Discovery & Registration Patterns

Why discovery exists

IPs and ports change on every deploy, autoscale event, and spot interruption. Clients need a stable name → current healthy instances.

Client-side vs server-side discovery

Pattern	Flow	Examples
Client-side	Client queries registry, picks instance, connects	gRPC + xDS, Netflix Eureka + Ribbon (legacy), Consul agent
Server-side	Client → load balancer → registry-backed pool	AWS ALB/NLB, K8s Service + kube-proxy, Envoy gateway

Server-side — simpler clients, LB is chokepoint. Client-side — smarter routing (latency-aware, subset), harder client libraries.

Service registry

Catalog of {service_name → [(host, port, metadata)]}.

Registration:

• Self-registration — service starts, registers with Consul/etcd, deregisters on shutdown
• Third-party — K8s controller watches pods, updates Endpoints

Deregistration on crash — TTL heartbeats; stale entries removed after missed beats.

Kubernetes model (interview staple)

Pod → labeled → Service (ClusterIP) → Endpoints object → kube-proxy/iptables or eBPF → DNS: my-svc.namespace.svc.cluster.local

• ClusterIP — internal VIP, not routable outside cluster
• Headless Service (clusterIP: None) — DNS returns pod IPs directly (StatefulSets, Cassandra)

Consul / etcd as registry

• Consul — service catalog + health checks + KV + mesh connect
• etcd — K8s' backing store; not usually called directly by apps

Distributed Systems → Consensus owns Raft/etcd correctness — here: apps consume the catalog API.

DNS-based discovery

Simple: payments.prod.internal → multiple A records (round-robin DNS).

Weaknesses: TTL caching, no health awareness, limited load balancing. Fine for stable fleets; K8s DNS is richer.

Service mesh discovery

Istio/Linkerd use service identity + control plane pushing endpoint lists to sidecars (xDS). App connects localhost:sidecar; sidecar knows real backends.

Failure modes (staff depth)

• Thundering herd — registry restart, all clients refresh simultaneously
• Stale cache — client holds dead instance until TTL
• Split brain registry — rare with Raft; know health check design matters

Cross-reference: Networking DNS and L7 LB; Communication → gRPC for why connection reuse needs L7-aware balancing.