The check for short paths under Windows via os.path.samefile, introduced in #11936, also found similar tests in symlinked tests in the GH Actions CI.
Fixes#12039.
Co-authored-by: Bruno Oliveira <bruno@soliv.dev>
Previously, if more than one fixture finalizer raised, only the first
was reported, and the other errors were lost.
Use an exception group to report them all. This is similar to the change
we made in node teardowns (in `SetupState`).
As detailed in
https://github.com/pytest-dev/pytest/issues/11475#issuecomment-1937043670,
currently with `--import-mode=importlib` pytest will try to import every
file by using a unique module name, regardless if that module could be
imported using the normal import mechanism without touching `sys.path`.
This has the consequence that non-test modules available in `sys.path`
(via other mechanism, such as being installed into a virtualenv,
PYTHONPATH, etc) would end up being imported as standalone modules,
instead of imported with their expected module names.
To illustrate:
```
.env/
lib/
site-packages/
anndata/
core.py
```
Given `anndata` is installed into the virtual environment, `python -c
"import anndata.core"` works, but pytest with `importlib` mode would
import that module as a standalone module named
`".env.lib.site-packages.anndata.core"`, because importlib module was
designed to import test files which are not reachable from `sys.path`,
but now it is clear that normal modules should be imported using the
standard mechanisms if possible.
Now `imporlib` mode will first try to import the module normally,
without changing `sys.path`, and if that fails it falls back to
importing the module as a standalone module.
This also makes `importlib` respect namespace packages.
This supersedes #11931.
Fix#11475Close#11931
(diff better viewed ignoring whitespace)
In pytest<8, the collection tree for `pyargs` arguments in an invocation
like this:
pytest --collect-only --pyargs pyflakes.test.test_undefined_names
looked like this:
```
<Package test>
<Module test_undefined_names.py>
<UnitTestCase Test>
<TestCaseFunction test_annotationUndefined>
... snipped ...
```
The pytest 8 collection improvements changed it to this:
```
<Dir pytest>
<Dir .tox>
<Dir venv>
<Dir lib>
<Dir python3.11>
<Dir site-packages>
<Package pyflakes>
<Package test>
<Module test_undefined_names.py>
<UnitTestCase Test>
<TestCaseFunction test_annotationUndefined>
... snipped ...
```
Besides being egregious (and potentially even worse than the above,
going all the way to the root, for system-installed packages, as is
apparently common in CI), this also caused permission errors when trying
to probe some of those intermediate directories.
This change makes `--pyargs` arguments no longer try to add parent
directories to the collection tree according to the `--confcutdir` like
they're regular arguments. Instead, only add the parents that are in the
import path. This now looks like this:
```
<Package .tox/venv/lib/python3.11/site-packages/pyflakes>
<Package test>
<Module test_undefined_names.py>
<UnitTestCase Test>
<TestCaseFunction test_annotationUndefined>
... snipped ...
```
Fix#11904.
pytest allows a fixture to request its own name (directly or
indirectly), in which case the fixture with the same name but one level
up is used.
To know which fixture should be used next, pytest keeps a mutable
item-global dict `_arg2index` which maintains this state. This is not
great:
- Mutable state like this is hard to understand and reason about.
- It is conceptually buggy; the indexing is global (e.g. if requesting
`fix1` and `fix2`, the indexing is shared between them), but actually
different branches of the subrequest tree should not affect each
other.
This is not an issue in practice because pytest keeps a cache of the
fixturedefs it resolved anyway (`_fixture_defs`), but if the cache is
removed it becomes evident.
Instead of the `_arg2index` state, count how many `argname`s deep we are
in the subrequest tree ("the fixture stack") and use that for the index.
This way, no global mutable state and the logic is very localized and
easier to understand.
This is slower, however fixture stacks should not be so deep that this
matters much, I hope.
According to my understanding, this code, which handles obtaining the
relevant fixturedefs when a dynamic `getfixturevalue` is used, has an
optimization where it only grabs fixturedefs that are visible to the
*parent* of the item, instead of the item itself, under the assumption
that a fixturedef can't be visible to a single item, only to a
collector.
