Read-only transactions in partitioned replicated databases

Abstract

Environments where read-only transactions predominate update transactions are considered. The notion of a group in replicated databases, the recognition of which maximizes the data availability for read-only transactions in error-prone environments, is developed. It is shown that the group identification problem is NP-complete. Replica control algorithms that attempt to approximately determine groups are presented. In one algorithm, groups are determined dynamically; in the other techniques, groups are predefined. It is shown formally that both approaches yield algorithms that preserve the consistency of the database. In the environments considered, the approaches presented are shown to yield significant improvements over other fault-tolerant replica control algorithms.<>