beginner
Note: This section keeps on growing! Keep an eye on it from time to time.
This document is meant to be an introduction to Functional Programming for people from all backgrounds. We’ll go through some of the key concepts and then dive into their implementation and use in real world cases.
Some similar documents, focused on explaining general concepts rather than Arrow’s versions, can be found for examples in JavaScript and in Scala.
A datatype is a class that encapsulates one reusable coding pattern. These solutions have a canonical implementation that is generalised for all possible uses.
Some common patterns expressed as datatypes are absence handling with Option
,
branching in code with Either
,
catching exceptions with Try
,
or interacting with the platform the program runs in using IO
.
Some of these patterns are implemented using a mix of sealed
classes where each inheritor is a data
class.
For example, the internal representation of an Option
is a sealed
class with two data
classes Some<A>(val a: A)
and None
,
and Ior
is a sealed
class with three data
class inheritors, Left(val a: A)
, Right(val b: B)
, and Both(val a: A, val b: B)
.
Datatypes that express patterns like deferred evaluation can do it by nesting themselves with every operation they chain. One example is IO
.
import arrow.effects.*
IO { 0 }
.continueOn(UI)
.flatMap { IO { it * 2 } }
.map { it + 1 }
// Map(source=Bind(cont=ContinueOn(cont=Suspend(thunk=() -> arrow.effects.IO.Pure<A>), cc=Unconfined), g=(A) -> arrow.effects.IO<B>), g=(kotlin.Int) -> kotlin.Int, index=0)
You can read more about all the datatypes that Arrow provides in its section of the docs.
Typeclasses are interface abstractions that define a set of extension functions associated to one type. These extension functions are canonical and consistent across languages and libraries; and they have inherent mathematical properties that are testable, such as commutativity or associativity.
Examples of behaviours abstracted by typeclasses are: comparability (Eq
),
composability (Monoid
),
its contents can be mapped from one type to another (Functor
),
or error recovery (MonadError
).
Typeclasses have two main uses:
Add new functionality to types. For example, if I know how to compare two objects I can add a new extension function to check for inequality.
Or if I know how to aggregate objects together, I can add an extension function for List
that aggregates all of its elements.
The number of extra extension functions that you get per typeclass can be from one in Eq
to 17 (!) in Foldable
.
Abstracting over behavior. Like any other interface, you can use them in your functions and classes as a way of talking about the capabilities of the implementation,
without exposing the details. This way you can create APIs that work the same for Option
, Try
, or Observable
.
You can read more about all the typeclasses that Arrow provides in its section of the docs.
Let’s dive in one example. The typeclass Eq
parametrized to F
defines equality between two objects of type F
:
interface Eq<F> {
fun F.eqv(b: F): Boolean
fun F.neqv(b: F): Boolean =
!eqv(b)
}
A single implementation of a typeclass for a specific datatype or class. Because typeclasses require generic parameters each implementation is meant to be unique for that parameter.
For example, given a class like this:
data class User(val id: Int) {
companion object
}
We can declare that instances of this class can be equated based on their id
property, and therefore that User
itself is an instance of the Eq
typeclass:
@extension
interface UserEqInstance: Eq<User> {
override fun User.eqv(b: User): Boolean = id == b.id
}
Note that classes must have companion objects for this to work. All typeclass instances provided by Arrow can be found in the companion object of the type they’re defined for, including platform types like String or Int.
import arrow.*
import arrow.core.*
import arrow.data.*
import arrow.instances.*
import arrow.typeclasses.*
import arrow.instances.option.functor.*
import arrow.instances.either.monadError.*
import arrow.instances.listk.traverse.*
String.eq()
// arrow.instances.StringKt$eq$1@3c293d05
Option.functor()
// arrow.instances.option.functor.OptionFunctorInstanceKt$functor$1@4e46fa3f
import arrow.instances.mapk.semigroup.*
MapK.semigroup<String, Int>(Int.semigroup())
// arrow.instances.mapk.semigroup.MapKSemigroupInstanceKt$semigroup$1@8c3058a
Either.monadError<Throwable>()
// arrow.instances.either.monadError.EitherMonadErrorInstanceKt$monadError$1@4c16c0e3
ListK.traverse()
// arrow.instances.listk.traverse.ListKTraverseInstanceKt$traverse$2@2fc877f7
If you’re defining your own instances and would like for them to be discoverable in their corresponding datatypes’ companion object, you can generate it by annotating them as @extension
, and Arrow’s annotation processor will create the extension functions for you.
