On 12/7/2016 9:25 AM, Jonathan Proulx wrote:
We've been using Ceph as ephemeral backend (and glance store and cinder backend) for > 2 years (maybe 3 ) and have been very happy. cinder has been rock solid on RBD side. Early on when we had 6 osd servers we lost one in production to a memory error. 1/6 is a large fraction to loose but Ceph handled it as designed and there was not noticable impact to any running instances. The ability to start VMs from snap shot of a glance image (which is how it's implemented if glance and nova are both cinder backed) makes start up super fast. Having shared storage also make live migration easy so we can do hardware maintenence (kernel and OS upgrades) without impacting running VMs As of Mitaka snapshotting running VMs is also fast, though for earlier releases VMs we suspended while the running RBD volumes was copied down to the the hypervisors and then restarted while the downloaded images was re-uploaded from the hypervisor to glance and back into Ceph. This could mean VMs down foro 15min or more if they had large root volumes. As I said this got fixed in Mitaka so if you're current this is no longer a problem.
This is the specific change mentioned above: https://review.openstack.org/#/c/205282/ -- Thanks, Matt Riedemann _______________________________________________ OpenStack-operators mailing list OpenStack-operators@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
