[C++][Python] hash_any/hash_all produce incorrect results on sliced boolean arrays

[Fly-a-Kite]

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

Environment

pyarrow: 24.0.0
Python: 3.12.3
Platform: Ubuntu24.04

24.0.0 was the latest PyPI release when reproduced on 2026-05-26.

Reproducer

import pyarrow as pa
import pyarrow.compute as pc

base = pa.table({"g": [99, 10, 10], "flag": [True, False, None]})
sliced = base.slice(1)
rebuilt = pa.table(sliced.to_pydict())

def grouped_any(table):
    return (
        table.group_by("g", use_threads=False)
        .aggregate([("flag", "any")])
        .column("flag_any")
        .to_pylist()
    )

print("offset:", sliced["flag"].chunk(0).offset)
print("scalar any:", pc.any(pa.array([False, None])).as_py())
print("sliced:", grouped_any(sliced))
print("rebuilt:", grouped_any(rebuilt))

Actual Output

offset: 1
scalar any: False
sliced: [True]
rebuilt: [False]

Expected Output

sliced: [False]
rebuilt: [False]

After slice(1), the only rows in group g=10 have flag values
[False, None]. The grouped any result should therefore agree with scalar
compute.any and the zero-offset rebuilt table.

Component(s)

Python

[AlenkaF]

Thanks for opening the issue and with such a clean reproducible example!
It looks to me like the offset in the aggregate kernel is ignored. So I asked Copilot to look into the C++ code for me and it came up with a simple fix for it in the cpp/src/arrow/compute/kernels/hash_aggregate.cc:

arrow/cpp/src/arrow/compute/kernels/hash_aggregate.cc

Line 1264 in e2b4437

Impl::UpdateGroupWith(reduced, *g, bit_util::GetBit(bitmap, position));

should be something like:

Impl::UpdateGroupWith(reduced, *g, bit_util::GetBit(bitmap, input.offset + position));

I haven't tested it but I would give it a try if anyone is interested in adding a fix. Also, I can confirm this bug locally! =)

[fenfeng9]

@AlenkaF, agree with your analysis. I've verified it locally. Happy to submit a PR.

[rok]

Issue resolved by pull request 50094
#50094