- ceph-users - lists.ceph.io

Yet another meltdown starting
by Frank Schilder 11 May '20

11 May '20

Hi all, another client-load induced meltdown. It is just starting and I hope we get it under control. This time, its the MGRs failing under the load. It looks like thay don't manage to get their beacons to the mons and are kicked out as unresponsive. However, the processes are fine and up. Its just an enormous load. I'm trying to increase # ceph config set global mon_mgr_beacon_grace 90 but the command doesn't complete. I guess because all the MGRs are out. Is there any way to force the MONs *not* to mark MGRs as unresponsive? Best regards, ================= Frank Schilder AIT Risø Campus Bygning 109, rum S14

2 4

Recover datas from pg incomplete
by Francois Legrand 11 May '20

11 May '20

Hi, After a major crash in which we lost few osds, we are stucked with incomplete pgs. At first, peering was blocked with peering_blocked_by_history_les_bound. Thus we set osd_find_best_info_ignore_history_les true for all osds involved in the pg and set the primary osd down to force repeering. It worked for one pg which is in a replica 3 pool, but for the 2 others pgs which are in a erasurce coding (3+2) pool, it didn't worked... and the pgs are still incomplete. We know that we will have data lost, but we would like to minimize it and save as much as possible. Also because this pg is part of the data pool of a cephfs filesystem and it seems that files are spread among a lot of pgs and loosing objects in a pg of the datapool means the loss of a huge number of files ! According to https://www.spinics.net/lists/ceph-devel/msg41665.html a way would be to : - stop each osd involved in that pg - export the shards with ceph-objectstore-tool - compare the size of the shards and select the biggest one (alternatively maybe we can also look at the num_objects returned by ceph pg query ?) - Mark it as complete - restart the osd - Wait for recover and finally get rid of the missing objects with ceph pg 10.2 mark_unfound_lost delete But on this other source https://github.com/TheJJ/ceph-cheatsheet/blob/master/README.md or here https://medium.com/opsops/recovering-ceph-from-reduced-data-availability-3-… it's suggested to remove the other parts (but I am not sure these threads are really related to EC pools). Could you confirm that we could follow this procedure (or correct it or suggests anything else) ? Thanks for your advices. F. PS: Here is a part of the ceph pg 10.2 query return : "state": "incomplete", "snap_trimq": "[]", "snap_trimq_len": 0, "epoch": 434321, "up": [ 78, 105, 90, 4, 41 ], "acting": [ 78, 105, 90, 4, 41 ], "info": { "pgid": "10.2s0", "state": "incomplete", "last_peered": "2020-04-22 09:58:42.505638", "last_became_peered": "2020-04-20 11:06:07.701833", "num_objects": 161314, "num_objects_missing_on_primary": 0, "num_objects_missing": 0, "num_objects_degraded": 0, "num_objects_misplaced": 0, "num_objects_unfound": 0, "num_objects_dirty": 