Anomaly Detection for Time Series Infrastructure Metric Data

Abstract

Time series metric data is used by monitoring systems to analyze the performance of hardware and software from a data center. This data can be easily displayed using line graphs and, as such, can be manually checked by operators for anomalies. Manual reviews have some major disadvantages, since they are time consuming and error prone. In this paper, we describe our approach for automatically detecting outliers in metric data, which have some unique properties like high seasonality and extreme seasonal spikes. We use the "Prophet" library to create a model of the time series data, which is then automatically analyzed with our own distance formula for detecting anomalies in the time series data. A part of our metrics data have a propensity for extreme recurrent spikes, for example network usage during daily backups that is millions of times higher than the average. To handle such special cases we have also implemented a seasonal spike detector.