Array chunk warning

[mattchurgin]

• [] Tests passed
• Passes black dask / flake8 dask

Pull request to address issue #6810 . Issue a warning when mapping a function with 'axis' arguments onto a dask array. This alerts users to possible unexpected behavior when applying a function with axis argument along a dask array axis comprised of more than 1 chunk. Unfortunately not all tests are passing (they passed in my env before this commit). @jsignell would you mind taking a look when you have a chance and seeing if my approach to warning makes sense?

[jsignell]

This seems like a nice addition, but you'll need to make sure that the existing tests pass and add a new test to check this new functionality.

[mattchurgin]

Thanks @jsignell for the feedback. I incorporated it and added a test.

I saw some tests were failing, and it looks like that is because the default behavior of pytest here is to throw an error on warnings. When I run pytest with -W ignore::UserWarning, all my tests pass. All tests also pass when I comment out all of my additions except for the warning. I'm not sure if it's possible/reasonable to include an ignore::UserWarning in the pytest config. I'm actually also not sure where pytest is configured in the codebase. What do you think?

[jsignell]

I think you'd want to treat each test separately by adding a decorator like @pytest.mark.filterwarnings("ignore:Operating along axis with nchunks>1").

But having seen how much this pattern shows up in the test codebase, I am unsure whether we really want this warning or not (sorry!). I think if there were a way to disable it or an existing pattern of enabling more warnings it could be really nice to have this kind of thing. Maybe someone can chime in with a second opinion? @dask/maintenance

[mattchurgin]

Thanks @jsignell for the feedback. I appreciate it. Happy to continue working on this if people think it's valuable. I see from the tests that it's assumed users have a good understanding of how the chunk structure of their array influences how computations on it will behave. At the same time, not everyone will do that. I agree we don't want to overdo it with warnings. I leave it up to you and the maintainers to decide whether this use case is overdoing it. Thanks!

[jsignell]

Ok I am going to close this - I think we want to err on the side of only really important warnings. Thank you for opening it though!