Weak consistency and pessimistic replica control

Abstract

In distributed systems replication is used to enhance availability and performance. Concurrent access to copies on different sites must be synchronized so transactions remain serializable. The main difficulty is the possibility of a partition of the network due to site or communication failures. Several protocols have been designed to synchronize transactions running in different components. Most pessimistic algorithms restrict access to a unique component per object and impose mutual consistency of copies. In this paper we show that this is not necessary for pessimistic control. We present a forwarding strategy for missing updates and a method to globally order conflicting transactions in a partitioned system. This enables consistent views of objects in minority components, logical conflicts in different components and one-copy serializability. We present an algorithm based on these ideas which achieves higher availability than other pessimistic protocols. This leads us to define a new concept for control as opposed to the traditional adaptation/recovery paradigm for replica control.