Open development, releases and quality

Every 6 months, as a cycle ends and we prepare the next, we look back at our release model and try to see how we can improve it. My opinion is that we need (once more) to evolve it, and here is why.

Objectives and past evolutions

Our main objective is to produce stable and usable software. Our secondary objective is let stable new features reach the hands of our users in a timely manner. Our tertiary objective is to empower developers to work efficiently on the features and improvements of tomorrow. In simple solutions, those three objectives are not really compatible, and in the past we tried to make adjustments without acknowledging the fundamental incompatibility between those objectives.

The Austin-Cactus system was a 3-month cycle with various freezes. The issues with this system were that it took too long to get features in the hands of testers and users, developers had trouble working during the frozen periods, and not so many people actually helped during the QA period.

For Diablo, we decided to switch to a 6-month cycle with monthly milestones. Those milestones were supposed to address the "get features into the hands of users early" issue. To empower developers, we introduced an almost-always-open trunk: we only had 2 weeks of feature freeze before the Essex branch opened (and coexisted with the Diablo release branch). The problem was the resulting quality was not good, since we accumulated 6 months' worth of features and only had 2 weeks' worth of QA on them.

So for Essex, there was a decision that each project would decide where to place its feature freeze and its first release candidate (which would be when the Folsom branch opens). Most of them decided to use the essex-3 milestone as a soft feature freeze and essex-4 as a hard one, which reintroduced the "closed trunk" issue. And the work needed to stabilize 6 months' worth of features is still daunting. And we saw lots of feature freeze exceptions for last-minute "almost there" features that can't afford to wait another 6 months.

So what's the solution ? Is there a good way to (at the same time) reduce the pain of "missing a release", have always-open development branches, and get higher quality ?

The kernel model

The only way to reduce the pain of missing a release is to have shorter time-based releases. The only way to have always-open development branches without sacrificing release quality is to separate them completely from release branches. Where do those truisms lead us ?

Another famous open innovation project has already been there, and that's the Linux kernel. Development happens in various always-open topic branches, and regularly a merge window opens to propose stable features for inclusion in the mainline kernel.

If we manage to efficiently separate OpenStack in topics, we could adopt the same model. We could have several topic branches where core developers on a specific area can collaborate and review code affecting that area. We could have frequent releases (every 6-8 weeks ?), and for each release we could have a "merge window" where stuff from topic branches can be proposed for release, if deemed stable enough. Between the moment the merge window closes and the moment the final release is cut, various release candidates can be produced, on which serious QA can be unleashed without blocking other developers.

This solves all objectives. If a feature is not ready yet (according to the team maintaining the topic branch or according to release management), it can bake until the next merge window, which is not far away. Regular releases ensure that improvements reach the users in a timely manner. Development branches are always open. And with a reasonable amount of new code in every release, QA work is facilitated, theoretically resulting in higher release quality.

Additional benefits

Splitting development into topics has an additional benefit: smaller groups developing an expertise on a specific area make better reviewers on that specific area than a random nova-core developer that can't be an expert in all things Nova. I'd say that each topic should be small enough to be manageable with a team of 6 core reviewers at the maximum. Of course, nothing prevents anyone from being active on multiple topics if they wish.

More frequent releases also allows you to set themes for each release. It's difficult to refuse some feature in a 6-month cycle, but it's easy to delay a feature to the next 6-week cycle. In the same way some kernel releases introduce large architectural changes and some others are more geared towards performance improvements or stability, we could also define themes for every release -- after all the next one is not so far away.

Challenges

There are a few issues with this model obviously. Code needs to be componentized enough so that merge pains can be limited. Changes from every release need to be merged back into topic branches. Code needs to be clearly separated into a set of relevant topics, each with its own set of core reviewers maintaining the topic branch. The release team must be staffed enough to be able to review proposed code for stability. Bug fixes need to easily end up in the release branch. And how do those "releases" match the 6-month period between summits ? What means "Folsom" in that respect ?

I hope we can use the following weeks to discuss the devil in the details of this possible evolution, and be ready to take a decision when the time comes for us to gather at the Folsom Design summit.