Track control plane join race conditions via kubeadm

[chuckha]

Kubeadm has a bug with joining multiple control plane nodes simultaneously. Occasionally, the control plane will fail to join because not enough etcd members are ready.

We can work around this by setting a concurrency limit of 1 to the controller that is responsible for joining the control plane nodes (usually the infrastructure controller). This fixes the issue by making joining a node is a blocking operation so only one node can join at a time. If users set a concurrency limit > 1 then there is about a 20% chance failure of a control plane join failing when two or more control planes attempt to join a cluster simultaneously.

CAPI is able to work around this by retrying the join after failure, but there will be ominous logs and a slow down in how long it takes all control plane nodes to become ready.

Let's track the work kubernetes/kubeadm#1793 that will improve simultaneous control plane joining.

/kind bug
/milestone Next
/priority backlog

[vincepri]

Would #2016 also be a potential fix to this issue?

[chuckha]

Hm yes, if folks aren't using the kubeadm control plane then they wouldn't run into this issue!

/close

This is a duplicate of #2016

[neolit123]

Kubeadm has a bug with joining multiple control plane nodes simultaneously. Occasionally, the control plane will fail to join because not enough etcd members are ready.

technically the problem is in etcd.