161314, "num_objects_recovered": 1290285, "peer_info": [ "peer": "4(3)", "pgid": "10.2s3", "state": "active+undersized+degraded+remapped+backfilling", "last_peered": "2020-04-25 13:25:12.860435", "last_became_peered": "2020-04-22 10:45:45.520125", "num_objects": 162869, "num_objects_missing_on_primary": 0, "num_objects_missing": 0, "num_objects_degraded": 85071, "num_objects_misplaced": 0, "num_objects_unfound": 0, "num_objects_dirty": 162869, "num_objects_recovered": 1368082, "peer": "9(2)", "pgid": "10.2s2", "state": "down", "last_peered": "2020-04-25 13:25:12.860435", "last_became_peered": "2020-04-22 10:45:45.520125", "num_objects": 162869, "num_objects_missing_on_primary": 0, "num_objects_missing": 0, "num_objects_degraded": 0, "num_objects_misplaced": 0, "num_objects_unfound": 0, "num_objects_dirty": 162869, "num_objects_recovered": 1368082, "peer": "41(4)", "pgid": "10.2s4", "state": "unknown", "last_peered": "0.000000", "last_became_peered": "0.000000", "num_objects": 0, "num_objects_missing_on_primary": 0, "num_objects_missing": 0, "num_objects_degraded": 0, "num_objects_misplaced": 0, "num_objects_unfound": 0, "num_objects_dirty": 0, "num_objects_recovered": 0, "peer": "46(4)", "pgid": "10.2s4", "state": "down", "last_peered": "2020-04-25 13:25:12.860435", "last_became_peered": "2020-04-22 10:45:45.520125", "num_objects": 162869, "num_objects_missing_on_primary": 0, "num_objects_missing": 0, "num_objects_degraded": 0, "num_objects_misplaced": 0, "num_objects_unfound": 0, "num_objects_dirty": 162869, "num_objects_recovered": 1368082, "peer": "52(3)", "pgid": "10.2s3", "state": "unknown", "last_peered": "0.000000", "last_became_peered": "0.000000", "num_objects": 0, "num_objects_missing_on_primary": 0, "num_objects_missing": 0, "num_objects_degraded": 0, "num_objects_misplaced": 0, "num_objects_unfound": 0, "num_objects_dirty": 0, "num_objects_recovered": 0, "peer": "69(0)", "pgid": "10.2s0", "state": "unknown", "last_peered": "0.000000", "last_became_peered": "0.000000", "num_objects": 84063, "num_objects_missing_on_primary": 0, "num_objects_missing": 78807, "num_objects_degraded": 0, "num_objects_misplaced": 0, "num_objects_unfound": 0, "num_objects_dirty": 84063, "num_objects_recovered": 0, "peer": "90(2)", "pgid": "10.2s2", "state": "down", "last_peered": "0.000000", "last_became_peered": "0.000000", "num_objects": 0, "num_objects_missing_on_primary": 0, "num_objects_missing": 0, "num_objects_degraded": 0, "num_objects_misplaced": 0, "num_objects_unfound": 0, "num_objects_dirty": 0, "num_objects_recovered": 0, "peer": "105(1)", "pgid": "10.2s1", "state": "incomplete", "last_peered": "0.000000", "last_became_peered": "0.000000", "num_objects": 0, "num_objects_missing_on_primary": 0, "num_objects_missing": 0, "num_objects_degraded": 0, "num_objects_misplaced": 0, "num_objects_unfound": 0, "num_objects_dirty": 0, "num_objects_recovered": 0, "recovery_state": [ "peering_blocked_by": [] "agent_state": {}

