Skip to main content

Architecture

Servers and Agents​

  • A server node is defined as a host running the k3s server command, with control-plane and datastore components managed by K3s.
  • An agent node is defined as a host running the k3s agent command, without any datastore or control-plane components.
  • Both servers and agents run the kubelet, container runtime, and CNI. See the Advanced Options documentation for more information on running agentless servers.

Single-server Setup with an Embedded DB​

The following diagram shows an example of a cluster that has a single-node K3s server with an embedded SQLite database.

In this configuration, each agent node is registered to the same server node. A K3s user can manipulate Kubernetes resources by calling the K3s API on the server node.

K3s Architecture with a Single ServerK3s Architecture with a Single Server

High-Availability K3s​

Single server clusters can meet a variety of use cases, but for environments where uptime of the Kubernetes control plane is critical, you can run K3s in an HA configuration. An HA K3s cluster comprises:

  • Three or more server nodes that will serve the Kubernetes API and run other control plane services
  • An embedded etcd datastore (as opposed to the embedded SQLite datastore used in single-server setups)
K3s Architecture with High-availability ServersK3s Architecture with High-availability Servers

Fixed Registration Address for Agent Nodes​

In the high-availability server configuration, each node can also register with the Kubernetes API by using a fixed registration address, as shown in the diagram below.

After registration, the agent nodes establish a connection directly to one of the server nodes.

Agent Registration HAAgent Registration HA

How Agent Node Registration Works​

Agent load-balancer​

Agent nodes are registered with a websocket connection initiated by the k3s agent process, and the connection is maintained by a client-side load balancer running as part of the agent process. Initially, the agent connects to the supervisor (and kube-apiserver) via the local load-balancer on port 6443. The load-balancer maintains a dynamic list of server endpoints, and uses these as the back-end target pool for the local load-balancer. The default (and initially only) endpoint is seeded by the hostname from the --server address. Once connected to the cluster, the agent retrieves a list of kube-apiserver addresses from the Kubernetes service endpoint list in the default namespace. Endpoint addresses are synced to the load-balancer, which then maintains stable connections to all servers in the cluster, providing a connection to the kube-apiserver that tolerates outages of individual servers.

Node-password secrets​

Agents register to the cluster using the the join token as credentials, along with a randomly generated password for the node. The agent stores its selected password locally at /etc/rancher/node/password, and the cluster stores a hash of the password in a Kubernetes secret. Node password secrets are stored in the kube-system namespace with names using the template <node-name>.node-password.k3s. In order to protect the integrity of node identities, any subsequent attempts to re-register or obtain certificates for a node that has a node password secret must use the same password.

Creation of node password secrets fails open, to allow recovery from outages affecting validating webhooks. Nodes must be allowed to join even during an outage so that pods for the webhook can be started. Once the outage is resolved, creation of node password secrets is resumed.

Node password secrets are deleted when their owning node resource is deleted. Node password secrets that never had an owning node resource created are periodically garbage-collected.

If the /etc/rancher/node directory of an agent is removed, or you wish to rejoin a node using an existing name, the node should be deleted from the cluster. This will clean up both the old node entry, and the node password secret, and allow the node to (re)join the cluster.

If you frequently reuse hostnames, but are unable to remove the node password secrets, a unique node ID can be automatically appended to the hostname by launching K3s servers or agents using the --with-node-id flag. When enabled, the node ID is also stored in /etc/rancher/node/.