Support for masked arrays (`numpy.ma.masked`)

[radoering]

When I try to store a masked array, I get only an incomprehensible error message:

ValueError: cannot reshape array of size 3 into shape (24,)

See "example 1" for details.

When I use zarr 2 (or zarr 3 with zarr_format=2), there is no error but the mask is just lost. See "example 2".

Example 1

import numpy
import zarr

masked_array = numpy.ma.masked_array([1, 2, 3], mask=[0, 1, 0])
zarr_array = zarr.array(masked_array)

Error Message:

ValueError: cannot reshape array of size 3 into shape (24,)

Example 2

import numpy
import zarr

masked_array = numpy.ma.masked_array([1, 2, 3], mask=[0, 1, 0])
print(masked_array)

zarr_array = zarr.array(masked_array, zarr_format=2)
print(zarr_array[:])

Output:

[1 -- 3]
[1 2 3]

[d-v-b]

saving a masked array requires saving 2 arrays: the data itself, and the mask. You can do this directly today by explicitly creating the two arrays separately, but that's not terribly convenient.

There's also a question about interoperability with other zarr implementations. I think if we support masked arrays, it would probably be via something like zarr-developers/zarr-extensions#33, but I'm not sure how well that would work for masked arrays specifically.