Experiment: tests: linkcheck: do not require 'freshenv'

[jayaddison]

Purpose

Thanks to the output of pytest's slowest-duration-tests output, as produced when the Sphinx test suite ran on Windows in GitHub Actions CI, I noticed that some linkcheck tests on Windows seem to be unexpectedly long-duration (reported as more than 4 seconds each).

I'm curious whether this is due to the use of the freshenv=True project setup test parameter -- so I'm experimenting with unsetting the fresenv parameter for those tests here.

The linkcheck tests in particular do seem to use it a lot in our testsuite:

$ grep -rwc freshenv tests|grep -vw 0
tests/test_transforms/test_transforms_post_transforms.py:3
tests/test_ext_napoleon/test_ext_napoleon_docstring.py:1
tests/test_writers/test_docutilsconf.py:2
tests/test_markup/test_smartquotes.py:12
tests/test_extensions/test_ext_viewcode.py:4
tests/test_extensions/test_ext_todo.py:3
tests/test_extensions/test_ext_duration.py:8
tests/test_intl/test_intl.py:3
tests/test_domains/test_domain_py_pyobject.py:1
tests/test_domains/test_domain_js.py:1
tests/test_domains/test_domain_py.py:6
tests/test_domains/test_domain_py_canonical.py:1
tests/test_environment/test_environment_indexentries.py:8
tests/test_environment/test_environment.py:1
tests/test_command_line.py:3
tests/test_ext_autodoc/test_ext_autodoc_typehints.py:1
tests/test_search.py:3
tests/test_builders/test_build_warnings.py:4
tests/test_builders/test_build_latex.py:1
tests/test_builders/test_build_linkcheck.py:33
tests/test_builders/test_incremental_reading.py:2

References

• Discovered from thread at linkcheck: Allow case-insensitive URL comparisons #14046 (comment)

[jayaddison]

Removing the freshenv=True argument does not appear to have made a difference: https://github.com/sphinx-doc/sphinx/actions/runs/19750249013/job/56591850933?pr=14112#step:9:2435

[AA-Turner]

@jayaddison could we just decouple the tests from Sphinx? As I see it there are two parts: that the links are properly detected from source files (requires Sphinx) and that links are properly marked as working/otherwise.

See the recent changes to the autodoc tests where most tests no longer require the app fixture, I wonder if there's something similar we could do here?

A

[jayaddison]

@AA-Turner that sounds worthwhile, although I haven't completely gotten my head around how to separate those two aspects of the tests yet. I think b9d5d2b looks like one of the refactors you mention, though? (test-performance-related, going by the commit message, and removes the app test fixture)