Deffinitelly in our case OSD were not the guilty ones, since all osd that where blocking requests allways from the same pool, worked flawlesly (and still do) after we deleted the pool where we always saw the blocked PG's.
Since the pool was accesed by just one client, and had almost no ops to it, i really dont know how to reproduce the issue, but surely scares me to happen ever again, and most, taking into consideration that blocked iops on a OSD could cascade through all the cluster and block all other pools.
It was technically hard to explain to the management that one 1.5GB pool locked almost 250 vms from different TB size pools, and most of it, not having a root cause (meaning, why only that pool generated blocked iops)
hope to hear some more technical insights or someone else that went through the same.
best.
On Thu, Mar 23, 2017 at 5:47 AM, Peter Maloney <peter.maloney@xxxxxxxxxxxxxxxxxxxx> wrote:
I think Greg (who appears to be a ceph committer) basically said he was interested in looking at it, if only you had the pool that failed this way.
Why not try to reproduce it, and make a log of your procedure so he can reproduce it too? What caused the slow requests... copy on write from snapshots? A bad disk? exclusive-lock with 2 clients writing at the same time maybe?
I'd be interested in a solution too... like why can't idle disks (non-full disk queue) mean that the osd op or whatever queue can still fill with requests not related to the blocked pg/objects? I would love for ceph to handle this better. I suspect some issues I have are related to this (slow requests on one VM can freeze others [likely blame the osd], even requiring kill -9 [likely blame client librbd]).
On 03/22/17 16:18, Alejandro Comisario wrote:
any thoughts ?
On Tue, Mar 14, 2017 at 10:22 PM, Alejandro Comisario <alejandro@xxxxxxxxxxx> wrote:
Greg, thanks for the reply.True that i cant provide enough information to know what happened since the pool is gone.
But based on your experience, can i please take some of your time, and give me the TOP 5 fo what could happen / would be the reason to happen what hapened to that pool (or any pool) that makes Ceph (maybe hapened specifically in Hammer ) to behave like that ?
Information that i think will be of value, is that the cluster was 5 nodes large, running "0.94.6-1trusty" i added two nodes running the latest "0.94.9-1trusty" and replication into those new disks never ended, since i saw WEIRD errors on the new OSDs, so i thought that packages needed to be the same, so i "apt-get upgraded" the 5 old nodes without restrting nothing, so rebalancing started to happen without errors (WEIRD).
after these two nodes reached 100% of the disks weight, the cluster worked perfectly for about two weeks, till this happened.After the resolution from my first email, everything has been working perfect.
thanks for the responses.
On Fri, Mar 10, 2017 at 4:23 PM, Gregory Farnum <gfarnum@xxxxxxxxxx> wrote:
On Tue, Mar 7, 2017 at 10:18 AM Alejandro Comisario <alejandro@xxxxxxxxxxx> wrote:
Gregory, thanks for the response, what you've said is by far, the most enlightneen thing i know about ceph in a long time.
What brings even greater doubt, which is, this "non-functional" pool, was only 1.5GB large, vs 50-150GB on the other effected pools, the tiny pool was still being used, and just because that pool was blovking requests, the whole cluster was unresponsive.
So , what do you mean by "non-functional" pool ? how a pool can become non-functional ? and what asures me that tomorrow (just becaue i deleted the 1.5GB pool to fix the whole problem) another pool doesnt becomes non-functional ?
Well, you said there were a bunch of slow requests. That can happen any number of ways, if you're overloading the OSDs or something.When there are slow requests, those ops take up OSD memory and throttle, and so they don't let in new messages until the old ones are serviced. This can cascade across a cluster -- because everything is interconnected, clients and OSDs end up with all their requests targeted at the slow OSDs which aren't letting in new IO quickly enough. It's one of the weaknesses of the standard deployment patterns, but it usually doesn't come up unless something else has gone pretty wrong first.As for what actually went wrong here, you haven't provided near enough information and probably can't now that the pool has been deleted. *shrug*-Greg
| Alejandro Comisario CTO | NUBELIU E-mail: alejandro@xxxxxxxxxxxCell: +54 9 11 3770 1857 _ www.nubeliu.com |
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
