The documented pattern for parametrizing conditional raising leaks a lot of memory

[emontnemery]

The documented pattern for parametrizing conditional raising leaks a lot of memory.

The problem is that the pytest.raises context manager is mutated with a reference to the raised exception and traceback when exiting the context manager: https://github.com/pytest-dev/pytest/blob/923044785f9fb7b1fac2c5d52df1fb8ea84ffc37/src/_pytest/python_api.py#L1047-L1048

This causes the traceback to be kept in memory until the test session terminates.

If this is indeed the correct way for this kind of parametrization, pytest should do some automatic clean up after the test finishes.

Otherwise, the documented parametrization should be adjusted to not pass in mutable objects, for example:

@pytest.mark.parametrize(
    "example_input,expectation",
    [
        (3, lambda: nullcontext(2)),
        (2, lambda: nullcontext(3)),
        (1, lambda: nullcontext(6)),
        (0, lambda: pytest.raises(ZeroDivisionError)),
    ],
)
def test_division(example_input, expectation):
    """Test how much I know division."""
    with expectation() as e:
        assert (6 / example_input) == e

The same problem happens if the parametrization involves exception objects which are raised during the test, the objects are mutated by Python which adds a __traceback__ to them:

@pytest.mark.parametrize("error", [Exception("Boom!"),])
def test_division(error):
    """Blow up."""
    raise error