[Note - I understand there are ongoing discussion that may lead to a proposal for an out-of-tree incubation process for new Neutron features. This is a complementary proposal that describes how our existing development process can be used to stabilize new features in-tree over the time frame of a release cycle or two. We should fully consider both proposals, and where each might apply. I hope something like the approach I propose here will allow the implementations of Neutron BPs with non-trivial APIs that have been targeted for the Juno release to be included in that release, used by early adopters, and stabilized as quickly as possible for general consumption.]

According to our existing development process, once a blueprint and associated specification for a new Neutron feature have been reviewed, approved, and targeted to a release, development proceeds, resulting in a series of patches to be reviewed and merged to the Neutron source tree. This source tree is then the basis for milestone releases and the final release for the cycle.

Ideally, this development process would conclude successfully, well in advance of the cycle's final release, and the resulting feature and its API would be considered fully "stable" in that release. Stable features are ready for widespread general deployment. Going forward, any further modifications to a stable API must be backwards-compatible with previously released versions. Upgrades must not lose any persistent state associated with stable features. Upgrade processes and their impact on a deployments (downtime, etc.) should be consistent for all stable features.

In reality, we developers are not perfect, and minor (or more significant) changes may be identified as necessary or highly desirable once early adopters of the new feature have had a chance to use it. These changes may be difficult or impossible to do in a way that honors the guarantees associated with stable features.

For new features that effect the "core" Neutron API and therefore impact all Neutron deployments, the stability requirement is strict. But for features that do not effect the core API, such as services whose deployment is optional, the stability requirement can be relaxed initially, allowing time for feedback from early adopters to be incorporated before declaring these APIs stable. The key in doing this is to manage the expectations of developers, packagers, operators, and end users regarding these new optional features while they stabilize.

I therefore propose that we manage these expectations, while new optional features in the source tree stabilize, by clearly labeling these features with the term "preview" until they are declared stable, and sufficiently isolating them so that they are not confused with stable features. The proposed guidelines would apply during development as the patches implementing the feature are first merged, in the initial release containing the feature, and in any subsequent releases that are necessary to fully stabilize the feature.

Here are my initial not-fully-baked ideas for how our current process can be adapted with a "preview feature" concept supporting in-tree stabilization of optional features:

* Preview features are implementations of blueprints that have been reviewed, approved, and targeted for a Neutron release. The process is intended for features for which there is a commitment to add the feature to Neutron, not for experimentation where "failing fast" is an acceptable outcome.

* Preview features must be optional to deploy, such as by configuring a service plugin or some set of drivers. Blueprint implementations whose deployment is not optional are not eligible to be treated as preview features.

* Patches implementing a preview feature are merged to the the master branch of the Neutron source tree. This makes them immediately available to all direct consumers of the source tree, such as developers, trunk-chasing operators, packagers, and evaluators or end-users that use DevStack, maximizing the opportunity to get the feedback that is essential to quickly stabilize the feature.

* The process for reviewing, approving and merging patches implementing preview features is exactly the same as for all other Neutron patches. The patches must meet HACKING standards, be production-quality code, have adequate test coverage, have DB migration scripts, etc., and require two +2s and a +A from Neutron core developers to merge.

* DB migrations for preview features are treated similarly to other DB migrations, forming a single ordered list that results in the current overall DB schema, including the schema for the preview feature. But DB migrations for a preview feature are not yet required to preserve existing persistent state in a deployment, as would be required for a stable feature.

* All code that is part of a preview feature is located under neutron/preview/<feature>/. Associated unit tests are located under neutron/tests/unit/preview/<feature>/, and similarly for other test categories. This makes the feature's status clear to developers and other direct consumers of the source tree, and also allows packagers to easily partition all preview features or individual preview features into separate optionally installable packages.

* The tree structures underneath these locations should make it straightforward to move the preview feature code to its proper tree location once it is considered stable.

* Tempest API and scenario tests for preview features are highly desirable. We need to agree on how to accomplish this without preventing necessary API changes. Posting WIP patches to the Tempest project may be sufficient initially. Putting Tempest-like tests in the Neutron tree until preview features stabilize, then moving them to Tempest when stabilization is complete, might be a better long term solution.

* No non-preview Neutron code should import code from anywhere under the neutron.preview module, unless necessary for special cases like DB migrations.

* URIs for the resources provided by preview features should contain the string "preview".

* Configuration file content related to preview features should be clearly labeled as "preview".

* Preview features should be documented similarly to any stable Neutron feature, but documents or sections of documents related to preview features should have an easily recognizable label that clearly identifies the feature as a "preview".

* Support for preview features in client libraries, and in other projects such as Horizon, Heat, and DevStack, are essential to get the feedback needed from early adopters during feature stabilization. They are implemented normally, but should be labeled "preview" appropriately, such as in GUIs, in CLI help strings and in documentation so that end user expectations regarding stability are managed.

* A process is needed to prevent long-term stagnation of features in the preview sub-tree. It is reasonable to expect a new feature to remain for one or two cycles, possibly with little change (other than bug fixes), before stabilizing. A suggested rule is that a new approved BP is required after two cycles, or the feature gets removed from the Neutron source tree (maybe moved (back) to an incubation repository).


I would appreciate feedback via this email thread on whether this "preview feature" concept is worth further consideration, refinement and potential usage for approved feature blueprints, especially during the Juno cycle. I've also posted the proposal text at https://etherpad.openstack.org/p/neutron-preview-features for those interested in helping refine the proposal.

Thanks,

-Bob


_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev

Reply via email to