From Interfaces to Traits

This is a post about how different languages handle the concept of interfaces. We’ll go over the classical OO way of handling them, with Java, to the more recent approaches of languages like Rust, as well as others in between.

Why do we want interfaces?

The problem interfaces address is polymorphism. Polymorphic code can work with different types of things in a flexible way. In practice this means functions that can accept different types, and work differently based on those types.

For example, we might have a function that can print out results to a file or directly to the terminal.

Polymorphism lets us write more reusable code. One function can operate on many types without having to be rewritten for each of them.

Functions that use only an interface, instead of all the implementation details of a type, are easier to understand. For example, instead of depending on all of a file’s operations, we only care about using it as another output stream. This allows our functions to worry about less things, and get closer to having a single responsibility.

Classical Interfaces

Interfaces in Java are a variant of inheritance, so let’s look over how that works first.

Inheritance

Java has classes, and these have methods. For example:

class Rectangle {
  private int height;
  private int width;

  public Rectangle(height, width) {
    this.height = height;
    this.width = width;
  }

  public int area() {
    return height * width;
  }
}

This is a class, with two fields, a constructor, and a method that makes use of both of those fields.

We can create a new instance of this class, and call its methods as follows:

var rectangle = new Rectangle(3, 4);
System.out.println(rectangle.area());

This program prints out 12, because we call this rectangle’s methods using the data contained in its fields.

We can also make new classes which inherit from another; for example:

class Square extends Rectangle {
  public Square(width) {
    super(width, width);
  }
}

This class can be used like this:

var square = new Square(3);
System.out.println(square.area());

This program prints out 9.

The Square class inherits all the methods and their implementations from its parent class Rectangle, and can use its parent’s constructor. This means that it already has all of Rectangle’s methods and their implementations from the start.

Classes can also change the implementation of certain methods. This is called overriding in Java.

  @Override
  public int height() {
    // new behavior
  }

Abstract Classes

Java also has a feature called abstract classes.

Abstract classes have one big difference from normal classes: they can choose not to provide an implementation for a given method.

For example:

abstract class Shape {
  abstract int height();

  int heightSquared() {
    var h = height();
    return h * h;
  }
}

We’ve left the height method abstract. We can’t actually create instances of the Shape class. Instead, we need to extend the class with another, and then we can create instances of that subclass.

Interfaces

Now that we’ve seen classes, and then abstract classes, we can move on to interfaces, as implemented in Java.

An interface is essentially an abstract class, where all the methods are abstract.

For example:

interface ShapeLike {
  int area();
}

We can then have different classes that implement this interface:

class Rectangle implements ShapeLike {
  int area() {
    return width * height;
  }
}

class Square implements ShapeLike {
  int area() {
    return width * width;
  }
}

This can be used for polymorphism, by declaring a function that accepts an interface instead of a specific type:

class ShapeUtils {
   static int areaSquared(ShapeLike shape) {
     var a = shape.area();
     return a * a;
   }
}

(We make a class with a static method because Java doesn’t like free functions).

One key thing to notice here is that each class has to explicitly declare that it implements a given interface. There’s no way to make an old class implement a new interface.

Java has many other ways of implementing polymorphism through inheritence, from subclassing to abstract classes to interfaces. All of these have allow a function to accept a given type without knowing whether that argument is of that exact type, or a given subtype. When accepting an interface, a function can only use the methods that interface provides, and is oblivious to the other details the various classes implementing that interface may have.

Middle Ground: Go

The main difference between Go and Java is that in Go, implementing an interface is implicit, whereas in Java, this is explicit.

Continuing with our geometry examples, in Go we might have code that looks like this:

package main

import "fmt"

type Shape interface {
  area() int
}

type Square struct {
  width int
}

func (s Square) area() int {
  return s.width * s.width
}

func main() {
  var s Shape
  s = Square{width: 3}
  fmt.Println(s.area())
}

(This is actually a complete Go program that can be run, and prints out 9)

The first part of this program declares a new interface type, named Shape. This interface is defined by the method area. With the way interfaces work in Go, any type that has a method named area with the right type signature can be used as that interface. Later on in the program, we assign a value of type Square to a variable of type Shape. This is allowed because Square has a method with the right name and types.

One downside of Java’s interfaces is that old types cannot implement new interfaces. If we notice a behavior we want to abstract over, we can’t make it work with existing types. Because Go has implicit interfaces, if we notice that multiple types already have a given method, we can abstract over that. We can create a function that accepts any type that has a given set of methods by using interfaces.

Rust’s Traits

Rust’s version of interfaces is called traits. These traits function quite similarly to Haskell’s typeclasses.

The main difference between Rust and the other two examples we’ve seen so far is that in Rust, traits are implemented explicitly, but can be implemented for existing types.

Let’s look at an example:

trait Shape {
  fn area(&self) -> i32;
}

struct Square {
  width: i32
}

impl Shape for Square {
   fn area(&self) -> i32 {
     width * width
   }
}

We’ve seen a similar example a few times. This is very similar to the Go version, except that instead of implementing an area method in the struct (which we can do in Rust), and then having the interface implementation be implicit, we have to explicitly implement that interface.

One advantage of explicit implementation is that the Rust compiler can warn us if we incorrectly implement an interface. In Java a similar thing happens as well.

Because the implementation block is separate from the declaration of the type itself, we can implement a trait for a type that already exists. This is very useful, because we can identify abstractions at any time.

There are two “guidelines” of sorts that constrain this a little bit. We should try and put a trait implementation either

For existing types, we have to put the implementation next to the trait, since we don’t have access to the original file.

For types we think up after having created the interface, we can implement that interface next to the type.

Interface Matrix

We can divide up the design space we’ve gone over so far like so:

Language Existing Types Explicit
Java No Yes
Go Yes No
Rust Yes Yes

I don’t know of a language with some kind of interface construct which doesn’t work with existing types, but has implicit implementation. I think that this may not even be possible. If the implementation is implicit, then it will pick up the methods that already exist for certain types.

My Opinions

Now we come to the meat of the post: why I think Rust’s position in the design space is the best.

Traits work with existing types

This is the most important aspect of traits in my view.

By being able to implement new traits for old types, you can discover and work on abstractions independently of other types. This allows you to work much more flexibly with other people’s code, since you can implement your own abstraction layer without having to write a bunch of wrapper types.

Traits are still explicit

In general, I prefer explicit behavior to implicit behavior.

One advantage of trait implementations being explicit is that it’s easy to tell that a type implements a trait correctly. In Rust, if your implementation uses the wrong name or method type, then you’ll catch it then and there. In Go, you’ll only catch this once you try and assign this type to a given interface.

The two guidelines for where trait implementations should lie keeps things much more orderly. It’s very easy to know what existing types a trait decided to provide implementations for, since they all have to be located next to the trait itself.

Summary

Different languages have different ways of implementing interface-like concepts. The main axes in the design space are explicit vs implicit implementation, and working with existing types or not.

Rust’s position of working with existing types in an explicit way is the best, in my opinion.