Debugging non-deterministic segfault on 0.61.0

Numba Support: How do I do ...?
1

Hi, I’m looking at upgrading from 0.60.0 to 0.61.0 and I’m hitting a segfault in our test suite. I don’t yet have a minimal reproducer, but I have a test that:

  • Consistently passes on 0.60.0 on my local macos machine and a linux intel machine
  • Consistently passes on 0.61.0 on my local macos machine
  • Consistently segfaults on 0.61.0 on my linux intel machine

It segfaults with:

Fatal Python error: Segmentation fault

Current thread 0x00007587aaccac40 (most recent call first):
  File "/home/.pyenv/versions/3.12.7/envs/env/lib/python3.12/site-packages/numba/experimental/jitclass/boxing.py", line 61 in wrapper
...

If I start pruning to narrow down the cause, the segfault ends up being non-deterministic where sometimes the test passes, other times it segfaults. Also, it segfaults if I run with NUMBA_OPT=1 but not with NUMBA_OPT=0. In short, I’m looking for some advice on where I might be able to go with this. Am I right in saying that this is a failure during compilation rather than runtime? Are there any further debug flags or options I can try? I’ve tried setting NUMBA_DEBUG=1 but nothing is printed before the segfault

Thanks