Hacker News new | ask | show | jobs
by rimunroe 1184 days ago
This is a hook version of the same code as near as I could guess it.

  function useViewport(initialValue) {
    const [state, setState] = useState(initialValue);
    useEffect(() => {
      return watchViewport((x, y) => setState(x + y))
    }, [])
    return state.get();
  }
  
  // usage
  
  function MyComponent() {
    const viewportSize = useViewport()
    // use in render
  }
I made a couple of assumptions here. From usage, I assume watchViewport is supposed to both subscribe and return a teardown function. I also assume that the viewportState.set/get are functions for getting and setting the tracked value in the component's state.

In my opinion, there are many advantages to the hook version and several major disadvantages to the class version:

1. There's no need for hooks to have a value method or React.Components to grow the use or addState methods because the hook is just a function call which returns a value. How it produces that value is up to the code inside the function--which does call out to React--but the fact the function will always be called when MyComponent is called (absent something throwing earlier in the function body of course). The value you see being returned by the hook will always match the value you get when you call useViewport() in your component. These two things are guaranteed to have the exact same behavior as any other JavaScript function call and thus you can reason about the "registration" and value passing without needing to learn any framework-specific APIs.

2. There's no need for an addOnUnmount method on the component object passed to the ViewportHook constructor, because the hook function can use the function component's equivalent to componentWillUnmount (the return value of a useEffect) in the exact same way as a function component can, but without need for external registration.

3. In order to implement the class API, you'd need to either (A) pass the component's actual instance to the hook, or (B) create a new type of value to represent the instance of the component the hook is registering against. Option (B) is yet another API to learn, as you have to learn a new type of object to deal with in a React application. Option (A) would mean figuring out a way to prevent people from calling those methods after construction, OR introducing the possibility of registering a new slice of state partway through. The latter might be possible, and maybe that's even what you intend, but I'd want to know what the expected impact on methods like shouldComponentUpdate or getDerivedStateFromProps would be. Speaking of those...

4. I can't think of any obvious way you could pass previous versions of hook-related instance properties to lifecycle methods in the same way that you can pass prevProps and prevState.

5. Concision: the hook version has a dramatic reduction in the amount of code you have to read

6. Bundle size: because the hook version relies on functions rather than class properties, it can be minified trivially and thus reduce bundle size even more than the obvious reduction in character count would imply

1 comments

The class version is

  class ViewportHook {
    constructor(component) {
      this.viewportState = component.addState(this, initialValue)
      component.addEffect(() => {
        return watchViewport((x, y) => this.viewportState.set({something: x + y})
      })
    }

    value() { return this.viewportState.get() }
  }

The hook version is

  function useViewport() {
    const [state, setState] = useState(initialValue);
    useEffect(() => {
      return watchViewport((x, y) => setState(x + y))
    }, [])
    return state;
  }
I changed the API to be more similar to the existing useEffect where you can return the unsubscribe function. That makes the difference in the size of code non-significant.

Again, its more about API design rather than classes or functions.

Another crucial benefit of the class version: you can use hooks in conditions or loops, in any order. This is because they're not called on every render. I also used a unique `key` to pass to `addState`, but its not really a requirement, since the hook constructors only get called once.

I don't understand points 3 and 4, can you elaborate? I'm assuming that when i call `component.use(HookClass)` the component creates a new instance of that hookclass. Regarding learning, I don't think the whole concept of (functional) hooks and their idiosyncracies are any easier to learn than learning one new type of class.

> Bundle size: because the hook version relies on functions rather than class properties, it can be minified trivially and thus reduce bundle size even more than the obvious reduction in character count would imply

I don't think this will make any meaningful difference. The full component methods API (addState, addEffect) is likely to be in use in any nontrivial project, so that can't be minified away. For 3rd party hook classes, they would be small and independend through `use` and easy to minify / dead-code-eliminate if not in use.

> I don't understand points 3 and 4, can you elaborate? I'm assuming that when i call `component.use(HookClass)` the component creates a new instance of that hookclass.

