[ https://issues.apache.org/jira/browse/HDFS-7996?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14387466#comment-14387466 ]
Colin Patrick McCabe edited comment on HDFS-7996 at 4/3/15 9:13 PM: -------------------------------------------------------------------- The {{BlockReceiver}} should either be closed, or not closed. Having lots of subtly different close functions just makes it hard to think about. It would be much better to add a function like {{BlockReceiver#claimReplicaHandler}} and have that function return {{BlockReceiver#replicaHandler}} (and set {{BlockReceiver#replicaHandler}} to null). Then {{finalizeBlock}} can hold on to the {{replicaHandler}} and its associated volume for as long as it wants. I also notice an existing bug in {{BlockReceiver#close}}... if there is an {{IOException}} closing the data or metadata streams, the volume reference will never be removed. We should fix this with a finally block. Similarly, if there is an IOE closing one stream, we should try to close the other stream rather than giving up. Otherwise we have a file descriptor leak on IO error. [Edit: actually, the streams will be closed on an IOE, but not on a RuntimeException/] was (Author: cmccabe): The {{BlockReceiver}} should either be closed, or not closed. Having lots of subtly different close functions just makes it hard to think about. It would be much better to add a function like {{BlockReceiver#claimReplicaHandler}} and have that function return {{BlockReceiver#replicaHandler}} (and set {{BlockReceiver#replicaHandler}} to null). Then {{finalizeBlock}} can hold on to the {{replicaHandler}} and its associated volume for as long as it wants. I also notice an existing bug in {{BlockReceiver#close}}... if there is an {{IOException}} closing the data or metadata streams, the volume reference will never be removed. We should fix this with a finally block. Similarly, if there is an IOE closing one stream, we should try to close the other stream rather than giving up. Otherwise we have a file descriptor leak on IO error. > After swapping a volume, BlockReceiver reports ReplicaNotFoundException > ----------------------------------------------------------------------- > > Key: HDFS-7996 > URL: https://issues.apache.org/jira/browse/HDFS-7996 > Project: Hadoop HDFS > Issue Type: Bug > Components: datanode > Affects Versions: 2.6.0 > Reporter: Lei (Eddy) Xu > Assignee: Lei (Eddy) Xu > Priority: Critical > Attachments: HDFS-7996.000.patch, HDFS-7996.001.patch, > HDFS-7996.002.patch > > > When removing a disk from an actively writing DataNode, the BlockReceiver > working on the disk throws {{ReplicaNotFoundException}} because the replicas > are removed from the memory: > {code} > 2015-03-26 08:02:43,154 INFO > org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.FsDatasetImpl: Removed > volume: /data/2/dfs/dn/current > 2015-03-26 08:02:43,163 INFO org.apache.hadoop.hdfs.server.common.Storage: > Removing block level storage: > /data/2/dfs/dn/current/BP-51301509-10.20.202.114-1427296597742 > 2015-03-26 08:02:43,163 WARN org.apache.hadoop.hdfs.server.datanode.DataNode: > IOException in BlockReceiver.run(): > org.apache.hadoop.hdfs.server.datanode.ReplicaNotFoundException: Cannot > append to a non-existent replica > BP-51301509-10.20.202.114-1427296597742:blk_1073742979_2160 > at > org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.FsDatasetImpl.getReplicaInfo(FsDatasetImpl.java:615) > at > org.apache.hadoop.hdfs.server.datanode.fsdataset.impl.FsDatasetImpl.finalizeBlock(FsDatasetImpl.java:1362) > at > org.apache.hadoop.hdfs.server.datanode.BlockReceiver$PacketResponder.finalizeBlock(BlockReceiver.java:1281) > at > org.apache.hadoop.hdfs.server.datanode.BlockReceiver$PacketResponder.run(BlockReceiver.java:1241) > at java.lang.Thread.run(Thread.java:745) > {code} > {{FsVolumeList#removeVolume}} waits all threads release {{FsVolumeReference}} > on the volume to be removed, however, in {{PacketResponder#finalizeBlock()}}, > it calls > {code} > private void finalizeBlock(long startTime) throws IOException { > BlockReceiver.this.close(); > final long endTime = ClientTraceLog.isInfoEnabled() ? System.nanoTime() > : 0; > block.setNumBytes(replicaInfo.getNumBytes()); > datanode.data.finalizeBlock(block); > {code} > The {{FsVolumeReference}} was released in {{BlockReceiver.this.close()}} > before calling {{datanode.data.finalizeBlock(block)}}. -- This message was sent by Atlassian JIRA (v6.3.4#6332)