REGR: period_range giving incorrect values for large datetimes

[dsaxton]

From the pandas-dev mailing list:

I encountered a bug in pandas 1.1.0 after updating from 1.0.5 that I wanted to pass along. If a period range is created that extends past the maximum timestamp date of 2262-04-11, then the period range is shifted back by an hour after this date. The screenshots below show an example with the shifted times highlighted in yellow.

import pandas as pd
from datetime import datetime

start = datetime(2000, 1, 1)
end = datetime(2500, 1, 1)

idx = pd.period_range(start, end, freq="6H")
assert idx[-1].hour % 2 == 0  # raises

idx
# PeriodIndex(['2000-01-01 00:00', '2000-01-01 06:00', '2000-01-01 12:00',
#              '2000-01-01 18:00', '2000-01-02 00:00', '2000-01-02 06:00',
#              '2000-01-02 12:00', '2000-01-02 18:00', '2000-01-03 00:00',
#              '2000-01-03 06:00',
#              ...
#              '2499-12-29 17:00', '2499-12-30 23:00', '2499-12-30 05:00',
#              '2499-12-30 11:00', '2499-12-30 17:00', '2499-12-31 23:00',
#              '2499-12-31 05:00', '2499-12-31 11:00', '2499-12-31 17:00',
#              '2500-01-01 23:00'],
#             dtype='period[6H]', length=730489, freq='6H')

pd.__version__
# '1.1.0'

[dsaxton]

577de1c is the first bad commit
commit 577de1c
Author: jbrockmendel [email protected]
Date: Wed Apr 15 12:20:54 2020 -0700

REF: de-duplicate code in libperiod (#33491)

pandas/_libs/tslibs/period.pyx | 108 +++++++++++++++--------------------------
1 file changed, 39 insertions(+), 69 deletions(-)

PR #33491
cc @jbrockmendel

[jorisvandenbossche]

@dsaxton if something is a regression compared to 1.0, you can always directly add a 1.1.x milestone (which doesn't mean it will get done of course, but then at least it shows up in the overview of issues that are good to have for the bug-fix release)