Not going through the obvious of that crush map is just not looking correct or even sane... or that the policy itself doesn't sound very sane - but I'm sure you'll understand the caveats and issues it may present...
what's most probably happening is that a (or several) pool is using those same OSDs and the requests to those PGs are also getting blocked because of the disk full. This turns that some (or all) of the remaining OSDs are waiting for that one to complete some IO, and whilst those OSDs have IOs waiting to complete it also stops responding to the IO that was only local. Adding more insanity to your architecture what should (the keyword here is should as I never tested, saw or even thought of such scenario) work would be OSDs to have local storage and OSDs to have distributed storage. As for the architecture itself, and not knowing much of your use-case, it may make sense to have local storage in something else than Ceph - you're not using any of the facilities it provides you, and having some overheads - or using a different strategy for it. IIRC there was a way to hint data locality to Ceph... On Wed, Aug 16, 2017 at 8:39 AM, Mandar Naik <mandar.p...@gmail.com> wrote: > Hi, > I just wanted to give a friendly reminder for this issue. I would appreciate > if someone > can help me out here. Also, please do let me know in case some more > information is > required here. > > On Thu, Aug 10, 2017 at 2:41 PM, Mandar Naik <mandar.p...@gmail.com> wrote: >> >> Hi Peter, >> Thanks a lot for the reply. Please find 'ceph osd df' output here - >> >> # ceph osd df >> ID WEIGHT REWEIGHT SIZE USE AVAIL %USE VAR PGS >> 2 0.04399 1.00000 46056M 35576k 46021M 0.08 0.00 0 >> 1 0.04399 1.00000 46056M 40148k 46017M 0.09 0.00 384 >> 0 0.04399 1.00000 46056M 43851M 2205M 95.21 2.99 192 >> 0 0.04399 1.00000 46056M 43851M 2205M 95.21 2.99 192 >> 1 0.04399 1.00000 46056M 40148k 46017M 0.09 0.00 384 >> 2 0.04399 1.00000 46056M 35576k 46021M 0.08 0.00 0 >> TOTAL 134G 43925M 94244M 31.79 >> MIN/MAX VAR: 0.00/2.99 STDDEV: 44.85 >> >> I setup this cluster by manipulating CRUSH map using CLI. I had a default >> root >> before but it gave me an impression that since every rack is under a >> single >> root bucket its marking entire cluster down in case one of the osd is 95% >> full. So I >> removed root bucket but that still did not help me. No crush rule is >> referring >> to root bucket in the above mentioned case. >> >> Yes, I added one osd under two racks by linking host bucket from one rack >> to another >> using following command - >> >> "osd crush link <name> <args> [<args>...] : link existing entry for >> <name> under location <args>" >> >> >> On Thu, Aug 10, 2017 at 1:40 PM, Peter Maloney >> <peter.malo...@brockmann-consult.de> wrote: >>> >>> I think a `ceph osd df` would be useful. >>> >>> And how did you set up such a cluster? I don't see a root, and you have >>> each osd in there more than once...is that even possible? >>> >>> >>> >>> On 08/10/17 08:46, Mandar Naik wrote: >>> >>> Hi, >>> >>> I am evaluating ceph cluster for a solution where ceph could be used for >>> provisioning >>> >>> pools which could be either stored local to a node or replicated across a >>> cluster. This >>> >>> way ceph could be used as single point of solution for writing both local >>> as well as replicated >>> >>> data. Local storage helps avoid possible storage cost that comes with >>> replication factor of more >>> >>> than one and also provide availability as long as the data host is alive. >>> >>> >>> So I tried an experiment with Ceph cluster where there is one crush rule >>> which replicates data across >>> >>> nodes and other one only points to a crush bucket that has local ceph >>> osd. Cluster configuration >>> >>> is pasted below. >>> >>> >>> Here I observed that if one of the disk is full (95%) entire cluster goes >>> into error state and stops >>> >>> accepting new writes from/to other nodes. So ceph cluster became unusable >>> even though