pytest-run-parallel

A simple pytest plugin to run tests concurrently

This pytest plugin takes a set of tests that would be normally be run serially and execute them in parallel.

The main goal of pytest-run-parallel is to discover thread-safety issues that could exist when using C libraries, this is of vital importance after PEP703, which provides a path for a CPython implementation without depending on the Global Interpreter Lock (GIL), thus allowing for proper parallelism in programs that make use of the CPython interpreter.

For more information about C thread-safety issues, please visit the free-threaded community guide at https://py-free-threading.github.io/

How it works

This plugin is not an alternative to pytest-xdist and does not run all of the tests in a test suite simultaneously in a thread pool. Instead, it runs many instances of the same test in a thread pool. It is only useful as a tool to do multithreaded stress tests using an existing test suite and is not useful to speed up the execution of a test suite via multithreaded parallelism.

Given an existing test taking arguments *args and keyword arguments **kwargs, this plugin creates a new test that is equivalent to the following Python code:

import threading
from concurrent.futures import ThreadPoolExecutor

def run_test(b, *args, **kwargs):
    for _ in range(num_iterations):
        b.wait()
        execute_pytest_test(*args, **kwargs)


with ThreadPoolExecutor(max_workers=num_parallel_threads) as tpe:
    b = threading.Barrier(num_parallel_threads)
    for _ in range(num_parallel_threads):
        tpe.submit(run_test, b, *args, **kwargs)

The execute_pytest_test function hides some magic to ensure errors and failures get propagated correctly to the main testing thread. Using this plugin avoids the boilerplate of rewriting existing tests to run in parallel in a thread pool. Note that args and kwargs might include pytest marks and fixtures, and the way this plugin is currently written, those fixtures are shared between threads.

Features

• Global CLI flags:

  • --parallel-threads to run a test suite in parallel
  • --iterations to run multiple times in each thread
  • --skip-thread-unsafe to skip running tests marked as or detected to be thread-unsafe.
  • --mark-warnings-as-unsafe and --mark-ctypes-as-unsafe to always skip running tests that use the warnings or ctypes modules, respectively. These are useful if you are adding support for Python 3.14 to a library that already runs tests under pytest-run-parallel on Python 3.13 or older.
  • --mark-hypothesis-as-unsafe to always skip running tests that use hypothesis. While newer version of Hypothesis are thread-safe, and versions which are not are automatically skipped by pytest-run-parallel, this flag is an escape hatch in case you run into thread-safety problems caused by Hypothesis, or in tests that happen to use hypothesis and were skipped in older versions of pytest-run-parallel.
  • --forever to keep running tests in an endless loop starting from the top when completing a test run, until a test crashes or the user explicitly ends the process with Ctrl-C. This is especially helpful when trying to reproduce thread safety bugs that might only occur rarely. Note that pytest's progress indicator will keep showing 100% forever after the first pass of the test suite. forever is not compatible with -n from pytest-xdist.
  • --ignore-gil-enabled to ignore the RuntimeWarning generated when the GIL is enabled at runtime on the free-threaded build and run the tests despite the fact that the GIL is enabled. This option has no effect if pytest is configured to treat warnings as errors.

• Five corresponding markers:

  • pytest.mark.parallel_threads_limit(n) to mark a single test to run in a maximum of n threads, even if the --parallel-threads command-line argument is set to a higher value. This is useful if a test uses resources that should be limited.
  • pytest.mark.force_parallel_threads(n) to force a test to always run in n threads, overriding both the --parallel-threads command-line argument and any parallel_threads_limit marker. Use this when a test must always be multi-threaded with a specific thread count.
  • pytest.mark.parallel_threads(n) (deprecated when n > 1 — use force_parallel_threads instead) to mark a test to always run in n threads. Note that this implies that the test will be multi-threaded just because pytest-run-parallel is installed, even if --parallel-threads is not passed at the command-line.
  • pytest.mark.thread_unsafe to mark a single test to run in a single thread. It is equivalent to using pytest.mark.parallel_threads(1)
  • pytest.mark.iterations(n) to mark a single test to run n times in each thread

• Four corresponding fixtures:

  • num_parallel_threads: The number of threads the test will run in
  • num_iterations: The number of iterations the test will run in each thread
  • thread_index: An index for the test's current thread.
  • iteration_index: An index for the test's current iteration.

