4. Deployment and Orchestration of ML Workflows (22%)

A model sitting in a notebook is worth nothing to the business — deployment is where ML generates value. This domain at 22% tests the full journey from trained model to production system: selecting the right endpoint type and compute (4.1), scripting reproducible infrastructure with IaC and auto scaling (4.2), and building CI/CD pipelines that automate the retrain-deploy cycle (4.3).

The sections are ordered to mirror a real deployment workflow. First you decide where the model runs (endpoint type, instance family, container strategy). Then you make that infrastructure repeatable and scalable (CloudFormation/CDK, auto scaling, VPC configuration). Finally you automate the lifecycle so model updates flow through pipelines rather than manual steps (CodePipeline, SageMaker Pipelines, deployment strategies). Expect the exam to present end-to-end scenarios that cross all three sections — a question about a canary deployment (4.3) may hinge on your auto scaling configuration (4.2) and endpoint type choice (4.1).