Logging and recovery in PRISMA

Abstract

The recovery methods for main-memory database systems are mostly based on maintaining a disk based log and checkpoint. The information written to the log usually consists of low level information, before-after images or single tuple updates. In this paper we examine the effect of writing higher level update information to the log. This results in less logging overhead during normal processing, but an increased replay overhead during recovery. We introduce a cost model for analyzing the effect of higher level logging on the transaction throughput, and present the results of a performance analysis based on this cost model.