[FEA] Proxy ndarrays don't pass an instancecheck() for np.ndarray

[shwina]

There's a lot of code out there that does something like:

if not isinstance(x, np.ndarray):
    raise TypeError("Not a numpy array")

This is a problem for cudf.pandas, because proxy ndarray types, such as those returned by pd.Series.values do not pass the isinstance() check.

Ideally, more projects would avoid doing a hard isinstance() check and instead use something like EAFP:

np.asarray(x)

..but that is not the world we live in today. Too many third-party libraries that people want to use with cudf.pandas use the pattern above, so it's on us to solve the problem right now.

[Matt711]

As far as I can see there isn't a fix for this because we cannot modify np.ndarray's __instancecheck__ which is called when you call isinstance(c, np.ndarray). For example,

import cudf.pandas
cudf.pandas.install()
import pandas as pd
import numpy as np

x = pd.Series([1,2]).to_numpy() # proxy for numpy.ndarray

isinstance(x, np.ndarray) # calls np.ndarray.__instancecheck__(x) which returns false

I thought you might be able to modify np.ndarray.__instancecheck__ but you get an error

TypeError: cannot set '__instancecheck__' attribute of immutable type 'numpy.ndarray'

[shwina]

That's correct. I don't think there's a very good solution here, short of proxying numpy in the same way we proxy pandas. But that comes with its own set of challenges.