ds.mean() deletes non numeric variables

[lukasbrunner]

Code Sample, a copy-pastable example if possible

import xarray as xr
ds = xr.Dataset(
    coords={'dim': range(5)},
    data_vars={'data': 'string'})
print(ds.mean('dim'))  # print 1

Problem description

I know one should probably not use strings as variables but since the possibility exists I use it every now and again because I find it convenient. However recently I encountered a somewhat strange case when taking the mean.

Calculating the mean over a dimension deletes all non-numeric variables even if they do not depend on the given dimension. Variables not depending on the dimension over which the mean is taken should be untouched I think?

# print 1
<xarray.Dataset>
Dimensions:  ()
Data variables:
    *empty*

Expected Output

I would expect the variable 'data' to still be in the dataset.

Output of xr.show_versions()

INSTALLED VERSIONS ------------------

commit: None
python: 3.6.5.final.0
python-bits: 64
OS: Linux
OS-release: 4.4.138-59-default
machine: x86_64
processor: x86_64
byteorder: little
LC_ALL: None
LANG: en_US.UTF-8
LOCALE: en_US.UTF-8

xarray: 0.10.7
pandas: 0.22.0
numpy: 1.14.3
scipy: 1.1.0
netCDF4: 1.3.1
h5netcdf: 0.5.0
h5py: 2.7.1
Nio: None
zarr: None
bottleneck: 1.2.1
cyordereddict: 1.0.0
dask: 0.17.2
distributed: 1.21.5
matplotlib: 2.2.2
cartopy: 0.16.0
seaborn: 0.8.1
setuptools: 39.0.1
pip: 9.0.3
conda: None
pytest: 3.5.0
IPython: 6.3.1
sphinx: 1.7.2

[shoyer]

This was changed somewhat recently by #2156. The way the logic works now, we try to reduce all data variables whether or not they include the relevant dimension.

The issue with the old behavior (which did keep non-numeric variables in these cases) is that some aggregations don't always preserve scalar values. Variance and standard deviation are two examples that come to mind off-hand:

>>> np.var(2)
0.0

[lukasbrunner]

Thanks for your reply and the explainer! I took me about 15min of thinking (its still early in the morning) until I realized my mistake ':) After a mean (var/std) call you expect the dataset to only contain means (variances/stds) no matter of the dimensions involved, right?!
My original thought was that 'data' in my example should not be touched at all since it is independent of the dimension I was handling.