⚠️ Upstream CI failed ⚠️

[github-actions]

Workflow Run URL

Python 3.10 Test Summary

dask/dataframe/tests/test_groupby.py::test_groupby_dropna_with_agg[disk]: AssertionError: MultiIndex level [0] are different

MultiIndex level [0] values are different (66.66667 %)
[left]:  Index(['a', 'b', nan], dtype='object', name='id1')
[right]: Index(['a', nan, 'b'], dtype='object', name='id1')
At positional index 1, first diff: b != nan
dask/dataframe/tests/test_groupby.py::test_groupby_dropna_with_agg[tasks]: AssertionError: MultiIndex level [0] are different

MultiIndex level [0] values are different (66.66667 %)
[left]:  Index(['a', 'b', nan], dtype='object', name='id1')
[right]: Index(['a', nan, 'b'], dtype='object', name='id1')
At positional index 1, first diff: b != nan

[crusaderky]

I've seen this quite a few times. It sounds like an upstream random issue in pandas although I cannot understand how you could get randomicity in it. Also worth noting that I've never seen it outside of CI.

[jrbourbeau]

Thanks @crusaderky. Where else have you seen this issue?

This issue is coming from our upstream CI build where we use the nightly version of pandas -- I wouldn't expect to see it without a dev version of pandas (though would be good to know if I'm wrong here)

[jrbourbeau]

FWIW I've pushed up a few PRs to start knocking some of these upstream compatibility issues out. @ncclementi @charlesbluca @douglasdavis @rjzamora just wanted to check in to see if you have interest / bandwidth to help get things ready for the pandas=2.0 release. Given the number of issues, a divide and conquer approach might be best 😄

[crusaderky]

Thanks @crusaderky. Where else have you seen this issue?

Sorry, I did not pay enough attention when it happened

[jrbourbeau]

Opened dask/fastparquet#832 for the dask/dataframe/io/tests/test_parquet.py::test_categories_unnamed_index[fastparquet] failure