React's UseRef Deep Dive
giovannibenussi.com
giovannibenussi.com
Mixing stateful effectful code inside what would otherwise be a pure declarative render function leads to so much complexity in an attempt to bridge the two paradigms. Junior engineers are also constantly tripped up by the subtleties of `useEffect` and `useState`.
Furthermore I think attempting to "encapsulate" side effects is a bad approach to writing testable programs. Most React components that use hooks end up having several chains of promises inside them but no way to await the final promise meaning tests have to be full of
await wait(0)
await wait(0)
await wait(0)
await wait(3)
Cycle.js is the only framework that seems to get this right by acknowledging that a component doesn't just output JSX, but actually outputs JSX, HTTP Requests, etc.
Look, I get that Redux was a pain in the ass with all the boilerplate, but how did we throw the baby (unidirectional data flow) out with the bathwater. I no longer advertise that I know frontend development because the entire react community seems to have lost its mind.
If many successful people voluntarily use something you cannot understand there is always the possibility that you are the one who are missing out on something.
(FWIW: I've been there myself)
I can see why some people prefer to use classes where the design is cleaner (even though you are more limited).
*: https://reactjs.org/docs/hooks-rules.html#:~:text=Only%20Cal....
Reading the react docs, I find this part interesting:
> Hooks were designed with static typing in mind. Because they’re functions, they are easier to type correctly than patterns like higher-order components. Importantly, custom Hooks give you the power to constrain React API if you’d like to type them more strictly in some way. React gives you the primitives, but you can combine them in different ways than what we provide out of the box.
(https://reactjs.org/docs/hooks-faq.html#do-hooks-work-with-s...)
Seems to me as if hooks try to fix the lack of a good programming language. These approaches usually go wrong eventually. Maybe it's time to ditch Javascript (and even Typescript) and use a proper language (purescript maybe?) so that all these things can be done with just functions.
You can probably build all this infrastructure yourself without hooks and only with plain functions. It'll involve some global store and some subscription mechanism. You'll end up with something really similar to Redux.
It sounds to me that such a concept would be very useful in a lot of contexts, so instead of specializing it only to react hooks, it might be better to just be a standalone library. E.g., why not use the same technique in Angular?
That way, developers don't have to learn framework-specific ways of doing things again and again but can reuse their knowledge. I think the fact that everyone is building their own thing is one of the reasons why many people have this notion of "need to learn a new framework every half a year". Backend technology seems to do better here, maybe because the pace is slower?
See: https://github.com/reduxjs/redux, among others.
The React ecosystem did exactly this for the past few years, but evidently it was decided that a more native approach like hooks were better.
Coming from the backend, I miss something like the effect system builtin in Haskell/Purescript or added as a library to languages like Scala (e.g. https://typelevel.org/cats-effect/) or Kotlin (e.g. https://arrow-kt.io/docs/effects/io/)
Personally, any time anybody has ever suggested using redux I have always fought it. I've never seen the news for a global state management library, since react has always had tools to manage state built in.
Hooks are functions. If that wasn't obvious, it's also explicit called out as a key thing about them in the excerpt from the docs that you present and then try to make a point with that entirely ignores the content of the excerpt.
Hooks are not _just_ functions, they are special functions where that (by the documentation) should not be called inside loops, conditions, or nested functions. These are severe restrictions compared to the "normal/pure" functions that I meant.
I find hooks/effects better than mobx or redux. hooks/effects, redux, mobx would be my preference order. Although redux works so well that it can be used with hooks/effects.
For global state I would just use a React.Context with it's own effects. Most components do not have any async effects.
I do admit that when jumping into a new codebase with hooks/effects I often find people have just butchered the concept. But I find that when I jump into codebases with mobx and redux too. The butchering that people do with mobx is much worse, with massive module importing of global stores, misuse of `name?: type` inside stores because they want to initialize the store at a later time.
You can write global reactions to state changes very easily.
Most components do not have any async effects.
I think this talk by Andre Staltz is the best explanation and solution of the problem https://www.youtube.com/watch?v=SXdtrhn8iII
mobx on the other hand is exactly that, the state is globally mutable, anyone can import it from anywhere to anywhere.
I can easily have all the user settings accessed inside a small effect and react to events or not depending on the settings and it's composable and reusable.
Testing with jest/enzyme is still the same. I'm writing selectors, simulating clicks and the component rerenders. Mocking the effects or functions that effects use is also simple. With jest 26 and modern fake timers it's easy to test async effects and changes of state too.
