On 14/12/2016 22:59, Dotan, Alon wrote: Galera is managed via a specific resource agent which takes care of all the stuff needed to start, stop and, most of all, recover. It automates a lot of tasks which otherwise will be manual. Haproxy is managed by Pacemaker because of the Vips. There are specific constraints that ensures that whether a Vip change the host, haproxy service will follow. Hope this helps, -- Raoul Scarazzini rasca(a)redhat.com

On 15/12/2016 09:42, Dotan, Alon wrote: It is a systemd resource managed by Pacemaker. -- Raoul Scarazzini rasca(a)redhat.com

Actually, we have been trying to move away from Pacemaker where possible, but certain services are not perfectly active/active, so we need an additional layer on top to ensure that connections go to the right server(s). Galera is one of the services that still benefits from pcs. Andrew Beekhof has a great overview of the changes in progress here [1]. From that doc: "Generally we use round-robin to distriute load amongst instances of active/active services, however Galera and Qpid use the stick-table options to ensure that incoming connections to the virtual IP (VIP) should be directed to only one of the available backends. In Galera's case, although it can run active/active, this helps avoid lock contention and prevent deadlocks. It is used in combination with the httpchk option that ensures only nodes that are in sync with its peers are allowed to handle requests. [snip] Although galera supports active/active configurations, we recommend active/passive (enforced by the load balancer) in order to avoid lock contention." Another service which is currently active/passive is Cinder. In theory, this can be made active/active, but there are lingering concerns about edge cases that make active/passive safer for now. We are likely moving away from Pacemaker entirely in the long run, but there are some technical hurdles to overcome before we can do that. The good news is that most services do not require pcs, and rely only on systemd. You can now distribute systemd services between multiple sets of servers, perhaps moving Neutron services to their own set of servers, or perhaps moving the active/active services to a horizontally scalable set of API servers. The remaining services which require pcs will have to be colocated on a 3-node Pacemaker cluster for now. [1] - https://github.com/beekhof/osp-ha-deploy/blob/master/ha-openstack.md -- Dan Sneddon | Senior Principal OpenStack Engineer dsneddon(a)redhat.com | redhat.com/openstack dsneddon:irc | @dxs:twitter On 12/14/2016 01:59 PM, Dotan, Alon wrote: