Data teams want fast, safe ways to ship models to production — but cloud budgets and managed platforms are not always an option. This talk presents a simple, fully on-prem MLOps reference setup built from open-source components: training with notebooks/Airflow, experiment tracking and registry with MLflow + Postgres, artifact storage on MinIO (S3), promotion via MLflow aliases (Test/Production), and per-alias serving using MLflow Serve containers.
The key idea: aliases act as stable contracts, so promotion and rollback become an instant alias switch while model versions change underneath. You’ll see a short demo of the workflow: train → register → switch alias → a new serving container starts → dashboards update.
We’ll also cover what this setup monitors today (health checks, logs via Loki, basic Grafana panels) and what to add next (latency/RPS/error metrics, drift and quality monitoring).