I guess the only flaw of hooks/effects is that they not efficient because of so many effects fetching the same state but they are efficient enough. For example, mobx would be extremely efficient in that regard (any store.value access is just that, not a useState call with new () => {} instantiated on every rerender.
They are, but they're also tripped up by the subtleties of class-based React components. The difference is that with hooks it's really obvious when they've tripped up (because React can tell, and warns you) while with classes it's really not obvious at all. This is one place where hooks are a clear win.
That's just the nature of being a junior engineer — you get tripped up. Tripped up by mutating objects, by stale closures, by comparing numbers to strings, by asynchrony, by thinking in terms of rxjs or xstream streams, by god knows what else.
That will also help seniors when they are tired or under time-pressure.
It seems like these are always drawn up by backend developers who lob fistfuls of aggression-poop over the fence for reasons beyond my fathoming, or junior developers (or UI folks who occasionally use javascript) that find it easier to trash on patterns rather than learn about them.
I have used React professionally for years and get the skepticism around hooks. It seemed really stupid to me before I learned about it. Why change what wasn't broken?
Now after learning about it I'll never go back to class components if it can be avoided and happily recommend hooks to all my clients who also end up loving them.
No idea what this promise chaining thing is that you're talking about. Sounds like a bad design pattern that has nothing to do with hooks. No idea why you're suggesting Redux has anything to do with hooks. They're completely different.
This meme is dead. Only a superficial understanding of React, or ingrained bad programming pattern habits, keep this meme alive.
> No idea what this promise chaining thing is that you're talking about.
If you have any asynchronous things going on in `useEffect`, you'll have to do something similar to that `await(0)` song and dance in tests. This specifically affects tests if you do things like update the UI by toggling loading spinners on await.
> Redux
s/Redux/higher order components. One of the motivations for hooks was that as a mechanism for logic composition, HOCs just felt awful to use. (So did render props, which everyone suddenly used for everything in a brief moment of collective insanity.)
> Only a superficial understanding of React
I think there's something in this. The fact is that good or bad, 1) hooks aren't intuitive, 2) hooks have basically doubled React's API surface area. Previously, React was so simple that a backend engineer could pick it up and get productive with it in half a week. That's much less the case these days. I've been onboarding devs to React for years, and these days there's a lot more "yeah, that's magic, you don't need to know how that works for now."
With hooks, I've literally seen almost completely novice programmers (e.g., people who’d hacked out some powershell while working desktop support previously) pick up React and be productive in a month or so, while also learning JS and Python and SQL and backend development from scratch in parallel; if backend “engineers” can't do it in half a week or so, they’re hacks.
You should have a core (pure) functional component, that doesn't perform network requests, and an outer component for side effects.
Your testing for the inner component becomes trivial: Do the correct props product the correct HTML? Fairly trivial to validate. Use Enzyme, react-testing-library, etc.
The outer component, which contains side effects, can be covered through an end-to-end testing tool like Cypress.
You should have a core (pure) functional component, that doesn't perform network requests, and an outer component for side effects.
IMO, this is the biggest thing we lost with hooks. Before there were two different types of component, and you couldn't make a pure functional component stateful or effectful without completely rewriting it.
I love hooks, but I'm constantly asking in code reviews "are you sure this is where you want this state to live?".
The other big problem with hooks is that there are no safeguards (and little documentation) about using them wrong. The post you're responding to mentions using `useEffect` to toggle a loading spinner. This might make sense if the loading spinner is outside the React component tree, but if you're using useEffect to toggle some piece of React state, yer doin it wrong. Once you start slipping these kinds of hacks into a codebase, it's a slippery slope.
Is there anybody that develops on the frontend professionally fulltime and makes these kinds of complaints?
Now after learning about it I'll never go back to class components.
IMO state and side effects belong outside of the render functions. Not mixed up inside them. This is exactly what redux, cycles.js, Elm, MVC, etc do.
No idea what this promise chaining thing is that you're talking about.
async () => {
await foo();
await bar();
}
() => foo().then(() => bar());
No idea why you're suggesting Redux has anything to do with hooks. They're completely different.
The hooks API not only suffer from weird wording (how is useEffect more expressive than componentDidMount) but it introduced some concepts like dependency declarations (again useEffect) and a new meta syntax.
It could have been a good development if the published preview would have been a real request for comments with the possibility of changing something. But instead it was introduced as stable in basically the same form.
