[BUG] INT96 Parquet timestamp decoding silently produces wrong values outside int64 range
Open
bug
cuIO
- Dominant language
- C++
- Stars
- 9.8k
- Forks
- 1.1k
- Avg merge
- 3d 6m
- Merged PRs (30d)
- 278
Description
**Describe the bug**
When reading Parquet files with INT96 timestamps, cuDF silently produces incorrect values for timestamps outside the representable range of the requested output type.
**Steps/Code to reproduce bug**
[TestCudfInt96.java](https://gist.github.com/piotrrzysko/700d361a8ed334112418bb32163c488b) demonstrates two scenarios in which the bug occurs.
**Expected behavior**
cuDF should be able to read the full INT96 range without silent data corruption
**Environment overview (please complete the following information)**
- Environment location: Cloud (AWS)
- Method of cuDF install: Maven dependency (`ai.rapids:cudf:26.06.0-cuda13`)
Contributor guide
Assessment
This issue has not been assessed yet.