The tested py.path.local.make_numbered_dir function is *not*
multiprocess safe, because is uses os.listdir which itself is not.
The os.listdir documentation explicitly states that:
> If a file is removed from or added to the directory during the call
> of this function, whether a name for that file be included is unspecified.
This can lead to a race when:
1. process A attempts to create directory N
2. the creation fails, as another process already created it in the meantime
3. process A calls listdir to determine a more recent maxnum
4. processes B+ repeatedly create newer directories and they delete directory N
5. process A doesn't have directory N or any newer directory in listdir result
6. process A attempts to create directory N again and raises
For details, see https://github.com/pytest-dev/pytest/issues/11603#issuecomment-1805708144
and bellow.
Additionally, the test itself has a race in batch_make_numbered_dirs.
When this functions attempts to write to repro-N/foo,
repro-N may have already been removed by another process.
For details, see https://github.com/pytest-dev/pytest/issues/11603#issuecomment-1804714313
and bellow.
---
The tested py.path.local.make_numbered_dir function is not used in pytest.
There is a different implementation in _pytest.pathlib.
Closes#11603
Logging has many global states, and we did foresee this by creating a ``cleanup_disabled_logging`` fixture,
however one might still forget to use it and failures leak later -- sometimes not even in the same PR, because the order
of the tests might change in the future, specially when running under xdist.
This problem surfaced during pytest-dev/pytest#11530, where tests unrelated to the change started to fail.
The current method as the following problem, described by Sadra
Barikbin:
The tests that request both `pytester` and `monkeypatch` and use
`monkeypatch.chdir` without context, relying on `monkeypatch`'s teardown
to restore cwd. This doesn't work because the following sequence of
actions take place:
- `monkeypatch` is set up.
- `pytester` is set up. It saves the original cwd and changes it to a
new one dedicated to the test function.
- Test function calls `monkeypatch.chdir()` without context.
`monkeypatch` saves cwd, which is not the original one, before
changing it.
- `pytester` is torn down. It restores the cwd to the original one.
- `monkeypatch` is torn down. It restores cwd to what it has saved.
The solution here is to have pytester use `monkeypatch.chdir()` itself,
then everything is handled correctly.
Tests if a captured exception group contains an expected exception.
Will raise `AssertionError` if the wrapped exception is not an exception group.
Supports recursive search into nested exception groups.
Starting with `resolve_package_path` and its associated tests,
this refactoring seeks to make path concatenation more
readable and consistent within tests/functions.
As discussed in #11413:
- code is free to use either `/` and `joinpath`
- consistency within a function is more important than consistency across the codebase
- it is nice to use `/` when it is more readable
- it is nice to use `joinpath` when there is little context
- be mindful that `joinpath` may be clearer when joining multiple segments
For packages, `import_path` receives the path to the package's `__init__.py` file, however module names (as they live in `sys.modules`) should not include the `__init__` part.
For example, `app/core/__init__.py` should be imported as `app.core`, not as `app.core.__init__`.
Fix#11306
Move handling of user_properties to `finalize()`.
Previously if a fixture failed during teardown, `pytest_runtest_logreport` would not be called with "teardown", resulting in the user properties not being saved on the JUnit XML file.
Fixes: #11367
This reverts commit e938580257.
Revert "improve error msg and test"
This reverts commit c0cf822ca1.
Revert "error msg"
This reverts commit ec1053cc16.
Revert "changelog"
This reverts commit d2dc8a70b5.
Revert "simplify code / take out user-gen typeerror case"
This reverts commit b9cb87d862.
To remove fixtures.py::add_funcargs_pseudo_fixture_def and add its logic
i.e. registering funcargs as params and making corresponding fixturedefs,
right to Metafunc.parametrize in which parametrization takes place.
To remove funcargs from metafunc attributes as we populate metafunc
params and make pseudo fixturedefs simultaneously and there's no need to
keep funcargs separately.