NOTE: If you’d like to use @extension
for transitive typeclasses, like a Show<List<A>>
that requires a function returning a Show<A>
, you’ll need for the function providing the transitive typeclass to have 0 parameters. This will make the transitive typeclass a parameter of the extension function.
Arrow provides a extensions
DSL making available in the scoped block all the functions and extensions defined in all instances for that datatype. Use the infix function extensions
on an object, or function, with the name of the datatype prefixed by For-.
ForOption extensions {
binding {
val a = Option(1).bind()
val b = Option(a + 1).bind()
a + b
}.fix()
}
//Option(3)
ForOption extensions {
map(Option(1), Option(2), Option(3), { (one, two, three) ->
one + two + three
})
}
//Option(6)
ForOption extensions {
listOf(Option(1), Option(2), Option(3)).k().traverse(this, ::identity)
}
//Option(ListK(1, 2, 3))
ForTry extensions {
binding {
val a = Try { 1 }.bind()
val b = Try { a + 1 }.bind()
a + b
}.fix()
}
//Success(3)
ForTry extensions {
map(Try { 1 }, Try { 2 }, Try { 3 }, { (one, two, three) ->
one + two + three
})
}
//Success(6)
ForEither<Throwable>() extensions {
listOf(just(1), just(2), just(3)).k().traverse(this, ::identity)
}
//Right<Throwable, ListK<Int>>(ListK(1,2,3))
If you defined your own instances over your own data types and wish to use a similar extensions
DSL you can do so for both types with a single type argument such as Option
:
object OptionContext : OptionMonadErrorInstance, OptionTraverseInstance {
override fun <A, B> Kind<ForOption, A>.map(f: (A) -> B): Option<B> =
fix().map(f)
}
infix fun <A> ForOption.Companion.extensions(f: OptionContext.() -> A): A =
f(OptionContext)
Or for types that require partial application of some of their type arguments such as Either<L, R>
where L
needs to be partially applied
class EitherContext<L> : EitherMonadErrorInstance<L>, EitherTraverseInstance<L> {
override fun <A, B> Kind<EitherPartialOf<L>, A>.map(f: (A) -> B): Either<L, B> =
fix().map(f)
}
class EitherContextPartiallyApplied<L> {
infix fun <A> extensions(f: EitherContext<L>.() -> A): A =
f(EitherContext())
}
fun <L> ForEither(): EitherContextPartiallyApplied<L> =
EitherContextPartiallyApplied()
NOTE: This approach to type constructors will be simplified if KEEP-87 is approved. Go vote!
A type constructor is any class or interface that has at least one generic parameter. For example,
ListK<A>
or Option<A>
.
They’re called constructors because they’re similar to a factory function where the parameter is A
, except type constructors work only for types.
So, we could say that after applying the parameter Int
to the type constructor ListK<A>
it returns a ListK<Int>
.
As ListK<Int>
isn’t parametrized in any generic value it is not considered a type constructor anymore, just a regular type.
Like functions, a type constructor with several parameters like Either<L, R>
can be partially applied for one of them to return another type constructor with one fewer parameter.
For example, applying Throwable
to the left side yields Either<Throwable, A>
, or applying String
to the right side results in Either<E, String>
.
Type constructors are useful when matched with typeclasses because they help us represent instances of parametrized classes — the containers — that work for all generic parameters — the content.
As type constructors is not a first class feature in Kotlin, Λrrow uses an interface Kind<F, A>
to represent them.
Kind stands for Higher Kind, which is the name of the language feature that allows working directly with type constructors.
In a Higher Kind with the shape Kind<F, A>
, if A
is the type of the content then F
has to be the type of the container.
