Uninterruptible Migration of Continuous Queries without Operator State Migration

Abstract

The elasticity brought by cloud infrastructure provides a promising solution for a data stream management system to handle its incoming workload, which can be highly variable: the system can scale out when heavily loaded, and scale in otherwise. In such a solution, the efficiency of the mechanism used to migrate a query from one node to another is very important. Generally, a stream application requires real-time outputs for its continuous queries, and downtime is not acceptable. Moreover, the migration should not add considerable processing cost to a node that could have been already overloaded. In this paper, we present our migration protocol, named UniMiCo, which satisfies those requirements. We implemented UniMiCo in a DSMS prototype and experimentally show that the protocol preserves correctness, while introducing no noticeable changes in the response time of the continuous query being migrated.