The /api/v2/analytics/predictiverouting/events channel is dropping roughly 15% of interaction_routed payloads when agents toggle queues faster than the 2-second heartbeat on our us-east-1 cluster. Sequence numbers jump from 482 to 511 right after the WebSocket reconnects, and the dashboard throughput just flatlines. Someone mentioned the v2.1 Notification API handles backpressure automatically, which isn’t quite right since the v2.3 spec explicitly requires manual lastReceivedSequence tracking for these streams. Latency spikes to 800ms and the queue metrics just freeze.