Spent a frustrating day trying to build a new test cluster, turned out I had jumbo frames set on the cluster-network only, but having re-wired the machines recently with a new switch, I forgot to check it could handle jumbo-frames (it can't).
Symptoms were stuck/unclean PGs - a small subset of PGs would go active but always a proportion would not, got side-tracked by using a ruleset set to OSD (it worked once) but would not work with host - all red-herrings I think. Anyhow, somewhere deep in Ceph a check might be useful at the network layer for fragmentation (or just remember this message). Thanks to Jean-Charles Lopez (JCL) on IRC for walking me through diagnosis (and sticking with me) while I circled around and around... _______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