Previously, when assigning a scope for a fully-indirect parameter set,
when there are multiple fixturedefs for a param (i.e. same-name fixture
chain), the highest scope was used, but it should be the lowest scope,
since that's the effective scope of the fixture.
* [pre-commit.ci] pre-commit autoupdate
updates:
- [github.com/PyCQA/flake8: 6.0.0 → 6.1.0](https://github.com/PyCQA/flake8/compare/6.0.0...6.1.0)
- [github.com/asottile/pyupgrade: v3.9.0 → v3.10.1](https://github.com/asottile/pyupgrade/compare/v3.9.0...v3.10.1)
* Use is instead of type comparison with equal to appease the linter
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Pierre Sassoulas <pierre.sassoulas@gmail.com>
Previously, the `_getconftestmodules` function was used both to load
conftest modules for a path (during `pytest_load_initial_conftests`),
and to retrieve conftest modules for a path (during hook dispatch and
for fetching `collect_ignore`). This made things muddy - it is usually
nicer to have clear separation between "command" and "query" functions,
when they occur in separate phases.
So split into "load" and "get".
Currently, `gethookproxy` still loads conftest itself. I hope to change
this in the future.
This test makes clear the need for the `_check_scope()` call in the
`pytest_setup_fixture` impl in fixtures.py, which otherwise seems
redundant with the one in `_compute_fixture_value`.
Dict comparsion in the ApproxMapping class did not check if values were None before attempting to subtract for max_abs_diff stat, which was throwing an TypeError instead of being handled by pytest error assertion. Check for None has been added before these calculations, so that None will properly show as Obtained/Expected in pytest assert message
The initial implementation (in #7246) introduced the `importlib` mode, which
never added the imported module to `sys.modules`, so it included a test
to ensure calling `import_path` twice would yield different modules.
Not adding modules to `sys.modules` proved problematic, so we began to add the imported module to `sys.modules`
in #7870, but failed to realize that given we are now changing `sys.modules`, we might
as well avoid importing it more than once.
Then #10088 came along, passing `importlib` also when importing application modules
(as opposed to only test modules before), which caused problems due to imports
having side-effects and the expectation being that they are imported only once.
With this PR, `import_path` returns the module immediately if already in
`sys.modules`.
Fix#10811Fix#10341
Fixes#11104.
See the issue for a description of the problem.
Now, we use the same logic for initial conftest paths as we do for
deciding the initial args, which was the idea behind checking
`namespace.file_or_dir` and `testpaths` previously.
This fixes the issue of `testpaths` being considered for initial
conftests even when it's not used for the args.
(Another issue in faeb16146b was that the
`testpaths` were not glob-expanded, this is also fixed.)
Currently, if `--confcutdir` is not set, `inipath.parent` is used, and
if `initpath` is not set, then `confcutdir` is None, which means there
is no cutoff.
Having no cutoff is not great, it means we potentially start probing
stuff all the way up to the filesystem root directory. So let's add
another fallback, to `rootpath`, which is always something reasonable.
In #10758 we introduced the support for the use of the walrus operator in the test cases. There was a case which was not handled that caused a bug report #11028. This PR aims to fix the issue and also to improve how the walrus operator is handled in the AssertionRewriter class.
Closes#11028
Added handling of %f directive to print microseconds in log format options, such as log-date-format. It is impossible to do with a standard logging.Formatter because it uses time.strftime which doesn't know about milliseconds and %f. In this PR I added a custom Formatter which converts LogRecord to a datetime.datetime object and formats it with %f flag. This behaviour is enabled only if a microsecond flag is specified in a format string.
Also added a few tests to check the standard and changed behavior.
Closes#10991
`--lf` has a feature where if a certain `Module` (python file) does not
contain any failed tests, it is skipped entirely at the collector level
instead of being collected and each item skipped individually. When this
happens the collection summary looks like this:
run-last-failure: rerun previous 1 failure (skipped 1 file)
However, this feature didn't work for `Module`s inside of `Package`s,
only for those directly beneath the `Session`.