1 0

adding block.db to OSD
by Stefan Priebe - Profihost AG 11 May '20

11 May '20

Hello, is there anything else needed beside running: ceph-bluestore-tool --path /var/lib/ceph/osd/ceph-${OSD} bluefs-bdev-new-db --dev-target /dev/vgroup/lvdb-1 I did so some weeks ago and currently i'm seeing that all osds originally deployed with --block-db show 10-20% I/O waits while all those got converted using ceph-bluestore-tool show 80-100% I/O waits. Also is there some tuning available to use more of the SSD? The SSD (block-db) is only saturated at 0-2%. Greets, Stefan

2 21

Re: Write Caching to hot tier not working as expected
by Steve Hughes 11 May '20

11 May '20

Thank you Eric. That 'sounds like' exactly my issue. Though I'm surprised to bump into something like that on such a small system and at such low bandwidth. But the information I can find on those parameters is sketchy to say the least. Can you point me at some doco that explains what they do, how to read the current values and how to set them? Cheers, Steve -----Original Message----- From: Eric Smith <Eric.Smith(a)vecima.com> Sent: Monday, 11 May 2020 8:00 PM To: Steve Hughes <steveh(a)scalar.com.au>; ceph-users(a)ceph.io Subject: RE: [ceph-users] Re: Write Caching to hot tier not working as expected It sounds like you might be bumping up against the default objecter_inflight_ops (1024) and/or objecter_inflight_op_bytes (100MB). -----Original Message----- From: steveh(a)scalar.com.au <steveh(a)scalar.com.au> Sent: Monday, May 11, 2020 5:48 AM To: ceph-users(a)ceph.io Subject: [ceph-users] Re: Write Caching to hot tier not working as expected Interestingly, I have found that if I limit the rate at which data is written the tiering behaves as expected. I'm using a robocopy job from a Windows VM to copy large files from my existing storage array to a test Ceph volume. By using the /IPG parameter I can roughly control the rate at which data is written. I've found that if I limit the write rate to around 30MBytes/sec the data all goes to the hot tier, zero data goes to the HDD tier, and the observed write latency is about 5msec. If I go any higher than this I see data being written to the HDDs and the observed write latency goes way up. _______________________________________________ ceph-users mailing list -- ceph-users(a)ceph.io To unsubscribe send an email to ceph-users-leave(a)ceph.io --

1 0

ceph-volume/batch fails in non-interactive mode
by Michał Nasiadka 11 May '20

11 May '20

Hello, I stumbled across that weird behaviour today - after initially creating disks and DB LVs on NVMe (using ceph-ansible) - second run of ceph-volume ends up with an error: --> RuntimeError: 4 devices were filtered in non-interactive mode, bailing out What is interesting - the same command in interactive mode (without —yes) ends up without that error. I raised a bug in Tracker - https://tracker.ceph.com/issues/45461 <https://tracker.ceph.com/issues/45461> I found a PR fixing a similar bug here: https://github.com/ceph/ceph/pull/33202 <https://github.com/ceph/ceph/pull/33202> - is that related? Kind regards, Michal

1 0

Write Caching to hot tier not working as expected
by steveh＠scalar.com.au 11 May '20

11 May '20

Hi all, I'm a newbie to Ceph. I'm an MSP and a small-scale cloud hoster. I'm intending to use Ceph as production storage for a small-scale private hosting cloud. We run ESXi as our HVs so we want to present Ceph as iSCSI. We've got Ceph Nautilus running on a 3-node cluster. Each node contains a pair of Bronze XEONs, 128G RAM, 6 x 10G NICs, 8 x 10TB spinners, 2 x 2TB SATA SSDs and a 4TB NVME. The HDDs and the 2TB SSDs give me an rbd pool of 24 OSDs (1024 PGs) with the SSDs partitioned and used to hold the DB and WAL. Each SSD hold the DB/WAL for 4 HDDs. The NVMes give me a cache pool of 3 OSDs (128 PGs) which I want to use as a hot tier. As far as I can tell I have followed the guidance given here: https://docs.ceph.com/docs/nautilus/rados/operations/cache-tiering/ The cluster is working, the iSCSI is working, and generally everything is looking pretty good. My only problem at this stage is that the tiering is not handling writes in the way I expect, and I really can't get my brain around why. For my test I start with the hot tier empty. I drained it by setting dirty_ratio=dirty_high_ratio=full_ratio = 0. I then set dirty_ratio = 0.5, dirty_high_ratio = 0.6 and full_ratio = 0.7 and start writing data to it at high speed (using a simple file copy) from a VM on ESXi. My expectation is that all inbound writes will land initially on the hot tier, resulting in low latency writes as seen by the ESXi hosts. I expect that once the hot tier fills to 50% that it will start to flush these writes down to the HDD storage. What I actually see is that as soon as I start throwing data at the cluster, the Ceph dashboard shows writes going to both the NVMEs and to the HDDs, and the write latency seen by ESX hits several hundred milliseconds. It seems that the hot tier is absorbing only a fraction of the writes. Here are my pool settings. [root@ceph00 ~]# ceph osd pool ls detail pool 4 'rbd' replicated size 3 min_size 2 crush_rule 1 object_hash rjenkins pg_num 1024 pgp_num 1024 autoscale_mode warn last_change 6971 lfor 6971/6971/6971 flags hashpspool,selfmanaged_snaps tiers 11 read_tier 11 write_tier 11 stripe_width 0 application rbd removed_snaps [1~3] pool 11 'cache' replicated size 3 min_size 2 crush_rule 3 object_hash rjenkins pg_num 128 pgp_num 128 autoscale_mode warn last_change 7065 lfor 6971/6971/6971 flags hashpspool,incomplete_clones,selfmanaged_snaps tier_of 4 cache_mode writeback target_bytes 3078632557772 hit_set bloom{false_positive_probability: 0.001, target_size: 0, seed: 0} 3600s x12 decay_rate 0 search_last_n 0 min_read_recency_for_promote 2 stripe_width 0 application rbd removed_snaps [1~3] pool 12 'pure_hdd' replicated size 3 min_size 2 crush_rule 1 object_hash rjenkins pg_num 1024 pgp_num 1024 autoscale_mode warn last_change 7059 flags hashpspool,selfmanaged_snaps stripe_width 0 application rbd removed_snaps [1~3] Is anyone able to point me towards a solution? Thanks, Steve

