pytest recipes

Skipping tests based on markers

A few packages use custom pytest markers to indicate e.g. tests requiring Internet access. These markers can be used to conveniently disable whole test groups, e.g.:

python_test() {
    epytest -m 'not network' dask

Skipping tests based on paths/names

There are two primary methods of skipping tests based on path (and name) in pytest: using --ignore and --deselect.

--ignore causes pytest to entirely ignore a file or a directory when collecting tests. This works only for skipping whole files but it ignores missing dependencies and other failures occurring while importing the test file.

--deselect causes pytest to skip the specific test or tests. It can be also used to select individual tests or even parametrized variants of tests.

Both options can be combined to get tests to pass without having to alter the test files. They are preferable over suggestions from skipping problematic tests when tests are installed as part of the package. They can also be easily applied conditionally to Python interpreter.

The modern versions of eclasses provide two control variables, EPYTEST_IGNORE and EPYTEST_DESELECT that can be used to list test files or tests to be ignored or deselected respectively. These variables can be used in global scope to avoid redefining python_test(). However, combining them with additional conditions requires using the local scope.

python_test() {
    local EPYTEST_IGNORE=(
        # ignore whole file with missing dep
        # deselect a single test
        # deselect a parametrized test based on first param
    [[ ${EPYTHON} == python3.6 ]] && EPYTEST_DESELECT+=(
        # deselect a test for py3.6 only

Avoiding the dependency on pytest-runner

pytest-runner is a package providing pytest command to setuptools. While it might be convenient upstream, there is no real reason to use it in Gentoo packages. It has no real advantage over calling pytest directly.

Some packages declare the dependency on pytest-runner in setup_requires. As a result, the dependency is enforced whenever is being run, even if the user has no intention of running tests. If this is the case, the dependency must be stripped.

The recommended method of stripping it is to use sed:

python_prepare_all() {
    sed -i -e '/pytest-runner/d' || die

Disabling plugin autoloading

Normally, when running a test suite pytest loads all plugins installed on the system. This is often convenient for upstreams, as it makes it possible to use the features provided by the plugins (such as async test function support, or fixtures) without the necessity to explicitly enable them. However, there are also cases when additional plugins could make the test suite fail or become very slow (especially if pytest is called recursively).

The modern recommendation for these cases is to disable plugin autoloading via setting the PYTEST_DISABLE_PLUGIN_AUTOLOAD environment variable, and then explicitly enable specific plugins if necessary.


Previously we used to recommend explicitly disabling problematic plugins via -p no:<plugin>. However, it is rarely obvious which plugin is causing the problems, and it is entirely possible that another plugin will cause issues in the future, so an opt-in approach is usually faster and more reliable.

The easier approach to enabling plugins is to use the -p option, listing specific plugins. The option can be passed multiple times, and accepts a plugin name as specified in the package’s entry_points.txt file:

python_test() {
    epytest -p asyncio -p tornado

However, this approach does not work when the test suite calls pytest recursively (e.g. you are testing a pytest plugin). In this case, the PYTEST_PLUGINS environment variable can be used instead. It takes a comma-separated list of plugin module names:

python_test() {
    local -x PYTEST_PLUGINS=xdist.plugin,xdist.looponfail,pytest_forked


Please note that failing to enable all the required plugins may cause some of the tests to be skipped implicitly (especially if the test suite is using async functions and no async plugin is loaded). Please look at skip messages and warnings to make sure everything works as intended.

Using pytest-xdist to run tests in parallel

pytest-xdist is a plugin that makes it possible to run multiple tests in parallel. This is especially useful for programs with large test suites that take significant time to run single-threaded.

Using pytest-xdist is recommended if the package in question supports it (i.e. it does not cause semi-random test failures) and its test suite takes significant time. This is done via setting EPYTEST_XDIST to a non-empty value prior to calling distutils_enable_tests. It ensures that an appropriate depedency is added, and that epytest adds necessary command-line options.

distutils_enable_tests pytest

Please note that some upstream use pytest-xdist even if there is no real gain from doing so. If the package’s tests take a short time to finish, please avoid the dependency and strip it if necessary.

Not all test suites support pytest-xdist. Particularly, it requires that the tests are written not to collide one with another. Sometimes, xdist may also cause instability of individual tests. In some cases, it is possible to work around this using the same solution as when dealing with flaky tests.

When only a few tests are broken or unstable because of pytest-xdist, it is possible to use it and deselect the problematic tests. It is up to the maintainer’s discretion to decide whether this is justified.

Dealing with flaky tests

A flaky test is a test that sometimes passes, and sometimes fails with a false positive result. Often tests are flaky because of too steep timing requirements or race conditions. While generally it is preferable to fix the underlying issue (e.g. by increasing timeouts), it is not always easy.

Sometimes upstreams use such packages as dev-python/flaky or dev-python/pytest-rerunfailures to mark tests as flaky and have them rerun a few minutes automatically. If upstream does not do that, it is also possible to force a similar behavior locally in the ebuild:

python_test() {
    # plugins make tests slower, and more fragile
    # some tests are very fragile to timing
    epytest -p rerunfailures --reruns=10 --reruns-delay=2

Note that the snippet above also disables plugin autoloading to speed tests up and therefore reduce their flakiness. Sometimes forcing explicit rerun also makes it possible to use xdist on packages that otherwise randomly fail with it.

Using pytest-timeout to prevent deadlocks (hangs)

pytest-timeout plugin adds an option to terminate the test if its runtime exceeds the specified limit. Some packages decorate specific tests with timeouts; however, it is also possible to set a baseline timeout for all tests.

