I notice that if I try to do assert_eq(x, dx) it fails, but I would have expected not to.

>>> x
masked_array(data=[--, --, --],
             mask=[ True,  True,  True],
       fill_value='?',
            dtype=object)
>>> dx 
dask.array<setitem, shape=(3,), dtype=object, chunksize=(2,), chunktype=numpy.MaskedArray>

>>>dx.compute() 
masked_array(data=[--, --, --],
             mask=[ True,  True,  True],
       fill_value='?',
            dtype=object)           

not sure why the assert_eq(x, dx) does not work.
But with assert_eq(x.mask, da.ma.getmaskarray(dx))
we are asserting the mask is the same, but not the resulting array like in all the other tests.

@jrbourbeau Do you have any thoughts?

Hi - I think it's because np.ma.allclose (https://github.com/dask/dask/blob/2022.10.2/dask/array/utils.py#L177)
treats masked elements as being equal to co-located non-masked elements:

>>> import numpy as np
>>> a = np.ma.arange(3)
>>> b = np.ma.arange(3)
>>> a[0] = np.ma.masked
>>> b[1] = np.ma.masked
>>> print(a)
[-- 1 2]
>>> print(b)
[0 -- 2]
>>> np.ma.allclose(a, b, masked_equal=True)
True

... I should also say that the test is checking that both arrays (np and dask) are masked everywhere, so unless we check the mask arrays themselves, we could get a false positive. Thanks.

Agreed, we should check that the arrays are the same. Hence me asking why trying assert_eq(x, dx) fails. We should find a way to check for this, ideally without having to compute the dask array.

Hi @ncclementi, I'm not sure if I need to do anything, here. I think that assert_eq(x, dx) is behaving as expected, in that it mimics how numpy behaves - so it's less that it fails, rather it's not designed to answer the question we're asking. I think that the only way two wholly masked arrays can be safely compared is by comparing their masks, which assert_eq can do easily. What do you think?

Hi again, just checking that if there's anything outstanding on this issue? It'd be nice to see it in a January release, if possible :). Thanks.