Run under Python 3 by default

As we agreed in the 2018-05-29 Python2 Deprecation Timeline resolution, the next phase of our adoption of python 3 is to begin running all jobs using python 3 by default, and only using python 2 to test operating under python 2 (via unit, functional, or integration tests). This goal describes the activities needed to move us to this “python 3 first” state.

Champions

Gerrit Topic

python3-first

Completion Criteria

  1. The Zuul settings to attach jobs have been moved from the openstack-infra/project-config repository into each project repository.

  2. Documentation build and publish jobs use python 3.

  3. Release notes build and publish jobs use python 3.

  4. Source code linter jobs (pep8, pylint, bandit, etc.) should use python 3.

  5. Release artifact build and publish jobs use python 3 and publish to PyPI.

  6. There are functional test jobs running under python 3.

  7. The wiki tracking page is up to date for the project.

  8. Projects are running python 3.6 unit test jobs.

The goal champions will start by preparing the patches for steps 1-5. That will give project teams more time to focus on steps 6-8, which will need the expertise of someone on the team. Project teams should be prepared to review the changes quickly, and assist if there are any questions about jobs with branch-specifiers (see below). Step 7, updating the wiki tracking page, is the responsibility of the project team. Steps 6 and 8 involve adding more test jobs and may also require code changes or tox settings updates, so those will be left to the project team as well.

Details

Moving settings from project-config

A big part of the work for this goal for each project is to configure different sets of jobs to run the same tasks differently on different branches. Although Zuul supports doing this using branch specifiers, it will be easier to maintain the job settings over time if we place them in each repository. This lets us set different jobs, or variants of jobs, to run against each branch just by modifying the Zuul settings in that branch.

The first step, therefore, is to take the project settings that control the branch-aware jobs out of the openstack-infra/project-config repository and put them into each git repository in the appropriate branch.

https://review.opendev.org/580736/ is an example of adding the project settings to the master branch of openstack/whereto. The patch was constructed by using the python3-first command from the openstack/goal-tools repository (see the README.rst there for directions).

Note

Not all of the job settings should be moved into the project tree:

  1. The line applying the translation-jobs project template (or any of its branch-specific variants) should stay in openstack-infra/project-config.

  2. Settings for jobs that should only ever run on the master branch should stay in openstack-infra/project-config.

  3. Any project templates that are not “branch-aware” (because they run based on a tag event) need to remain configured in zuul.d/projects.yaml in openstack-infra/project-config.

Each repository needs a patch like 574393 on the master branch. The settings can go into any of the Zuul configuration files (.zuul.yaml, zuul.yaml, .zuul.d/project.yaml, etc.) and should be organized based on the way the project team has their Zuul configuration set up already.

Similar patches should then be created for the stable/rocky, stable/queens, stable/pike, and stable/ocata branches. Sometimes it will be possible to just backport the change from the master branch, but many projects have jobs configured with branch specifiers to control which branches run the tests. We do not want to mix branch specifiers with branch-specific configuration, so the settings will need to be modified or removed. This may mean that the versions of the change on stable branches are different from the version on master. The python3-first jobs extract command in the goal-tools repository should correctly compute the set of jobs to move into each branch.

After all of the settings are in place within the branches of a repository, they can be removed from openstack-infra/project-config. https://review.opendev.org/580737/ shows how to remove the settings for whereto.

Note

Due to the large number of changes involved, we want to update the openstack-infra/project-config repository with 1 patch per project team. That patch must not be approved before all of the settings are added to each repository in all of the relevant branches. Normally we would manage that requirement using Depends-On to link the patches together, but for some teams that will produce a very very long list of inter-patch dependencies (Oslo would take 128 patches). Rather than listing them all as dependencies, therefore, we will wait to submit the update to project-config until after all of the changes to the project repositories have been approved.

At this point, most of the job configuration for each repository can be managed directly by the project team, and each branch will have a separate list of jobs. This will make future job management much simpler and should make it possible for the project team to do the rest of the work for the goal without waiting for anyone else to review their patches.

