Thanks Greg, that helped get the last stuck PGs back online, and everything looks normal again. Here's the promised post-mortem. It might contain only a little of value to developers, but certainly a bunch of face-palming for readers (and a big red face for me). This mess started during a reorganization: migrating OSDs to better disk controllers and from xfs to btrfs, moving journals from unreliable SSDs to partitions on the OSD disks (better SSDs added later), size=4 CRUSH maps distributing two replicas to each of two nodes (see thread from last month) and unrelated things such as a 10gb network upgrade. Most OSDs were migrated in a reasonable way, but for forgotten reasons I took a short-cut with osd.1, and copied the data outside of Ceph from the old xfs partition to the new btrfs partition, figuring if it didn't work, with size=3 there would be two other replicas, right? Indeed, the newly copied OSD did start up and appear to be working. In the meantime, I clobbered two more (full) OSDs by carelessly migrating the journal device from the old SSD to the OSD *data* partition rather than the journal partition. Oops! The filesystems were immediately clobbered upon starting the OSDs, and with two OSDs down and size=3, many PGs had only one replica. While the clobbered OSDs were recovering, I restarted osd.1, apparently functioning normally after its initial start. However, since the OSD's 'cluster' directory didn't exist as a btrfs subvolume, Ceph's snapshots contained the whole filesystem instead of just the data directory. After the restart, ceph-osd tried to roll back the snapshot into the 'cluster' subvolume, but since 'cluster' existed instead as a normal directory, ceph-osd errored out with the problem described in the earlier thread, "error ioctl(BTRFS_IOC_SNAP_CREATE) failed: (17) File exists". At this point, with three OSDs offline, a number of PGs were stuck in 'incomplete' state. I made a careful copy of osd.1 (data + journal) to experiment with. Once the snapshots were sorted out, starting osd.1 failed with complaints about a corrupt journal, which I can't explain except by the recurring pattern of stupidity. Zapping the journal and rolling back the snapshot got the OSD back online, but with the remaining eight 'incomplete' PGs described in the below quoted message. It turns out that Ceph marked these PGs 'incomplete' because there were already partial (with 0MB-100MB of 6GB-14GB each) replicas on other OSDs with newer versions from the brief initial period osd.1 was online, before it was restarted and before its journal was zapped. Even though the complete PG data was still on osd.1, presumably because it was marked as current from earlier epochs, Ceph considered it 'stray'. Greg's tip about the dev tools led to the resolution of these final incomplete PGs. The ceph_objectstore_tool, built from latest git sources, was used to remove the newer, incomplete PG replicas from all OSDs except osd.1. With only the single copy of those PGs on osd.1, Ceph then happily backfilled them to the rest of the cluster. Since the VMs had been shut off during the whole mess, data consistency has apparently been maintained, and everything is back to normal. Whew! I have personally learned a lot of lessons, both about Ceph specifically, and in general about my cowboy approach that led to this fiasco. These don't need to be told here (but readers feeling the urge to lecture me should feel free). If I had any lame 'blame the equipment' excuses, I'd suggest some idiot-proofing: OSDs should refuse to start when the btrfs 'cluster' subvolume doesn't exist, and ceph-osd should refuse to write the journal and data onto the same partition. John On 08/29/2014 12:56 PM, Gregory Farnum wrote: > Hmm, so you've got PGs which are out-of-date on disk (by virtue of being > an older snapshot?) but still have records of them being newer in the > OSD journal? > That's a new failure node for me and I don't think we have any tools > designed for solving. If you can *back up the disk* before doing this, I > think I'd try flushing the OSD journal, then zapping it and creating a > new one. > You'll probably still have an inconsistent store, so you'll then need > to, for each incomplete pg (and for every OSD at the same point on it) > use the OSD store tool extract the "pginfo" entry and rewrite it to the > versions which the pg actually is on disk. > > These are uncharted waters and you're going to need to explore some dev > tools; good luck! > (Unless Sam or somebody has a better solution.) > -Greg > > On Thursday, August 28, 2014, John Morris <john at zultron.com > <mailto:john at zultron.com>> wrote: > > Greg, thanks for the tips in both this and the BTRFS_IOC_SNAP_CREATE > thread. They were enough to get PGs 'incomplete' due to 'not enough > OSDs hosting' resolved by rolling back to a btrfs snapshot. I promise > to write a full post-mortem (embarrassing as it will be) after the > cluster is fully healthy. > > As is my kind of luck, the cluster *also* suffers from eight of the > *other* 'missing log' sort of 'incomplete' PGs: > > 2014-08-28 23:42:03.350612 7f1cc9d82700 20 osd.1 pg_epoch: 13085 > pg[2.5c( v 10143'300715 (9173'297714,10143'300715] local-les=11890 > n=3404 ec=1 les/c 11890/10167 12809/12809/12809) [7,3,0,4] > r=-1 lpr=12809 pi=8453-12808/138 lcod 0'0 > inactive NOTIFY] > handle_activate_map: Not dirtying info: > last_persisted is 13010 while current is 13085 > > The data