Department of Computer Science: MSc Thesis Presentation
Kubernetes-Based Architecture For An On-premises Machine Learning Platform
Author: Swati Choudhary
Supervisor: Juho Rousu
Date: Thursday 9 September 2021
Abstract of thesis:
This thesis introduces an architecture where a machine learning platform can be set up on an unmanaged Kubernetes cluster in an on-premises environment, using Kubeflow and Kubeflow Pipelines. The proposed architecture allows the solution to be platform-agnostic and avoid vendor lock-in. The goal for providing an on-premises architecture is to be able to have control over the security of the platform and conform the machine learning service to custom Service Level Agreements (SLAs). While designing the architecture, open-source technologies are used which allow for review of the source code to identify potential vulnerabilities and weaknesses of the platform. This is especially vital in use-cases revolving around critical security and confidentiality. As part of experiments, it is shown how a machine learning model can be trained using Kubeflow Pipelines and deploy the model using Kubernetes in order to serve real-time user requests for prediction. Finally, the complexity of the proposed solution versus the existing solution is compared both with respect to the development of the solution as well as migration of the solution to other vendors.