Service Discovery
Client-side vs server-side discovery, DNS-based discovery in Kubernetes, registries like Consul and Eureka, and health-aware routing.
What it is
In dynamic environments, instances come and go (autoscaling, deployments, failures), so callers cannot hard-code addresses. Service discovery maps a logical service name to the current healthy instances.
| Pattern | How | Examples |
|---|---|---|
| Server-side discovery | Caller hits a stable address (LB/VIP); the LB knows instances | Kubernetes Service (ClusterIP), AWS ALB target groups, Cloud Map + ALB |
| Client-side discovery | Caller queries a registry and load balances itself | Netflix Eureka + Spring Cloud LoadBalancer, Consul, gRPC xDS |
| DNS-based | Service name resolves to instance or VIP addresses | Kubernetes CoreDNS, AWS Cloud Map, Consul DNS |
| Mesh-based | Sidecar or proxyless clients get endpoints from a control plane | Istio, Linkerd, Consul Connect, Cloud Service Mesh |
flowchart LR
subgraph K8s cluster
A[order-svc pod] -->|http://payment-svc| DNS[CoreDNS]
DNS --> SVC[payment-svc ClusterIP]
SVC --> P1[payment pod 1]
SVC --> P2[payment pod 2]
SVC --> P3[payment pod 3]
CP[Endpoints controller<br/>watches readiness] -.updates.-> SVC
endReal-world example
On Kubernetes, order-service calls http://payment-service.payments.svc.cluster.local. CoreDNS resolves the Service name; kube-proxy or the eBPF data plane (e.g. Cilium) routes to ready pods. Across clusters or to VMs, a mesh or AWS Cloud Map/Consul provides a shared registry.
Advantages and disadvantages
| Approach | Advantages | Disadvantages |
|---|---|---|
| Platform (K8s Service, LB) | No client libraries, language-neutral | Extra hop for some data planes; limited LB algorithms |
| Client-side registry | Smart routing, no extra hop | Library per language, registry is critical |
| Mesh | Rich traffic control, mTLS | Operational complexity, resource overhead |
Architect-level considerations
[!ARCHITECT] On Kubernetes, use native Services and DNS; adding Eureka on top is usually redundant. Use a registry or mesh when spanning clusters, VMs and multiple platforms.
Interview questions and answers
Is Eureka still needed with Kubernetes?
Generally no. Kubernetes provides discovery and load balancing through Services. Spring Cloud Kubernetes can integrate config and discovery if needed.
How does discovery interact with health checks?
Only ready instances are registered or receive traffic. Readiness probes in Kubernetes remove pods from Service endpoints; registries use heartbeats or health checks and deregister failed instances.