Single stream partition throttles DynamoDB

DynamoDB table keyed on streamId (partition key) + ts (sort key) storing live-stream chat events. During a single huge livestream, the app starts getting ProvisionedThroughputExceededException / throttling on writes for that stream while overall table-level consumed capacity is well under the provisioned total. CloudWatch shows one partition's throttle count spiking; everything else is fine. The stream has 5M concurrent viewers chatting. Triage, mitigate live, and redesign so it can't recur.