I have been reading the architecture section of ceph document. One thing has
not been clear to me is how the data HA works when we encounter OSD or server
failure. Does the Crush algorithm recalculate based on the new cluster map and
point the data to the 2nd or 3rd replica for existing data block read or write?
Given the 2nd or 3rd replica's location (OSDs) were calculated by the primary
OSD instead of the client, this is not clearly to me if and how this is being
done.
A related question to the data HA mechanism, if client (librados) does
recalculate the primary OSD location and point to the 2nd OSD, how much latency
or how long the IO hang will client (e.g., VM) experience on an average load
scenario? In our traditional commercial hypervisor environment, we experienced
SCSI time out and Linux guest OS file system turns to readonly mode due to NFS
Datastore/Network hiccups.
Thanks. --weiguo
_______________________________________________
ceph-users mailing list
[email protected]
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
