Ceilometer-agent-compute service not running after Scale out of nova-cloud-controller application

On a deployment all but 2 (out of 114 units) of the ceilometer agent units are blocked in a state with "Services not running that should be: ceilometer-agent-compute".

This is mostly due to nova-compute.service is dependency of the ceilometer-agent.service which are not running or being restarted frequently.

The workaround was simply to run a systemctl restart ceilometer-agent-compute.service on the affected units.

System is ussuri/focal, charm revs.:
 cs:ceilometer-282, cs:ceilometer-agent-271
cs:~openstack-charmers-next/nova-cloud-controller-549
cs:nova-compute-334
(FCE template 21.10)

Event flow on one machine:
 - ceilometer-agent-compute.service has been restarted 40 times within 5 minutes after running juju add-unit nova-cloud-controller --to 24:lxd
 - last entry in this journalctl ceilometer-agent-compute.service is Dependency failed for Ceilometer Agent Compute (at 17:26:58)
 - nova-compute-service is full with Failed with result 'exit-code', Scheduled restart job, restart counter is at 35 (starting from 1). This series of service restart retries span the time of the ceilometer-agent.service is being restarted. At 17:26:58 there is an Failed to start Openstack Compute. event logged, but then at 17:30:13, next event is Started Openstack Compute. At this time the ceilometer-agent is not being started anymore and resulting stuck failed/stopped.
 - at between 17:30:04 and 17:30:16 the nova-compute juju application unit's cloud-compute:483 relation is being executed and finished. Unit is ready.

But there is no communication between the nova-compute unit and the ceilometer-agent juju units, hence the ceilometer-agent is never started back again and stuck stopped.

Getting out logs from the systems are hard, it is a secured environment going through locked down jump hosts.