A malformed Higher Kind would use the whole type constructor to define the container, duplicating the type of the content .
This incorrect representation has large a number of issues when working with partially applied types and nested types.Kind<Option<A>, A>
What Λrrow does instead is define a surrogate type that’s not parametrized to represent F
.
These types are named same as the container and prefixed by For-, as in ForOption
or ForListK
.
You have seen these types used in the Syntax section above!
class ForOption private constructor() { companion object {} }
sealed class Option<A>: Kind<ForOption, A>
class ForListK private constructor() { companion object {} }
data class ListK<A>(val list: List<A>): Kind<ForListK, A>
As ListK<A>
is the only existing implementation of Kind<ForListK, A>
, we can define an extension function on Kind<ForListK, A>
to do the downcasting safely for us.
This function by convention is called fix()
, as in, fixing a type from something generic into concrete.
fun Kind<ForListK, A>.fix() = this as ListK<A>
This way we have can to convert from ListK<A>
to Kind<ForListK, A>
via simple subclassing and from Kind<ForListK, A>
to ListK<A>
using the function fix()
.
Being able to define extension functions that work for partially applied generics is a feature from Kotlin that’s not available in Java.
You can define fun Kind<ForOption, A>.fix()
and fun Kind<ForListK, A>.fix()
and the compiler can smartly decide which one you’re trying to use.
If it can’t it means there’s an ambiguity you should fix!
The function fix()
is already defined for all datatypes in Λrrow, alongside a typealias for its Kind<F, A>
specialization done by suffixing the type with Of, as in ListKOf<A>
or OptionOf<A>
. If you’re creating your own datatype that’s also a type constructor and would like to create all these helper types and functions,
you can do so simply by annotating it as @higherkind
and the Λrrow’s annotation processor will create them for you.
@higherkind
data class ListK<A>(val list: List<A>): ListKOf<A>
// Generates the following code:
//
// class ForListK private constructor() { companion object {} }
// typealias ListKOf<A> = Kind<ForListK, A>
// fun ListKOf<A>.fix() = this as ListK<A>
Note that the annotation @higherkind
will also generate the integration typealiases required by KindedJ as long as the datatype is invariant. You can read more about sharing Higher Kinds and type constructors across JVM libraries in KindedJ’s README.
Now that we have a way of representing generic constructors for any type, we can write typeclasses that are parametrised for containers.
Let’s take as an example a typeclass that specifies how to map the contents of any container F
. This typeclass that comes from computer science is called a Functor
.
interface Functor<F> {
fun <A, B> Kind<F, A>.map(f: (A) -> B): Kind<F, B>
}
See how the class is parametrized on the container F
, and the function is parametrized to the content A
. This way we can have a single representation that works for all mappings from A
to B
.
Let’s define an instance of Functor
for the datatype ListK
, our own wrapper for lists.
@extension
interface ListKFunctorInstance : Functor<ForListK> {
override fun <A, B> Kind<F, A>.map(f: (A) -> B): ListK<B> {
val list: ListK<A> = this.fix()
return list.map(f)
}
}
This interface extends Functor
for the value F
of ListK
. We use an annotation processor @extension
to generate an object out of an interface with all the default methods already defined, and to add an extension function to get it into the companion object of the datatype.
The @extension
processor also projects all type class declared functions into the data type that it’s extending as extensions functions.
These extensions functions may be imported a la carte when working with concrete data types.
@extension
interface ListKFunctorInstance : Functor<ForListK>
// Somewhere else in the codebase
import arrow.instances.listk.functor.*
ListK.functor()
// arrow.instances.listk.functor.ListKFunctorInstanceKt$functor$1@5a816e76
The signature of map
once the types have been replaced takes a parameter Kind<ForListK, A>
, which is the receiver, and a mapping function from A
to B
. This means that map will work for all instances of ListK<A>
for whatever the value of A
can be.
override fun <A, B> Kind<ForListK, A>.map(f: (A) -> B): ListK<B>
The implementation is short. On the first line we downcast Kind<ForListK, A>
to ListK<A>
using fix()
. Once the value has been downcasted, the implementation of map inside the ListK<A>
we have obtained already implements the expected behavior of map.
val list: ListK<A> = this.fix()
return list.map(f)
Higher kinds are also used to model functions that require for a datatype to implement a typeclass. This way you can create functions that abstract behavior (defined by a typeclass) and allow callers to define which datatype they’d like to apply it to.
