Sharing data and work across queries in analytical workloads

Traditionally, query execution engines in relational databases have followed a query-centric model: They optimize and execute each incoming query using a separate execution plan, independent of other concurrent queries. For workloads with low contention for resources, or workloads with short-lived queries, this model makes the optimization phase faster and creates efficient execution plans. For workloads with heavy contention, or workloads with long-running analytical queries, this model cannot exploit the sharing opportunities that might exist among concurrent queries in order to save I/O, CPU and RAM resources. We argue that exploiting these sharing opportunities is a crucial step towards handling these increasingly common workloads. In this paper, we study three research prototype systems that employ various methodologies for sharing data and work: (a) The QPipe query execution engine [1], which employs a circular scan per table and shares work through simultaneous pipelining, (b) the DataPath system [2], which employs an uninterrupted linear scan per disk and shares work through a global query plan, and (c) the SharedDB system [3], which employs a circular scan per table partition, shares work through a global query plan, uses batched execution, and services both OLTP and OLAP workloads under response time guarantees. We classify these methodologies, analyze their commonalities and differences, and identify their strengths and shortcomings.