Fixed in #802 and #947.

If you try that doing that indexing with a pandas.Series, you actually get an error message:

``` In [71]: s.loc['bar', slice(4000, 4100.3)]

...

/Users/shoyer/conda/envs/xarray-dev/lib/python3.5/site-packages/pandas/core/indexing.py in _getitem_axis(self, key, axis) 1363 # nested tuple slicing 1364 if is_nested_tuple(key, labels): -> 1365 locs = labels.get_locs(key) 1366 indexer = [ slice(None) ] * self.ndim 1367 indexer[axis] = locs

/Users/shoyer/conda/envs/xarray-dev/lib/python3.5/site-packages/pandas/core/index.py in get_locs(self, tup) 5692 if not self.is_lexsorted_for_tuple(tup): 5693 raise KeyError('MultiIndex Slicing requires the index to be fully lexsorted' -> 5694 ' tuple len ({0}), lexsort depth ({1})'.format(len(tup), self.lexsort_depth)) 5695 5696 # indexer KeyError: 'MultiIndex Slicing requires the index to be fully lexsorted tuple len (2), lexsort depth (0,) ```

I guess it's also worth investigating get_locs as an alternative or companion to get_loc_level.

From this point of view I agree that da.sel(band_wavenumber={'band': 'bar'}) is a nicer solution! I'll follow your suggestion of returning a new pandas.Index object from convert_label_indexer.

Unless I miss a better solution, we can use pandas.MultiIndex.get_loc_level to get both the indexer and the new pandas.Index object. However, there may still be some advanced cases where it won't behave as expected. For example, selecting both the band 'bar' and a range of wavenumber values (that doesn't exactly match the range of that band)

da.sel(band_wavenumber={'band': 'bar', 'wavenumber': slice(4000, 4100.3)})`

will a-priori return a stacked DataArray with the full multi-index:

In [32]: idx = da.band_wavenumber.to_index() In [33]: idx.get_loc_level(('bar', slice(4000, 4100.3)), level=('band', 'wavenumber')) Out[33]: (array([False, False, True, True, False], dtype=bool), MultiIndex(levels=[['bar', 'foo'], [4050.2, 4050.3, 4100.1, 4100.3, 4100.5]], labels=[[0, 0], [2, 3]], names=['band', 'wavenumber']))

The good news about writing our own custom way to select levels is that because we can avoid the stack/unstack, we can simply omit unused levels without worrying about doing dropna with unstack. So as long as we are implementing this in own other method (e.g., sel or xs), we can default to drop_level=True.

I would be OK with xs, but da.xs('bar', dim='band_wavenumber', level='band') feels much more verbose to me than da.sel(band_wavenumber={'band': 'bar'}). The later solution involves inventing no new API, and because dictionaries are not hashable there's no potential conflict with existing functionality.

Last year at the SciPy conference sprints, @jonathanrocher was working on adding similar dictionary support into .loc in pandas (i.e., da.loc[{'band': 'band'}]). I don't think he ever finished up that PR, but he might have a branch worth looking at as a starting point.

I think that this solution is better than, e.g., directly providing index level names as arguments of the sel method. This may be confusing and there may be conflict when different dimensions have the same index level names.

This is a fair point, but such scenarios are unlikely to appear in practice. We might be able to, for example, update our handling of MultiIndexes to guarantee that level names cannot conflict with other variables. This might be done by inserting dummy-variables of some sort into the _coords dict whenever a MultiIndex is added. It would take some work to ensure this works smoothly, though.

Besides this, It would be nice if the drop_level=True behavior could be applied by default to any selection (i.e., also when using loc, sel, etc.), like with Pandas. I don't know how Pandas does this (I'll look into that), but at first glance this would here imply checking for each dimension if it has a multi-index and then checking the labels for each index level.

Yes, agreed. Unfortunately the pandas code that handles this is a complete mess of spaghetti code (see pandas/core/indexers.py). So are welcome to try decoding it, but in my opinion you might be better off starting from scratch. In xarray, the function convert_label_indexer would need an updated interface that allows it to possibly return a new pandas.Index object to replace the existing index.

OK, I've read more carefully the discussion you referred to, and now I understand why it is preferable to call dropna explicitely. My last suggestion above is not compatible with this.

The xs method (not sure about the name) may still provide a concise way to perform a selection with explicit unstack and dropna. Maybe it is more appropriate to use dropna instead of drop_level:

python da.xs('bar', dim='band_wavenumber', level='band', dropna=True)

Thinking about this issue, I'd like to know what you think of the suggestions below before considering any pull request.

The following line code gives the same result than in my previous comment, but it is more explicit and shorter:

python da.unstack('band_wavenumber').sel(band='bar').dropna('wavenumber', how='any')

A nice shortcut to this would be adding a new xs method to DataArray and Dataset, which would be quite similar to the xs method of Pandas but here with an additional dim keyword argument:

python da.xs('bar', dim='band_wavenumber', level='band', drop_level=True)

Like Pandas, the default value of drop_level would be True. But here drop_level rather sets whether or not to apply dropna to all (unstacked) index levels of dim except the specified level.

I think that this solution is better than, e.g., directly providing index level names as arguments of the sel method. This may be confusing and there may be conflict when different dimensions have the same index level names.

Another, though less elegant, solution would be to provide dictionnaries to the sel method:

python da.sel(band_wavenumber={'band': 'bar'})

Besides this, It would be nice if the drop_level=True behavior could be applied by default to any selection (i.e., also when using loc, sel, etc.), like with Pandas. I don't know how Pandas does this (I'll look into that), but at first glance this would here imply checking for each dimension if it has a multi-index and then checking the labels for each index level.

Thanks for the tip. So I finally obtain the desired result when selecting the band 'bar' by doing this:

In [21]: (da.sel(band_wavenumber='bar') ...: .unstack('band_wavenumber') ...: .dropna('band', how='all') ...: .dropna('wavenumber', how='any') ...: .sel(band='bar')) Out[21]: <xarray.DataArray (wavenumber: 3)> array([ 1.20000000e-04, 1.00000000e-04, 8.50000000e-05]) Coordinates: band <U3 'bar' * wavenumber (wavenumber) float64 4.1e+03 4.1e+03 4.1e+03

But it's still a lot of code to write for such a common operation.

I'd be happy to think more deeply about this and contribute to the development of this great package ! (within the limits of my skills)

This is a really good point that honestly I had not thought carefully about before. I agree that it would be very nice to have this behavior, though. This will require a bit of internal refactoring to pass on the level information to the MultiIndex during indexing.

To remove unused levels after unstacking, you need to add an explicit dropna, e.g., da.unstack('band_wavenumber').dropna('band_wavenumber'). This is definitely a break from pandas, but for a good reason (IMO). See here for discussion on this point.

I raised another issue for the bug related to copying MultiIndex that you had in the earlier version of this PR (#769).

More broadly, if you care about MultiIndex support, it would be great to get some help pushing it. I'm happy to answer questions, but I'm at a new job and don't have a lot of time to work on new development.

Mmm now I'm wondering if the problem I explained above isn't just related to the 3rd TODO item in #719 (make levels accessible as coordinate variables).

Sorry for the post if it is the case.