@liewegas . My cluster always met this bug:
40148 2016-12-23 00:09:24.590117 7fed61f8aa40 10 bluefs _read h 0x56294761a600 0x50000~8000 from file(ino 20 size 0x100000 mtime 2016-12-23 00:09:13.077397 bdev 1 extents [1:0x900000+100000])
40149 2016-12-23 00:09:24.590127 7fed61f8aa40 20 bluefs _read left 0xb0000 len 0x8000
40150 2016-12-23 00:09:24.590133 7fed61f8aa40 20 bluefs _read got 32768
40151 2016-12-23 00:09:24.590138 7fed61f8aa40 3 rocksdb: db/000012.log: dropping 743 bytes; Corruption: missing start of fragmented record(2)

I add debug message in rocksb: the type of this fragment is kReclylabelasttype. So this mean rocksb enable log recycling function.

The last flush infos of ino=20 is:
2016-12-23 00:09:14.761784 7f0b1d847700 10 bluefs _flush 0x558649842500 0x4cbfa1cd7 to file(ino 20 size 0x100000 mtime 2016-12-23 00:09:13.077397 bdev 1 extents [1:0x900000+100000])
2016-12-23 00:09:14.761791 7f0b1d847700 10 bluefs _flush_range 0x558649842500 pos 0x4cbfa 0x4cbfa1cd7 to file(ino 20 size 0x100000 mtime 2016-12-23 00:09:13.077397 bdev 1 extents [1:0x900000+100000])
2016-12-23 00:09:14.761813 7f0b1d847700 20 bluefs _flush_range file now file(ino 20 size 0x100000 mtime 2016-12-23 00:09:13.077397 bdev 1 extents [1:0x900000+100000])

ca20994 merged manually to kraken branch

Do you have a procedure for reproducing this?