Kubernetes for Skeptics: When You Actually Need It

Kubernetes is the platform every CTO eventually considers. It is also, for most companies, overkill. Here is how to decide honestly.

What Kubernetes actually buys you

1. Declarative orchestration — you describe the desired state; the controller reconciles.
2. Self-healing — failed pods restart, failed nodes get drained.
3. Horizontal scale — replicas, autoscalers, multi-zone topology.
4. A standard API — every cloud, every on-prem provider speaks K8s.
5. A massive ecosystem — Helm, Argo, Istio, Knative, KEDA, cert-manager.

What it costs

• A dedicated platform engineer (or 0.5 FTE minimum).
• 3 control-plane nodes you don’t run business logic on.
• A cluster upgrade every 4 months that requires planning.
• A 200-page learning curve before anyone is productive.

The decision matrix

You have…	Use
1–10 services, one team	Fly.io, Render, Railway, ECS Fargate, Cloud Run
10–50 services, multiple teams, multi-region	Kubernetes (managed: EKS, GKE, AKS)
Stateful workloads with operators (Postgres, Kafka)	Kubernetes
One monolith + a few jobs	Docker Compose on a beefy VM
Strict compliance (FedRAMP, on-prem)	Kubernetes (it is the standard)

Lighter alternatives worth knowing

• Nomad — same scheduling primitives without the API surface area; pairs well with Consul.
• Fly.io / Railway / Render — “scale to zero PaaS”; you write a Dockerfile, they handle the rest.
• Coolify / Dokploy — self-hosted PaaS layer on a single VM.
• AWS ECS + Fargate — Kubernetes-shaped without the control plane.

If you do go Kubernetes

• Start with a managed cluster. Always.
• Use GitOps from day one (ArgoCD or Flux).
• Adopt Kustomize or Helm, not both.
• Run policy-as-code (Kyverno, OPA) before anyone deploys to prod.
• Budget for observability — logs, metrics, traces, costs.

TL;DR

Kubernetes is a great answer to “we have many teams shipping many services”. For everything else, simpler is faster, cheaper and on-call-friendlier.