This section provides strategies and techniques to assist with broken model upgrades. See the main Upgrading models page for background information.
It may occur that an agent does not restart upon upgrade. One thing that may
help is the inspection and modification of its
agent.conf file. Comparing it
with its file before upgrading can be very useful.
Installing a different or modified configuration file will require a restart of the daemon. For example, for a machine with an ID of '2':
juju ssh 2 'ls -lh /etc/init/juju*'
This will return something similar to:
-rw-r--r-- 1 root root 506 Sep 2 00:57 /etc/init/jujud-machine-2.conf -rw-r--r-- 1 root root 533 Sep 2 00:57 /etc/init/jujud-unit-mysql-0.conf
Therefore, if the agent for machine '2' is not coming up you can connect to the machine in this way:
juju ssh 2
Then modify or restore the agent file
/var/lib/juju/agents/machine-2/agent.conf), and while still connected to the
machine, restart the agent:
sudo service jujud-machine-2 restart
The restart of an agent, due to invoking
upgrade-model or by manual means (as
above) may cause a hook for that particular unit/machine to be called. That can
sometimes lead to hook failures. Connect to that unit using the
command, see what is wrong, and retry the hook using the
juju debug-hooks etcd/2
In a different terminal retry the failed hook:
juju resolved etcd/2
See Debugging charm hooks for more information.
When the running agent software that is more than 1 patch point behind the targeted upgrade version the upgrade process will abort.
One very common reason for "agent version skew" is that during a previous upgrade the agent could not be contacted and, therefore, was not upgraded along with the rest of the agents.
For example, the following error message will be printed when attempting to upgrade from 2.2.1 to 2.2.2 when an agent is still running, say, 2.2.0:
ERROR some agents have not upgraded to the current model version 2.2.1: machine-0, unit-ubuntu-0
To overcome this situation you may force the upgrade by ignoring the agent version check:
juju upgrade-model --ignore-agent-versions
If an attempted upgrade results in failure it may prove difficult to return to a working setup and you may be compelled to start anew. Doing so will make the old controller completely inert and you should consider it a data loss situation.
Begin by removing the controller with
juju destroy-controller or
juju kill-controller. If this is insufficient you may need to ask Juju to
simply "forget" about the controller. This is done with the
Once the above is completed, a new controller can be created.