k3s/docs/adrs/agent-join-token.md

3.7 KiB

Support kubeadm-style Bootstrap Token Secrets

Date: 2022-12-20

Status

Accepted

Context

K3s Token Types and Use

K3s currently supports two tokens that can be used to join nodes to the cluster:

  • --token: This is the default token, and a random value is generated during initial cluster startup if not specified by the user. This token is also used as the passphrase input to the PBKDF2 function used to generate the encryption key for cluster bootstrap data. For this reason, all server nodes must use the same token value once the cluster has been started, and the token value cannot be changed.
  • --agent-token: By default, this is set to the same as the --token value. If set, this token can be used to join new agents to the cluster, but not servers. This token value can be changed after the cluster has beens started, but doing so requires coordinatating reconfiguration and restart of all of the servers in the cluster.

Internally, these tokens are used as the password for HTTP Basic authentication to the K3s supervisor when the agent bootstraps its configuration and certificates. Servers use a username of server, while agents (including servers local agents) use node. Once nodes join the cluster they also populate a node password secret that prevents other nodes from using the same node name, but this is unrelated to the token.

Security Considerations

Users have requested the ability to generate single-use or limited-duration tokens that can be used to join nodes to the cluster, but can be deleted or automatically expire in order to reduce the impact should the token be compromised. Currently, compromise of the server token would require a complete rebuild of the cluster in order to use a new token. Compromise of the agent token would require a coordinated restart of all nodes in the cluster.

Existing Work

kubeadm includes a kubeadm token create command that creates secrets of type bootstrap.kubernetes.io/token, which is a core upstream type that is not restricted for use by kubeadm.

There are helpers for interacting with bootstrap token secrets in the k8s.io/cluster-bootstrap package, and upstream Kubernetes includes two controllers (tokencleaner and bootstrapsigner) to support use of cluster bootstrap secrets. The latter controller is not relevant for our use case, as it serves the same function as existing K3s supervisor routes - making initial cluster CA certificates and a client kubeconfig available for bootstrapping nodes. The boostrap-tokens documentation can be referenced for more information.

Decision

  • K3s will allow joining agents to the cluster using bootstrap token secrets.
  • K3s will NOT allow joining servers to the cluster using bootstrap token secrets.
  • K3s will include a k3s token subcommand that allows for token create/list/delete operations, similar to the the functionality offered by kubeadm.
  • K3s will enable the tokencleaner controller, in order to ensure that bootstrap token secrets are cleaned up when their TTL expires.
  • K3s agent bootstrap functionality will allow a agent to connect the cluster using existing Node Authorization to authenticate to the cluster during startup, even after its join token has been invalidated.
  • K3s agent bootstrap functionality will NOT allow an agent to connect to the cluster if it does not have a valid token, and its Node object has been deleted from the cluster.

Consequences

This will require additional documentation, CLI subcommands, and QA work to validate use of bootstrap token secret auth.