I have SSD and NVMe drives in the cluster ranging from 960GB to 15TB. After we had an unfortunate power event a few weeks ago, the cluster was grinding through misplaced and degraded objects. This prevented rebalancing from happening, which eventually filled several of the 960GB osds completely and knocked them offline. On Fri, Nov 29, 2024 at 6:03 AM Igor Fedotov <igor.fedotov@xxxxxxxx> wrote: > Hi Frederic, > > >My question was more about why bluefs would still fail to allocate 4k > chunks after being allowed to do so by > https://tracker.ceph.com/issues/53466 (John's case with v17.2.6 actually) > My hypothesis is that it's facing real "no-space" case not the one I > explained above. But again - we haven't performed thorough analysis of > John's case, so we are just speculating... > > >Is BlueFS aware of the remaining space and maybe using some sort of > reserved blocks/chunks like other filesystems to handle full/near null > situations ? > > What remaining space are you talking here - if the one consisting of small > (<64K) extents only then BlueFS hopefully uses it since v17.2.6. > Reserving more spare space at Bluefs on mkfs? Previously at Bluestore team > we had discussions on something like that to permit easier recovery from > "no-space" cases. No final solution has been taken yet and in fact this > provides no complete fix for the issue anyway - that spare space might end > at some point as well... > > >If so, then it should never crash, right? > What we have here is not a real crash (although looks like that) - it's an > expected assertion. We just don't have good enough [automatic] scenario to > exit from this state. Hence "crying" about that aloud. > > The problem is that by its design RocksDB has to write out some data > (either due to the need for internal maintenance or to fulfill client > requests) on any update access. So at some point we have no space to record > such a transaction. How one can proceed in that case - refuse to execute it > and return an error? OK, but what's next? Any followup data removal would > need some transactions to be recorded in RocksDB as well. And DB/FS would > need more space for that. Use some reserved spare space? But that's > conceptually similar to stalling OSD writes at some free space threshold we > already have at OSD level - as you can see this protection doesn't work > from time to time. > So IMO it's better polish existing protection means then. > > Not to mention - I'm pretty sure that someone would abuse additional spare > space mechanics if any and finally face the same problem at some point. > Thus triggering another iteration of it.. ;) > > Thanks, > Igor > > > On 29.11.2024 10:20, Frédéric Nass wrote: > > Hi Igor, > > Thank you for taking the time to explains the fragmentation issue. I had > figured out the most part of it by reading the tracker and the PR but it's > always clearer when you explain it. > > My question was more about why bluefs would still fail to allocate 4k > chunks after being allowed to do so by > https://tracker.ceph.com/issues/53466 (John's case with v17.2.6 actually) > > Is BlueFS aware of the remaining space and maybe using some sort of > reserved blocks/chunks like other filesystems to handle full/near null > situations ? If so, then it should never crash, right? > Like other filesystems don't crash, drives's firmwares dont crash, etc. > > Thanks, > Frédéric. > > ----- Le 28 Nov 24, à 12:52, Igor Fedotov <igor.fedotov@xxxxxxxx> > <igor.fedotov@xxxxxxxx> a écrit : > > Hi Frederic, > > here is an overview of the case when BlueFS ıs unable to