Reconfiguration Primitives for Self-Adapting Overlays in Distributed Publish-Subscribe Systems

Most distributed applications involve some form of event-based interaction, often implemented using a publish-subscribe (pub-sub) infrastructure. To improve scalability, the acts of matching events against subscriptions and delivery them are performed collaboratively by a set of brokers connected into an overlay network. Recent research has proposed several approaches to support the self-adaptation of such overlay network to adapt it to changes in application traffic. However these approaches focus on the monitor, analyze, plan parts of the self-adaptation loop, without considering the issues that arise in the execution part. This paper proposes a set of primitives that fills the gap in the execution phase. Compared to existing work, our approach: (i) is transparent w.r.t. the routing policies of the middleware, (ii) preserves existing properties and guarantees of the middleware, such as no duplication of events, causal ordering, and minimal delays for the events delivered during a reconfiguration. We discuss the correctness of our primitives and implement them in a simulated environment to measure their cost in terms of network overhead.