Survivable consensus objects

Abstract

Reaching consensus among multiple processes in a distributed system is fundamental to coordinating distributed actions. We present a new approach to building survivable consensus objects in a system consisting of a (possibly large) collection of persistent object servers and a transient population of clients. Our consensus object implementation requires minimal support from servers, but at the same time enables clients to reach coordinated decisions despite the arbitrary (Byzantine) failure of any number of clients and up to a threshold number of servers.