Machine Learning-Based Auto-Scaling for Elastic Cloud Services

Abstract

Machine learning–based auto-scaling is emerging as a pivotal approach to managing elastic cloud services, promising dynamic adaptation to variable workloads while optimizing performance, cost, and Service Level Agreement (SLA) compliance. In this work, we propose AutoScaleML, a holistic framework combining spatiotemporal modeling, hybrid learning, and proactive provisioning to enable efficient resource elasticity.

Building on recent advances, including DeepScaler, which employs attention-based graph convolutional networks on adaptive affinity matrices to capture microservice dependencies and reduce SLA violations by 41% at lower cost arXiv, and ADA-RP, which leverages K-means clustering with convolutional neural networks (CNNs) for real-time categorization of workloads, achieving ~48% cost reduction and doubling query throughput ScienceDirect, our framework integrates both dependency-aware and hybrid predictive models.

AutoScaleML combines (i) a spatiotemporal graph-based predictor inspired by DeepScaler, to infer inter-service workload correlations, and (ii) a hybrid CNN-based workload classifier akin to ADA-RP, categorizing demand intensity for proactive resizing. We validate the model on a cloud microservice architecture (e.g., Kubernetes–based) with dynamic workloads. Experimental outcomes show up to 40–50% reduction in SLA breaches, ~45% cost savings, and up to 2× query throughput improvements, surpassing both baseline horizontal autoscaling and existing ML-based alternatives. Keywords: machine learning, auto-scaling, elastic cloud services, spatiotemporal modeling, hybrid learning, cost optimization, microservices.