Bulldozer00's Blog

“Jane, you ignorant slut” – Dan Ackroid

In this EETimes post, I Don’t Need No Stinkin’ Requirements, Jon Pearson argues for the commencement of coding and testing before the software requirements are well known and recorded. As a counterpoint, Jack Ganssle wrote this rebuttal: I Desperately Need Stinkin Requirements. Because software industry illuminaries (especially snake oil peddling methodologists) often assume all software applications fit the same mold, it’s important to note that Pearson and Gannsle are making their cases in the context of real-time embedded systems – not web based applications. If you think the same development processes apply in both contexts, then I urge you to question that assumption.

I usually agree with Ganssle on matters of software development, but Pearson also provides a thoughtful argument to start coding as soon as a vague but bounded vision of the global structure and base behavior is formed in your head. On my current project, which is the design and development of a large, distributed real-time sensor that will be embedded within a national infrastructure, a trio of us have started coding, integrating, and testing the infrastructure layer before the application layer requirements have been nailed down to a “comfortable degree of certainty”.

The simple figures below show how we are proceeding at the current moment. The top figure is an astronomically high level, logical view of the pipe-filter archetype that fits the application layer of our system. All processes are multi-threaded and deployable across one or more processor nodes. The bottom figure shows a simplistic 2 layer view of the software architecture and the parallel development effort involving the systems and software engineering teams. Notice that the teams are informally and frequently synchronizing with each other to stay on the same page.

The main reason we are designing and coding up the infrastructure while the application layer requirements are in flux is that we want to measure important cross-cutting concerns like end-to-end latency, processor loading, and fault tolerance performance before the application layer functionality gets burned into a specific architecture.

So, do you think our approach is wasteful? Will it cause massive, downstream rework that could have been avoided if we held off and waited for the requirements to become stable?