Fix#11054.
Forces requested `caplog` logging levels to be enabled if they were disabled via `logging.disable()`
`[attr-defined]` mypy error ignored in `logging.py` because there were existing errors with the imports
and `loggin.Logger.manager` is an attr set at runtime. Since it's in the standard lib I can't really fix that.
Ignored an attr-defined error in `src/_pytest/config/__init__.py` because the re-export is necessary.
Fixes#8711
`_set_initial_conftests` could break on some systems if a very long
option was passed, because the `Path.exists()` call raises an
`OSError` instead of returning `False`.
Fix#10169
Since `Traceback.getcrashentry` takes the `ExceptionInfo`, it is not
really independent of it and is in the wrong layer. Prevent nonsensical
mistakes by inlining it.
TracebackEntry needs the excinfo for the `__tracebackhide__ = callback`
functionality, where `callback` accepts the excinfo.
Currently it achieves this by storing a weakref to the excinfo which
created it. I think this is not great, mixing layers and bloating the
objects.
Instead, have `ishidden` (and transitively, `Traceback.filter()`) take
the excinfo as a parameter.
Closes#10831.
This fixes a small bug where running tests that contained
`pytest.fail(pytrace=False)` with the `--tb=line` flag set results in
an output of "None" in the Failures section of the output, and adds
a test to ensure the behavior is correct.
By "empty traceback" I mean a traceback all of whose entries have been
filtered/cut/pruned out.
Currently, if an empty traceback needs to be repr'ed, the last entry
before the filtering is used instead (added in
accd962c9f).
Showing a hidden frame is not so good IMO. This commit does the
following instead:
1. Shows details of the exception.
2. Shows a message about how the full trace can be seen.
Example:
```
_____________ test _____________
E ZeroDivisionError: division by zero
All traceback entries are hidden. Pass `--full-trace` to see hidden and internal frames.
```
Also handles `--tb=native`, though there the `--full-trace` bit is not
shown.
This commit contains some pieces from
431ec6d34e (which has been reverted).
Helps towards fixing issue # 1904.
Co-authored-by: Felix Hofstätter <Felhof1@hotmail.com>
The old-style `sys.exc_info()` triplet is redundant nowadays with
`(type(exc), exc, exc.__traceback__)`, and is beginning to get
soft-deprecated in Python 3.12.
Add a nicer API to ExceptionInfo which takes just the exc instead of the
triplet. There are already a few internal uses which benefit.
pytest could crash given pathological AST position attributes, which shouldn't happen when testing real Python code, but could happen when testing AST produced by e.g. Hylang.
Another example of the failure is in the nightly CI for the JAX project: https://github.com/google/jax/actions/runs/4607513902/jobs/8142126075
Co-authored-by: Bruno Oliveira <nicoddemus@gmail.com>
Co-authored-by: Jake VanderPlas <jakevdp@google.com>
Benjamin Schubert
Patrick Lannigan
Miro Hrončok
dependabot[bot]
Sharad Nair
Ran Benita
Bruno Oliveira
Benjamin Schubert
ryanpudd
Tanya Agarwal
Reagan Lee
Chris Mahoney
Mihail Milushev
Mihail Milushev
Bruno Oliveira
Aleksandr Brodin
Warren Markham
Marc Mueller
Sadra Barikbin
Ronny Pfannschmidt
Israel Fruchter
Sean Malloy
Jon Parise
Zac Hatfield-Dodds
DetachHead
Tyler Smart
pre-commit-ci[bot]
Isaac Virshup
Kenny Y
Hugo van Kemenade
Lesnek
akhilramkee
Cheukting
Tushar Sadhwani
Thomas Grainger
nondescryptid
Alessio Izzo
theirix
Alex Lambson
Ville Skyttä
Adam J. Stewart
Bruno Oliveira
Brian Larsen
Alex
Kodi Arfer