Description
I have a test suite that started life in nosetests, and a lot of the tests still use xunit style with a setup_method that creates fixture objects and stores them in the instance of the test class. I ran into a problem with tests exhausting resources, and it seems like pytest is hanging on to references to the instances of the test classes, preventing these fixture objects from being garbage collected.
Here's a proof-of-concept:
conftest.py:
import gc
def pytest_runtest_logfinish(nodeid, location):
import test_leak
gc.collect()
print(f"\nExpensive objects: {test_leak.ExpensiveObject.count}", end="")test_leak.py:
import gc
from typing import ClassVar
import pytest
class ExpensiveObject:
count: ClassVar[int] = 0
def __init__(self) -> None:
ExpensiveObject.count += 1
def __del__(self) -> None:
ExpensiveObject.count -= 1
class TestClass:
def setup_method(self) -> None:
self.obj = ExpensiveObject()
@pytest.mark.parametrize("x", range(5))
def test(self, x: int) -> None:
passOutput:
test_leak.py::TestClass::test[0] PASSED [ 20%]
Expensive objects: 1
test_leak.py::TestClass::test[1] PASSED [ 40%]
Expensive objects: 2
test_leak.py::TestClass::test[2] PASSED [ 60%]
Expensive objects: 3
test_leak.py::TestClass::test[3] PASSED [ 80%]
Expensive objects: 4
test_leak.py::TestClass::test[4] PASSED [100%]
Expensive objects: 5
So despite the calls to gc.collect(), the number of ExpensiveObjects is increasing. In contrast, if I use a pytest.fixture, they are all cleaned up immediately.
Environment
Package Version
-------------- -------
exceptiongroup 1.1.3
iniconfig 2.0.0
packaging 23.1
pip 23.2.1
pluggy 1.3.0
pytest 7.4.0
setuptools 68.1.2
tomli 2.0.1
wheel 0.41.2
- Pytest version: 7.4.0
- OS version: Ubuntu 22.04
- Python version: 3.10.12
Description
I have a test suite that started life in nosetests, and a lot of the tests still use xunit style with a
setup_methodthat creates fixture objects and stores them in the instance of the test class. I ran into a problem with tests exhausting resources, and it seems like pytest is hanging on to references to the instances of the test classes, preventing these fixture objects from being garbage collected.Here's a proof-of-concept:
conftest.py:
test_leak.py:
Output:
So despite the calls to
gc.collect(), the number of ExpensiveObjects is increasing. In contrast, if I use apytest.fixture, they are all cleaned up immediately.Environment
pip listfrom the virtual environment you are using