This fixes running `pytest tests/test_foo.py::test_bar`, where `tests`
is a symlink to `project/app/tests`: previously
`project/app/conftest.py` would be ignored for fixtures then.
The problem was that _matchnodes would receive two items: [DoctestModule, Module]. It would then collect the first one, *cache it*, and fail to match against the name in the command line. Next, it would reuse the cached item (DoctestModule) instead of collecting the Module which would eventually find the "test" name on it.
Added the type of the node to the cache key to avoid this problem, although I'm not a big fan of caches that have different key types.
Fix#3843
Now each in-process pytest run saves a snapshot of important global Python
state and restores it after the test completes, including the list of loaded
modules & the Python path settings.
Previously only the loaded package data was getting restored, but that was
also reverting any loaded package changes done in the test triggering the
pytest runs, and not only those done by the pytest runs themselves.
Updated acceptance tests broken by this change, which were only passing before
by accident as they were making multiple pytest runs with later ones depending
on sys.path changes left behind by the initial one.
os.symlink might fail on Windows because users require a special policy
to create symlinks (argh).
This is not a problem on AppVeyor because it is logged in as an admin,
but would be surprising for Windows users running pytest's test
suite on their computer.
The leak was caused by the (unused) `FixtureRequest._fixture_values`
cache.
This was introduced because the `partial` object (created to call
FixtureDef.finish() bound with the Request) is kept alive
through the entire session when a function-scoped fixture depends
on a session-scoped (or higher) fixture because of the nested
`addfinalizer` calls.
FixtureDef.finish() started receiving a request object in order to
obtain the proper hook proxy object (#2127), but this does not seem
useful at all in practice because `pytest_fixture_post_finalizer`
will be called with the `request` object of the moment the fixture value
was *created*, not the request object active when the fixture value
is being destroyed. We should probably deprecate/remove the request
parameter from `pytest_fixture_post_finalizer`.
Fix#2981
Anthony Sottile
Bruno Oliveira
Thomas Hisch
Daniel Hahler
Niklas JQ
Jeffrey Rackauckas
CrazyMerlyn
Ronny Pfannschmidt
Ronny Pfannschmidt
Jurko Gospodnetić
Henk-Jaap Wagenaar
hugovk
Christian Boelsen
Andras Tim