Dimensioning Microservices on Kubernetes Platforms Using Machine Learning Techniques

Abstract: In recent years, cloud computing and containerization have become increasingly popular for various applications. However, optimizing resource usage and minimizing costs while providing reliable and efficient service to users can be a challenge. One such challenge is scaling containers according to the current system load. This thesis focuses on the problem of properly dimensioning pods in Kubernetes, a widely used platform for managing containerized applications. The motivation behind this problem is that the standard solution, Horizontal Pod Autoscaler (HPA), often results in wasted resources or poor performance due to overprovisioning or underprovisioning. This thesis proposes a proactive approach to scaling applications by utilizing machine learning models to predict future resource requirements based on concurrent user counts on the platform. This approach involves collecting and modifying data from HPA and training various machine learning models to forecast and predict the system's future behavior. The results of this thesis demonstrate that machine learning models can predict future resource requirements based on trends in the application. Based on the data evaluation, it was found that overprovisioning negatively affects the response time. Additionally, the machine learning models were accurate enough to predict future CPU needs. However, further research and experimentation are necessary to improve the accuracy and reliability of these models.