Structural Typing vs. Duck Typing

Introduction

A while ago, I explained why I thought that Duck Typing was dangerous. More recently, Scala popularized a different type of typing called Structural Typing, which is often described as being "type safe Duck Typing". Let's take a closer look at Structural Typing and see if it delivers on this promise.

With Duck Typing, you send a message to an object in absolute darkness. You don't know whether this object knows this message and you just trust the caller to have passed you an object that does:

def test(o)
  log o.getName   # Let's hope this will work
end

The code above might be broken, but you will only find out at runtime.

Structural typing allows you to be a bit more explicit about your expectations on this object. In the following Scala example, I declare that the object passed in parameter should have at least one method, called getName, that should return a String:

def test(f: { def getName(): String }) {
  log(f.getName)
}

This doesn't seem like much, but it does buy us a lot of type safety. For example, if I try to pass an object that doesn't define a getName method, the Scala compiler complains:

type mismatch;
 found   : Test
 required: AnyRef{def getName(): String}

I get the same error message if I try to pass an object that has a getName method that returns an int instead of a String.

From that standpoint, Structural Typing is indeed superior to Duck Typing and one of my major objections to Duck Typing ("it's not type safe") goes away. Structural Typing doesn't solve everything, though. Here is a quick side by side comparison of the main techniques available:

1. I'm conflating classes, interfaces and traits. They have different semantics, but mean the same in the context of typing.
2. More details on the impossibility to refactor Duck Typed code can be found in this blog entry.
3. Any attempt to rename the getName method in either of the three locations (inside the class for the instance passed in parameter, in the method signature or inside the method body) will cause all the other locations to be properly renamed as well.
4. Since you declare no types at all with duck typing, there is no repetition.
5. If you need to declare more than one method that accepts a parameter with a getName method on it, you will have to repeat the entire (f: { def getName(): String } statement for each one.
6. Since you declare exactly one type with a Class, there is no repetition.

As you can see, Structural Typing is doing pretty well and the only problem is with the violation of the DRY principle that it requires as soon as you need to use it more than once for a signature. This can be easily avoided if you simply avoid using Structural Typing in this case and instead, encapsulate the method in a Class/Interface/Trait.

Bad

def test(f: { def getName(): String }) {
  log(f.getName)
}

def toXml(f: { def getName(): String }) {
  log("" + f.getName + "");
}

The repetition notwithstanding, the problem with the code above is that renaming one of the getName methods will not cause the other one to be automatically renamed as well since the compiler has no way of knowing that these methods are identical. In some way, it's ironic that the so-called Structural Typing doesn't preserve... the structure of the type we're actually passing.

Better

trait HasName {
  def getName : String = { ... }
}

def test(HasName f) = {
  log(f.getName)
}

def toXml(HasName f) = {
  log("<log>" + f.getName + "</log>");
}

So, Structural Typing... Good or Bad?

I have mixed feelings.

Structural typings can be handy for occasional pieces of code that might not be worth creating a Class/Trait/Interface for, but as soon as you want to reuse the signature more than once or if that signature contains more than one method, you are better off creating a carefully named type that captures that signature and use it instead.

What do you think?