Multi node support

[flaeppe]

Backwards compatibility

The changes here are not backwards compatible in the (odd?) case of previously having had this playbook run on multiple hosts.
We'd now expect a hosts file to refer to any nodes involved in a single swarm.

Backwards compatibility can be achieved, when above is false, by declaring any hosts deployed with the previous version under both docker_swarm_managers and docker_swarm_workers:

e.g.

Previous hosts file:

[managers]
host1

New hosts file:

[managers]
host1

[docker_swarm_managers]
host1

[docker_swarm_workers]
host1

As far as I see, it should only be necessary to change the hosts file.

Additional info

• The swarm.yml task set asserts that there's an odd number of managers declared in the docker_swarm_managers group.

• Node labels can be declared via the docker_swarm_labels (per host/group/etc) if wanted

• Three new (default) variables: docker_swarm_interface, docker_swarm_addr and docker_swarm_port have been added

• Would it be overkill to add a guard that validates that the worker group is non empty if all managers are declared as "manager only" nodes?

Some drawbacks

Demoting a manager node

As far as I could dig, I found no reasonable task/steps to let the playbook demote a manager (i.e. removing host from the manager group while having it in the worker group). Although handling that here could be quite an edge case, or perhaps at most a nice to have.

As of now, one could rearrange the hosts file so that the host only appears in the worker group and then manually run docker node demote NODE on the manager node to sync the swarm with the hosts file layout.

hosts groups declaration order

As some steps are run with flags run_once and delegate_to we expect the docker_swarm_managers group to be declared before the docker_swarm_workers group. Otherwise I suspect that a register variable is undefined and stuff crashes.

At least that's what I could find online. I haven't actually tested with a real hosts file just yet.

We should really test if this assumption is correct, would be nice if it isn't.

EDIT: This should no longer be an issue. See changes in: a9164fd for how it was resolved. Worth noting is also that the playbook with those fixes has been tested with a multi node setup (1 worker node, 1 manager only node) where the worker group appeared before the manager group in the hosts file

[flaeppe]

I have now tested the playbook with a multi node setup(1 worker node, 1 manager only node) successfully.

(hosts groups declaration order is no longer an issue either, this was also tested)

[beshrkayali]

Nitpick: Would be nice to make tags either use dash - or _ as a word separator. I think all existing ones use a - so maybe we should make swarm task tags also use that? :)