Anomaly Detection Via Kpis for Software Performance Failures

As cloud services have become more popular, high-reliability requirements for distributed storage systems have also increased significantly. However, the software systems often suffer performance failures during long-term running. Anomaly detection based on performance indicators is an effective approach for performance failures. To handle the complexity of the distributed databases, this paper proposes a novel anomaly detection methodology via key performance indicators monitored at different levels. Then, a variance inflation factor (VIF) feature selection method is proposed to reduce the multicollinearity problem in the multivariate time series of performance indicators. For a long-running software system, some unavoidable noise will be introduced when monitoring these performance indicators. Therefore, it is necessary to explore the suitable feature processing method to enhance features and eliminate noise interference from indicators. Discrete wavelet analysis has a superior performance in feature extraction in signal analysis. Thus, in this paper, discrete wavelet transform is modified to create an adaptive DWT for noise reduction and feature extraction by taking advantage of the multiresolution analysis. Then, unsupervised anomaly detection algorithms are used to detect the expected indicator outliers. The anomaly detection performance of the proposed method is evaluated in terms of precision, recall, and the F1-score. Finally, this approach is verified through the experiments conducted on Cassandra, and the F1-score is 0.974 which means this approach is effective.