Proactive adaptation of microservice-based applications

Abstract

Proactive auto-scaling of Microservice-based Applications has become popular in industry and academia. Proactive systems analyse historical data patterns to estimate future trends, assuming they will occur again. Early detection of potential problems, like high latency, enables prompt action, including service replication, to fix the issues before they arise. Several studies propose proactive auto-scaling systems for microservices. However, they have design limitations in their forecasting systems that may negatively impact forecast runtime accuracy. For example, all these systems rely on a single forecasting model for the prediction task. Using a single forecasting model increases the risk of inaccurate estimates, leading to unsuitable interventions that could harm the customer experience. This work presents PMA (Proactive Microservices Auto-scaler), a MAPE-K-based auto-scaling system that uses forecasting models to anticipate and avoid microservices performance issues. PMA offers three models to address existent design limitations: univariate, multivariate and a Multiple Predictor Systems strategy that uses multiple models for prediction. Several experiments were performed to evaluate PMA and compare its performance to Predict Kube (PK), a leading adaptive industry tool. In 93.75% of the experiments, PMA outperformed PK for managing the applications. This work aims to improve proactive microservices auto-scaling systems, addressing some of their current design limitations to develop a more accurate and reliable forecasting system.