Implementing typeclasses in Scala I

I’ve been learning Scala recently and have been very impressed with the power of some of its features as well as the way in which it seamlessly combines object-oriented and functional programming. Nevertheless, some of these features can seem a bit tricky and learning resources can be relatively hard to find. In particular, I had some difficulty wrapping my head around implicits. Things came together when I began to ask myself: what are the analogs of Haskell’s typeclasses in Scala?

It turns out that, unlike in Haskell, typeclasses are not built-in to the language; rather, they can be implemented using implicits. Because of this, they’re actually more powerful than in Haskell: you can, for example, create multiple coexisting typeclass instances and pass them around as objects. Moreover, due to Scala’s object-oriented nature, typeclasses can inherit from one another.

Note that there already exist some Scala libraries, such as Scalaz and Cats, that develop many useful typeclasses, but for the purpose of understanding how this is done I think it’s best to start from scratch.

Note. I’ll make use of some basic Scala features. Implicits are the only feature I’ll really explain (I also briefly explain “options”). Most of the others are similar to ones found in Java (traits are like interfaces, objects replace the static keyword, pattern matching is somewhat like the switch statement). In the later posts I’ll implement functors and monads, but these won’t be discussed here.

I may also do things in slightly non-idiomatic ways at times. I’m also not an expert on programming languages. If you find anything wrong or that could be improved in this post, please let me know.


In this post, I’ll start by discussing typeclasses, options, and implicits. Then I’ll introduce Measurable, a made-up typeclass that we’ll work with, and show how to make ordinary types like Int and String members of it. This will lead to a discussion of context bounds, which will allow us to add type constraints and implement an instance of Option for Measurable. I’ll wrap up with a brief discussion of implicit classes.

Why typeclasses?

Rather than giving a general explanation of what typeclasses are, I’ll try to explain some of the basic ideas in the examples that follow. Much better introductions to typeclasses than I could write can easily be found online. All I’ll say here is that typeclasses are “like (Java) interfaces” but much more general and powerful. In particular, we’ll see the following examples of what typeclasses can do that interfaces can’t:

  1. Types can be made members of typeclasses outside of their own definitions; this means, for example, that we can make built-in types members of typeclasses.
  2. Typeclass instantiation of parameterized types can be specialized depending on the type parameter (“ad hoc polymorphism”).
  3. So-called “higher-kinded” types can be made members of typeclasses.


I’ll be using options extensively as examples, so I’ll briefly explain what they are: Options are basically a type-safe replacement for null. For any type A, the parameterized type Option[A] can take on either the value None or the value Some(x) for any value x of type A. For example, you could use options to write a function that reads user input and expects an integer like this.

def getInteger(): Option[Int] = {
  val s = StdIn.readLine()
  try {
  } catch {
    case _: NumberFormatException => None

Here None is used as a type-safe way of signalling that the user did not input a number. This is different from returning null because we’re actually returning an object and avoiding the possibility of a NullPointerException.

Scala implicits

Implicits obviate the need to (explicitly) cast or convert objects from one type to another. In fact, this is how Scala deals with different numerical types: when you pass, say, an Int to a function of a Double, Scala performs an implicit conversion (note that Int is not a subtype of Double).

There are several ways to define implicit conversions. Rather than provide an exhaustive list, I’ll sprinkle these throughout the post. Here’s one way to do it.

Implicit conversion I. Define an implicit function or method that performs the implicit conversion

Suppose we wanted to pass an ordinary (non-option) value a to a function f of an option and have it automatically understood as Some(x). Here’s what we’d do.

implicit def any2option[A](x: A): Option[A] = Some(x)

def f[A](option: Option[A]): Unit = option match {
  case None => println("None")
  case Some(x) => println(x)

Now (as long as any2option is in scope) we can call, for example, f(10), which will print “10” to the screen. When the compiler encounters our function call, it notices that 10 is not of the right type to be passed to f. In most statically-typed languages, this would immediately lead to a compilation error. The Scala compiler, however, searches within the “implicit scope” for an implicit conversion; in this case, it finds any2option, which can convert an Int to an Option[Int], which is of the right type to be passed to f. The compiler then replaces our ordinary function call f(10) by f(any2option(10)).

Note that if we called f(Some(10)), the compiler would still print “10” to the screen (rather than “Some(10)”) because an implicit conversion is not necessary (hence not used) in this case.

The fact that any2option works with parameters of any type is referred to as parametric polymorphism. This isn’t especially relevant right now, but will be good to know later. This is a great term if you want to sound smart, because it’s made up of two words of Greek origin.

Beginning to define a typeclass

I decided to start with a very simple, made-up typeclass Measurable. Members of Measurable have values that can be “measured” in some way. It doesn’t really matter how but some ideas that come to mind are: the measure of a numeric value could be its absolute value; the measure of a string could be its length; the measure of a list of measurable values could be the sum of their measures; and so on. We begin our definition with a trait declaration.