Remove this optimization for the following reasons:
- It doesn't save much (one loop iteration in `matchfactories`)
- It slightly complicates the complex fixtures code
- If some plugin wants to make a fixture visible only to a single item,
why not let it?
- In the static case (`getfixtureclosure`), this optimization is not
done (despite the confusing name `parentnode`, it is *not* the parent
node). This is inconsistent.
Allow for the output of test case execution to be controlled independently from the application verbosity level.
`verbosity_test_case` is the new ini setting to adjust this functionality.
Fix#11639
Passing a short path in the command line was causing the matchparts check to fail, because ``Path(short_path) != Path(long_path)``.
Using ``os.path.samefile`` as fallback ensures the comparsion works on Windows when comparing short/long paths.
Fix#11895
In `CallSpec2.setmulti` the `params` and `_arg2scope` dicts are always
set together.
Further, the `get_parametrized_fixture_keys` accesses `_arg2scope` for
all argnames without a check, which I think rules out that the code
protects against plugin shenanigans.
After removing the suppress, adjust the comment and code to make more
sense.
Fix#12021.
Reopens#11706.
This reverts commit 12b9bd5801.
This change caused a bad regression in pytest-xdist:
https://github.com/pytest-dev/pytest-xdist/issues/1024
pytest-xdist necessarily has special handling of `--maxfail` and session
fixture teardown get executed multiple times with the change.
Since I'm not sure how to adapt pytest-xdist myself, revert for now.
I kept the sticky `shouldstop`/`shouldfail` changes as they are good
ideas regardless I think.
Dicts these days preserve order, so the sort is no longer needed to
achieve determinism.
As shown by the `test_dynamic_parametrized_ordering` test, this can
change the ordering of items, but only in equivalent ways (same number
of setups/teardowns per scope), it will just respect the user's given
ordering better (hence `vxlan` items now ordered before `vlan` items
compared to the previous ordering).
Follow up to #12006, let's put some comments clarifying `is_in_confcutdir` semantics, as this is not the first time someone misunderstands it.
Also removed an obsolete comment in `_loadconftestmodules`: we already set the `confcutdir` based on `rootdir`/`initfile` if not explicitly given.
Co-authored-by: Ran Benita <ran@unusedvar.com>
As the comment above it states, we need to keep the comparison simple so mypy can understand it, otherwise it will fail to properly handle this on Windows and will flag ``os.getuid()`` missing.
Test:
`warnings.warn()` expects that its first argument is a `str` or a
`Warning`, but since 9454fc38d3
`pytest.warns()` no longer allows `Warning` instances unless the first
argument the `Warning` was initialized with is a `str`. Furthermore, if
the `Warning` was created without arguments then `pytest.warns()` raises
an unexpected `IndexError`. The new tests reveal the problem.
Fix:
`pytest.warns()` now allows using `warnings.warn()` with a `Warning`
instance, as is required by Python, with one exception. If the warning
used is a `UserWarning` that was created by passing it arguments and the
first argument was not a `str` then `pytest.raises()` still considers
that an error. This is because if an invalid type was used in
`warnings.warn()` then Python creates a `UserWarning` anyways and it
becomes impossible for `pytest` to figure out if that was done
automatically or not.
[ran: rebased on previous commit]
Today `pyproject.toml` is the standard for declaring a Python project root, so seems reasonable to consider it for the ini configuration (and specially `rootdir`) in case we do not find other suitable candidates.
Related to #11311
Previously this would trigger an `AssertionError`.
While it could be considered a bug-fix, but given it now can be relied upon, it is probably better to consider it an improvement.
