People seem to forget that components (stateless components, stateless functional components or whatever you want to call them) are just functions. One function calls another function, that's function composition. In plain JavaScript we do it all the time. But when React is involved people are suddenly all like: no you can't call that function directly, you have to name it 'Component' with a capital C and use like '<Component />' and to compose two Components you have to use a HOC. It's like they all forget how to JavaScript…
An HOC is a plain JavaScript function that accepts a React component (and maybe other things) and returns a React component. I don’t think your complaint applies to HOCs, because they are literally an example of your suggestion to just use plain functions.
As a note, it is not safe for a component using a render prop to implement shouldComponentUpdate. The problem is that the render prop may close over data that the component is unaware of. The Asana core components used to the use the render prop pattern until we discovered that issue and also realized it made tests harder to write. Now we use React.cloneElement instead.
It also means you're passing in a new closure to the render prop on every render. You can move it to a member on your component's class, but that makes your code far hairier. It's also easy to forget to put it on the class, so you end up setting up eslint rules, and next thing you know, you've made everything worse.
I came up with the pattern after discovering the issues with render callbacks and before Higher Order Components were common. You can achieve the same result with a Higher Order Component. The only benefit of this approach over an HOC is that you write a regular Component.
To implement the pattern, you make you Component expect a single child for it's `children` prop. Let's assume that the child has a `data` prop which is not set in the parent Component. In render, you return `React.cloneElement(this.props.children, { data: this.state.valueToInject })`.
The value of a Higher Order Component is that you can usually define a map to props function.
I started today by setting out to write a response of the form "Certainly you couldn't rewrite MY HOC library to use render props, look at how it [etc]" and ended today having rewritten my HOC library to use render props. In the process, I was able to dramatically simplify the API and remove about a third of the overall code.
This is how the technique was first popularised by Cheng Lou in react-motion. But it was generally found that using children made it really inaccessible to people unfamiliar with the pattern.
I've literally had good developers not understand them until I switched an example from using children to using a render prop, at which point there's a big light bulb moment.
My biggest issue with HOCs is the hard time you have when things are nested to high hell. A component like withThis(withThat(enrich(withState(withPropsOnChange(withPropsOnChange(withPropsOnChange(withPropsOnChange(....... when this component gives you an issue you begin to wonder what HOCs are really doing for you.
Having a render prop is slightly better but even this escape hatch isn't foolproof and you'll still end up needing things like onClickOutside.
I think a lot of HOC libraries were designed assuming decorators would be standardized soon. The connect method from react-redux is definitely in that camp:
@connect(mapStateToProps, mapDispatchToProps)
class Component extends React.Component { ... }
To some extent, it was - if you look at the earliest versions of Redux, `connect()` and its predecessor forms were indeed being used as a decorator.
However, I personally advise against using `connect()` as a decorator, for several reasons:
- It's still a Stage 2 proposal. Now, the Class Properties syntax isn't final either (currently Stage 3), which the React team (and I) highly recommend using. However, the Class Properties syntax seems to be much more stable, the behavior it's implementing is a lot simpler, and if by chance it happens to change in the future, it should be relatively easy to code-mod (and the React team has said they would release a code-mod if that happens). Meanwhile, the decorators spec has changed several times (including recently), and the Babel plugins have also had to change behavior and implementation over time.
- It obscures the real class definition. The standard advice for testing Redux-connected components is to export the "plain" class separately as a named export, and `export default connect()(MyComponent)`, then import and test the plain version. If you use @connect() as a decorator, the plain version isn't accessible, and testing becomes more difficult.
- Going along with that, I've seen many questions about why defaultProps and propTypes don't work right when @connect() is used, and it's because those get applied to the wrapper component, not the plain component, and thus things don't work the way you would want them to.
I see no advantages to using connect as a decorator. I encourage people to write their mapState functions separately anyway for clarity and testability (instead of inline as an argument to connect), so it's just a matter of moving the line with `connect` elsewhere in the file and changing the syntax slightly.
I definitely wasn't recommending that usage, just illustrating what it might look like. I personally think it looks much cleaner, which I would call an advantage.