Sure! You show ViewportHook's constructor receiving a "component" prop. Since, in the calling component, you call this.use(ViewportHook), the `use` method is supposed to pass some sort of reference of the calling component into ViewportHook's constructor. My question was about what the type of that parameter is. Internally, is `use` something like `use(Hook) { return new Hook(this); }`? If so, you're passing a direct reference to the class instance. I had thought that maybe the framework could pass a more limited delegate for the instance to the constructor to prevent people from doing something silly with it, but that would prevent you from assigning a property safely anyway.

> I don't think this will make any meaningful difference. The full component methods API (addState, addEffect) is likely to be in use in any nontrivial project, so that can't be minified away. For 3rd party hook classes, they would be small and independend through `use` and easy to minify / dead-code-eliminate if not in use.

Object properties can't be safely minified. Any user-defined component using lifecycles will still need to have "constructor", "render", etc in the generated source, whereas identifiers (like "MyComponent" in `const MyComponent = ...`) aren't accessed by being looked up on an object, and thus can safely be minified to one or two character names. This was one of the motivations of the design of hooks. I believe it's called out in the original talk from React Conf 2018.

[Edit]

And regarding point 4: componentDidUpdate receives prevProps and prevState, and shouldComponentUpdate receives nextProps and nextState. How would you provide information about the previous or next version of that hook-based state if it's being tracked as an instance property rather than in the component's state object?

> I had thought that maybe the framework could pass a more limited delegate for the instance to the constructor to prevent people from doing something silly with it, but that would prevent you from assigning a property safely anyway.

That seems like a good idea. I'll admit my illustration is not a fully fleshed out design, it was meant more to be a sketch of how it would be possible to make a class based design much more capable than the original one was.

> Any user-defined component using lifecycles will still need to have "constructor", "render", etc in the generated source

Ahh I got it - this is not about DCE but about minified names. Not sure why my mind went with dead code elimination.

My first gut feeling is that this wouldn't be as important for everyday users if they already have code splitting / DCE / gzip. I'll look up the video, would be interesting to see some numbers - I'm hoping thats part of the presentation.

> componentDidUpdate receives prevProps and prevState, and shouldComponentUpdate receives nextProps and nextState. How would you provide information about the previous or next version of that hook-based state if it's being tracked as an instance property rather than in the component's state object?

I'm giving up having a single component state with the new API - this class based hook API also allows you to have multiple states. So I'm going to focus on the props part.

In the hook constructor, you would be able to use a listener for componentUpdate:

  component.onUpdate((prevProps, props) => { run code });
You could also do something like

  component.watch(() => [otherHook.someValue(), this.someState.get()], () => {
    // callback
  })
to get render-time change detection (equivalent to `useEffect(fn, [otherHookValue, someStateValue])`).

Its all a bit wordier, no doubt, but I feel that most of the capabilities can be implemented.

I imagine most of them could be, but this is an API which would require a lot of additions to the existing React.Component API, and would end up duplicating a bunch of already-existing functionality for something which in the end would probably need to conform to the same rules hooks do, but with what certainly feel to me like clunkier ergonomics. You'd also still be left with the problems of class components like poor minifiability and unreliability for hot module replacement.

I have to go to bed, but if you'd like to see a much better explanation about why hooks were adopted, I highly recommend the [checks notes] 1,401 comment-long discussion[1] on the PR for adopting the hooks RFC back in 2018, as this sort of design was brought up frequently. Especially worthwhile is Sebastian Markbåge's ending summary about why the team was going with hooks[2].

[1] https://github.com/reactjs/rfcs/pull/68 [2] https://github.com/reactjs/rfcs/pull/68#issuecomment-4393148...

I was trying to discuss the merits or demerits of hooks overall - just wanted to point out that a class based design doesn't need to be significantly less powerful or HoC-level awkward to use.

(I can't resist commenting I suppose - I did follow that thread as much as time permitted back then, and I couldn't agree with the conclusion from the arguments presented. Specifically, I was unconvinced that dispatch performance and file size were that dramatically different. In my experience, classes are very efficient in modern engines and code-splitting means much more than minification, especially after gzip. Even if they are the right trade-offs for Facebook, its unclear whether they're the right trade-offs for the rest of React users and the community as a whole - from my experience it has errected a bit of a barrier to front end work for backend engineers because there is a steep and weird learning curve at the start)