the old approach was to change the behavior globally. The new approach is to make it opt-in when you use specific new features that rely on it.
At a guess, I'd say "Side effects in render functions". People might be doing things in render that aren't generally considered safe or wise, and forcing concurrent rendering on them would run that code multiple times immediately, instead of the expected "once per render".
The new version apparently doesn't do that, unless you specifically use one of the new features, which means that you have accepted the other caveats for that scenario.
I just wanna give a shoutout to Facebook for finding a balance between rapid, innovative development on a major library without constantly introducing breaking changes. On the backend I'm a Clojure dev, and this approach has long been a standard in clojure.core which has earned my loyalty for many years. Wish more projects could be like this.
Whilst the major changes are not technically breaking, they are culturally breaking; e.g. ask for help how to do something with classes, and you are likely to get someone telling you to "just use hooks instead". I've never encountered another library whose users are such opinionated assholes.
Who are these jerk users? I'm going to give them a noogie!
In all seriousness, I'm sorry you had such a bad experience. That's not what we want the React community to be. Next time, feel free to tag me or someone else on the core team and we'll tell them to cut it out.
I'm excited about concurrent mode. We're using it in production to help make a search function 'feel' faster, and it's very effective. Of course our focus is still making things actually faster, but concurrent mode goes a long way way in improving user experience which is essential.
One thing I need to wrap my head around is avoiding unnecessary work with useDeferredValue, for example. I've got something set up which will defer multiple values, but I know the user will only want the most recent value by the time something can render. As a result, once values arrive it'll kind of cycle through each rendered value rapidly before stopping at the last one. I guess I need to implement some sort of throttling and/or debouncing under the hood better - this probably isn't up to concurrent mode to solve for me.
I'm glad to see batching being focused on. I have code which implements batching to some degree, and it's always struck me as so awkward that I can't rely on React to make good decisions for me in this area. The abstraction required in a context provider or component to accomplish this is fairly ugly and really gets in the way of focusing on the component's core responsibility. I think this will be a great improvement.
Automatic batching appears to cause problems, because the screen doesn't update when it needs to.
Here's the scenario: User is holding down the down arrow key. On each key event you have to update the screen so that appears to scroll up. But React gets in the way by batching setStates, and this causes the screen updates to be jumpy. Now you have to find a workaround. This is an example of React doing too much.
>Note: React only batches updates when it’s generally safe to do. For example, React ensures that for each user-initiated event like a click or a keypress, the DOM is fully updated before the next event. This ensures, for example, that a form that disables on submit can’t be submitted twice.
If you see a problem like the one you're describing, please file an issue in the main React repo, and we'll have a look. I'm not sure how you could have experienced this behavior since we released the alpha an hour ago.
To clarify, I am not talking about the React 18 feature. I have this problem in React 17. On Pagedown key, I throttle the setState calls because calling it too fast causes React to batch the updates. I'll try to make an standalone repro and open an issue in the React repo.
We have a mechanism to handle this exact scenario. We classify certain user input events, like keypresses and clicks, as "discrete" events. Updates that are triggered from discrete event will always be processed sequentially. We don't batch multiple events. Although we do batch multiple setStates within the same event, which is different. "Automatic batching" refers to the updates themselves, not the events.
We call non-discrete events (like mousemove) "continuous", and for those we will batch updates across multiple events.
We've been testing this at Facebook for a while and we think our model is really solid, handling all the common cases and a bunch of edge cases, too. If you try it out and find something unexpected, please let us know and we'll fix it!
I agree the React implementation is very complex, but in a way that's by design: we add features to React when it allows us to move complexity out of our product code and into a framework.
Like, yes, implementing batching in React was complicated, but in exchange the developer doesn't have to think about it anymore. It just works!
We do intend to ship a ES module build of React in the near future. Because that's a breaking change that could require significant changes to existing apps (i.e. to update the imports), we'll likely wait until the next major release cycle.
However, note that tree-shaking probably won't help much in the case of React. React isn't a library of independently useful features; it's a cohesive framework where features work together. We also have a really small API surface area. So there's no much to "shake" apart.
Arguably the main `react` package would benefit slightly from tree shaking, but those exports are already really small. The `react` package doesn't contain that much implementation; it mostly just calls into internal React DOM to schedule updates.
However, there really isn't anything to tree-shake in React. The renderer implementation is effectively a single consolidated set of logic that handles all work for all types of components. So, there's no unused functions to get shaken out.
I use it 2014-style. Class-based components, componentWillMount, render, callbacks, and that's about it.
I love this workflow, and generally structure all my UIs this way regardless of platform - have a state struct and then fully render the UI as a function of the state. To me, that's the big hurrah for using React.
But I have no clue why React is at version 18 and why it keeps getting new features. I'm perfectly happy making quite complicated UIs with the pattern I know well. Should I be following along? Does it matter?
