Affects Version/s: 0.9.2, 0.12.0
Fix Version/s: 0.12.0
Cluster of 10 machines, running Hadoop 0.9.2 + Nutch
Cluster is set up with tasktrackers running on the same machines as datanodes. Tasks create heavy load in terms of local CPU/RAM/diskIO. I noticed a lot of the following messages from the datanodes in such situations:
2007-02-15 05:30:53,298 WARN dfs.DataNode - Failed to transfer blk_-4590782726923911824 to xxx.xxx.xxx/10.10.16.109:50010
java.net.SocketException: Connection reset
java.io.IOException: Block blk_71053993347675204 has already been started (though not completed), and thus cannot be created.
My reading of the code in DataNode.DataXceiver.writeBlock() and FSDataset.writeToBlock() + FSDataset.java:459 suggests the following scenario: there is no cleanup of temporary files in /tmp that are used to store the incomplete blocks being transferred. If the datanode is CPU-starved and drops the connection while creating this temp file, the source datanode will attempt to transfer it again - but there is already a file under this name in /tmp, because when the connection was dropped the target datanode didn't bother to cleanup.
I also see that this section is unchanged in trunk/.
The solution to this would be to check the age of the physical file in the /tmp dir, in FSDataset.java:436 - if it's older than a few hours or so, we should delete it and proceed as if there were no ongoing create op for this block.