Source time Edition date
IndustryAWS Machine Learning Blog
Orchestrating SageMaker AI endpoint auto scaling with AWS CDK
使用 AWS CDK 编排 SageMaker AI 端点自动扩缩容
Teams deploying SageMaker AI endpoints via AWS CDK must weigh the operational burden of a custom-resource workaround against the risk that rolling updates fail outright when Application Auto Scaling is registered.
Judgment
Platform and MLOps teams that own SageMaker endpoint CDK stacks must decide whether to build and maintain a reusable custom-resource construct that deregisters and re-registers Application Auto Scaling during updates, or keep manual intervention and accept failed or blocked instance-type, instance-count, and environment-variable rollouts.
Basis
Lauren Eterno's AWS Machine Learning Blog post (23 OCT 2026) states that changing an Amazon SageMaker AI real-time endpoint's instance type, instance count, or environment variables makes CloudFormation or AWS CDK deployments fail. The mechanism: SageMaker AI requires the scalable target to be deregistered before the endpoint update, but deregistering prevents CloudFormation from tracking the Application Auto Scaling resources. The proposed fix is a reusable CDK construct using CloudFormation custom resources to deregister and register auto scaling automatically during endpoint updates, so endpoint configuration and auto-scaling parameters can be updated in one deployment without manual intervention. The post notes several failure modes depend on the change and workaround, but the excerpt stops before listing them.
Unclear The source does not quantify the construct's maintenance cost, deployment duration, failure rates, or which failure modes occur for each endpoint change.
Next Check the published construct's CloudFormation custom-resource code and IAM permissions, then test one instance-count update in a non-production stack to confirm auto scaling is re-registered.