Getting started with fp-ts: Monad

gcanti

Giulio Canti

Posted on April 9, 2019

Getting started with fp-ts: Monad

In the last post we saw that we can compose an effectful program f: (a: A) => M<B> with a pure n-ary program g by lifting g, provided that M admits an applicative functor instance

Program f Program g Composition
pure pure g ∘ f
effectful pure, n-ary liftAn(g) ∘ f
where `liftA1 = lift`

However we must solve one last case: what if both programs are effectful?

f: (a: A) => M<B>
g: (b: B) => M<C>
Enter fullscreen mode Exit fullscreen mode

What's the "composition" of such f and g?

In order to handle this last case we need something more powerful than Functor since it's quite easy to end up with nested contexts.

The problem: nested contexts

To better explain why we need something more, let's see some examples.

Example (M = Array)

Say we want to retrive the followers of the followers of a Twitter user:

interface User {
  followers: Array<User>
}

const getFollowers = (user: User): Array<User> => user.followers

declare const user: User

const followersOfFollowers: Array<Array<User>> = getFollowers(user).map(getFollowers)
Enter fullscreen mode Exit fullscreen mode

There's something wrong here, followersOfFollowers has the type Array<Array<User>> but we want Array<User>.

We need to flatten the nested arrays.

The flatten: <A>(mma: Array<Array<A>>) => Array<A> function exported by fp-ts comes in handy

import { flatten } from 'fp-ts/Array'

const followersOfFollowers: Array<User> = flatten(getFollowers(user).map(getFollowers))
Enter fullscreen mode Exit fullscreen mode

Nice! What about other data structures?

Example (M = Option)

Say we want to calculate the inverse of the head of a numeric list

import { Option, some, none, option } from 'fp-ts/Option'
import { head } from 'fp-ts/Array'

const inverse = (n: number): Option<number> => (n === 0 ? none : some(1 / n))

const inverseHead: Option<Option<number>> = option.map(head([1, 2, 3]), inverse)
Enter fullscreen mode Exit fullscreen mode

Opss, I did it again, inverseHead has the type Option<Option<number>> but we want Option<number>.

We need to flatten the nested Options.

import { isNone } from 'fp-ts/Option'

const flatten = <A>(mma: Option<Option<A>>): Option<A> => (isNone(mma) ? none : mma.value)

const inverseHead: Option<number> = flatten(option.map(head([1, 2, 3]), inverse))
Enter fullscreen mode Exit fullscreen mode

All those flatten functions... It's not a coincidence, there's a functional pattern under the hood.

Indeed all those type constructors (and many others) admit a monad instance and

flatten is the most peculiar operation of monads

So what's a monad?

This is how monads are often presented...

Definition

A monad is defined by three things:

(1) a type constructor M which admits a Functor instance

(2) a function of with the following signature

of: <A>(a: A) => HKT<M, A>
Enter fullscreen mode Exit fullscreen mode

(3) a function flatMap with the following signature

flatMap: <A, B>(f: (a: A) => HKT<M, B>) => ((ma: HKT<M, A>) => HKT<M, B>)
Enter fullscreen mode Exit fullscreen mode

Note: recall that the HKT type is the fp-ts way to represent a generic type constructor, so when you see HKT<M, X> you can think to the type constructor M applied to the type X (i.e. M<X>).

The functions of and flatMap are required to obey three laws:

  • flatMap(of) ∘ f = f (Left identity)
  • flatMap(f) ∘ of = f (Right identity)
  • flatMap(h) ∘ (flatMap(g) ∘ f) = flatMap((flatMap(h) ∘ g)) ∘ f (Associativity)

where f, g, h are all effectful functions and is the usual function composition.

Ok but... why?

Back in the day when I first saw such a definition my first reaction was bewilderment.

All these questions were spinning around in my head:

  • why those two particular operations and why they have those types?
  • why the name "flatMap"?
  • why the laws? What do they mean?
  • but above all, where's my flatten?

This post will try to answer each question.

Let's get back to our problem: what's the composition of two effectful functions (also called Kleisli arrows)?

two Kleisli arrows, what's their composition?

