As you would expect in a conference entitled velocity, and in a follow-on devops day, speeding up things was an overarching theme. In the context of devops, the theme primarily manifested itself in lively discussions about the number of deploys per day. Comments such as the following reply to my post Ops Driven Dev were typical:
Conceptually, I move the whole business application configuration into the source code…
The theme that was missing for me in many of the presentations and discussions on the subject was the striking of a balance between velocity and quality. The classical trade-off in process control is between production rate and product quality (and safety, but that aspect [safety] is beyond the scope of this post). IMHO this trade-off applies to software just as it applies to mechanical or chemical processes.
The heart of the “deploy early and often” strategy hailed by advocates of continuous deployment is known deployment state to known deployment state. You don’t let the deployment evolve from one state to another before it has stabilized to a robust state. The power of this incremental deployment is in dealing with single-piece (or as small number of pieces as possible) flow rather than dealing with the effects of multiple-piece flow. When the deployment increments are small enough, rollback, root cause analysis and recovery are relatively straightforward if a deployment turns sour. It is a similar concept to Agile development, extending continuous integration to continuous deployment.
While I am wholeheartedly behind this devops strategy, I believe it needs to be reinforced through rigorous quality criteria the code must satisfy prior to deployment. The most straightforward way for so doing is through embedding technical debt criteria in the release/deploy process. For example:
- The code will not be deployed unless the overall technical debt per line of code is lower than $2.
- To qualify for deployment, code duplication levels must be kept under 8%.
- Code whose Cyclomatic complexity per Java class is higher than 15 will not be accepted for deployment.
- 50% unit test coverage is the minimal level required for deployment.
- Many others…
I have no doubt whatsoever that code which does not satisfy these criteria might be successfully deployed in a short-term manner. The problem, however, is the accumulative effect over the long haul of successive deployments of code increments of inadequate quality. As Figure 1 demonstrates, a Java file with Cyclomatic complexity of 38 has a probability of 50% to be error-prone. If you do not stop it prior to deployment through technical debt criteria, it is likely to affect your customers and play havoc with your deployment quite a few times in the future. The fact that it did not do so during the first hour of deployment does not guarantee that such a file will be “well-behaved” in the future.
Figure 1: Error-proneness as a Function of Cyclomatic Complexity (Source: http://www.enerjy.com/blog/?p=198)
To attain satisfactory long-term quality and stability, you need both the right process and the right code. Continuous deployment is the “right process” if you have developed the deployment infrastructure to support it. The “right code” in this context is code whose technical debt levels are quantified and governed prior to deployment.
