Retraining triggers and knowledge-base refresh
Lesson 4 of 4 in MLOps Pipelines: CI/CD for Models.
A deployed model is a snapshot of the world that keeps aging. Domain 3 closes with the loop that fights the aging: something notices (a schedule, or drift caught by monitoring), EventBridge carries the signal, a pipeline retrains, evaluates, and registers, a gate approves, and the endpoint updates — then monitoring starts the next lap. Every piece is a service you have already met; this lesson is about the wiring.
| SageMaker Pipelines | Step Functions | Amazon MWAA (Airflow) | |
|---|---|---|---|
Shape | ML-native DAG: processing, training, tuning, condition, register-model steps; step caching; lineage | General serverless state machine with retries, error handling, and integrations across virtually all AWS services | Managed Apache Airflow: Python DAGs, schedules, the broadest OSS operator ecosystem |
The tell | "Pure ML workflow on SageMaker AI" — the default for this domain | "The workflow spans many services" (Lambda, Glue, Bedrock, approvals) or needs complex branching/error flows | "The team already runs Airflow" / existing DAGs to migrate |
Key terms: SageMaker Model Registry, SageMaker Pipelines, Deployment guardrails, Blue/green deployment, Canary deployment, Bedrock Prompt Management
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.