Optimizing voting-type algorithms for replicated data

The main objectives of data replication are improved availability and reduced communications cost for queries. Maintaining the various copies consistent, however, increases the communications cost incurred by updates. For a given degree of replication, the choice of a specific concurrency control algorithm can have a significant impact on the total communications cost. In this paper we present various models for analyzing and understanding the trade-offs between the potentially opposing objectives of maximum resiliency and minimum communications cost in the context of the quorum consensus class of algorithms. It is argued that an optimal vote assignment is one which meets given resiliency goals and yet incurs the least communications cost compared with all other alternative assignments. A mathematical model for vote assignment is developed, and optimal algorithms are presented. It is demonstrated that significant cost savings can be realized from these approaches.