So, I attended this seminar recently, in which the usual suspects were paraded in front of a saucer-eyed audience; There was Amazon, Salesforce, local hopes and established players in equal measures. The first presentation was by a Swedish guy named Joakim Lindbom, CTO at Cap Gemini in Sweden - a really sharp guy, by the way.
His talk was about Amazon EC2, and I had heard a lot of it before, but not everything. Joakim focused on the architectural choices that Amazon had made when creating their different services, being an architect himself. He argued successfully the common knowledge that it really isn't possible to scale a relational database to do what Amazon have to do, and that the same apply to services.
In general a cloud-based architecture means that you use a distributed queueing service both to store incoming requests and to communicate between system parts. Yes, a kind of ESB, if you will, but simple!
It also means using schemaless databases that don't mind adding another column even though there's already 1.0E9 items in the table.
And then it struck me; If it is absolutely critical to use ___SIMPLE___ (meaning non-complex, OK?) services which are both asynchronous (queued, event-driven) and dynamically or non-typed, to be able to build a system over a certain size - why is this not critical when building smaller system?
Or, to phrase the questions differently; How comes it is possible to build systems that does not comply with these traits, since they have been proven to be just critical? Size, is the answer of course.
But to expand that answer is to get to the core of the problem; Smaller sized systems can skimp on doing things right, because it doesn't show up until they get larger. And what small system doesn't?
So, essentially, what has been discovered to be critical in building large-scaled 'cloud' systems is just as essential in building smaller system, it just isn't as evident. Which of course can be translated as saying that it isn't critical. And, no, sure, it isn't critical. But what I'm saying is that even smaller system will by necessity cost less to build, if built according to the same principles the larger-scale system use, if nothing else but the simple fact that they will grow over time.
One of the successes of Erlang has been the reational approach to distributed proecessing; Since interprocess communication in Erlang is message-based, putting those messages on a bus á la Amazon's simple queue service, will be transparent to the processes using it, and also, those very processes need not know where they are and can so be safely dotted out in the cloud somewhere.
No wait, nobody has to, right. It usable that way right off the bat.
Have a nice weekend all y'all.