(two Kleisli arrows, what's their composition?)

I don't even know what its type is.

Wait... we already encountered an abstraction that is all about composition. Do you remember what I said about categories?

Categories capture the essence of composition

We can turn our problem into a category problem: can we find a category that models the composition of Kleisli arrows?

The Kleisli category

Let's try to build a category K (named Kleisli category) which only contains effectful functions:

  • the objects are the same objects of the TS category, that is all the TypeScript types.
  • the morphisms are built like this: whenever there's a Kleisli arrow f: A ⟼ M<B> in TS we draw an arrow f': A ⟼ B in K

above the TS category, below the K construction

(above the _TS_ category, below the _K_ construction)

So what would be the composition of f' and g' in K? That's the dashed arrow labeled h' in the image below

above the composition in the TS category, below the composition in the K construction

(above the composition in the _TS_ category, below the composition in the _K_ construction)

Since h' is an arrow that goes from A to C, there should be a corresponding function h from A to M<C> in TS.

So a good candidate for the composition of f and g in TS is still an effectful function with the following signature: (a: A) => M<C>.

How can we build such a function? Well, let's try!

In which we build the composition step by step

The point (1) of the monad defintion says the M admits a functor instance, so we can lift the function g: (b: B) => M<C> to a function lift(g): (mb: M<B>) => M<M<C>> (here I'm using its synonym map)

where flatMap comes from

(where `flatMap` comes from)

And now we are stuck: there's no legal operation on the functor instance which is able to flatten a value of type M<M<C>> to a value of type M<C>, we need an additional flatten operation.

If we can define such operation then we can obtain the composition we are looking for

h = flatten ∘ map(g) ∘ f

But wait, flatten ∘ map(g) is flatMap, that's where the name comes from!

h = flatMap(g) ∘ f

We can now update our "composition table"

Program f Program g Composition
pure pure g ∘ f
effectful pure, n-ary liftAn(g) ∘ f
effectful effectful flatMap(g) ∘ f
where `liftA1 = lift`

What about of? Well, of comes from the identity morphisms in K: for each identity morphism 1A in K there should be a corresponding function from A to M<A> (i.e. of: <A>(a: A) => M<A>).

where of comes from

(where `of` comes from)

The laws

Last question: where do the laws come from? They are just the category laws in K translated into TS:

Law K TS
Left identity 1Bf' = f' flatMap(of) ∘ f = f
Right identity f' ∘ 1A = f' flatMap(f) ∘ of = f
Associativity h' ∘ (g' ∘ f') = (h' ∘ g') ∘ f' flatMap(h) ∘ (flatMap(g) ∘ f) = flatMap((flatMap(h) ∘ g)) ∘ f

Monads in fp-ts

In fp-ts the flatMap function is modeled by a variant called chain, which is basically flatMap with the arguments rearranged

flatMap: <A, B>(f: (a: A) => HKT<M, B>) => ((ma: HKT<M, A>) => HKT<M, B>)
chain:   <A, B>(ma: HKT<M, A>, f: (a: A) => HKT<M, B>) => HKT<M, B>
Enter fullscreen mode Exit fullscreen mode

Note that chain can be derived from flatMap (and viceversa).

Now if we get back to the examples showing the problem with nested contexts we can fix them by using chain

import { array, head } from 'fp-ts/Array'
import { Option, option } from 'fp-ts/Option'

const followersOfFollowers: Array<User> = array.chain(getFollowers(user), getFollowers)

const headInverse: Option<number> = option.chain(head([1, 2, 3]), inverse)
Enter fullscreen mode Exit fullscreen mode

Conclusion

Functional programming provides universal ways to compose functions with effects: functors, applicative functors and monads are all abstractions that offer principled tools for composing different kind of programs.

TLDR: functional programming is really all about composition

💖 💪 🙅 🚩
gcanti
Giulio Canti

Posted on April 9, 2019

Join Our Newsletter. No Spam, Only the good stuff.

Sign up to receive the latest update from our blog.

Related