Hi Simon, Do you mean both standalone DB and(!!) standalone WAL devices/partitions by having SSD DB/WAL? If so then BlueFS might eventually overwrite some data at you DB volume with BlueFS log content. Which most probably makes OSD crash and unable to restart one day. This is quite random and not very frequent event which is to some degree dependent from cluster loading. And the period between actual data corruption and any evidence of this is non-zero most of the time - we tend to see it mostly when RocksDB was performing compaction. Other OSD configuration which might suffer from the issue is main device + WAL devices. Much less failure probability exists for main + DB layout. It requires almost full DB to get any chances to appear. Main-only device configurations aren't under the threat as far as I can tell. Thanks, Igor On 11/15/2019 12:40 PM, Simon Ironside wrote: > Hi, > > I have two new-ish 14.2.4 clusters that began life on 14.2.0 , all > with HDD OSDs with SSD DB/WALs but neither have experienced obvious > problems yet. > > What's the impact of this? Does possible data corruption mean possible > silent data corruption? > Or does the corruption cause the OSD failures mentioned on the tracker > and you're basically ok if you either haven't had a failure or if you > keep on top of failures the way you would if they were normal disk > failures? > > Thanks, > Simon > > On 14/11/2019 16:10, Sage Weil wrote: >> Hi everyone, >> >> We've identified a data corruption bug[1], first introduced[2] (by yours >> truly) in 14.2.3 and affecting both 14.2.3 and 14.2.4. The corruption >> appears as a rocksdb checksum error or assertion that looks like >> >> os/bluestore/fastbmap_allocator_impl.h: 750: FAILED >> ceph_assert(available >= allocated) >> >> or in some cases a rocksdb checksum error.  It only affects BlueStore >> OSDs >> that have a separate 'db' or 'wal' device. >> >> We have a fix[3] that is working its way through testing, and will >> expedite the next Nautilus point release (14.2.5) once it is ready. >> >> If you are running 14.2.2 or 14.2.1 and use BlueStore OSDs with >> separate 'db' volumes, you should consider waiting to upgrade >> until 14.2.5 is released. >> >> A big thank you to Igor Fedotov and several *extremely* helpful users >> who >> managed to reproduce and track down this problem! >> >> sage >> >> >> [1] https://tracker.ceph.com/issues/42223 >> [2] >> https://github.com/ceph/ceph/commit/096033b9d931312c0688c2eea7e14626bfde0ad… >> [3] https://github.com/ceph/ceph/pull/31621 >> _______________________________________________ >> ceph-users mailing list -- ceph-users(a)ceph.io >> To unsubscribe send an email to ceph-users-leave(a)ceph.io > _______________________________________________ > ceph-users mailing list -- ceph-users(a)ceph.io > To unsubscribe send an email to ceph-users-leave(a)ceph.io

Hi Simon, On 11/15/2019 6:02 PM, Simon Ironside wrote: right yep For your case it makes sense to check this. And then safely wait for 14.2.5 if its not full. this is probably the last resort when you already start seeing this issue and is absolutely uncomfortable with data loss probability. Not a panacea anyway though as one can have already broken data but still undiscovered data corruption at multiple OSDs b. May be. We tend to see it during DB compaction (mostly triggered by DB write access) but IMO it can be detected during scrubbing and/or store fsck as well. As per multiple reports there are some chances to lose OSD data. E.g. we've got reports about reproducing 1-2 OSD failures per day under some stress(!!!) loading testing. That's probably not the general case and production clusters might suffer from this much less frequently. E.g. for our multiple QA activities we've observed the issue just once since it had been introduced. Anyway it's possible to lose multiple OSDs simultaneously. Probability is rather not that large but it's definitely non-zero. But as fix is almost ready I'd recommend to wait for it and apply ASAP. > Thanks, > Simon. > _______________________________________________ > ceph-users mailing list -- ceph-users(a)ceph.io > To unsubscribe send an email to ceph-users-leave(a)ceph.io

Any word on 14.2.5? Nervously waiting here . . . Thanks, Simon. On 18/11/2019 11:29, Simon Ironside wrote: > I will sit tight and wait for 14.2.5. > > Thanks again, > Simon.