Oversubscription is an essential cost management strategy in multi-tenant, cloud Database-as-a-Service (DBaaS), and its importance is magnified by the emergence of serverless databases. In the FRA project, we have developed novel resource management techniques that enables cloud DBaaS providers to oversubscribe resources in DBaaS clusters while controlling the impact on performance and availability. The controls are realized at different levels in the cluster infrastructure: across cores within a single node, across nodes within a cluster, and across clusters within a data center.
The concrete techniques proposed in this context were developed and validated in the context of a commercial service: Microsoft’s Azure SQL Database. The project is a collaboration between Microsoft Research, the Azure SQL DB team, and has benefited from the contributions of multiple interns. A paper providing an overview of the project and key flexible resource allocation mechanisms and policies appears in VLDB 2024 (opens in new tab) (opens in new tab). Additional publications covering specific aspects of the projects can be found in the Publications tab.