Kubernetes on Indian infrastructure: when you need it, when you don't

Kubernetes is the most popular container orchestrator in the world. It is also the most over-applied. Most workloads do not need it. Here is a practical decision tree for engineering teams evaluating K8s in 2026.

When you need K8s

• You run 5+ stateless services that need to scale independently based on traffic (e.g. API tier + worker tier + cron tier).
• You have a multi-team org and need namespace isolation
  • per-team resource quotas.
• You do blue-green + canary deploys and need the declarative deployment object (Deployment, Service, Ingress) to manage them.
• You need to autoscale on a metric (CPU, memory, queue depth) and the cloud-provider autoscaler doesn’t fit.

When you don’t

• You run 1-2 long-running services. Just run them on Apps T1-T10 and skip the orchestration layer.
• You have a single team. A docker-compose file + a reverse proxy is enough.
• You do rolling deploys. A docker run on the next instance is enough.
• Your traffic is steady (not bursty). Static capacity beats elastic capacity when the curve is flat.

What it costs on Indian infrastructure

K8s is free; the cluster it runs on is not. The cost ranges we see in 2026:

• Small dev cluster (1 control plane + 2 workers, 4 vCPU + 8 GB each): Rs 12,000-15,000 / month on a Linux Dedicated L2 (which gives you 6 vCPU + 16 GB). You over-provision by ~50%, which is the right tradeoff for a dev cluster.
• Production cluster (3 control planes + 4 workers, 8 vCPU + 32 GB each): Rs 80,000-1,00,000 / month on 2-3x Linux Dedicated L3 instances.
• High-availability cluster (3 control planes + 6 workers, 16 vCPU + 64 GB each): Rs 2,00,000-2,50,000 / month on 3x Linux Dedicated L4 + a load balancer.

When you should use a managed K8s service instead

If you don’t have a platform team that can run a K8s control plane 24x7, don’t run your own K8s. Use a managed service (EKS, GKE, AKS) or our upcoming managed K8s offering (Qs2 2026). The cost of the control plane engineer is greater than the cost of the cluster.

If your workload fits in 1-3 containers, run it on Apps T1-T10 and use our auto-restart + auto-backup. The K8s tax (learning curve, control plane, observability) is not worth it for a single-digit-container workload.