Location and replication independent recovery in a highly available database

Abstract

An increasing number of database applications demand high availability combined with scalable throughput and load balancing. We present a logging and recovery method which meets these goals by utilizing the log as a replication means from primaries to hot standbys. To be both scalable and highly available, recovery is performed independently of the location and replica of the original operations. The recovery method is designed to allow for load balancing in all situations of a highly available database, during normal processing, takeover and online self-repair. Logical tuple access is used both for redo and undo, and together with state identifiers connected to tuples this is the basis for the replication and location independent recovery. Unlike existing replication methods, this logging and recovery method allows for scalable replication by facilitating asynchronous parallel streams of log records from primaries to hot standbys. To provide high availability we also need fast takeover and takeback.