BUG: dask.array: asarray(..., copy=None) copies in dask==2024.12

[ev-br]

closes gh-209

[ev-br]

Two issues:

• on python 3.9 we install an earlier dask version, which does not have a breaking change
• array-api-tests hang, dask==2024.12 tests hang on CI #210

[betatim]

In the standard for asarray it says that for copy=None you should not copy if possible. Do you think "in dask we always make a copy" is compatible with that? And if no, is there anything we can do in array-api-compat to fix this? It would be nice if we could organise (through the compat layer) for users to get the same behaviour no matter what version of dask they are using.

If there is nothing we can do to fix this then I think updating the test is fine.

[lucascolley]

Yes, I think it is compatible for Dask to say that they will always copy, and just raise an error for copy=False.

[ev-br]

The issue with 2024.12 is copy=None, actually.

[lucascolley]

The issue with 2024.12 is copy=None, actually.

It is compatible for them to always copy with copy=None.

EDIT: the wording in the spec is

must reuse existing memory buffer if possible

and the Dask maintainers appear to be saying 'reusing the existing memory buffer is never possible' (or at least never supported)

[ev-br]

CI is green with three tweaks

• asarray(..., copy=None) for array-likes always copies
• the test for copy=None is updated to match
• a unrelated test needs to be skipped on CI: was xfailed and was hanging with dask 2024.12

[crusaderky]

@lucascolley @rgommers this PR is blocking everything

[ev-br]

Let's unblock everything then :-)