Production ML

Under development

This lesson is part of the course scaffold and is being actively written. The learning objectives and outline below define its final scope.

Learning objectives

By the end of this lesson you will be able to:

• Implement online learning updates and decide, with evidence, when incremental updates beat scheduled full retrains
• Detect concept drift with statistical monitors on features and predictions, and define the action each drift signal triggers
• Run champion/challenger retraining with explicit promotion criteria and a rollback path
• Version models in a registry so any live prediction can be traced to an exact model artifact, training window, and feature set

Outline

1. Online learning — incremental updates vs full retrains
2. Concept drift — detection on features, predictions, and outcomes
3. Model monitoring — performance, stability, and staleness in production
4. Champion/challenger retraining — promotion criteria and rollback
5. Model registries and versioning — artifacts, lineage, and reproducibility
6. Wiring model health into the live monitoring and alerting stack

Prerequisites

• Tree Ensembles
• Part VI — Monitoring, Logging, Alerting