allocate more > space at main/shared device albeıt free space is available. Below I'm > talking about stuff exısted before fıxıng > https://tracker.ceph.com/issues/53466. > > First of al - BlueFS's minimal allocation unit for shared device was > bluefs_shared_alloc_size (=64K by default). Which means that it was unable > to use e.g. 2x32K or 16x4K chunks when it needed additional 64K bytes. > > Secondly - sometimes RocksDB performs recovery - and some other > maintenance tasks that require space allocation - on startup. Which > evidently triggers allocation of N*64K chunks from shared device. > > Thirdly - a while ago we switched to 4K chunk allocations for user data > (please not confuse with BlueFS allocation). Which potentially could result > ın specific free space fragmentation pattern when there ıs limited (or even > empty) set of long (>=64K) chunks free. Still technically having enough > free space available. E.g. free extent list could look like (off~len, both > in hex): > > 0x0~1000, 0x2000~1000, 0x4000~2000, 0x10000~4000, 0x2000~1000, etc... > > In that case original BlueFS allocator implementation was unable to locate > more free space which in turn was effectively breaking both RockDB and OSD > boot up. > > One should realize that the above free space fragmentation depends on a > bunch of factors, none of which is absolutely dominating: > > 1. how user write/remove objects > > 2. how allocator seeks for free space > > 3. how much free space is available > > So we don't have full control on 1. and 3. and have limited opportunities > in tuning 2. > > Small device sizes and high space utilization severely increase the > probability for the issue to happen but theoretically even a large disk > with mediocre utilization could reach "bad" state over time if used (by > both clients and allocator) "improperly/inefficiently". Hence tuning > thresholds can reduce the issue's probability to occur (at cost of > additional spare space waste) but it isn't a silver bullet. > > https://tracker.ceph.com/issues/53466 fixes (or rather works around) the > issue by allowing BlueFS to use 4K extents. Plus we're working on making > better resulting free space fragmentation on aged OSDs by improving > allocation strategies, e.g. see : > > - https://github.com/ceph/ceph/pull/52489 > > - https://github.com/ceph/ceph/pull/57789 > > - https://github.com/ceph/ceph/pull/60870 > > Hope this is helpful. > > Thanks, > > Igor > > On 27.11.2024 16:31, Frédéric Nass wrote: > > > > ----- Le 27 Nov 24, à 10:19, Igor Fedotov <igor.fedotov@xxxxxxxx> > <igor.fedotov@xxxxxxxx> a écrit : > > Hi Istvan, > > first of all let me make a remark that we don't know why BlueStore is out > of space at John's cluster. > > It's just an unconfirmed hypothesis from Frederic that it's caused by high > fragmentation and BlueFS'es inability to use chunks smaller than 64K. In > fact fragmentation issue is fixed since 17.2.6 so I doubt that's the > problem. > > Hi Igor, > > I wasn't actually pointing this as the root cause (since John's already > using 17.2.6) but more to explain the context, but while we're at it... > > Could you elaborate on circumstances that could prevent BlueFS from being > able to allocate chunks in collocated OSDs scenario? Does this ability > depend on near/full thresholds being reached or not? If so then icreasing > these thresholds by 1-2% may help avoiding the crash, no? > > Also, if BlueFS is aware of these thresholds, shouldn't an OSDs be able to > start and live without crashing even when it's full and simply (maybe > easier said than done...) refuse