Once a sysadmin, always a sysadmin. I'm not really sure why (although I suspect the folks on alt.sysadmin.recovery would call it a character defect), but even though I'm mostly a programmer this year, visions of well-managed servers are still dancing through my head. Maybe they'll stop someday.

Anyway, I've been stewing over some interesting ideas on server management recently, and I think we've been doing it all wrong. We've been trying to build strong, resilient, flexible servers that we can maintain for years, adding and removing services as needed. This is an outgrowth of the way programmers are trained to think, with maintainability prized as once of the highest virtues of software. The XP people have a slightly different take on things, but underneath it all they still prize maintainable systems. On c2.com's wiki, the discussion of Christoper Alexander's A Pattern Language is informative, as it shows parallels between software design and building design, concentrating on features in buildings that give them a long, useful life.

I'm not sure that we really want any of that for servers, though, or at least not for standard, cookie-cutter services like DNS. We should be aiming for disposable servers, where the only things that we care about are the configuration state, data, performance, and security. The actual system files, and even OS should be irrelevant, and even ignored if possible.

I'm working on a demonstration, along with a design for a larger-scale system that have some bizarrely nice properties. I don't see why it wouldn't work, frankly, and it'll make a huge change in the amount of work needed to implement small services in networks.