Doctests don't pass with pandas 3

[jsignell]

What is your issue?

It isn't showing up in CI because we aren't running doctests in an env with pandas 3, but if you update to pandas 3 and run doctests a lot of them fail. This is despite the fact that none of the regular tests fail.

The biggest source of difference is that the default unit has changed for pd.date_range when the input is a string. It is now us rather than ns. This dtype shows up in the repr so it is very common to see datetime64[ns] in docstrings.

There are a couple solutions that works across both pandas 2 and pandas 3:

• set the units in every pd.date_range call.
• use xr.date_range (which I just learned existed) more often

The downside of those options is that they are slightly less legible. The third option is to just switch the expected output to datetime[us], but that has the downside of only passing if pandas >=3.

Note that there are several other places where the dtype is str rather than object with pandas>=3 so we might end up only getting doctests that pass with a particular major version of pandas anyways.

[jsignell]

I put up #11130 to demonstrate the scope of changes if we just make doctests pass with pandas >= 3.

[dcherian]

if we just make doctests pass with pandas >= 3.

fine by me!

pytest-accept should ease the pain of figuring out the changes :)

[jsignell]

pytest-accept should ease the pain of figuring out the changes :)

It wasn't too painful :)