detect-test-pollution

a tool to detect test pollution

installation

pip install detect-test-pollution

what is test pollution?

test pollution is where a test fails due to the side-effects of some other test in the test suite.

it usually appears as a "test flake" something where the test fails mysteriously but passes when run by itself.

a simple example of this is the following python code:

k = 1

def test_k():
    assert k == 1

def test_k2():
    global k

    k = 2
    assert k == 2

now this example is a little bit silly, you probably wouldn't write code this poorly but helps us demonstrate the problem here.

when run normally -- these tests pass:

$ pytest -q t.py
..                                                                       [100%]
2 passed in 0.00s

but, if the tests were run in some other order (due to something like pytest-randomly or pytest-xdist) then the pollution would be apparent:

$ pytest -q t.py::test_k2 t.py::test_k
.F                                                                       [100%]
=================================== FAILURES ===================================
____________________________________ test_k ____________________________________

    def test_k():
>       assert k == 1
E       assert 2 == 1

t.py:4: AssertionError
=========================== short test summary info ============================
FAILED t.py::test_k - assert 2 == 1
1 failed, 1 passed in 0.03s

often this flake happens in a codebase with hundreds or thousands of tests and it's difficult to track down which test is causing the global side-effects.

that's where this tool comes in handy! it helps you find the pair of tests which error when run in order.

usage

once you have identified a failing test, you'll be able to feed it into detect-test-pollution to find the causal test.

the basic mode is to run:

detect-test-pollution \
    --failing-test test.py::test_id_here \
    --tests ./tests

where test.py::test_id_here is the identifier of the failing test and ./tests is the directory where your testsuite lives.

if you've already narrowed down the list of testids further than that, you can specify a --testids-file instead of --tests to speed up discovery:

detect-test-pollution \
    --failing-test test.py::test_id_here \
    --testids-file ./testids

you can usually get a list of testids via pytest --collect-only -q (though you'll need to strip some unrelated lines at the end, such as timing and warning info).

then detect-test-pollution will bisect the list of tests to find the failing one. here's an example bisection from a bug in pytest

$ detect-test-pollution --tests ./testing --failing-test testing/io/test_terminalwriter.py::test_should_do_markup_FORCE_COLOR
discovering all tests...
-> discovered 3140 tests!
ensuring test passes by itself...
-> OK!
ensuring test fails with test group...
-> OK!
running step 1:
- 3139 tests remaining (about 12 steps)
running step 2:
- 1570 tests remaining (about 11 steps)
running step 3:
- 785 tests remaining (about 10 steps)
running step 4:
- 393 tests remaining (about 9 steps)
running step 5:
- 197 tests remaining (about 8 steps)
running step 6:
- 99 tests remaining (about 7 steps)
running step 7:
- 50 tests remaining (about 6 steps)
running step 8:
- 25 tests remaining (about 5 steps)
running step 9:
- 12 tests remaining (about 4 steps)
running step 10:
- 6 tests remaining (about 3 steps)
running step 11:
- 3 tests remaining (about 2 steps)
double checking we found it...
-> the polluting test is: testing/test_terminal.py::TestTerminal::test_report_teststatus_explicit_markup

fuzzing

detect-test-pollution can also be used to "fuzz" out failing tests.

it does this by shuffling the test ids and running the testsuite until it fails.

here's an example execution on a silly testsuite:

$ detect-test-pollution --fuzz --tests t.py
discovering all tests...
-> discovered 1002 tests!
run 1...
-> OK!
run 2...
-> found failing test!
try `detect-test-pollution --failing-test t.py::test_k --tests t.py`!

afterwards you can use the normal mode of detect-test-pollution to find the failing pair.

supported test runners

at the moment only pytest is supported -- though in theory the tool could be adapted to support other python test runners, or even other languages.

Oven