> Op 12 juli 2016 om 12:35 schreef Simon Murray > <simon.mur...@datacentred.co.uk>: > > > Hi all. > > I'm about to perform a rather large reorganization of our cluster and > thought I'd get some insight from the community before going any further. > > The current state we have (logically) is two trees, one for spinning rust, > one for SSD. Chassis are the current failure domain, and are all chassis > types are organized under a single rack. > > The machines have been physically been relocated across 4 racks and I have > re-written the crush map to organize it so the chassis are correctly > located in the correct racks. I intend to also change the rules so that > the failure domain is now at the rack level so we can tolerate more severe > power and switching failure. > > Question is how is the best way to do this? > > 1) The pragmatist in me says commit the new crush map, let things > rebalance, then apply a new rule set to each pool and again let things > rebalance. >
Yes, that would be easy to do. Also, you can always revert to the old situation if you need to. > 2) It would however mean much less scheduled maintenance and keep customers > happier if I could just do everything as a big bang and do everything at > once e.g. rename the existing rule sets to replicated_rack_leaf(_ssd), and > change the chooseleaf option to type rack, and hope for the best :) > > Is the latter safe or just plain crazy? A big change can be easier for Ceph. You simply limit the backfill and recovery to one at a time and let it run. You could also create new rulesets which do the mapping in a different way and switch per pool to the new rulesets. Wido > > Si > > -- > DataCentred Limited registered in England and Wales no. 05611763 > _______________________________________________ > ceph-users mailing list > ceph-users@lists.ceph.com > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com _______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
