Check out my Latest articles

An ode to the kind-projector and to the partial-unification of Scala

12 Apr 201817 min read
➡️ scala partial-unification kind-projector cats scalaz dotty typeclass hkt

The kind-projector Scalac plugin is quite ubiquitous in the Scala world. So much that we won’t need it in the future! Its features will be native in Dotty/Scala 3 as we’ll see. 🙏 But we’re not there yet.

This plugin introduces new keywords to provide some syntactic sugar for advanced type-related usage. It does not provide any new feature but simplifies how to write and read code that is hardly readable without.

If you’ve seen code like:

def makeFunctor[U] = new Functor[Either[?, U]] { ... }
def compose[G[_]: Applicative]: Applicative[λ[α => F[G[α]]]] = { ... }

It’s the kind-projector in action. It’s way less verbose and explicit than what we would have to write without it (we’ll see).

We’ll also learn how it’s linked to the Partial Unification feature of scalac and all its caveats.

The kind-projector and the partial-unification are powerful features that push the limits of the Scala type system. It lets us —the developers— abstract our work to reuse (compose) it easily. It limits the scope of possibilities when we code: when everything is abstract, it’s less bug-prone. It lets the compiler help us to check if the types make sense.

They are heavily used in libraries because they need to be the most generic possible for us to work with them (and even with themselves…). But they will also find their ways into real projects and applications. We must be aware of what the partial-unification is. We must consider the kind-projector as native, nor fear it, and totally embrace it. 😻


Summary


Partial Unification

First, we’ll study partial-unification on its own by going through some examples. That will help us to understand what it is all about. We can forget about the kind-projector for now.

What’s the problem?

Let’s say with have a simple function that works with any type constructor F[_] (* -> * / a higher kinded type / a HKT):

def show[F[_], A](f: F[A]) = println(f) // we can't do much more with `f` except use it as `Any`

We can give it instances that fits, such as List[_] or Option[_]:

show(List(1, 2, 3)) // "List(1, 2, 3)"
show(Some("hello")) // "Some(hello)"

But we can’t pass it variables which types doesn’t fit F[_] such as Either[_, _] (* -> * -> * doesn’t fit with * -> *):

val err: Either[Error, String] = Left(new Error("boom"))
show(err) // doesn't not compile!

Without partial-unification, we can’t compile this code: 😢

no type parameters for method show: (f: F[A])Unit exist so that it can be applied to arguments (Either[Error,String]).
 --- because ---
argument expression's type is not compatible with formal parameter type;
 found   : Either[Error,String]
 required: ?F[?A]

There is a workaround to make it work. We can create a type that has only one parametric type, and use type ascription to pass the value (it’s checked at compile-time, it’s not a runtime cast):

type ErrorEither[A] = Either[Error, A]
show(err: ErrorEither[String]) // "Left(java.lang.Error: boom)"

We can also generalize this type to ignore its left parametric type (all three notations do the same):

type R[A] = Either[_, A] // basically, "Any"
type R[A] = ({ type F[E] = Either[E, A] })#F[_] // type lambda and projection
type R[A] = Either[E, A] forSome { type E }     // existential types

show(Right(5): R[Int])          // "Right(5)"
show(Right("hello"): R[String]) // "Right(hello)"
show(Left(5): R[String])        // "Left(5)"

Partial unification to the rescue

With partial-unification flag enabled, the same code works without declaring the useless types or any types trick: 💪

// build.sbt
scalacOptions += "-Ypartial-unification"
val err: Either[Error, String] = Left(new Error("boom"))
show(err) // "Left(java.lang.Error: boom)"

The compiler successfully unified the call-site types with the function signature types: Either[_, _] into F[_]. How’s that? One type is missing! The compiler had to choose one. This is why it’s called partial unification, because it’s not total. Note that it’s used for type constructor unification only (the unification is done per parameter, a simple A doesn’t need any partial application!).

