DVC

Building Secure MLOps Platforms in Air-Gapped Environments for DevOps Engineers

The article explains how DevOps engineers can construct a minimal yet complete MLOps platform inside a closed, air-gapped network without relying on any cloud services. It highlights the cultural gap between data scientists and DevOps teams and shows how MLOps practices such as data versioning, experiment tracking, and model registries bridge that gap. The architecture separates GPU training nodes from a Kubernetes serving cluster and uses open-source components including ArgoCD for GitOps, MLflow with PostgreSQL, MinIO, DVC, OpenBao, and External Secrets Operator. Special attention is given to secrets management with SOPS and age keys, image scanning with Trivy and Bandit, and secure Docker access via socket proxy on the training server. The setup avoids Kubeflow and HashiCorp Vault due to complexity and licensing concerns, favoring lighter, license-friendly alternatives suitable for isolated perimeters.