With pandas 3, pd.Series(datetime.datetime(...)) will produce a Series with np.datetime64[us] values instead of np.datetime64[ns] values, so this conversion now warns.

As far as I can tell, whether reference_dates started as datetime.datetime objects or np.datetime64[ns] values was not material to this test, so I removed the conversion to datetime.datetime to avoid the conversion warning under pandas 3 (the times would previously get converted back to datetime64[ns] values in the DataArray constructor).

da.to_dict() produces datetime.datetime objects, which under pandas 3 lead to a conversion warning in the DataArray constructor.

if we have this pattern in multiple modules, it might be worth adding the code as a special context manager to xarray.tests.__init__. Something like this might work (I didn't check):

from contextlib import contextmanager
import warnings

@contextmanager
def ignore_warnings(category=None, pattern=None):
    if category is None and pattern is None:
        raise ValueError("need at least one of category and pattern")

    try:
        with warnings.catch_warnings():
            warnings.filterwarnings("ignore", message=pattern, category=category)
            yield
    finally:
        pass

Thanks—I ended up switching to marking these tests with @pytest.mark.filterwarnings("ignore:Converting non-nanosecond"), since that is a pattern we use elsewhere in the tests already.

Use of datetime.datetime objects was immaterial to this test, so we use pd.date_range to produce the dates instead to avoid the non-nanosecond conversion warning.

This is needed since dates sometimes consists of datetime.datetime objects, which leads to a conversion warning under pandas 3.

pandas.to_datetime will infer the precision from the input in pandas 3, so we explicitly specify the desired precision now.