arrow: [C++] arrow-compute-* tests segfaulting on verify-rc-source-windows job

Describe the bug, including details regarding any error messages, version, and platform.

Fairly recently, some of the arrow-compute-* tests started segfaulting on the verify-rc-source-windows job. See an example run https://github.com/ursacomputing/crossbow/actions/runs/16131946653/job/45520878681

And a sample of a single test failure

[ RUN      ] TestPivotKernel.AllInputKinds
unknown file: error: SEH exception with code 0xc0000005 thrown in the test body.
Stack trace:


[  FAILED  ] TestPivotKernel.AllInputKinds (0 ms)

And a summary:

The following tests FAILED:
	 24 - arrow-compute-scalar-if-else-test (Failed)        arrow-compute-tests unittest
	 25 - arrow-compute-scalar-temporal-test (Failed)       arrow-compute-tests unittest
	 26 - arrow-compute-scalar-math-test (Failed)           arrow-compute-tests unittest
	 28 - arrow-compute-vector-test (Failed)                arrow-compute-tests unittest
	 29 - arrow-compute-vector-sort-test (Failed)           arrow-compute-tests unittest
	 30 - arrow-compute-vector-selection-test (Failed)      arrow-compute-tests unittest
	 31 - arrow-compute-vector-swizzle-test (Failed)        arrow-compute-tests unittest
	 32 - arrow-compute-aggregate-test (Failed)             arrow-compute-tests unittest

Related, I guess the missing stack traces in the above output are a known thing and might require some work: https://github.com/apache/arrow/issues/27455.

Component(s)

C++

About this issue

  • Original URL
  • State: closed
  • Created a year ago
  • Comments: 17 (17 by maintainers)

Commits related to this issue

Most upvoted comments

so not sure how much help I can offer besides saying “I think you are on to something”

Your support is enough 😃

At the moment I am trying to find differences between AMD64 Windows 2022 AVX2 and the verify-rc-source-windows job which use the same Windows GitHub runner. I am triggering tests on that PR to try and find out the culprit. I’ll update if I find anything.

I was able to reproduce the crash locally and could see it was due to not finding the gtest and other DLL.s. Reverting https://github.com/apache/arrow/pull/46839 fixed the issue and the C++ tests now pass for me like @raulcd found on https://github.com/apache/arrow/pull/47035.

So we need what @kou is working on in https://github.com/apache/arrow/pull/47049. I’m looking at the PyArrow test failures now.

@raulcd the main difference I see with those jobs is that the rc job run cmake from a .bat file and doesn’t use the bash script AND it also uses conda for dependencies (which is probably the more relevant difference)

Yes, I agree is not the meson change, I think is the change on the Gtest main that was included on that PR mainly because reverting provides a successful build. It seems to fail for some combination of Windows builds.