How do services find each other when instances come and go?

Static config breaks the first time the autoscaler adds an instance nobody listed, or a rolling deploy replaces every IP. A service registry fixes it: each instance registers itself (or the orchestrator registers it) on start with its address and a health endpoint; it renews a heartbeat with a TTL (say 10–30 s); the registry drops entries whose TTL expires or whose health check fails. Service A queries the registry (or a local cache of it) and load-balances across healthy instances.

Two shapes: client-side discovery, where A holds the list and picks an instance (fewer hops, smarter routing, but every language needs the client library), and server-side, where A calls a stable name and a load balancer or proxy does the lookup (simpler clients, one more hop). Kubernetes services are server-side via DNS plus kube-proxy; a service mesh puts a sidecar next to each instance and does discovery, retries and mTLS there — see Service Discovery.