Well, useEffect cross-cuts way more of the component lifecycle than just componentDidMount: it also handles componentDidUpdate and componentWillUnmount, and a few other things that didn't quite get lifecycle functions before. I suppose you could call it useWhenComponentMountsUpdatesOrWillUnmount, but "effect" as in "side effect" for when the component moves through lifecycle events isn't an uncommon name for that sort of thing (even outside of React).
Lately when working with all this I feel as if the same code would be simpler and easier to follow if it wasn't shoehorned into JS. It's like the uncanny valley of functional programming.
But more fundamentally, dealing with frontend application code is mentally exhausting in a way I haven't experienced before. I feel like I'm no longer learning Javascript as a language, but just keeping up with the novel abstraction of the month.
I should add that I don't hate it, and there's plenty still to appreciate. Some of it is a joy and it's refreshing to work on projects that embrace more functional styles over the typical CRUD and OOP taxonomy construction you get in a typical backend job. Different set of problems once you get away from the typical framework stuff (react, redux, saga boilerplate).
Any codebase, using any language, framework, feature, and technology, needs to be managed properly. There's really no getting around that. And without that, you're almost always going to end up with things turning into something along the lines of "complete untestable spaghetti". Especially when you have junior developers involved. Hooks are no different, nor is the entire Javascript landscape.
In general these complaints are getting tiring and often feel like they are coming from folks who don't have the appropriate level of experience or knowledge in a particular subject area to be making these generalizations. I don't know what your level of competence is, but this line:
> Look, I get that Redux was a pain in the ass with all the boilerplate, but how did we throw the baby (unidirectional data flow) out with the bathwater.
gives me pause, because Hooks aren't meant to replace Redux in any way. There are hooks that maintain state, sure, but generally they're meant to replace class component lifecycle functions. Hooks and Redux still coexist quite well.
React hooks definitely threw a wrinkle in the testability of react components. However, the ideas around how the FE should be tested has changed to the point where current thinking is that integration tests are the most valuable tests you can write. Testing what a react component does in isolation of redux or its side-effects is easy to do and also not incredibly useful.
Furthermore, many modern codebases still use redux. I still advocate for it. Using hooks and redux are not mutually exclusive.
I wish there was something better than hooks, it’s very easy to forget a dependency or get into infinite loops, but they do make life much easier. No more mapStateToProps or mapDispatchToProps HoCs. No more “smart” vs “dumb” components. No more function/class components. There’s only one component.
Also, we are seeing an increase in “headless” react libraries, where the logic of some functionality is disconnected from the visual design of a set of components. It makes composition, extendability, and maintainability much easier.
Also, why even bother giving an example of useCallback if all the example does is log something to the console? Especially immediately following a section called "Real world use cases". This is how newbs get the mistaken idea that you need shit like useCallback to show and hide an element based on state and you end up with overengineered code everywhere.
Examples of library functionality should be of something that you should actually use that functionality for (because it's impossible or impractical to write in other ways). Not just because it makes the docs better, but because it gives you a chance to run into a big red flag when you've implemented something and realise you can't actually find a use case for it.
Then again half the people writing docs for libraries in React land don't even use the libraries they're documenting.
I still vastly prefer JSX for frontend dev but I think really that’s about all I care for in this ecosystem.
useRef is a hook to serve just that purpose as is useState and useEffect.
Building new hooks off of these base hooks lets you write your own custom hooks to handle whatever side effects you need.
The existence of a hook is what should alert you to the fact that there may be some side effect thing happening here.
Not sure if this applies to you but I've found that most people who criticize hooks haven't made the leap to using custom hooks and have instead only used ones provided to them (useState, useEffect, etc).
Hooks are an incredible way of encapsulating behavior, separating out that behavior from a component, and exposing that behavior to any component that needs the same behavior. Check out this video for a practical example: https://www.youtube.com/watch?v=nUzLlHFVXx0
I'm in the middle of a big refactor and writing reusable custom hooks are an absolute savior to share functionality across components that may have nothing to do with each other.
Hooks are functions, which is why we can compose them to build new behavior from smaller ones. That's the point. Just cause a thing is a function doesn't mean its pure. Like I said above, hooks are often meant to contain the impure parts of your otherwise declartive and pure UI expressing the UI as a function of state.
Clearly hooks are not a bad idea given how quickly they've caught on in the react/front end community too. So while the majority certainly isn't always right, there's usually a common reasoning worth understanding
Similar view here, to an extent. But... with many topics, a majority of what I find is fanboy posts that strawman "problems" with earlier tech/tools, gloss over problems with the new way, and fawn over "hello world" examples. It can take a lot of time to get deep reasons, that may have some impactful value on your own problem spaces, and... at some point, I have to draw the line on "investigating all the bandwagons".
