Service Discovery in the Mesh

The Mesh as Discovery Consumer

 meshes don't replace discovery platforms — they CONSUME them:

 kubernetes: watches Services/Endpoints/EndpointSlices
   → builds its internal service registry
 cloud/VM estates: integrate via registry adapters or
   WorkloadEntries (declared external endpoints)

 flow:
 [k8s API: pods change]──watch──►[control plane]
                                      │ computes config (EDS)
                                      ▼
                        all proxies updated in seconds

 proxies then load-balance across CURRENT healthy endpoints
 without DNS round-robin staleness problems.

What Mesh Discovery Adds Over Plain K8s

 □ UNIFORM ACROSS CLUSTERS/CLOUDS: one namespace view
   spanning k8s + VMs + other clusters (federation)
 □ HEALTH-AWARE LB built-in: passive outlier ejection +
   active health checks beyond kubelet probes
 □ LOCALITY AWARENESS: region/zone/subzone routing priority:
     same-zone first → same-region fallback → anywhere
   (latency AND cross-zone cost optimization, declaratively)
 □ RICH SUBSETS: route to version-labeled endpoint groups
   (v1/v2 splits are discovery-level, not app-level)

Multi-Cluster Federation Patterns

 three trust/topology models:

 SINGLE network: clusters share one flat network;
   endpoints directly routable. simplest.
 MULTI-network: east-west gateways bridge clusters;
   traffic crosses via tunnels; identity preserved.
 CENTRALIZED control plane watching many remote APIs vs
 EACH cluster with local control plane + trust exchange:
   
   scale/intuition: local control planes per cluster +
   trust bundles exchanged = fewer catastrophic couplings.

 split-horizon gotchas worth naming:
 - service NAME collisions across clusters
 - locality labels must be truthful or zone-affinity lies
 - endpoint propagation lag between clusters (seconds)

Need	Mesh feature
VMs and k8s one mesh	workload entries / auto-registration
Zone-cost reduction	locality-aware LB weighting
Blue/green at fleet level	subsets + routing
Cross-region DR	failover priority configuration

Interview Framing

“How does your 3-cluster mesh keep proxy state current?” scored shape: platform-watch→control-plane→xDS-push pipeline drawn, federation model chosen with reasoning (multi-network + local control planes for scale), locality-aware routing named with WHY (latency+cost), subset-based canaries tied back to discovery. Mesh-discovery questions test whether you see the mesh as a distributed SYSTEM over registries rather than a magic box.