homepage

When checking whether a class implements all abstractmethods (to know whether the class can be instantiated), one should only consider methods that come *before* the abstractmethod in the MRO -- methods that come after cannot be said to override the abstractmethod. Indeed, this is currently the case:
 from abc import ABC, abstractmethod
 class NeedsFoo(ABC):
 foo = abstractmethod(lambda self: None)
 class HasFoo(ABC):
 foo = lambda self: None
 class FooImplFirst(HasFoo, NeedsFoo): pass
 class FooImplSecond(NeedsFoo, HasFoo): pass
 FooImplFirst()
 try: FooImplSecond()
 except TypeError: pass
 else: raise Exception("Expected error")
Here FooImplFirst correctly overrides the foo method (using the HasFoo mixin first), so can be instantiated; FooImplSecond doesn't (by the MRO, FooImplSecond().foo would resolve to the abstract implementation), and we get a TypeError on instantiation.
However, this is not the case when considering builtins:
 from abc import ABC, abstractmethod
 class NeedsKeys(ABC):
 keys = abstractmethod(lambda self: None)
 HasKeys = dict # dict has a keys method.
 class KeysImplFirst(HasKeys, NeedsKeys): pass
 class KeysImplSecond(NeedsKeys, HasKeys): pass
 KeysImplFirst()
 try: KeysImplSecond()
 except TypeError: pass
 else: raise Exception("Expected error")
This example differs from the first only by having dict be the mixin that provides the keys method. However, running this example shows that KeysImplSecond() will incorrectly succeed: the ABC machinery does not realize that the keys method has not been overridden.
(Alternatively, one could say that "providing the method later in the MRO" is also sufficient; I think that goes against the expectations about ABCs but at least consistency between the non-builtin and builtin cases would be better.)

See also issue37927.