For those interested in this issue. We've been seeing OSDs with corrupted wals after they had a suicide time out. I've updated the ticket created by William with some of our logs. https://tracker.ceph.com/issues/48827#note-16
We're using ceph 15.2.2 in this cluster. Currently we are contemplating a way forward, but it looks like the wals are being corrupted under load. Kind regards, Wout 42on ________________________________________ From: William Law <w...@stanford.edu> Sent: Tuesday, 19 January 2021 18:48 To: ceph-users@ceph.io Subject: [ceph-users] Nautilus Cluster Struggling to Come Back Online I guess as a sort of follow up from my previous post. Our Nautilus (14.2.16 on ubuntu 18.04) cluster had some sort of event that caused many of the machines to have memory errors. The aftermath is that initially some OSDs had (and continue to have) this error https://tracker.ceph.com/issues/48827 others won't start for various reasons. The OSDs that *will* start are badly behind the current epoch for the most part. It sounds very similar to this: https://blog.noc.grnet.gr/2016/10/18/surviving-a-ceph-cluster-outage-the-hard-way/ We are having trouble getting things back online. I think the path forward is to: -set noup/nodown/noout/nobackfill/and wait for the OSDs that run to come up; we were making good progress yesterday until some of the OSDs crashed with OOM errors. We are again moving forward but understandably nervous. -export the PGs from questionable OSDs and and then rebuild the OSDs; import the PGs if necessary (very likely). Repeat until we are up. Any suggestions for increasing speed? We are using noup/nobackfill/norebalance/pause but the epoch catchup is taking a very long time. Any tips for keeping the epoch from moving forward or speeding up the OSDs catching up? How can we estimate how long it should take? Thank you for any ideas or assistance anyone can provide. Will _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io
