regression in 0.23.0 on rolling max with DatetimeIndex.

[lexual]

Behaviour of .rolling() with datetimeindex appears to have changed or regressed between 0.22 & 0.23 versions of pandas.

See below for cut-down test case exhibiting behaviour.

Also see below for different output produced on same code on same data with 2 different pandas versions.

import numpy as np
import pandas as pd


def main():
    print('pandas', pd.__version__)
    n = 10
    index = pd.date_range(
        start='2018-1-1 01:00:00',
        freq='1min',
        periods=n,
    )
    s = pd.Series(
        data=0,
        index=index,
    )
    s.iloc[1] = np.nan
    s.iloc[-1] = 2
    print(s)
    maxes = s.rolling(window=f'{n}min').max()
    print(maxes.value_counts(dropna=False))


if __name__ == '__main__':
    main()

ACTUAL OUTPUT

pandas 0.22.0      
2018-01-01 01:00:00    0.0             
2018-01-01 01:01:00    NaN             
2018-01-01 01:02:00    0.0             
2018-01-01 01:03:00    0.0             
2018-01-01 01:04:00    0.0             
2018-01-01 01:05:00    0.0             
2018-01-01 01:06:00    0.0             
2018-01-01 01:07:00    0.0             
2018-01-01 01:08:00    0.0             
2018-01-01 01:09:00    2.0             
Freq: T, dtype: float64                
0.0    9           
2.0    1           
dtype: int64       
pandas 0.23.0      
2018-01-01 01:00:00    0.0             
2018-01-01 01:01:00    NaN             
2018-01-01 01:02:00    0.0             
2018-01-01 01:03:00    0.0             
2018-01-01 01:04:00    0.0             
2018-01-01 01:05:00    0.0             
2018-01-01 01:06:00    0.0             
2018-01-01 01:07:00    0.0             
2018-01-01 01:08:00    0.0             
2018-01-01 01:09:00    2.0             
Freq: T, dtype: float64                
0.0    10          
dtype: int64       

[jreback]

can you create the input programatically here (and update the top )

[lexual]

done @jreback , updated simpler test-case, completely in code in 1st comment.

[lexual]

$ git bisect bad 
39e7b6916b07982240bac87132848fb2665806a2 is the first bad commit
commit 39e7b6916b07982240bac87132848fb2665806a2
Author: Matt Kirk <[email protected]>
Date:   Wed Feb 14 18:13:19 2018 +0700

    Performance increase rolling min max (#19549)

:040000 040000 6543b93964c04b7f71c14a9992c8f70606e02c77 a7f4578db8843fdb0f9dbe02322584ced5628dd9 M      asv_bench
:040000 040000 3f68b43b26722b2f7a2874433e88965ca5d8bfda 2eaff46202589c76a8d85251cfdb052dd38660f1 M      doc
:040000 040000 e28eb2f13f13846afe64b82ce492a5a877b98be7 18784432eb0d9655cca3de34a02f5f514c44eaff M      pandas
:100755 100755 2332503e558ed7ee9d57e3b3153ae674c1d292ad c66979dd19ef039cd4b5172c65276d1340507631 M      setup.py

[lexual]

cc @hexgnu

[mroeschke]

This looks fixed on master and could still use a regression test.

In [9]: pd.__version__
Out[9]: '0.24.0.dev0+911.g24ab22f75'

In [10]: s
Out[10]:
2018-01-01 01:00:00    0.0
2018-01-01 01:01:00    NaN
2018-01-01 01:02:00    0.0
2018-01-01 01:03:00    0.0
2018-01-01 01:04:00    0.0
2018-01-01 01:05:00    0.0
2018-01-01 01:06:00    0.0
2018-01-01 01:07:00    0.0
2018-01-01 01:08:00    0.0
2018-01-01 01:09:00    2.0
Freq: T, dtype: float64

In [11]: maxes = s.rolling(window=f'{n}min').max()

In [12]: maxes.value_counts(dropna=False)
Out[12]:
0.0    9
2.0    1
dtype: int64