Use definition of DTypeLike from Numpy if found

[jenshnielsen]

This allows application of numpy ufuncs to DataArrays to typecheck correctly with Numpy 1.20

E.g. without this fix:

import xarray as xr
import numpy as np

def applyufunctodataarray() -> xr.DataArray:
    a = xr.DataArray(np.array([-1,2,3]))
    b = np.abs(a)
    return b

will result in:

test.py:9: error: Argument 1 to "__call__" of "ufunc" has incompatible type "DataArray"; expected "Union[Union[int, float, complex, str, bytes, generic], Sequence[Union[int, float, complex, str, bytes, generic]], Sequence[Sequence[Any]], _SupportsArray]"
Found 1 error in 1 file (checked 1 source file)

While it will type check without issues with this fix.

I have not yet documented this or added tests.
Please let me know if this change is acceptable and I am happy to do that.

• Tests added
• Passes pre-commit run --all-files
• User visible changes (including notable bug fixes) are documented in whats-new.rst

[pep8speaks]

Hello @jenshnielsen! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2021-02-22 08:09:19 UTC

[max-sixty]

Thanks for the PR @jenshnielsen , we're happy to have you as a contributor.

This looks good! We do have a mechanism for checking versions elsewhere in the code using LooseVersion, but this approach works too. IIUC the existing code may break type-checking, but the code will still run on numpy<1.20; please let me know if that's not the case.

I'm not that well-versed here, so I'll wait a day for merging to see whether others have comments.

[jenshnielsen]

@max-sixty Thanks :)

IIUC the existing code may break type-checking, but the code will still run on numpy<1.20; please let me know if that's not the case.

Yes the existing code which is now the fallback could be considered wrong as it does not match the definitions that numpy ships.
We could consider syncing that with numpy but that seems like more trouble than its worth since no one is likely to typecheck this against the definitions of numpy functions without running numpy 1.20 or newer.

The numpy definition is:

DTypeLike = Union[
    np.dtype,
    # default data type (float64)
    None,
    # array-scalar types and generic types
    type,  # TODO: enumerate these when we add type hints for numpy scalars
    # anything with a dtype attribute
    "_SupportsDType[np.dtype[Any]]",
    # character codes, type strings or comma-separated fields, e.g., 'float64'
    str,
    _VoidDTypeLike,
]

So one could add None and type but should probably omit the rest in order to not replicate too much of numpys internal type definitions

[mathause]

Thanks. Looks good. IMHO typing is a nice-to-have far-from-complete part of xarray & I wouldn't bother about backwards compatibility (of the typing). Up to you.

[max-sixty]

Great, merging. Thanks @jenshnielsen !

[jenshnielsen]

Thanks @max-sixty @mathause