Conflict classes for replicated databases: a case-study

Abstract

The major challenge in fault-tolerant replicated transactional databases is providing efficient distributed concurrency control that allows non-conflicting transactions to execute concurrently. A common approach is to partition the data according to the data access patterns of the workload, assuming that this will allow operations in each partition to be scheduled independently and run in parallel. The effectiveness of this approach hinges on the characteristics of the workload: (i) the ability to identify such partitions and (ii) the actual number of such partitions that arises. Performance results that have been presented to support such proposals are thus tightly linked to the simplistic synthetic benchmarks that have been used. This is worrisome, since these benchmarks have not been conceived for this purpose and the resulting definition of partitions might not be representative of real applications. In this paper we contrast a more complex synthetic benchmark (TPC-E) with a real application in the same area (financial brokerage), concluding that the real setting makes it much harder to determine a correct partition of the data and that sub-optimal partitioning severely constrains the performance of replication.