On 9/26/19 9:45 PM, Eugen Block wrote: Yes, this works since Luminous 12.2.11 k

Update: I expanded all rocksDB devices, but the warnings still appear: BLUEFS_SPILLOVER BlueFS spillover detected on 10 OSD(s)      osd.0 spilled over 2.5 GiB metadata from 'db' device (2.4 GiB used of 30 GiB) to slow device      osd.19 spilled over 66 MiB metadata from 'db' device (818 MiB used of 15 GiB) to slow device      osd.25 spilled over 2.2 GiB metadata from 'db' device (2.6 GiB used of 30 GiB) to slow device      osd.26 spilled over 1.6 GiB metadata from 'db' device (1.9 GiB used of 30 GiB) to slow device      osd.27 spilled over 2.6 GiB metadata from 'db' device (2.5 GiB used of 30 GiB) to slow device      osd.28 spilled over 2.4 GiB metadata from 'db' device (1.3 GiB used of 30 GiB) to slow device      osd.29 spilled over 2.9 GiB metadata from 'db' device (1.7 GiB used of 30 GiB) to slow device      osd.31 spilled over 2.2 GiB metadata from 'db' device (2.7 GiB used of 30 GiB) to slow device      osd.32 spilled over 2.4 GiB metadata from 'db' device (1.7 GiB used of 30 GiB) to slow device      osd.33 spilled over 2.2 GiB metadata from 'db' device (2.0 GiB used of 30 GiB) to slow device OSD.19 can be ignored as it's currently not in use, but the other devices have been expanded from 20 to 30 GB (following the explanations about the compaction levels). According to the OSD logs these are the sizes we're dealing with: Level  Size L0     31.84 L1     183.86 L2     923.67 L3     3.62 Sum    4.74 Int    0.00 Is there any sign that these OSDs would require even larger bdev devices (300GB)? Which would not be possible with the currently used SSDs, unfortunately.

generally expanding existing DB devices isn't enough to immediately eliminate the spillover alert. As spilled over data is already there and doesn't immediately move by such an expansion. Theoretically allert will eventually disappear after RocksDB completely rewrites all the data at slow device. So you should either wait for a while for things to stabilize. May be monitoring spilled over volumes from the alerts which presumably should decrease (or at least doesn't grow). Please note this will most probably happen under some load only. Or migrate bluefs data from main device with ceph-bluestore-tool. Which I'd recommend in case of emergency only.

Try it with a little bit large db devices. The db devices also holds the WAL, which has a default size of 1 GB afaik. And you also need to consider gigabyte vs. gibibytes. We ran into the same problem in our setup..... You might also want to increase the size even further, since rocksdb needs some free space during compaction. The worst case scenario is ~60 GB per device to take compaction into account. If you do not have an extreme metadata/omap data loaded workload, you won't need more capacity for these partitions.

Hi Eugen, generally expanding existing DB devices isn't enough to immediately eliminate the spillover alert. As spilled over data is already there and doesn't immediately move by such an expansion. Theoretically allert will eventually  disappear after RocksDB completely rewrites all the data at slow device. Never experimented how this happens in the reality though... So you should either wait for a while for things to stabilize. May be monitoring spilled over volumes from the alerts which presumably should decrease (or at least doesn't grow). Please note this will most probably happen under some load only. Or migrate bluefs data from main device with ceph-bluestore-tool. Which I'd recommend in case of emergency only. Or try to compact DB with ceph-kvstore-tool. Which is unlikely to help. Thanks, Igor On 9/27/2019 11:54 AM, Eugen Block wrote: > Update: I expanded all rocksDB devices, but the warnings still appear: > > BLUEFS_SPILLOVER BlueFS spillover detected on 10 OSD(s) >      osd.0 spilled over 2.5 GiB metadata from 'db' device (2.4 GiB > used of 30 GiB) to slow device >      osd.19 spilled over 66 MiB metadata from 'db' device (818 MiB > used of 15 GiB) to slow device >      osd.25 spilled over 2.2 GiB metadata from 'db' device (2.6 GiB > used of 30 GiB) to slow device >      osd.26 spilled over 1.6 GiB metadata from 'db' device (1.9 GiB > used of 30 GiB) to slow device >      osd.27 spilled over 2.6 GiB metadata from 'db' device (2.5 GiB > used of 30 GiB) to slow device >      osd.28 spilled over 2.4 GiB metadata from 'db' device (1.3 GiB > used of 30 GiB) to slow device >      osd.29 spilled over 2.9 GiB metadata from 'db' device (1.7 GiB > used of 30 GiB) to slow device >      osd.31 spilled over 2.2 GiB metadata from 'db' device (2.7 GiB > used of 30 GiB) to slow device >      osd.32 spilled over 2.4 GiB metadata from 'db' device (1.7 GiB > used of 30 GiB) to slow device >      osd.33 spilled over 2.2 GiB metadata from 'db' device (2.0 GiB > used of 30 GiB) to slow device > > > OSD.19 can be ignored as it's currently not in use, but the other > devices have been expanded from 20 to 30 GB (following the > explanations about the compaction levels). > According to the OSD logs these are the sizes we're dealing with: > > Level  Size > L0     31.84 > L1     183.86 > L2     923.67 > L3     3.62 > Sum    4.74 > Int    0.00 > > > Is there any sign that these OSDs would require even larger bdev > devices (300GB)? Which would not be possible with the currently > used SSDs, unfortunately. > > Is there anything else I can do without recreating the OSDs? > > Thanks, > Eugen > > > Zitat von Konstantin Shalygin <k0ste(a)k0ste.ru>ru>: > >> On 9/26/19 9:45 PM, Eugen Block wrote: >>> I'm following the discussion for a tracker issue [1] about >>> spillover warnings that affects our upgraded Nautilus cluster. >>> Just to clarify, would a resize of the rocksDB volume (and >>> expanding with 'ceph-bluestore-tool bluefs-bdev-expand...') >>> resolve that or do we have to recreate every OSD? >> >> Yes, this works since Luminous 12.2.11 >> >> >> >> k > > > _______________________________________________ > ceph-users mailing list -- ceph-users(a)ceph.io > To unsubscribe send an email to ceph-users-leave(a)ceph.io

On 9/27/19 3:54 PM, Eugen Block wrote: After expanding you should tell to OSD compact command, like `ceph tell osd.0 compact`. k