Dropping indices doesn't work for dimensions without coordinates

[xlambein]

MCVE Code Sample

Calling the following:

arr = xr.DataArray(
    np.arange(10),
    dims=['x']
)

arr.sel(x=[1, 2])  # This works

arr.drop([1, 2], dim='x')  # Raises an exception

Produces this traceback:

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
~/anaconda3/envs/fibricheck-py36/lib/python3.6/site-packages/xarray/core/dataset.py in drop(self, labels, dim, errors)
   3135             try:
-> 3136                 index = self.indexes[dim]
   3137             except KeyError:

~/anaconda3/envs/fibricheck-py36/lib/python3.6/site-packages/xarray/core/indexes.py in __getitem__(self, key)
     32     def __getitem__(self, key):
---> 33         return self._indexes[key]
     34

KeyError: 'x'

During handling of the above exception, another exception occurred:

ValueError                                Traceback (most recent call last)
<ipython-input-1-d0f96f591c69> in <module>
      9 arr.sel(x=[1, 2])  # This works
     10
---> 11 arr.drop([1, 2], dim='x')  # Raises an exception

~/anaconda3/envs/fibricheck-py36/lib/python3.6/site-packages/xarray/core/dataarray.py in drop(self, labels, dim, errors)
   1690         if utils.is_scalar(labels):
   1691             labels = [labels]
-> 1692         ds = self._to_temp_dataset().drop(labels, dim, errors=errors)
   1693         return self._from_temp_dataset(ds)
   1694

~/anaconda3/envs/fibricheck-py36/lib/python3.6/site-packages/xarray/core/dataset.py in drop(self, labels, dim, errors)
   3137             except KeyError:
   3138                 raise ValueError(
-> 3139                     'dimension %r does not have coordinate labels' % dim)
   3140             new_index = index.drop(labels, errors=errors)
   3141             return self.loc[{dim: new_index}]

ValueError: dimension 'x' does not have coordinate labels

Expected Output

Calling drop should return the following DataArray:

<xarray.DataArray (x: 8)>
array([0, 3, 4, 5, 6, 7, 8, 9])
Dimensions without coordinates: x

Problem Description

Since I can index arr with dimension x, I would expect drop to work as well. The current behaviour seems inconsistent to me.

Output of xr.show_versions()

INSTALLED VERSIONS
------------------
commit: None
libhdf5: 1.10.2
libnetcdf: 4.6.3

xarray: 0.12.3
pandas: 0.24.2
numpy: 1.16.4
scipy: 1.1.0
netCDF4: 1.5.1.2
pydap: None
h5netcdf: None
h5py: 2.7.0
Nio: None
zarr: None
cftime: 1.0.3.4
nc_time_axis: None
PseudoNetCDF: None
rasterio: None
cfgrib: None
iris: None
bottleneck: None
dask: 2.1.0
distributed: None
matplotlib: 3.0.3
cartopy: None
seaborn: 0.9.0
numbagg: None
setuptools: 40.8.0
pip: 19.2.1
conda: None
pytest: 4.5.0
IPython: 7.4.0
sphinx: None

[max-sixty]

Thanks for the clear report @xlambein .

In your example, x is a data variable. drop operates either on variables or indexes / 'coordinate labels'. There are no indexes here.

Does that make sense?

[shoyer]

There is still something to be said for making sel() and drop() consistent. So given that sel() is OK with integer inputs, perhaps drop() should be, too?

[max-sixty]

@xlambein can I clarify something you said above:

Calling drop should return the following DataArray:

<xarray.DataArray (x: 8)>
array([0, 3, 4, 5, 6, 7, 8, 9])
Dimensions without coordinates: x

Is this a mistype and actually you mean it should return arr (i.e. the random numbers) with two values removed?

[max-sixty]

There is still something to be said for making sel() and drop() consistent. So given that sel() is OK with integer inputs, perhaps drop() should be, too?

Yes. I worry a bit about drop becoming overloaded - it already means both variables and indexes - but given we've done that on sel, I agree.

If there's any support for splitting drop's two functions apart into two different methods, I'd also been keen on that

[xlambein]

@max-sixty Yep sorry! I got my testing code mixed up after making a change. I meant to have np.arange(10) as the values of arr (to highlight which values were removed). I edited the original comment.

[xlambein]

The way I understand it, dimensions without coordinates act as if they had np.arange(...) as their coordinates. Hence, I don't see it as overloading drop. I feel like there would still be only two behaviours to drop: on variables and on indexes.

That being said, it might be nice indeed to split drop into drop_variable and drop_index (or have any of them just be called drop).