Let's say you've got a class called Car. There are two subclasses of Car: Convertible and Sedan. And it turns out that all cars are either convertibles or sedans. (Who knew?) So really, there's no reason that a car object wouldn't be an instance of Convertible or Sedan, and in fact an object that's a direct instance of Car itself won't even work correctly.

class Car
  def go_forward
    # ...
  end
end

class Convertible < Car
  def door_count
    2
  end
end

class Sedan < Car
  def door_count
    4
  end
end

How would you write Car#doors? You wouldn't, because unlike moving forward, that behavior isn't shared across all cars. Car is an abstract class: a class that should never be instantiated directly.

But what's stopping us? Nothing. And that's a problem. So let's fix it:

$ sudo gem install abstraction

Then:

require 'abstraction'

class Car
  abstract
  
  def go_forward
    # ...
  end
end

Car.new
#> AbstractClassError: Car is an abstract class and cannot be instantiated

But:

Convertible.new  # => #<Convertible:0x8fdf4>

Awesome.

Ok, let's take it back a step. You've got a Car class with no subclasses. They haven't been necessary, and you don't want to add complexity you don't need. Good for you.

class Car
  def go_forward
    # ...
  end
end

Cars go forward, and that's about it. Except, now some parts of the code want to know how many doors a car has. Remember, that depends on the kind of car it is, so we'll need Car to know about its type. And since convertibles are pretty rare in your code, you have cars be sedans by default.

class Car
  attr_reader :type
  
  def initialize(type=:sedan)
    @type = type
  end
  
  def door_count
    case type
    when :convertible
      2
    when :sedan
      4
    end
  end
  
  def go_forward
    # ...
  end
end

Pretty soon you realize its time to refactor this puppy. You want a refactoring called Replace Type Code with Subclasses. What you end up with is the set of classes we saw at the beginning:

class Car
  def go_forward
    # ...
  end
end

class Convertible < Car
  def door_count
    2
  end
end

class Sedan < Car
  def door_count
    4
  end
end

The problem is that all of your tests are passing, but none of your code is using the subclasses yet. You could probably grep or ack through your source to find all of the times you use Car.new; in fact, you should. But you should still be testing that you've done it right. Also, if Car is backed by an ORM, it might be creating Car objects for you.

But as we've seen, Abstraction clears that all up. Just make the class abstract...

class Car
  abstract
  
  def go_forward
    # ...
  end
end

...and watch your tests fail. When they pass again, you've completed the refactoring.

(Warning: this section is a bit of a tease.)

Traditionally, abstract classes are found in strongly typed languages, where the compiler makes sure they're never created by type checking. In the Ruby world, the test suite is essentially our type checker. No complier can statically prove that an abstract Ruby class will never be instantiated, but we can exercise the test suite and see if it ever happens.

Abstract classes usually have a way to notate abstract methods. These are methods which are declared in the superclass, but don't have an implementation there. A concrete subclass has to implement all of the abstract methods. Again, this is checked by the type checker.

We have an example of an abstract method above, it's just not denoted in any way: #door_count. Similar to the case of abstract classes, we can't prove statically that abstract methods are implemented in the concrete subclasses. We have to run the tests and see if they're defined when they're called.

But: if they're called and there's no implementation, we'll get a NoMethodError anyway. The declaration of an abstract method in the superclass is really only useful to the type checker, to tell it that, for instance, any Car object has a #door_count method. We don't have a type checker. So we don't need to declare abstract methods.

It would be, if it were meaningful. The problem is: what does it mean to implement a method? In Ruby you really can't know whether a method is implemented until you send it the message and see if it raises a NoMethodError. There's no way to determine whether a class "implements" all of its superclasses abstract methods without making assumptions like that the class doesn't use method_missing or that instances of the class won't get their own singleton implementations. And the nail in the coffin: there's no time when a Ruby class is done being implemented, so there's no time to check.

On the other hand, maybe there's a way to make it useful. If there is, it certainly belongs here, so drop me a line or just fork away.

Written by Peter Jaros at drop.io.
Copyright 2009 drop.io, Inc.