Leveraging machine learning and feature engineering for optimal data-driven scaling decision in serverless computing

Abstract

Serverless computing offers scalability and cost-efficiency, but balancing performance and cost remains challenging, particularly in scaling decisions that can lead to cold starts or resource misallocation. This research is motivated by the need to minimize the impact of cold starts and optimize resource utilization in serverless applications by developing intelligent, data-driven scaling decisions. We delve into using machine learning and feature engineering to model and simulate predictions for optimal scaling decisions for Azure Function Apps (AFA). Our focus lies in predicting the ideal timing for provisioning or de-provisioning the Function App’s environment. Using historical invocation data, we applied a sliding window to transform the time-series data into patterns categorized as load or unload classes, considering various target periods. To identify the most effective model, we compared the performance of various baseline models with and without calibration (isotonic and sigmoid) to enhance precision. In addition, we assess multiple feature extraction methods in invocation patterns and explore the use of Principal Component Analysis (PCA) for dimensionality reduction to reduce computation costs. Using the best-identified configurations, we model and simulate the class patterns over time to compare the actual classes with the predicted ones, focusing on memory usage and the costs associated with cold starts. The proposed model is thoroughly evaluated using various metrics under different setups, revealing notable improvements in scaling decisions achieved by applying calibration and feature engineering methods. These findings demonstrate the potential of machine learning for intelligent, data-driven scaling decisions in serverless computing, offering valuable insights for cloud providers to optimize resource allocation and for developers to build more efficient and responsive serverless applications. Specifically, the proposed method can be integrated into serverless platforms to automatically adjust resource provisioning based on predicted workload demands, reducing cold start latency and improving cost-effectiveness.