MLOps

MLOps, short for machine learning operations, applies engineering and operational practices across a machine-learning system's lifecycle. It connects development decisions with the work of running, changing, and eventually retiring the system. It is a practice, not the name of a mandatory platform.

What must travel with a forecast?

Suppose a shop trains a model to predict tomorrow's orders. Saving the model file is only one step. The team also needs the definition of an order, input transformations, evaluation evidence, the runtime environment, and a way to identify which release produced a prediction. Otherwise, a successful experiment can become an unexplained production result.

Evaluation determines whether a candidate meets agreed requirements. Deployment makes an accepted candidate available in its target environment. Monitoring then connects observed behavior to investigation and an owned response.

Automation does not decide the objective

An automated training run can finish successfully while producing an unsuitable model. The release process must still apply quality and compatibility checks. A changed input may call for repairing data rather than retraining.

A small team can start with repeatable scripts, versioned records, and a named responder. More automation helps when it makes those responsibilities reliable; installing additional tools does not itself establish them.

Reference: Google Cloud: MLOps.