Having several machineSets with taints configured and one of them does not have any taint.

Run a workload without toletarions (to prevent it from running on the nodes of the rest of the machineSets) and only scheduling possibilities are the "no-taints" machineSet.

When you need more pods and the new ones appears as Pending, the clusterAutoscaler scale-up the wrong machineSet until new nodes are running, trigering a new scale-up activity until a valid node is ready (randomly??)

ROSA 4.12.35

Always

4.12 ROSA Cluster
Additional machineSet with taints
A deployment (no nodeSelector, no tolerations) with high resource request
Scale the deployment to have some pending pods

The AutoScaler scale-up the additional machineSet and the new node can't be used by your workload 

The pending pod can run after some scale-up activities, generating unneeded nodes

Scale-up the right machineSet where the pod will be able to be scheduled