An available copy protocol tolerating network partitions

Maintaining in a consistent state multiple copies of the same data is a complex task, especially when the copies reside on sites that can be separated from each other by network partitions. All existing replication controls that tolerate network partitions use quorums to provide mutual exclusion and prevent inconsistent updates. Unfortunately these protocols require a minimum of n + 2 voting sites to guarantee that the data will remain accessible in the presence of n site failures. As a result, they provide much lower data availabilities than protocols that exclude communication failures. We present here a replication protocol that extends the available copy approach to environments where communication failures may cause network partitions. Our protocol assumes that replicas monitor the communication paths linking them with their peers and can therefore detect network partitions. As a result, each individual replica can safely establish whether it has remained up to date or is likely to have missed some updates. We evaluate under standard Markovian assumptions the availability of a replicated file consisting of two replicas managed by our robust available copy protocol when the two replicas are separated by one gateway. We find our protocol to perform much better than (1) dynamic-linear voting with three replicas, (2) dynamic-linear voting with two replicas and one witness, (3) voting with ghosts with two replicas, and (4) voting with bystanders with two replicas.