Exploring System and Machine Learning Performance Interactions when Tuning Distributed Data Stream Applications

Abstract

Deploying machine learning (ML) applications over distributed stream processing engines (DSPEs) such as Apache Spark Streaming is a complex procedure that requires extensive tuning along two dimensions. First, DSPEs have a vast array of system configuration parameters (such as degree of parallelism, memory buffer sizes, etc.) that need to be optimized to achieve the desired levels of latency and/or throughput. Second, each ML model has its own set of hyper-parameters that need to be tuned as they significantly impact the overall prediction accuracy of the trained model. These two forms of tuning have been studied extensively in the literature but only in isolation from each other. This position paper identifies the necessity for a combined system and ML model tuning approach based on a thorough experimental study. In particular, experimental results have revealed unexpected and complex interactions between the choices of system configuration and hyper-parameters, and their impact on both application and model performance. These findings open up new research directions in the field of self-managing stream processing systems.