Leveraging Time Series Autocorrelation Through Numerical Differentiation for Improving Failure Prediction

Given the complexity of modern software systems, it is no longer possible to detect every fault before deployment. Such faults can eventually lead to failures at runtime, compromising the business process and causing significant risk or losses. Online Failure Prediction (OFP) is a complementary fault-tolerance technique that tries to predict failures in the near future, by using past data and the current state of the system. However, modern systems are comprised of many components and thus a proper characterization of its state requires hundreds of system metrics. As the system evolves through time, these data can be seen as multivariate time series, where the value of a system metric at a given time is related to its previous value. Although various techniques exist for leveraging this autocorrelation, they are often either simplistic (e.g., sliding-window), or too complex (e.g., Long-Short Term Memory (LSTM)). In this paper we propose the use of numerical differentiation, computing the first and second derivative, as a means to extract information concerning the underlying function of each system metric to support the development of predictive models for OFP. We conduct a comprehensive case using a Linux failure dataset that was generated through fault injection. Results suggest that numerical differentiation can be a promising approach to improve the performance of Machine Learning (ML) models for dependability-related problems with similar sequential characteristics (e.g., intrusion detection).