Btw: After the storm, I highly suggest you to consider to use Jumboframe.
It works like a charm.

[]'s
Arthur (aKa Guilherme Geronimo)
On 04/09/2019 15:50, Guilherme Geronimo wrote:

I see that you have many inactive PGs, probably because the 6 OSD OUT+DOWN.

Problems with "flapping" OSD I use to solved:

* setting NOUP flag
* restarting the "fragile" OSDs
* check if everything is ok look ing their logs
* taking off the NOUP flag

Another solution is:
* Setting NOIN and NOUP flag
* Taking the fragile OSD out
* restarting the "fragile" OSDs
* check if everything is ok look ing their logs
* taking off the NOUP flag
* Take a coffee and wait till all data are drain

[]'s
Arthur (aKa Guilherme Geronimo)
On 04/09/2019 15:32, Bryan Stillwell wrote:
We are not using jumbo frames anywhere on this cluster (all mtu 1500).  The cluster was originally built in October of 2016 and has the following history:

2016-10-04: Created with Hammer (0.94.3)
2017-05-03: Upgraded to Hammer (0.94.10)
2017-10-09: Upgraded to Jewel (10.2.9)
2017-11-02: Upgraded to Jewel (10.2.10)
2018-04-30: Upgraded to Luminous (12.2.5)
2018-09-05: Upgraded to Luminous (12.2.8)
2019-04-05: Upgraded to Luminous (12.2.11)
2019-04-18: Upgraded to Luminous (12.2.12)
2019-07-26: Upgraded to Nautilus (14.2.2)

It wasn't until after the Nautilus upgrade when this problem started showing up.

Here's the output you requested:

[root@a2mon002 ~]# ceph -s
  cluster:
    id:     XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX
    health: HEALTH_ERR
            nodown,norebalance,noscrub,nodeep-scrub flag(s) set
            1 nearfull osd(s)
            19 pool(s) nearfull
            1 scrub errors
            Reduced data availability: 6014 pgs inactive, 3 pgs down, 5958 pgs peering, 83 pgs stale
            Possible data damage: 1 pg inconsistent
            Degraded data redundancy: 1601/81648846 objects degraded (0.002%), 4 pgs degraded, 5 pgs undersized
            1048 slow requests are blocked > 32 sec

  services:
    mon: 3 daemons, quorum a2mon002,a2mon003,a2mon004 (age 17m)
    mgr: a2mon004(active, since 53m), standbys: a2mon003, a2mon002
    mds: cephfs:2 {0=a2mon004=up:active(laggy or crashed),1=a2mon003=up:active(laggy or crashed)} 1 up:standby
    osd: 143 osds: 141 up, 137 in; 486 remapped pgs
         flags nodown,norebalance,noscrub,nodeep-scrub

  data:
    pools:   20 pools, 6288 pgs
    objects: 27.22M objects, 98 TiB
    usage:   308 TiB used, 114 TiB / 422 TiB avail
    pgs:     0.048% pgs unknown
             95.611% pgs not active
             1601/81648846 objects degraded (0.002%)
             53012/81648846 objects misplaced (0.065%)
             5379 peering
             495  remapped+peering
             269  active+clean
             75   stale+peering
             46   activating
             7    stale+remapped+peering
             3    unknown
             3    active+undersized+degraded
             3    down
             2    activating+remapped
             1    activating+undersized
             1    active+clean+scrubbing
             1    remapped+inconsistent+peering
             1    activating+undersized+degraded
             1    stale+activating
             1    creating+peering

[root@a2mon002 ~]# ceph versions
{
    "mon": {
        "ceph version 14.2.2 (4f8fa0a0024755aae7d95567c63f11d6862d55be) nautilus (stable)": 3
    },
    "mgr": {
        "ceph version 14.2.2 (4f8fa0a0024755aae7d95567c63f11d6862d55be) nautilus (stable)": 3
    },
    "osd": {
        "ceph version 14.2.2 (4f8fa0a0024755aae7d95567c63f11d6862d55be) nautilus (stable)": 141
    },
    "mds": {
        "ceph version 14.2.2 (4f8fa0a0024755aae7d95567c63f11d6862d55be) nautilus (stable)": 1
    },
    "overall": {
        "ceph version 14.2.2 (4f8fa0a0024755aae7d95567c63f11d6862d55be) nautilus (stable)": 148
    }
}


