Understanding Types, Subtypes and Type Variance

This is a cross-post from my blog. You can find the original here.

Most of us when we just start with programming, get confused with concepts of types, subtypes, inheritance, etc. But just as with the process of learning any language this all starts to come naturally to us, we grow up to casually use subtyping wherever we want in a program. In fact, it becomes almost unimaginable to code without these.

Subtyping is nothing but writing one type in terms of another. Only when I started exploring other languages I realized that there are deeper theories to it, and maybe I do not understand subtypes to the extent I thought I did. I am writing this post in order to pass on this programming existential crisis to my fellow readers (but also attempting to develop an understanding along the way). We shall take subtyping a level further and understand how the assignment and substitution behaviour of our types affects our code behaviour (and subsequently ours).

Let's begin with an example. Let's say we have a class Animal. It can speak and eat.

class Animal(val name: String){
  def speak = "animal speaking"
  def eat = "animal eats"
  override def toString = "animal: " + name
}

While Animal has general properties, it can also be of two types: Cat and Dog. Cat and Dog speak, eat in their own way. Also, a dog has its own special function as well, dogonly. Similarly, a cat is special in its own way, say catonly.

class Cat(override val name: String) extends Animal(name) {
  override def speak = "meow"
  override def eat = "fish"
  def catonly : String = "cat only"
}

class Dog(override val name: String) extends Animal(name) {
  override def speak = "woof"
  override def eat = "bone"
  def dogonly : String = "dog only"
}

(I do not have experience with pets, so pardon me if I did not implement Dog's and Cat's functions correctly...)

With the above examples in mind, let's go over some concepts formally.

Type

A type or a data type represents the type of data that tells the compiler/interpreter as to how the programmer intends to use that data. So, a type can be Integer, String, Boolean. We can create our own types as well, such as Animal, Cat and Dog as above. A type, the way it is defined, indicates what values it can take and what operations can be done on it.

Subtyping

Subtyping talks about the relationship between different types. We have a subtype and we have a supertype.

We can say that a type B is a subtype of A if:

• we can substitute B subtype in place of A supertype.
• we can perform all the operations on B that can be performed on A.

This is often written as A :> B, meaning that an instance of type B can be safely used wherever type A is expected.

Let's be very clear on this as this will be the basis of our article:

If Animal :> Cat, then we can use Cat wherever we expect an Animal.

So, if we have a list that expects objects of Animal type:

val animal1 = new Animal("animal1")
val animal2 = new Animal("animal2")
val animals : List[Animal] = List(animal1, animal2)

Then I can add a Cat type and a Dog type object to it as well:

val cat1 = new Cat("cat1")
val dog1 = new Dog("dog1")
val allTypesOfAnimals : List[Animal] = cat1 :: dog1 :: animals // this works as expected

But what about the reverse case? Does this relation hold true when the directions are reversed?
Let's try passing an object of type A where B is expected when A :> B:

val cats : List[Cat] = List(cat1, cat2)
val brokenCats : List[Cat] = animal1 :: cats // error

Using the above example, it becomes clear that while we can add a Cat to a List[Animal], we cannot add an Animal to a List[Cat]. The intuition behind this is that a subtype is special in some sense, and a supertype cannot replace it. Here, if we were to successfully pass Animal to a List[Cat], it would break in case we call catonly() on List[Cat].

Working in OOP based languages makes this sound kinda obvious as we think about the same concept it in terms of Inheritance: an object of subclass cannot be passed where an object of superclass is required.

Subtyping v/s Inheritance

At this point you may ask me, "Amita, why are you confusing us with subtyping and inheritance? Aren't we talking about inheritance here?" Well, actually, yes and no. I found this article here that can help:

Subtyping refers to compatibility of interfaces. A type B is a subtype of A if every function that can be invoked on an object of type A can also be invoked on an object of type B.

Inheritance refers to reuse of implementations. A type B inherits from another type A if some functions for B are written in terms of functions of A.

In OOP, inheritance and subtyping usually go hand in hand, the syntax being the same makes it more confusing. For example, in our case, Cat and Dog are subtypes of Animal: any operations done on Animal can be done on Cat and Dog as well. Also, Cat and Dog inherit from Animal: we have redefined (overridden) the function in Animal in Dog and Cat.

Inheritance helps you reuse the implementations of methods/instance variables inside the super class, on the other hand, Subtyping deals with the safe behaviour of this class (a type) on the whole when it is passed to another data structure or function (formally called Variance, which we talk about later).

Subclassing doesn't guarantee Subtyping.

When we talk about subtyping here, we don't discuss reusing implementations, but actually delve into the behaviours of types in different conditions.

Variance - Subtyping for complex types

Now that we have established how subtypes and supertypes are used and how they behave, let's take it a bit further. What about complex types?

Complex Types

Complex types are types that are composed of other types. e.g. Lists, Maps, Options, Functions etc... A List alone has no meaning. They are to be used as a list of some type, say List[Integer], List[Animal] etc. Similarly, a function has a type signature, say, func (a : Integer, b : String) : Boolean. Here a function func has the type signature as (Integer -> String) -> Boolean.

Variance defined formally is

Variance is the correlation of subtyping relationships of complex types and the subtyping relationships of their component types.

Subtyping with Lists: An instance of Covariance

We are going to continue with our example above to understand this concept step by step. Let's take a complex type, a List. If we have types A and B such that A:>B, how do we relate List[A] to a List[B]?

Reiterating on what we established above, if A:>B, we can pass B wherever we expect A.

Subtyping on complex types gives rise to 2 cases:

Case 1: pass List[B] where we expect List[A]

I have a function that expects a List[Animal] and does something with it:

def expectingListOfSupertype(animals : List[Animal]): Unit ={
  println(animals.map(_.toString))
}

Can I pass cats: List[Cat] in place of animals: List[Animal]?

expectingListOfSupertype(cats) // yes, I can!

Case 2: pass List[A] where we expect List[B]

This is the converse case to case 1.

def expectingListOfSubtype(cats: List[Cat]): Unit = {
  println(cats.map(_.catonly))
}

expectingListOfSubtype(animals) //type mismatch compile time error

No, I can't. This makes sense because if we could pass List[Animal] to the function where List[Cat] is expected, it could fail on calling the specific function catonly on an animal instance.

Hence, in case of Lists, we can substitute List[B] in place of List[B] when A:>B but not the opposite.

This proves a direct relationship:

if A:>B, then List[A]:>List[B]

The above phenomenon is called Covariance.

Subtyping with Functions: An instance of Contravariance

Let's take another example of a complex type Function. If we have types A and B such that A:>B, how do we relate Function[A] to a Function[B]?

Reiterating our subtyping relation of component types A and B, if A:>B, we can pass B wherever we expect A. Can this relation hold true for complex type Functions as well?

Let's define a function with supertype, and another function that calls this function.

val superTypeFunction : Animal => String = animal => s"${animal.name} is here"

def functionThatExpectsSupertypeFunction(fun: Animal => String) = {
    // calls fun somewhere
}

Similarly, we define a function with subtype, and another function that calls this function

val subTypeFunction : Cat => String = cat => s"${cat.name} cat is here and has special ${cat.catonly} behaviour"

def functionThatExpectsSubtypeFunction(fun: Cat => String) = {
    // calls fun somewhere
}

How are the functions going to behave in the following cases?

Case1: pass f[B] where we expect f[A]

Lets try passing the subtype function to functionThatExpectsSupertypeFunction:

functionThatExpectsSupertypeFunction(subTypeFunction) // type mismatch!

The above scenario in case of functions doesn't compile. But why?

Let's consider the below implementation of functionThatExpectsSupertypeFunction:

def functionThatExpectsSupertypeFunction(fun: Animal => String) = {
  val fish = new Animal("nemo")
  println(fun(fish))
}

If the compiler could allow passing subtype function to the above, it would break coz internally it would call fun: Cat => String which can then try to call catonly on an Animal. Ouch!

Case2: pass f[A] where we expect f[B]

Now let's try passing supertype function to functionThatExpectsSubtypeFunction:

functionThatExpectsSubtypeFunction(superTypeFunction) // compiles!

On the other hand, we can pass a function of supertype where subtype is expected.

The intuition can be understood by thinking as to how this works in an example implementation. Suppose I implement functionThatExpectsSubtypeFunction like this:

def functionThatExpectsSubtypeFunction(fun: Cat => String) = {
  val cat = new Cat("kitty")
  println(fun(cat))
}

In the above scenario, even if we pass an Animal => String function to fun parameter, it always supplies the subtype is supplied to it. And by definition of subtyping, Cat can be used wherever Animal is expected. Hence this works!

You can rightly observe, that Functions have the opposite subtyping behaviour as compared to Lists.

if A:>B, then Function[B]:>Function[A]

This is called Contravariance.

Subtyping with Arrays: An instance of Invariance

Let's talk about another data structure: Arrays. These are mutable data structures, so for an Array[Animal] we can write Array[Cat] anytime. At the same time while reading we can expect Array[i] can be an instance of Animal, Cat or Dog...

Thus, we can say that it can neither be safe for arrays to be Covariant or Contravariant. Confused? Wikipedia explains it better:

If we wish to avoid type errors, then only the third choice is safe. Clearly, not every Animal[] can be treated as if it were a Cat[], since a client reading from the array will expect a Cat, but an Animal[] may contain e.g. a Dog. So the contravariant rule is not safe.

Conversely, a Cat[] cannot be treated as an Animal[]. It should always be possible to put a Dog into an Animal[]. With covariant arrays this cannot be guaranteed to be safe, since the backing store might actually be an array of cats. So the covariant rule is also not safe—the array constructor should be invariant.

val animalarray: Array[Animal] = Array(animal1, animal2)
val catarray: Array[Cat] = Array(cat1)

def expectingArrayOfSupertype(animals: Array[Animal])= {
  // some read/write operations
}
def expectingArrayOfSubtype(cats: Array[Cat]) = {
  // some read/write operations
}

expectingArrayOfSupertype(catarray) // type mismatch!!
expectingArrayOfSubtype(animalarray) // type mismatch!!

Both our cases fail for Arrays, they being Invariant.

Special case of Variance of Arrays in Java

For Java / C# developers, it is worth noting that Arrays are covariant. The reason is in detail here. Hint: it's because Generics was included much later in the language than Arrays...

Rule of Thumb

Well, at this point it can be too much to digest. But there can be a simple rule of thumb to it:

• Read-only data types/sources can be covariant, e.g. immutable Lists
• write-only data types/sinks can be contravariant, e.g. functions
• Mutable data types which act as both sources and sinks should be invariant. e.g. Arrays

Contravariance is quite unintuitive in nature, as it is difficult to imagine a type with just write-only capabilities. In case of functions(the case where we're only talking about it accepting a type once), we can think that it took a subtype, but did not return it, hence making it 'write-only'. Usually, a complex type that accepts a simple type and simply consumes it, can be treated as contravariant.

Try this code

This whole post makes more sense on running the code snippets. Checkout this gist which contains the entire code. I would recommend running them one by one and observe the magic of types yourself!

Further Reading

• This Wikipedia article is my starting point, and the Animal example is inspired from here.
• Liskov's Substitution Principle
• SubTypes v/s Subclasses
• Subtyping vs Inheritance
• Scala Docs: Variance