Re: dealing with spillovers

[Reed Dier <reed.dier@xxxxxxxxxxx>]

Thanks for sticking with me Igor.

Attached is the ceph-kvstore-tool stats output.

Hopefully something interesting in here.

Thanks,

Reed

Attachment: kvstoretool.log
Description: Binary data

On Jun 12, 2020, at 6:56 AM, Igor Fedotov <ifedotov@xxxxxxx> wrote:

Hi Reed,

thanks for the log.

Nothing much of interest there though. Just a regular SST file that RocksDB instructed to put at "slow" device. Presumably it belongs to a higher level hence the desire to put it that "far". Or (which is less likely) RocksDB lacked free space when doing compaction at some point and spilled some data out. So I was wrong - ceph-kvstore's stats command output might be helpful...

Thanks,

Igor

On 6/11/2020 5:14 PM, Reed Dier wrote:

Apologies for the delay Igor,

Hopefully you are still interested in taking a look.

Attached is the bluestore bluefs-log-dump output.

I gzipped it as the log was very large.

Let me know if there is anything else I can do to help track this down.

Thanks,

Reed

On Jun 8, 2020, at 8:04 AM, Igor Fedotov <ifedotov@xxxxxxx> wrote:

Reed,

No, "ceph-kvstore-tool stats" isn't be of any interest.

For the sake of better issue understanding it might be interesting to have bluefs log dump obtained via ceph-bluestore-tool's bluefs-log-dump command. This will give some insight what RocksDB files are spilled over.  It's still not clear what's the root cause for the issue. It's not that frequent and dangerous though so no active investigation on that...

Wondering if migration has helped though?

Thanks,

Igor

On 6/6/2020 8:00 AM, Reed Dier wrote:

The WAL/DB was part of the OSD deployment.

OSD is running 14.2.9.

Would grabbing the ceph-kvstore-tool bluestore-kv <path-to-osd> stats as in that ticket be of any usefulness to this?

Thanks,

Reed

On Jun 5, 2020, at 5:27 PM, Igor Fedotov <ifedotov@xxxxxxx> wrote:

This might help -see comment #4 at https://tracker.ceph.com/issues/44509

And just for the sake of information collection - what Ceph version is used in this cluster?

Did you setup DB volume along with OSD deployment or they were added later as  was done in the ticket above?

Thanks,

Igor

On 6/6/2020 1:07 AM, Reed Dier wrote:

I'm going to piggy back on this somewhat.

I've battled RocksDB spillovers over the course of the life of the cluster since moving to bluestore, however I have always been able to compact it well enough.

But now I am stumped at getting this to compact via $ceph tell osd.$osd compact, which has always worked in the past.

No matter how many times I compact it, I always spill over exactly 192KiB.

BLUEFS_SPILLOVER BlueFS spillover detected on 1 OSD(s)

     osd.36 spilled over 192 KiB metadata from 'db' device (26 GiB used of 34 GiB) to slow device

     osd.36 spilled over 192 KiB metadata from 'db' device (16 GiB used of 34 GiB) to slow device

     osd.36 spilled over 192 KiB metadata from 'db' device (22 GiB used of 34 GiB) to slow device

     osd.36 spilled over 192 KiB metadata from 'db' device (13 GiB used of 34 GiB) to slow device

The multiple entries are from different time trying to compact it.

The OSD is a 1.92TB SATA SSD, the WAL/DB is a 36GB partition on NVMe.

I tailed and tee'd the OSD's logs during a manual compaction here: https://pastebin.com/bcpcRGEe

This is with the normal logging level.

I have no idea how to make heads or tails of that log data, but maybe someone can figure out why this one OSD just refuses to compact?

OSD is 14.2.9.

OS is U18.04.

Kernel is 4.15.0-96.

I haven't played with ceph-bluestore-tool or ceph-kvstore-tool but after seeing the above mention in this thread, I do see ceph-kvstore-tool <rocksdb|bluestore-kv?> compact, which sounds like it may be the same thing that ceph tell compact does under the hood?

compact

Subcommand compact is used to compact all data of kvstore. It will open the database, and trigger a database's compaction. After compaction, some disk space may be released.

Also, not sure if this is helpful:

osd.36 spilled over 192 KiB metadata from 'db' device (13 GiB used of 34 GiB) to slow device

ID   CLASS WEIGHT    REWEIGHT SIZE    RAW USE  DATA    OMAP    META    AVAIL   %USE  VAR  PGS STATUS TYPE NAME

  36   ssd   1.77879  1.00000 1.8 TiB  1.2 TiB 1.2 TiB 6.2 GiB 7.2 GiB 603 GiB 66.88 0.94  85     up             osd.36

You can see the breakdown between OMAP data and META data.

After compacting again:

osd.36 spilled over 192 KiB metadata from 'db' device (26 GiB used of 34 GiB) to slow device

ID   CLASS WEIGHT    REWEIGHT SIZE    RAW USE  DATA    OMAP    META    AVAIL   %USE  VAR  PGS STATUS TYPE NAME

  36   ssd   1.77879  1.00000 1.8 TiB  1.2 TiB 1.2 TiB 6.2 GiB  20 GiB 603 GiB 66.88 0.94  85     up             osd.36

So the OMAP size remained the same, while the metadata ballooned (while still conspicuously spilling over 192KiB exactly)

These OSDs have a few RBD images, cephfs metadata, and librados objects (not RGW) stored.

The breakdown of OMAP size is pretty widely binned, but the GiB sizes are definitely the minority.

Looking at the breakdown with some simple bash-fu

KiB = 147

MiB = 105

GiB = 24

To further divide that, all of the GiB sized OMAPs are SSD OSD's:

	SSD	HDD	TOTAL
KiB	0	147	147
MiB	36	69	105
GiB	24	0	24

I have no idea if any of these data points are pertinent or helpful, but I want to give as clear a picture as possible to prevent chasing the wrong thread.

Appreciate any help with this.

Thanks,

Reed

On May 26, 2020, at 9:48 AM, thoralf schulze <t.schulze@xxxxxxxxxxxx> wrote:

hi there,

trying to get around my head rocksdb spillovers and how to deal with
them … in particular, i have one osds which does not have any pools
associated (as per ceph pg ls-by-osd $osd ), yet it does show up in ceph
health detail as:

    osd.$osd spilled over 2.9 MiB metadata from 'db' device (49 MiB
used of 37 GiB) to slow device

compaction doesn't help. i am well aware of
https://tracker.ceph.com/issues/38745 , yet find it really
counter-intuitive that an empty osd with a more-or-less optimal sized db
volume can't fit its rockdb on the former.

is there any way to repair this, apart from re-creating the osd? fwiw,
dumping the database with

ceph-kvstore-tool bluestore-kv /var/lib/ceph/osd/ceph-$osd dump >
bluestore_kv.dump

yields a file of less than 100mb in size.

and, while we're at it, a few more related questions:

- am i right to assume that the leveldb and rocksdb arguments to
ceph-kvstore-tool are only relevant for osds with filestore-backend?
- does ceph-kvstore-tool bluestore-kv … also deal with rocksdb-items for
osds with bluestore-backend?

thank you very much & with kind regards,
thoralf.

_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx

_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx

Attachment: smime.p7s
Description: S/MIME cryptographic signature

_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx