Service Registry
The database of currently available service instances and their addresses, maintained by registration and pruned by health checking.
The component that makes discovery possible, and a piece of infrastructure whose own availability becomes a system-wide dependency.
Two registration models. Self-registration: instances register on startup and heartbeat to stay listed. Simple, and it requires registry logic in every application. Third-party registration: the platform registers instances on your behalf — what Kubernetes does through the API server and Endpoints — which keeps applications ignorant of the registry entirely.
The property that actually determines availability is how fast an unhealthy instance leaves rotation, since that interval is dead time in every caller's request stream. It is set by the health check interval times the failure threshold, plus propagation delay, and it is frequently tens of seconds when people assume it is one.
The failure mode to design against is registry unavailability. A caller that cannot resolve a service and therefore fails has made the registry a hard dependency of every request. Statically stable designs cache the last known good instance list and keep using it, on the reasonable assumption that a stale list is better than no list.