Choosing Version Numbers

As part of the push to release code from the oslo incubator in stand-alone libraries, we have had several different discussions about versioning and release schedules. This is an attempt to collect all of the decisions we have made in those discussions and to lay out the rationale for the approach we are taking.

Problem Description

We have two types of Oslo libraries. Libraries like oslo.config and oslo.messaging were created by extracting incubated code, updating the public API, and packaging it. Libraries like cliff and taskflow were created as standalone packages from the beginning, and later adopted by the Oslo team to manage their development and maintenance.

Incubated libraries have been released at the end of a release cycle, as with the rest of the integrated packages. Adopted libraries have historically been released “as needed” during their development.

The first release of oslo.config was 1.1.0, as part of the grizzly release. The first release of oslo.messaging was 1.2.0, as part of the havana release. oslo.config was also updated to 1.2.0 during havana. All “adopted” libraries (created elsewhere and brought into the Oslo project) had release numbers less than 1.0.0 when the original draft of this policy was written.

Proposed Policy

At the Juno summit, Josh Harlow proposed that we use semantic versioning (SemVer) for oslo libraries. Part of that proposal also included ideas for allowing breaking backwards compatibility at some release boundaries, and this policy does not yet address that issue. The first step is choosing a rational release versioning scheme.

SemVer is widely used and gives us relatively clear guidelines about choosing new version numbers. Oslo started using pbr’s modified SemVer for new releases, beginning in the Juno cycle.

SemVer Life Cycle

New libraries should start with version 0.1.0, incrementing following the SemVer policies through the end of the cycle with the goal of reaching 1.0.0 by the end of their first full development cycle.

Existing libraries will follow SemVer, incrementing from the version they had at the start of Kilo.

Frequent Releases

While we can run gate jobs using the master branch of Oslo libraries, developers will have to take extra steps to run unit tests this way locally. To reduce this process overhead, while still ensuring that developers use current versions of the code, we produce releases of libraries during the release cycle fairly frequently. We have a weekly check-up during the Oslo team meetings, and tag releases early on Mondays when deemed necessary. Waiting until Monday prevents us from introducing a gate issue just before the weekend starts.

Patch Releases for Stable Branches

Updates to existing library releases can be made from stable branches. Checking out stable/icehouse of oslo.config for example would allow a release 1.3.1. We don’t have a formal policy about whether we will create patch releases, or whether applications are better off using the latest release of the library.

All libraries will need to maintain stable branches to support these patch releases. We will cap the versions of Oslo libs used in stable branches to allow patch releases but not updates with minor version number changes.

Alternatives & History

Synchronizing with the Rest of OpenStack

The Oslo team could release libraries at any point, without concern for the release schedule of the rest of OpenStack. We have to be prepared for libraries we do not maintain to be updated at any point, so this wouldn’t be adding a new aspect to our release and testing processes. However, Oslo is part of OpenStack and so we initially wanted to be on the same schedule.

Mark McLoughlin has written a good justificiation for this, which is summarized as “my instinct is to do everything just like any of the other core projects except in those cases where Oslo really is a special case.” With Oslo following the release schedule of the other projects, we get all of the benefits (shifting focus from features to bugs; stable branches; synchronization with the users of our libraries; the OpenStack release manager).

When we stopped creating Alpha releases, we stopped full release synchronization. We do still release a final version for a given major and minor version number at the end of a release, and we do still follow the feature freeze process.

Alpha Releases

In the past, alpha releases of Oslo libraries have been distributed as tarballs on an openstack server, with official releases going to PyPI. Applications that required the alpha release specified the tarball in their requirements list, followed by a version specifier. This allowed us to prepare alpha releases, without worrying that their release would break continuous-deployment systems by making new library releases available to pip. This approach still made it difficult for an application developer to rely on new features of an oslo library, until an alpha version was produced.

When the PyPI mirror was introduced in our CI system, relying on tarballs not available on the mirror conflicted with our desire to have the gate system install only from the package mirror. When we started installing only from the mirror, we needed to publish our alpha releases in a format that will work with the mirror, and so we started using alpha version numbers of predicted final versions during a release.

At the Kilo summit, a vocal group of consumers of Oslo libraries requested that we stop using alpha versioning and switch to simple SemVer. We started doing that in Kilo, with mixed results (breaking changes made it into the stable branch test environments). At this point, there is no sense in going back to alpha releases during Kilo, so we will stick with the current plan and work through the resulting issues.

https://etherpad.openstack.org/p/kilo-oslo-alpha-versioning

We decided that new libraries should start with version 0.1.0, incrementing following the SemVer policies through the end of the cycle with the goal of reaching 1.0.0 by that time.

Existing libraries will follow SemVer, incrementing from the version they had at the start of Kilo.

Juno Policy

The versions for existing libraries oslo.config and oslo.messaging will be incremented from their Icehouse versions but updating the minor number (1.x.0) at the end of the Juno cycle.

All adopted libraries using numbers less than 1.0 will be released as 1.0.0 at the end of the Juno cycle, based on the fact that we expect deployers to use them in production.

Releases of new libraries graduated during Juno will be tagged with regular release numbers < 1.0. This allows us to add them to our requirements list (which won’t accept alphas of packages with no other release). At the end of Juno, we will tag the libraries 1.0.0.

Releases of existing libraries during Juno should all be marked as alphas of the anticipated upcoming SemVer-based release number (1.0.0.0aN or 1.2.0.0aN or whatever). The new CI system can create packages as Python wheels and publish them to the appropriate servers, which means projects will no longer need to refer explicitly to pre-release tarballs. pip won’t install alpha libraries unless you explicitly request them with a command line flag to install any alphas available or you explicitly require the alpha version. pip <= 1.3 didn’t support the flag for controlling alphas (they were always seen and installed), but also didn’t support wheels, so we publish alphas only as wheels to ensure that older pips don’t see them.

Cross-Project Unit Testing in the Gate

We had a blueprint for Juno to add cross-project unit test gating for applications and oslo libraries. This would have allowed us to verify that tests for applications do not break then Oslo libraries change, but also that those tests do not make assumptions about Oslo library implementation details. However, this level of testing was deemed too expensive in terms of test servers, and so the plan was dropped.

Capping Requirements in Stable Branches

We do not typically use upper bounds on the requirements specifications for Oslo libraries, so new releases may automatically be adopted by continuous-deployment systems building packages for stable branches of OpenStack applications. Although we have been careful about API compatibility in the past, there is a chance that a new release could break an older application. Applications could add an upper bound using SemVer numbering if they choose, although that may prevent them from seeing bug fixes so it is not recommended.

During the Kilo summit we discussed capping versions of requirements in stable branches. The initial attempt to do this right after the summit failed because it prevented some upgrades from working correctly. Work to apply caps is ongoing, and is outside of the scope of this policy.

Tagging milestones in libraries

We don’t tag libraries at the milestones like we do with applications, since the tags we use for milestones (e.g., 2014.2.b1) aren’t valid versions for libraries and would be out of order with other releases anyway. We may tag an alpha release around the time of the milestone, but since we do those on demand anyway there’s no strict rule that we must do it at that time.

Implementation

Author(s)

Primary author: doug-hellmann

Other contributors: markmc

Milestones

The current policy was put into place for Kilo.

Work Items

N/A

References

Revision History

Revisions

Release Name

Description

Juno

Introduced

Note

This work is licensed under a Creative Commons Attribution 3.0 Unported License. http://creativecommons.org/licenses/by/3.0/legalcode