[C++] Support DST-aware timestamp + interval addition

[icexelloss]

Describe the enhancement requested

In Pandas, there is a way to do "day offset shift" that is day light saving aware, e.g.

import pyarrow as pa
import pandas as pd
import pyarrow.compute as pc

begin = pd.Timestamp('20250308 9:30', tz='America/New_York')
print(begin)
print(begin + pd.DateOffset(days=1))
print(begin + pd.DateOffset(days=2))

Outputs:

2025-03-08 09:30:00-05:00
2025-03-09 09:30:00-04:00
2025-03-10 09:30:00-04:00

However, there is no way to do such operations as far as I am aware with arrow compute, the closest I found is this

begin = pa.scalar(pd.Timestamp('20250309', tz='America/New_York'))
pc.add(begin, pa.scalar((0, 1, 0), type=pa.month_day_nano_interval()))

However, it seems not supported, I also didn't find any specific arrow compute function to do this.

IMO this would be really useful to deal with time in non-UTC timezone.

Component(s)

C++

[pitrou]

@rok Did we have any plans to add this? Are there any difficulties to know of?

[pitrou]

I think we can follow Pandas' lead where days and larger units are considered logical while hours and smaller units are considered physical:

>>> begin
Timestamp('2025-03-08 09:30:00-0500', tz='America/New_York')

# Logical calendar addition
>>> begin + pd.DateOffset(days=2)
Timestamp('2025-03-10 09:30:00-0400', tz='America/New_York')
>>> begin + pd.DateOffset(months=1)
Timestamp('2025-04-08 09:30:00-0400', tz='America/New_York')

# Physical duration addition
>>> begin + pd.DateOffset(hours=48)
Timestamp('2025-03-10 10:30:00-0400', tz='America/New_York')
>>> begin + pd.DateOffset(seconds=48*3600)
Timestamp('2025-03-10 10:30:00-0400', tz='America/New_York')

[rok]

@pitrou I don't think we currently support this, but we definitely should. Currently one could perhaps use temporal rounding as a workaround, so something like pc.round(pc.add(zoned_timestamps + some_duration), unit='day') (doing this from memory, it's likely wrong API-wise). But a proper "add_temporal_unit" kernel would be nicer.

edit: or indeed considering units larger than hour logical for add, but that would be a behavior change.

[pitrou]

Also, we should be careful to follow Pandas' logic when adding/subtracting a mixed interval:

>>> end = begin + pd.DateOffset(days=1, hours=24)
>>> end - pd.DateOffset(days=1, hours=24) == begin
True

>>> end == begin + pd.DateOffset(days=1) + pd.DateOffset(hours=24)
True
>>> end == begin + pd.DateOffset(hours=24) + pd.DateOffset(days=1)
False

>>> end - pd.DateOffset(days=1) - pd.DateOffset(hours=24) == begin
False
>>> end - pd.DateOffset(hours=24) - pd.DateOffset(days=1) == begin
True

[rok]

I just recalled we do have several temporal subtraction kernels that (I believe) look calendar units for units greater than hour.