// Measurable.scala

trait Measurable[A] {
  def measure(x: A): Double

In addition to this, we’ll also define a “law” for Measurable. This is a contract that clients of measurable are expected to obey but that cannot be enforced by the type system. Numerous common typeclasses derived from mathematics, such as functors and monads, have laws of this kind. Our Measurable law will be the following.

Measurable law. The measure function must produce non-negative values.

This requirement cannot be expressed, at least in plain Scala, because it requires refinement types (however, there does appear to be a library for refinement types in Scala).

Instantiating typeclass members

So far that was pretty simply. All we did was define a trait with a single method. Now you might complain: Traits are just Scala’s versions of interfaces. So how is a typeclass different from an interface? Well, for starters, we’re actually not done with the typeclass definition. But more importantly, typeclasses are different in how we use them. Rather than extending them, we’ll implement them implicitly.

Suppose we wanted to make Int a member of Measurable. Since Int is already defined elsewhere, we can’t change it. However, we can define a separate object that realizes the instance of Measurable we would like.

val int2measure: Measurable[Int] = x => Math.abs(x)

Syntactic sugar. Here we’ve used a bit of syntactic sugar. Since Measurable contains only one abstract method, we can define an anonymous class implementing it just by specifying how that method is implemented. Moreover, this method can be given as an anonymous function (because its name, measure, is already determined). That’s why the right-hand side looks like an anonymous function (with the same signature as measure when A = Int). We’ve also used Scala’s string interpolation.

This is all well and good, but how does it help us? Let’s remind ourselves of what we’re trying to accomplish. We’d like to be able to pass an Int to Measurable.measure. The latter is abstract, but we’d like it to be implicitly defined by int2measure.measure. We could call the latter directly, but this isn’t really satisfactory. For example, suppose we implemented another implicit conversion string2measure: Measurable[String]. Now we’d have to call string2measure.measure("hello"). We’d like to be able to unify the measure methods of int2measure, string2measure, and so on into a single function.

Ad hoc polymorphism

What we want is a polymorphic function measure[A] that works for any A for which an instance of Measurable[A] exists. This is not parametric polymorphism, because the type A is constrained. We need some way to express this constraint. This is where the notion of an implicit parameter comes in.

Implicit conversion II. Add an implicit parameter to the type signature of a function that may require an implicit conversion.

Let’s define our measure function and then explain how it works. Note that we’ve placed this function in the companion object for Measurable.

// Measurable.scala // ...

object Measurable {
  def measure[A](x: A)(implicit measureInstance: Measurable[A]): Double =

Even though measure is generic, it takes an implicit parameter of type Measurable[A]. Because the parameter is implicit, we don’t have to explicitly pass in a value for it. Rather, the compiler will look for an implicit object in the current scope that it can use, depending on the type parameter A. For instance, we would like it to be able to find int2measure, so we make that value implicit.

// MeasurableClient.scala

import Measurable._

object MeasurableInt {
  implicit val int2measure: Measurable[Int] = x => Math.abs(x)

Now if we call measure(10) (with int2measure in scope), the compiler will automatically find int2measure and pass it in as the value of measureInstance; thus, measure(10) will become measure(10)(int2measure), which evaluates to int2measure.measure(10). Similarly, measure("hello") would become string2measure.measure("hello").

In a way, measure is not fully generic even though it appears to be declared generically. Rather, measure[A](x) will only work for types A for which an instance Measurable[A] exists. This is the basic idea of ad hoc polymorphism. This is an even better term than parametric polymorphism for sounding smart, because it’s made up of of a Latin phrase together with a word of Greek origin.

Enforcing laws. At this point, we could implement the Measurable law using an assert statement as follows.

object Measurable {
  def measure[A: Measurable](x: A): Double = {
    val result = implicitly[Measurable[A]].measure(a)
    assert(result >= 0)

However, this won’t work for other typeclasses.

Defining a typeclass

Putting together what we saw above, here’s how we define the Measurable typeclass.

// Measurable.scala

trait Measurable[A] {
  def measure(x: A): Double

object Measurable {
  def measure[A](x: A)(implicit measureInstance: Measurable[A]): Double =

To summarize:

  • A typeclass is implemented using a trait together with a companion object whose methods implicitly implement the trait’s methods via ad hoc polymorphism.
  • An instance of a typeclass is an implicit instantiation of the trait in the typeclass’s implementation.
  • A member of a typeclass is any type with an instance of the typeclass.

Context bounds

At this point we’re pretty much done. We’ve implemented the Measurable typeclass and demonstrated how to make other types members of it. However, there’s a few things we get “for free” at this point that are worth adding.

The type signature of measure says that “A must have a type that implements the Measurable typeclass”. This precise idea can be expressed using the following syntactic sugar supported by Scala: instead of declaring measure as above, we write def measure[A: Measurable](x: A). The constraint A: Measurable is called a context bound.

