How to add an array of timedeltas with cftime index?

[ahuang11]

This raises an error

import cftime
import numpy as np
import xarray as xr
import pandas as pd
da = xr.DataArray(
    data=xr.CFTimeIndex([cftime.DatetimeProlepticGregorian(1, 1, 1)])
)
td = pd.to_timedelta(['2H', '3H']).values

da + td

UFuncTypeError: ufunc 'add' cannot use operands with types dtype('O') and dtype('<m8[ns]')

Is the only solution to loop through?

[spencerkclark]

Oops I seemed to have missed this question. To add timedeltas to cftime objects, you must use datetime.timedelta objects instead of timedelta64[ns] values (this is what cftime objects use for timedelta arithmetic internally). E.g. the following works for your last line:

da + td.to_pytimedelta()

However, there is currently no way to store datetime.timedelta objects directly within a DataArray (xarray will automatically cast them to timedelta64[ns] values). The above example happens to work because we are adding a NumPy array of datetime.timedelta objects to a DataArray.

This is not ideal, but I do not know what the best solution would be. Similar to what we have in CFTimeIndex itself, we could add logic to cast timedelta64[ns] values to datetime.timedelta objects when doing addition or subtraction of timedelta-like DataArrays from cftime DataArrays. This would be helpful in a lot of cases, but that said, it would not work for timedelta objects greater than ~292 years due to the limits on the range of timedelta64[ns] values.

[spencerkclark]

This would be helpful in a lot of cases, but that said, it would not work for timedelta objects greater than ~292 years due to the limits on the range of timedelta64[ns] values.

With new developments in pandas (see more discussion in #7493) we may no longer need to worry about the nanosecond-precision limitation, which would conveniently eliminate this issue.