FFWD : performance-aware event stream processing via domain-specific load-shedding policies

Tools and applications for event stream processing and real-time analytics are getting a huge hype these days [1], due to their ability to analyze massive amount of data in short time, extracting information useful for a wide range of application scenarios, from the smallest IoT embedded sensor to the most popular Social Network feed. Of course, real-time analysis is key to obtain useful insights; however, streaming sys- tems expose only limited computational capabilities that have to face with a theoretically infinite amount of data. The intrinsic variability of streaming flows can lead those applications to overloading conditions, characterized by system saturation and uncontrolled loss of events. To avoid this critical condition, streaming application should monitor their QoS metrics (e.g. latency, performance, resource utilization), possibly self-adapting to cope with sudden input spikes. In this thesis work we propose FFWD, a framework for the runtime load adaptation of streaming application with QoS constraints and output quality requirements. FFWD tackles this problem by means of a general and customizable approach, relying at the same time on a solid abstraction of the streaming system and on the unique domain- specific details of the case study at hand. The framework leverages a Load Manager able to observe the system status and to achieve the QoS requirement; a Policy Wrapper, which hosts a domain-specific policy able to derive the quantity of load to be shed; a Shedding Plan, which hosts the shedding probabilities computed by the policy; and a LS Filter, which discards events directly on the input stream. We evaluated our approach in two case studies: (1) real-time sentiment analysis with latency constraints and (2) system and application monitoring with CPU usage con- straints. In both cases we show how our approach is able to maintain limited the QoS metric while achieving good results in output estimation.