it’s only >>> >>> 32% full. The writes are blocked even for pools which are not touching >>> the full osd. >>> >>> >>> I have tried playing around crush hierarchy but it did not help. So is it >>> possible to store data in the above >>> >>> manner with Ceph ? If yes could we get cluster state in usable state >>> after one of the node is full ? >>> >>> >>> >>> # ceph df >>> >>> >>> GLOBAL: >>> >>> SIZE AVAIL RAW USED %RAW USED >>> >>> 134G 94247M 43922M 31.79 >>> >>> >>> # ceph –s >>> >>> >>> cluster ba658a02-757d-4e3c-7fb3-dc4bf944322f >>> >>> health HEALTH_ERR >>> >>> 1 full osd(s) >>> >>> full,sortbitwise,require_jewel_osds flag(s) set >>> >>> monmap e3: 3 mons at >>> {ip-10-0-9-122=10.0.9.122:6789/0,ip-10-0-9-146=10.0.9.146:6789/0,ip-10-0-9-210=10.0.9.210:6789/0} >>> >>> election epoch 14, quorum 0,1,2 >>> ip-10-0-9-122,ip-10-0-9-146,ip-10-0-9-210 >>> >>> osdmap e93: 3 osds: 3 up, 3 in >>> >>> flags full,sortbitwise,require_jewel_osds >>> >>> pgmap v630: 384 pgs, 6 pools, 43772 MB data, 18640 objects >>> >>> 43922 MB used, 94247 MB / 134 GB avail >>> >>> 384 active+clean >>> >>> >>> # ceph osd tree >>> >>> >>> ID WEIGHT TYPE NAME UP/DOWN REWEIGHT PRIMARY-AFFINITY >>> >>> -9 0.04399 rack ip-10-0-9-146-rack >>> >>> -8 0.04399 host ip-10-0-9-146 >>> >>> 2 0.04399 osd.2 up 1.00000 1.00000 >>> >>> -7 0.04399 rack ip-10-0-9-210-rack >>> >>> -6 0.04399 host ip-10-0-9-210 >>> >>> 1 0.04399 osd.1 up 1.00000 1.00000 >>> >>> -5 0.04399 rack ip-10-0-9-122-rack >>> >>> -3 0.04399 host ip-10-0-9-122 >>> >>> 0 0.04399 osd.0 up 1.00000 1.00000 >>> >>> -4 0.13197 rack rep-rack >>> >>> -3 0.04399 host ip-10-0-9-122 >>> >>> 0 0.04399 osd.0 up 1.00000 1.00000 >>> >>> -6 0.04399 host ip-10-0-9-210 >>> >>> 1 0.04399 osd.1 up 1.00000 1.00000 >>> >>> -8 0.04399 host ip-10-0-9-146 >>> >>> 2 0.04399 osd.2 up 1.00000 1.00000 >>> >>> >>> # ceph osd crush rule list >>> >>> [ >>> >>> "rep_ruleset", >>> >>> "ip-10-0-9-122_ruleset", >>> >>> "ip-10-0-9-210_ruleset", >>> >>> "ip-10-0-9-146_ruleset" >>> >>> ] >>> >>> >>> # ceph osd crush rule dump rep_ruleset >>> >>> { >>> >>> "rule_id": 0, >>> >>> "rule_name": "rep_ruleset", >>> >>> "ruleset": 0, >>> >>> "type": 1, >>> >>> "min_size": 1, >>> >>> "max_size": 10, >>> >>> "steps": [ >>> >>> { >>> >>> "op": "take", >>> >>> "item": -4, >>> >>> "item_name": "rep-rack" >>> >>> }, >>> >>> { >>> >>> "op": "chooseleaf_firstn", >>> >>> "num": 0, >>> >>> "type": "host" >>> >>> }, >>> >>> { >>> >>> "op": "emit" >>> >>> } >>> >>> ] >>> >>> } >>> >>> >>> # ceph osd crush rule dump ip-10-0-9-122_ruleset >>> >>> { >>> >>> "rule_id": 1, >>> >>> "rule_name": "ip-10-0-9-122_ruleset", >>> >>> "ruleset": 1, >>> >>> "type": 1, >>> >>> "min_size": 1, >>> >>> "max_size": 10, >>> >>> "steps": [ >>> >>> { >>> >>> "op": "take", >>> >>> "item": -5, >>> >>> "item_name": "ip-10-0-9-122-rack" >>> >>> }, >>> >>> { >>> >>> "op": "chooseleaf_firstn", >>> >>> "num": 0, >>> >>> "type": "host" >>> >>> }, >>> >>> { >>> >>> "op": "emit" >>> >>> } >>> >>> ] >>> >>> } >>> >>> >>> >>> -- >>> Thanks, >>> Mandar Naik. >>> >>> >>> _______________________________________________ >>> ceph-users mailing list >>> ceph-users@lists.ceph.com >>> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >>> >>> >>> -- >>> >>> -------------------------------------------- >>> Peter Maloney >>> Brockmann Consult >>> Max-Planck-Str. 2 >>> 21502 Geesthacht >>> Germany >>> Tel: +49 4152 889 300 >>> Fax: +49 4152 889 333 >>> E-mail: peter.malo...@brockmann-consult.de >>> Internet: http://www.brockmann-consult.de >>> -------------------------------------------- >> >> >> >> >> -- >> Thanks, >> Mandar Naik. > > > > > -- > Thanks, > Mandar Naik. > > _______________________________________________ > ceph-users mailing list > ceph-users@lists.ceph.com > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > _______________________________________________ ceph-users mailing list ceph-users@lists.ceph.com http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com