Serializable snapshot isolation for replicated databases in high-update scenarios

Abstract

Many proposals for managing replicated data use sites running the Snapshot Isolation (SI) concurrency control mechanism, and provide 1-copy SI or something similar, as the global isolation level. This allows good scalability, since only ww -conflicts need to be managed globally. However, 1-copy SI can lead to data corruption and violation of integrity constraints [5]. 1-copy serializability is the global correctness condition that prevents data corruption. We propose a new algorithm Replicated Serializable Snapshot Isolation (RSSI) that uses SI at each site, and combines this with a certification algorithm to guarantee 1-copy serializable global execution. Management of ww -conflicts is similar to what is done in 1-copy SI. But unlike previous designs for 1-copy serializable systems, we do not need to prevent all rw -conflicts among concurrent transactions. We formalize this in a theorem that shows that many rw -conflicts are indeed false positives that do not risk non-serializable behavior. Our proposed RSSI algorithm will only abort a transaction when it detects a well-defined pattern of two consecutive rw -edges in the serialization graph. We have built a prototype that integrates our RSSI with the existing open-source Postgres-R(SI) system. Our performance evaluation shows that there is a worst-case overhead of about 15% for getting full 1-copy serializability as compared to 1-copy SI in a cluster of 8 nodes, with our proposed RSSI clearly outperforming the previous work [6] for update-intensive workloads.