I've considered the first point, but hadn't thought about the second and third. I'm guessing the last point isn't a concern if you're using class properties for those?
Given all this, if you were redesigning the API today, would you now make connect take the component as the first parameter instead?
No. The reason why it's written as not just a HOC, but a curried-ish function is so that you could potentially reuse the "connection" definition for multiple components:
Admittedly, I suspect that use case isn't popping up very often. Most of the time what I see is that a given component type is only connected once, and a given connection setup is only used with one component. I also don't remember seeing specific comments by Dan or Andrew saying that's why it's written this way - I'm inferring the intent from the API definition. Still, it's a potentially useful capability in the API, so no reason to throw it away.
Theoretically that could be useful, but I've also never seen connections used this way. I have seen components connected in more than one way, but of course that's no easier with the curried style. And, I'd point out, if someone actually wanted to share the configuration, it's trivial enough to write:
Not that I'd argue against currying generally -- it's only that it turns out to be a little awkward here with the various optional parameters to connect, and also that the connection setup in practice seems to be tightly coupled to the connected component.
What alternative would you propose for a component that genuinely needed all the functionality of that nested HOC example you provided?
The render prop pattern will have just as much nesting. Mixins or plain JavaScript class composition have major well-accepted problems. You could just write one big component with all that functionality, but the fact that it would ever be split into layers of HOCs implies that various functionality is being reused (and probably provided by third party libraries).
The technique discussed is a kind of inversion of control, a concept which has a pretty long history (~30 years according to the wikipedia article). I've been writing React for a couple of months now and had settled on this method without knowing it was considered new or in anyway unusual.
First, could you clarify what _you_ mean by "Redux boilerplate" in this case? The phrase gets thrown around frequently, but it means different things to different people. Just yesterday, I tweeted some thoughts on how you can use as much abstraction as you want with Redux, and linked to examples of reusable action creator/reducer logic [0]. I also wrote a pair of blog posts that discuss the intent behind Redux's design [1], and why common usage patterns exist [2].
It's worth noting that React-Redux was originally written using a "render props"-type approach, but was changed to be a Higher-Order Component before it hit 1.0. In fact, there was a thread a month ago that discussed render props-based reimplementations of `connect` [3], and in that thread I linked to several recent examples of people reinventing that wheel as well as prior discussion of why React-Redux wound up as a HOC.
Finally, earlier this year I opened up an issue to discuss ways that we can improve both the "getting started" experience for Redux users, as well as build more powerful abstractions on top of Redux [4]. I'd love more feedback and ideas (and ideally people from the community volunteering to help us make things better).
Have you looked at any of the existing libraries in the ecosystem? There's dozens of tools for generating reusable functions like action creators and reducers [0], managing complex collections [1], and even higher-level abstractions on top of Redux like Kea [2].
Ultimately, you can put as much or as little abstraction on top of Redux as you want.
"Higher order" sounds pretty functional to me. Higher order functions and all. Although I guess the term higher order function is hardly even spoken of in real functional languages because they're taken for granted.
The term “higher order” generally indicates complexity and indirection. If you can solve something first order, then that is much more preferable to a higher order (function, logic, object) solution. Higher order functions fall into the same category (you can have them if you want, but you must understand what you are getting into in terms of complexity).
> The term “higher order” generally indicates complexity and indirection.
As do all abstractions.
> If you can solve something first order, then that is much more preferable to a higher order solution.
a.k.a. "To abstract, or not to abstract?". It really depends, of course. But I know for certain I would not want to solve "mapping over an array" without the Array.prototype.map higher-order function.
It really depends. Higher order debugging is more of an art than reality, so if the mapping logic is tricky, I would definitely convert a map into a loop just so I had better access to debugging. It should never be “use map whenever possible.” Also, if you drift out of well known functional origami to more niche uses of function arguments, well, there be the dragons.
Higher order function is a functional pattern.
In some OOP languages where functions were not available (Java7), it was not possible to create Higher order function without classes.
https://en.wikipedia.org/wiki/Higher-order_function
