ETA prediction service

Design an ETA prediction service that powers a navigation app serving 1M route requests/minute. ETAs must reflect live traffic that changes minute-to-minute, must be returned in <100ms p99 for a route that may span hundreds of road segments, and must stay accurate across rush hour, incidents, and weather. You have a fleet of probe vehicles feeding live speeds. Describe how you ingest live traffic onto the road graph, compute the ETA, serve it at low latency, and reconcile the physics-based segment model with a learned correction.