[
https://issues.apache.org/jira/browse/HBASE-15019?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Matteo Bertozzi updated HBASE-15019:
------------------------------------
Attachment: HBASE-15019-v1_branch-1.2.patch
I haven't found any good way to write a unit test for this, but it's fairly
easy to reproduce it manually. but Replication IT with DataNode killing should
do the job.
the simplest way to reproduce it constantly to me was:
replace the this.writer.close() in FSHLog and throw an exception to get to the
catch "Riding over failed WAL close".
{code}
try {
if (this.writer != null) {
Trace.addTimelineAnnotation("closing writer");
- this.writer.close();
- Trace.addTimelineAnnotation("writer closed");
+ //this.writer.close();
+ this.closeErrorCount.set(0);
+ throw new IOException("writer not closed");
+ //Trace.addTimelineAnnotation("writer closed");
}
- this.closeErrorCount.set(0);
} catch (IOException ioe) {
{code}
add a peer, and do some puts. (it's easier to reproduce with archived files.
so, roll/flush a lot)
{code}
create 'testtb', {NAME => 'f', REPLICATION_SCOPE => 1}
for i in 1..1000
put 'testtb', "row-#{i}", 'f:q', "value-#{i}"
sleep 4
if i % 2 == 0
wal_roll 'localhost,16201,1451331420432'
flush 'testtb'
end
end
{code}
then stop/restart HDFS. and the "Try to recover the WAL lease" from the
replication source will show up. and rows will be replicated.
> Replication stuck when HDFS is restarted
> ----------------------------------------
>
> Key: HBASE-15019
> URL: https://issues.apache.org/jira/browse/HBASE-15019
> Project: HBase
> Issue Type: Bug
> Components: Replication, wal
> Affects Versions: 2.0.0, 1.2.0, 1.1.2, 1.0.3, 0.98.16.1
> Reporter: Matteo Bertozzi
> Assignee: Matteo Bertozzi
> Attachments: HBASE-15019-v0_branch-1.2.patch,
> HBASE-15019-v1_branch-1.2.patch
>
>
> RS is normally working and writing on the WAL.
> HDFS is killed and restarted, and the RS try to do a roll.
> The close fail, but the roll succeed (because hdfs is now up) and everything
> works.
> {noformat}
> 2015-12-11 21:52:28,058 ERROR
> org.apache.hadoop.hbase.regionserver.wal.ProtobufLogWriter: Got IOException
> while writing trailer
> java.io.IOException: All datanodes 10.51.30.152:50010 are bad. Aborting...
> at
> org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer.setupPipelineForAppendOrRecovery(DFSOutputStream.java:1147)
> at
> org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer.processDatanodeError(DFSOutputStream.java:945)
> at
> org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer.run(DFSOutputStream.java:496)
> 2015-12-11 21:52:28,059 ERROR
> org.apache.hadoop.hbase.regionserver.wal.FSHLog: Failed close of HLog writer
> java.io.IOException: All datanodes 10.51.30.152:50010 are bad. Aborting...
> at
> org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer.setupPipelineForAppendOrRecovery(DFSOutputStream.java:1147)
> at
> org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer.processDatanodeError(DFSOutputStream.java:945)
> at
> org.apache.hadoop.hdfs.DFSOutputStream$DataStreamer.run(DFSOutputStream.java:496)
> 2015-12-11 21:52:28,059 WARN org.apache.hadoop.hbase.regionserver.wal.FSHLog:
> Riding over HLog close failure! error count=1
> {noformat}
> The problem is on the replication side. that log we rolled and we were not
> able to close
> is waiting for a lease recovery.
> {noformat}
> 2015-12-11 21:16:31,909 ERROR
> org.apache.hadoop.hbase.regionserver.wal.HLogFactory: Can't open after 267
> attempts and 301124ms
> {noformat}
> the WALFactory notify us about that, but there is nothing on the RS side that
> perform the WAL recovery.
> {noformat}
> 2015-12-11 21:11:30,921 WARN
> org.apache.hadoop.hbase.regionserver.wal.HLogFactory: Lease should have
> recovered. This is not expected. Will retry
> java.io.IOException: Cannot obtain block length for
> LocatedBlock{BP-1547065147-10.51.30.152-1446756937665:blk_1073801614_61243;
> getBlockSize()=83; corrupt=false; offset=0; locs=[10.51.30.154:50010,
> 10.51.30.152:50010, 10.51.30.155:50010]}
> at
> org.apache.hadoop.hdfs.DFSInputStream.readBlockLength(DFSInputStream.java:358)
> at
> org.apache.hadoop.hdfs.DFSInputStream.fetchLocatedBlocksAndGetLastBlockLength(DFSInputStream.java:300)
> at org.apache.hadoop.hdfs.DFSInputStream.openInfo(DFSInputStream.java:237)
> at org.apache.hadoop.hdfs.DFSInputStream.<init>(DFSInputStream.java:230)
> at org.apache.hadoop.hdfs.DFSClient.open(DFSClient.java:1448)
> at
> org.apache.hadoop.hdfs.DistributedFileSystem$3.doCall(DistributedFileSystem.java:301)
> at
> org.apache.hadoop.hdfs.DistributedFileSystem$3.doCall(DistributedFileSystem.java:297)
> at
> org.apache.hadoop.fs.FileSystemLinkResolver.resolve(FileSystemLinkResolver.java:81)
> at
> org.apache.hadoop.hdfs.DistributedFileSystem.open(DistributedFileSystem.java:297)
> at org.apache.hadoop.fs.FilterFileSystem.open(FilterFileSystem.java:161)
> at org.apache.hadoop.fs.FileSystem.open(FileSystem.java:766)
> at
> org.apache.hadoop.hbase.regionserver.wal.HLogFactory.createReader(HLogFactory.java:116)
> at
> org.apache.hadoop.hbase.regionserver.wal.HLogFactory.createReader(HLogFactory.java:89)
> at
> org.apache.hadoop.hbase.regionserver.wal.HLogFactory.createReader(HLogFactory.java:77)
> at
> org.apache.hadoop.hbase.replication.regionserver.ReplicationHLogReaderManager.openReader(ReplicationHLogReaderManager.java:68)
> at
> org.apache.hadoop.hbase.replication.regionserver.ReplicationSource.openReader(ReplicationSource.java:508)
> at
> org.apache.hadoop.hbase.replication.regionserver.ReplicationSource.run(ReplicationSource.java:321)
> {noformat}
> the only way to trigger a WAL recovery is to restart and force the master to
> trigger the lease recovery on WAL split.
> but there is a case where restarting will not help. If the RS keeps going
> rolling and flushing the unclosed WAL will be moved in the archive, and at
> that point the master will never try to do a lease recovery on it.
> since we know that the RS is still going, should we try to recover the lease
> on the RS side?
> is it better/safer to trigger an abort on the RS, so we have only the master
> doing lease recovery?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
