WORLDS ’ 05 : Second Workshop on Real , Large Distributed Systems

Joel Wein described Akamai’s Content Distribution Network (CDN) as having 15,000 servers in 1,100 third-party networks, with a NOCC managed by a day crew of eight and a night crew of three. The focus of this paper is not on CDN but on Akamai’s experience in its seven-year experiment: in particular, keeping its distributed system running using Recovery Oriented Computing. In a single day, it is not unusual to lose servers, racks of servers, and even several data centers. The base assumption is that there will be a significant and constantly changing number of component or other failures occurring at all times in the network. The development philosophy is that their software must continue to work seamlessly despite numerous failures.