Yesterday we surveyed 4 inference options. Most common and exam-heavy is real-time endpoints. Almost every online service where users await results deploys via real-time endpoints. Today we cover real-time endpoint setup (Model → EndpointConfig → Endpoint), auto-response to traffic (auto-scaling), and instance choices (CPU/GPU/accelerator).
Key insight: SageMaker real-time deployment is a 3-layer structure. Model bundles artifact and inference container; EndpointConfig specifies instance type/count; Endpoint handles actual HTTPS traffic. This separation enables zero-downtime deployment and A/B testing.
Real-time endpoints are built from 3 resources: