Repeatable crashes and misbehavior in GitHub Windows CI

[dabrahams]

Description

This step reliably caused crashes and other weird behaviors in my GitHub CI, on Windows only. The exact behavior could always be repeated by re-triggering the workflow run without deleting existing caches, but the misbehavior for any particular cache configuration varies. Most often there's an unceremonious crash during building with no diagnostic, but I've also seen my final step, which just runs, e.g., hylo/.build/debug/hc --version, fail, with Powershell reporting that hc is not a program, cmd or cmdlet. This is after everything built and passed all tests… so the executable must be there, right?

Reproduction

fork https://github.com/hylo-lang/hylo at 7a152a454, enable actions, and push some trivial change to trigger CI. When the run succeeds, then make another change and push that.

Expected behavior

success

Environment

GitHub runner, windows-latest, Swift 5.9

Additional information

No response

[dabrahams]

Update: this has nothing to do with caches. I'm still seeing the behavior after disabling the caches for Windows.

[finagolfin]

If this is happening without github caching too, is it reproducible off github, ie on a local Windows machine? If not, maybe this is just something related to the constrained resources of the free github Windows CI or its documented performance issues, actions/runner-images#7320.

[dabrahams]

I don't know. My latest bet is that the problem is in SPM on Windows, which is even less supported than SPM on Mac and Linux. I'm in the process of transitioning to CMake and if that makes the problem go away, we'll have confirmation.

[finagolfin]

#72221 suggests this isn't just a github issue for Windows, at least with the performance aspect.