But also, at some point, one of the reasons to adopt something is "everyone else is", and there's some value in using "the latest trend" (hiring, bug solving, new tools, etc). And that may be enough value, but rarely does anyone justify their decision with "everyone else is doing it" - they try to give deeper 'technical' reasons, when (whether there are good/strong reasons) sometimes they just aren't able to do a good job.
I know there’s an aversion to classes in a lot of the current React community but the Component design in React was really much easier to reason with.
What would it have looked like if class-based components got a v2 rev instead? I feel like a lot of the crazy with class-based components came from the myriad low-level callbacks. useEffect() is a better abstraction, but it seems like some sort of class-based analogue could have been created (perhaps registering effects in the constructor).
The other nice thing about hooks is composability. I haven't given it a ton of thought but maybe that could have been addressed somehow too? Dunno.
I would be willing to accept some extra verbosity if it meant we could relax or eliminate some of the rules of hooks. The conditional rendering problem sucks.
This certainly seems like it would have been a better design, it would definitely be easier to reason about. Or even just an additional function boundary between where hooks are called and where JSX is rendered (semantically it’s basically the same thing but less verbose). Or even just changing the hooks interface to be a factory, accepting a component function as a parameter, where that component receives props `P & HookState<FooHook>`, which probably would be a lighter lift for React’s implementation because it could simply be a wrapper around the current behavior.
interface HookClassComponent {
render(): React.JSX
}
interface HookClassComponentConstructor {
new(): HookClassComponent // all hooks should happen here
}
function renderHookClass(hookClass: HookClassComponentConstructor) {
const instance = new hookClass()
return hookClass.render()
}
I'd be curious though to see more people try an "HOC" like this, though, and report back what sort of results they get.
I think it would fit more my preferred style as well as the kind of interface that people find attractive in hooks. It would work well with generics in TypeScript. And it wouldn’t be as fussy about the weird implicit behavior of component naming that came with wrapped/generated classes and decorators in past experiments with composition in past React fads.
Hooks are the place to segment out effectful code from otherwise pure code
If that's your preference React is probably the wrong tool to begin with. React is just the view in MVC. Plenty of frameworks exist to help build an MVC with react that provide their own structure and guidelines.
React is incredibly flexible and can be used in many ways. All this shows is the need for strong architecture decisions and guidance because React is so hands off in making recomendations of how it should be used or what tools it should be used with.
That's a strength. It absolutely requires more up front work on yours and mine part to construct an architecture that fits our needs, but it gives us the ability to completely tailor the solution which is the best part.
> If that's your preference React is probably the wrong tool to begin with.
That’s... a strange take, considering that’s how almost everyone wrote React UIs before hooks.
I feel like the unspoken difference in this thread is between people who have IO/business in their view code, and people whose taste is to strongly separate them, leaving component state for strict view modeling like animation timers, etc. I had this difference with my former tech lead.
The problem I have with hooks is fundamentally that they perform magic side effects that change the semantics of both the call site and its return type: what would otherwise be a pure function of props returning a data structure (referential transparency) is transformed into a constructor/factory type thing that returns a stateful thing which will render differently for the same props when its internal state has changed.
I generally would prefer to avoid state, and model as many changes as possible with functional data flow, but even where I do need to reach for state I’d prefer its initialization be expressed entirely outside the render function body. It would be a very small API change to achieve the same expressiveness of hooks, with just a small bit of additional syntax.
I’d argue that test is not that useful. Instead what we do in the react world today is focus on integration tests. Testing what happens when you navigate to a page, click on elements, and make sure it does what it is supposed to.
Just as you probably don’t get much value out of testing a function that increments an input number, but you may get a lot of value out of testing, say, a function that generates a series of smooth quadratic curves over a series of points.
And that’s a much easier thing to test at the unit level than integration.
It seems to me that either:
- your usage of props and rendering isn’t particularly complex, but your interaction model is
- you’re using state to manage that complexity where someone else may model it without state
I’m not passing judgment if it’s the latter, it’s fairly common in the frontend world to model that way. But it’s certainly not my preference.
I know what hooks are. Neither of those things is true. The hooks are called/instantiated in the same function body that returns the render. This performs some kind of stateful magic which makes that function itself inherently impure. Then that function returns a render which calls into that magic to perform further side effects, so that’s not pure either.
> The existence of a hook is what should alert you to the fact that there may be some side effect thing happening here.
That is true, and is for sure a benefit. But a class sends the same signal, but with the benefit of actually being fairly straightforward to reason about (notwithstanding the complexity of lifecycle, but lifecycle can be complex with hooks too).
