Adaptive and Efficient Event Stream Processing through Relaxed Semantics

Licentiate thesis, 2026

Stream processing has become a fundamental data processing mechanism in modern edge-to-cloud systems, enabling the transformation of continuous data streams into timely aggregated results and actionable insights, e.g., in traffic or energy consumption monitoring. In practice, deploying and executing streaming applications, referred to as continuous queries, is often challenged by fluctuating workloads, changing resource availability, and varying application requirements. For example, a traffic monitoring system running on a roadside device might observe a surge in events reported by cars and, as a result, might need to produce outputs more frequently to support timely control decisions, despite limited computational power. This necessitates that Stream Processing Engines (SPEs) -- software systems for executing continuous queries on unbounded data streams -- dynamically adjust their behavior at runtime rather than relying on fixed configurations. Yet existing SPEs offer limited support for such adaptive capabilities, and even core operations such as data aggregation are no exception.

This thesis addresses this limitation by exploring how stream Aggregates -- stateful operators that summarize streaming data -- can be relaxed at runtime to adaptively trade performance against resource consumption and output guarantees. On the one hand, stream Aggregate states can vary significantly and shift over time in response to variations in the incoming data, and not all states are accessed with equal frequency; therefore, those states updated infrequently could be compressed. However, deciding when and what to compress is not trivial, as the optimal trade-off between memory savings and processing overhead varies with workload. We thus design an on-demand compression mechanism that uses Reinforcement Learning (RL) to dynamically tune Aggregate state compression levels at runtime, balancing memory consumption and processing efficiency under a target latency threshold, and identify policies that balance feedback timeliness and learned quality. On the other hand, Aggregates are defined by two key parameters -- window advance (the output frequency) and window size (the interval length) -- which typically remain fixed throughout execution, limiting their ability to adapt to changing workloads and resource conditions. Therefore, we support dynamic reconfiguration of both parameters at runtime, allowing analysts to specify a set of acceptable advances and sizes to trade performance and guarantees via weaker (at-most-once) or stronger (exactly-once) reconfiguration semantics. We provide alternative reconfiguration strategies for both in-order and out-of-order tuple arrival models, and empirically show that the approach matches fixed-configuration baselines' performance while supporting reconfiguration in negligible time.