A simple approach to shared storage database servers

Abstract

This paper introduces a generic technique to obtain a shared-storage database cluster from an off-the-shelf database management system, without needing to heavily refactor server software to deal with distributed locking, buffer invalidation, and recovery from partial cluster failure. Instead, the core of the proposal is the combination of a replication protocol and a surprisingly simple modification to the common copy-on-write logical volume management technique: One of the servers is allowed to skip copy-on-write and directly update the original backing store. This makes it possible to use any shared-nothing database server software in a shared or partially shared storage configuration, thus allowing large cluster configurations with a small number of copies of data.