DEVELOPMENT OF A DEEP LEARNING-BASED AUTOSCALER ON KUBERNETES

• Main Author: Fadhil Al Hafidz, Rozan
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85029
• Institution: Institut Teknologi Bandung

Kubernetes has become the primary platform for container management in modern computing environments. However, the built-in Kubernetes autoscaler, HorizontalPodAutoscaler (HPA), still has limitations in responding to sudden workload changes. This study aims to develop a deep learning-based autoscaler on Kubernetes that can respond more quickly and perform scaling by predicting upcoming workload metrics to prevent the delayed queueing effect. The study involves several key stages: designing and implementing a system for collecting, processing, and storing raw metric data from Kubernetes; designing and implementing a deep learning-based autoscaler using a Recurrent Neural Network (RNN); and comparing the performance of the developed autoscaler with HPA. The testing results indicate that although the deep learning-based autoscaler has potential in predicting workload metrics, its predictions are not yet accurate enough. Consequently, the performance of the autoscaler is still not better than that of HPA. This is due to the limited amount of training data and the premature prediction of metric decreases. Recommendations for future research include increasing the amount and variety of training data, implementing automatic fine-tuning, using a stabilization window for scaling stability, and integrating the autoscaler as a Kubernetes extension. This research demonstrates that using deep learning for autoscaling in Kubernetes holds promise, but further research is needed to achieve better performance than HPA.