comparing vs. subtracting different calendars

[mathause]

The following snipped used to raise a TypeError. With the changed behavior of __richcompare__ this now raises a ValueError in __sub__ which pandas ignores:

import pandas as pd
import cftime

time_1 = cftime.DatetimeGregorian(2000, 1, 1)
time_2 = cftime.DatetimeProlepticGregorian(2001, 1, 1)

pd.Series([time_1, time_2]).rank(method="dense", numeric_only=False)

This is used in xarray to avoid combining datasets with different calendars, which does not really work anymore with this change. I don't know why pandas ignores the error (and it's not silent). So it's not really a cftime issue but I thought I report it here anyway.

• Do you plan to allow __sub__ between different calendars?
• Should __sub__ raise a TypeError instead of a ValueError

[jswhit]

PR #234 enabled __richcmp__ between different calendars. You are right that the error should be changed to a TypeError. I will create a PR for that. As far as enabling __sub__ between different calendars, that's a tricky one. It's possible, but which calendar type would be returned? PR #234 also added a change_calendar method that can be used to achieve this - I think it's probably better to force users to explicity do d1 - d2.change_calendar(d1.calendar).

[jswhit]

PR #236 merged. @mathause please let us know whether that fixes the xarray tests

[spencerkclark]

Great, thanks @jswhit -- that indeed fixed the issue @mathause identified. The removal of utime (#235) resulted in some nc-time-axis-related test failures in xarray as well (it appears nc-time-axis makes extensive use of it).

It was fairly straightforward to remove nc-time-axis's use of utime (see my updates to jswhit/nc-time-axis#1). Do you think we should try and get SciTools/nc-time-axis#51 merged there? It sounds like they would make a new release after that: SciTools/nc-time-axis#54 (comment).

[mathause]

Unfortunately it did not work. Actually pandas ignores the TypeError silently 🤦

https://github.com/pandas-dev/pandas/blob/029907c9d69a0260401b78a016a6c4515d8f1c40/pandas/_libs/algos.pyx#L87

Still I think TypeError is the right one to raise - so I guess we have to fix this in xarray & am not sure how to best do this. Pity, the old version was quite elegant.

---

Edit: we can probably do

pd.Series([time_1, time_2]).diff()

which will raise the appropriate error.

[spencerkclark]

Oops, I don't know how I missed that that test still failed in the xarray test suite...thanks @mathause for correcting me.

I agree we can work around this in xarray.

[mathause]

I think from my side this can be closed - thanks @jswhit & @spencerkclark