Skip to main content

High Availability

Kube-Hetzner is designed with high availability in mind. By default, it creates clusters with multiple control plane nodes to ensure no single point of failure.

Default HA Configuration​

The default configuration includes:

  • 3 Control Plane Nodes: Provides etcd quorum and API server redundancy
  • 2 Agent Nodes: Ensures application availability
  • Load Balancer: Distributes traffic across control plane nodes
  • Automatic Failover: Built-in redundancy for all critical components

Etcd Quorum​

Kubernetes uses etcd for storing cluster state. For etcd to function properly, it requires a quorum (majority) of nodes to be available. With 3 control plane nodes:

  • Can tolerate 1 node failure and maintain quorum
  • Requires an odd number of nodes (3, 5, 7, etc.)
  • Never use an even number of control plane nodes

Control Plane Load Balancing​

Kube-Hetzner automatically creates a Hetzner Load Balancer that distributes traffic across all control plane nodes. This ensures:

  • No single point of failure for API access
  • Even distribution of API requests
  • Automatic removal of unhealthy control plane nodes

Automatic Upgrades and Reboots​

One of the key features of Kube-Hetzner is automatic OS and Kubernetes upgrades. These are performed in a rolling fashion to maintain availability:

  1. Nodes are drained of workloads
  2. Node is upgraded/rebooted
  3. Node is restored to service
  4. Process repeats for next node

This ensures that only one node is down at any given time, maintaining cluster availability.

Recommendations for HA​

Minimum Configuration​

For production workloads, we recommend at least:

  • 3 Control Plane Nodes
  • 2 Agent Nodes
  • Automatic upgrades enabled

Scaling for Higher Availability​

For mission-critical applications:

  • 5 Control Plane Nodes (can tolerate 2 failures)
  • Multiple Agent Nodepools across different locations
  • Enable autoscaling for automatic capacity adjustment

Disabling HA​

If you need to reduce costs for development or testing, you can create a single-node cluster:

# In your kube.tf
control_plane_nodepools = [
  {
    name        = "control-plane",
    server_type = "cx22",
    location    = "nbg1",
    count       = 1
  }
]

agent_nodepools = [
  {
    name        = "agent",
    server_type = "cx22",
    location    = "nbg1",
    count       = 1
  }
]

Note that single-node clusters:

  • Do not provide high availability
  • Should have automatic OS upgrades disabled
  • Use the built-in k3s service load balancer instead of Hetzner LB