Fix drop_sel for a MultiIndex

[owena11]

This is a proposed fix to #10862 to allow consistent usage between sel and drop_sel for MultiIndexes. The current implementation seems to convert the labels to an array to handle the edge case of labels of type xr.DataArray. This change makes that edge case more explict and delegates the reponsibility for converting to an array if needed down to the indexer.

Exisitng indexes already do this and so the change shouldn't have any preformance implications:

• pandas.Index.drop - here

• pandas.MultiIndex.drop - here

• Closes Differing behaviour between sel and drop_sel for MultiIndexes #10862

• Tests added

[welcome]

Thank you for opening this pull request! It may take us a few days to respond here, so thank you for being patient.
If you have questions, some answers may be found in our contributing guidelines.

[owena11]

I've imported DataArray here as that seems to fit with the style in other parts of the codebase (here), however I'm not entirely sure why this is done.

[shoyer]

I think that dataarray.py imports dataset.py already (historically Dataset predates DataArray slightly), so this avoids a recursive import.

[owena11]

Great, thanks for the explanation👍

[max-sixty]

I don't know this area that well, though it does seem to solve the immediate problem.

would anyone have alternative approaches that get at the root of the issue?

otherwise I would suggest merging this, maybe with a comment explaining, and then nothing prevents refactoring later...

(adding @benbovy , the indexing Czar, though others will also know more)

[owena11]

I don't know this area that well,

I have to admit the same! To try and provide more context I've dived down the git blame rabbit hole to explain the original reason for the cast to an array type within the code.

It was introduced in this PR #3177. The PR intorduces a whole bunch of type hints for the old drop method before the split to drop_vars / drop_sel refereced indexes: 'OrderedDict[Any, pd.Index]'. At that time the Index.drop didn't have a type signature apart from a docstring in pandas unless they're stored elsewhere. So the best theory I have is that is was related to matching the types from somewhere in the chain.

[shoyer]

Thanks @owena11 !

[shoyer]

I think that dataarray.py imports dataset.py already (historically Dataset predates DataArray slightly), so this avoids a recursive import.

[shoyer]

If you wanted to make this a little safer, could add:

Suggested change

	# Most conversion to arrays is better handled in the indexer, however
	# DataArrays are a special case where the underlying libraries don't provide
	# a good conversition.
	if isinstance(labels_for_dim, DataArray):
	labels_for_dim = np.asarray(labels_for_dim)
	# Most conversion to arrays is better handled in the indexer, however
	# DataArrays are a special case where the underlying libraries don't provide
	# a good conversition.
	if isinstance(labels_for_dim, DataArray):
	if labels_for_dim.dims not in ((), (dim,)):
	raise ValueError(
	"cannot use drop_sel() with DataArray values with "
	"along dimensions other than the dimensions being "
	f"indexed along: {labels_for_dim}"
	)
	labels_for_dim = np.asarray(labels_for_dim)

But this LGTM to me! Definitely an incremental improvement.

[owena11]

Commit and then reverted, highlighed by the tests it might break peoples current usage where a DataArray gets assigned the default dim names (i.e dim_0 etc) .

Also despite thinking this would nudge sel and drop_sel to be more consistent. After checking neither enforcing the alignment between dims for selecting with a DataArray:

>>> data = xr.Dataset({"x": ["a", "b"]})
>>> data.sel(x=xr.DataArray(["a",], dims=("y",)))
<xarray.Dataset> Size: 4B
Dimensions:  (y: 1)
Coordinates:
    x        (y) <U1 4B 'a'
Dimensions without coordinates: y
Data variables:
    *empty*

So would propose leaving this change for now.

[shoyer]

Right, sel() imposes the dimensions and coordinates of the indexer rather than checking for alignment. It is not obvious to me what the inverse of that would be!

[welcome]

Congratulations on completing your first pull request! Welcome to Xarray! We are proud of you, and hope to see you again!

[owena11]

Thanks for the time reviewing @shoyer & @max-sixty!