Workaround: install lxml from binary wheel when unit tests run on MacOS

[jayaddison]

A pull request to attempt to narrow down on the version of Python 3.10.x where MacOS unit test failure noted at #628 (comment) bug could have been introduced.

Reading and comparing the 3.10.x and 3.11.x changelogs (since the failure affects both lineages) hasn't produced any definitive answers yet.

Problem identified: when lxml is installed from source on MacOS, it will link against the system-provided libxml2 at build-time. This appears to be the cause of the unit test failures we noticed, and this does not occur when using binary wheels from PyPy.

As a workaround, ensure that lxml is installed from binary wheel during MacOS unit test runs.

[jayaddison]

My guess is that this isn't a runtime issue; I think something's wrong with the lxml dependency installation.

Mon, 10 Oct 2022 21:40:51 GMT   Running setup.py install for lxml: still running...
Mon, 10 Oct 2022 21:41:51 GMT   Running setup.py install for lxml: still running...
Mon, 10 Oct 2022 21:41:54 GMT   Running setup.py install for lxml: finished with status 'done'

...and yet extruct is extracting nothing...

{'microdata': [], 'json-ld': []}

[jayaddison]

It seems that there is a behaviour change between (one of the versions of?) libxml2 shipped with some recent releases of MacOS and the behaviour of the library (or bindings?) when compiled during pip install lxml.

The binary wheel on PyPi -- which was statically compiled once and then uploaded for distribution -- appears to work. So we can use pip install --only-binary=lxml lxml to install that version for affected unit test runs.

Why doesn't this affect other versions of Python during unit testing? I'm unsure.

[jayaddison]

Why doesn't this affect other versions of Python during unit testing? I'm unsure.

Also now noted in the pull request description, but for context here too: the reason that this only appeared to affect Python 3.10 and Python 3.11-rc.1 on MacOS was that binary wheels of lxml were not available for those version combinations.

So in fact, there wasn't necessarily a behaviour change introduced by either MacOS or Python. lxml built-from-source on Python 3.8 exhibits the same parsing problem.