any I/Os? Sorry for the noob questions. :-) > > This topic is particularly important when using NVMe drives as > 'collocated' OSDs, expecially since they often host critical metadata pools > (cephfs, rgw index). > > Cheers, > Frédéric. > > Thanks, > > Igor > On 27.11.2024 4:01, Szabo, Istvan (Agoda) wrote: > > Hi, > > This issue should not happen anymore from 17.2.8 am I correct? In this > version all the fragmentation issue should have gone even with collocated > wal+db+block. > > ------------------------------ > *From:* Frédéric Nass <frederic.nass@xxxxxxxxxxxxxxxx> > <frederic.nass@xxxxxxxxxxxxxxxx> > *Sent:* Wednesday, November 27, 2024 6:12:46 AM > *To:* John Jasen <jjasen@xxxxxxxxx> <jjasen@xxxxxxxxx> > *Cc:* Igor Fedotov <igor.fedotov@xxxxxxxx> <igor.fedotov@xxxxxxxx>; > ceph-users <ceph-users@xxxxxxx> <ceph-users@xxxxxxx> > *Subject:* Re: down OSDs, Bluestore out of space, unable to > restart > > Email received from the internet. If in doubt, don't click any link nor > open any attachment ! > ________________________________ > > Hi John, > > That's about right. Two potential solutions exist: > 1. Adding a new drive to the server and sharing it for RocksDB metadata, or > 2. Repurposing one of the failed OSDs for the same purpose (if adding more > drives isn't feasible). > > Igor's post #6 [1] explains the challenges with co-located OSDs > (DB+WAL+data on the same device) when they run out of space, where > significant fragmentation occurs and BlueFS and BlueStore block sizes are > misaligned. The solution (included in 17.2.6) was to allow BlueFS to > allocate 4k extents when it couldn't find 64k contiguous extents. However, > it seems that even with this fix, these OSDs still can't boot up. > > Therefore, the recommendation is to extend the RocksDB volume to another > device as a temporary workaround. > > Before proceeding, I recommend checking the failed OSDs' > bluefs_shared_alloc_size value. If it's 64k, you might want to try lowering > this to 32k or even 4k, as some users reported [2] that reducing this value > helped failed OSDs boot up and remain stable for a period of time. Might be > worth checking and trying. > > Regards, > Frédéric. > > [1] https://tracker.ceph.com/issues/53466#note-6 > [2] https://github.com/rook/rook/issues/9885#issuecomment-1761076861"; > > > ________________________________ > De : John Jasen <jjasen@xxxxxxxxx> <jjasen@xxxxxxxxx> > Envoyé : mardi 26 novembre 2024 18:50 > À : Igor Fedotov > Cc: ceph-users > Objet : Re: down OSDs, Bluestore out of space, unable to > restart > > Let me see if I have the approach right'ish: > > scrounge some more disk for the servers with full/down OSDs. > partition the new disks into LVs for each downed OSD. > Attach as a lvm new-db to the downed OSDs. > Restart the OSDs. > Profit. > > Is that about right? > > > On Tue, Nov 26, 2024 at 11:28 AM Igor Fedotov <igor.fedotov@xxxxxxxx> > <igor.fedotov@xxxxxxxx> wrote: > > > Well, so there is a single shared volume (disk) per OSD, right? > > > > If so one can add dedicated DB volume to such an OSD - one done OSD will > > have two underlying devices: main(which is original shared disk) and new > > dedicated DB ones. And hence this will effectively provide additional > > space for BlueFS/RocksDB and permit OSD to start up. > > > > I'm not aware of all the details how to do that with cephadm (or whatever > > RH uses) but on bare metal setup this could be achieved by issuing > > 'ceph-volume lvm new-db' command which will attach new LV (provided by > > user) to specific OSD. > > > > > > Thanks, > > > > Igor > > > > > > On 