Let’s use the typeclass Applicative
, that contains the constructor function just()
.
interface Applicative<F>: Functor<F> {
// Constructs the current datatype with a value of type A inside
fun <A> just(a: A): Kind<F, A>
/* ... */
}
Once we have this typeclass behavior define we can now write a function that’s parametrized for any F
that has one instance of Applicative
. The function uses the constructor just
to create a value of type Kind<F, User>
, effectively generifying the return on any container F
.
fun <F> Applicative<F>.randomUserStructure(f: (Int) -> User): Kind<F, User> =
AP.just(f(Math.random()))
Now lets create a simple example instance of Applicative
where our F
is ListK
. This implementation of a just
constructor is trivial for lists, as it just requires wrapping the value.
@instance
interface ListKApplicativeInstance : Applicative<ForListK> {
override fun <A> just(a: A): Kind<ForListK, A> = ListK(listOf(a))
/* ... */
}
And now we can show how this function randomUserStructure()
can be used for any datatype that implements Applicative
. As the function returns a value Kind<F, User>
the caller is responsible of calling fix()
to downcast it to the expected value.
val list = ListK.applicative().randomUserStructure(::User).fix()
//[User(342)]
val option = Option.applicative().randomUserStructure(::User).fix()
//Some(User(765))
val either = Either.applicative<Unit>().randomUserStructure(::User).fix()
//Right(User(221))
Passing the instance in every function call seems like a burden. So, because randomUserStructure
is an extension function for Applicative
we can omit the implicit parameter as long as we are within the scope of an Applicative instance. You can use the standard functions with
and run
for this.
with (ListK.applicative()) {
// Lots of Kotlin here
// Multiple calls
randomUserStructure(::User).fix()
}
// [User(342)]
Option.applicative().run {
tupled(randomUserStructure(::User), randomUserStructure(::User))
}
// Some(value = Tuple2(a = User(765), b = User(127)))
It is also possible to use a form of Dependency Injection
to make the typeclass scope available to a whole class. For example, using simple delegation:
class UserFetcher<F>(AP: Applicative<F>): Applicative<F> by AP {
fun genUser() = randomUserStructure(::User)
}
UserFetcher(Option.applicative()).genUser().fix()
// Some(value = User(943))
To learn more about this Typeclassless
technique you should head to the Dependency Injection
documentation.
A side-effect is statement that changes something in the running environment. Generally this means setting a variable, displaying a value on screen, writing to a file or a database, logging, start a new thread…
When talking about side-effects, we generally see functions that have the signature (...) -> Unit
, meaning that unless the function doesn’t do anything, there’s at least one side-effect. Side-effects can also happen in the middle of another function, which is an undesirable behavior in Functional Programming.
Side-effects are too general to be unitested for because they depend on the environment. They also have poor composability. Overall, they’re considered to be outside the Functional Programming paradigm, and are often referred as “impure” functions.
Because side-effects are unavoidable in any program FP provides several datatypes for dealing with them! One way is by abstracting their behavior. The simplest examples of this are the Writer
datatype, which allows you to write to an information sink like a log or a file buffer; or State
datatype, which simulates scoped mutable state for the duration of an operation.
For more complicated side-effects that can throw or jump threads we need more advanced datatypes, called Effects, that wrap over impure operations. Some of these datatypes may be already familiar to you, like rx.Observable
, kotlinx.coroutines.Deferred
, or Arrow’s IO
. These Effects compose, catch exceptions, control asynchrony, and most importantly can be run lazily. This gets rid of the issues with side-effects.
Although one can also write the whole program in an imperative way inside a single Effect wrapper that wouldn’t be very efficient as you don’t get any of its benefits :D