Documentation Jobs

The publish-openstack-docs-pti project template defines a set of jobs for testing and building documentation using tox as a wrapper around Sphinx, and replaces the old publish-openstack-sphinx-docs project template. The new jobs use a docs tox environment, defined in each repository, so the tox settings can be used to provide a self-testing patch to the repository to switch the version of python used.

https://review.opendev.org/580738 shows how to change the project settings for a repository to use the new documentation jobs. The change will be self-testing, and should only be made on the master branch.

https://review.opendev.org/572895 shows how to update the tox.ini settings in the project to set the basepython variable for the docs environment, used for developer testing. This change should only be made on the master branch.

Common issues to anticipate:

  • Under python 3 the output of subprocess.check_output() is a bytes instance, but sphinx expects values for version and timestamps to be str objects. Doc builds that do things in conf.py (or extensions) like extract the modification date from the most recent commit will need to properly decode the return values. https://review.opendev.org/#/c/575483 shows one example of how to fix this sort of problem.

Release Notes Jobs

The release-notes-jobs-python3 project template defines a set of jobs for testing and building release notes using python 3. The release notes jobs do not use tox, but the tox settings should still be updated.

In the Zuul configuration on the master branch of the repository, change the project template release-notes-jobs to release-notes-jobs-python3. If the patch to change the project template does not run the new job, it may be necessary to add a dummy release note to make the patch self-testing.

https://review.opendev.org/#/c/572895/ shows how to update the tox.ini settings in the project to set the basepython variable for the releasenotes environment, used for developer testing. This change should only be made on the master branch.

Common issues to anticipate:

  • Under python 3 the output of subprocess.check_output() is a bytes instance, but sphinx expects values for version and timestamps to be str objects. Doc builds that do things in conf.py (or extensions) like extract the modification date from the most recent commit will need to properly decode the return values.

Source Code Linter Jobs

Most of the jobs we have that run source code linters do use tox to control the versions of the linter tool. These jobs typically have names like openstack-tox-linters or openstack-tox-pep8.

https://review.opendev.org/#/c/572895/ shows how to update the tox.ini settings in the project to set the basepython variable for the environments. All of the linter jobs running against python source code in the master branch should be updated. These changes should be self-testing.

Any linter jobs that use python-based tools to check other sorts of source should also be updated to use python 3, if possible.

Common issues to anticipate:

  • The built-in file``() no longer exists under python 3, so using it causes pylint to report an undefined name. Use ``open() instead.

  • The built-in unicode no longer exists under python 3, so using it causes pylint to report an undefined name. Use six.text_type instead.

  • Under python 3 the flake8/hacking/pep8/pylint tools run different or additional checks. This may mean new code formatting issues will have to be fixed as part of changing the linter jobs over.

  • There is a bug in the older version of pylint that many projects are using that prevents it from working correctly under python 3. Pylint will have to be upgraded as part of this transition; version 1.9.2 is known to work.

    The error message from the broken version is:

    AttributeError: 'Call' object has no attribute 'starargs'
    

    Updating the version of pylint brings new rules, and will require modifications either to source code or to the pylint configuration. https://review.opendev.org/#/c/573024/ is an example of updating to the latest version of pylint in the freezer repository, with a combination of fixes and disabling rules.

Release Artifact Publishing

We will be making several changes to artifact publishing for Python-based projects simultaneously. The job settings for the release artifact publishing need to be defined in openstack-infra/project-config/zuul.d/projects.yaml rather than in each project repository, because those jobs are not “branch aware” and therefore we do not want different versions of the jobs on different branches.

First, a new job that uses setuptools to validate the packaging metadata for a repository will run in the check and gate queues when README.rst, setup.cfg, or setup.py are modified. This will be an early warning for issues that may come up as part of publishing the build artifacts, and runs the same step that was added recently to the validation job in openstack/releases as well as actually building an sdist and a wheel.