1 1

radosgw Swift bulk upload
by Martin Zurowietz 11 May '20

11 May '20

Hi, I'm interested in the Swift extract archive/bulk upload feature [1]. First I thought that this feature is not implemented in radosgw, as it is not mentioned anywhere in the docs [2] and some test requests of mine were not successful. However, while browsing the issue tracker, I found some issues that mention a bulk upload feature [3, 4]. Can someone please clarify whether this feature exists in radosgw or not? Thanks a lot, Martin [1]: https://docs.openstack.org/swift/latest/middleware.html#module-swift.common… [2]: https://docs.ceph.com/docs/master/radosgw/swift/ [3]: https://tracker.ceph.com/issues/19645 [4]: https://tracker.ceph.com/issues/8945 -- Biodata Mining Group Faculty of Technology Bielefeld University https://biodatamining.cebitec.uni-bielefeld.de/ Office: V10-103 +49 521 106 60 58

1 0

OSD Inbalance - upmap mode
by Ashley Merrick 11 May '20

11 May '20

I have a cluster running 15.2.1, was originally running 14.x, the cluster is running the balance module in upmap mode (I have tried crush-compat in the past) Most OSD's are around the same & used give or take 0.x, however there is one OSD that is down a good few % and a few that are above average by 1 or 2 %, I have been trying to get the balance to fix this. I have tried running a manual osdmaptool command on an export of my map, but it lists no fixed however does display the underfall OSD in it's output (overfull 3,4,5,6,7,8,9,10,11,12,13,14,15,18,19,20 underfull [36]) The debug output is just lots of: 2020-05-05T06:15:39.172+0000 7f3dfb0c3c40 10 trying 2.55 2020-05-05T06:15:39.172+0000 7f3dfb0c3c40 10 2.55 [12,3,7,6,33,34,30,35,21,18] -> [12,3,7,6,33,34,30,35,21,16] 2020-05-05T06:15:39.172+0000 7f3dfb0c3c40 10 will try adding new remapping pair 18 -> 16 for 2.55 NOT selected osd 2020-05-05T06:15:39.172+0000 7f3dfb0c3c40 10 stddev 528.667 -> 528.667 2020-05-05T06:15:39.172+0000 7f3dfb0c3c40 10 Overfull search osd.7 target 170.667 deviation 9.33327 Is there anything I can to try and balance the overfull onto the underful OSDs to balance out the last bit.

