Diverse Database Replication Based on Snapshot Isolation – Performance Implications of Improved Dependability

Numerous database replication schemes are built on the crash failure assumption where majority of failures are self-evident as defined in [1]. The study in [1] convincingly refuted this common assumption showing that many of the faults in relational Database Management Systems (DBMSs) cause systematic non-crash failures. Similar results were obtained in the subsequent study [2]. Consequently, the existing database replication solutions, which typically use the same DBMS, are ineffective fault-tolerant mechanisms. Conversely, using diverse DBMSs is a suitable way of protecting against non-crash failures. We have built a middleware-based database replication protocol, DivRep, and deployed it with diverse database servers (DivSQL), for improved fault tolerance. DivSQL provides strict Snapshot Isolation (SI) guarantees, and assumes “incorrect results” failure model (IRFM) – the most realistic one based on the extensive experimental analyses of DBMS faults ([1], [2]). The dependability gain comes with the inherent performance overhead. We provide a comprehensive performance evaluation of DivSQL using 3 diverse DBMSs (two are leaders in the field).