In this way react is choose your own adventure. There is no current signal that the way class components work will change and I don’t exactly expect that to happen any time soon if at all. Rather then new features are aimed at developers with more complex use cases and make heavy use of asynchronous data fetching in a way that lends itself to concurrency.
If this isn’t you I’m sure at least the general performance improvements will be nice otherwise I don’t see why you can’t just keep doing what you’re doing
As someone who was initially resistant to hooks, I actually find that they're a more expressive API to how I was already thinking about my applications. There's footguns for sure, just like there are with class components, but I found that rewriting my class components with hooks instead actually made the code cleaner and easier to reason about. That being said, it's possible the benefit simply came from the rewrite, but even with newer components, I feel like hooks are a better API.
I personally fell in love with hooks after initially disliking them, you should give it a try, maybe in a toy personal project, to see if you like it. It is great for separation of concerns.
You basically stick to writing pure functional components focused only on how to render, and try to abstract reusable logic into hooks, and expose only what you need
For example, you could make a useApi() hook and inside your component just do const { data, loading, error } = useApi('/endpoint'), so you can hide everything you don't need away from the rendering logic
They may not be pure, but they are still functional. The hooks follow similar laws to Monads. They aren't entirely monadic and it would have been nice if they were and used a more monadic combinator for bind [1], but they are a relatively pragmatic solution for a language without a strong type system to encode things like algebraic effects and monadic bindings in.
[1] Mostly useless aside: using a .then()able based plumbing would have opened up the possibility of using the async/await combinator language. The names async/await would give the wrong impression especially prior to the actual concurrency changes to eventually ship in React 18, but would have potentially been a more monadic fit.
Yes, I stated that. Those functions are impure. Purity is a lovely goal, but purity also isn't a defining characteristic of functional programming, it's a modern pursuit. I've met enough classic LISP practitioners that lived their whole programming lives in the most impure of functional programming, that I would never want to tell them to their face that what they did didn't count as functional programming because of all the leaky impurity in LISP (due to pragmatic considerations of the time).
It may be defining of pure functional programming. But not for "functional programming". Some arguments for why (and indeed some discussion for why not, too !) :
https://wiki.c2.com/?FunctionalProgramming
Ed: and I suppose if a function call allocates a stack frame it's no longer a pure function? Or are all functions of type crash-or-value?
Not necessarily. The IO Monad (Haskell et al.) is impure, a monad is just a functional abstraction that doesn't have anything to do with purity at all.
If you get really technical, any program is impure as running the code has the side effect of increasing the temperature outside the cpu from running the program. That's a side effect. And running a program is going to change the temperature in a different way each time
That's to say, at a certain point, all this is arbitrary.
Pure code is generally easier, but a program that's only pure code is pretty useless. You always need an escape hatch which is what some monads like IO or Futures provide. Monads and other functional paradigms allow us to write and interact with impure code in a pure way and mindset. That's exactly what hooks do as does the react mindset of, your view is a function of state.
Hooks easily let us separate that statefulness into their own abstraction - just as futures do for async calls, and IOs do for the file system. This lets us model our view as a function of state (and props which are just arguments to a function)
It lets us think like that despite it not being true.
You need state to do anything useful. A lot of functional programming revolves around the idea of isolating impure aspects of a language from the pure parts.
That's why we have Futures, IOs, Options, Eithers etc.
No one is saying they are pure, the types are an indicator that something impure is happening. They are abstractions that allow us to pretend they're pure and work with them regardless of the context of the type.
I can map over a Future, IO, or Option regardless of whether the Future resolved, the IO was successful, or the Option contains a value.
That's a functional mindset where the impurity is isolated to the data type and methods are exposed on the type that lets us ignore certain possibilities.
I don't care about the context when I map over one of these types, whether its a Future, Option, Array, IO etc. I can write my pure function and pass it off to map. Who cares if the Future really contains an error, the Option is None, or the Array is empty. Map handles all those cases for me and lets me interact with these impure aspects in a pure way.
The function I define and pass to map is pure and lets me interact with impure parts of the code in a pure way.
useState and useEffect should be thought of in the same way. Clearly it denotes something impure is happening, but the stateful value can be treated like a state machine that can be declaratively tested against a finite list of possible states.
In the case of a Promise, it's successful, pending, or error. For an Option its Some<T> or None. Etc etc.
If my hook returns 3 values, result, isPending, error (and if you really want, each value can be an option), I can easily handle each scenario deterministically.
That allows for pattern match like solutions (in JS it can be through cases clauses since there's no FP pattern matching).
This allows for writing components in a pure like approach even when its obviously impure.
