singledispatch chain gives type errors

[lopuhin]

I believe this is different from other issues about singledispatch (#2904).

This is the program that does not type-check:

from singledispatch import singledispatch

class A:
    pass

class B(A):
    pass

@singledispatch
def base(x, **kw):
    pass

@base.register(A)
@singledispatch
def a(x):
    print('a')

@a.register(B)
def b(x):
    print('b')

base(A())
# want "base" and "a" both dispatch to "b"
base(B())
a(B())

Expected output from running it:

a
b
b

running mypy 0.501 with Python 3.5 gives the following errors:

t.py:13: error: Argument 1 has incompatible type _SingleDispatchCallable[Any]; expected Callable[..., Any]
t.py:18: error: Callable[..., Any] has no attribute "register"

Next is just the motivation part - why do we need this :)

Why such a complicated program: I want both "base" and "a" dispatch to "b", because "base" is a more generic function that works for any type supported by our library, and "a" is less generic, and all functions that dispatch from "a" share a common signature. So the user can import "a" and get more help from the IDE (and mypy) about supported arguments:

@singledispatch
def base(x, **kwargs):
    pass

@base.register(A)
@singledispatch
def a(x, foo=None):
    print('a')

@a.register(B)
def b(x, foo=None):
    print('b')

Real code that fails to typecheck is here: https://github.com/TeamHG-Memex/eli5/blob/724c63a9b97e44eef8264305e66e808d249810c4/eli5/sklearn/explain_weights.py#L114

[lopuhin]

But there is a workaround that works in almost the same way (for us it's good enough) and typechecks:

from singledispatch import singledispatch

class A:
    pass

class B(A):
    pass

@singledispatch
def base(x, **kwargs):
    print('base')

@singledispatch
def a(x, foo=None):
    print('a')

def register(cls):
    def deco(f):
        return base.register(cls)(a.register(cls)(f))
    return deco

@register(B)
def b(x, foo=None):
    print('b')

base(A())
# want "base" and "a" both dispatch to "b"
base(B())
a(B())

[graingert]

see #2904

[AlexWaygood]

Cannot reproduce on 0.920+; this appears to now be fixed.