Trading Freshness for Performance in a Cluster of Replicated Databases

Abstract

In the Application Service Provider (ASP) model, applications and databases are hosted at the provider site and are accessed by customers, typically through the Internet, who are no longer concerned with data and application maintenance tasks. Through replication of customers’ databases at several nodes, a cluster system can yield high-availability and high-performance at low cost. Our objective is to demonstrate the viability of the ASP model using a cluster system for pharmacy applications in France. In particular, we must support mixed workloads composed of front-o.ce update-intensive transactions (e.g. drug sales) and back-o.ce read-intensive queries (e.g. statistics on drugs sold). Our approach is to capture application semantics for optimizing load balancing within the cluster system. We use a mono-master replication scheme: updating transaction are all sent to a single master node while read-only queries may be sent to any node. Slave nodes are updated asynchronously through refresh transactions and consistency reduces to freshness. Application semantics mainly consist of a freshness level, which express how much the result of a given query may di.er from the expected results if the query was sent to the master node.