测试人社区

Pytest-Flaky tests

A “flaky” test is one that exhibits intermittent or sporadic failure,
that seems to have non-deterministic behaviour. Sometimes it passes,
sometimes it fails, and it’s not clear why. This page discusses pytest
features that can help and other general strategies for identifying,
fixing or mitigating them.

Why flaky tests are a problem

Flaky tests are particularly troublesome when a continuous integration
(CI) server is being used, so that all tests must pass before a new code
change can be merged. If the test result is not a reliable signal –
that a test failure means the code change broke the test – developers
can become mistrustful of the test results, which can lead to
overlooking genuine failures. It is also a source of wasted time as
developers must re-run test suites and investigate spurious failures.

Potential root causes

System state

Broadly speaking, a flaky test indicates that the test relies on some
system state that is not being appropriately controlled - the test
environment is not sufficiently isolated. Higher level tests are more
likely to be flaky as they rely on more state.

Flaky tests sometimes appear when a test suite is run in parallel (such
as use of pytest-xdist). This can indicate a test is reliant on test
ordering.

  • Perhaps a different test is failing to clean up after itself and
    leaving behind data which causes the flaky test to fail.
  • The flaky test is reliant on data from a previous test that doesn’t
    clean up after itself, and in parallel runs that previous test is
    not always present
  • Tests that modify global state typically cannot be run in parallel.

Overly strict assertion

Overly strict assertions can cause problems with floating point
comparison as well as timing issues.
pytest.approx
is useful here.

Pytest features

Xfail strict

pytest.mark.xfail ref{.interpreted-text role=“ref”} with
strict=False can be used to mark a test so that its failure does not
cause the whole build to break. This could be considered like a manual
quarantine, and is rather dangerous to use permanently.

PYTEST_CURRENT_TEST

PYTEST_CURRENT_TEST{.interpreted-text role=“envvar”} may be useful for
figuring out “which test got stuck”. See
pytest current test env{.interpreted-text role=“ref”} for more
details.

Plugins

Rerunning any failed tests can mitigate the negative effects of flaky
tests by giving them additional chances to pass, so that the overall
build does not fail. Several pytest plugins support this:

Plugins to deliberately randomize tests can help expose tests with state
problems:

Other general strategies

Split up test suites

It can be common to split a single test suite into two, such as unit vs
integration, and only use the unit test suite as a CI gate. This also
helps keep build times manageable as high level tests tend to be slower.
However, it means it does become possible for code that breaks the build
to be merged, so extra vigilance is needed for monitoring the
integration test results.

Video/screenshot on failure

For UI tests these are important for understanding what the state of the
UI was when the test failed. pytest-splinter can be used with plugins
like pytest-bdd and can save a screenshot on test
failure
,
which can help to isolate the cause.

Delete or rewrite the test

If the functionality is covered by other tests, perhaps the test can be
removed. If not, perhaps it can be rewritten at a lower level which will
remove the flakiness or make its source more apparent.

Quarantine

Mark Lapierre discusses the Pros and Cons of Quarantined
Tests

in a post from 2018.

CI tools that rerun on failure

Azure Pipelines (the Azure cloud CI/CD tool, formerly Visual Studio Team
Services or VSTS) has a feature to identify flaky
tests

and rerun failed tests.

Research

This is a limited list, please submit an issue or pull request to expand
it!

  • Gao, Zebao, Yalan Liang, Myra B. Cohen, Atif M. Memon, and Zhen
    Wang. “Making system user interactive tests repeatable: When and
    what should we control?.” In Software Engineering (ICSE), 2015
    IEEE/ACM 37th IEEE International Conference on
    , vol. 1, pp. 55-65.
    IEEE, 2015. PDF
  • Palomba, Fabio, and Andy Zaidman. “Does refactoring of test smells
    induce fixing flaky tests?.” In Software Maintenance and Evolution
    (ICSME), 2017 IEEE International Conference on
    , pp. 1-12.
    IEEE, 2017. PDF in Google
    Drive
  • Bell, Jonathan, Owolabi Legunsen, Michael Hilton, Lamyaa Eloussi,
    Tifany Yung, and Darko Marinov. “DeFlaker: Automatically detecting
    flaky tests.” In Proceedings of the 2018 International Conference
    on Software Engineering
    . 2018.
    PDF

Resources