Streamlining ML Training in Kubernetes: An MLOps Architecture with Kubeflow

Ioannis Korontanis, Athina Zacharia, Antonios Makris, Maria Pateraki, Konstantinos Tserpes

Abstract: Machine Learning Operations (MLOps) is essential for automating the deployment, monitoring, and management of ML models. By integrating MLOps with DevOps practices, developers can create automated training pipelines. This paper explores using Kubeflow as an MLOps platform and GitHub Actions as a CI/CD pipeline for training and deploying ML models. Kubeflow provides a scalable framework for orchestrating ML workflows in containers, with Kubernetes enabling efficient resource management. Containerization ensures consistency, portability, and reproducibility across environments, while GitHub Actions automates testing, version control, and deployment. A real-world case study demonstrates this architecture and discusses challenges and best practices for modern MLOps workflows.

External IDs:doi:10.1145/3770501.3771304