Amazon’s just issued their report on last week’s disaster. The report’s pretty dense with turgid geekspeak, but the upshot seems to be that a mistake reconfiguring a network router inside the data center led to a cascading series of failures that took hours to unfold. Given that we’re talking silicon working at gigahertz speed, it must have been an event of staggering complexity.
My interpretation of it, overall, is that it was in fact a complexity disaster, a system that had grown up over several years to a gigantic scale, with an intricate and highly-interconnected infrastructure. Where historically you think of a data center as a loosely-coupled collection of discrete computers connected by discrete wires, a cloud is a large cluster of tightly-coupled specialized nodes, some nodes providing storage, some providing generalized (Linux) or specialized (database, email) logic. And every node is a member of multiple networks providing different kinds of connections to different groups of other entities inthe cloud.
I’m tempted to call the cloud a “brain” because of that neural-like structure, but that’s a little over the top, despite its digital componentry. But it is clearly behaving like an organism. The report described several distinct phases of the “event”, which sound like the organism going through massive internal shifts to re-achieve homeostatis, things like “remirrong storms” and “races” as large numbers of nodes shifted from one error-recovery strategy to another, like flocks of fishes or birds suddenly veering this way or that in perfect unison. The report talked of the humans taking actions to influence the behavior of the cloud, but it’s telling that they couldn’t simply type in a few commands to direct the cloud to do what they wanted. They made small changes, watched what happened, then poked it again. The treated it as a living thing that couldn’t just be shut down and rebooted.
The report described a lot of what I interpet as emergent behavior, and I think that it’s taken the tightly-interconnected nature of the cloud to produce it. I’ve thought in a vague way for years that a true digital intelligence can’t be constructed, but can only emerge from a sufficiently complex digital milieu, and this event provides a glimpse of forces at work in a cloud that might just satisfy the preconditions for emergence of intelligence. Someday.