Endpoint anatomy: variants, instances, containers

Lesson 2 of 4 in Serving Predictions: Endpoints and Batch.

"Deploy to an endpoint" is really three API objects stacked, and the exam checks you know which fact lives where. A model names the container image (from Amazon ECR), the model artifact in S3, and the IAM role the container runs with. An endpoint config lists one or more production variants — each variant pairs a model with an instance type, an initial instance count, and a traffic weight. An endpoint brings a config to life as a persistent HTTPS resource. Updating production means creating a new config and calling UpdateEndpoint — SageMaker AI stands up the new fleet before draining the old one, so the swap itself causes no downtime.

Control plane builds it, data plane calls it — two different API clients
CallPlane (Boto3 client)What it carries

CreateModel

control (sagemaker)

Container image URI, model artifact S3 path, execution role

CreateEndpointConfig

control (sagemaker)

The variant list: model name, instance type and count, weight per variant

CreateEndpoint / UpdateEndpoint

control (sagemaker)

Binds a config to a named, persistent HTTPS endpoint

InvokeEndpoint

data (sagemaker-runtime)

The payload itself — plus optional TargetVariant routing

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.