Which types did it used in show? It must respect some rules, it’s not random. We can find out by using some scala-reflect compile-time reflection (or -Xprint:typer):

import scala.reflect.runtime.universe._
def show[F[_], A: TypeTag](f: F[A])(implicit tt: TypeTag[F[A]]) = {
    println(s"${typeOf[F[A]]} and A: ${typeOf[A]}")
}

show(List(1, 2, 3)) // "List[Int] and A: Int"
show(Some("hello")) // "Some[String] and A: String"
val err: Either[Error, String] = Left(new Error("boom"))
show(err)           // "Either[Error,String] and A: String"

The compiler used the right type of Either[_, _], the left one does not exist from the show() perspective.

Let’s try with more complex types: 🙀

show(Triple(1, "2", true)) // "Triple[Int,String,Boolean] and A: Boolean"
show(List(List(1, 2, 3)))  // "List[List[Int]] and A: List[Int]"
show(Free.liftF(IO.pure(Some("hello")))) // Free[IO,Some[String]] and A: Some[String]

It really accepts anything!

show(5)         // "Any and A: Nothing"
show("hello")   // "Comparable[String] and A: String"
                // This one is "funny" because: String extends Comparable[String]
show((str: String) => str.length) // "String => Int and A: Int" (Function[String, Int])

If we resume how the unification works:

def show[F[_], A](f: F[A]) = ???

                   List[Int]: F[_],A = List, Int      // perfect match
              Option[String]: F[_],A = Option, String // perfect match
       Either[Error, String]: F[_],A = [T]Either[Error, T], String
Triple[Int, String, Boolean]: F[_],A = [T]Triple[Int, String, T], Boolean
Free[IO[_], Option[String]]]: F[_],A = [T]Free[IO, T], Option[String]
               String => Int: F[_],A = [T]String => T, Int
                         Int: F[_],A = Any, Nothing

As we see, when the types don’t match, the compiler fixes the left-most type parameters until a global match is found.

Here are more examples with more type parameters:

def show2[F[_, _], A, B](f: F[A, B]) = ???

val err: Either[Error, String] = Left(new Error("boom"))
show2(err)                         // F[_,_],A,B = Either, Error, String: // perfect match
show2((str: String) => str.length) // F[_,_],A,B = Function1, String, Int // perfect math
show2(Triple(1, "2", true))        // F[_,_],A,B = [T, U](Int, T, U), String, Boolean

When we try to trigger the partial unification on a nested type constructor, it’s not what we expect:

def show3[F[_[_]], G[_]](f: F[G]) = ???
show3(Some(Triple(1, "2", true)))

I would hope it’s trying to match: F as Some and G as Triple (and doing partial unification on its args), but it seems we’re at the edge of the type system:

scala.reflect.internal.Types$NoCommonType:
 lub/glb of incompatible types: [X] <: scala.collection.GenTraversable[X] and

It’s ill-kinded because Option has the shape F[_] whereas our function expects a F[_[_]].

We can do this instead:

def show4[F[_], G[_], A](fa: F[G[A]]]) = ??? // F[G[A]] = ((* -> *) -> *)
show4(Some((1, true)))           // F[_],G[_] = Some, [T](Int, T) -- partial unification!
                                 // Note that Intellij IDEA thinks it's a type mismatch.

Anyway, we get to the very important rule that will bite us later (when using the kind-projector, we’ll see):

The type constructors are partially applied from left to right leaving the rightmost type parameters free, therefore favorising the right-biaised type constructors.

Paraphrasing @djspiewak:

You should always, always order the type parameters in your type constructor from least-to-most specific. In other words, if your type constructor takes multiple parameters, you should order them such that the parameters you expect users to vary (i.e. write a map function over, change frequently between functions, etc) should be right-most, while the parameters you expect to be consistent (i.e. the same at multiple unrelated places in the code) should be left-most.

At the end, it’s a feature of the compiler to treat F[A, B] as F[A][B] and apply currying when it needs to match types with a given object.

History and deprecation

We have seen what’s partial-unification is all about. Let’s resume its history and what’s going to happen:

Scala 2.12.6 won’t need the -Ypartial-unification flag anymore.

Kind Projector

Now we can tackle the other part of this article. The kind-projector is unrelated to the partial-unification and is more fun! The combinaison of both features make the code really verbose saving. (and can sometimes let Intellij troubled when it tries to validate the code, red everywhere)

As said, the kind-projector is simply a scalac plugin. It is NOT part of scalac for now, but this will eventually arrive in Dotty/Scala 3 with some minor adjustements I guess. 🙏

What does it offer?