Fix#11311
Should be around 40% faster according to this simplified small benchmark:
python -m timeit "a=[0, 1, 2, 3, 4];b=list((e if i in {0, len(a) -1} else str(e)) for i, e in enumerate(a))"
200000 loops, best of 5: 1.12 usec per loop
python -m timeit "a=[0, 1, 2, 3, 4];b=list((a[0], *(str(e) for e in a[1:-1]), a[-1]))"
500000 loops, best of 5: 651 nsec per loop
python -m timeit "a=[0, 1, 2, 3, 4,5,6,7,8,9,10,11,12,13,14,15,16];b=list((e if i in {0, len(a) -1} else str(e)) for i, e in enumerate(a))"
100000 loops, best of 5: 3.31 usec per loop
python -m timeit "a=[0, 1, 2, 3, 4,5,6,7,8,9,10,11,12,13,14,15,16];b=list((a[0], *(str(e) for e in a[1:-1]), a[-1]))"
200000 loops, best of 5: 1.72 usec per loop
* Improve error message when using @pytest.fixture twice
While obvious in hindsight, this error message confused me. I thought my fixture
function was used in a test function twice, since the wording is ambiguous.
Also, the error does not tell me *which* function is the culprit.
Finally, this adds a test, which wasn't done in
cfd16d0dac where this was originally implemented.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Fix#11929.
Figured out what's going on. We have the following collection tree:
```
<Dir pyspacewar>
<Dir src>
<Package pyspacewar>
<Package tests>
<DoctestModule test_main.py>
<DoctestItem pyspacewar.tests.test_main.doctest_main>
```
And the `test_main.py` contains an autouse fixture (`fake_game_ui`) that
`doctest_main` needs in order to run properly. The fixture doesn't run!
It doesn't run because nothing collects the fixtures from (calls
`parsefactories()` on) the `test_main.py` `DoctestModule`.
How come it only started happening with commit
ab63ebb3dc07b89670b96ae97044f48406c44fa0? Turns out it mostly only
worked accidentally. Each `DoctestModule` is also collected as a normal
`Module`, with the `Module` collected after the `DoctestModule`. For
example, if we add a non-doctest test to `test_main.py`, the collection
tree looks like this:
```
<Dir pyspacewar>
<Dir src>
<Package pyspacewar>
<Package tests>
<DoctestModule test_main.py>
<DoctestItem pyspacewar.tests.test_main.doctest_main>
<Module test_main.py>
<Function test_it>
```
Now, `Module` *does* collect fixtures. When autouse fixtures are
collected, they are added to the `_nodeid_autousenames` dict.
Before ab63ebb3dc, `DoctestItem` consults
`_nodeid_autousenames` at *setup* time. At this point, the `Module` has
collected and so it ended up picking the autouse fixture (this relies on
another "accident", that the `DoctestModule` and `Module` have the same
node ID).
After ab63ebb3dc, `DoctestItem` consults
`_nodeid_autousenames` at *collection* time (= when it's created). At
this point, the `Module` hasn't collected yet, so the autouse fixture is
not picked out.
The fix is simple -- have `DoctestModule.collect()` call
`parsefactories`. From some testing I've done it shouldn't have negative
consequences (I hope).
Currently, `DoctestModule` does `import_path` on its own. This changes
it to use `importtestmodule` as `Module` does. The behavioral changes
are:
- Much better error messages on import errors.
- Handles a few more error cases (see `importtestmodule`). This
technically expands the cover of `--doctest-ignore-import-errors` but
I think it makes sense.
- Considers `pytest_plugins` in the module.
- Populates `self.obj` as properly (without double-imports) as is
expected from a `PyCollector`.
This is also needed for the next commit.
ruff is faster and handle everything we had prior.
isort configuration done based on the indication from
https://github.com/astral-sh/ruff/issues/4670, previousely based on
reorder-python-import (#11896)
flake8-docstrings was a wrapper around pydocstyle (now archived) that
explicitly asks to use ruff in https://github.com/PyCQA/pydocstyle/pull/658.
flake8-typing-import is useful mainly for project that support python 3.7
and the one useful check will be implemented in https://github.com/astral-sh/ruff/issues/2302
We need to keep blacken-doc because ruff does not handle detection
of python code inside .md and .rst. The direct link to the repo is
now used to avoid a redirection.
