Ceph

I met the error too. (submit_transaction_sync always return -1)

I added logs in RocksDBStore::submit_transaction_sync , found out rocksdb return error:
2016-04-08 08:46:02.064797 7f1093600700 -1 rocksdb: submit_transaction_sync IO error: /opt/ceph-bluestore/var/lib/blue/osd/osd4/db/024586.log: No such file or directory

My ceph version is 10.0.5
BlueStore_bluefs = false

rocksdb include these files:
[root@lj20 db]# pwd
/opt/ceph-bluestore/var/lib/blue/osd/osd4/db
[root@lj20 db]# ls
006585.sst 019654.sst 020975.sst 021884.sst 022248.sst 023086.sst 023565.sst 024015.sst 024205.sst 024421.sst 024487.sst 024523.sst 024540.sst 024556.sst 024571.sst LOCK
015462.sst 019826.sst 020976.sst 021886.sst 022249.sst 023087.sst 023566.sst 024018.sst 024206.sst 024422.sst 024488.sst 024525.sst 024541.sst 024557.sst 024572.sst MANIFEST-021513
016094.sst 019921.sst 021164.sst 021998.sst 022310.sst 023088.sst 023567.sst 024019.sst 024207.sst 024423.sst 024489.sst 024526.sst 024543.sst 024558.sst 024573.sst OPTIONS-021510
016181.sst 020179.sst 021165.sst 021999.sst 022311.sst 023197.sst 023568.sst 024054.sst 024268.sst 024424.sst 024508.sst 024527.sst 024544.sst 024559.sst 024574.sst OPTIONS-021516
016183.sst 020502.sst 021166.sst 022000.sst 022312.sst 023198.sst 023780.sst 024055.sst 024269.sst 024425.sst 024511.sst 024528.sst 024545.sst 024560.sst 024575.sst
016224.sst 020503.sst 021167.sst 022001.sst 022313.sst 023309.sst 023781.sst 024080.sst 024270.sst 024427.sst 024512.sst 024529.sst 024546.sst 024561.sst 024576.sst
016386.sst 020504.sst 021334.sst 022002.sst 022403.sst 023313.sst 023782.sst 024082.sst 024353.sst 024455.sst 024513.sst 024530.sst 024547.sst 024562.sst 024577.sst
017622.sst 020725.sst 021386.sst 022003.sst 022829.sst 023314.sst 023783.sst 024135.sst 024399.sst 024479.sst 024514.sst 024531.sst 024548.sst 024563.sst 024578.sst
017845.sst 020726.sst 021387.sst 022005.sst 022969.sst 023498.sst 023784.sst 024136.sst 024400.sst 024480.sst 024516.sst 024532.sst 024549.sst 024564.sst 024579.sst
017846.sst 020727.sst 021388.sst 022066.sst 022970.sst 023499.sst 023962.sst 024137.sst 024401.sst 024481.sst 024517.sst 024534.sst 024550.sst 024565.sst 024580.sst
017912.sst 020728.sst 021453.sst 022067.sst 022971.sst 023500.sst 023964.sst 024186.sst 024402.sst 024482.sst 024518.sst 024535.sst 024551.sst 024566.sst 024581.sst
019294.sst 020790.sst 021609.sst 022068.sst 023081.sst 023560.sst 023965.sst 024200.sst 024404.sst 024483.sst 024519.sst 024536.sst 024552.sst 024567.sst 024584.log
019295.sst 020854.sst 021670.sst 022070.sst 023082.sst 023561.sst 023966.sst 024201.sst 024405.sst 024484.sst 024520.sst 024537.sst 024553.sst 024568.sst 024585.sst
019420.sst 020973.sst 021671.sst 022245.sst 023083.sst 023562.sst 023967.sst 024202.sst 024406.sst 024485.sst 024521.sst 024538.sst 024554.sst 024569.sst CURRENT
019421.sst 020974.sst 021883.sst 022246.sst 023084.sst 023564.sst 023968.sst 024204.sst 024407.sst 024486.sst 024522.sst 024539.sst 024555.sst 024570.sst IDENTITY
[root@lj20 db]# ls al | grep 'log'
-rw-r--r- 1 root root 4015151 Apr 8 08:46 024584.log

Hi guys, I am trying to fix this problem by introducing a mechanism of _txc_abort(). It will reply an -EIO to client if we met any problem in transaction submitting.

does that sounds good enough?

This is rocksdb's bug.
Rocksdb report:
2016-04-14 14:37:13.705520 7f0d4f4e3700 2 rocksdb: Waiting after background compaction error: IO error: /opt/ceph-bluestore/var/lib/blue/osd/osd1/db/407669.log: No such file or directory, Accumulated background error counts: 1

Return -EIO to client maybe is not proper.

shasha lu wrote:

This is rocksdb's bug.
Rocksdb report:
2016-04-14 14:37:13.705520 7f0d4f4e3700 2 rocksdb: Waiting after background compaction error: IO error: /opt/ceph-bluestore/var/lib/blue/osd/osd1/db/407669.log: No such file or directory, Accumulated background error counts: 1

I am not sure is that a bug in socksdb here, but I am sure objectstore has no method to abort a transaction. That's what I am working for.

Return -EIO to client maybe is not proper.

Why not, please consider this scenario, there is something wrong in the device of rocksdb and we got an EIO from writing data into it. Then we return a -EIO to user, why not proper?

https://github.com/ceph/ceph/pull/8599

Hi, guys, does this commit solve it?

yes, osd will abort when submit_transaction met error.

BTW, the default bluestore rocksdb options is
bluestore_rocksdb_options = compression=kNoCompression,max_write_buffer_number=16,min_write_buffer_number_to_merge=3,recycle_log_file_num=16

I remove the option recycle_log_file_num=16, with ceph.conf
bluestore_rocksdb_options = compression=kNoCompression,max_write_buffer_number=16,min_write_buffer_number_to_merge=3

The error no longer appears.
Maybe the options recycle_log_file_num should be killed in config_opts.h