A trained model is only the start of MLOps. Real systems need a repeatable pipeline that pulls fresh data, retrains and versions a model, rolls it out safely, and watches it in production. This scene lays that pipeline out in 3D: a data lake feeds a training station, trained models are versioned into a registry, traffic is split between a small canary release and the stable production model, and a monitoring dial tracks how far live data has drifted from what the model was trained on.
"Data drift" and "concept drift" are the two most common reasons a production model quietly gets worse over time — the world keeps changing even when the model's weights don't, which is why monitoring and scheduled retraining are core MLOps practices rather than optional extras.
A 3D MLOps pipeline where data flows from a data lake through training into a versioned model registry, splits across a canary and production deployment, and gets watched by a live drift monitor.
How production ML systems are more than a trained model: versioned training runs, staged canary rollout, live serving traffic, and continuous drift monitoring that can trigger automatic retraining.
Set the drift rate, traffic volume and canary split, then watch accuracy fall as drift rises. Leave auto-retrain on to see the pipeline recover itself, or force a retrain manually at any time.
Most production model failures aren't bugs in the code — they're silent data or concept drift, which is why monitoring dashboards and scheduled retraining pipelines are considered core MLOps infrastructure.