kill-controller removes machines from migrated model

[Nate Finch]

(Now monitored at: http://reports.vapour.ws/releases/issue/5893ebba749a565681529d1e)

migrate a model from lxd1 to lxd2
kill-controller lx1 -yt 0

machines in the migrated model get taken down with kill-controller

Not sure if this is unique to the lxd provider.

Note that I had previously tried to migrate the model back to lxd1 which failed, saying "ERROR target prechecks failed: model is being migrated out of target controller" (a known bug), which is why I killed lxd1. Not sure if that matters.

[Curtis Hovey (sinzui)]

Given that the model-migration test is take a long time to delete all the models that were created, we know we want to delete unneeded models as soon as they are unneeded. This issues shows that we should not only delete unused models, but verify only the old model is deleted; the new model is kept.

[Christian Muirhead (2-xtian)]

This is happening because in LXD and some other providers we tag the instances with the controller UUID, and when destroying the controller at the provider level we check all instances to see if they still refer to the controller and if they do we destroy them too. The migration code isn't updating the tags on migrated instances to point to the new controller.

The providers that do track instance -> controller are: azure, ec2, gce, lxd, maas and openstack.
I'm adding a MoveInstancesToController method to each provider's environ that will update the tags for the migrated machines.

I'll add the method to the Environ interface once all of the implementations are done, and then update the migrationmaster worker to call it when the migration minions switch over.

[Christian Muirhead (2-xtian)]

The GCE change turned out to be bigger than I expected, so I'm going to start doing each of the providers as separate PRs.

[Christian Muirhead (2-xtian)]

Note for myself - the rackspace provider is a thin wrapper around the openstack one, so it doesn't need the new method added (although the fake rackspace provider used in tests still does).

[Christian Muirhead (2-xtian)]

PR to add a do-nothing implementation of the new method to controllers that don't track instance->controller tags: https://github.com/juju/juju/pull/6810
PR for gce provider: https://github.com/juju/juju/pull/6813

[Christian Muirhead (2-xtian)]

PR for lxd: https://github.com/juju/juju/pull/6827

axw suggests AdoptInstances as a better name than MoveInstancesToController, so I'll do that.

[Christian Muirhead (2-xtian)]

PR for maas: https://github.com/juju/juju/pull/6829

[Christian Muirhead (2-xtian)]

PR for azure: https://github.com/juju/juju/pull/6868

[Christian Muirhead (2-xtian)]

PR for ec2: https://github.com/juju/juju/pull/6900

[Christian Muirhead (2-xtian)]

PR for openstack: https://github.com/juju/juju/pull/6943

[Christian Muirhead (2-xtian)]

PR to add AdoptResources to the migrationtarget facade: https://github.com/juju/juju/pull/6950

[Christian Muirhead (2-xtian)]

PR to call AdoptResources from the migrationmaster worker: https://github.com/juju/juju/pull/6952

[Christian Muirhead (2-xtian)]

I've just seen a bug in the azure provider change - it updates the controller tags for everything except the resource group. Fixing it now.

[Christian Muirhead (2-xtian)]

PR for azure fix: https://github.com/juju/juju/pull/6961

[Anastasia (anastasia-macmood)]

Final fix committed to develop (2.2) as part of the merge https://github.com/juju/juju/pull/6968