• Modifications to existing fixtures:

  • tmp_path and tmpdir: Patched to be thread-safe, with individual subdirectories being created for each thread. Subdirectories are not created for each iteration, so you may see issues with reused temporary directions when using --iterations.

While pytest-run-parallel has special handling for the tmp_path and tmpdir fixtures to ensure that each thread has a private temporary directory, the plugin only does this if a test requests these fixtures directly. If another fixture requests tmp_path or tmpdir, then all threads will share a temporary directory in that fixture.

When using the fixtures thread_index and iteration_index, they should be requested directly by tests, and will return 0 when requested by other fixtures.

Note: It's possible to specify --parallel-threads=auto, pytest.mark.force_parallel_threads("auto"), or pytest.mark.parallel_threads_limit("auto") which will let pytest-run-parallel choose the number of logical CPU cores available to the testing process. If that cannot be determined, the number of physical CPU cores will be used. If that fails as well, it will fall back to running all tests single-threaded.

Requirements

pytest-run-parallel depends exclusively on pytest. Optionally installing psutil will help with identifying the number of logical cores available to the testing process in systems where that's not possible with the Python stdlib.

Installation

You can install "pytest-run-parallel" via pip from PyPI:

$ pip install pytest-run-parallel

If you want to additionally install psutil you can run:

$ pip install pytest-run-parallel[psutil]

Caveats

Pytest itself is not thread-safe and it is not safe to share stateful pytest fixtures or marks between threads. Existing tests relying on setting up mutable state via a fixture will see the state shared between threads. Tests that dynamically set marks or share marks will also likely not be thread-safe. See the pytest documentation for more detail and the community-maintained free threaded Python porting guide for more detail about using pytest in a multithreaded context on the free-threaded build of Python.

We suggest marking tests that are incompatible with this plugin's current design with @pytest.mark.thread_unsafe or @pytest.mark.thread_unsafe(reason="...").

The following functions and modules are known to be thread-unsafe and pytest-run-parallel will automatically skip running tests using them in parallel:

• The pytest capsys, capteesys, capsysbinary, capfd and capfdbinary fixtures
• The pytest monkeypath fixture

The following fixtures are known to be thread-unsafe on Python 3.13 and older, or on 3.14 and newer if Python isn't configured correctly:

• pytest.warns
• pytest.deprecated_call
• The pytest recwarn fixture
• warnings.catch_warnings
• unittest.mock
• ctypes

If an older version of hypothesis that is known to be thread-unsafe is installed, tests using hypothesis are skipped.

Additionally, if a set of fixtures is known to be thread unsafe, tests that use them can be automatically marked as thread unsafe by declaring them under the thread_unsafe_fixtures option under pytest INI configuration file:

[pytest]
thread_unsafe_fixtures =
    fixture_1
    fixture_2
    ...

Or under the section tool.pytest.ini_options if using pyproject.toml:

[tool.pytest.ini_options]
thread_unsafe_fixtures = [
    'fixture_1',
    'fixture_2',
    ...
]

Similarly, if a function is known to be thread unsafe and should cause a test to be marked as thread-unsafe as well, its fully-qualified name can be registered through the thread_unsafe_functions option in the INI file (or under tool.pytest.ini_options when using pyproject.toml):

[pytest]
thread_unsafe_functions =
    module.submodule.func1
    module.submodule2.func2
    ...

You can also blocklist entire modules by using an asterisk:

[pytest]
thread_unsafe_functions =
    module1.*
    module2.submodule.*
    ...

Also, if you define a __thread_safe__ = False attribute on a function that is called by a test and is up to two levels below in the call stack, then pytest-run-parallel will automatically detect that a thread-unsafe function is being used and will mark the test as thread-unsafe.

