Thanks for the report! Confirmed on main.

first bad commit: [0283df6] REF: implement Index._should_partial_index (#42227)

cc @jbrockmendel

@simonjayhawkins you never ping me on my good commits

The bad is copy and pasted from the git bisect output and is the default label. All your commits are good, but some are maybe less good than others?

I cc you on these since IIRC you requested me to do so.

Just kidding around simon. my bad that wasn't clear.

first bad commit: [0283df6] REF: implement Index._should_partial_index (#42227)

in #42227, if not self._should_compare(target) and not is_interval_dtype(self.dtype): was changed to if not self._should_compare(target) and not self._should_partial_index(target): in Index.get_indexer in pandas/core/indexes/base.py

This is the changed code that the code sample in the OP hits.

it appears is_interval_dtype(self.dtype) is True whereas self._should_partial_index(target) is False for the code sample accounting for the changed behavior.

In _bins_to_cuts there is a if isinstance(bins, IntervalIndex): fastpath. When we get there with the OP example, we have converted the Timestamp to a float64. The bins.get_indexer(x) result is now correct, it's the x we have at this point that is the problem.

moving to 1.4.3

Is there a recommended workaround? Some of my code stopped working and it took me a couple of hours of debugging to realize that pd.cut was broken in recent Pandas releases, and then find this page.

@dom-insytesys comments like this are not helpful. pandas is a community project. Feel free to post a workaround for the benefit of other users or submit a PR to fix and it will be included in the next pandas patch release.

@simonjayhawkins I'm so sorry if my comment annoyed you. I didn't intend to be rude.

What I was hoping to convey was firstly that this is not an obscure corner case. One of the most common uses of pd.cut() is to bin time series data into windows. So I assume this bug affects a lot of Pandas users, myself included.

Until Pandas 1.4.3 is released, I need to figure out some sort of a workaround. Obviously, because Pandas is open-source (hoorah!) I can theoretically get the source code, find the bug, and fix it myself. Or I can use pd.cut() source as a guide to writing my own alternative version.

However, that has a pretty steep learning curve (I'm not very proficient in Python, nor familiar with Pandas internals, so I'd guess it would take me a day or more), and I was hoping that there was some sort of quick workaround that would be childishly obvious to guys like you who have a good understanding of this code.

For example, @jbrockmendel provides an explanation above that I think says that the bug involves the conversion of Timestamp data to float64. That suggests to me that if I convert the timeseries data that needs binning from datetime64[ns] to a float64 (seconds since epoch?), that will avoid the bug. I can't figure out how to do this (e.g. "astype('float64') produces TypeError) , but I bet it is trivial for you. That's why I was hoping one of the Pandas team would say, "here's a quick way to avoid the bug: just convert your data from datetime64 to float64 using XXXX function". That would help affected users get quickly back on course.

Again, I apologize wholeheartedly if I caused offense. I hugely appreciate the work of the volunteer community that works on Pandas. It's an amazing project. I'm just asking for some help, with the expectation that any guidance would probably be useful to other users who end up stumbling to this page for the same reason.

The quick fix is described in my bug report. Use strings.

In the spirit of attempting to help out others, here's my crude effort at a workaround.

Firstly, some code to illustrate the bug:

# Example code to illustrate pd.cut() bug
df = pd.DataFrame({"clock":pd.date_range(start='2020/1/1', end='2020/1/2', periods=25), "randomdata":[random.random() for _ in range(25)]})
windows = pd.IntervalIndex.from_breaks(pd.date_range(start='2019/12/31 23:30', end='2020/1/1 23:30', periods=13))
bins = pd.cut(df.clock, bins=windows)
df['randomdata'].groupby(bins).mean()

This produces the following output:

clock
(2019-12-31 23:30:00, 2020-01-01 01:30:00] NaN
(2020-01-01 01:30:00, 2020-01-01 03:30:00] NaN
(2020-01-01 03:30:00, 2020-01-01 05:30:00] NaN
(2020-01-01 05:30:00, 2020-01-01 07:30:00] NaN
(2020-01-01 07:30:00, 2020-01-01 09:30:00] NaN
(2020-01-01 09:30:00, 2020-01-01 11:30:00] NaN
(2020-01-01 11:30:00, 2020-01-01 13:30:00] NaN
(2020-01-01 13:30:00, 2020-01-01 15:30:00] NaN
(2020-01-01 15:30:00, 2020-01-01 17:30:00] NaN
(2020-01-01 17:30:00, 2020-01-01 19:30:00] NaN
(2020-01-01 19:30:00, 2020-01-01 21:30:00] NaN
(2020-01-01 21:30:00, 2020-01-01 23:30:00] NaN
Name: randomdata, dtype: float64

If I understand correctly, the bug relates to the fact that df.clock has a Timestamp-like dtype. This means it gets on a fasttrack code path. That can be avoided by converting the datetime to a string. Changing the line:

bins = pd.cut(df.clock.dt.strftime("%Y/%m/%d %H:%M:%S.%f'"), bins=windows)

Produces the desired output:

clock
(2019-12-31 23:30:00, 2020-01-01 01:30:00] 0.686914
(2020-01-01 01:30:00, 2020-01-01 03:30:00] 0.462383
(2020-01-01 03:30:00, 2020-01-01 05:30:00] 0.667617
(2020-01-01 05:30:00, 2020-01-01 07:30:00] 0.503656
(2020-01-01 07:30:00, 2020-01-01 09:30:00] 0.615401
(2020-01-01 09:30:00, 2020-01-01 11:30:00] 0.428048
(2020-01-01 11:30:00, 2020-01-01 13:30:00] 0.631899
(2020-01-01 13:30:00, 2020-01-01 15:30:00] 0.357134
(2020-01-01 15:30:00, 2020-01-01 17:30:00] 0.624220
(2020-01-01 17:30:00, 2020-01-01 19:30:00] 0.495276
(2020-01-01 19:30:00, 2020-01-01 21:30:00] 0.617120
(2020-01-01 21:30:00, 2020-01-01 23:30:00] 0.112893
Name: randomdata, dtype: float64

This seems to be working for me. But I'm a little concerned that it is fragile. (What happens, for example if IntervalIndex is in a different timezone? etc.)

moving to 1.4.4

Closing as a duplicate of #46218. Discussion of a solution seems further along there

I think you closed the wrong one.

Thanks @sergiykhan. Reopened

I opened #47771 with a potential fix.

It doesn't fix the root cause inside cut of converting the datetime64 data to numeric but then still passing that to pd.cut (although in hindsight that should actually also be a straightforward change), but it restores the behaviour of IntervalIndex.get_indexer to accept numeric target values for datetime64 intervals.