oversk7 opened a new issue, #54906:
URL: https://github.com/apache/arrow/issues/54906
### Describe the bug, including details regarding any error messages,
version, and platform.
`ascii_upper`, `ascii_lower` and `ascii_swapcase` return an invalid array
when the input has `offset == 0` but its first value offset is not 0. Such an
input is valid. The format only says that the first offset is generally 0, and
`ValidateFull()` accepts the input.
Reproducer (pyarrow 26.0.0 from PyPI, Python 3.12, Windows 11):
```python
import struct
import pyarrow as pa
import pyarrow.compute as pc
offsets = pa.py_buffer(struct.pack("<3i", 5, 8, 8))
data = pa.py_buffer(b"xxxxxabc")
arr = pa.Array.from_buffers(pa.string(), 2, [None, offsets, data])
arr.validate(full=True) # OK
print(arr.to_pylist()) # ['abc', '']
out = pc.ascii_upper(arr)
out.validate(full=True)
# pyarrow.lib.ArrowInvalid: Offset invariant failure: offset for slot 1 out
of bounds: 8 > 3
```
Expected: `['ABC', '']`.
- The same happens with `ascii_lower` and `ascii_swapcase`, and with
`large_string` input.
- `utf8_upper` and `utf8_lower` are not affected.
The cause is in `StringDataTransform`
([scalar_string_ascii.cc](https://github.com/apache/arrow/blob/373d64c02cb23663b63c5853099fca43a46c39fc/cpp/src/arrow/compute/kernels/scalar_string_ascii.cc#L453-L487)):
- When `input.offset == 0`, the input offsets buffer is reused unchanged.
- The transformed data, however, is written to a new buffer of
`offsets[length] - offsets[0]` bytes, starting at position 0.
- So the reused offsets (5, 8, 8) point past the end of the 3-byte output
data.
Reusing the offsets is only correct when `offsets[0] == 0`. Otherwise the
existing `else` branch, which rebases the offsets by `offsets[0]`, gives the
right result.
This was found with an AI-assisted analysis of the C++ code against the
columnar format specification. The reproducer above was run against pyarrow
26.0.0.
### Component(s)
C++
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
