Scatter-gather based approach in scaling complex event processing systems for stateful operators

With the introduction of Internet of Things (IoT), scalable Complex Event Processing (CEP) and stream processing on memory, CPU, and bandwidth constraint infrastructure have become essential. While several related work focuses on replication of CEP engines to enhance scalability, they do not provide expected performance while scaling stateful queries for event streams that do not have predefined partitions. Most of the CEP systems provide scalability for stateless queries or for the stateful queries where the event streams can be partitioned based on one or more event attributes. These systems can only scale up to the pre-defined number of partitions, limiting the number of events they can process. Meanwhile, some CEP systems do not support cloud-native and microservices features such as startup time in milliseconds. In this research, we address the scalability of CEP systems for stateful operators such as windows, joins, and pattern by scaling data processing nodes and connecting them as a directed acyclic graph. This enabled us to scale the processing and working memory using the scatter and gather based approach. We tested the proposed technique by implementing it using a set of Siddhi CEP engines running on Docker containers managed by Kubernetes container orchestration system. The tests were carried out for a fixed data rate, on uniform capacity nodes, to understand the processing capacity of the deployment. As we scale the nodes, for all cases, the proposed system was able to scale almost linearly while producing zero errors for patterns, 0.1% for windows, and 6.6% for joins, respectively. By reordering events the error rate of window and join queries was reduced to 0.03% and 1% while introducing 54ms and 260ms of delays, respectively