In real-world FP programs there will be some impure code. But that impure code should be in thin outer layers. The core needs to be pure. Why? Because otherwise it defeats the purpose of FP. The purpose of FP is easy verification and ease of reasoning about code. That's only possible if most of the code is pure functions.
There is really no need shoehorn something that's not a good fit for FP into the FP paradigm. It is not like customers are going to pay more money if it is FP. So don't bother.
> You basically stick to writing pure functional components focused only on how to render
If you’re using hooks like useState or useEffect I wouldn’t consider those functional components pure.
Hooks provide some nice syntactic sugar, which is not nothing, and make it easier to share logic between components. In my experience hooks have made my code easier to read, but they haven’t simplified it. The side effects and state might not be inlined but they’re still there. If anything they’ve made it less explicit because some of the details are hidden behind the magic that the React runtime uses to add state to functions.
But the beauty of hooks is as you say reusable statefull behavior. It also lets you treat your functional components more like a pure function in my opinion.
If the hooks you use are poorly designed, then yeah, it may be more difficult. But well designed and implemented hooks make your component appear more like a pure function.
The state is (more) decoupled from the UI code than before and lets me easily reuse behaviors across components.
They're technically not pure, but they're as close as we can get to functional-style state handling: declare dependencies upfront so the caller knows exactly what they'll get when they execute your function. Fully state-less code achieves basically nothing; you need to "escape the hatch" eventually to work with the rest of the architecture. Hooks are very beautiful escape hatches that let you deal with persistence without having to sacrifice the rest of your function's body to it.
And if you really wanted to make it 'pure', you could easily write a hook with an explicit return type that wraps the type of the initial value. If you're using typescript, you can then limit the possible types of the generic to only monads.
type State<T> = val as T;
const useStatex = <T>(initialState: T) => {
const [state, setState] = useState(initialState);
return [state, updateState] as [State<T>, (val: T) => void];
}
This new hook, useStatex, has an explicit type definition that indicates something stateful is taking place. Same as Futures. You can take this even further in the typing limiting the generic to only allow certain types for the underlying values to be monads, or even more simply
type State<T> = Promise<T> | Array<T> | Option<T>
So you avoid an extra layer of nesting like
State<Future<Array<T>>>
Point is, there are lots of ways to easily resolve any its not pure! arguments same as any other functional programming concept
eta
The simplest way to limit the return type might even be to just do this
I find the function-based components much more elegant than class-based components. The semantics of Javascript classes are confusing, and the new notation is very concise. (Especially with an IDE, such that "refactor this pile of tags into a function" is a one-click process.)
The useState hooks are a tiny bit clearer than the setState semantics. The effect hooks are less clear, but all told, the new mechanisms for handle global state and side effects better without Redux. (Redux is neat, but after working with it for years, I just have to conclude that in most cases it's just too much mental overhead compared to a more idiomatically Javascript solution.)
There's nothing wrong with continuing to use class-based components, but I think you'll find that everybody else is going to gradually deprecate it. The key facts about the work flow -- state to render function to surprisingly efficient DOM reconciliation -- remains the same.
I am with you. React as originally designed made sense to me. You render the screen point-in-time and React takes care of updating the screen efficiently. This is all I need from React. Everything else is bloat.
Wasn't react first designed in standard ml, without classes? Then classes were bolted on in the js rewrite - and hooks are closer to the original design?
> So I continued to explore framework-izing these ideas, and I had implemented several iterations of what eventually became React, in a few languages - one of the first explorations began as a rough implementation of the reconciler and component model in Standard ML (CreateClass was a "module function"). This was really great because SML embraces immutability by default which is natural when building React style components. We naturally wanted to deploy UI to web browsers, and at the time, the compile-to-JS landscape was not as mature as it is now - I don't even think source maps existed yet. So it made sense to port that exploration to JavaScript (...)
> One thing I noticed that when people were creating point to point bindings in their more traditional "MVC" app structure, it would almost always end up requiring "computable" bindings which invoke a function anytime a mutable cell had received a "change event". All these "computable bindings" ended up chaining together and small changes would end up causing large recomputation of the majority of the UI. I realized that functions already do transform inputs into outputs, and if we could just find a way to reinvoke those functions repeatedly, and quickly enough, that we could be much more expressive and concise, at not much more performance cost that the chain reaction that "computed bindings" would result in anyways.
I think it's less about being pure or functional, but more about being "ergonomic". I abhored class-based components from day 1, but on hooks everything just clicked.
>".. I use it 2014-style ..."
are you using any type system eg TypeScript or Flow ?
I am asking because I do not remember them being used back then, so wondering if you had upgraded there.
Also what do you use for 'whole app' state handling (rather than component level) ?
Asking because React Context feature seems to be the right fit for the 'whole app' state handling, but it was not there back in 2014.
