State Compensation: A No-cost Scheme for Scalable Failure Recovery in Tree-based Overlay Networks

Tree-based overlay networks (TBŌNs) have become important for scalable data multicast and aggregation. This infrastructure’s generality has lead to widespread usage in large scale and widely distributed environments – environments in which reliability must be addressed. This paper presents state compensation, a novel reliability concept for TBŌN environments that avoids explicit state replication (such as checkpoints) for failure recovery by leveraging general properties of TBŌN computations that allow computational state from non-failed processes to compensate for state lost from failed ones. In this paper, we present our state compensation mechanisms, prove sufficient properties of distributed computations that make these mechanisms feasible and show how to derive computation-specific recovery primitives from these properties. We also present a case study of the recovery process. The result is a general TBŌN recovery model that requires no additional storage, network, or computational resources during normal operation.