TRACK: Optimizing Artificial Neural Networks for Anomaly Detection in Spark Streaming Systems

Abstract

Due to the growth of Big Data processing technologies and cloud computing services, it is common to have multiple tenants share the same computing resources, which may cause performance anomalies. There is an urgent need for an effective performance anomaly detection method that can be used within the production environment to avoid any late detection of unexpected system failures. To address this challenge, we introduce, TRACK, a new black-box training workload configuration optimization with a neural network driven methodology to identify anomalous performance in an in-memory Big Data Spark streaming platform. The proposed methodology revolves around using Bayesian optimization to find the optimal training dataset size and configuration parameters to train the model efficiently. TRACK is validated on a real Apache Spark streaming system and the results show that the TRACK achieves the highest performance (95% for F-score) and reduces the training time by 80% to efficiently train the proposed anomaly detection model in the in-memory streaming platform.