Improving the Scalability of Cloud-Based Resilient Database Servers

摘要

Many rely now on public cloud infrastructure-as-a-service for database servers, mainly, by pushing the limits of existing pooling and replication software to operate large shared-nothing virtual server clusters. Yet, it is unclear whether this is still the best architectural choice, namely, when cloud infrastructure provides seamless virtual shared storage and bills clients on actual disk usage. This paper addresses this challenge with Resilient Asynchronous Commit (RAsC), an improvement to a well-known shared-nothing design based on the assumption that a much larger number of servers is required for scale than for resilience. Then we compare this proposal to other database server architectures using an analytical model focused on peak throughput and conclude that it provides the best performance/cost trade-off while at the same time addressing a wide range of fault scenarios.