Deployment & Orchestration

A trained model earns nothing until it is serving predictions, and this domain is the arc that gets it there. The single organizing question is where you are on that arc, because each answer lives in a different subtopic. First you pick how the model is served: which Amazon SageMaker AI inference option fits the traffic. Then you decide what runs it: the compute, the scaling behaviour, and the infrastructure-as-code (IaC) that provisions it repeatably. Then you automate the whole thing so a retrained model reaches production the same way code does. The classic exam trap collapses these into one question, offering a heavier answer than the stage needs, such as a full CI/CD pipeline where the scenario only asks how to serve one model, or a persistent real-time endpoint where the traffic is plainly offline. Name the stage the scenario is actually at (select, provision, or automate) and the right service almost always follows from that alone.

Read this page as a map, then follow the three subtopics in order. Deployment Infrastructure covers the select step: it walks the four SageMaker AI serving options (real-time, serverless, asynchronous, batch transform), single- versus multi-model endpoints, the container and compute choice, and SageMaker Neo for edge or constrained hardware, so you can match a serving pattern to traffic shape, payload size, and latency need. Provisioning & Scaling covers the provision step: on-demand versus reserved capacity, Application Auto Scaling policies (target-tracking, step, scheduled), the load-proportional metric to track, defining infrastructure as code with AWS CloudFormation or the AWS Cloud Development Kit (AWS CDK), and keeping the endpoint private in your VPC. CI/CD Pipelines for ML covers the automate step: modelling the release with AWS CodePipeline, choosing the ML-native orchestrator, gating deployment on the SageMaker AI Model Registry approval, updating a live endpoint with deployment guardrails, and closing the loop so drift can trigger retraining. Each subtopic carries the mechanisms, the exact limits, and the traps; this overview just shows how they fit together.

Across all three steps the exam rewards the same instinct: choose the most managed option that meets the requirement, and no more. That means a serverless endpoint over an always-on one when traffic is intermittent, an auto-scaling policy over a fixed instance count when demand varies, IaC over click-ops for anything you must reproduce, and SageMaker Pipelines over a hand-rolled orchestrator for a pure SageMaker workflow. The right-sized answer is rarely the biggest one, so let the requirement in the scenario (steady or spiky traffic, a latency SLA, a need to reproduce or roll back) pick the option, not a reflex to reach for the most powerful service.