Implement process for importing backends

[machow]

Related to

• #86
• #54
• #141

Problem

We might have generic singledispatch function, with a concrete implementation in another file.

generics.py:

from functools import singledispatch

@singledispatch
def f(x):
    raise NotImplementedError()

pandas_backend.py:

import pandas as pd

from generics import f

@f.register(pd.DataFrame)
def _f_df(x):
    print("some DF function")
    return x

Where pandas_backend is optional, as one of several backends. This setup lets us avoid importing pandas when it's not needed.

However, if a user has imported pandas, it's not a guarantee they'll get the concrete implementations they need, or an informative error message:

import pandas as pd

df = pd.DataFrame({'x': [1,2,3]})

from generics import f

# generic NotImplementedError message, no knowledge to tell users about pandas_backend.
f(df)

Existing approaches

• sqlalchemy dialects - setuptools entrypoints
• ibis setuptools entrypoints
• import hooks - probably a bad idea