Revolutionizing Cloud Management with AI-Powered Kubernetes Autoscaling Solutions

Mundigehalla Prabhakar, Chethan (2025) Revolutionizing Cloud Management with AI-Powered Kubernetes Autoscaling Solutions. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (2MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

The proliferation of microservices and containerization, orchestrated predominantly by Kubernetes, has introduced significant challenges in cloud resource management. Traditional autoscaling mechanisms, such as the Horizontal Pod Autoscaler (HPA), are inherently reactive, adjusting resources only after performance metrics have already breached predefined thresholds. This reactive nature often causes periods of inefficient application performance or unnecessary over-provisioning of resources leading to buggy user experiences and excess operational costs. This work addresses the inefficiencies of the current autoscaling practices by designing, implementing, and evaluation a proactive, AI-based autoscaling solution for Kubernetes. The architecture we designed for this project implements a hybrid AI approach, using Machine Learning (ML) for workload forecasting, and Reinforcement Learning (RL) for rapid, intelligent, forward-looking scaling decisions. We designed a custom Kubernetes controller that encapsulates the AI pipeline using a Custom Resource Definition (CRD). An additional backoff controller was also included to ensure production-grade reliability - a large challenge when relying on distributed, AI systems - the backoff controller defaults back to standard HPA when there is low confidence in the AI model or instability in the system. Lastly an A/B testing controller automates the safe deployment and validation and promotion of new AI models. The evaluation on a Google Kubernetes Engine (GKE) cluster, demonstrated that the system was able to continually anticipate workload fluctuations, make smart scaling decisions to retain application performance and resource efficiency, and showed robust resilience. Overall the results confirm that a unified, AI-driven, approach can meaningfully push the state of cloud management further by fundamentally changing autoscaling from a reactive strategy, to a proactive optimization engine.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Mijumbi, Rashid UNSPECIFIED
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence T Technology > T Technology (General) > Information Technology > Cloud computing Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning
Divisions:	School of Computing > Master of Science in Cloud Computing
Depositing User:	Ciara O'Brien
Date Deposited:	30 Mar 2026 10:56
Last Modified:	30 Mar 2026 10:56
URI:	https://norma.ncirl.ie/id/eprint/9246

Actions (login required)

View Item