The kind-projector provides advanced type features and reserves some identifiers in the language such as:

  • Lambda/λ: hopefully, Intellij can change to λ;
  • ? keyword: to not be confused with _, both have different purposes (well, _ has a tons of them according to where it is used, that’s another story).

Lambda/λ is a generalization of ?. We’ll used the latter to keep it simple. 😳

After “_” we have “?”

// build.sbt
addCompilerPlugin("org.spire-math" %% "kind-projector" % "0.9.4")

As we said, it provides syntactic sugar around advanced typing. For future reference, all these lines are equivalent:

({ type F[E] = Either[E, A] })#F // lambda types and type projection "#F", scalac native
Either[?, A]
Lambda[E => Either[E, A]]
λ[α => Either[α, A]]

Note that I just lied because we can’t use it in type aliases where the lambda types can:

type R[A] = ({ type F[E] = Either[E, A] })#F[_] // OK!
type R[A] = Either[_, A] // OK!

type R[A] = Either[?, A] // BOOM! As we said, it's not a replacement for '_'

The kind-projector is used to type functions return type or through inheritance, we’ll see why.

Usage in functions

Consider this example:

def makeListFunctor = new Functor[List] {
  override def map[A, B](fa: List[A])(f: A => B): List[B] = fa.map(f)
}
makeListFunctor.map(List(1, 2))(_ + 1) // List(2, 3)

It works because List has the kind * -> *: that’s what Functor expects ((* -> *) -> *).

If we want a Functor[Either], it’s different because Either has the kind * -> * -> *: one degree more. We need to reduce it by one. This is where the kind-projector shines (hence its name):

// Left-biased functor for Either (generally, everything is right-biased but we're messing around)
def makeFunctor[U] = new Functor[Either[?, U]] { // Same: new Functor[λ[α => Either[α, U]]]
  override def map[A, B](fa: Either[A, U])(f: A => B): Either[B, U] = fa match {
    case Left(value: A) => f(value).asLeft
    case Right(value: U) => value.asRight
  }
}

val f: Functor[Either[?, String]] = makeFunctor[String]
f.map(Left(5))(_ * 2)                            // Left(10). Works with Int.
f.map(Left("hey"))(_.length)                     // Left(3).  Works with String.
f.map(Left(new Exception("boom")))(_.getMessage) // Left(boom). Anything!

We fix one type and let the other one free, represented by the ? (it’s like the partial-unification system). We chose to fix the right one (that’s why it’s left-biased) by using the type parameter U of the function makeFunctor.

We can create any Functor[Either] with this function. Functor[F[_]] uses the free degree ? to have the kind F[_], and makeFunctor fixes the other degree [U]. That’s how we use the kind-projector.

What’s interesting is the signature vs the implementation of map:

def map[A, B](fa: F[A])(f: A => B): F[B]
=>
override def map[A, B](fa: Either[A, U])(f: A => B): Either[B, U] = ???

F[A] is mapped to Either[A, U]: A is free and U is fixed by the outside (same for F[B]), thus the contract is respected.

Either[A, U] maps to F[A]:

  • A is free to form F[_];
  • U comes from the outside (same for F[B]), thus the kinds contract is respected.

It works with implicit def of course! That’s even one of the biggest usage (see cats or scalaz).

Here, we’ll create a right-biased Functor[Either] automatically at the call-sites of the method f (to avoid some trouble with partial unification, we’ll explain just after):

implicit def makeFunctor[U]: Functor[Either[U, ?]] = new Functor[Either[U, ?]] {
    override def map[A, B](fa: Either[U, A])(f: A => B): Either[U, B] = fa match {
      case Left(value) => Left(value)
      case Right(value) => Right(f(value))
    }
  }

def f[F[_]: Functor, A, B](a: F[A], f: A => B) = implicitly[Functor[F]].map(a)(f)

f(Left(42): Either[Int, String], (_: String).length) // Left(42)
f(Right(42): Either[Error, Int], (_: Int) + 1)       // Right(43)

For the first call, we pass an Either[Int, String] to f asking for a F[A]. What’s happening:

  • through the partial-unification, f is using F[_]=[T]Either[Int, T] and A=String;
  • therefore, it provides us with a Functor[Either[Int, ?]];
  • Int is fixed on left;
  • map expects the free degree ? (represented by the right part of the Either) to be String.

