[HBASE-11620] Record the class name of Writer in WAL header so that only proper Reader can open the WAL file

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 0.98.4
• Fix Version/s: 0.99.0, 0.98.5
• Hadoop Flags: Reviewed
• Assignee: Ted Yu
• Reporter: Ted Yu
• Created: 30/Jul/14 20:11
• Updated: 06/Apr/18 17:50
• Resolved: 01/Aug/14 17:23

Description

Reported by Kiran in this thread: "HBase file encryption, inconsistencies observed and data loss"

After step 4 ( i.e disabling of WAL encryption, removing SecureProtobufReader/Writer and restart), read of encrypted WAL fails mainly due to EOF exception at Basedecoder. This is not considered as error and these WAL are being moved to /oldWALs.

Following is observed in log files:

2014-07-30 19:44:29,254 INFO  [RS_LOG_REPLAY_OPS-HOST-16:15264-1] wal.HLogSplitter: Splitting hlog: hdfs://HOST-16:18020/hbase/WALs/HOST-16,15264,1406725441997-splitting/HOST-16%2C15264%2C1406725441997.1406725444017, length=172
2014-07-30 19:44:29,254 INFO  [RS_LOG_REPLAY_OPS-HOST-16:15264-1] wal.HLogSplitter: DistributedLogReplay = false
2014-07-30 19:44:29,313 INFO  [RS_LOG_REPLAY_OPS-HOST-16:15264-1] util.FSHDFSUtils: Recovering lease on dfs file hdfs://HOST-16:18020/hbase/WALs/HOST-16,15264,1406725441997-splitting/HOST-16%2C15264%2C1406725441997.1406725444017
2014-07-30 19:44:29,315 INFO  [RS_LOG_REPLAY_OPS-HOST-16:15264-1] util.FSHDFSUtils: recoverLease=true, attempt=0 on file=hdfs://HOST-16:18020/hbase/WALs/HOST-16,15264,1406725441997-splitting/HOST-16%2C15264%2C1406725441997.1406725444017 after 1ms
2014-07-30 19:44:29,429 DEBUG [RS_LOG_REPLAY_OPS-HOST-16:15264-1-Writer-0] wal.HLogSplitter: Writer thread Thread[RS_LOG_REPLAY_OPS-HOST-16:15264-1-Writer-0,5,main]: starting
2014-07-30 19:44:29,429 DEBUG [RS_LOG_REPLAY_OPS-HOST-16:15264-1-Writer-1] wal.HLogSplitter: Writer thread Thread[RS_LOG_REPLAY_OPS-HOST-16:15264-1-Writer-1,5,main]: starting
2014-07-30 19:44:29,430 DEBUG [RS_LOG_REPLAY_OPS-HOST-16:15264-1-Writer-2] wal.HLogSplitter: Writer thread Thread[RS_LOG_REPLAY_OPS-HOST-16:15264-1-Writer-2,5,main]: starting
2014-07-30 19:44:29,591 ERROR [RS_LOG_REPLAY_OPS-HOST-16:15264-1] codec.BaseDecoder: Partial cell read caused by EOF: java.io.IOException: Premature EOF from inputStream
2014-07-30 19:44:29,592 INFO  [RS_LOG_REPLAY_OPS-HOST-16:15264-1] wal.HLogSplitter: Finishing writing output logs and closing down.
2014-07-30 19:44:29,592 INFO  [RS_LOG_REPLAY_OPS-HOST-16:15264-1] wal.HLogSplitter: Waiting for split writer threads to finish
2014-07-30 19:44:29,592 INFO  [RS_LOG_REPLAY_OPS-HOST-16:15264-1] wal.HLogSplitter: Split writers finished
2014-07-30 19:44:29,592 INFO  [RS_LOG_REPLAY_OPS-HOST-16:15264-1] wal.HLogSplitter: Processed 0 edits across 0 regions; log file=hdfs://HOST-16:18020/hbase/WALs/HOST-16,15264,1406725441997-splitting/HOST-16%2C15264%2C1406725441997.1406725444017 is corrupted = false progress failed = false

To fix this, we need to propagate EOF exception to HLogSplitter. Any suggestions on the fix?

-------- (end of quote from Kiran)

In BaseDecoder#rethrowEofException() :

    if (!isEof) throw ioEx;
    LOG.error("Partial cell read caused by EOF: " + ioEx);
    EOFException eofEx = new EOFException("Partial cell read");
    eofEx.initCause(ioEx);
    throw eofEx;

throwing EOFException would not propagate the "Partial cell read" condition to HLogSplitter which doesn't treat EOFException as an error.

I think IOException should be thrown above - HLogSplitter#getNextLogLine() would translate the IOEx to CorruptedLogFileException.