Observed during testing of candidate-4.15 image as of 2024-02-08.

This is an incomplete report as I haven't verified the reproducer yet or attempted to get a must-gather. I have observed this multiple times now, so I am confident it's a thing. I can't be confident that the procedure described here reliably reproduces it, or that all the described steps are required.

I have been using MCO to apply machine config to masters. This involves a rolling reboot of all masters.

During a rolling reboot I applied an update to CPMS. I observed the following sequence of events:

• master-1 was NotReady as it was rebooting
• I modified CPMS
• CPMS immediately started provisioning a new master-0
• CPMS immediately started deleting master-1
• CPMS started provisioning a new master-1

At this point there were only 2 nodes in the cluster:

• old master-0
• old master-2

and machines provisioning:

• new master-0
• new master-1