SaSPartitioner: A Self-Adaptive Streaming Partitioner Using Deep Reinforcement Learning

Stream processing has been widely adopted in online services, such as e-commerce platforms and real-time query systems, due to its ability to leverage distributed parallel computing for low-latency, large-scale data processing. However, data skew, which results in imbalanced load distribution across computational nodes, can significantly increase processing latency and degrade resource efficiency. Most existing stream processing systems adopt partitioning strategies based on data volume, estimating system load using the volume of data in each partition. As a result, they often struggle to adapt promptly to dynamic changes in data distribution caused by streaming workloads. To this end, we propose SaSPartitioner, a self-adaptive partitioning framework based on system-metric modeling that leverages deep reinforcement learning (DRL) and real-time system metrics. Specifically, we formulate the streaming partitioning problem as a Markov decision process (MDP), collecting real-time metrics such as CPU and memory usage of streaming operators. Besides, we design both offline and online reinforcement learning models to enable seamless adaptation to dynamic shifts in data distribution. To further enhance efficiency, we introduce two acceleration methods, i.e., partition masking and hot key selection, to prune the state and action space of the DRL model. Extensive experiments on Apache Flink show that SaSPartitioner significantly outperforms existing stream partitioning methods, achieving up to a $44 \times$ increase in throughput while maintaining robustness under rapidly changing data distributions.