AWS describes a reusable CDK construct that sequences SageMaker endpoint updates and auto scaling registration through CloudFormation custom resources.
AWS published a reference implementation on October 23, 2026, for updating Amazon SageMaker AI endpoints with auto scaling managed in the same AWS Cloud Development Kit deployment. The guidance addresses a deployment conflict: SageMaker requires an endpoint’s scalable target to be deregistered before certain configuration changes, while CloudFormation may lose track of the scaling resources if they are removed manually.
The proposed construct uses three CloudFormation custom resources to coordinate the operation. First, it deregisters the scalable target. CDK then updates the endpoint configuration. After the endpoint reaches “InService,” another resource registers the scalable target, and a final step attaches the scaling policy once that target can be queried. Explicit dependencies enforce this sequence.
The post says this arrangement lets teams change endpoint settings and scaling parameters in one deployment, without manually deregistering and restoring auto scaling. A shared configuration hash triggers the orchestration when either set of values changes. The example includes settings such as instance type, minimum and maximum capacity, target value, and scale-in and scale-out cooldowns; these are configurable examples, not prescribed defaults.
AWS also details the operational footprint: the sample stack uses three custom resources, two Step Functions state machines, and 12 Lambda functions, including functions generated by the CDK provider framework. Asynchronous completion handlers wait for the endpoint to become ready and for the scaling target to become available, with stated time limits of one hour for target registration and 15 minutes for policy attachment.
The implementation is presented as a reusable sample, not a change to SageMaker’s underlying service behavior. Teams starting from the example can deploy its sample stack; teams with an existing CDK stack can add the construct alongside their endpoint resources. AWS notes that deployments require the relevant service permissions and that deleting the sample stack removes the endpoint and associated resources.