We had seen the slow peering shortly after the Nautilus upgrade, but it eventually recovered.  We then started filling the cluster up to test another Nautilus bug (https://tracker.ceph.com/issues/41255), but then a disk started to die (which caused the inconsistent PG).  When we marked it out we ran into this peering problem again, but it seems much worse this time.

Thanks,
Bryan

On Sep 4, 2019, at 11:55 AM, Guilherme Geronimo <guilherme.geronimo@gmail.com> wrote:

Notice: This email is from an external sender.



Hey Bryan,

I suppose all nodes are using jumboframes (mtu 9000), right?
I would suggest to check OSD->MON communication.

Can you send the output os these commands for us?
* ceph -s
* ceph versions

[]'s
Arthur (aKa Guilherme Geronimo)

On 04/09/2019 14:18, Bryan Stillwell wrote:
Our test cluster is seeing a problem where peering is going incredibly slow shortly after upgrading it to Nautilus (14.2.2) from Luminous (12.2.12).

From what I can tell it seems to be caused by "wait for new map" taking a long time.  When looking at dump_historic_slow_ops on pretty much any OSD I see stuff like this:

# ceph daemon osd.112 dump_historic_slow_ops
[...snip...]
       {
           "description": "osd_pg_create(e180614 287.4b:177739 287.75:177739 287.1c3:177739 287.1cf:177739 287.1e1:177739 287.2dd:177739 287.2fc:177739 287.342:177739 287.382:177739)",
           "initiated_at": "2019-09-03 15:12:41.366514",
           "age": 4800.8847047119998,
           "duration": 4780.0579745630002,
           "type_data": {
               "flag_point": "started",
               "events": [
                   {
                       "time": "2019-09-03 15:12:41.366514",
                       "event": "initiated"
                   },
                   {
                       "time": "2019-09-03 15:12:41.366514",
                       "event": "header_read"
                   },
                   {
                       "time": "2019-09-03 15:12:41.366501",
                       "event": "throttled"
                   },
                   {
                       "time": "2019-09-03 15:12:41.366547",
                       "event": "all_read"
                   },
                   {
                       "time": "2019-09-03 15:39:03.379456",
                       "event": "dispatched"
                   },
                   {
                       "time": "2019-09-03 15:39:03.379477",
                       "event": "wait for new map"
                   },
                   {
                       "time": "2019-09-03 15:39:03.522376",
                       "event": "wait for new map"
                   },
                   {
                       "time": "2019-09-03 15:53:55.912499",
                       "event": "wait for new map"
                   },
                   {
                       "time": "2019-09-03 15:59:37.909063",
                       "event": "wait for new map"
                   },
                   {
                       "time": "2019-09-03 16:00:43.356023",
                       "event": "wait for new map"
                   },
                   {
                       "time": "2019-09-03 16:20:50.575498",
                       "event": "wait for new map"
                   },
                   {
                       "time": "2019-09-03 16:31:48.689415",
                       "event": "started"
                   },
                   {
                       "time": "2019-09-03 16:32:21.424489",
                       "event": "done"
                   }
               ]
           }

It always seems to be in osd_pg_create() with multiple "wait for new map" messages before it finally does something.  What could be causing it so long to get the OSD map?  The mons don't appear to be overloaded in any way.

Thanks,
Bryan
_______________________________________________
ceph-users mailing list -- ceph-users@ceph.io
To unsubscribe send an email to ceph-users-leave@ceph.io
_______________________________________________
ceph-users mailing list -- ceph-users@ceph.io
To unsubscribe send an email to ceph-users-leave@ceph.io