Add support for Array API in NamedArray

[Illviljan]

• Closes #xxxx
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst
• New functions/methods are listed in api.rst

[andersy005]

@Illviljan, i'm curious to hear your thoughts on adopting a similar structure/organization as that used in Numpy so as to it easier to incorporate new array API compliant functionality in the future. the new structure would group functionality under /namedarray/_array_api/*.py instead of of putting everything in _array_api.py.

e.g.

├── _creation_functions.py
├── _data_type_functions.py
├── _dtypes.py
├── _elementwise_functions.py
├── _indexing_functions.py
├── _manipulation_functions.py
├── _searching_functions.py
├── _set_functions.py
├── _sorting_functions.py
├── _statistical_functions.py
├── linalg.py

[Illviljan]

Sure! I just thought it was too few functions to bother with it right now.

Cubed has similar structure too.

[andersy005]

perfect! i intend to add a few more functions to the array_api and i'm happy to put this structure in place

[dcherian]

@Illviljan can you comment on why this approach is useful if we can't advertise array_api compliance based on upstream input?

[Illviljan]

I notice sklearn for example is slowly getting array api support and I think it would be a great win to get these kind of packages to just work out of the box.
I think it's lazy of us not to give it a serious attempt considering how much our maintainers have pushed for array standardization in the past.

There's two distinct programming styles between array api vs. xarray:

• function-focused, xp.mean(x) - array api, requires axis but it is fine to add the optional kwarg dim and encourage that one over axis in the docs.
• method-focused, x.mean() - xarray/namedarray, we can do anything so I think we should enforce using dim instead of axis.

Due to this, I'm not yet convinced it is impossible for us to not support both. Most stuff xarray do is simply an extension of the narrow array api. Internally we have to deal with array apis and the axis parameter anyway so we will have most tools available to easily make these functions.

namedarray imports duck_array_ops.py, don't you think it's very similar to _array_api.py?

xarray/xarray/namedarray/_aggregations.py

Line 9 in 7c1bb8d

from xarray.core import duck_array_ops

[Illviljan]

This particular PR has simply grown because of the lack of typing in .reduce.
It was supposed to simply just add a mean function in _array_api.py but then you go down the typing rabbit hole...

[dcherian]

I'm not yet convinced it is impossible for us to not support both.

Sure, but the way to proceed would be to build consensus around this in an issue discussion or at the bi-weekly meeting.

requires axis but it is fine to add the optional kwarg dim and encourage that one over axis in the docs.

For example, this assertion has had very little support within the core team in discussions.

[Illviljan]

I'm not yet convinced it is impossible for us to not support both.

Sure, but the way to proceed would be to build consensus around this in an issue discussion or at the bi-weekly meeting.

It's hard to change peoples mind sometimes, so I like to have some working examples for discussion.
I don't feel like these array api focused PRs improving on a non-public module is disturbing anyone.

requires axis but it is fine to add the optional kwarg dim and encourage that one over axis in the docs.

For example, this assertion has had very little support within the core team in discussions.

Still, there is duck_array_ops.py that does basically the same thing.

It is the only way to be compliant to both ways and xarray in general seem to enjoy the method-chaining style more. So adding array_api compatible equivalents wont change current workflows much but will however allow considerable more workflows automatically. Being able to pass around attrs and dims might be enough for users.

[dcherian]

It's hard to change peoples mind sometimes, so I like to have some working examples for discussion.

Don't you have enough already?

Still, there is duck_array_ops.py that does basically the same thing.

These ops are applied to the wrapped array not to an xarray object.

[Illviljan]

Still, there is duck_array_ops.py that does basically the same thing.

These ops are applied to the wrapped array not to an xarray object.

Yes, but it doesn't look insurmountable to me for namedarray to replace duck_array_ops.py with the _array_api + some extra functions to get the same result and the widened possibilities.

Thank you for the input @dcherian. I won't merge this without approval.
I might continue testing stuff on this branch until I come to the same conclusion as you.

[dcherian]

generally expanding duck_array_ops to use the array_api more widely would be a great improvement :) still not sure about handling namedarray at that level though