Open Stack

On Fri, Nov 2, 2018 at 9:39 AM Dan Prince <dprince at redhat.com> wrote:
>
> I pushed a patch[1] to update our containerized deployment
> architecture docs yesterday. There are 2 new fairly useful sections we
> can leverage with TripleO's stepwise deployment. They appear to be
> used somewhat sparingly so I wanted to get the word out.
>
> The first is 'deploy_steps_tasks' which gives you a means to run
> Ansible snippets on each node/role in a stepwise fashion during
> deployment. Previously it was only possible to execute puppet or
> docker commands where as now that we have deploy_steps_tasks we can
> execute ad-hoc ansible in the same manner.
>
> The second is 'external_deploy_tasks' which allows you to use run
> Ansible snippets on the Undercloud during stepwise deployment. This is
> probably most useful for driving an external installer but might also
> help with some complex tasks that need to originate from a single
> Ansible client.

+1


> The only downside I see to these approaches is that both appear to be
> implemented with Ansible's default linear strategy. I saw shardy's
> comment here [2] that the :free strategy does not yet apparently work
> with the any_errors_fatal option. Perhaps we can reach out to someone
> in the Ansible community in this regard to improve running these
> things in parallel like TripleO used to work with Heat agents.

It's effectively parallel across one role at a time at the moment, up
to the number of configured forks (default: 25). The reason it won't
parallelize across roles, is because it's a different task file used
with import_tasks for each role. Ansible won't run that in parallel
since the task list is different.

I was able to make this parallel across roles for the pre and post
deployments by making the task file the same for each role, and
controlling the difference with group and host vars:
https://review.openstack.org/#/c/574474/
>From Ansible's perspective, the task list is now the same for each
host, although different things will be done depending on the value of
vars for each host.

It's possible a similar approach could be done with the other
interfaces you point out here.

In addition to the any_errors_fatal issue when using strategy:free, is
that you'd also lose the grouping of the task output per role after
each task finishes. This is mostly cosmetic, but using free does
create a lot more noisier output IMO.

-- 
-- James Slagle
--