OK! Success of a sort.
I removed and re-installed each of the failed services in turn using the
"ceph orch rm" command followed by "ceph orch apply". They came up with
default settings (1 server), but they did come up.
Finally, I tried it with prometheus. This gave me:
prometheus ?:9095 0/1 - 10s
count:1
However, in order for the dashboard to be happy, I had to supply more
info. Since ceph orch ls wouldn't tell me /where/ the new prometheus was
deployed, I used the hosts tab in the dashboard to find it.
Following that, I had to set the following:
ceph config set mgr mgr/prometheus/server_addr 10.0.1.58
ceph config set mgr mgr/prometheus/server_port 9095
ceph dashboard set-prometheus-api-host 10.0.1.58 (ceph08)
ceph dashboard set-prometheus-api-port 909
Once all of the above were set, the dashboard stopped complaining about
being able to access the prometheus API.
However, one last wart remains. Despite being up and running (and
confirmed listening on ceph09 port 9095), I do get this:
# ceph health detail
HEALTH_ERR Module 'prometheus' has failed: gaierror(-2, 'Name or service
not known'); too many PGs per OSD (648 > max 560)
[ERR] MGR_MODULE_ERROR: Module 'prometheus' has failed: gaierror(-2,
'Name or service not known')
Module 'prometheus' has failed: gaierror(-2, 'Name or service not
known')
[WRN] TOO_MANY_PGS: too many PGs per OSD (648 > max 560)
On 3/28/25 08:53, Tim Holloway wrote:
Actually, I did deploy a new mds node yesterday. But I followed your
instructions and successfully removed and re-installed ceph-exporter
(4 nodes). So that part works.
On 3/28/25 07:28, Eugen Block wrote:
Okay, next I would keep prometheus disabled to see if the mgr works
properly. So disable the module again, and also reset the dashboard
setting to an empty value:
ceph dashboard reset-prometheus-api-host
Then see if you get an mds daemon deployed. Or test it by removing
and redeploying ceph-exporter or crash or something, anything to test
if the mgr is able to remove and deploy other services.
Zitat von Tim Holloway <t...@mousetech.com>:
Thanks for the info on removing stubborn dead OSDs. The actual
syntax required was:
cephadm rm-daemon --name osd.2 --fsid <fsid>--force
On the "too many pgs", that's because I'm down 2 OSDs. I've got new
drives, but they were waiting to clear out the dead stuff. I know
it's risky, but I have backups.
Recall that the start of this thread was on a HEALTH_OK system and
prometheus was not activating. The OSD stuff was just a distraction.
I did notice that the attempt to add a new mds did work after I did
a "ceph mgr fail", so it's only prometheus that's a permanent problem.
Here's the latest health after clearing out the dead OSDs:
# ceph health detail
HEALTH_ERR Module 'prometheus' has failed: gaierror(-2, 'Name or
service not known'); too many PGs per OSD (648 > max 560)
[ERR] MGR_MODULE_ERROR: Module 'prometheus' has failed: gaierror(-2,
'Name or service not known')
Module 'prometheus' has failed: gaierror(-2, 'Name or service
not known')
[WRN] TOO_MANY_PGS: too many PGs per OSD (648 > max 560)
And yes, disabling prometheus will make the "name or service not
known" errors go away.
On 3/28/25 02:49, Eugen Block wrote:
Did you disable the prometheus module? I would expect the warning
to clear if you did.
Somewhere deep inside ceph, those deleted OSDs still exist. Likely
because ceph08 hasn't deleted the systemd units that run them.
Or do you still see those OSDs in 'cephadm ls' output on ceph08? If
you do, and if those OSDs are really already drained/purged, you
can remove them with 'cephadm rm-daemon --name osd.2'. And I would
try to get the MGR into a working state first, before you try to
deploy prometheus again. So my recommendation is to get into
HEALTH_OK first. And btw, "TOO_MANY_PGS: too many PGs per OSD (648
> max 560)" is serious, you can end up with inactive PGs during
recovery, so I'd also consider checking the pools and their PGs.
Zitat von Tim Holloway <t...@mousetech.com>:
Thanks for your patience.
host ceph06 isn't referenced in the config database. I think I've
finally purged it. I also reset the dashboard API host address
from ceph08 to dell02. But since prometheus isn't running on
dell02 either, there's no gain there.
I did clear some of that lint out via "ceph mgr fail".
So here's the latest. There are strange things happening at the
base OS level that keep host ceph08 from running its OSDs anymore.
At boot, device /dev/sdb suddenly changes to /dev/sdd (????) and
there seem to be I/O errors. It's really strange, but I'm going to
replace the physical drive and that will hopefully cure that.
The problem is, reef and earlier releases seem to have a lot of
trouble in deleting OSDs that aren't running. As I've noted
before, they tend to get permanently stuck in the "deleting"
state. When I cannot restart the OSD, the only cure for that has
been to run around the system and apply brute force until things
clear up.
I did a dashboard purge of the OSDs on ceph08 and that removed
them from the GUI (they'd already drained). I also banged on
things until I got them out of the OSD tree display and then did a
crush delete on host ceph08. And, incidentally, the OSD tree works
on simple host names, not FQDNs like the rest of ceph!
So in theory, I'm ready to jack in new drives and add new OSDs to
ceph08. Except:
# ceph health detail
HEALTH_ERR 2 failed cephadm daemon(s); Module 'prometheus' has
failed: gaierror(-2, 'Name or service not known'); too many PGs
per OSD (648 > max 560)
[WRN] CEPHADM_FAILED_DAEMON: 2 failed cephadm daemon(s)
daemon osd.2 on ceph08.internal.mousetech.com is in error state
daemon osd.4 on ceph08.internal.mousetech.com is in error state
[ERR] MGR_MODULE_ERROR: Module 'prometheus' has failed:
gaierror(-2, 'Name or service not known')
Module 'prometheus' has failed: gaierror(-2, 'Name or service
not known')
[WRN] TOO_MANY_PGS: too many PGs per OSD (648 > max 560)
Somewhere deep inside ceph, those deleted OSDs still exist. Likely
because ceph08 hasn't deleted the systemd units that run them.
I'm going to try removing/re-installing prometheus. since it's now
showing up in ceph health. I think last time I had zombie OSDs I
had to brute-force delete their corresponding directories under
/var/lib/ceph.
On 3/27/25 14:01, Eugen Block wrote:
ceph config-key rm
mgr/cephadm/host.ceph06.internal.mousetech.com.devices.0
_______________________________________________
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
_______________________________________________
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
_______________________________________________
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
