When using openstack where there are several separately routed floating ip pools, the pool has to be specified in order to get a floating IP belonging to the same availability zone as the instance it will be assigned to.
When no pool is specified, one seems to be chosen at random.
Juju has no (documented?) way of specifying this, and it does not match the availability zone to the pool internally.
Juju shows (unless you happen to get an IP from the right pool by chance):
agent-state: error
agent-state-info: |-
cannot assign public address <floating ip> to instance "<instance id>": failed to add floating ip <floating ip> to server with id: <instance id>
caused by: request (https://keystone/servers/<instance id>/action) returned unexpected status: 400; error info: {"badRequest": {"message": "Unable to associate floating ip <floating ip> to fixed ip <fixed ip> for instance <instance id>. Error: External network <network id> is not reachable from subnet <subnet id>. Therefore, cannot associate Port <port id> with a Floating IP.", "code": 400}}
This is the same as trying to associate it manually:
External network <network id> is not reachable from subnet <subnet id>. Therefore, cannot associate Port <port id> with a Floating IP.
As a workaround, it is possible to allocate an IP from the right pool manually before deploying with juju. Juju uses this automatically.
This is related to, but not solved by bug #1183831.