Charm does not always fulfil client requests

More often than not in a deployment I have charms related to keystone k8s (either by identity_service or identity_credentials) stuck having not received their credentials from keystone.

For example:
$ juju status
...
glance/0* waiting idle 10.1.214.150 (identity-service) integration incomplete
horizon/0* waiting idle 10.1.214.142 (workload) Not all relations are ready
keystone/0* active idle 10.1.214.154

This appears to be because the requests from the client were deferred and then the keystone pod was recreated by juju causing all outstanding deferred events to be lost:

$ juju debug-log -m openstack --replay | grep keystone | grep -E "pebble poll failed| run stop hook"
unit-keystone-0: 13:57:37 ERROR juju.worker.uniter pebble poll failed for container "keystone": failed to send pebble-ready event: terminated
unit-keystone-0: 13:57:37 DEBUG juju.worker.uniter.operation running operation run stop hook for keystone/0