Perpustakaan Digital ITB

Kubernetes is an orchestration platform that is commonly used to manage containerized applications in microservices architectures. One of its core components, Horizontal Pod Autoscaler (HPA), enables dynamic resource scaling by adjusting the number of pods ba-sed on real-time system load. However, HPA is a reactive approach and does not optimally utilize historical application metrics. This leaves room for improvement through proactive provisioning strategies that anticipate resource needs before actual load spike occur. The limitations of this reactive mechanisms can lead to reduced service availability and degra-ded application responsiveness, especially under fluctuating traffic conditions. This study proposes a proactive provisioning approach based on machine learning to address dyna-mically changing traffic patterns. The proposed solution integrates SARIMA-based model using historical CPU usage data, a warm pool of pods to reduce pod initialization time, and weight-based dynamic routing to efficiently distribute traffic to ready resources. There are four continuous phases for the proposed solution: machine learning training and data collection, model updating and traffic prediction, dynamic routing and spike management, and spike recovery. Evaluation for the system is done under simulated traffic scenarios including gradual pattern, cyclical pattern, stable loads, and random loads. The result de-monstrates that the proposed approach enables more proactive provisioning than the default HPA mechanism. It also improves resource allocation efficiency without compromising system stability by accurately predicting load patterns and anticipating spikes in advance. The results highlight the effectiveness of combining proactive and reactive strategies for managing dynamic workloads in Kubernetes environment.