Sometimes there are zombie placement allocations records left for
hypervisors, meaning there are no mapping VMs running on those hosts,
but the placement allocation records still show up in the database.
Does this mean that nova conductor/scheduler would still think those
resources have already been allocated, while in reality, there's no
existing VMs mapped to those zombie placement allocations records?
The scheduler will indeed consider the resources as existing if there are orphaned allocations. That said, nova has a 'nova-manage placement audit' command that looks at orphaned allocations, so you can use it for checking whether those allocations are wrong and then delete them if you want.
-S