Protecting Replicated Objects Against Media Failures

We present a replication control protocol that provides excellent data availabilities while guaranteeing that all writes to the object are recorded in at least two replicas. The protocol, robust dynamic voting (RDV) accepts reads and writes as long as at least two replicas remain available. The replicated object remains inaccessible until either the two last available replicas recover or one of the two last available replicas can collect the votes of a majority of replicas. We evaluate the read and write availabilities of replicated data objects managed by the RDV protocol and compare them with those of replicated objects managed by majority consensus voting, dynamic voting and hybrid dynamic voting protocols. We show that RDV can provide extra protection against media failures with no siginificant loss of availability.