NO to left-biased!

If we still had our left-biased Functor[Either[?, U]], we wouldn’t be able to compile: 😢

  • through the partial-unification, f is (still) using F[_]=[T]Either[Int, T] and A=String;
  • therefore, it provides us with a Functor[Either[Int, ?]];
  • … and it fails! ⛔️ Because our implicit provides only Functor[Either[?, Int]].

The partial-unification can’t find a left-biased instance. We need to hijack the game as we saw earlier:

type R[A] = Either[A, _] // Prevent partial-unification
f(Left(42): R[Int], (_: Int) + 10)            // Left(52)
f(Right(42): R[Error], (_: Error).getMessage) // Right(42)

This is the major pain point of the relation between kind-projector and the partial-unification.

Now we understand better why @djspiewak said “You should always, always order the type parameters in your type constructor from least-to-most specific.”.

We present a way to deal with the left-biased types afterwards. ⬇️

Usage in inheritance

Let’s create some type class we will inherit from:

trait Stringify[F[_]] { // it's like Functor
  def stringify[A](a: F[A]): String
}
// to help us out
implicit class RichF[F[_], A](val a: F[A]) extends AnyVal {
  def stringify(implicit s: Stringify[F]): String = s.stringify(a)
}
IO.pure(42).stringify // Good! "could not find implicit value for parameter s: Stringify[IO]"

Let’s say we want to implement this typeclass for Either. Same problem, * -> * -> * doesn’t fit * -> *! So we fix one type with the outside (the class here) and kind-project the other:

class EitherStringify[T] extends Stringify[Either[T, ?]] {
  override def stringify[A](a: Either[T, A]): String = a.fold(l => s"Left: $l", r => s"Right: $r")
}
// automatically build the instance when needed
implicit def eitherStringify[T]: Stringify[Either[T, ?]] = new EitherStringify[T]

"toto".asRight[Throwable].stringify // "Right: toto"
42.asRight[Throwable].stringify     // "Right: 42"
42.asLeft[String].stringify         // "Left: 42"

The creation of the class was not necessary, we could have used an anonymous one. Here are some more examples:

implicit def eitherStringify[T]: Stringify[Either[T, ?]] = new Stringify[Either[T, ?]] {
  override def stringify[A](a: Either[T, A]): String = a.fold(l => s"Left: $l", r => s"Right: $r")
}

// The classic case here
implicit def listStringify: Stringify[List] = new Stringify[List] {
  override def stringify[A](a: List[A]): String = s"List: ${a.mkString(",")}"
}

Here, we need one more parameter, a Comonad, to be able to run the Free[S[_], A]:

implicit def freeStringify[F[_]: Comonad]: Stringify[Free[F, ?]] = new Stringify[Free[F, ?]] {
  override def stringify[A](a: Free[F, A]): String = s"Free: ${a.run.toString}"
}
Free.liftF(3.some).stringify // Free: 3

Here, we use a Kleisli which requires a Stringify instance of the type it contains and a value A (thus the method is not implicit):

def kleisliStringify[F[_]: Stringify, A](a: A): Stringify[Kleisli[F, A, ?]] =
  new Stringify[Kleisli[F, A, ?]] {
    override def stringify[B](k: Kleisli[F, A, B]): String = s"Kleisli: ${k.run(a).stringify}"
  }

val k: Kleisli[List, String, String] = Kleisli(_.split(" ").toList)
k.stringify(kleisliStringify("hello world")) // Kleisli: List: hello,world

// partial-unification, once again!
def kEither[A]: Kleisli[Either[Throwable, ?], A, A] = Kleisli(_.asRight[Throwable])
kEither.stringify(kleisliStringify("hello world")) // Kleisli: Right: hello world

As we can see, it’s exactly the same principle with inheritance or not. I hope all those examples have triggered your “Eureka”! 😀

