Predicting Failures in Autoscaling Distributed Applications

This paper is a research article that presents Preface, a novel approach designed to predict failures in cloud applications utilizing autoscaling. The objective of Preface is to enhance failure prediction capabilities in production environments, particularly for microservice-based distributed applications that operate on dynamically configured resources. Traditional failure prediction techniques struggle with the variability of key performance indicators (KPIs) in autoscaling scenarios. Preface addresses this limitation by integrating a Rectifier layer that processes KPI sets of varying sizes, generating a fixed set of descriptive statistics. This enables the neural network-based predictor to identify anomalies indicative of potential failures. The authors validate Preface through experiments on two applications, Alemira and TrainTicket, demonstrating its effectiveness in predicting failures with a success rate ranging from 41% to 99%. The results indicate that Preface can timely activate countermeasures to mitigate the impact of failures on users, thereby contributing to advancements in software reliability and engineering.