Written January 20, 2016

Inherent Complexity of Interface Specfication

In Brook’s ‘No Silver Bullet’ paper he argues, impotently, that “much complexity [in software] comes from conformation to other interfaces”. The whole argument, which can be found in the “Conformity” section reduces to:

  1. Software needs to interact with the world and the world is complex, therefore software is complex, and/or
  2. Software is easy to change and thus must be the element which changes, therefore software is still complex.

To phrase it another way … water is wet.

Throughout his paper Brooks paints a likable characterization of the software engineering process, one that plays to the ego, to the self-centered. I fell for it. I drank up the Koolaid like so many a hapless lead engineer or inexperienced manager. It behooves us to. We are compelled. The argument makes us feel important. Unfortunately, the thought, which appears so palatable, is nominally vacuous. A conversation of the complexity introduced by interfaces, encompassing Brooks’ comment that re-engineering cannot eliminate the complexities of consuming interfaces, is implicit in the design process and is thus hardly worth mentioning in such isolated terms. Brooks misses the underpinnings of interface complexity completely with his hand-waving. Let me explain …

If we are to believe point one where the world’s man made interfaces are muddy, fully captured devices born of human flaw then one would suppose that the natural driving instinct of all system architects would be to shutter protocols capturing volumes of complexity fast inside components. This would have the immediate result of reducing the number, and likely complexity, of interfaces and necessitates the creation of a few large functional units where domain complexity is captured internally to those units.

Running somewhere between completely counter and moderately orthogonal to Brooks’ first point is his second. For the sake of re-iteration - because software is easy to change, it must be the element to change and conforming to extrinsic pressure drives complexity. Intuition says that Brooks is correct.
We enhance the changeability of our software by increasing the the count, and possibly the complexity of, the interfaces of a system. This concept may be exacted at different levels of abstraction, equally workable at the method level and all the way up to intra-system communications.

In either case however, capturing the notion of complexity arising from conforming to or confirming an interface is unimportant, and I would go so far as to say that holds true even nestled in conversation of essential vs accidental complexity. Let’s enumerate again the design options that arise from Brooks’ argument:

  1. We make large functional units that aptly capture domain complexity but resist change
  2. We make small functional units which enhance change, capture domain complexity but bring additional protocol complexity to bear.

Each point makes assumptions, alluding to the presence of external motivators. Point one subsumes that our software is in need of future change. Point two suggests that the addition of protocol complexity is somehow burdensome and therefore must be avoided. What’s missing in Brooks’ argument, underpinning the entire conversation about interfaces, are financial considerations and enabling options for the future.

For point one where we assume that changeability is a requirement it means that we have not only good faith but predictive financial assessments dictating that: 1) the system will require future modification and 2) that optimizing for those costs is necessary burden that must be undertaken now.
In the absence of either data point, the so called Majestic Monolith rises as a clear path forward. Though future engineers and managers will likely damn the decision, it will have been made with the best of intentions, fully informed.

For point two we must tackle the notion that the complexity wrought by a multitude of interfaces is, in fact, burdensome. In this decision path the additional complexity is inherent, unavoidable in the requirements, the financial burden is therefore implicit. The decision left on the table is to implement custom protocols or to leverage pre-established protocols to optimize the situation. In some cases the financial burden of custom protocols will be warranted, in others that will be the case (that discussion is beyond the scope of this paper).

In either case though, the complexity of capturing the domain persists and remains the core consideration of interface design, irrespective of the architectural/design choices made through implementation which are, for all intents and purposes, purely financial decisions, by products of requirements. Furthermore, even in the consideration of capturing someone else’s poorly designed API through abstraction is a financial relief in terms of avoiding implementing the captured behavior yourself.

Brooks rightly captured the dimensions of interface specification, though he was hardly the first. Unfortunately he identifies merely the canopy of the conversation missing wholly the financial motivators and underpinnings of interface specification.