clearly exists in the renegade OSD's data directory. There > are no reported 'unfound' objects, so 'mark_unfound_lost revert' doesn't > apply. No apparently useful data from 'ceph pg ... query', but an > example is pasted below. > > Since the beginning of the cluster rebuild, all ceph clients have been > turned off, so I believe there's no fear of lost data by reverting to > these PGs, and besides there're always backup tapes. > > How can Ceph be told to accept the versions it sees on osd.1 as the > most current version, and forget the missing log history? > > John > > > $ ceph pg 2.5c query > { "state": "incomplete", > "epoch": 13241, > "up": [ > 7, > 1, > 0, > 4], > "acting": [ > 7, > 1, > 0, > 4], > "info": { "pgid": "2.5c", > "last_update": "10143'300715", > "last_complete": "10143'300715", > "log_tail": "9173'297715", > "last_backfill": "0\/\/0\/\/-1", > "purged_snaps": "[]", > "history": { "epoch_created": 1, > "last_epoch_started": 11890, > "last_epoch_clean": 10167, > "last_epoch_split": 0, > "same_up_since": 13229, > "same_interval_since": 13229, > "same_primary_since": 13118, > "last_scrub": "10029'298459", > "last_scrub_stamp": "2014-08-18 17:36:01.079649", > "last_deep_scrub": "8323'284793", > "last_deep_scrub_stamp": "2014-08-15 17:38:06.229106", > "last_clean_scrub_stamp": "2014-08-18 17:36:01.079649"}, > "stats": { "version": "10143'300715", > "reported_seq": "1764", > "reported_epoch": "13241", > "state": "incomplete", > "last_fresh": "2014-08-29 01:35:44.196909", > "last_change": "2014-08-29 01:22:50.298880", > "last_active": "0.000000", > "last_clean": "0.000000", > "last_became_active": "0.000000", > "last_unstale": "2014-08-29 01:35:44.196909", > "mapping_epoch": 13223, > "log_start": "9173'297715", > "ondisk_log_start": "9173'297715", > "created": 1, > "last_epoch_clean": 10167, > "parent": "0.0", > "parent_split_bits": 0, > "last_scrub": "10029'298459", > "last_scrub_stamp": "2014-08-18 17:36:01.079649", > "last_deep_scrub": "8323'284793", > "last_deep_scrub_stamp": "2014-08-15 17:38:06.229106", > "last_clean_scrub_stamp": "2014-08-18 17:36:01.079649", > "log_size": 3000, > "ondisk_log_size": 3000, > "stats_invalid": "0", > "stat_sum": { "num_bytes": 0, > "num_objects": 0, > "num_object_clones": 0, > "num_object_copies": 0, > "num_objects_missing_on_primary": 0, > "num_objects_degraded": 0, > "num_objects_unfound": 0, > "num_read": 0, > "num_read_kb": 0, > "num_write": 0, > "num_write_kb": 0, > "num_scrub_errors": 0, > "num_shallow_scrub_errors": 0, > "num_deep_scrub_errors": 0, > "num_objects_recovered": 0, > "num_bytes_recovered": 0, > "num_keys_recovered": 0}, > "stat_cat_sum": {}, > "up": [ > 7, > 1, > 0, > 4], > "acting": [ > 7, > 1, > 0, > 4]}, > "empty": 0, > "dne": 0, > "incomplete": 1, > "last_epoch_started": 10278}, > "recovery_state": [ > { "name": "Started\/Primary\/Peering", > "enter_time": "2014-08-29 01:22:50.132826", > "past_intervals": [ > [...] > { "first": 12809, > "last": 13101, > "maybe_went_rw": 1, > "up": [ > 7, > 3, > 0, > 4], > "acting": [ > 7, > 3, > 0, > 4]}, > [...] > ], > "probing_osds": [ > 0, > 1, > 2, > 3, > 4, > 5, > 7], > "down_osds_we_would_probe": [], > "peering_blocked_by": []}, > { "name": "Started", > "enter_time": "2014-08-29 01:22:50.132784"}]} > > > > > On Wed, Aug 27, 2014 at 12:40 PM, Gregory Farnum <greg at inktank.com > <javascript:;>> > wrote: > > > On Tue, Aug 26, 2014 at 10:46 PM, John Morris <john at > zultron.com <http://zultron.com>> > > wrote: > > > In the docs [1], 'incomplete' is defined thusly: > > > > > > Ceph detects that a placement group is missing a necessary period > > > of > > > history from its log. If you see this state, report a bug, > and try > > > to start any failed OSDs that may contain the needed information. > > > > > > However, during an extensive review of list postings related to > > > incomplete PGs, an alternate and oft-repeated definition is > > > something > > > like 'the number of existing replicas is less than the min_size of > > > the > > > pool'. In no list posting was there any acknowledgement of the > > > definition from the docs. > > > > > > While trying to understand what 'incomplete' PGs are, I simply set > > > min_size = 1 on this cluster with incomplete PGs, and they continue > > > to > > > be 'incomplete'. Does this mean that definition #2 is incorrect? > > > > > > In case #1 is correct, how can the cluster be told to forget the > > > lapse > > > in history? In our case, there was nothing writing to the cluster > > > during the OSD reorganization that could have caused this lapse. > > > > Yeah, these two meanings can (unfortunately) both lead to the > > INCOMPLETE state being reported. I think that's going to be fixed in > > our next major release (so that INCOMPLETE means "not enough OSDs > > hosting" and "missing log" will translate into something else), but > > for now the "not enough OSDs" is by far the more common. In your case > > you probably are missing history, but you don't want to recover from > > it using any of the cluster tools because they're likely to lose more > > data than necessary. (Hopefully, you can just roll back to a slightly > > older btrfs snapshot, but we'll see). > > -Greg > _______________________________________________ > ceph-users mailing list > ceph-users at lists.ceph.com <javascript:;> > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > > > > -- > Software Engineer #42 @ http://inktank.com | http://ceph.com
 |