Post-partition reconciliation protocols for maintaining consistency

Abstract

This paper addresses design exploration for protocols that are employed in systems with availability-consistency tradeoffs. Distributed data is modelled as states of objects replicated across a network, and whose updates require satisfaction of integrity constraints over multiple objects. Upon detection of a partition, such a network will continue to provide delivery of services in parallel partitions; but only for updates with non-critical integrity constraints. Once the degraded mode ends, the parallel network partitions are reconciled to arrive at one partition.Using a formal treatment of the reconciliation process, three algorithms are proposed and studied in terms of their influence on service outage duration. The longer the reconciliation time, the lower is system availability; since the interval in which no services are provided is longer. However, the reconciliation time in turn is affected by the time to construct the post-partition system state. The shorter the construction time the higher is the number of updates that took place in the degraded mode but that will not be taken up in the reconciled partition. This will lead to a longer interval for rejecting/redoing these operations and thereby increase reconciliation time.