Why is client-side load balancing (each service instance discovering peers itself) sometimes preferred over a centralized load balancer?

A centralized load balancer is a single hop every request must pass through, adding latency and becoming a potential bottleneck/single point of failure at scale. Client-side load balancing (via a service registry like Eureka) lets each caller query the registry once and then call instances directly, spreading the discovery load and removing the extra network hop per request.