Segmentation fault on `import juliacall` in a CI job

[mtsokol]

Affects: JuliaCall

Describe the bug

Hi! I recently started using JuliaCall to call Julia from Python. In my CI I have several jobs: ubuntu, macos and windows. One of my steps installs Julia and the other one builds python package and runs pytest.

A segmentation fault happens on import juliacall, inside juliacall.__init__.py::init() function, here, but only on ubuntu jobs. Macos and windows complete successfully. Julia 1.10.2 gets installed.

Do you know what might be the reason?

---

Failing ubuntu output: https://github.com/pydata/sparse/actions/runs/8252820808/job/22573248981?pr=647#step:7:116
JuliaCall setup: https://github.com/willow-ahrens/finch-tensor/blob/main/src/finch/julia.py
JuliaCall version: https://github.com/willow-ahrens/finch-tensor/blob/1bf21a28d28a19ba1cea59c6f5a719cb8914e395/pyproject.toml#L11
CI definition that installs Julia: https://github.com/pydata/sparse/blob/4bfea8fa5b66393a1ff7c2db45218ba41f46baec/.github/workflows/ci.yml#L44

[hameerabbasi]

Just to note, I can reproduce this in Docker outside of GitHub Actions, using the content in pydata/sparse#649 and this Dockerfile:

FROM --platform=x86_64 condaforge/miniforge3:latest

ADD ci/environment.yml .
RUN conda env create -f environment.yml

RUN mkdir -p /root/workdir
WORKDIR /root/workdir
ADD . .
RUN conda run --live-stream -n sparse-dev pip install -e .[tests]
RUN conda run --live-stream -n sparse-dev pytest --pyargs sparse/tests/test_backends.py

Also, switching to Poetry resolves the issue somehow.

[hameerabbasi]

Valgrind output

77.76 ERROR conda.cli.main_run:execute(124): `conda run pytest --pyargs sparse/tests/test_backends.py` failed. (See above for error)
78.52 ==7== 
78.52 ==7== HEAP SUMMARY:
78.52 ==7==     in use at exit: 529,620 bytes in 1,132 blocks
78.52 ==7==   total heap usage: 34,393 allocs, 33,261 frees, 38,464,700 bytes allocated
78.52 ==7== 
78.56 ==7== 528 bytes in 1 blocks are possibly lost in loss record 564 of 761
78.56 ==7==    at 0x483B7F3: malloc (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
78.56 ==7==    by 0x2244CD: _PyObject_GC_NewVar (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D0CB: _PyEval_MakeFrameVector (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC3E: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24D447: object_vacall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x25BBCA: _PyObject_CallMethodIdObjArgs (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x17D3C4: PyImport_ImportModuleLevelObject.cold (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x241668: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x2E281F: _PyEval_Vector (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x2E2766: PyEval_EvalCode (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x2E9969: builtin_exec (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DE32: cfunction_vectorcall_FASTCALL (in /opt/conda/bin/python3.10)
78.56 ==7== 
78.56 ==7== 544 bytes in 1 blocks are possibly lost in loss record 565 of 761
78.56 ==7==    at 0x483B7F3: malloc (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
78.56 ==7==    by 0x2244CD: _PyObject_GC_NewVar (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D0CB: _PyEval_MakeFrameVector (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC3E: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x241F29: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7== 
78.56 ==7== 568 bytes in 1 blocks are possibly lost in loss record 570 of 761
78.56 ==7==    at 0x483B7F3: malloc (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
78.56 ==7==    by 0x2244CD: _PyObject_GC_NewVar (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D0CB: _PyEval_MakeFrameVector (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC3E: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23DA3B: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23DA3B: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7== 
78.56 ==7== 664 bytes in 1 blocks are possibly lost in loss record 625 of 761
78.56 ==7==    at 0x483DFAF: realloc (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
78.56 ==7==    by 0x2EB828: _PyObject_GC_Resize (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D1C9: _PyEval_MakeFrameVector (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC3E: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7== 
78.56 ==7== 1,664 bytes in 2 blocks are possibly lost in loss record 719 of 761
78.56 ==7==    at 0x483B7F3: malloc (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
78.56 ==7==    by 0x2244CD: _PyObject_GC_NewVar (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D0CB: _PyEval_MakeFrameVector (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC3E: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x23D62F: _PyEval_EvalFrameDefault (in /opt/conda/bin/python3.10)
78.56 ==7==    by 0x24DC6B: _PyFunction_Vectorcall (in /opt/conda/bin/python3.10)
78.56 ==7== 
78.56 ==7== LEAK SUMMARY:
78.56 ==7==    definitely lost: 0 bytes in 0 blocks
78.56 ==7==    indirectly lost: 0 bytes in 0 blocks
78.56 ==7==      possibly lost: 3,968 bytes in 6 blocks
78.56 ==7==    still reachable: 525,652 bytes in 1,126 blocks
78.56 ==7==         suppressed: 0 bytes in 0 blocks
78.56 ==7== Reachable blocks (those to which a pointer was found) are not shown.
78.56 ==7== To see them, rerun with: --leak-check=full --show-leak-kinds=all
78.56 ==7== 
78.56 ==7== Use --track-origins=yes to see where uninitialised values come from
78.56 ==7== For lists of detected and suppressed errors, rerun with: -s
78.56 ==7== ERROR SUMMARY: 15 errors from 6 contexts (suppressed: 16 from 2)

[hameerabbasi]

Okay, I've managed to reduce it down to this (Purely conda, juliapkg and juliacall):

environment.yml

name: sparse-dev
channels:
  - conda-forge
  - nodefaults
dependencies:
  - python
  - numpy
  - julia
  - pyjuliacall
  - pyjuliapkg

Dockerfile

FROM --platform=x86_64 condaforge/miniforge3:latest

ADD environment.yml .
RUN conda env create -f environment.yml
RUN mkdir -p /root/workdir
WORKDIR /root/workdir
ADD test.py /root/workdir/test.py
RUN conda run --live-stream -n sparse-dev python test.py

test.py

import juliacall

[cjdoris]

I think this is a duplicate of #464 (please comment if not and I'll reopen).