More robust checks for cftime input to xr.date_range

[kafitzgerald]

What is your issue?

I recently ran into a case where the logic in xr.date_range isn't quite robust enough (detail below).

This is something we use in a downstream package and while I can definitely add some additional checks there I figured I'd log an issue here and see if you all might be open to addressing this in Xarray. Happy to work on a PR if that's helpful (though I'm not terribly familiar with the code).

The following example should replicate the failure:

import xarray as xr

times = xr.cftime_range(start='2020-01-01', end='2021-12-31', freq='D')
xr.date_range(start=times[0], end=times[-1], freq='D', calendar=times.calendar)

Resulting in:

TypeError: Cannot convert input [2020-01-01 00:00:00] of type <class 'cftime._cftime.DatetimeGregorian'> to Timestamp

Because of the "standard" calendar and an error that doesn't match the one Xarray is looking for it uses the pd.date_range rather than the cftime equivalent and then fails.

[spencerkclark]

Thanks @kafitzgerald—indeed I think we'd take a PR to fix this. It makes the dispatch logic a bit more complicated, but it seems natural that we should use the cftime_range code path in the event that cftime objects are passed to start and / or end, regardless of the specified calendar, if use_cftime=None.