Manual fixes:
- Lines that became too long
- % formatting that was not done automatically
- type: ignore that were moved around
- noqa of hard to fix issues (UP031 generally)
- fmt: off and fmt: on that is not really identical
between black and ruff
- autofix re-order in pre-commit from faster to slower
Co-authored-by: Ran Benita <ran@unusedvar.com>
- Previously, `getpass.getuser()` would leak an ImportError if the
USERNAME environment variable was not set on Windows because the `pwd`
module cannot be imported.
- Starting in Python 3.13.0a3, it only raises `OSError`.
Fixes#11874
Change our mypy configuration to disallow untyped defs by default, which ensures *new* files added to the code base are fully typed.
To avoid having to type-annotate everything now, add `# mypy: allow-untyped-defs` to files which are not fully type annotated yet.
As we fully type annotate those modules, we can then just remove that directive from the top.
The function currently uses `find_suffixes` which iterates the entire
directory searching for files with the given suffix. In some cases
though, like in pytest's selftest, the directory can get big:
$ ls /tmp/pytest-of-ran/pytest-0/
7686
and iterating it many times can get slow.
This doesn't fix the underlying issue (iterating the directory) but at
least speeds it up a bit by using `os.scandir` instead of
`path.iterdir`. So `make_numbered_dir` is still slow for pytest's
selftests, but reduces ~10s for me.
We have a use case for this in the next commit.
The name can be obtained by using `manager.get_name(plugin)`, however
this is currently O(num plugins) in pluggy, which would be good to
avoid. Besides, it seems generally useful.
This removes one thing that directory collectors need to worry about.
This adds one hook dispatch per `__pycache__` file, but I think it's
worth it for consistency.
(Diff better viewed ignoring whitespace)
Since e1c66ab0ad, conftest loading is
handled at the directory level before sub-nodes are collected, so there
is no need for the doctest plugin to handle it specially.
This was probably the case even before
e1c66ab0ad, but I haven't verified this.
This is a useful addition to the existing `listchain`. While `listchain`
returns top-to-bottom, `iterparents` is bottom-to-top and doesn't require
an internal full iteration + `reverse`.
Since we're removing nose support, let's also drop support for this
attribute.
From doing a code search on github, this seems completely unused outside
of nose, except for some projects which used to use it, but no longer
do.
Refs #11662.
--- Problem
Each fixture definition has a "visibility", the `FixtureDef.baseid`
attribute. This is nodeid-like string. When a certain `node` requests a
certain fixture name, we match node's nodeid against the fixture
definitions with this name.
The matching currently happens on the *textual* representation of the
nodeid - we split `node.nodeid` to its "parent nodeids" and then check
if the fixture's `baseid` is in there.
While this has worked so far, we really should try to avoid textual
manipulation of nodeids as much as possible. It has also caused problem
in an odd case of a `Package` in the root directory: the `Package` gets
nodeid `.`, while a `Module` in it gets nodeid `test_module.py`. And
textually, `.` is not a parent of `test_module.py`.
--- Solution
Avoid this entirely by just checking the node hierarchy itself. This is
made possible by the fact that we now have proper `Directory` nodes
(`Dir` or `Package`) for the entire hierarchy.
Also do the same for `_getautousenames` which is a similar deal.
The `iterparentnodeids` function is no longer used and is removed.
Instead of modifying user objects like modules and classes that we
really shouldn't be touching, use the new `_register_fixture` internal
API to do it directly.
Add a function on the `FixtureManager` to register a fixture with
pytest. Currently this can only be done through `parsefactories`.
My aim is to eventually make something like this available to plugins,
as it's a pretty common need.
* Fix handling empty values of NO_COLOR and FORCE_COLOR
Fix handling NO_COLOR and FORCE_COLOR environment variables to correctly
be ignored when they are set to an empty value, as defined
in the specification:
> Command-line software which adds ANSI color to its output by default
> should check for a NO_COLOR environment variable that, when present
> *and not an empty string* (regardless of its value), prevents
> the addition of ANSI color.
(emphasis mine, https://no-color.org/)
The same is true of FORCE_COLOR, https://force-color.org/.
