Document Type

Technical Report

Publication Date


Technical Report Number



Replication of data at more than one site in a distributed database has been reported to increase the availability in data in systems where sites and links are subject to failure. We have shown in results summarized in this paper that in many interesting cases the advantage is slight. A well-placed single copy is available to transactions almost as much of the time as is correct replicated data no matter how ingeniously it is managed. We explain these findings in terms of the behavior of the partitions that form in networks where components fail. We also show that known and rather simple protocols for the maintenance of multiple copies are essentially best possible by comparing them against an unrealizable "protocol" that knows the future. We complete our study of these questions by reporting that while computing the availability of data is #P-complete, nonetheless there is a tight analytical bound on the amount replication can improve over a well-located single copy. We close with some observations regarding system design motivated by this work.