Second, the new packaging test, build, and publish jobs will all run under python 3.

Third, all python-based deliverables will have their sdists and wheels published to PyPI. This will simplify dependency management between plugins and server projects and will streamline the number of variations of release jobs that we have.

In order to make this change, project teams may first need to register their project name on PyPI. Refer to the creator’s guide in the infra manual for details of how to do this.

Note

The original goal contained an additional step: “After the name is configured on PyPI, change any existing release project template to publish-to-pypi-python3. https://review.opendev.org/580740 shows an example of changing the job setting.”

Infra now has made publish-to-pypi using python3, so all repos use again publish-to-pypi and thus the step mentioned above is not needed anymore.

It is not possible to test the job change, because it needs to be made in the openstack-infra/project-config repository. Therefore, after the first change merges it will be useful to create a second patch in the project repository with a whitespace or other typo-fix change in the README.rst to trigger the packaging test job in this patch to ensure everything works properly. https://review.opendev.org/580741 shows an example of such a change.

Note

Teams using release jobs that rely on python to publish artifacts for projects not written in python (and therefore not covered by publish-to-pypi-python3) should work with the release and infra teams to update their release jobs to use python 3.

Common issues to anticipate:

  • Projects that have not published to PyPI before may need to fix their README.rst file if it uses RST directives only defined by Sphinx and not by docutils. The new test job will catch any issues.

  • Projects that cannot reserve their project name on PyPI because it is owned by another community may need to change the sdist name in their setup.cfg in order to be able to publish to PyPI under a different name. That will not change how the code is imported, but it will change package names and may require setting tarball-base in the release settings managed in openstack/releases. The release management team can help if you end up needing to change names, so contact them before starting to make the change.

Functional Test Jobs

Updating the functional test jobs for a project will require more knowledge of the jobs that exist, which ones need to be duplicated under python 3, and which can be changed to python 3 without being run under python 2. Changing the job configuration will require knowledge of the job implementation details. For these reasons, the analysis and implementation work for updating the functional test jobs is left up to each project team.

Libraries used by services that run in the default integrated gate can add the lib-forward-testing-python3 project template to ensure they have full integration tests run. https://review.opendev.org/#/c/575927/ shows an example of doing this for oslo.config.

Where possible, when modifying existing jobs, a variable should be added to control the version of python so that the same job implementation (playbooks, roles, etc.) can be used instead of duplicating the entire job definition. This will simplify cleaning up the old job definitions when python 2 support is finally dropped.

It should be possible to update functional and integration test jobs that run through tox by setting basepython = python3 for the appropriate tox environment, as in https://review.opendev.org/#/c/572895/.

Wiki Tracking Page

We have been using https://wiki.openstack.org/wiki/Python3 to track the status of support in each project. Teams should keep the page up to date with information about blockers, test jobs, etc. as they work on this goal (and after, ideally).

Python 3.6 Unit Test Jobs

On the mailing list Zane proposed updating to test with Python 3.6 when it is available. Adding those test jobs will be easier after the Zuul configuration is moved out of the project-config repository, so this step is left for last. Because adding the test job may require code changes, it will be up to each project team to take this step by adding openstack-python36-jobs to the list of templates associated with the project on the master branch. The change will be self-testing, and can either be structured to include the code changes (if they are trivial) or end a series of patches (if the code changes are significant).

Note

We do not plan to update the minimum version of python 3 we support as part of this goal. Projects already running python 3.5 jobs should continue to do so.

References

Current State / Anticipated Impact

A significant number of patches to update the tox settings for projects have already been proposed and many have been merged:

https://review.opendev.org/#/q/topic:python3-first

Some of the Oslo libraries are using the python 3 versions of these jobs already.

Because the goal champion team will prepare a lot of the patches to move the Zuul settings, we expect project teams to be able to focus on unique aspects of their testing such as branch-specific jobs or functional jobs.