The cost of increased transactional correctness and durability in distributed databases

In this paper we investigate the problem of providing durability for Web Service transactions in the presence of system failures. We show that the popular lazy replica update propagation method is vulnerable to loss of transactional updates in the presence of hardware failures. We propose an extension to the lazy update propagation approach to reduce the risk of data loss. Our approach is based on the “buddy” system, requiring that updates are preserved synchronously in two replicas, called buddies. The rest of the replicas are updated using lazy update propagation protocols. Our method provides a balance between durability (i.e., effects of the transaction are preserved even if the server, executing the transaction, crashes before the update can be propagated to the other replicas) and efficiency (i.e., our approach requires a synchronous update between two replicas only, adding a minimal overhead to the lazy replication protocol). Moreover, we show that our method of selecting the buddies ensures correct execution and can be easily extended to balance workload, and reduce latency observable by the client.