MuSE Graphs for Flexible Distribution of Event Stream Processing in Networks

Complex event processing (CEP) enables reactive and predictive applications through the continuous evaluation of queries over streamsof event data. In a network of event sources, efficient query evaluation is achieved through distribution: Queries are split into operators(query decomposition), which are then assigned to some of the nodes(operator placement). Yet, existing solutions show two fundamentalissues: They limit the decomposition to the operator hierarchy of aquery, ignoring possible rewritings of it, and place each operator atexactly one node in the network. That neglects potential optimizations related to the composition of patterns through multiple queriesas results are always gathered at a single sink node.In this paper, we propose a new evaluation model for CEP, coinedMulti-Sink Evaluation (MuSE) graphs. It incorporates arbitrary projections of queries for distribution and assigns them to potentiallymany nodes. We prove correctness of query evaluation with MuSEgraphs and provide a cost model to assess its efficiency. Since theconstruction of cost-optimal MuSE graphs is intractable, we presentan approximation algorithm and several pruning strategies. Our evaluation results show that MuSE graphs reduce network transmissioncosts by up to three orders of magnitude over baseline strategies,while an implementation in a state-of-the-art framework for dishu-berlin.detributed computing highlights significant latency improvements.