Question about k3s agent load balancer

[aisensiy]

There are articles talking about the embedded etcd HA solution of k3s like this. One of the key behavior is adding a load balancer solution (EIP like this article or LB from the clound provider) between the agents and masters:

k3s agent --> load balancer --> master

And the architecture of k3s also show that a Fixed Registration Address is necessary.

While, after some research I found that k3s (at least v1.21.5+k3s2) have a internal agent load balancer (config at /var/lib/rancher/k3s/agent/etc/k3s-agent-load-balancer.yaml) which will auto update the master k8s api server list in it. So the out side load balancer is unnecessary?

And I also do a experiment like this:

1. Create a k3s cluster with 3 master nodes (master0 master1 master2) and 1 worker (worker0), and the worker agent joined the cluster from the master0 node.
2. Create a ssh session (session0) to worker0 with command journal -u k3s-agent -f to watch the log going.
3. Create a ssh session (session1) to master0 and run command systemctl stop k3s, then session0 (from the log tail) will show that the master0 is down and the agent load balancer is updated.
4. Create a ssh session (session2) to worker0 and run command cat /var/lib/rancher/k3s/agent/etc/k3s-agent-load-balancer.yaml will show there are only two records of master api endpoints over there.
5. Run command systemctl start k3s in session1 (the master0), the log in session0 will show that Agent load balancer updated and run command cat /var/lib/rancher/k3s/agent/etc/k3s-agent-load-balancer.yaml again in session2 there will show three ServerAddresses.

This behavior is so cool but I don't see any docs in k3s talking about this. So...do I misunderstand this behavior or there is little information about this behavior so that people still using a lb between agent and master?

[brandond]

So the out side load balancer is unnecessary?

Our documentation lists a requirement for a "fixed registration endpoint" so that nodes do not rely on a single server being online in order to join the cluster. This endpoint could be a load-balancer or a DNS alias, it's up to you. This is only needed when nodes are registering to the cluster; once they have successfully joined, they use the client load-balancer to communicate directly with the servers without going through the registration endpoint.

[aisensiy]

This is a very clear and helpful reply. 👍