Remember that it’s still possible to not use ? in our code, and replace it with the lambda types but…:

implicit def eitherStringify[T]: Stringify[Either[T, ?]] = new Stringify[Either[T, ?]] { ... }
implicit def eitherStringify[T] = new Stringify[({ type F[E] = Either[T, E] })#F] { ... }

Enjoy the partial-unification for left-biased structures

Let’s say we have a left-biased Functor for Tuple3:

implicit def leftBiasedFunctor[B, C] = new Functor[(?, B, C)] {
  override def map[T, U](fa: (T, B, C))(f: T => U): (U, B, C) =
   (f(fa._1), fa._2, fa._3)
}

As we seen, we can’t rely on partial-unification:

(1, "2", true).map(i => i + 10) // does NOT compile.
                                // It unifies as [A](Int, String, A), the Functor is not available

We can apply some trick (see this SO) to reverse the types to enjoy the right-biaised types. Let’s make our code compile!

We need a structure to invert the types: 😈

implicit class InferToLeft[M[_, _, _], A, B, C](a: M[A, B, C]) {
  def lefty: InferToLeft.U[M, C, B, A] = a
}

object InferToLeft {
  type U[M[_, _, _], C, B, A] = M[A, B, C]
}

(1, "2", true).lefty.map(i => i + 10) // (11,2,true) It works!
val lefty: InferToLeft.U[Tuple3, Boolean, String, Int] = (1, "2", true).lefty
lefty.map(i => i + 10)

It works because:

  • U acts like Id (type Id[A] = A), it’s a “virtual wrapper”;
  • InferToLeft.U[Tuple3, Boolean, String, Int] <=> (Int, String, Boolean);
  • it’s messed up (Intellij sees red);
  • the partial-unification resolves: [T]InferToLeft.U[Tuple3, Boolean, String, T] which is equal to [T](T, String, Boolean);
  • therefore it can create the Functor: new Functor[(?, B, C)], brillant! 👏

Mindblow

Note this another trick if we want to do the same with a Bifunctor:

implicit def leftBiasedBifunctor[C]: Bifunctor[(?, ?, C)] = new Bifunctor[(?, ?, C)] {
  override def bimap[A, B, T, U](fab: (A, B, C))(f: A => T, g: B => U): (T, U, C) =
    (f(fab._1), g(fab._2), fab._3)
  }

This won’t work:

(1, "2", true).lefty.bimap(i => i + 10, s => s.length)

We need another variant of lefty:

implicit class InferToLeft[M[_, _, _], A, B, C](a: M[A, B, C]) {
  def lefty: InferToLeft.U[M, C, B, A] = a
  def lefty2: InferToLeft.V[M, C, A, B] = a // C, A, B, not C, B, A
}

object InferToLeft {
  type U[M[_, _, _], C, B, A] = M[A, B, C]
  type V[M[_, _, _], C, A, B] = M[A, B, C] // C, A, B, not C, B, A
}

(1, "2", true).lefty2.bimap(_ + 10, _.length) // it works!

The partial-unification resolves: [T, U]InferToLeft.V[Tuple3, Boolean, T, U] <=> [T, U](T, U, Boolean), the definition of our Bifunctor. 💪

With lefty, it would have resolved [T, U]InferToLeft.V[Tuple3, Boolean, T, U] <=> [T, U](U, T, Boolean), NOT our definition!

Real cases study: cats and scalaz

Now that we understand both concepts, let’s practice on existing code from cats and scalaz.

In cats, the Traverse[F[_]] typeclass has a compose method:

def compose[G[_]: Traverse]: Traverse[λ[α => F[G[α]]]] = ???

It’s using the λ syntax because it’s not possible to use ? here. Why?

Traverse[F[G[?]]] won’t compile: F[[?$2$]G[?$2$]] takes no type parameters, expected: one: it doesn’t fit what’s expected: Traverse[[?]F[G[?]]].

Traverse expects a F[_], but G[?] is interpreted as … G[A]! As the documentation states: ? always binds at the tightest level. Therefore it’s still giving F[G[A]] instead of F[G].

We can correlate this with what is generated when we use one level only:

// one nested: OK!
def f: Traverse[List[?]]
➔ def f: Traverse[[?]List[?]]