Now if you try making that replacement, you’ll encounter an issue on the right-hand side: measureInstance is no longer bound to anything. However, if A does have a Measurable instance (represented by an implementation of Measurable[A]), an implicit parameter will be passed to measure and can be retrieved within its scope by referring to implicitly[Measurable[A]]. Thus, the definition of measure becomes the following.

// Measurable.scala
// ...

object Measurable {
  def measure[A: Measurable](x: A): Double =

This is not only cleaner, but better captures the fact that measure uses ad hoc rather than parametric polymorphism.

Class constraints

Equipped with our knowledge of context bounds, we can now also make Option[A] a member of Measurable… that is, so long as A itself is a member of Measurable. With context bounds, this constraint is incredibly easy to express! We just need something like int2measure, but polymorphic.

def option2measure[A: Measurable](x: A): Measurable[Option[A]] = {
  case None => 0
  case Some(x) => measure(x)

Syntactic sugar. In addition to specifying an anonymous instance of a single-method trait by an anonymous method (as we did with int2measure), we’re also specifying this anonymous method by an anonymous pattern match.

Type variance

There’s still an issue we should address. If we set val x: Option[Int] = Some(10) and call measure(x), then everything works fine. But measure(Some(10)) doesn’t compile. That’s because Some(10) has type Some[Int], which isn’t the same as Option[Int] (although it is a subtype).

We need measure[Some[Int]] to accept an implicit parameter of type Measurable[Option[Int]]. More generally, we need measure[A] to accept an implicit parameter whose type is Measurable[B] where B is any supertype of A. This can be accomplished if Measurable[B] <: Measurable[A] whenever A <: B. This is known as contravariance of the type parameter to Measurable.

In general in Scala, a relationship between A and B doesn’t imply any relationship between corresponding parameterized types (there are good reasons for this). However, a contravariance relation can be imposed by prefixing a type parameter with - and, similarly, a covariance relationship by prefixing with a +. In this case, we’ll change our Measurable trait as follows.

// Measurable.scala

trait Measurable[-A]

Implicit classes

One more thing we can do is use implicit classes to automatically convert objects that are members of Measurable to objects with a measure method.

Implicit conversion III. Define an implicit class. Such a class can implicitly convert from the type of its constructor parameter to the class that it itself defines.

// Measurable.scala

object Measurable {
  implicit class MeasurableOps[A: Measurable](x: A) {
    def measure: Double = Measurable.measure(x)

Now we can call 10.measure or Some(10).measure and it will behave as expected! In a way, we’ve “implicitly” extended the Int class and Some[A] (for A: Measurable) classes.

Let’s think about how this works. When 10.measure is called, the compiler looks for a measure method in the Int class. It doesn’t find one, so it looks in the for an implicitly defined type that has a measure method and to which Int can be converted. It finds the implicit class MeasurableOps and sees that MeasurableOps requires a parameter of a type A that is a member of Measurable. Since A = Int, the compiler looks for an implicit conversion of Int to Measurable and finds int2measure (One sometimes says that the type checker has proved that Int is a member of Measurable, i.e. that the requirement Int: Measurable can be satisfied). It passes 10 explicitly and int2measure implicitly to the constructor of MeasurableOps, which returns a new MeasurableOps object equipped with a method measure. Since implicitly[Measurable[A]] is bound to the implicit constructor parameter int2measure, the measure method of this new object calls int2measure.measure(10).

What’s next?

In the next post I’ll discuss higher-kinded types and the Functor and Monad typeclasses. I’ll also introduce a tree-like data structure to test these ideas on.

4 thoughts on “Implementing typeclasses in Scala I”

  1. Thank you for this great post about typeclasses from the ground up. Very useful indeed!

    I am stuck after defining the trait Measureabe as contravariant (“trait Measureable[-A] { … }”):
    However, even if I then run “Measurable.measure(Some(-10))” for example (on the Measurable object), I still get:
    “error: could not find implicit value for parameter m: typeclasses.Measurable[Some[Int]]”
    Can you be more specific abut the contravariant setup need to make using “Some(10)” as parameter possible or explain me what else I am missing to change?

    // Measureable.scala
    trait Measurable[-A] {
    def measure(x: A): Double
    object Measurable {
    def measure[A: Measurable](a: A): Double = implicitly[Measurable[A]].measure(a)
    // Usage in REPL
    import typeclasses.Measurable
    implicit val int2measure: Measurable[Int] = _.abs
    Measurable.measure(-10) // works fine…
    implicit def option2measure[A: Measurable](x: A): Measurable[Option[A]] = {
    case None => 0
    case Some(x) => Measurable.measure(x)
    Measurable.measure(Some(-10)) // error…


      1. Indeed helped – it turns out that having the code in a file and importing it into a worksheet somehow is the problem. If I put the code all into the worksheet (incl. the trait definition and the object), then it runs. Worksheets are weird…? (And I am too lazy investigating why right now ;-)) Thanks for helping me resolve this puzzle.

        Liked by 1 person

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s