Partly generated by an agent, but I did double-check with my brain on 😉.
cc @itamarst in case you have some insights or you have seen something similar when putting this empirical tests together.
I may try to investigate further on my Windows VM to see if I can at least reproduce.
The problem
The Windows pylatest_conda_forge_openblas CI job intermittently crashes while running:
PYTHONPATH=. python -X faulthandler tests/empirical_scope_observation.py blas
+ python -X faulthandler tests/empirical_scope_observation.py blas
Windows fatal exception: access violation
Thread 0x00000f48 [ThreadPoolExecutor-0_0Windows fatal exception: ]access violation
(most recent call first):
Current thread's C stack trace (most recent call first):
<cannot get C stack on this system>
File "D:\a\threadpoolctl\threadpoolctl\tests\empirical_scope_observation.py", line 58 in blas_math
File "C:\Users\runneradmin\miniconda3\envs\testenv\Lib\concurrent\futures\thread.py", line 73 in run
File "C:\Users\runneradmin\miniconda3\envs\testenv\Lib\concurrent\futures\thread.py", line 86 in run
File "C:\Users\runneradmin\miniconda3\envs\testenv\Lib\concurrent\futures\thread.py", line 119 in _worker
File "C:\Users\runneradmin\miniconda3\envs\testenv\Lib\threading.py", line 1024 in run
Windows fatal exception: access violation
Current thread's C stack trace (most recent call first):
<cannot get C stack on this system>
./continuous_integration/run_tests.sh: line 34: 1491 Segmentation fault PYTHONPATH=. python -X faulthandler tests/empirical_scope_observation.py blas
Error: Process completed with exit code 139.
Unfortunately, faulthandler doesn't show anything useful on Windows hence the cannot get C stack on this system
Environment
The September 24 failure and September 26 successful run both use:
- Python 3.14.7, with the GIL enabled
- NumPy 2.5.3
- conda-forge OpenBLAS 0.3.34, pthreads build
- The same threadpoolctl commit,
f6bef5dca966327941a87f4a92673cd9c0096add
CI logs
It seems to happen quite often
To investigate further
It could be useful to capture a native crash dump with ProcDump and inspect it with WinDbg.
LLM suggestion
For this CI crash, use ProcDump to capture a dump, then inspect it with WinDbg. After downloading Microsoft ProcDump, run this in PowerShell with the conda environment activated and the repository as the working directory:
$env:PYTHONPATH = "."
New-Item -ItemType Directory -Force dumps | Out-Null
.\procdump64.exe -accepteula -ma -e 1 -f C0000005 -n 1 `
-x dumps "$env:CONDA_PREFIX\python.exe" `
-X faulthandler tests/empirical_scope_observation.py blas
This captures a full dump at the first access violation (C0000005), before application exception handlers run. Upload dumps/*.dmp as a CI artifact, using if: always() on the upload step. ProcDump options
Open the dump in WinDbg and run:
.symfix
.reload
!analyze -v
.ecxr
kv
~* kb
.ecxr selects the exception context; kv shows its native stack; ~* kb shows every thread’s stack.
Microsoft debugging guidance
Symbol availability determines the detail: without matching symbols, you may get DLL names and offsets rather than internal function names or source lines. Preserve the exact Python, NumPy and OpenBLAS binaries from the failing environment, plus any matching debug symbols. Even an unsymbolized trace can help establish which native library is crashing.
Partly generated by an agent, but I did double-check with my brain on 😉.
cc @itamarst in case you have some insights or you have seen something similar when putting this empirical tests together.
I may try to investigate further on my Windows VM to see if I can at least reproduce.
The problem
The Windows
pylatest_conda_forge_openblasCI job intermittently crashes while running:Unfortunately, faulthandler doesn't show anything useful on Windows hence the
cannot get C stack on this systemEnvironment
The September 24 failure and September 26 successful run both use:
f6bef5dca966327941a87f4a92673cd9c0096addCI logs
It seems to happen quite often
To investigate further
It could be useful to capture a native crash dump with ProcDump and inspect it with WinDbg.
LLM suggestion
For this CI crash, use ProcDump to capture a dump, then inspect it with WinDbg. After downloading Microsoft ProcDump, run this in PowerShell with the conda environment activated and the repository as the working directory:
This captures a full dump at the first access violation (C0000005), before application exception handlers run. Upload dumps/*.dmp as a CI artifact, using if: always() on the upload step. ProcDump options
Open the dump in WinDbg and run:
.ecxrselects the exception context;kvshows its native stack;~* kbshows every thread’s stack.Microsoft debugging guidance
Symbol availability determines the detail: without matching symbols, you may get DLL names and offsets rather than internal function names or source lines. Preserve the exact Python, NumPy and OpenBLAS binaries from the failing environment, plus any matching debug symbols. Even an unsymbolized trace can help establish which native library is crashing.