> 4 juni 2017 kl. 23:23 skrev Roger Brown <rogerpbr...@gmail.com>: > > I'm a n00b myself, but I'll go on record with my understanding. > > On Sun, Jun 4, 2017 at 3:03 PM Benoit GEORGELIN - yulPa > <benoit.george...@yulpa.io <mailto:benoit.george...@yulpa.io>> wrote: > Hi ceph users, > > Ceph have a very good documentation about technical usage, but there is a lot > of conceptual things missing (from my point of view) > It's not easy to understand all at the same time, but yes, little by little > it's working. > > Here are some questions about ceph , hope someone can take a little time to > point me where I can find answers : > > - Backup : > Do you backup data from a CEPH cluster or you consider a copy as a backup of > that file ? > Let's say I have replica size of 3 . Somehow , my crush map will keep 2 copy > in my main rack and 1 copy to another rack in another datacenter > Can I consider the third copy as a backup ? What would be your position ? > > Replicas are not backups. Just ask GitLab after accidental deletion. source: > https://www.theregister.co.uk/2017/02/01/gitlab_data_loss/ > <https://www.theregister.co.uk/2017/02/01/gitlab_data_loss/> > > > - Writing process of ceph object storage using radosgw > Simple question, but not sure about it. > The more replica the more slower will be my cluster ? Does CEPH have to > acknowledge the number of replica before saying it's good ? > From what I read, CEPH will write and acknowledge the de primary OSD of the > pool , So if that the cas, I does not matter how many replica I want and how > far are situated the others OSD that would work the same. > Can I chose myseft the primary OSD in my zone 1 , have a copy on zone 2 > (same rack) and a third zone 3 in another datacenter that might have some > latency . > > More replicas make slower cluster because it waits for all devices to > acknowledge write before reporting back. source: ?
I’d say stick with 3 replicas in one DC, then if you want to add another DC for better data protection (note, not backup), you’ll just add asynchronous mirroring between DCs (http://docs.ceph.com/docs/master/rbd/rbd-mirroring/ <http://docs.ceph.com/docs/master/rbd/rbd-mirroring/>) with another cluster there. That way you’ll have a quick cluster (especially if you use awesome disks like NVME SSD journals + SSD storage or better) with a location redundancy. > > > - Data persistance / availability > If crush map is by hosts and I have 3 hosts with replication of 3 > This means , I will have 1 copy on each hosts > Does it means I can lost 2 hosts and still have my cluster working, at least > on read mode ? and eventually in write too if i say , osd pool default min > size = 1 > > Yes, I think. But best practice is to have at least 5 hosts (N+2) so you can > lose 2 hosts and still keep 3 replicas. > Keep in mind that you ”should” have enough storage free as well to be able to loose 2 nodes. If you fill 5 nodes to 80% and loose 2 nodes you won’t be able to repair it all until you get them up and running again. > > Thanks for your help. > - > > Benoît G > _______________________________________________ > ceph-users mailing list > ceph-users@lists.ceph.com <mailto:ceph-users@lists.ceph.com> > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > <http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com> > > Roger > > _______________________________________________ > 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