Calling parent class __init__ with multiple inheritance, what's the right way?

The answer to your question depends on one very important aspect: Are your base classes designed for multiple inheritance?

There are 3 different scenarios:

1. The base classes are unrelated, standalone classes.

   If your base classes are separate entities that are capable of functioning independently and they don't know each other, they're not designed for multiple inheritance. Example:

   class Foo:    def __init__(self):        self.foo = 'foo'class Bar:    def __init__(self, bar):        self.bar = bar

   Important: Notice that neither Foo nor Bar calls super().__init__()! This is why your code didn't work correctly. Because of the way diamond inheritance works in python, classes whose base class is object should not call super().__init__(). As you've noticed, doing so would break multiple inheritance because you end up calling another class's __init__ rather than object.__init__(). (Disclaimer: Avoiding super().__init__() in object-subclasses is my personal recommendation and by no means an agreed-upon consensus in the python community. Some people prefer to use super in every class, arguing that you can always write an adapter if the class doesn't behave as you expect.)

   This also means that you should never write a class that inherits from object and doesn't have an __init__ method. Not defining a __init__ method at all has the same effect as calling super().__init__(). If your class inherits directly from object, make sure to add an empty constructor like so:

   class Base(object):    def __init__(self):        pass

   Anyway, in this situation, you will have to call each parent constructor manually. There are two ways to do this:

   • Without super

     class FooBar(Foo, Bar):    def __init__(self, bar='bar'):        Foo.__init__(self)  # explicit calls without super        Bar.__init__(self, bar)

   • With super

     class FooBar(Foo, Bar):    def __init__(self, bar='bar'):        super().__init__()  # this calls all constructors up to Foo        super(Foo, self).__init__(bar)  # this calls all constructors after Foo up                                        # to Bar

   Each of these two methods has its own advantages and disadvantages. If you use super, your class will support dependency injection. On the other hand, it's easier to make mistakes. For example if you change the order of Foo and Bar (like class FooBar(Bar, Foo)), you'd have to update the super calls to match. Without super you don't have to worry about this, and the code is much more readable.

2. One of the classes is a mixin.

   A mixin is a class that's designed to be used with multiple inheritance. This means we don't have to call both parent constructors manually, because the mixin will automatically call the 2nd constructor for us. Since we only have to call a single constructor this time, we can do so with super to avoid having to hard-code the parent class's name.

   Example:

   class FooMixin:    def __init__(self, *args, **kwargs):        super().__init__(*args, **kwargs)  # forwards all unused arguments        self.foo = 'foo'class Bar:    def __init__(self, bar):        self.bar = barclass FooBar(FooMixin, Bar):    def __init__(self, bar='bar'):        super().__init__(bar)  # a single call is enough to invoke                               # all parent constructors        # NOTE: `FooMixin.__init__(self, bar)` would also work, but isn't        # recommended because we don't want to hard-code the parent class.

   The important details here are:

   • The mixin calls super().__init__() and passes through any arguments it receives.
   • The subclass inherits from the mixin first: class FooBar(FooMixin, Bar). If the order of the base classes is wrong, the mixin's constructor will never be called.

3. All base classes are designed for cooperative inheritance.

   Classes designed for cooperative inheritance are a lot like mixins: They pass through all unused arguments to the next class. Like before, we just have to call super().__init__() and all parent constructors will be chain-called.

   Example:

   class CoopFoo:    def __init__(self, **kwargs):        super().__init__(**kwargs)  # forwards all unused arguments        self.foo = 'foo'class CoopBar:    def __init__(self, bar, **kwargs):        super().__init__(**kwargs)  # forwards all unused arguments        self.bar = barclass CoopFooBar(CoopFoo, CoopBar):    def __init__(self, bar='bar'):        super().__init__(bar=bar)  # pass all arguments on as keyword                                   # arguments to avoid problems with                                   # positional arguments and the order                                   # of the parent classes

   In this case, the order of the parent classes doesn't matter. We might as well inherit from CoopBar first, and the code would still work the same. But that's only true because all arguments are passed as keyword arguments. Using positional arguments would make it easy to get the order of the arguments wrong, so it's customary for cooperative classes to accept only keyword arguments.

   This is also an exception to the rule I mentioned earlier: Both CoopFoo and CoopBar inherit from object, but they still call super().__init__(). If they didn't, there would be no cooperative inheritance.

Bottom line: The correct implementation depends on the classes you're inheriting from.

