xarray Datasets are not correctly serialized to netcdf

[bocklund]

import numpy as np
import xarray as xr
import datetime
data = xr.DataArray(np.random.randn(2, 3), coords={'x': ['a', 'b']}, dims=('x', 'y'))
ds = xr.Dataset({'z': data})
ds.to_netcdf('netcdf_test.nc')  # works
ds.attrs['created'] = datetime.datetime.now()
ds.to_netcdf('netcdf_test_time.nc')  # fails

I thought about opening this issue upstream, but their solution for handling datetimes as DataArrays is precisely to use the attrs as metadata for that array. I don't think they would support generic serialization of datetimes that are in attrs.

It seems like we should either not use datetimes or just serialize/deserialize ourselves.

[richardotis]

I would be comfortable converting datetimes to strings on creation. We don't use them in any library code, from a readability standpoint it's the same for the user, and it's trivial for downstream code to convert the string to a datetime object if programmatic access is needed. Also, it would be a one line change.