Dask.array deepcopy does not preserve masking since 0.18.2

[pp-mo]

If 'a' is a dask array wrapped around a numpy masked array (with "from _array(..., asarray=False)",
then the result of "copy.deepcopy(a).compute()" is now (wrongly) not a masked array,

Though the original "a.compute()", and even "copy.copy(a)" are correct.

Example code to show :

import dask.array as da
import numpy.ma as ma
import copy
t = ma.masked_array([1, 2], mask=[0, 1])
a = da.from_array(t, chunks=t.shape, asarray=False)
print(copy.deepcopy(a).compute())

For example, in dask 0.18.1

>>> import dask
>>> import dask.array as da
>>> import numpy.ma as ma
>>> import copy
>>> t = ma.masked_array([1, 2], mask=[0, 1])
>>> a = da.from_array(t, chunks=t.shape, asarray=False)
>>> print(dask.__version__)
0.18.1
>>> print(copy.copy(a).compute())
[1 --]
>>> print(copy.deepcopy(a).compute())
[1 --]
>>> 

But in 0.18.2 ...

>>> import dask
>>> import dask.array as da
>>> import numpy.ma as ma
>>> import copy
>>> 
>>> t = ma.masked_array([1, 2], mask=[0, 1])
>>> a = da.from_array(t, chunks=t.shape, asarray=False)
>>> 
>>> print(dask.__version__)
0.18.2
>>> print(copy.copy(a).compute())
[1 --]
>>> print(copy.deepcopy(a).compute())
[1 2]
>>> 

[pp-mo]

This is causing us to pin dask in iris 😢
as it breaks various unit tests for us.

I suspect the change here, but I don't know really enough to suggest a fix ...
( but surely this anyway does not respect the comment on the previous line, as noted ? )

[mrocklin]

cc @jcrist

…

On Fri, Aug 3, 2018 at 11:32 AM, Patrick Peglar ***@***.***> wrote: This has caused us to pin dask in iris <https://github.com/SciTools/iris> 😢 as it breaks various unit tests for us. I suspect the change here <https://github.com/dask/dask/pull/3754/files#r207583443>, but I don't know really enough to suggest a fix ... but surely this did not respect the comment anyway ? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#3848 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AASszOdhXSMxZJ5q2rlYeXCsOgYehTD9ks5uNG0lgaJpZM4VuMMc> .

[jcrist]

Looks like the np.copy in the map_blocks should be replaced with dask.utils.M.copy (which would call the .copy method on the array).

[mrocklin]

@pp-mo it looks like there is a proposed fix to your problem in #3852 . Do you have time to take a look?