* Streamline testing for FORCE_COLOR and NO_COLOR
Streamline the tests for FORCE_COLOR and NO_COLOR variables, and cover
all possible cases (unset, set to empty, set to "1"). Combine the two
assert functions into one taking boolean parameters. Mock file.isatty
in all circumstances to ensure that the environment variables take
precedence over the fallback value resulting from isatty check (or that
the fallback is actually used, in the case of both FORCE_COLOR
and NO_COLOR being unset).
* Put a 'reset' color in front of the highlighting
When doing the highlighting, some lexers will not set the initial color
explicitly, which may lead to the red from the errors being propagated
to the start of the expression
* Add syntactic highlighting to the error explanations
This updates the various error reporting to highlight python code when
displayed, to increase readability and make it easier to understand
--- Current main
In current main (before pervious commit), calls to gethookproxy/ihook
are the trigger for loading non-initial conftests. This basically means
that conftest loading is done almost as a random side-effect,
uncontrolled and very non-obvious. And it also dashes any hope of making
gethookproxy faster (gethookproxy shows up prominently in pytest
profiles).
I've wanted to improve this for a while, #11268 was the latest step
towards that.
--- PR before change
In this PR, I ran into a problem.
Previously, Session and Package did all of the directory traversals
inside of their collect, which loaded the conftests as a side effect. If
the conftest loading failed, it will occur inside of the collect() and
cause it to be reported as a failure.
Now I've changed things such that Session.collect and Package.collect no
longer recurse, but just collect their immediate descendants, and
genitems does the recursive expansion work.
The problem though is that genitems() doesn't run inside of specific
collector's collect context. So when it loads a conftest, and the
conftest loading fails, the exception isn't handled by any CollectReport
and causes an internal error instead.
The way I've fixed this problem is by loading the conftests eagerly in a
pytest_collect_directory post-wrapper, but only during genitems to make
sure the directory is actually selected.
This solution in turn caused the conftests to be collected too early;
specifically, the plugins are loaded during the parent's collect(), one
after the other as the directory entries are collected. So when the
ihook is hoisted out of the loop, new plugins are loaded inside the
loop, and due to the way the hook proxy works, they are added to the
ihook even though they're supposed to be scoped to the child collectors.
So no hoisting.
--- PR after change
Now I've come up with a better solution: since now the collection tree
actually reflects the filesystem tree, what we really want is to load
the conftest of a directory right before we run its collect(). A
conftest can affect a directory's collect() (e.g. with a
pytest_ignore_collect hookimpl), but it cannot affect how the directory
node itself is collected. So I just moved the conftest loading to be
done right before calling collect, but still inside the CollectReport
context.
This allows the hoisting, and also removes conftest loading from
gethookproxy since it's no longer necessary. And it will probably enable
further cleanups. So I'm happy with it.
Previously the error report would have all sections glued together:
- The assertion representation
- The error explanation
- The full diff
This makes it hard to see at a glance where which one starts and ends.
One of the representation (dataclasses, tuples, attrs) does display a
newlines at the start already.
Let's add a newlines before the error explanation and before the full
diff, so we get an easier to read report.
This has one disadvantage: we get one line less in the least verbose
mode, where the output gets truncated.
When running `tox -e py-lsof` I get a deluge of this warning:
```
src/pytest/.tox/py-lsof-numpy-pexpect/lib/python3.11/site-packages/_pytest/pytester.py:130: EncodingWarning: UTF-8 Mode affects locale.getpreferredencoding(). Consider locale.getencoding() instead.
```
Use `locale.getencoding` instead.
Bruno Oliveira
mrbean-bremen
Ran Benita
Bruno Oliveira
Patrick Lannigan
robotherapist
Eric Larson
Ben Leith
Reagan Lee
Eero Vaher
Pierre Sassoulas
Florian Bruhin
whysage
Clément Robert
Russell Martin
Dương Quốc Khánh
clee2000
faph
woutdenolf
Loïc Estève
Franck Charras
Fabian Sturm
Ben Brown
Marc Bresson
Michał Górny
Benjamin Schubert
Arthur Richard