my face
paulgray.net

Redux Selectors are Monads

Using fp-ts to unlock advanced composition patterns.

May 17, 2019
javascriptreduxfunctormonadfunctional programming

Realizing this unlocks some advanced composition patterns that we may have not considered before. Such as Do (a haskell-like do notation for JavaScript):

const userPostsSelector = 
  Do(selectorMonad)
    .bind('user', userSelector)
    .bindL('posts', ({user}) => postsSelector(user.id))
    .done()

Here, we've composed userSelector and postsSelector into another selector, userPostsSelector, which selects a user and that user's posts, returning an object like:

{
  user: { id: 1, name: 'bob' },
  posts: [...]
}

The Monad interface

A monad is a wrapper around some value, which has a chain method that takes a callback, which takes the wrapped value, and returns another wrapper around some other value, and 'unwraps' it, leaving you with only one layer of wrapper. If we were to write out the type signature, it would look something like this:

class Wrapper<A> {
  chain(f: (a: A) => Wrapper<B>): Wrapper<B>
}

Wrapper can be any generic type. You've undoubtedly used this pattern before. Take a look at a (simplified) definition of a Promise's then:

class Promise<A> {
  then(f: (a: A) => Promise<B>): Promise<B>
}

Here we've just plugged Promise into Wrapper, and you're probably familiar with this function already.

The same is true for Arrays:

class Array<A> {
  flatMap(f: (a: A) => Array<B>): Array<B>
}

Although arrays name this function flatMap and promises name this function then, they both have similar shapes!

Selectors could also have a chain method which is shaped just like promise's then and array's flatMap:

class Selector<State, A> {
  chain(f: (a:A) => Selector<State, B>): Selector<State, B>
}

This chain/flatMap/then method is the mark of a monad. Of course, there are a few more details which I'll describe later, but this one is the most important.

In fp-ts, the Monad interface is modelled as a typeclass (This just means that the methods we need to implement exist on a separate object, instead of on the class itself).

In our Selector example, we showed selectors as a class with a chain method. In reality, selectors are functions that are created outside of our control. Normally, If we wanted selectors to implement an interface, we'd have to wrap them with a thin wrapper that implements that interface, which gets clunky. With typeclasses we simply define extra functions on a standalone object, freeing us from this restriction:

type Selector<State, A> = (s: State) => A;

const selectorMonad = {
  ...
  chain<State, A, B>(
    selector: Selector<State, A>,
    f: (a: A) => Selector<State, B>
  ): Selector<State, B> {
    ???
  }
}

Take a minute and try to implement this method. It might help to try and articulate in your head what the arguments are, and what the return type is, before thinking about an implementation.

Do

Do works for all monads, not just selectors, arrays, and promises. How does it know when to use flatMap, then, or chain? The first parameter to Do is an instance of the Monad typeclass. This object contains all the logic.

For example, if we wanted to use Do with selectors, we'd take the selectorMonad typeclass instance and give it to Do:

Do(selectorMonad)

Do returns a "builder," which allows us to incrementally bind the result of our selectors to identifiers that can be used by subsequent selectors. For example, lets take our userSelector and bind it to the "user" identifier:

Do(selectorMonad)
  .bind("user", userSelector)

Remember, userSelector is still a normal selector function at this point. We could still use this like normal, say: userSelector(state), which would return a user. Since functions don't have a chain method defined on them, Do uses the chain method defined on selectorMonad.

Next, we can compose another selector that uses the output from userSelector. For our example, suppose we have a postsSelector which can select posts out of state for a particular user:

const postsSelector: (u: User) => Selector<State, Post[]> =
  user => state => {
    return state.posts[user.id]
  }

postsSelector is curried, which just means that we'll invoke it like:

postsSelector(myUser)(state) // returns posts[] for this user

This will make it easier to compose with other selectors, and to use in Do.

The bindL method (short for bindLamda) allows us to pass a function which takes all of the previously bound values as it's parameter. In our case, the user (which is the user returned from our userSelector) is already bound, so we can destructure it:

Do(selectorMonad)
  .bind("user", userSelector)
  .bindL("posts", ({ user }) => postsSelector(user.id))

From the function we pass to bindL we return a selector, and Do will use the chain method from our selectorMonad to compose the postsSelector with the userSelector.

This lambda form of bind is incredibly powerful, as it lets us compose wrappers which depend on values that come from wrappers higher in the change.

If this sounds familiar, it should. This was the impetus behind the async/await syntax in Javascript. Consider this example where, instead of selecting users and posts from a redux store, we're fetching users and posts from a server:

async function fetchUser() {
  const user = await fetchUser();
  const posts = await fetchPosts(user.id);
  return posts;
}

Here, we're composing chains of promises, where later promises rely on the output of previous promises. Can you see how it's similar to the example where we composed selectors?

We could also build a monad typeclass for promises and use Do to compose them:

function fetchUser() {
  return Do(promiseMonad)
    .bind('user', fetchUser())
    .bindL('posts', ({user}) => fetchPosts(user.id))
    .done()
}

Promises aren't technically monads, since they don't behave the monad laws (they are not referentially transparent). If that doesn't make sense to you right now, don't fret; we can still use them with Do, since they match the Monad shape.

For future reference, it might make more sense to use from fp-ts, which are just lazy promises.

Of course, Using Do to compose promises is a little more clumsy than the equivalent async/await (However you might finid that error-handling is much more straightforward). Why would we ever use Do then?

Unlike async/await, Do works for all monads, not just promises.

Consider getting all posts from all users:

Do(arrayMonad)
  .bind('user', usersArray)
  .bindL('posts', ({ user }) => user.postsArray)

Or getting a user's image from a user that's potentially null:

Do(maybeMonad)
  .bind('user', maybeUser)
  .bindL('image', ({ user }) => user.maybeImage)

Or even using multiple render props/hocs/hooks together:

Do(renderPropMonad)
  .bind('token', withAuthToken)
  .bind('currentUser', ({ token }) => withLoadable(fetchUser(token)))
  .do(({ token, user }) => ensureAdmin(user))
  .done()
  .render(({ token, user }) => (
    <div>Welcome, {user.name}!</div>
  ))

Do unlocks new syntax sugar for composing one of the most ubiquitous patterns in functional programming: monads. This is so common, in fact, that many functional programming languages provide syntax sugar for this very case. For example, there's do in haskell, and for-comprehensions in scala.

One day, I hope that Javascript can also have this type of syntax sugar. I've written a proposal for this syntax here: https://github.com/pfgray/ecma-proposal-chainable-do-syntax#generalization, and even a babel plugin which gets us most of the way there: https://github.com/pfgray/babel-plugin-monadic-do.

One of the best features of fp-ts is that it uses typeclasses, which means you don't need to adjust anything about the way you're using selectors. You simply define the selectorMonad object, and you can instantly start composing selectors that you made with the reselect library (or any other library for that matter!).

Async await was a great improvement (syntax-wise) for composing chains of promises. Do aims to bring that convenience to wealth of other types. You just need to be open to seeing those types through a different lens.