Monday, October 08, 2012

An architect's view of the MySQL Development model

When I stepped into my current job as Senior MySQL Architect in the summer of 2007 one of the first things on the agenda for discussion was the development model. We had released 5.0 almost 2 years ago and we were still more than a year from releasing MySQL 5.1. The model we used for developing MySQL 5.0 and MySQL 5.1 didn't work as well we wanted. So we had many discussions on how to improve this model.

A quick step 5 years forward to today, we are now using the Milestone release model and are just about to release the second GA version that have been developed using this model.

Differences in old and new development model

So what is the main difference between the new development model and the old development model?

I think the most important attribute of our current model is that the development tree always have to be in a releasable state every 2-4 months and this releaseable state have to be possible to bring to a GA version in around 6 months of stabilisation period. Our definition of releasable state is that the quality is at Release Candidate level. The old model effectively only required a releasable state when the version was reaching Release Candidate state since also in the beta stage new features were added, so a releasable state was only reached every 2 years instead of every 2-4 months.

Impact on new code in MySQL

So what does this mean for how we need to develop new code in MySQL. It means that new code must be added in fairly small chunks, even more it means that the new code cannot be too intermingled with old code since that would introduce too many new bugs. So if we need to add a new feature in an area where the complexity of the MySQL Server is high, then we first need to reengineer this area before we can proceed pushing the actual new feature. So in a sense the highest priority is always given to maintaining a stable development tree. In the past to get a new feature into the MySQL Server could sometimes have higher priority than the stability of the development tree.

So interestingly, by prioritizing tree stability before new features, we are able to develop more new features! But we are not necessarily able to develop specific new features as quick as we want sometimes. If a new feature requires tampering with complex MySQL Server code, then it might require a number of preparatory projects before they can actually be turned into a real feature. But anything that can be done outside of the complex parts of the MySQL Server can be added quite quickly.

Development of major features in new model

How does one handle addition of new major features in this model. The main consideration is that all new code must be fairly well structured. So when adding a new feature we can't simply start coding along, we first need to consider if the current code structure can be used to develop this new feature. As an example one of the last things added in MySQL 5.6 was a split of LOCK_open. If we would have tried to add this feature in earlier versions of the MySQL Server we would have had a new feature that touches many functional parts of the MySQL Server and it would have been introduced a significant set of new bugs to fix before any new GA version could have been released. Instead we went about things differently. First a number of reengineering projects isolated the main feature of LOCK_open to only protect metadata operations on tables and caching of metadata in connections. When these steps were completed about halfways through the MySQL 5.6 development, then it was very straightforward to make the actual split of LOCK_open in a fairly small patch which was a safe change.

Inspiration to new development model

Our selection of this new development model obviously was inspired by lessons from other successful open source projects such as the Linux kernel.

Change from old to new model

The change from the old model to the new model were probably the most difficult part. At the time of the change we had a new development tree which had some fairly significant new features implemented in it. But the tree was very far away from getting to a releasable state, so the tough decision was taken to restart development with a new tree based on the MySQL 5.1 development tree. This tree was the start of development using the new milestone model. The features from the old development tree was step by step moved over to the new development tree. This process actually took a few years since there are very high quality requirements on new features making it into the new development tree. But now we've moved all features from this old development tree that we will move, some of the features have been dropped or simply completely reimplemented.

Impact on future MySQL development

So what does this mean for current and future MySQL development? In MySQL 5.6 we added 200 new features, so obviously the new development model leads to a high pace of new interesting changes that will benefit our users and customers. We expect that the development model will continue to benefit us and that we will continue delivering new interesting features at a high pace to the MySQL community. As an architect I am proud to see how the MySQL development team continously improves the MySQL architecture and thus making it possible to add new features at an even higher pace.


杜修文 (Ivan Tu) said...

Great article, not only depict how endeavor from Oracle improve MySQL quality but also a good tool to communicate with the community, a Chinese translation available on

杜修文 (Ivan Tu) said...

Great post, shows how endeavor from Oracle improve MySQL quality, a Chinese translation of this is available from