[Python-Dev] Enum: subclassing?

[Ethan Furman]

On 05/01/2013 02:48 PM, Eli Bendersky wrote:
>
>     >     Am 01.05.2013 20:04, schrieb Eli Bendersky:
>     >
>     >     > Actually, in flufl.enum, IntEnum had to define a magic __value_factory__
>     >     > attribute, but in the current ref435 implementation this isn't needed, so
>     >     > IntEnum is just:
>     >     >
>     >     > class IntEnum(int, Enum):
>     >     >     '''
>     >     >     Class where every instance is a subclass of int.
>     >     >     '''
>     >     >
>     >     > So why don't we just drop IntEnum from the API and tell users they should
>     >     do the
>     >     > above explicitly, i.e.:
>     >     >
>     >     > class SocketFamily(int, Enum):
>     >     >   AF_UNIX = 1
>     >     >   AF_INET = 2
>     >     >
>     >     > As opposed to having an IntEnum explicitly, this just saves 2 characters
>     >     > (comma+space), but is more explicit (zen!) and helps us avoid the
>     >     special-casing
>     >     > the subclass restriction implementation.
>     >
>     >     Wait a moment... it might not be immediately useful for IntEnums (however,
>     >     that's because base Enum currently defines __int__ which I find questionable),
>     >     but with  current ref435 you *can* create your own enum base classes with your
>     >     own methods, and derive concrete enums from that.  It also lets you have a
>     >     base class for enums and use it in isinstance().
>     >
>     >     If you forbid subclassing completely that will be impossible.
>     >
>     >
>     > I'm not sure what you mean, Georg, could you clarify?
>     > This works:
>     >
>     >>>> from ref435 import Enum
>     >>>> class SocketFamily(int, Enum):
>     > ...   AF_UNIX = 1
>     > ...   AF_INET = 2
>     > ...
>     >>>> SocketFamily.AF_INET
>     > SocketFamily.AF_INET [value=2]
>     >>>> SocketFamily.AF_INET == 2
>     > True
>     >>>> type(SocketFamily.AF_INET)
>     > <Enum 'SocketFamily'>
>     >>>> isinstance(SocketFamily.AF_INET, SocketFamily)
>     > True
>     >
>     > Now, with the way things are currently implemented, class IntEnum is just
>     > syntactic sugar for above. Guido decided against allowing any kind of
>     > subclassing, but as an implementation need we should keep some restricted form
>     > to implement IntEnum. But is IntEnum really needed if the above explicit
>     > multiple-inheritance of int and Enum is possible?
>
>     Well, my point is that you currently don't have to inherit from int (or IntEnum)
>     to get an __int__ method on your Enum, which is what I find questionable.  IMO
>     conversion to integers should only be defined for IntEnums.  (But I haven't
>     followed all of the discussion and this may already have been decided.)
>
>
> Good point. I think this may be just an artifact of the implementation - PEP 435 prohibits implicit conversion to
> integers for non-IntEnum enums. Since IntEnum came into existence, there's no real need for int-opearbility of other
> enums, and their values can be arbitrary anyway.
>
> Ethan - unless I'm missing something, __int__ should probably be removed.

The reason __int__ is there is because pure Enums should be using plain ints as their value 95% or more of the time, and 
being able to easily convert to a real int for either database storage, wire transmission, or C functions is a Good Thing.

IntEnum is for when the enum item *must* be a real, bonafide int in its own right, and the use case here is backwards 
compatibility with APIs that are already using real ints -- and this is really the *only* time IntEnum should be used).

The downside to IntEnum is you lose all Enum type protection; so if you don't need a real int, use a fake int, er, I 
mean, Enum, which can easily be int'ified on demand due to its handy dandy __int__ method.