Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

Thanks @brandon-b-miller !

/merge

I have a failing test locally - looks like there's an existing issue with the way we read in frequencies that is being exposed by this. Looking into it

Looks like we have some conflicts, probably from #18751.

Argh, there's a couple tricky cases that have bubbled up through the tests that aren't simple fixes.

Firstly, pandas adjusts the frequency when slicing with a step. Theres no equivalent computation in cuDF python, instead we simply apply the original frequency which causes a mismatch.

>>> ps
2001-01-01    1
2001-01-02    2
2001-01-03    3
2001-01-04    4
2001-01-05    5
Freq: D, dtype: int64
>>> ps[::2]
2001-01-01    1
2001-01-03    3
2001-01-05    5
Freq: 2D, dtype: int64

Secondly, pandas has deprecated certain frequencies such as M in favor of others such as ME (month end) and automatically promotes them, probably to push a user towards using an exact number of days or explicitly request the end of the month

>>> pd.date_range(start="1990-01-01", periods=10, freq="M").freq
<python-input-11>:1: FutureWarning: 'M' is deprecated and will be removed in a future version, please use 'ME' instead.
<MonthEnd>

However cudf.DateOffset doesn't implement MonthEnd and if we just try and keep track of it as months, we error later on anytime we hit to_pandas as pandas tries to validate the frequency against the data itself and you'll get

Inferred frequency ME from passed values does not conform to passed frequency <DateOffset: months=1> 

Finally, [this section of the code here ] attempts to validate that there is equal spacing between datetime values when initting a datetimeindex. This wasn't being hit before, because freq wasn't carried through to this point, but now that it's being hit, it doesn't look like the logic works for months which can have len({28, 30, 31}) == 3 days.

This will likely take a little more time to fix since carrying it through has elucidated a few areas where our freq support isn't complete/working more generally.

Converted this back to in progress in light of the recent findings, pushed some changes, still some failures.