Hi Dan,
Unfortunately 'ceph osd down osd.0' doesn't help - it is marked down and
soon after back up, but it doesn't peer still. I tried reweighting the
OSD to half its weight, 4.0 instead of 0.0, and that results in about
half the PGs staying degraded. So this is not specific to zero weight.
Hi Frank,
Thanks for the explanation - I wasn't aware of this subtle point. So
when some OSDs are down, one has to be very careful with changing the
cluster then. I guess one could even end up with incomplete PGs this
way that ceph can't recover from in an automated fashion?
Andras
On 5/1
Hi Andras,
To me it looks like the osd.0 is not peering when it starts with crush weight 0.
I would try forcing the re-peering with `ceph osd down osd.0` when the
PGs are unexpectedly degraded. (e.g start the osd when crush weight is
0, then obverve the PGs are still degraded, then force the re-p
Hi Andreas,
the cluster map and crush map are not the same thing. If you change the crush
map while the cluster is in degraded state, you basically modify this history
of cluster maps explicitly and have to live with the consequences (keeping
history under crush map changes is limited to up+in
Hi Frank,
My understanding was that once a cluster is in a degraded state (an OSD
is down), ceph stores all changed cluster maps until the cluster is
healthy again exactly for the reason of finding missing objects. If
there is a real disaster of some kind, and many OSDs go up and down at
vari
Hi Andreas,
I made exactly the same observation in another scenario. I added some OSDs
while other OSDs were down.
This is expected.
The crush map is an a-priory algorithm to compute the location of objects
without contacting a central server. Hence, *any*change of a crush map while an
OSD is