Skip to main content

Description

Identify Amazon Bedrock Provisioned Throughputs in the InService or Updating state whose current foundation model is in the LEGACY lifecycle state. This preventive requirement is to use an Active model and prepare migration before retirement.

Rationale​

A Legacy foundation model is scheduled for retirement. Moving inference workloads to an Active model before the regional EOL date reduces the risk of requests failing when the old model becomes unavailable. AWS does not migrate callers automatically.

Impact​

Migration can change response quality, latency, model behavior, capacity requirements, and cost. Provisioned Throughput by Model Units associated with a base model cannot have its model changed in place; provision a replacement as needed, validate it, and update inference callers. Review existing commitments before purchasing or retiring capacity.

Audit​

This policy marks an AWS Bedrock Provisioned Model Throughput as INCOMPLIANT if its Status is InService or Updating and its current Model resolves to an AWS Bedrock Foundation Model whose Model Lifecycle: Status is LEGACY. It is COMPLIANT when that lifecycle status is ACTIVE. The related Model ARN must exactly match the throughput's current Model ARN, including the AWS Region and any context-length suffix. Legacy means deprecated and scheduled for retirement; it does not establish EOL, paid extended access, or actual request traffic. Existing Legacy endpoints may still be valid for inference; the finding calls for migration planning rather than declaring an outage or an invalid deployment.

Deployments with Status Creating or Failed are INAPPLICABLE; creation has not completed or a deployment operation has failed, so deployment health must be assessed separately. This exclusion does not prove that a failed update stopped all inference or that its retained model is supported. Updating remains in scope because AWS permits inference during updates and the old model can continue serving until deployment completes. Missing or unrecognized statuses are UNDETERMINED. Disappeared throughputs are automatically DISAPPEARED.

Custom current model ARNs are INAPPLICABLE, including the CMDB route with an empty Model ARN and a populated Custom Model ARN. This excludes custom endpoints, including imported custom-model routes, from the foundation-model requirements.