Kubernetes in 2026 is mature, opinionated, and surprisingly stable. The frontier moved on (WASM, agentic infra) — but Kubernetes is still where most real workloads live. This post is a quick state-of-affairs.

What’s the same

  • Pod, Deployment, Service are still the atoms.
  • Helm still dominates packaging.
  • kubectl is still your terminal.
  • etcd still backs control-plane state.
  • CRDs still extend the API.

If you learned Kubernetes in 2020, your knowledge mostly transfers.

What changed

Cilium default

EKS, GKE, AKS all default to Cilium as the CNI now. eBPF-based networking + service mesh without sidecars. See Cilium and eBPF in Production .

Gateway API replaces Ingress

apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata: { name: api }
spec:
  parentRefs: [{ name: external }]
  rules:
    - matches: [{ path: { type: PathPrefix, value: /api } }]
      backendRefs: [{ name: api-service, port: 80 }]

Better expressivity, proper TLS / TCP routing, multi-tenant gateway sharing. Ingress is in maintenance.

Sidecarless meshes

Istio Ambient and Cilium Service Mesh removed sidecars. Memory savings: 50–100MB per pod. Operational: simpler.

KEDA for event-driven scaling

apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata: { name: worker-scaler }
spec:
  scaleTargetRef: { name: worker }
  triggers:
    - type: kafka
      metadata: { topic: jobs, lagThreshold: "100" }

Scale on queue depth, Kafka lag, Prometheus metric — beyond CPU. KEDA is the standard.

KCP — multi-tenant control planes

For platform teams: KCP virtualizes the K8s API so you can give every team their own logical control plane on shared compute.

eBPF observability

Pixie, Tetragon, Cilium Hubble — the trio that gives you traces, security, network observability without app instrumentation.

What I’d ship today

For a typical 2026 startup adopting Kubernetes:

  1. Managed Kubernetes — pick one of EKS / GKE / AKS, default Cilium.
  2. Gateway API for ingress.
  3. Argo CD for GitOps .
  4. External Secrets Operator for secrets.
  5. KEDA for scaling.
  6. Cilium Hubble + Pixie for observability.
  7. Cert-manager for TLS.
  8. Velero for backups.

When NOT to use Kubernetes

  • Very small (<5 services) — Fly.io, Render, Railway are simpler.
  • Very large stateful (databases) — managed services often better.
  • Edge / serverless — Cloudflare Workers, Lambda, etc. better fit. See Cloudflare Workers + D1 + Durable Objects .

Operational realities

  • Kubernetes is a job description. Don’t expect part-time admin to keep it healthy.
  • Costs surprise. Right-size requests/limits aggressively.
  • Upgrades are real work. Plan major-version upgrades quarterly.
  • Security has many layers — RBAC, network policies, pod security, admission control. Use them all.

Read this next

If you want a 2026 K8s starter (managed cluster + Argo + Cilium + observability), it’s at rajpoot.dev .

Building something AI-, backend-, or data-heavy and want a second pair of eyes? I do consulting and freelance work — see my projects and ways to reach me at rajpoot.dev .