AIOPS FRAMEWORK FOR ALERTING PERFORMANCE ISSUES IN MICROSERVICES USING TIME SERIES FORECASTING

Abstract

For large scale applications based on Microservices architecture and hosted on scalable platforms like Pivotal Cloud Foundry or OpenShift, it is imperative to monitor the performance of microservices on a continuous basis and trigger performance critical alerts upfront when abnormal patterns are observed in the immediate past. One of the critical application performance indicators is Major Garbage Collection (Major GC), an increase of which causes application performance bottlenecks and might indicate insufficient memory allocation to the JVM or a potential memory leak. In this paper, we discuss a solution built using machine learning Algorithms that forecasts Major GC cycles (Number of Major Garbage Collection) of business critical microservices for the near future time periods. This paper discusses various algorithms within python’s stats models. tsa package that were explored and details the concepts and results of forecast models trained using Seasonal Autoregressive Integrated Moving Average, or SARIMA algorithm that was finally considered for forecasting based on model performance. The paper also elucidates the Forecast accuracy of the model, showing results of the forecast vs actual occurrences of Major Garbage collection. The objective of this solution based on AIOps framework is to enable development and production support teams get proactive notification when forecast Major GC patterns breach the set threshold limits, so that corrective action can be taken to improve the application performance by reducing frequency of Major GC cycles.