Affects Version/s: 2.6.0
Fix Version/s: None
CDH5.5.2, CentOS 6.7
Somehow (I haven't found root cause yet) we ended up with blocks that have corrupt replicas where the replica count is inconsistent between the blockmap and the corrupt replicas map. If we try to hdfs fsck any parent directory that has a child with one of these blocks, fsck will exit with something like this:
So I start at the top, fscking every subdirectory until I find one or more that fails. Then I do the same thing with those directories (our top level directories all have subdirectories with date directories in them, which then contain the files) and once I find a directory with files in it, I run a checksum of the files in that directory. When I do that, I don't get the name of the file, rather I get:
but since the files are in order, I can figure it out by seeing which file was before the NPE. Once I get to this point, I can see the following in the namenode log when I try to checksum the corrupt file:
2016-08-23 20:24:59,627 WARN org.apache.hadoop.hdfs.server.blockmanagement.BlockManager: Inconsistent number of corrupt replicas for blk_1335893388_1100036319546 blockMap has 0 but corrupt replicas map has 1
2016-08-23 20:24:59,627 WARN org.apache.hadoop.ipc.Server: IPC Server handler 23 on 8020, call org.apache.hadoop.hdfs.protocol.ClientProtocol.getBlockLocations from 192.168.1.100:47785 Call#1 Retry#0
At which point I can delete the file, but it is a very tedious process.
Ideally, shouldn't fsck be able to emit the name of the file that is the source of the problem - and (if -delete is specified) get rid of the file, instead of exiting without saying why?