Hello, we have CEPH cluster with 144 NVMe "disks", "background" network is RoCE. CEPH cluster is version 18.2.2 and is installed via CEPH orchestrator cephadm, container daemon is podman. OS is Debian bookworm, podman is in the version 4.3.1+ds1-8+b1, now we are installed version 4.3.1+ds1-8+deb12u1 (standard Debian bookworm package) We have sometimes problem with the laggy OSD. This Sunday we had problem with osd.25. It was laggy, some other OSDs reported it to MON, MON dismissed it from crush map, but osd.25 complaint about it, MON tried to add it again to the crush map - but this behavior caused, that whole CEPH cluster was laggy and had slow ops, which was a "kiss of death" to some ProxMox VMs :-(. When I restarted osd.25 daemon, after a couple minutes cluster goes back to HEALTHY state. I have prepared logs from 3 machines: c2-dc1-elected-mon-osd.25.txt - machine c2-dc1 had elected MON and 12 OSDs, same of them complait's about osd.25. c1-dc1-mon-osd.25.txt - machine with another MON and 12 OSDs c4-dc2-osd.25.txt - machine with osd.25 daemon (and other 11 OSDs) Problem was, that the first laggy daemon occured around 6 a.m., but problem insisted up to cca 4:20 p.m., when I've restarted osd.25. There is a question, if will be better to restart laggy OSD daemon by ceph orchestrator? For example: If OSD daemons from 3 host (now it is from 2 hosts by default) and after 20s reported some OSD as a dead, MON give a command to mgr to restart this OSD? Somethink like STONITH in corosync/pcs cluster ( https://en.wikipedia.org/wiki/STONITH )? Logs are very big (around 23MB tar.bz2), I've prepared it here: https://hazard.jcu.cz/osd.25.tar.bz2 We thinking about some problem with network, but on the same machine, there is another 11 OSDs and they parformed well. This is not the first occurence of this problem, when it happened the first time, we tried to reload whole server, but then we found, that reload of OSD container is enough... Sincerely Jan Marek -- Ing. Jan Marek University of South Bohemia Academic Computer Centre Phone: +420389032080 http://www.gnu.org/philosophy/no-word-attachments.cs.html