issue_comments

Will work on this issue in the next three weeks if no one else does till then.

@shoyer A fix in pandas that I can imagine is to modify the pd.Index to retain the dtype similar to the PandasIndexAdapter in xarray/core/indexing. However, it seems to me that their design it to coerce any float to float64, so I did not feel they would consider this to be a bug. I would try to dig into such solution more if you recommend so.

I added a test case in which the coords are float16 (in addiction to the scalar case)to show that casting to the coords type is required.

The lowest level in which I can access the coords type is in the indexing/remap_label_indexers where I could get the coords from the data_obj. In the latest code, casting is done just before calling convert_label_indexer

Sorry if I am missing the point, and thank you in advance for any further feedback.

The case also arises if coords are float16. e.g.

data_values = np.arange(4) coord_values_16 = np.asarray(float_values, dtype='float16') array_16 = DataArray(data_values, [('float16_coord', coord_values_16)]) array_16.sel(float16_coord=float_values[1:3])

@jfpiolle Thank you very much for the clarification

@shoyer @max-sixty I really appreciate your time and feedback. I will work on it within a week.

@jfpiolle Would you please share the files, or the code generating them?

As far as I understand we cannot have an attribute of type list. The closest thing I can think of is a variable length type as below

from netCDF4 import Dataset f = Dataset("example.nc","w") vlen_t = f.createVLType(numpy.int32, "phony_vlen") x = f.createDimension("x",3) y = f.createDimension("y",4) vlvar = f.createVariable("phony_vlen_var", vlen_t, ("y","x")) data =[] for n in range(0, len(x)*len(y)): data.append(numpy.arange(random.randint(1,10),dtype="int32")) data = numpy.reshape(data,(len(y),len(x))) vlvar[:] = data f.close() Using files generated this way, I did not get an error using mf_dataset.

@dcherian I sent a PR last week. Any feedback would be appreciated

I will handle this issue this Saturday if no one else handle it by then.

@Max-sixty I am delighted to be here. Thank you very much!

I modified as per your comments, please let me know if there are further modifications .

I am getting a different error for linux py36

Linux py36 UnsatisfiableError: The following specifications were found to be in conflict: - cfgrib[version='>=0.9.2'] - h5py - pynio

@dcherian I would like to send a PR and work on this issue if there is no problem in this.

@shoyer I would like to work on this issue (hopefully my first contribution).

I believe the error should indicate that we cannot groupby an empty group. This would be consistent with the documentation:

``` DataArray.groupby(group, squeeze: bool = True, restore_coord_dims: Optional[bool] = None) Parameters groupstr, DataArray or IndexVariable Array whose unique values should be used to group this array. If a string, must be the name of a variable contained in this dataset. ````

For the groupby class, we can raise an error in case the groupby object instantiation does not specify values to be grouped by. In the case above:

``` ValueError: variable to groupby must not be empty

```

If you find such a solution acceptable, I would create a pull request.

Thank you in advance