2005/12/27

Platform standards and loose coupling

Recently people have been talking about organizational standards for web application platforms (like Linux/Apache/Tomcat for example, or ASP.net for another).  Personally, I'm a big fan of the "small pieces loosely joined" concept.  Smallpieces are exponentially easier to build, test, deploy, and upgrade. Loose coupling gives flexibility and risk mitigation -- components canfail or be replaced without major impacts to the entire structure.  Allof these things help us cope with schedule and product risks.  Thetechnical tradeoff is a performance (latency) hit; for webapplications, I think the industry has proven that this is usually agood tradeoff.

I guess I should be clear here that I'm interested optimizing for effectiveness, notefficiency.  By effectiveness I mean that speed of development, qualityof service, time to market, flexibility in the face of changingbusiness conditions, and ability to adapt in general are much moreimportant than overall number of lines of code produced or even averagefunction points per month.  That is, a function delivered next week isoften far more valuable than ten delivered six months from now.

To do this, you need to start with the organization: Architecture reflects the organization that produces it.  So, first you need to create an organization of loosely coupled small pieces,with a very few well chosen defining principles that let theorganization work effectively.

Each part of the organization should decide on things like the application server platform they should use individually.  They're the ones with the expertise, and if there really is a best answer for a given situation they should be looking for it.  On the other hand, if there's no clear answer and if there's a critical mass of experience with one platform, that one will end up being the default option.  Which is just what we want; there's no top-down control needed here.


So the only case where there's an actual need for top-down organizational platform standards is to get acritical mass of people doing the same thing, where the benefit accruesmostly because of the critical mass, not because of individual projectbenefits.  There's not much benefit to bulk orders of Apache/Tomcat, soif you're avoiding vendor lock-in the main reason to do thisis toenable interoperation.  But that can be accomplished by picking openstandards and protocols -- pick some basic, simple, straightforwardstandards, make sure that teams know about them and are applying themwhere appropriate, and they'll be able to talk together.  This is a risk mitigation strategy rather than an optimization strategy; in other words, you know you can always get something working with a known amount of effort using the loosely coupled strategy.  When you're tightly coupled to anything, this is no longer true -- you inherit its risks.  Tightly coupling an entire organization to an application server platform also creates a monoculture, making some things very efficient but also increasing the risk that you'll be less able to adapt to new environments.