Scaling and Continuous Availability in Database Server Clusters through Multiversion Replication

In this paper, we study replication techniques for scaling and continuous operation for a dynamic content server. Our focus is on supporting transparent and fast reconfiguration of its database tier in case of overload or failures. We show that the data persistence aspects can be decoupled from reconfiguation of the database CPU. A lightweight in-memory middleware tier captures the typically heavyweight read-only requests to ensure flexible database CPU scaling and fail-over. At the same time, updates are handled by an on-disk database back-end that is in charge of making them persistent. Our measurements show instantaneous, seamless reconfiguration in the case of single node failures within the flexible in-memory tier for a web site running the most common, shopping, workload mix of the industry-standard e- commerce TPC-W benchmark. At the same time, a 9-node in-memory tier improves performance during normal operation over a stand-alone InnoDB on-disk database back- end. Throughput scales by factors of 14.6, 17.6 and 6.5 for the browsing, shopping and ordering mixes of the TPC-W benchmark, respectively.