Aaah, a bit weird and probably wrong. It would seem (from the code) that the thought it that a buffer/memoryview is something like a void python object (which to some degree may even be true, though I have my doubts). Just to note though, its all in numpy/core/numerictypes.py....

Worse yet, gh-9898 means that this only happens on python 3

Are there any plans to fix this soon?

From the docs of isscalar

In almost all cases np.ndim(x) == 0 should be used instead of this function, as that will also return true for 0d arrays.

I think that returning false on 0d memoryviews is actually correct, since it's consistent with 0d arrays.

# 0-D and 1-D arrays
>>> x0 = np.array(0)
>>> x1 = np.array([])
>>> np.isscalar(x0)
False
>>> np.isscalar(x1)
False

# A NumPy scalar
>>> dt0 = np.float64(0)
>>> np.isscalar(dt0)
True

# When going through `memoryview` for all those things
>>> np.isscalar(memoryview(dt0))
True
>>> np.isscalar(memoryview(x0))
True
>>> np.isscalar(memoryview(x1))
True

# Using `.ndim`:
>>> x0.ndim
0
>>> x1.ndim
1
>>> dt0.ndim
0
>>> memoryview(x0).ndim
0
>>> memoryview(x1).ndim
1
>>> memoryview(dt0).ndim
0

So conclusion: yes, memorview combined with isscalar indeed results in an inconsistency. However, .ndim is preferred and does do the correct thing. The isscalar docstring already discusses why .ndim == 0 should be preferred.

Not sure that this is worth keeping open; isscalar should perhaps get an even stronger warning.