A timeout causes the test run to fail, and therefore using it is not generally necessary for test suites that are working correctly. If individual tests are known to suffer from unfixable hangs, it is preferable to deselect them. However, setting a general timeout is recommended when a package is particularly fragile, or has suffered deadlocks in the past. A proactive setting can prevent it from hanging and blocking arch testing machines.

The plugin can be enabled via setting EPYTEST_TIMEOUT to the timeout in seconds, prior to calling distutils_enable_tests. This ensures that an appropriate depedency is added, and that epytest adds necessary command-line options.

distutils_enable_tests pytest

The timeout applies to every test separately, i.e. the above example will cause a single test to time out after 30 minutes. If multiple tests hang, the total run time will multiply consequently.

When deciding on a timeout value, please take into the consideration that the tests may be run on a low performance hardware, and on a busy system, and choose an appropriately high value.

It is a good idea to use the default assignment form, as in the snippet above, as that permits the user to easily override the timeout if necessary.


EPYTEST_TIMEOUT can also be set by user in make.conf or in the calling environment. This can be used as a general protection against hanging test suites. However, please note that this does not control dependencies, and therefore the user may need to install dev-python/pytest-timeout explicitly.

Avoiding dependencies on other pytest plugins

There is a number of pytest plugins that have little value to Gentoo users. They include plugins for test coverage (dev-python/pytest-cov), coding style (dev-python/pytest-flake8) and more. Generally, packages should avoid using those plugins.


As of 2022-01-24, epytest disables a few undesirable plugins by default. As a result, developers have a good chance of experiencing failures due to hardcoded pytest options first, even if they have the relevant plugins installed.

If your package really needs to use the specific plugin, you need to pass -p <plugin> explicitly to reenable it.

In some cases, upstream packages only list them as dependencies but do not use them automatically. In other cases, you will need to strip options enabling them from pytest.ini or setup.cfg.

src_prepare() {
    sed -i -e 's:--cov=wheel::' setup.cfg || die

TypeError: _make_test_flaky() got an unexpected keyword argument ‘reruns’

If you see a test error resembling the following:

TypeError: _make_test_flaky() got an unexpected keyword argument 'reruns'

This means that the tests are being run via flaky plugin while the package in question expects pytest-rerunfailures. This is because both plugins utilize the same @pytest.mark.flaky marker but support different set of arguments.

To resolve the problem, explicitly disable the flaky plugin and make sure to depend on dev-python/pytest-rerunfailures:

    test? (

python_test() {
    epytest -p no:flaky


An ImportPathMismatchError generally indicates that the same Python module (or one that supposedly looks the same) has been loaded twice using different paths, e.g.:

E   _pytest.pathlib.ImportPathMismatchError: ('path', '/usr/lib/pypy3.7/site-packages/path', PosixPath('/tmp/portage/dev-python/jaraco-path-3.3.1/work/jaraco.path-3.3.1/jaraco/'))

These problems are usually caused by pytest test discovery getting confused by namespace packages. In this case, the jaraco directory is a Python 3-style namespace but pytest is treating it as a potential test directory. Therefore, instead of loading it as jaraco.path relatively to the top directory, it loads it as path relatively to the jaraco directory.

The simplest way to resolve this problem is to restrict the test discovery to the actual test directories, e.g.:

python_test() {
    epytest test


python_test() {
    epytest --ignore jaraco

Failures due to missing files in temporary directories

As of 2024-01-05, epytest overrides the default temporary directory retention policy of pytest. By default, directories from successful tests are removed immediately, and the temporary directories from the previous test run are replaced by the subsequent test run. This frequently reduces disk space requirements from test suites, but it can rarely cause tests to fail.

If you notice test failures combined with indications that a file was not found, and especially regarding the pytest temporary directories, try if overriding the retention policy helps, e.g.:

python_test() {
    epytest -o tmp_path_retention_policy=all

fixture ‘…’ not found

Most of the time, a missing fixture indicates that some pytest plugin is not installed. In rare cases, it can signify an incompatible pytest version or package issue.

The following table maps common fixture names to their respective plugins.

Fixture name













pytest captures all warnings from the test suite by default, and prints a summary of them at the end of the test suite run:

=============================== warnings summary ===============================
asgiref/ 1 warning
tests/ 5 warnings
tests/ 12 warnings
tests/ 1 warning
  /tmp/asgiref/asgiref/ DeprecationWarning: There is no current event loop
    self.main_event_loop = asyncio.get_event_loop()

However, some projects go further and use filterwarnings option to make (some) warnings fatal:

==================================== ERRORS ====================================
_____________________ ERROR collecting tests/ ______________________
tests/ in <module>
    class ASGITest(TestCase):
tests/ in ASGITest
    async def test_wrapped_case_is_collected(self):
asgiref/ in __init__
    self.main_event_loop = asyncio.get_event_loop()
E   DeprecationWarning: There is no current event loop
=========================== short test summary info ============================
ERROR tests/ - DeprecationWarning: There is no current event loop
!!!!!!!!!!!!!!!!!!!! Interrupted: 1 error during collection !!!!!!!!!!!!!!!!!!!!
=============================== 1 error in 0.13s ===============================

Unfortunately, this frequently means that warnings coming from a dependency trigger test failures in other packages. Since making warnings fatal is relatively common in the Python world, it is recommended to:

  1. Fix warnings in Python packages whenever possible, even if they are not fatal to the package itself.

  2. Do not enable new Python implementations if they trigger any new warnings in the package.

If the warnings come from issues in the package’s test suite rather than the installed code, it is acceptable to make them non-fatal. This can be done either through removing the filterwarnings key from setup.cfg, or adding an ignore entry. For example, the following setting ignores DeprecationWarning in test directory:

filterwarnings =