Service discovery and config distribution

Design dynamic service discovery + config distribution for a microservices fleet of ~3,000 instances across regions. When an instance starts it must register so callers can find it; when it dies or fails health checks it must be removed quickly so traffic stops; and operators must be able to push a config change (e.g., a feature flag or a traffic weight) and have all instances converge within seconds. How do you build the coordination layer and avoid it becoming a single point of failure or a thundering-herd hotspot?