⚠️ Nightly upstream-dev CI failed ⚠️

[github-actions]

Workflow Run URL

Python 3.12 Test Summary

xarray/tests/test_missing.py::test_scipy_methods_function[barycentric]: FutureWarning: 'd' is deprecated and will be removed in a future version. Please use 'D' instead of 'd'.
xarray/tests/test_missing.py::test_scipy_methods_function[krogh]: FutureWarning: 'd' is deprecated and will be removed in a future version. Please use 'D' instead of 'd'.
xarray/tests/test_missing.py::test_scipy_methods_function[pchip]: FutureWarning: 'd' is deprecated and will be removed in a future version. Please use 'D' instead of 'd'.
xarray/tests/test_missing.py::test_scipy_methods_function[spline]: FutureWarning: 'd' is deprecated and will be removed in a future version. Please use 'D' instead of 'd'.
xarray/tests/test_missing.py::test_scipy_methods_function[akima]: FutureWarning: 'd' is deprecated and will be removed in a future version. Please use 'D' instead of 'd'.
xarray/tests/test_missing.py::test_interpolate_pd_compat_non_uniform_index: FutureWarning: 'd' is deprecated and will be removed in a future version. Please use 'D' instead of 'd'.

[keewis]

this looks real, and possibly related to #9081. I think this can be fixed by specifying the precision in the strategy, but pandas raises lots of warnings if I add dtype="datetime64[ns]".

cc @ilan-gold, @dcherian

[ilan-gold]

@keewis Thanks for the cc, but I really have no idea what is going on here. As I said in my PR that fixed a small roundtrip bug, this whole date-time business makes very little sense to me. It should in theory be possible to handle these as extension arrays and preserve the format, but that breaks a bunch of tests and code where the format is not expected (or this other custom data structure is expected) in very insane ways that I couldn't wrap my head around.

[keewis]

I think the issue is not in the code, but in what we expect to get as the result of the roundtrip: if you look at the generated df, the dtype is object, while xarray returns an extension array (the former appears to have changed with pandas' main). So really, I believe the fix would be to somehow make sure we get hypothesis to generate a datetime column with time zones.