Decoupled Bedrock Model Cascade with SQS and Fargate

Why the Answer Is Correct

Option C routes incoming requests into an Amazon SQS queue, decoupling producers from processing so sudden demand spikes are absorbed by the queue buffer rather than overwhelming the models. AWS Fargate scales the consumer tasks automatically without server management, invoking the small low-latency Bedrock model first to keep high-confidence results fast and cheap. When the confidence score falls below 0.65, the request is placed into a second SQS queue, which allows the larger, more expensive model to run asynchronously and only when required, directly satisfying the cost-optimization requirement. This two-queue design gives high resiliency through decoupled components, retries, and independent scaling of each stage.

Why the Other Options Are Wrong

Option A escalates synchronously to the larger model in the request path, which fails the explicit requirement that low-confidence requests be processed asynchronously and creates latency and availability coupling between the two models. Option B runs both the small and large model in parallel branches for every request, discarding all cost savings because the expensive model is invoked 100% of the time. Option D places models on Amazon EC2 instances with custom phrase-length and keyword routing, which reintroduces undifferentiated instance management, lacks managed decoupling, and replaces Bedrock's confidence-based escalation with a brittle hand-written heuristic instead of the actual score threshold.

Community Comment Notes

A single community comment identified the same verdict, selecting C because "Send requests to an Amazon SQS queue. Use AWS Fargate to process messages." The comment's reasoning aligns with the requirement set—decoupled queues for asynchronous large-model escalation—though no further discussion was provided to corroborate or challenge the choice.