[openstack-dev] [tripleo] The Future of Service Orchestration with Puppet & Pacemaker

Michele Baldessari michele at acksyn.org
Wed Apr 6 08:22:53 UTC 2016


Hi,

On Tue, Mar 29, 2016 at 03:48:29PM -0400, Emilien Macchi wrote:
> Some TripleO folks are currently working hard on making automation for
> upgrading TripleO between major OpenStack releases.
> One of the biggest challenges is how to deal with Puppet who usually
> notify Services when configuration change and Pacemaker who actually
> manage the services.
> 
> Currently, we do this horrible thing that is disabling service
> management when Pacemaker manages the service.
> It's terrible because Puppet wants to manage the service with
> 'systemd' but we disable it to let Pacemaker deal with that.
> The main limitation is that most of Puppet modules (specially in
> OpenStack) know how to do orchestration for deployments & upgrades,
> but we don't let it do because we disable service management.

agreed. The split between managing some resources via puppet and
some by hand with pcs commands is error-prone and cumbersome.

> A few months ago, we initiated a discussion about writing a Puppet
> Resource Service Provider in openstack/puppet-pacemaker that will deal
> with service management.
> What will it change?
> * we will enable service management again in THT.
> * Puppet won't use systemd to deal with services, but use 'pcs' tool.
> * We'll take the benefit of puppet-* modules orchestration. Example:
> this configuration change will restart this service (something we had
> hard time to do before).
> 
> We are currently working on it:
> https://review.openstack.org/#/c/286124/
> 
> But this is work in progress:
> * we need to implement the restart
> * we need to create a special Puppet fact, that will make sure we only
> run pcs on a single node (master by preference)
> * investigate how to find who is master and stop hardcoding it in THT
> (good for a bootstrap, but not good for upgrades).
> 
> Here is the kind of scenario we expect at the end:
> 
> Deployment of Neutron Server service
> * openstack/puppet-neutron will configure neutron.conf
> * openstack/puppet-neutron will notify neutron-server service
> * THT is overriding the neutron-server service resource to use 'pcs'
> provider, implemented in openstack/puppet-pacemaker
> * once neutron.conf is configured on controllers, pcs will start
> neutron-server, only from one node.
> 
> Upgrade of Neutron Server service
> * we change RabbitMQ parameters in neutron.conf with THT
> * openstack/puppet-neutron will update neutron.conf and notify
> neutron-server service
> * 'pcs' provider will restart neutron-server from one node, after the
> change in the neutron.conf
> 
> That will be, in my opinion the most elegant way to make Puppet &
> Pacemaker working together, any comment / feedback is highly welcome,
> it will be a big change and we want to make it during Newton cycle.
> Note: by designing this change, I found out we will also be able to
> reduce the number of steps and also drop some bash code in the upgrade
> scripts, which are both good news at making TripleO more consistent
> and fast to deploy.

While I haven't yet played with this changeset, my first question would
be the following:
In this pcs-provider scenario, do we care about the fact that
if you stop 'resource A', also all the resourced that have 'A' as an
ordering constraint will be stopped? Or should the users of the module
care about this aspect?

Thanks for working on this. I will try to look at the proposed changes
this week and provide some feedback.

cheers,
Michele
-- 
Michele Baldessari            <michele at acksyn.org>
C2A5 9DA3 9961 4FFB E01B  D0BC DDD4 DCCB 7515 5C6D



More information about the OpenStack-dev mailing list