DumpError: strings must be utf-8 decodable

[jaraco]

In pypa/distutils#183, I've stumbled onto an issue with xdist. In that issue, distutils/setuptools are moving from a stdout-based logging system to the Python logging framework. As a result, the quiet() context, which suppresses writes to sys.stdout no longer has the effect of suppressing logs.

One of the things that setuptools is logging is a filename containing surrogate escapes. It logs the name so the user can identify which filename was failing.

With pytest-xdist, however, the test fails with an INTERNALERROR in gateway_base.

A minimal test is to run pytest -n auto on the following:

def test_log_non_utf8():
    __import__('logging').getLogger().warn(
        'föö'.encode('latin-1').decode('utf-8', errors='surrogateescape')
    )
    __import__('pytest').fail('failing to emit messages')

That test seems legitimate and capturing behavior that provides value to the user. The issue doesn't occur without pytest-xdist.

My instinct is that pytest-xdist shouldn't be putting constraints on the allowed outputs for logging. Can something be done to be more lenient about legitimate non-encodeable values being passed? If the encoding is only an internal implementation detail between worker and supervisor, it should serialize any strings in a way that they're deserialized with fidelity to the original.

[RonnyPfannschmidt]

xdist uses stdout/stdin as communication channels both locally and via ssh, currently there is no way to work around that

[RonnyPfannschmidt]

i think we need to move this issue to execnet, as the real issue is about its serializer