The constructor is part of a class's public interface. If the class is designed as a mixin or for cooperative inheritance, that must be documented. If the docs don't mention anything of the sort, it's safe to assume that the class isn't designed for cooperative multiple inheritance.

Both ways work fine. The approach using super() leads to greater flexibility for subclasses.

In the direct call approach, C.__init__ can call both A.__init__ and B.__init__.

When using super(), the classes need to be designed for cooperative multiple inheritance where C calls super, which invokes A's code which will also call super which invokes B's code. See http://rhettinger.wordpress.com/2011/05/26/super-considered-super for more detail on what can be done with super.

[Response question as later edited]

So it seems that unless I know/control the init's of the classes I inherit from (A and B) I cannot make a safe choice for the class I'm writing (C).

The referenced article shows how to handle this situation by adding a wrapper class around A and B. There is a worked-out example in the section titled "How to Incorporate a Non-cooperative Class".

One might wish that multiple inheritance were easier, letting you effortlessly compose Car and Airplane classes to get a FlyingCar, but the reality is that separately designed components often need adapters or wrappers before fitting together as seamlessly as we would like :-)

One other thought: if you're unhappy with composing functionality using multiple inheritance, you can use composition for complete control over which methods get called on which occasions.

Either approach ("new style" or "old style") will work if you have control over the source code for A and B. Otherwise, use of an adapter class might be necessary.

Source code accessible: Correct use of "new style"

class A(object):    def __init__(self):        print("-> A")        super(A, self).__init__()        print("<- A")class B(object):    def __init__(self):        print("-> B")        super(B, self).__init__()        print("<- B")class C(A, B):    def __init__(self):        print("-> C")        # Use super here, instead of explicit calls to __init__        super(C, self).__init__()        print("<- C")

>>> C()-> C-> A-> B<- B<- A<- C

Here, method resolution order (MRO) dictates the following:

• C(A, B) dictates A first, then B. MRO is C -> A -> B -> object.
• super(A, self).__init__() continues along the MRO chain initiated in C.__init__ to B.__init__.
• super(B, self).__init__() continues along the MRO chain initiated in C.__init__ to object.__init__.

You could say that this case is designed for multiple inheritance.

Source code accessible: Correct use of "old style"

class A(object):    def __init__(self):        print("-> A")        print("<- A")class B(object):    def __init__(self):        print("-> B")        # Don't use super here.        print("<- B")class C(A, B):    def __init__(self):        print("-> C")        A.__init__(self)        B.__init__(self)        print("<- C")

>>> C()-> C-> A<- A-> B<- B<- C

Here, MRO does not matter, since A.__init__ and B.__init__ are called explicitly. class C(B, A): would work just as well.

Although this case is not "designed" for multiple inheritance in the new style as the previous one was, multiple inheritance is still possible.

Now, what if A and B are from a third party library - i.e., you have no control over the source code for A and B? The short answer: You must design an adapter class that implements the necessary super calls, then use an empty class to define the MRO (see Raymond Hettinger's article on super - especially the section, "How to Incorporate a Non-cooperative Class").

Third-party parents: A does not implement super; B does

class A(object):    def __init__(self):        print("-> A")        print("<- A")class B(object):    def __init__(self):        print("-> B")        super(B, self).__init__()        print("<- B")class Adapter(object):    def __init__(self):        print("-> C")        A.__init__(self)        super(Adapter, self).__init__()        print("<- C")class C(Adapter, B):    pass

>>> C()-> C-> A<- A-> B<- B<- C

Class Adapter implements super so that C can define the MRO, which comes into play when super(Adapter, self).__init__() is executed.

And what if it's the other way around?

Third-party parents: A implements super; B does not

class A(object):    def __init__(self):        print("-> A")        super(A, self).__init__()        print("<- A")class B(object):    def __init__(self):        print("-> B")        print("<- B")class Adapter(object):    def __init__(self):        print("-> C")        super(Adapter, self).__init__()        B.__init__(self)        print("<- C")class C(Adapter, A):    pass

>>> C()-> C-> A<- A-> B<- B<- C

Same pattern here, except the order of execution is switched in Adapter.__init__; super call first, then explicit call. Notice that each case with third-party parents requires a unique adapter class.

So it seems that unless I know/control the init's of the classes I inherit from (A and B) I cannot make a safe choice for the class I'm writing (C).

Although you can handle the cases where you don't control the source code of A and B by using an adapter class, it is true that you must know how the init's of the parent classes implement super (if at all) in order to do so.