need to be more flexible in ceph role assignments
We currently default RadosGW and monitors to the controllers. As we have learned at customer's deployments this is not a brilliant idea. Both openstack and ceph are full of bugs and we do nothing to prevent bad neighbor on the controllers. So a run away monitor process at Hawaii took out a controllers and through Fuel HA recipe bug took out both OpenStack and Ceph clusters. We need to allow for monitors and radosgw to be separately definable roles which can be flexibly assigned to controllers, computes or standalone nodes
Blueprint information
- Status:
- Started
- Approver:
- Andrew Woodward
- Priority:
- Medium
- Drafter:
- Gregory Elkinbard
- Direction:
- Approved
- Assignee:
- Dmitry Borodaenko
- Definition:
- Discussion
- Series goal:
- Accepted for future
- Implementation:
-
Slow progress
- Milestone target:
-
next
- Started by
- Vladimir Kuklin
- Completed by
Related branches
Related bugs
Sprints
Whiteboard
Two new roles have to be created: ceph-mon and ceph-rgw. The roles should be defined in openstack.yaml (with no role-specific volumes), the ceph Puppet module should be updated to further decouple ceph::mon and ceph::radosgw classes.
N.B.: ceph::radosgw currently assumes that apache is already installed, need to verify if that assumption has to be amended on a standalone ceph-rgw node.
angdraug: also see superceded blueprint https:/
Work Items
Work items:
Create ceph-mon role in openstack.yaml: INPROGRESS
-- needs better description: TODO
Create ceph-rgw role in openstack.yaml: INPROGRESS
-- also needs better description: TODO
Extract Fuel specific parts of ceph Puppet module into ceph::fuel: INPROGRESS
Make ceph::mon class self-sufficient: INPROGRESS
Make ceph::radosgw class self-sufficient: INPROGRESS
Add post-deployment checks for ceph-mon and ceph-rgw: TODO
Serializer changes: INPROGRESS
-- update tests: TODO
-- ceph-primary-mon FIRST: DONE
-- ceph-mons second : DONE
-- ceph-osds third: POSTPONED
-- ceph-rgw LAST - Not needed : DONE
-- generate rbd_secret_uuid: POSTPONED
-- generate fsid: POSTPONED
-- generate mon_hosts & mon_ip_addresses : POSTPONED
Make sure the primary ceph-mon is deployed before other nodes: TODO
Dependency tree
* Blueprints in grey have been implemented.
