EndpointSlices and Service Topology

💡 Quick Answer: EndpointSlices replace Endpoints for clusters with 100+ pod backends. Enable topology-aware routing with service.kubernetes.io/topology-mode: Auto to prefer same-zone backends — reducing cross-zone data transfer costs by 60-80%.

The Problem

With the legacy Endpoints API, every endpoint change (pod scale, rolling update) sends the FULL endpoint list to every node. For a Service with 1000 pods, that’s 1000 endpoints × N nodes = massive etcd and network overhead. EndpointSlices split this into manageable chunks.

The Solution

EndpointSlices

apiVersion: discovery.k8s.io/v1
kind: EndpointSlice
metadata:
  name: web-frontend-abc12
  labels:
    kubernetes.io/service-name: web-frontend
addressType: IPv4
endpoints:
  - addresses: ["10.244.1.5"]
    conditions:
      ready: true
    zone: "eu-west-1a"
    nodeName: worker-1
  - addresses: ["10.244.2.8"]
    conditions:
      ready: true
    zone: "eu-west-1b"
    nodeName: worker-3
ports:
  - name: http
    port: 8080
    protocol: TCP

Each EndpointSlice holds up to 100 endpoints. A Service with 500 pods = 5 EndpointSlices.

Topology-Aware Routing

apiVersion: v1
kind: Service
metadata:
  name: web-frontend
  annotations:
    service.kubernetes.io/topology-mode: Auto
spec:
  selector:
    app: web-frontend
  ports:
    - port: 80
      targetPort: 8080

With topology-mode: Auto, kube-proxy routes traffic to same-zone endpoints when possible:

graph TD
    subgraph Zone A
        CLIENT[Client Pod<br/>Zone A] -->|Prefer local| BACKEND_A[Backend Pod<br/>Zone A]
    end
    subgraph Zone B
        BACKEND_B[Backend Pod<br/>Zone B]
    end
    subgraph Zone C
        BACKEND_C[Backend Pod<br/>Zone C]
    end
    
    CLIENT -.->|Fallback if Zone A<br/>has no ready pods| BACKEND_B
    CLIENT -.->|Fallback| BACKEND_C

Endpoints vs EndpointSlices Performance

Common Issues

Topology-aware routing sends all traffic to one zone

If zone capacity is imbalanced (10 pods in zone A, 2 in zone B), Kubernetes falls back to cluster-wide routing. Ensure even pod distribution with topology spread constraints.

EndpointSlice not updating — stale endpoints

Check EndpointSlice controller: kubectl get events -n kube-system | grep endpointslice. Ensure the Service selector matches pod labels exactly.

Best Practices

• EndpointSlices are the default since K8s 1.21 — no action needed for most clusters
• Enable topology-aware routing for multi-zone clusters — reduces cross-zone costs
• Combine with topology spread constraints — even distribution enables topology routing
• Monitor EndpointSlice count — too many slices (>50 per Service) indicates churn
• Legacy Endpoints API still works — but causes performance issues above ~300 pods per Service

Key Takeaways

• EndpointSlices are the scalable replacement for the Endpoints API
• Each slice holds up to 100 endpoints — updates are incremental, not full-list
• Topology-aware routing prefers same-zone backends — reduces cross-zone data transfer
• topology-mode: Auto falls back to cluster-wide if zone capacity is imbalanced
• Essential for large clusters (100+ pods per Service) — 10x less API server and etcd load