Question

A media company runs a real-time content-tagging service on Amazon Bedrock using Anthropic Claude. During a viral event, traffic surged from ~50 requests/second to peaks of ~400 requests/second for roughly 90 minutes, then dropped back. The team observed frequent ThrottlingExceptions on their on-demand invocations during the surge, degrading user experience. They want the MOST cost-effective approach to handle these unpredictable, short-lived spikes without over-provisioning for the baseline load. Which approach should they implement?

Accepted Answer

Configure a cross-region inference profile to distribute on-demand invocations across multiple Regions, increasing aggregate throughput during spikes, combined with exponential backoff with jitter on retries.

Answer

Purchase Provisioned Throughput with a 6-month commitment sized to the 400 requests/second peak so capacity is always guaranteed.

Answer

Switch all real-time invocations to Bedrock batch inference jobs so the workload is decoupled from throughput limits.

Answer

Purchase no-commitment Provisioned Throughput sized to baseline (50 requests/second) and rely on it to absorb the 8x surge automatically.