MLOps with MLflow — From Experiment to Production Model

“Which model was it? The one with learning rate 0.01 or 0.001?” — if that sounds familiar, you need MLOps. We chose MLflow.

The Problem: ML Without Process¶

Jupyter notebooks, local copies of data, “best models” on a local disk. Reproducing last week’s experiment = hours of detective work.

MLflow’s Four Components¶

• Tracking — parameters, metrics, artifacts + git hash
• Projects — packaging ML code for reproducibility
• Models — standard format for deployment
• Registry — versioning, stage management (None → Staging → Production → Archived)

Automated Pipeline¶

In Airflow: data ingestion → feature engineering → training → evaluation. If the new model outperforms the production one (AUC > current + 0.02), it’s registered as Staging. Manual approval for Production.

MLOps = ML + DevOps Discipline¶

Without MLOps, ML is an experiment. With MLOps, it’s an engineering discipline.