// nested of nested using the inner scope: FAIL!
def f: Traverse[List[Future[?]]] // the equivalent of what we saw: Traverse[F[G[?]]]Traverse[List[[?]Future[?]]] // FAIL!

// nested of nested using the outer scope: OK!
def f: Traverse[λ[A => List[Future[A]]]]
➔ def f: Traverse[[A]List[Future[A]]]

Finally, it all makes sense: Traverse.compose composes two Traverse[F[_]] into one (x => traverseF(traverseG(x))). The end type would result in Traverse[Traverse[F[_]]].

Therefore, to fit the type parameter of Traverse[F[_]], compose needs to project the inner Traverse[_[_]] into a simple higher-kinded type _[_]:


A last and more complicated example in appearance, from scalaz:

// Tag.scala
def subst1[G[_], F[_[_]], T](fa: F[G]): F[λ[α => G[α] @@ T]] = ???

This one is part of the “tag” feature of scalaz. It allows to tag a type with another type. We can create different versions from the same type at compile-time without writing extra classes.

sealed trait Kg
val kgTag = Tag.of[Kg]

// will only accept List[Int] tagged with Kg
def process(kgs: List[Int] @@ Kg) = ???

// different level of tagging possible:
       kgTag(List(1, 2, 3)) : List[Int] @@ Kg                // F[_] @@ tag
 kgTag.subst(List(1, 2, 3)) : List[Int @@ Kg]                // F[_ @@ tag]
kgTag.subst1(Functor[List]) : Functor[λ[α => List[α] @@ Kg]] // F[G @@ tag] with G[_]

// our signature is actually this one:
Tag.subst1[Option, Functor, Kg](Functor[Option]) : Functor[λ[α => List[α] @@ Kg]]

Looking at the last line, we can understand the signature of Tag.subst1. It returns a F[G @@ T] where: G has one type parameter G[_] and is tagged with T.

But we must undertand what is a tag: it’s actually a type alias of a structural type 😰:

type @@[T, Tag] = Tagged[T, Tag]
type Tagged[A, T] = {type Tag = T; type Self = A}

If we replace the alias:

def subst1[G[_], F[_[_]], T](fa: F[G]): F[λ[α => Tagged[G[α], T]]] = ???

We stumbled upon exactly the same case as with our cats example!

// look at the return types
def subst1[G[_], F[_[_]], T](fa: F[G]): F[λ[α => Tagged[G[α], T]]]
           def compose[G[_]: Traverse]: Traverse[λ[α => F[G[α]]]]

If we squeeze our eyes enough (ie: in compose, replace F by Tagged and Traverse by F): we get the same case! (ignoring the T) So I don’t need to explain more of it, we already did it. 😁

I hope you enjoyed those examples. At the end, everything looks the same, it’s always the same logic, from a different angle.

Dotty

Dotty will probably implement ?. Type lambdas are already implemented.

For instance, we should write our previous cats example as:

// original: def compose[G[_]: Traverse]: Traverse[λ[α => F[G[α]]]] = ???
def compose[G[_]: Traverse]: Traverse[[α] => F[G[α]]] = ???

Conclusion

The partial-unification works with higher-kinded types and allows us to pass more complex types (such as Either[_,_]) to functions (expecting a F[_] for instance). The compiler will always favor right-biased type constructors by leaving the right-most types free (ignoring the left-most). We can use the InferToLeft trick to revert the types for the partial-unification to work with the left-biased types. In Scala 2.12.6, it’s enabled by default and the flag is gone. Until then: scalacOptions += "-Ypartial-unification". 😩

The kind-projector allows us to declare free types in higher-kinded types to “fit” their container, using ? or λ for more complex cases. ? is NOT _. Always favor right-most free types to avoid partial-unification issues. Cats and scalaz relies a LOT on that, but as we saw, it’s not that complicated to decrypt. Read the documentation of the kind-projector, it goes deeper than what we saw here (such as the polymorphic lambdas).

Don’t be scared of Intellij Idea: it’s often all red but it compiles. 😁

Wait for Dotty/Scala3. 🙏 The documentation of Dotty contains tons of interesting insights about types and is quite accessible. For instance, about the Kind polymorphism here (expand the menus on the left).