Kubernetes
The standard heavyweight orchestrator. Powerful when warranted.
Kubernetes is no longer the default choice for new infra in 2026 — but it's still the right answer when you have multiple teams shipping to one platform, or genuinely need its primitives.
What you actually need to run k8s
A real production cluster has ~15 moving parts:
- Distribution: k3s, k0s, kubeadm, Talos, EKS, GKE, AKS, DigitalOcean Kubernetes, etc.
- Ingress: ingress-nginx, Traefik, Contour, Cilium Gateway.
- Cert management: cert-manager.
- External DNS sync: ExternalDNS.
- CNI: Cilium, Calico, Flannel.
- CSI / storage: Longhorn, Rook-Ceph, OpenEBS, EBS / GCE PD CSI.
- Secrets: External Secrets Operator + a backend.
- GitOps / CD: ArgoCD or Flux.
- Observability: Prometheus + Grafana + Loki + Tempo.
- Service mesh (optional): Istio, Linkerd, Cilium.
- Policy: Kyverno or OPA Gatekeeper.
- DB operators: CloudNativePG for Postgres.
- Backup: Velero.
- Cost: Kubecost or OpenCost.
- Local dev: Kind, k3d, Tilt, Skaffold.
Hosted vs. self-managed
- ★ Hosted control plane (EKS, GKE, AKS, DigitalOcean K8s, Linode LKE, OVH Kube) — pay $70/mo for the control plane, bring your own nodes (or theirs).
- ★ k3s on your own boxes — single-binary "Kubernetes lite" that's actually full Kubernetes underneath. Run on Hetzner for ~$30/mo total.
- kubeadm — the official "build your own cluster" path. More steps; you learn a lot.
- Talos Linux — immutable, API-driven, no SSH; the modern self-host pick.
- Managed-but-not-cloud providers (Civo, Vultr Kubernetes, Scaleway Kapsule) — generous free tiers, often.
When k8s is right
- Multiple teams shipping multiple services on one platform.
- You need actual scheduling primitives (bin-packing across nodes, taints/tolerations, etc.).
- Many stateful workloads with operators (Postgres, Kafka, Elasticsearch, Cassandra).
- Compliance frameworks that benefit from declarative-everything.
- You want portable workloads across cloud providers.
When k8s is wrong
- Solo founder / small team — the ops tax is brutal.
- Few services, low ops capacity — Coolify or Kamal will get you 90% of the value.
- "We might need k8s in two years" — wait until you do.
- Hot path code that wants bare-metal performance — Nomad or direct VMs may be simpler.
Common modern stack
A typical 2026 production cluster:
- Talos Linux for nodes (immutable, declarative).
- Cilium as CNI + Gateway API ingress (replaces ingress-nginx + Calico + kube-proxy).
- ArgoCD with App-of-Apps + Renovate-driven Helm chart updates.
- CloudNativePG for Postgres.
- External Secrets + Vault or Doppler for secrets.
- Prometheus + Grafana + Loki + Tempo (kube-prometheus-stack chart).
- cert-manager + ExternalDNS for TLS + DNS.
- Velero for backups.
- Kyverno for policy.
You can stand this up in a weekend; you can also spend two months on it. Plan accordingly.
Pick this if…
- Multiple teams shipping to a shared platform.
- You actually need autoscaling, taints, anti-affinity, etc.
- Many stateful operators in scope (Kafka, ES, Postgres clusters, etc.).
- Compliance / portability matters more than ops cost.
Otherwise consider self-host PaaS, self-host fleet, or managed cloud first.