Better signature for Mapping.get().

[gvanrossum]

Fixes #278.

The approach here is different from what we thought up before. According
to this commit, there are three valid forms of Mapping[K, V].get():

  d.get(k)        # returns Optional[V]
  d.get(k, None)  # also returns Optional[V]  (seen in the wild)
  d.get(k, v)     # requires v to have type V, and returns V

So we don't allow d.get(k, d) where d is a type unrelated to V. This is
a compromise, but I don't think there's much use for d.get(d) with a
default that is differently-typed than V, and when I tried it my
application code (that didn't use this) wouldn't type-check correctly.

[gvanrossum]

@JukkaL @matthiaskramm Any thoughts on this interminable issue?

Should we also support the signature (K, Optional[V]) -> Optional[V]?

[matthiaskramm]

Seems fine to me.

Isn't (K, Optional[V]) -> Optional[V] already covered by the other cases?

[JukkaL]

The overload doesn't work well with mypy, since mypy doesn't support @overloads outside stub files. If a user defines a subclass of Mapping, they can't implement get with the correct signature.

[gvanrossum]

The overload doesn't work well with mypy, since mypy doesn't support @Overloads outside stub files.

We should fix the latter. But until then, what do you suggest?

[JukkaL]

I don't know really. Overloading seems to be the only way to give get() a precise signature, except maybe for the following alternative (if we stretch the semantics a a bit):

_DT = TypeVar('_DT')

class Mapping(...):
    def get(self, key: _KT, default: _DT = None) -> Union[_VT, _DT]: ...

This won't work either with current mypy, but this is arguably a simpler signature and might be easier to support. Here's how we might make this might work:

• _DT is only used for the default value of get.
• d.get(key) should result in inferring None as the value of _DT, since the default isn't given an explicit value in the call.
• d.get(key, '') should infer str, not Optional[str] as the value of _DT, since the default value None is not used.

It would be good if mypy would support one of these alternatives properly (@overload or my alternative proposal) as otherwise we are likely to break user code. So I wonder if this can wait a bit? Another idea is to somehow have a different get signature in mypy's strict optional mode.

[gvanrossum]

OK, I'll just close this PR until we have the mypy side of things sorted out.