Prepare to Pick Two



Prepare to Pick Two

Bill de hÓra

SoMETiMES ACCEpTing A ConSTRAinT or giving up on a property can lead to a better architecture, one that is easier and less expensive to build and run. Like buses, desirable properties tend to come in threes, and trying to define and build a system that supports all three can result in a system that does noth- ing especially well.

A famous example is Brewer’s conjecture, also known as Consistency, Avail- ability, and Partitioning (CAP), which states that there are three properties that are commonly desired in a distributed system—consistency, availability, and partition tolerance—and that it is impossible to achieve all three. Trying to have all three will drastically increase the engineering costs and typically increase complexity without actually achieving the desired effect or business goal. If your data must be available and distributed, achieving consistency becomes increasingly expensive and eventually impossible. Likewise, if the system must be distributed and consistent, ensuring consistency will lead at first to latency and performance problems and eventually to unavailability since the system cannot be exposed as it tries to reaches agreement.

It’s often the case that one or more properties are considered inviolate: data cannot be duplicated, all writes must be transactional, the system must be



100% available, calls must be asynchronous, there must be no single point of failure, everything must be extensible, and so on. Apart from being naïve, treating properties as religious artifacts will stop you from thinking about the problem at hand. We start to talk about architectural deviation instead of prin- cipled design and we confuse dogmatism with good governance. Instead we should ask, why must these properties hold? What benefit is to be had by doing so? When are these properties desirable? How can we break up the system to achieve a better result? Be ever the skeptic, because architectural dogma tends to undermine delivery. The inevitability of such tradeoffs is one of the most difficult things to accept in software development, not just as architects, but also as developers and stakeholders. But we should cherish them; it’s far better than having limitless choice, and accepting tradeoffs often induces a creative and inventive result.

Bill de hÓra is chief architect with NewBay Software, where he works on large scale web and mobile systems. He is co-editor of the Atom Publishing Protocol and previously served on the W3C RDF Working Group. He is a recognized expert on REST style and message-passing architectures and protocol design.