Abstract Base Classes in Python

What the ABC?

So what are Abstract Base Classes good for? A while ago I had a discussion at work about which pattern to use for implementing a maintainable class hierarchy in Python. More specifically, the goal was to define a simple class hierarchy for a service backend in the most programmer-friendly and maintainable way.

We had a BaseService class that defined a common interface and several concrete implementations. The concrete implementations do different things but all of them provide the same interface (MockService, RealService, and so on). To make this relationship explicit, the concrete implementations all subclass BaseService.

To make this code as maintainable and programmer-friendly as possible we wanted to make sure that:

• instantiating the base class is impossible; and
• forgetting to implement interface methods in one of the subclasses raises an error as early as possible.

When to Use Python’s abc Module

Now why would you want to use Python’s abc module to solve this problem? The above design is pretty common in more complex systems. To enforce that a derived class implements a number of methods from the base class, something like this Python idiom is typically used:

class Base:
    def foo(self):
        raise NotImplementedError()

    def bar(self):
        raise NotImplementedError()

class Concrete(Base):
    def foo(self):
        return 'foo() called'

    # Oh no, we forgot to override bar()...
    # def bar(self):
    #     return "bar() called"

So, what do we get from this first attempt at solving the problem? Calling methods on an instance of Base correctly raises NotImplementedError exceptions:

>>> b = Base()
>>> b.foo()
NotImplementedError

Furthermore, instantiating and using Concrete works as expected. And, if we call an unimplemented method like bar() on it, this also raises an exception:

>>> c = Concrete()
>>> c.foo()
'foo() called'
>>> c.bar()
NotImplementedError

This first implementation is decent, but it isn’t perfect yet. The downsides here are that we can still:

• instantiate Base just fine without getting an error; and
• provide incomplete subclasses—instantiating Concrete will not raise an error until we call the missing method bar().

With Python’s abc module that was added in Python 2.6, we can do better and solve these remaining issues. Here’s an updated implementation using an Abstract Base Class defined with the abc module:

from abc import ABCMeta, abstractmethod

class Base(metaclass=ABCMeta):
    @abstractmethod
    def foo(self):
        pass

    @abstractmethod
    def bar(self):
        pass

class Concrete(Base):
    def foo(self):
        pass

    # We forget to declare bar() again...

This still behaves as expected and creates the correct class hierarchy:

assert issubclass(Concrete, Base)

Yet, we do get another very useful benefit here. Subclasses of Base raise a TypeError at instantiation time whenever we forget to implement any abstract methods. The raised exception tells us which method or methods we’re missing:

>>> c = Concrete()
TypeError:
"Can't instantiate abstract class Concrete \
with abstract methods bar"

Without abc, we’d only get a NotImplementedError if a missing method was actually called. Being notified about missing methods at instantiation time is a great advantage. It makes it more difficult to write invalid subclasses. This might not be a big deal if you’re writing new code, but a few weeks or months down the line, I promise it’ll be helpful.

This pattern is not a full replacement for compile-time type checking, of course. However, I found it often makes my class hierarchies more robust and more readily maintainable. Using ABCs states the programmer’s intent clearly and thus makes the code more communicative. I’d encourage you to read the abc module documentation and to keep an eye out for situations where applying this pattern makes sense.