Jeff Quast

Python Engineer

On using tox in CI environments

From a single tox.ini file, all stages of software development may be defined and easily shared. We've removed the need of programming in any kind of OS-dependent bash, batch, or Makefile. The barriers for discovering how to prepare an environment and execute tests is reduced for newcomers, and reduces the cost of making changes in the test->build->release cycle. We may easily test new versions of python as they are released with minimal changes.

With this, we now have 3 basic targets and can use tox to execute those specified by the envlist option, or list and execute individual targets. The target testenv is special, allowing invocation of explicit python versions. For example, to execute the same tests using python2.7, we run:

tox -epy27

even though py27 is not explicitly defined here, it has an implicit relationship with testenv. If no target is specified, py34 is used as defined by envlist.

  • a fully featured example:

    commands = {envbindir}/pytest \
                   --log-format='%(levelname)s %(relativeCreated)2.2f %(filename)s:%(lineno)d %(message)s' \
                   --cov={toxinidir}/telnetlib3 \
                   --cov-config={toxinidir}/.coveragerc \
                   --strict --verbose --verbose \
                   } telnetlib3/tests
               coverage combine
               cp \
                   {toxinidir}/.coverage \
  • Interestingly, we can create a tox target command to execute only {posargs} when given instead of the standard default target:

    deps = pytest
    commands = {posargs:pytest --looponfail --verbose --verbose}
    looponfailroots = qwack

    This allows us to maintain a common environment from which to execute commands in. For example, we could specify a flake8 dependency along with the project's requirements.txt file and execute our preferred emacs or vim editor:

    tox -edevelop -- vim qwack/

    (idea contributed by signalpillar)

  • Our CI systems can make use of the {posargs:default value} to generate a great deal of detail, here is a more complex py.test command that produces a jUnit xml file and coverage report by envname:

    deps = pytest-cov
    commands = pytest {posargs:\
                   --strict --verbose --verbose --color=yes \
                   --junit-xml=results.{envname}.xml \
                   --cov qwack qwack/tests}

    Or, we can override it, maybe, tox -e pytest -- -k mytest

  • To create test groups, or a "quick test" feature, we can use the tox setenv and passenv options to specify a TEST_QUICK environment variable, conditioned by pytest.mark.skipif:

    import pytest
    @pytest.mark.skipif(os.environ.get('TEST_QUICK', None) is not None,
                        reason="TEST_QUICK specified")
    def test_something_very_slow():

    Would allow command invocation from bash as QUICK=1 tox -epy35 when used with passenv, or a specific quicktest target using setenv to skip such long-running tests in our testenv target.

  • When integrating with a CI system, one should still provide simple,, or targets, even if they drive tox, this allows us track precisely which tox targets are used in pairing stages of a build. This way, if the CI system reports a failure in the publish step, a developer should be able to execute locally to reproduce, while also allowing changes in build stages to be tracked by version control. Release engineers should avoid placing special shell code inside HTML textarea windows, which would fail to accommodate making historic builds, or allow changes to the build steps without peer review.

  • We can use passenv to allow our CI system to publish our coverage to the service, or communicate with Jira or GitHub without storing them in VCS by using environment variables:

    deps = coveralls
    commands = coveralls

    The environment variable COVERALLS_REPO_TOKEN would be hidden from any non-administrator accounts of the CI system, and outside of VCS.