issue_comments

@dcherian No problem! Thanks for all your help!

LGTM. The reason for requires_pint_0_15 would be nice but I don't feel strongly about it so it's up to you.

Thanks for being patient, @rpmanser.

No worries, thanks for the reviews!

I meant is_duck_array_or_ndarray, not is_duck_dask_array_or_ndarray. I don't suppose there's a way to fix that at this point?

Tests are failing in test_formatting.py for test_diff_attrs_repr_with_array. This looks very similar to the previous problem that was occurring for duck array equality due to NEP18. Is a similar xfail appropriate here?

I'm not sure why FormattingBlack failed. It's claiming that 45 files would be reformatted. I used the pre-commit hook, which passed.

I have erred on the side of leaving the isinstance checks in place where @jthielen reviewed them. If it is decided that no further tests are needed, then this should be ready for review.

On another note, if it is appropriate to apply these changes:

Revisiting it I now realize that is_array_like is actually not a good name for that function: officially, "array_like" means that it can be casted using np.asarray. Maybe we should rename that to is_duck_array? But then we'd also need to check for __array_ufunc__ and __array_function__ because we also need those to properly wrap duck arrays.

As long as that doesn't break any of the current uses, I think that would be the best way forwards. This would require xarray to be on NumPy 1.16+ (in order to ensure hasattr(np.ndarray, "__array_function__"), but that's only 9 days away by NEP 29, so hopefully that isn't too much of an issue. Then, most of the checks for "can this be wrapped by xarray" should be able to become something like

python is_duck_array(x) and not isinstance(x, (DataArray, Variable))

with is_duck_dask_array only being needed for checking Dask Array-like support.

I can do that as well. If not, should a new issue be opened for this?

@rpmanser For what it's worth, I'd think doing tests like test_dask.py but with a post-https://github.com/hgrecco/pint/pull/1129 Pint Quantity or suitably mocked wrapper class between xarray and Dask would be best. But, I think this is more of a maintainer-level question.

I'm getting a bit lost within the testing suite while trying to figure out the implications of this change and what kinds of tests are necessary for it. Would a test module for this - Individually test all functions that include the newly implemented is_duck_dask_array()? - Test that a duck dask Array has attributes of a dask Array and redirects the appropriate methods to those defined by Variable, DataArray, and Dataset? - Be similar to test_dask.py? - All, some, or none of the above?

Any suggestions would be greatly appreciated.

Revisiting it I now realize that is_array_like is actually not a good name for that function: officially, "array_like" means that it can be casted using np.asarray. Maybe we should rename that to is_duck_array? But then we'd also need to check for __array_ufunc__ and __array_function__ because we also need those to properly wrap duck arrays.

As long as that doesn't break any of the current uses, I think that would be the best way forwards. This would require xarray to be on NumPy 1.16+ (in order to ensure hasattr(np.ndarray, "__array_function__"), but that's only 9 days away by NEP 29, so hopefully that isn't too much of an issue. Then, most of the checks for "can this be wrapped by xarray" should be able to become something like

python is_duck_array(x) and not isinstance(x, (DataArray, Variable)) with is_duck_dask_array only being needed for checking Dask Array-like support.

Although, while we're at it, do we also need more careful handling of upcast types? I'm not sure if there even are any out there right now (not sure if a HoloViews Dataset counts here), but that doesn't necessarily mean there never will be any.

Also, a broader discussion that's I've seen hinted to in the past, but is brought to the forefront by this PR: how should xarray be checking for general duck array types it can wrap?

Right now, it looks like there is a mix of

python is_array_like(data) and python hasattr(data, "__array_function__") or isinstance(data, dask_array_type)

and it would be nice to bring consistency. However, both these checks seem like they'd let too many types through. For example, is_array_like as currently implemented would still let through too much, such as MemoryCachedArray and xarray Variables/DataArrays, and __array_function__ doesn't truly indicate a NumPy duck array on its own (and it really only works to only capture downcast types right now since xarray does not yet implement __array_function__ itself, #3917).