26.11.2024 19:16, John Jasen wrote: > > > > They're all bluefs_single_shared_device, if I understand your question. > > There's no room left on the devices to expand. > > > > We started at quincy with this cluster, and didn't vary too much from the > > Redhat Ceph storage 6 documentation for setting it up. > > > > > > On Tue, Nov 26, 2024 at 4:48 AM Igor Fedotov <igor.fedotov@xxxxxxxx> > <igor.fedotov@xxxxxxxx> > > wrote: > > > >> Hi John, > >> > >> you haven't described your OSD volume configuration but you might want > >> to try adding standalone DB volume if OSD uses LVM and has single main > >> device only. > >> > >> 'ceph-volume lvm new-db' command is the preferred way of doing that, see > >> > >> https://docs.ceph.com/en/quincy/ceph-volume/lvm/newdb/ > >> > >> > >> Thanks, > >> > >> Igor > >> > >> On 25.11.2024 21:37, John Jasen wrote: > >> > Ceph version 17.2.6 > >> > > >> > After a power loss event affecting my ceph cluster, I've been putting > >> > humpty dumpty back together since. > >> > > >> > One problem I face is that with objects degraded, rebalancing doesn't > >> run > >> > -- and this resulted in several of my fast OSDs filling up. > >> > > >> > I have 8 OSDs currently down, 100% full (exceeding all the full ratio > >> > settings on by default or I toggled to try and keep it together), and > >> when > >> > I try to restart them, they fail out. Is there any way to bring these > >> back > >> > from the dead? > >> > > >> > Here's some interesting output from journalctl -xeu on the failed OSD: > >> > > >> > ceph-osd[2383080]: bluestore::NCB::__restore_allocator::No Valid > >> allocation > >> > info on disk (empty file) > >> > ceph-osd[2383080]: bluestore(/var/lib/ceph/osd/ceph-242) > >> > _init_alloc::NCB::restore_allocator() failed! Run Full Recovery from > >> ONodes > >> > (might take a while) ... > >> > > >> > ceph-osd[2389725]: bluefs _allocate allocation failed, needed 0x3000 > >> > > >> > ceph-6ab85342-53d6-11ee-88a7-e43d1a153e91-osd-242[2389718]: -2> > >> > 2024-11-25T18:31:42.070+0000 7f0adfdef540 -1 bluefs _flush_range_F > >> > allocated: 0x0 offset: 0x0 length: 0x230f > >> > ceph-osd[2389725]: bluefs _flush_range_F allocated: 0x0 offset: 0x0 > >> length: > >> > 0x230f > >> > > >> > Followed quickly by an abort: > >> > > >> > > >> > /home/jenkins-build/build/workspace/ceph-build/ARCH/x86_64/AVAILABLE_ARCH/x86_64/AVAILABLE_DIST/centos8/DIST/centos8/MACHINE_SIZE/gigantic/release/17.2.6/rpm/el8/BUILD/ceph-17.2.6/src/os/bluestore/BlueFS.cc: > >> > In funct> > >> > > >> > > >> > /home/jenkins-build/build/workspace/ceph-build/ARCH/x86_64/AVAILABLE_ARCH/x86_64/AVAILABLE_DIST/centos8/DIST/centos8/MACHINE_SIZE/gigantic/release/17.2.6/rpm/el8/BUILD/ceph-17.2.6/src/os/bluestore/BlueFS.cc: > >> > 3380: ce> > >> > > >> > ceph > >> version > >> > 17.2.6 (d7ff0d10654d2280e08f1ab989c7cdf3064446a5) quincy (stable) > >> > 1: > >> > (ceph::__ceph_abort(char const*, int, char const*, > >> > std::__cxx11::basic_string<char, std::char_traits<char>, > >> > std::allocator<char> > const&)+0xd7) [0x559bf4361d2f] > >> > 2: > >> > (BlueFS::_flush_range_F(BlueFS::FileWriter*, unsigned long, unsigned > >> > long)+0x7a9) [0x559bf4b225f9] > >> > 3: > >> > (BlueFS::_flush_F(BlueFS::FileWriter*, bool, bool*)+0xa2) > >> [0x559bf4b22812] > >> > 4: > >> > (BlueFS::fsync(BlueFS::FileWriter*)+0x8e) [0x559bf4b40c3e] > >> > 5: > >> > (BlueRocksWritableFile::Sync()+0x19) [0x559bf4b51ed9] > >> > 6: > >> > (rocksdb::LegacyWritableFileWrapper::Sync(rocksdb::IOOptions const&, > >> > rocksdb::IODebugContext*)+0x22) [0x559bf507fbd2] > >> > 7: > >> > (rocksdb::WritableFileWriter::SyncInternal(bool)+0x5aa) > [0x559bf51a880a] > >> > 8: > >> > (rocksdb::WritableFileWriter::Sync(bool)+0x100) [0x559bf51aa0a0] > >> > 9: > >> > (rocksdb::SyncManifest(rocksdb::Env*, rocksdb::ImmutableDBOptions > >> const*, > >> > rocksdb::WritableFileWriter*)+0x10b) [0x559bf51a3bfb] > >> > 10: > >> > > >> > (rocksdb::VersionSet::ProcessManifestWrites(std::deque<rocksdb::VersionSet::ManifestWriter, > >> > std::allocator<rocksdb::VersionSet::ManifestWriter> >&, > >> > rocksdb::InstrumentedMutex*, rocksdb::FSDirectory*, bool, rocks> > >> > 11: > >> > > >> > (rocksdb::VersionSet::LogAndApply(rocksdb::autovector<rocksdb::ColumnFamilyData*, > >> > 8ul> const&, rocksdb::autovector<rocksdb::MutableCFOptions const*, > 8ul> > >> > const&, rocksdb::autovector<rocksdb::autovector<rocksdb::> > >> > 12: > >> > (rocksdb::VersionSet::LogAndApply(rocksdb::ColumnFamilyData*, > >> > rocksdb::MutableCFOptions const&, rocksdb::VersionEdit*, > >> > rocksdb::InstrumentedMutex*, rocksdb::FSDirectory*, bool, > >> > rocksdb::ColumnFamilyOptions const> > >> > 13: > >> > (rocksdb::DBImpl::DeleteUnreferencedSstFiles()+0xa30) [0x559bf50bd250] > >> > 14: > >> > (rocksdb::DBImpl::Recover(std::vector<rocksdb::ColumnFamilyDescriptor, > >> > std::allocator<rocksdb::ColumnFamilyDescriptor> > const&, bool, bool, > >> bool, > >> > unsigned long*)+0x13f1) [0x559bf50d3f21] > >> > 15: > >> > (rocksdb::DBImpl::Open(rocksdb::DBOptions const&, > >> > std::__cxx11::basic_string<char, std::char_traits<char>, > >> > std::allocator<char> > const&, > >> std::vector<rocksdb::ColumnFamilyDescriptor, > >> > std::allocator<rocksdb::Colu> > >> > 16: > >> > (rocksdb::DB::Open(rocksdb::DBOptions const&, > >> > std::__cxx11::basic_string<char, std::char_traits<char>, > >> > std::allocator<char> > const&, > >> std::vector<rocksdb::ColumnFamilyDescriptor, > >> > std::allocator<rocksdb::ColumnFa> > >> > 17: > >> > (RocksDBStore::do_open(std::ostream&, bool, bool, > >> > std::__cxx11::basic_string<char, std::char_traits<char>, > >> > std::allocator<char> > const&)+0x77a) [0x559bf503766a] > >> > 18: > >> > (BlueStore::_open_db(bool, bool, bool)+0xbb4) [0x559bf4a4bff4] > >> > 19: > >> > (BlueStore::_open_db_and_around(bool, bool)+0x500) [0x559bf4a766e0] > >> > 20: > >> > (BlueStore::_mount()+0x396) [0x559bf4a795d6] > >> > 21: > >> > (OSD::init()+0x556) [0x559bf44a0eb6] > >> > 22: > main() > >> > 23: > >> > __libc_start_main() > >> > 24: > >> _start() > >> > > >> > *** Caught signal (Aborted) ** > >> > in > thread > >> > 7f0adfdef540 thread_name:ceph-osd > >> > > >> > ceph > >> version > >> > 17.2.6 (d7ff0d10654d2280e08f1ab989c7cdf3064446a5) quincy (stable) > >> > 1: > >> > /lib64/libpthread.so.0(+0x12cf0) [0x7f0addff1cf0] > >> > 2: > >> gsignal() > >> > 3: > abort() > >> > 4: > >> > (ceph::__ceph_abort(char const*, int, char const*, > >> > std::__cxx11::basic_string<char, std::char_traits<char>, > >> > std::allocator<char> > const&)+0x197) [0x559bf4361def] > >> > 5: > >> > (BlueFS::_flush_range_F(BlueFS::FileWriter*, unsigned long, unsigned > >> > long)+0x7a9) [0x559bf4b225f9] > >> > 6: > >> > (BlueFS::_flush_F(BlueFS::FileWriter*, bool, bool*)+0xa2) > >> [0x559bf4b22812] > >> > 7: > >> > (BlueFS::fsync(BlueFS::FileWriter*)+0x8e) [0x559bf4b40c3e] > >> > 8: > >> > (BlueRocksWritableFile::Sync()+0x19) [0x559bf4b51ed9] > >> > 9: > >> > (rocksdb::LegacyWritableFileWrapper::Sync(rocksdb::IOOptions const&, > >> > rocksdb::IODebugContext*)+0x22) [0x559bf507fbd2] > >> > 10: > >> > (rocksdb::WritableFileWriter::SyncInternal(bool)+0x5aa) > [0x559bf51a880a] > >> > 11: > >> > (rocksdb::WritableFileWriter::Sync(bool)+0x100) [0x559bf51aa0a0] > >> > 12: > >> > (rocksdb::SyncManifest(rocksdb::Env*, rocksdb::ImmutableDBOptions > >> const*, > >> > rocksdb::WritableFileWriter*)+0x10b) [0x559bf51a3bfb] > >> > 13: > >> > > >> > (rocksdb::VersionSet::ProcessManifestWrites(std::deque<rocksdb::VersionSet::ManifestWriter, > >> > std::allocator<rocksdb::VersionSet::ManifestWriter> >&, > >> > rocksdb::InstrumentedMutex*, rocksdb::FSDirectory*, bool, rocks> > >> > 14: > >> > > >> > (rocksdb::VersionSet::LogAndApply(rocksdb::autovector<rocksdb::ColumnFamilyData*, > >> > 8ul> const&, rocksdb::autovector<rocksdb::MutableCFOptions const*, > 8ul> > >> > const&, rocksdb::autovector<rocksdb::autovector<rocksdb::> > >> > 15: > >> > (rocksdb::VersionSet::LogAndApply(rocksdb::ColumnFamilyData*, > >> > rocksdb::MutableCFOptions const&, rocksdb::VersionEdit*, > >> > rocksdb::InstrumentedMutex*, rocksdb::FSDirectory*, bool, > >> > rocksdb::ColumnFamilyOptions const> > >> > 16: > >> > (rocksdb::DBImpl::DeleteUnreferencedSstFiles()+0xa30) [0x559bf50bd250] > >> > 17: > >> > (rocksdb::DBImpl::Recover(std::vector<rocksdb::ColumnFamilyDescriptor, > >> > std::allocator<rocksdb::ColumnFamilyDescriptor> > const&, bool, bool, > >> bool, > >> > unsigned long*)+0x13f1) [0x559bf50d3f21] > >> > 18: > >> > (rocksdb::DBImpl::Open(rocksdb::DBOptions const&, > >> > std::__cxx11::basic_string<char, std::char_traits<char>, > >> > std::allocator<char> > const&, > >> std::vector<rocksdb::ColumnFamilyDescriptor, > >> > std::allocator<rocksdb::Colu> > >> > 19: > >> > (rocksdb::DB::Open(rocksdb::DBOptions const&, > >> > std::__cxx11::basic_string<char, std::char_traits<char>, > >> > std::allocator<char> > const&, > >> std::vector<rocksdb::ColumnFamilyDescriptor, > >> > std::allocator<rocksdb::ColumnFa> > >> > 20: > >> > (RocksDBStore::do_open(std::ostream&, bool, bool, > >> > std::__cxx11::basic_string<char, std::char_traits<char>, > >> > std::allocator<char> > const&)+0x77a) [0x559bf503766a] > >> > 21: > >> > (BlueStore::_open_db(bool, bool, bool)+0xbb4) [0x559bf4a4bff4] > >> > 22: > >> > (BlueStore::_open_db_and_around(bool, bool)+0x500) [0x559bf4a766e0] > >> > 23: > >> > (BlueStore::_mount()+0x396) [0x559bf4a795d6] > >> > 24: > >> > (OSD::init()+0x556) [0x559bf44a0eb6] > >> > 25: > main() > >> > 26: > >> > __libc_start_main() > >> > 27: > >> _start() > >> > NOTE: a > >> copy > >> > of the executable, or `objdump -rdS <executable>` is needed to > interpret > >> > this. > >> > _______________________________________________ > >> > 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 > _______________________________________________ > ceph-users mailing list -- ceph-users@xxxxxxx > To unsubscribe send an email to ceph-users-leave@xxxxxxx > > > ------------------------------ > This message is confidential and is for the sole use of the intended > recipient(s). It may also be privileged or otherwise protected by copyright > or other legal rules. If you have received it by mistake please let us know > by reply email and delete it from your system. It is prohibited to copy > this message or disclose its content to anyone. Any confidentiality or > privilege is not waived or lost by any mistaken delivery or unauthorized > disclosure of the message. All messages sent to and from Agoda may be > monitored to ensure compliance with company policies, to protect the > company's interests and to remove potential malware. Electronic messages > may be intercepted, amended, lost or deleted, or contain viruses. > > > > _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
 |