fix: Support pandas ExtensionArray ordering

[flying-sheep]

Fixes #6452

[codecov]

Codecov Report

Attention: Patch coverage is 92.30769% with 2 lines in your changes missing coverage. Please review.

Project coverage is 88.76%. Comparing base (b9243ba) to head (a2cb75f).
Report is 5 commits behind head on main.

Files with missing lines	Patch %	Lines
holoviews/element/util.py	87.50%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6481      +/-   ##
==========================================
- Coverage   88.76%   88.76%   -0.01%     
==========================================
  Files         323      323              
  Lines       68681    68698      +17     
==========================================
+ Hits        60963    60978      +15     
- Misses       7718     7720       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[hoxbro]

I left one comment, but this PR looks good otherwise.

[hoxbro]

This type-annotation does not seem right to me. The function where it is used can take more than these two, e.g., pd.Series or a cupy.Array. I think we should use something like np.typing.ArrayLike, but I haven't done any testing.

[flying-sheep]

Dataset.dimension_values() is documented as:

Returns:

NumPy array of values along the requested dimension

This documentation isn’t quite correct, e.g. my code relies on it returning a pd.ExtensionArray, but a lot of your code relies on dimension_values’s output e.g.

• having a .dtype attribute
• having the methods .astype, .min, .max, …
• being iterable
• being sliceable

I’m pretty sure it can’t be a pd.Series, for these your code passes on the inner np.ndarray | pd.ExtensionArray, that’s what I’m relying on here!

np.typing.ArrayLike also includes types like list or int, which I’m mostly sure isn’t correct. I mean, some parts of your code base check isinstance(list, vals), but others access the methods above without checking.

So can we figure out which types you actually support and actually fix the typing?

[hoxbro]

As mentioned in the original issue, much of the code was written before the invention of typing, so it is a huge lift.

Also worth mentioning is that when written, most of the underlying data structure was numpy. This is no longer the case, as it can be something else, like here with pd.ExtensionArray.

I tried to see if I could get out some other types than the array type and couldn't.

Your changes LGTM. Thank you for the PR!