> ## Documentation Index
> Fetch the complete documentation index at: https://docs.rafftechnologies.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Kubernetes control plane and HA

> The managed control plane, what HA adds, and how we keep etcd fast

<sub>Updated August 24, 2026</sub>

The control plane — API server, scheduler, controllers, etcd — is run by Raff. The single-master control plane is **free**; the HA control plane (3 masters) costs a flat **\$30/month**. Workers and storage nodes are billed separately.

## Standard vs HA

|                                           | Standard                               | HA (\$30/month)                                          |
| ----------------------------------------- | -------------------------------------- | -------------------------------------------------------- |
| Control-plane nodes                       | 1                                      | 3 (replicated etcd, quorum)                              |
| API availability through a master failure | No                                     | Yes                                                      |
| Gateways                                  | 1 (restarted automatically on failure) | 2 with automatic IP failover (VRRP)                      |
| Endpoint-path SLA                         | Infrastructure (VM) level              | [99.90% monthly](/products/build/kubernetes/details/sla) |
| Recommended for                           | Dev, staging                           | Production                                               |

You can create a cluster with HA or [upgrade later](/products/build/kubernetes/quickstart-guides/upgrade-to-ha); upgrades are not reversible. HA covers VM- and host-level failures within the region. For production: enable HA, run workers across at least 2–3 nodes, and use `ClusterIP` Services behind [Traefik ingress](/products/build/kubernetes/quickstart-guides/expose-a-service) — a fully redundant ingress path with no standalone load balancer.

## etcd on local NVMe

etcd is the one latency-critical component in Kubernetes: every write waits for an fsync. Raff places control-plane disks on **hypervisor-local NVMe** rather than network storage, keeping etcd fsync latency in single-digit milliseconds. The dashboard's control-plane health card shows the live etcd latency p99 and 24-hour restart counts for your cluster — the same signals our own alerting watches.