Regression in 2.5.7 causing false positives in matplotlib

[chrisburr]

Steps to reproduce

I'm seeing a lot of false positives inside matplotlib with the latest release of astroid but they all appear to haver a common cause. To reproduce create a file named test.py and install matplotlib 3.4.2:

from matplotlib import pyplot as plt
ax = plt.Axes(plt.gcf(), [0.0, 0.0, 1.0, 1.0])
ax.set_yticks([0, 1])

It looks like the issue is coming from an advanced use of wrappers to define the set_yticks methods:

    get_yticks = _axis_method_wrapper("yaxis", "get_ticklocs")
    set_yticks = _axis_method_wrapper("yaxis", "set_ticks")

https://github.com/matplotlib/matplotlib/blob/v3.4.2/lib/matplotlib/axes/_base.py#L34-L92

Current behavior

With astroid 2.5.7:

$ pylint -E test.py
************* Module test
test.py:3:0: E1102: ax.set_yticks is not callable (not-callable)

Expected behavior

With astroid 2.5.6:

$ pylint -E test.py

[nelfin]

Yeah, it appears that matplotlib uses a custom __set_name__ (see https://docs.python.org/3/reference/datamodel.html#creating-the-class-object) to dynamically create these descriptors. There's no explicit support for this in astroid. As to the behaviour you're seeing, as far as pylint is concerned it's seen these attributes go from Uninferable to Instance of ._axis_method_wrapper in 2.5.7. IMO, a custom __set_name__ (or probably any custom behaviour in the type.__new__ chain, e.g. __init_subclass__, __prepare__, etc.) should be grounds for marking that node Uninferable.

[Pierre-Sassoulas]

@nelfin should we change this issue title to "A custom __set_name__ should mark a node as Uninferable" ?

[nelfin]

@Pierre-Sassoulas, maybe. I'm not 100% sold on the title. I guess I'd ask "how would someone search for this issue in the future" before considering changing the title. If instead your thoughts are "what's a good short description of this change so I can keep track of it in an upcoming milestone", then I guess I see your point.