OSD failed to subscribe skipped osdmaps after "ceph osd pause"
per Wido's comment in #16982-7,
I tried adding new OSDs to the cluster and they also have to catch up, which never happens until I restart them over and over.
osd.136 in this case is a fresh OSD. You can see it jumps with 1.000 maps (osd_map_message_max), but then just waits.
I restart the osd, it goes 1k maps forward and waits. I restart, etc, etc.
and the root cause is analyzed at #16982-11.
in short, the problem here, due to "ceph osd pause", the subscription sent by objecter always gets in the way of OSD, so the latter cannot subscribe for the older osdmap to catch up with the cluster.
so a workaround is to not "ceph osd pause".
#2 Updated by Wido den Hollander over 2 years ago
To clarify why we did the 'osd pause'.
The upgrade from Hammer -> Jewel didn't go very well. We had machines go OOM, CPU 100% busy and even a disk failing under the pressure.
We want to be back up and running asap, but data integrity was even more important. I set the pause flag to prevent any changes to the data so that the cluster could make all PGs active again before I thought it was safe to continue.