AWS AppConfig lets you manage model-selection settings as feature flags that Lambda functions retrieve at runtime through the AWS AppConfig Agent Lambda extension, which polls for updated configuration data over a local HTTP endpoint, so no code changes or redeployments are needed to change which Bedrock model is used. AppConfig deployment strategies support a Linear growth type, where you define a growth factor (percentage of targets that receive the new configuration at each step, e.g., 20%) and a deployment duration, giving the required staged rollout starting at 20% of traffic. AppConfig also supports validators and CloudWatch alarm-based automatic rollback: if a configured CloudWatch alarm (such as one tracking error rate crossing a 1% threshold) enters the ALARM state during the deployment or the subsequent bake time, AppConfig automatically reverts the configuration to the last-known-good version, preventing continued exposure to a faulty model configuration. This combination of feature flags, Lambda extension retrieval, linear deployment strategy, and alarm-triggered rollback is documented in the AWS AppConfig user guide and directly satisfies every stated requirement, which the other options (custom DynamoDB/EventBridge logic, Lambda version/alias traffic shifting via CodeDeploy, or manual CloudWatch/Parameter Store logic) do not fully provide without extra custom engineering or without violating the no-redeploy requirement.
Community Discussion