ha-mode.md

HA endpoints for K8s

The following components require a highly available endpoints:

• etcd cluster,
• kube-apiserver service instances.

The latter relies on a 3rd side reverse proxy, like Nginx or HAProxy, to achieve the same goal.

Etcd

The etcd clients (kube-api-masters) are configured with the list of all etcd peers. If the etcd-cluster has multiple instances, it's configured in HA already.

Kube-apiserver

K8s components require a loadbalancer to access the apiservers via a reverse proxy. Kubespray includes support for an nginx-based proxy that resides on each non-master Kubernetes node. This is referred to as localhost loadbalancing. It is less efficient than a dedicated load balancer because it creates extra health checks on the Kubernetes apiserver, but is more practical for scenarios where an external LB or virtual IP management is inconvenient. This option is configured by the variable loadbalancer_apiserver_localhost (defaults to True. Or False, if there is an external loadbalancer_apiserver defined). You may also define the port the local internal loadbalancer uses by changing, loadbalancer_apiserver_port. This defaults to the value of kube_apiserver_port. It is also important to note that Kubespray will only configure kubelet and kube-proxy on non-master nodes to use the local internal loadbalancer.

If you choose to NOT use the local internal loadbalancer, you will need to use the kube-vip ansible role or configure your own loadbalancer to achieve HA. By default, it only configures a non-HA endpoint, which points to the access_ip or IP address of the first server node in the kube_control_plane group. It can also configure clients to use endpoints for a given loadbalancer type. The following diagram shows how traffic to the apiserver is directed.

A user may opt to use an external loadbalancer (LB) instead. An external LB provides access for external clients, while the internal LB accepts client connections only to the localhost. Given a frontend VIP address and IP1, IP2 addresses of backends, here is an example configuration for a HAProxy service acting as an external LB:

listen kubernetes-apiserver-https
  bind <VIP>:8383
  mode tcp
  option log-health-checks
  timeout client 3h
  timeout server 3h
  server master1 <IP1>:6443 check check-ssl verify none inter 10000
  server master2 <IP2>:6443 check check-ssl verify none inter 10000
  balance roundrobin

Note: That's an example config managed elsewhere outside of Kubespray.

And the corresponding example global vars for such a "cluster-aware" external LB with the cluster API access modes configured in Kubespray:

apiserver_loadbalancer_domain_name: "my-apiserver-lb.example.com"
loadbalancer_apiserver:
  address: <VIP>
  port: 8383

Note: The default kubernetes apiserver configuration binds to all interfaces, so you will need to use a different port for the vip from that the API is listening on, or set the kube_apiserver_bind_address so that the API only listens on a specific interface (to avoid conflict with haproxy binding the port on the VIP address)

This domain name, or default "lb-apiserver.kubernetes.local", will be inserted into the /etc/hosts file of all servers in the k8s_cluster group and wired into the generated self-signed TLS/SSL certificates as well. Note that the HAProxy service should as well be HA and requires a VIP management, which is out of scope of this doc.

There is a special case for an internal and an externally configured (not with Kubespray) LB used simultaneously. Keep in mind that the cluster is not aware of such an external LB and you need no to specify any configuration variables for it.

Note: TLS/SSL termination for externally accessed API endpoints' will not be covered by Kubespray for that case. Make sure your external LB provides it. Alternatively you may specify an externally load balanced VIPs in the supplementary_addresses_in_ssl_keys list. Then, kubespray will add them into the generated cluster certificates as well.

Aside of that specific case, the loadbalancer_apiserver considered mutually exclusive to loadbalancer_apiserver_localhost.