pytest-isolate

Isolate each test in a subprocess - pytest forked replacement, based on pytest-forked.

This pytest plugin was generated with Cookiecutter along with @hackebrot's cookiecutter-pytest-plugin template.

Features

• Run each test in an forked, isolated subprocess
• Captures stdout & stderr of crashing processes
• Add Timeout to a forked test
• Limit memory used by test
• Limit CPU time used by test
• Manage GPU resources with CUDA_VISIBLE_DEVICES, with support for fractional requests (1/2, 1/4, 1/8, 1/16)
• Plays nice with pytest-xdist
• Shows warnings, even with xdist!
• Create visual timeline of test execution (isolated or not)

Requirements

• pytest

Optional Dependencies

For GPU resource management:

• pynvml (optional, for automatic GPU detection)

Support

• Operating Systems: Linux (tested), macOS (not tested but should work). Windows is not supported, and will probably not work as we are based on process forking.
• Python Versions: 3.9, 3.10, 3.11, 3.12

Installation

You can install "pytest-isolate" via pip from PyPI

pip install pytest-isolate

For GPU resource management support:

pip install pytest-isolate[gpu]

Usage

pytest --isolate

To run every test in its own forked subprocess.

Or:

pytest --isolate-timeout 10 --isolate-mem-limit 1000000 --isolate-cpu-limit 10

To set a timeout to every test in addition to forking, and limit to 10 cpu seconds.

Or:

pytest --timeline

With possible combination of the above, to generate a timeline of test execution. The timeline can be viewed in chrome://tracing.

To disable the pulgin, you can use the --no-isolate option:

pytest --no-isolate

Note:

Since this plugin uses fork, it will not work on operating systems without fork support (e.g. Windows).

The flags --timeout or --forked will also be respected such that pytest-isolate is a drop-in replacement forked pytest forked and pytest timeout.

If pytest-forked or pytest-timeout are installed, then they will take precedence. Uninstall them to use pytest-isolate.

Unlike pytest-timeout, timeout in pytest-isolate is implemented by forking the test to a separate subprocess, and setting timeout for that subprocess.

Using the isolate marker

You can use a mark to isolate or time limit the memory and/or cpu usage test:

@pytest.mark.isolate(timeout=10, mem_limit=10**6, cpu_limit=10)
def test_something():
    pass

The isolate marker can also be used to request gpus for a test on a gpu machine:

# Request 2 GPUs for this test
@pytest.mark.isolate(resources={'gpu': 2})
def test_with_gpus():
    # The test will have CUDA_VISIBLE_DEVICES set to the allocated GPU IDs
    assert len(os.environ.get("CUDA_VISIBLE_DEVICES").split(",")) == 2

# Request 1/2 of a GPU for this test, up to 2 tests that request 1/2 of a GPU will be allocated the same GPU
@pytest.mark.isolate(resources={'gpu': 1/2})
def test_with_half_gpu():
    # The test will have CUDA_VISIBLE_DEVICES set to the allocated GPU IDs
    assert len(os.environ.get("CUDA_VISIBLE_DEVICES").split(",")) == 1

Configuration Options

The options can be set in an pytest configuration file, e.g:

[tool.pytest.ini_options]
isolate_timeout=10
isolate_mem_limit=1000000
isolate_cpu_limit=10

CUDA_VISIBLE_DEVICES Handling

If CUDA_VISIBLE_DEVICES is already set when pytest starts, the plugin will respect this setting and only allocate from the GPUs specified there. This works even without pynvml installed.

For example:

• If CUDA_VISIBLE_DEVICES=0,1,2 is set, tests will only use GPUs 0, 1, and 2.
• If CUDA_VISIBLE_DEVICES= is set (empty), no GPUs will be used.

PyTorch and fork

torch.cuda.is_available() opens a CUDA context, and a CUDA context does not survive fork(). Since every isolated test is forked, one availability check before that point breaks the child, so on import this plugin sets:

PYTORCH_NVML_BASED_CUDA_CHECK=1

PyTorch then answers from NVML, opening no context, and falls back to the usual check if NVML is unavailable.

It is set at import rather than from a hook or an option, because pytest loads plugins before importing conftest.py and no configuration is parsed that early. To override, set PYTORCH_NVML_BASED_CUDA_CHECK yourself, or PYTEST_ISOLATE_NO_NVML_CUDA_CHECK=1 to leave the variable untouched.

Contributing

Contributions are very welcome. Tests can be run with tox, please ensure the coverage at least stays the same before you submit a pull request.

License

Distributed under the terms of the MIT, pytest-isolate is free and open source software

Issues

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