I also like you picked up preact for smaller project.. just because the size is so small. But for lager things (eg over 20K lines of JS) I seem to be staying with React.
Really looking forward to new server side components and suspense.
Also, kudos the the React / facebook team for building such an exceptional framework. React has really enabled small teams like ours to build large applications and even a framework! It takes a lot of dedication and careful planning to create such a versatile lib for a broad audience. 8 years down the road and the team is still keeping it together while innovating web-dev to new frontiers.
Interesting, I never did notice this bit in my previous readings of this interview:
> I had also put the foundation in place, to ensure that React components could be server rendered - and that feature was more so inspired by demand and constraints at the time than any specific technology.
"Jordan Walke
Transcript from Thursday January 26th, 2017
Facebook Engineer | Creator of React.js & Reason"
Two things about React will never stop being funny to me,
1. An entire generation of front end developers have staked their careers on a framework developed by Facebook. Facebook!
2. The hooks API. I honestly thought this was a joke when I first saw it. Developers will talk about the elegance of functional components then go and use the hooks API.
I know this is controversial, but I find them hard to understand. When I call `useState()`, what's the identity of the state that I'm using? What ties the state together between subsequent invocations of my component? For example, I know that strange things happen if you conditionally invoke `useState()`. People like to talk about "functional components", but hooks seem to live and breathe on mysterious side effects that I don't understand.
If only there was some way to discover how things worked... Some website where you could type “how react hooks work” and get dozens of articles and blog posts explaining it.
They are apparently this panacea of elegant stateless code, but the developer is expected to know that multiple calls to useState must be made in the same order every time. Typically we associate APIs that are fragile to out-of-order calls with tightly coupled stateful systems.
Not to mention the hideous use APIs. Thanks you, React, for this array which apparently contains a state and a setter?
I think your first point gets to a tradeoff that isn't much discussed about hooks- they can be elegant and simple, but only assuming quite a bit of upfront knowledge about how they work.
It's almost like taking the verbosity of class lifecycle methods, and 'pre compiling' it into a form which is terser, yes, but that now only a sufficiently informed reader can grok - the lifecycle is still there, it just has to be modelled by the reader instead of by the code.
Sometimes this is a good tradeoff, but not always.
The developer isn't just "expected to know", because they'll get an error if they don't. React's debugger is really good at pointing out mistakes.
As someone who started off thinking hooks were too "magic" to be taken seriously, such concerns are a thing of the past.
If you just think of calls to useState as variable declarations, it makes zero sense that you'd ever want to call useState in a different order.
this -
let userName = 'bob';
let favouriteColour = 'purple';
is this
let [ userName, setUserName ] = useState('bob)';
let [ favouriteColour , setFavouriteColour ] = useState('purple)';
Hooks enable you to reason about the state your component is in in a way that class components couldn't.
Hooks felt extremely awkward to me at first, because it was forcing me to write code in ways that felt unnatural. Once I stopped fighting it, the benefits became apparent, and now the old way feels awkward.
I've seen this a lot, when is an example of a time when you would want to call hooks out of order? I can't think of a single time where this would be a good thing to do
Sometimes you would like to return out of a series of hooks calls early or throw to an error boundary. That also leads to conditional calls.
Example: you have a {data, error} = fetchData() and then a useEffect(..., [data]) or 3. It would be very handy if you could just throw the error before the useEffect because you already know it's going to be irrelevant.
So then you put everything after in a new component (because that can be rendered conditionally) or keep checking for errors throughout the effects code. Both are kind of meh.
I'm not sure if this could maybe even work, actually, because there are no real out of order calls, only incomplete calls? But rules of hooks say no, and I have gotten the "hooks called out of order" error for it.
let [ data, setData ] = useState();
useEffect(() => {
let result = fetchData();
if (result.error) {
//do something, or abort?
return;
}
setData(result.data);
}, []);
useEffect(() => {
if (!data) return;
//perform secondary chained action using the fetched data
let result = fetchMoreData(data.guid);
}, [ data ]);
IMO there are far greater concerns with centralized API services, which are nearly always opaque, closed-source, and incentivized towards vendor lock-in (AWS being the elephant in the room).
"Staking careers" is also a little much; React isn't that hard to learn (it's near-trivial compared to .NET or iOS development), and any developer worth their salt will have to frequently retrain into new patterns and frameworks during their career.
There can be legitimate concerns with power concentrations of open-source software based on the incentives of its stewards (Chrome comes to mind); but ultimately React carries a relatively small footprint, and can always be forked [0] if it goes in a direction that developers don't like.
I'd love to see web components support greatly enhanced in React 18. It's literally the only major JS library used on the frontend which still offers poor WC support.
Does anyone know if the alpha builds will work with react native? I'm really hoping so because the flickering between transitions when using latest relay is super annoying to get around.
reply