Real-time data stream performance improvement with reinforcement learning

Abstract

Real-time data streaming has become a fundamental aspect of modern datacentric applications, including a wide range of areas such as financial analytics and the Internet of Things (IoT). As real-time data streams become integral to these applications, optimizing their performance is paramount. This thesis aims to improve the efficiency of real-time data streams by employing a reinforcement learning (RL) technique to reduce the latency, especially in the widely used real-time data streaming platform, Kafka. Using reinforcement learning (RL), the system adjusts Kafka settings in real time to minimise latency while remaining within the resource limitation. The main breakthrough comes in the RL agent's capacity to acquire knowledge from previous encounters and input to make intelligent, immediate choices on parameter modifications. The agent engages with the environment, overseeing system metrics and latency feedback, and constantly adjusts configuration parameters to get optimal performance. This adaptive approach guarantees that the data stream remains highly efficient and responsive, even when faced with fluctuating workload intensities and resource constraints. The efficiency of the RL-based technique in improving real-time data stream performance across various workloads and circumstances is demonstrated through experimental assessments. The study examines the compromises between reducing latency and optimising throughput, thoroughly examining how varying levels of workload intensity and resource limitations affect performance. The results of this study provide a substantial contribution to the field of real-time data stream processing. This study utilises reinforcement learning techniques to provide a data-driven and adaptable solution that improves the efficiency and responsiveness of Kafka-based streaming applications in many scenarios. This methodology not only enhances our comprehension of optimising data streams in real time but also offers valuable guidance for implementing reinforcement learning in dynamic contexts with limited resources.