> Not sure if this applies to you but I've found that most people who criticize hooks haven't made the leap to using custom hooks and have instead only used ones provided to them (useState, useEffect, etc).
I try not to use them at all, to minimize state and side effects as much as possible, and to encapsulate stateful code from rendering code.
I find them hard to reason about, whether in the core interface or as abstractions around the core. They fundamentally change the behavior of a function component from “this is a function which receives input (props) and returns a rendered data structure (JSX)” to “this sets up some initial state and returns a reference type of some kind that’s used to track that state over time”, but with the same syntax and semantics. The former is what attracted me to React and JSX, the latter is what I wanted to get away from. But making the same semantics do both is even worse than just having different semantics for both.
> Hooks are functions
Well not really, they’re routines expressed as functions, but they cause side effects on their enclosing context. That’s why, for example, you can’t conditionally call a hook. If you do, React won’t know to treat this component differently from components which don’t call hooks.
const myReactiveComputation = () => {
if (getReactiveValue() > 5){
console.log(getOtherReactiveValue());
}
}
useEffect(() => {
if (iWantToRunThis(){
...
};
}, [])
if (iWantToRunThis()) useEffect(() => {...}, []);
Hooks have been a savior for me. I agree that there are a lot of sharp edges with hooks, but they're not magic. They're powerful and they compose really well. If you're doing anything behaviorally complicated on a React app, it's easy to build whatever abstraction is appropriate to your project.
For example, if you wanted to instantiate a charting library like uPlot (https://github.com/leeoniya/uPlot) you might instantiate a uPlot object storing state and methods for the chart. You can perform the chart initialization in a useEffect hook, and use a ref to keep a reference to the instance you've created.
https://reactjs.org/docs/integrating-with-other-libraries.ht...
Ah, I also forgot to mention, what I'd like to see is more documentation on how to hide an imperative API behind a more declarative React/Angular etc. API. That's not always trivial but often what you want.
I think it's especially important because while you often can find ready-made wrapper for popular pure JS libraries they are usually abandoned (and thus don't keep up with the latest versions) or have some flaw that bites you sooner or later. My current recommendation is to write your own wrapper unless the wrapper is an official one of the authors of the underlying JS library. Naturally, that's not always feasible for complex libraries, but matter of fact you don't always need the full functionality of that library so you only have to wrap what you need.
One common pattern I end up using for a lot of imperative calls is a "useAsyncFunction" hook which runs any async function, returns state for loading, error, and the result of that function.
Unfortunately each library handles DOM manipulation, configuration options, callbacks, instantiation, etc. in its own way, and so I don't think there's any generalized solution.
This isn't correct; it's well documented [1], but people seem to often think this is true; if you reference it in multiple places or not is irrelevant.
Passing a ref in to the dependencies of `useEffect` doesn't work either.
Mostly I like hooks, but this one in particular seems to trip people up a lot; I'd go as far as to say that using callback refs should be the only supported way of using them; the normal one is pretty much broken by default for any component that isn't entirely trivial.
[1] - I think https://medium.com/@teh_builder/ref-objects-inside-useeffect... covers this pretty carefully.
I'm glad the article also shows how to use the more traditional API as well, even if it is a bit more verbose.
[1]: https://www.infoq.com/presentations/Simple-Made-Easy/ [2]: https://www.youtube.com/watch?v=4anAwXYqLG8
Before hooks, if you wanted a stateful component you had to use class components with all of their specific methods (didMount, didUpdate, shouldUpdate, render etc).
Hooks allow functional components to have state, and a functional component in and of itself has an even smaller API than the class. The React API vastly decreases. A single hook, useEffect, replaces at least 3 methods (didMount, didUpdate, willUnmount) that previously had to written out separately in class components. Each of those methods had to contain logic for ALL your state and side effects so they could very easily grow in scope and be responsible for handling many things.
Now each individual concern can be packaged up in its own useEffect call. If I have state that is relevant to the useEffect call, than I just build my own custom hook that binds the stateful data with the useEffect.
Because hooks are just functions built up from other hooks, I'm no longer constrained by the React API (its effectively gone) and I can so much more easily the abstractions I need that can be shared across components.
I mean, for testing, having too many class methods is clearly annoying so I try encapsulate as much functionality into small and separate functions.
But yeah, apart from that I never saw the need to use hooks. I mean either I write a functional component without state or I write a class component. That's it.
I hope they don't deprecate lifecycle methods. Such a great interface!