> -----Original Message----- > From: Studziński Krzysztof > Sent: Wednesday, July 24, 2013 1:18 AM > To: 'Gregory Farnum'; Yehuda Sadeh > Cc: ceph-devel@vger.kernel.org; ceph-us...@lists.ceph.com; Mostowiec > Dominik > Subject: RE: [ceph-users] Flapping osd / continuously reported as failed > > > -----Original Message----- > > From: Gregory Farnum [mailto:g...@inktank.com] > > Sent: Wednesday, July 24, 2013 12:28 AM > > To: Studziński Krzysztof; Yehuda Sadeh > > Cc: ceph-devel@vger.kernel.org; ceph-us...@lists.ceph.com; Mostowiec > > Dominik > > Subject: Re: [ceph-users] Flapping osd / continuously reported as failed > > > > On Tue, Jul 23, 2013 at 3:20 PM, Studziński Krzysztof > > <krzysztof.studzin...@grupaonet.pl> wrote: > > >> On Tue, Jul 23, 2013 at 2:50 PM, Studziński Krzysztof > > >> <krzysztof.studzin...@grupaonet.pl> wrote: > > >> > Hi, > > >> > We've got some problem with our cluster - it continuously reports > failed > > >> one osd and after auto-rebooting everything seems to work fine for > some > > >> time (few minutes). CPU util of this osd is max 8%, iostat is very low. > > >> We > > tried > > >> to "ceph osd out" such flapping osd, but after recovering this behavior > > >> returned on different osd. This osd has also much more read operations > > than > > >> others (see file osd_reads.png linked at the bottom of the email; at > about > > >> 16:00 we switched off osd.57 and osd.72 started to misbehave. Osd.108 > > >> works while recovering). > > >> > > > >> > Extract from ceph.log: > > >> > > > >> > 2013-07-23 22:43:57.425839 mon.0 10.177.64.4:6789/0 24690 : [INF] > > osd.72 > > >> 10.177.64.8:6803/22584 boot > > >> > 2013-07-23 22:43:56.298467 osd.72 10.177.64.8:6803/22584 415 : [WRN] > > map > > >> e41730 wrongly marked me down > > >> > 2013-07-23 22:50:27.572110 mon.0 10.177.64.4:6789/0 25081 : [DBG] > > osd.72 > > >> 10.177.64.8:6803/22584 reported failed by osd.9 10.177.64.4:6946/5124 > > >> > 2013-07-23 22:50:27.595044 mon.0 10.177.64.4:6789/0 25082 : [DBG] > > osd.72 > > >> 10.177.64.8:6803/22584 reported failed by osd.78 10.177.64.5:6854/5604 > > >> > 2013-07-23 22:50:27.611964 mon.0 10.177.64.4:6789/0 25083 : [DBG] > > osd.72 > > >> 10.177.64.8:6803/22584 reported failed by osd.10 > 10.177.64.4:6814/26192 > > >> > 2013-07-23 22:50:27.612009 mon.0 10.177.64.4:6789/0 25084 : [INF] > > osd.72 > > >> 10.177.64.8:6803/22584 failed (3 reports from 3 peers after 2013-07-23 > > >> 22:50:43.611939 >= grace 20.000000) > > >> > 2013-07-23 22:50:30.367398 7f8adb837700 0 log [WRN] : 3 slow > requests, > > 3 > > >> included below; oldest blocked for > 30.688891 secs > > >> > 2013-07-23 22:50:30.367408 7f8adb837700 0 log [WRN] : slow request > > >> 30.688891 seconds old, received at 2013-07-23 22:49:59.678453: > > >> sd_op(client.44290048.0:125899 .dir.4168.2 [call > rgw.bucket_prepare_op] > > >> 3.9447554d) v4 currently no flag points reached > > >> > 2013-07-23 22:50:30.367412 7f8adb837700 0 log [WRN] : slow request > > >> 30.179044 seconds old, received at 2013-07-23 22:50:00.188300: > > >> sd_op(client.44205530.0:189270 .dir.4168.2 [call rgw.bucket_list] > > 3.9447554d) > > >> v4 currently no flag points reached > > >> > 2013-07-23 22:50:30.367415 7f8adb837700 0 log [WRN] : slow request > > >> 30.171968 seconds old, received at 2013-07-23 22:50:00.195376: > > >> sd_op(client.44203484.0:192902 .dir.4168.2 [call rgw.bucket_list] > > 3.9447554d) > > >> v4 currently no flag points reached > > >> > 2013-07-23 22:51:36.082303 mon.0 10.177.64.4:6789/0 25159 : [INF] > > osd.72 > > >> 10.177.64.8:6803/22584 boot > > >> > 2013-07-23 22:51:35.238164 osd.72 10.177.64.8:6803/22584 420 : [WRN] > > map > > >> e41738 wrongly marked me down > > >> > 2013-07-23 22:52:05.582969 mon.0 10.177.64.4:6789/0 25191 : [DBG] > > osd.72 > > >> 10.177.64.8:6803/22584 reported failed by osd.20 10.177.64.4:6913/4101 > > >> > 2013-07-23 22:52:05.587388 mon.0 10.177.64.4:6789/0 25192 : [DBG] > > osd.72 > > >> 10.177.64.8:6803/22584 reported failed by osd.9 10.177.64.4:6946/5124 > > >> > 2013-07-23 22:52:05.610925 mon.0 10.177.64.4:6789/0 25193 : [DBG] > > osd.72 > > >> 10.177.64.8:6803/22584 reported failed by osd.78 10.177.64.5:6854/5604 > > >> > 2013-07-23 22:52:05.610951 mon.0 10.177.64.4:6789/0 25194 : [INF] > > osd.72 > > >> 10.177.64.8:6803/22584 failed (3 reports from 3 peers after 2013-07-23 > > >> 22:52:20.610895 >= grace 20.000000) > > >> > 2013-07-23 22:52:05.630821 mon.0 10.177.64.4:6789/0 25195 : [DBG] > > osd.72 > > >> 10.177.64.8:6803/22584 reported failed by osd.10 > 10.177.64.4:6814/26192 > > >> > 2013-07-23 22:53:47.203352 mon.0 10.177.64.4:6789/0 25300 : [INF] > > osd.72 > > >> 10.177.64.8:6803/22584 boot > > >> > 2013-07-23 22:53:46.417106 osd.72 10.177.64.8:6803/22584 474 : [WRN] > > map > > >> e41742 wrongly marked me down > > >> > > > >> > Could you please take a look at our config and suggest some > > >> improvements? > > >> > See attached "ceph pg <pg_id> query" for two groups during recovery > > and > > >> parts of our config file. > > >> > Our cluster's size: 6 hosts, 26 HDD each, 156 osds, 6488 pgs, mostly in > > one > > >> bucket having 9M objects, 3342 GB data, 11173 GB used, 31690 GB / > 42864 > > GB > > >> avail. > > >> > > >> I'm surprised you're running into it at 9m objects but this is almost > > >> certainly the problem. Right now the index for each RGW bucket lives > > >> on a single OSD; you're probably having issues with whichever OSD is > > >> receiving the bucket index reads. Is it feasible for you to shard the > > >> contents into multiple buckets and see if things calm down? > > >> -Greg > > >> Software Engineer #42 @ http://inktank.com | http://ceph.com > > > > > > I was afraid of that answer. :-/ > > > We'll try to delete some objects for the beginning (as they are > transformed > > other objects and can be re-created when needed) and then try to shard it > > as you suggest. > > > What is your opinion about max number of objects in one bucket ? > > > > It depends more on how much activity you're throwing at the bucket, I > > think. We haven't done any large-scale tests and perhaps we need to. > > Leveldb (which we're using under the covers to store this stuff) > > should not have any trouble with the amount of data that's in there, > > but if you're trying to do frequent enough object lookups or puts then > > you might just saturate the disk/node's ability to keep up. I should > > mention that Yehuda just started a discussion on handling this in the > > thread "[ceph-users] rgw bucket index". > > > > When did you start noticing this trouble? > > Most of our objects are images: originals (mostly 5-10MB) and its > transformations (compressed, crops, filters, etc, usually <<1MB). We also > have plenty of small text files. We mostly read them (objects are cached by > nginx outside the cluster), but yesterday (15h ago) we had pretty large > deployment of new service that could create plenty of transformed images > with different sizes. Unfortunately I cannot say right now what is the > difference in bucket size between now and yesterday, I will try to check it > out. > > From my stats I can see that in peaks we had about 1.5K PUT and 8K GET > operations per minute. Most common use-case while PUTting an object in > our apps is this sequence: > - HEAD (checking if transformed object already exist), > - GET the original > - PUT transformed object and its meta-data > We also have some amount of PUTs of new objects, but I think there won't > be many by now (it is 1am in Poland). > Right now there are 80 PUT and 1.5K GET operations per minute and the osd > is still flapping.
Additional info about our index growth: the deployment I mentioned in previous mail didn't influence the speed of index grow. Currently we have 60K of new objects daily, that speed is almost constant for the last 7 days. -- Best regards, Krzysztof Studzinski