Running several replicas only helps if they don't all fail together. Pod anti-affinity keeps Pods away from Pods with matching labels within a topology domain named by a node label: kubernetes.io/hostname means one per node, topology.kubernetes.io/zone one per zone. Required anti-affinity caps replicas at the number of domains; preferred anti-affinity spreads as far as it can and then doubles up. Pod affinity does the opposite and pulls Pods together, for example a cache beside its app.
Topology spread constraints express "as even as possible": maxSkew is the largest allowed difference in matching Pods between domains, and whenUnsatisfiable is DoNotSchedule (hard) or ScheduleAnyway (soft).
When nodes are full, priority decides who runs. A PriorityClass is a cluster-wide number; Pods name it in priorityClassName, higher values are scheduled first and can preempt lower ones, and a Pod naming a class that doesn't exist is refused.
No node more than one replica ahead of another.