1 1

Elasticsearch Sync module bug ?
by Luca Cervigni 11 May '20

11 May '20

Hello, I have a two zones cluster. Default zone + a ES sync zone where I would like to index metadata and do searches via the ES tier gateway. Struggling to get this service running (documentation is bad and very, very minimal) I thought that in the end I made it work. The requests are getting correctly to the ES zone, I can see the payload sent to ES with debug rgw = 20, ES returns OK then an error is trigger, apparently a JSON parsing error taking the ES search and giving it back to the user. I tried with Nautilus, Octopus, and different versions of ES, from 6 to 7. I have no possibilities of downgrades as our production cluster is on nautilus. Here the logs: https://pastebin.com/hnj8YysK I would think this is a bug, but would like your opinion. If not a bug how shall I solve this? I am using a handmade S3 tool to query as boto is not supported anymore, obo just does not work, and boto3 cannot send any arbitrary requests anymore. Also I would like to know if those Sync modules are still supported or not. The documentation is very poor and if they still are in the roadmap, it should be improved. Thanks Cheers, Luca Cervigni Pawsey Supercomputing centre

1 0

RGW crashed
by Zhenshi Zhou 11 May '20

11 May '20

Hi all, I deployed a multi-site cluster in order to sync object from an old cluster to a brand new cluster. It seems good cause I can see the data syncing. However, when I check the cluster health, it shows warn messages "2 daemons have recently crashed". I get the crash info by 'sudo ceph crash info $id': { "os_version_id": "7", "utsname_release": "3.10.0-957.27.2.el7.x86_64", "os_name": "CentOS Linux", "entity_name": "client.rgw.ceph-node7", "timestamp": "2020-05-09 15:17:59.482502Z", "process_name": "radosgw", "utsname_machine": "x86_64", "utsname_sysname": "Linux", "os_version": "7 (Core)", "os_id": "centos", "utsname_version": "#1 SMP Mon Jul 29 17:46:05 UTC 2019", "backtrace": [ "(()+0xf5f0) [0x7f32b1bdf5f0]", "(RGWCoroutine::set_sleeping(bool)+0xc) [0x555eeb1351ac]", "(RGWOmapAppend::flush_pending()+0x2d) [0x555eeb13acad]", "(RGWOmapAppend::finish()+0x10) [0x555eeb13acd0]", "(RGWDataSyncShardCR::stop_spawned_services()+0x2b) [0x555eeb0a185b]", "(RGWDataSyncShardCR::incremental_sync()+0x72a) [0x555eeb0a9baa]", "(RGWDataSyncShardCR::operate()+0x9d) [0x555eeb0ab33d]", "(RGWCoroutinesStack::operate(RGWCoroutinesEnv*)+0x60) [0x555eeb136520]", "(RGWCoroutinesManager::run(std::list<RGWCoroutinesStack*, std::allocator<RGWCoroutinesStack*> >&)+0x236) [0x555eeb137196]", "(RGWCoroutinesManager::run(RGWCoroutine*)+0x78) [0x555eeb138098]", "(RGWRemoteDataLog::run_sync(int)+0x1cf) [0x555eeb08851f]", "(RGWDataSyncProcessorThread::process()+0x46) [0x555eeb1e71a6]", "(RGWRadosThread::Worker::entry()+0x115) [0x555eeb1b6195]", "(()+0x7e65) [0x7f32b1bd7e65]", "(clone()+0x6d) [0x7f32b10e188d]" ], "utsname_hostname": "ceph-node7", "crash_id": "2020-05-09_15:17:59.482502Z_b80d7bee-faa0-4d2f-9d86-a1b3f4d4802e", "ceph_version": "14.2.8" } AND { "os_version_id": "7", "utsname_release": "3.10.0-957.27.2.el7.x86_64", "os_name": "CentOS Linux", "entity_name": "client.rgw.ceph-node7", "timestamp": "2020-05-10 16:23:13.375063Z", "process_name": "radosgw", "utsname_machine": "x86_64", "utsname_sysname": "Linux", "os_version": "7 (Core)", "os_id": "centos", "utsname_version": "#1 SMP Mon Jul 29 17:46:05 UTC 2019", "backtrace": [ "(()+0xf5f0) [0x7f409f42e5f0]", "(RGWCoroutine::set_sleeping(bool)+0xc) [0x55e3f45e01ac]", "(RGWOmapAppend::flush_pending()+0x2d) [0x55e3f45e5cad]", "(RGWOmapAppend::finish()+0x10) [0x55e3f45e5cd0]", "(RGWDataSyncShardCR::stop_spawned_services()+0x2b) [0x55e3f454c85b]", "(RGWDataSyncShardCR::incremental_sync()+0x72a) [0x55e3f4554baa]", "(RGWDataSyncShardCR::operate()+0x9d) [0x55e3f455633d]", "(RGWCoroutinesStack::operate(RGWCoroutinesEnv*)+0x60) [0x55e3f45e1520]", "(RGWCoroutinesManager::run(std::list<RGWCoroutinesStack*, std::allocator<RGWCoroutinesStack*> >&)+0x236) [0x55e3f45e2196]", "(RGWCoroutinesManager::run(RGWCoroutine*)+0x78) [0x55e3f45e3098]", "(RGWRemoteDataLog::run_sync(int)+0x1cf) [0x55e3f453351f]", "(RGWDataSyncProcessorThread::process()+0x46) [0x55e3f46921a6]", "(RGWRadosThread::Worker::entry()+0x115) [0x55e3f4661195]", "(()+0x7e65) [0x7f409f426e65]", "(clone()+0x6d) [0x7f409e93088d]" ], "utsname_hostname": "ceph-node7", "crash_id": "2020-05-10_16:23:13.375063Z_9e70a0c0-929e-445f-b4cd-8d29e909fe2f", "ceph_version": "14.2.8" } So I fetch and check the file "ceph-client.rgw.ceph-node7.log". The log has huge amount of errors like: -732> 2020-05-09 23:17:53.476 7f328b7ff700 0 RGW-SYNC:data:sync:shard[98]:entry[harbor-registry:f70a5eb9-d88d-42fd-ab4e-d300e97094de.4620.94:23]:bucket[harbor-registry:f70a5eb9-d88d-42fd-ab4e-d300e97094de.4620.94:23]:inc_sync[harbor-registry:f70a5 eb9-d88d-42fd-ab4e-d300e97094de.4620.94:23]: ERROR: lease is not taken, abort AND -723> 2020-05-09 23:17:56.388 7f328b7ff700 5 RGW-SYNC:data:sync:shard[88]:entry[harbor-registry:f70a5eb9-d88d-42fd-ab4e-d300e97094de.4620.94:13]:bucket[harbor-registry:f70a5eb9-d88d-42fd-ab4e-d300e97094de.4620.94:13]: incremental sync on bucket fa iled, retcode=-125 AND -215> 2020-05-09 23:17:58.809 7f328b7ff700 5 RGW-SYNC:data:sync:shard[10]:entry[pf2-harbor-swift:f70a5eb9-d88d-42fd-ab4e-d300e97094de.4608.101:113]:bucket[pf2-harbor-swift:f70a5eb9-d88d-42fd-ab4e-d300e97094de.4608.101:113]: full sync on bucket failed, retcode=-125 AND 2020-05-09 23:18:24.048 7f4085867700 1 robust_notify: If at first you don't succeed: (110) Connection timed out 2020-05-09 23:18:24.048 7f4083863700 0 ERROR: failed to distribute cache for shubei.rgw.log:datalog.sync-status.shard.f70a5eb9-d88d-42fd-ab4e-d300e97094de.5 2020-05-09 23:28:49.181 7f407e859700 1 heartbeat_map reset_timeout 'RGWAsyncRadosProcessor::m_tp thread 0x7f407e859700' had timed out after 600 2020-05-10 03:12:01.905 7f409708a700 -1 received signal: Hangup from killall -q -1 ceph-mon ceph-mgr ceph-mds ceph-osd ceph-fuse radosgw rbd-mirror And finally it crashed. I'm not sure where the problem is. Were the crashes caused by the network? Thanks

1 0