Undesired automatic conversion of 'dtype' based on variable units 'days'?

[tlogan2000]

import xarray as xr
from glob import os
import numpy as np
import urllib.request
xr.set_options(enable_cftimeindex=True)

url = 'https://github.com/Ouranosinc/xclim/raw/master/tests/testdata/NRCANdaily/nrcan_canada_daily_tasmax_1990.nc'
infile = r'~/nrcan_canada_daily_tasmax_1990.nc'
urllib.request.urlretrieve(url,infile)

freqs = ['MS']  # , 'QS-DEC', 'YS']
ds = xr.open_dataset(infile)
su = (ds.tasmax > 25.0+273.15) * 1.0
for f in freqs:

    output = su.resample(time=f).sum(dim='time')
    output.attrs['units'] = 'days'
    output.attrs['standard_name'] = 'summer_days'
    output.attrs['long_name'] = 'summer days'
    output.attrs['description'] = 'Number of days where daily maximum temperature exceeds 25℃'

    # use original file as template
    ds1 = xr.open_dataset(infile, drop_variables=['tasmax', 'time','time_vectors','ts'])
    ds1.coords['time'] = output.time.values
    ds1['su'] = output

    comp = dict(zlib=True, complevel=5)
    encoding = {var: comp for var in ds1.data_vars}
    print(os.path.basename(infile).replace('.nc', '_SummerDays-' + f) + ' : writing ' + f + ' to netcdf')
    # with dask.config.set(pool=ThreadPool(4)):

    ds1.to_netcdf('~/testNRCANsummerdays.nc', format='NETCDF4', encoding=encoding)

    ds2 = xr.open_dataset('~/testNRCANsummerdays.nc')
    print(ds1)
    print(ds2)
    print(ds1.su.max())
    print(ds2.su.max())

Problem description

I am calculating a climate index 'summer days' ('su') from daily maximum temperatures. Everything goes fine but when I reread a newly created output .nc file my 'su' dtype has changed from float to timedelta64. This seems to be due to the units ('days') that I assign to my newly created variable which when read by xarray must trigger an automatic conversion? If I alter the units to 'days>25C' everything is ok.

Is there a way to avoid this behavior and still keep my units as 'days' which is the CF_standard for this climate index calculation? (note there a large number of cases such as this - wet days, dry-days etc etc all of which have 'days' as the expected unit

[shoyer]

see related issues #1621 (and a work-in-progress fix #2105)

[tlogan2000]

Great thx