Usage

This plugin has two modes of operation, one via the --parallel-threads and --iterations pytest CLI flags, which allows a whole test suite to be run in parallel:

$ pytest --parallel-threads=10 --iterations=10 tests

By default, the value for both flags will be 1, thus not modifying the usual behaviour of pytest except when the flag is set.

Note that using pytest-xdist and setting iterations to a number greater than one might cause tests to run even more times than intended.

The other mode of operation occurs at the individual test level, via the pytest.mark.force_parallel_threads and pytest.mark.iterations markers:

# test_file.py
import pytest

@pytest.fixture
def my_fixture():
    ...

@pytest.mark.force_parallel_threads(2)
@pytest.mark.iterations(10)
def test_something_1():
    # This test will be run in parallel using two concurrent threads
    # and 10 times in each thread
    ...

@pytest.mark.parametrize('arg', [1, 2, 3])
@pytest.mark.force_parallel_threads(3)
def test_fixture(my_fixture, arg):
    # pytest markers and fixtures are supported as well
    ...

Both modes of operations are supported simultaneously, i.e.,

# test_something_1 and test_fixture will be run using their set number of
# threads; other tests will be run using 5 threads.
$ pytest -x -v --parallel-threads=5 test_file.py

You can skip tests marked as or detected to be thread-unsafe by passing --skip-thread-unsafe in your pytest invocation. This is useful when running pytest-run-parallel under Thread Sanitizer. Setting --skip-thread-unsafe=True will avoid unnecessarily running tests where thread sanitizer cannot detect races because the test is not parallelized.

Older versions of pytest-run-parallel always marked tests using the warnings and ctypes modules as thread-unsafe, since both were not thread-safe until Python 3.14. If you are adding support for Python 3.14 and would like to continue marking tests that use warnings or ctypes, pass --mark-warnings-as-unsafe or --mark-ctypes-as-unsafe, respectively, in your pytest invocation.

Additionally, pytest-run-parallel exposes the num_parallel_threads and num_iterations fixtures which enable a test to be aware of the number of threads that are being spawned and the number of iterations each test will run:

# test_file.py
import pytest

def test_skip_if_parallel(num_parallel_threads):
    if num_parallel_threads > 1:
        pytest.skip(reason='does not work in parallel')
    ...

The thread_index and iteration_index fixtures are also available, which enable tests to display different behavior between threads and iterations.

# test_file.py
import numpy as np

def test_unique_rng_streams(thread_index):
    # create an RNG stream with a seed that is deterministic
    # but still unique to this thread
    rng = np.random.default_rng(thread_index)
    ...

Finally, the thread_comp fixture allows for parallel test debugging, by providing an instance of ThreadComparator, whose __call__ method allows to check if all the values produced by all threads during an specific execution step are the same:

# test_file.py
def test_same_execution_values(thread_comp):
    a = 2
    b = [3, 4, 5]
    c = None
    # Check that the values for a, b, c are the same across tests
    thread_comp(a=a, b=b, c=c)

Tracing

If you run pytest with verbose output (e.g. by passing -v in your pytest invocation), you will see that tests are annotated to either "PASS" or "PARALLEL PASS". A "PASS" indicates the test was run on a single thread, whereas "PARALLEL PASS" indicates the test passed and was run in a thread pool. If a test was not run in a thread pool because pytest-run-parallel detected use of thread-unsafe functionality, the reason will be printed as well.

If you are running pytest in the default configuration without -v, then tests that pass in a thread pool will be annotated with a slightly different dot character, allowing you to visually pick out when tests are not run in parallel.

For example in the output for this file:

tests/test_kx.py ·....·

Only the first and last tests are run in parallel.

In order to list the tests that were marked as thread-unsafe and were not executed in parallel, you can set the PYTEST_RUN_PARALLEL_VERBOSE environment variable to 1.

Contributing

Contributions are very welcome. Tests can be run with uv run pytest. If you want to run pytest with coverage enabled you can run uv run coverage run and then uv run coverage report -m to get the coverage report. Please ensure the coverage at least stays the same before you submit a pull request.

License

Distributed under the terms of the MIT license, "pytest-run-parallel" is free and open source software

Issues

If you encounter any problems, please file an issue along with a detailed description.