IoT telemetry ingestion

Design an ingestion pipeline for 50 million IoT devices (smart meters, EV chargers, sensors) each emitting a small telemetry record every 10 seconds, ~5M writes/sec at peak, bursty when a firmware rollout reboots a region simultaneously. Records must be durably stored, queryable as time series, and trigger near-real-time alerts (e.g., over-temperature). Devices have weak/intermittent connectivity and limited buffering. How do you handle the massive fan-in without dropping data or melting the backend?