Overload Management in Data Stream Processing Systems with Latency Guarantees

Stream processing systems are becoming increasingly important to analyse real-time data generated by modern applications such as online social networks. Their main characteristic is to produce a continuous stream of fresh results as new data are being generated at real-time. Resource provisioning of stream processing systems is difficult due to time-varying workload data that induce unknown resource demands over time. Despite the development of scalable stream processing systems, which aim to provision for workload variations, there still exist cases where such systems face transient resource shortages. During overload, there is a lack of resources to process all incoming data in real-time; data accumulate in memory and their processing latency grows uncontrollably compromising the freshness of stream processing results. In this paper, we present a feedback control approach to design a nonlinear discrete-time controller that has no knowledge of the system to be controlled or the workload for the data and is still able to control the average tuple end-to-end latency in a single-node stream processing system. The results, of our evaluation on a prototype stream processing system, show that our method controls the average tuple end-to-end latency despite the time-varying workload demands and increasing number of queries.