CKS: nothing detects drift between a cluster's VM map, its VMs and its Kubernetes nodes #14665

problem

CKS keeps the list of VMs of each Kubernetes cluster in kubernetes_cluster_vm_map. Scale, upgrade, stop/start and the cluster status scanner all trust that list. Nothing checks it against the VMs that actually exist or against the nodes Kubernetes knows about, so when the two drift apart nobody is told:

• a VM that is in the map but has been destroyed or expunged, or no longer exists, keeps the cluster in Alert (the scanner compares the map's size and VM states against the expected node count), and the next scale or upgrade works from a wrong list;
• a VM that runs as a node of the cluster but is not in the map is invisible to CloudStack: it is not scaled, upgraded, stopped or deleted with the cluster;
• a Kubernetes node with no VM behind it, or a mapped running VM that is not a Kubernetes node, means the cluster Kubernetes runs is not the one CloudStack thinks it manages.

The API and UI refuse to destroy a VM that belongs to a CKS cluster, so the map does not drift through normal use. It drifts when:

• a node is rebuilt or replaced by hand (a new VM joined to the cluster, the old one removed outside CKS);
• the database is edited by hand;
• a destroy or expunge fails during a scale-down, which removed the VM from the map but could leave the VM behind. That path is fixed in CKS: fix node drain on scale-down and upgrade, make a failed upgrade resumable #14611.

A production cluster was found sitting in Alert with one of its control nodes missing from the map. The only way to find out why was to compare the map, the VMs and kubectl get nodes by hand.

versions

4.22 and main.

The steps to reproduce the bug

1. Create a CKS cluster and wait for it to be Running.
2. Remove one of its rows from kubernetes_cluster_vm_map (or replace a worker node by hand with a new VM that joins the cluster).
3. The cluster moves to Alert, or keeps a node CloudStack does not know about, and no alert or event says which VM or node is the cause.

What to do about it?

Have the cluster status scanner compare the map with the cluster's VMs and its Kubernetes nodes, and raise an alert and an event naming the cluster and the VM or node for each mismatch, once per mismatch. Detection only, the map is not changed.