SQS FIFO queue throughput collapses

An SQS FIFO queue account-commands processes per-account commands (MessageGroupId = account_id) and guarantees in-order processing per account. At 10:00, overall throughput collapses: ApproximateNumberOfMessagesVisible climbs to 90k, but the worker fleet is mostly idle (low CPU, few in-flight). Dashboards show one message group is stuck, its handler keeps failing and the message is being retried (it's in flight, then back, repeatedly), and maxReceiveCount is high so it never DLQs. Recent context: one specific account triggered a command that hits a code path with a null-pointer bug deployed this morning. Explain why the whole queue's throughput collapsed and how you mitigate.