[Python] floor_temporal and ceil_temporal bugs on DST changeover.

[haraldnh]

Describe the bug, including details regarding any error messages, version, and platform.

When running a ceil_temporal operation that results in a timestamp hitting the spring leap forward point, pyarrow errors out with 'Local time does not exist'. Similar during the fall back switchover, both floor_temporal and ceil_temporal will error with 'Local time is ambiguous'.

Observed with pyarrow 19.0.1 on MacOS/arm64 and Linux/amd64.

Component(s)

Python

[raulcd]

Could you provide a small reproducible example?
cc @rok @AlenkaF

[rok]

This is unfortunately the current behavior. We do have an effort to handle ambiguous/nonexistent local times that I'd like to complete at some point.

How much in a hurry are you @haraldnh ?

As a workaround perhaps you could devise something with pc.is_dst kernel, but it could be quite unwieldy.

[haraldnh]

Just to start with the example:

import pyarrow as pa
import pyarrow.compute as pc
from datetime import datetime
from zoneinfo import ZoneInfo

tz = ZoneInfo('America/New_York')
data1 = pa.array([datetime(2025,3,9,1,30,tzinfo=tz)])
fail1 = pc.ceil_temporal(data1, unit='hour')

data2 = pa.array([datetime(2025,11,2,1,30,tzinfo=tz)])
fail2 = pc.floor_temporal(data2, unit='hour')
fail3 = pc.ceil_temporal(data2, unit='hour')

The two datetime hh:mm values from python are: data1 is at 01:30 EST (06:30 UTC),
while data2 is at 01:30 EDT (05:30 UTC).

[haraldnh]

As a workaround perhaps you could devise something with pc.is_dst kernel, but it could be quite unwieldy.

I'll probably start with documenting it as a problem, and asking my users to "don't do that". 😸

When making an hourly or shorter time series it can just as well be done in UTC, so I can make some workarounds for that (except for the half-hour zones with dst, etc.). While daily series only hit the problem in South America (where those that still do often start DST at midnight).

The main problem is series running on 'n' hour frequency, where it's a bit impossible to know up-front if there will be a problem or not.