Compilation failures cause installation hang

[nedbat]

SUMMARY

On a Mac without libssh, the installation produces error messages, but never completes. More details at astral-sh/uv#16079

ISSUE TYPE

• Bug Report

PYLISSH and LIBSSH VERSION

OS / ENVIRONMENT

STEPS TO REPRODUCE

EXPECTED RESULTS

ACTUAL RESULTS

[webknjaz]

Thanks, I've been meaning to see if there's something in the build backend that needs to be handled better: #762 / #699 (comment).

[webknjaz]

Using a clean F43 container does not reproduce this — building fails just fine:

$ podman run --rm -it -v "${PWD}:/io" --workdir /io fedora:43 bash -c 'dnf install -y gcc git tox && tox run -e build-dists'
[..]
[1/1] Cythonizing /tmp/.tmp-ansible-pylibssh-pep517-pclxdil0/src/src/pylibsshext/_libssh_version.pyx
/tmp/.tmp-ansible-pylibssh-pep517-pclxdil0/src/src/pylibsshext/_libssh_version.c:1236:10: fatal error: libssh/libssh.h: No such file or directory
 1236 | #include "libssh/libssh.h"
      |          ^~~~~~~~~~~~~~~~~
compilation terminated.
error: command '/usr/bin/gcc' failed with exit code 1

ERROR Backend subprocess exited when trying to invoke build_wheel
build-dists: exit 1 (7.63 seconds) /io> .tox/build-dists/bin/python -m build --outdir /io/dist/ /io pid=54
  build-dists: FAIL code 1 (7.64=setup[0.02]+cmd[7.63] seconds)
  evaluation failed :( (7.67 seconds)

Perhaps, it's a macOS-only bug.

[webknjaz]

I think this may be cython/cython#3262 / cython/cython#3263. But I'm yet to verify this.

[webknjaz]

That Cython PR was released but this is still happening. I've found that setting

# pyproject.toml
[tool.local.cythonize.kwargs]
parallel = 1  # or 0

makes it work.

But anything higher than 1 (including the default of the CPU number auto-detection) makes multiprocessing hang within the internal machinery of Cython.
Having stared at it long enough, I realized that we use a different entry point which has the same problem that hasn't been fixed in that patch.

[webknjaz]

Hopefully, cython/cython#7183 will take care of this upstream. I'll need to apply a workaround here, meanwhile.

[webknjaz]

@nedbat would you mind validating #770 on your machine?