MLOps: shipping, monitoring and updating models

In short: MLOps is the craft of bringing models reliably into operation and keeping them healthy there. A model in a notebook is a start, not a finished product.

The life cycle

1. Data: version, check, provide reproducibly.
2. Training: automated, with recorded settings and results (experiment tracking).
3. Testing: on test data and on cases that caused errors earlier, see AI evaluation.
4. Delivery: as a service behind an interface, as a container (Podman and Docker compared) or on the device. New versions first for a few users (canary), then for all, with a way back.
5. Monitoring: response time, errors, cost and above all quality.
6. Retraining: when data or the world change.

Drift

• Data drift: the inputs change (different customers, new sensors).
• Concept drift: the relationship itself changes (fraud adapts).

Without monitoring quality drops unnoticed. You compare input distributions with those from training and measure quality as soon as the real outcomes are known.

Tools

MLflow (tracking), DVC (data versions), feature stores, orchestration such as Airflow. For language models prompt versions, cost control and prompt caching come on top.

See also: Inference servers