[ https://issues.apache.org/jira/browse/HDFS-101?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12792261#action_12792261 ]
Todd Lipcon commented on HDFS-101: ---------------------------------- I just tried changing that if statement to == 0 instead of > 0, and it seems to have fixed the bug for me. I reran the above test and it successfully ejected the full node: 09/12/17 19:59:21 WARN hdfs.DFSClient: DFSOutputStream ResponseProcessor exception for block blk_-1132852588861426806_1405java.io.IOException: Bad response 1 for block blk_-1132852588861426806_1405 from datanode 10.251.43.82:50010 at org.apache.hadoop.hdfs.DFSClient$DFSOutputStream$ResponseProcessor.run(DFSClient.java:2427) 09/12/17 19:59:21 WARN hdfs.DFSClient: Error Recovery for block blk_-1132852588861426806_1405 bad datanode[1] 10.251.43.82:50010 09/12/17 19:59:21 WARN hdfs.DFSClient: Error Recovery for block blk_-1132852588861426806_1405 in pipeline 10.250.7.148:50010, 10.251.43.82:50010, 10.251.66.212:50010: bad datanode 10.251.43.82:50010 Is it possible that clientName.length() would ever be equal to t0 in handleMirrorOutError? I'm new to this area of the code, so I may be missing something, but as I understand it, clientName.length() is 0 only for inter-datanode replication requests. For those writes, I didn't think any pipelining was done. > DFS write pipeline : DFSClient sometimes does not detect second datanode > failure > --------------------------------------------------------------------------------- > > Key: HDFS-101 > URL: https://issues.apache.org/jira/browse/HDFS-101 > Project: Hadoop HDFS > Issue Type: Bug > Affects Versions: 0.20.1 > Reporter: Raghu Angadi > Assignee: Hairong Kuang > Priority: Blocker > Fix For: 0.21.0 > > Attachments: detectDownDN-0.20.patch, detectDownDN.patch, > detectDownDN1.patch, hdfs-101.tar.gz > > > When the first datanode's write to second datanode fails or times out > DFSClient ends up marking first datanode as the bad one and removes it from > the pipeline. Similar problem exists on DataNode as well and it is fixed in > HADOOP-3339. From HADOOP-3339 : > "The main issue is that BlockReceiver thread (and DataStreamer in the case of > DFSClient) interrupt() the 'responder' thread. But interrupting is a pretty > coarse control. We don't know what state the responder is in and interrupting > has different effects depending on responder state. To fix this properly we > need to redesign how we handle these interactions." > When the first datanode closes its socket from DFSClient, DFSClient should > properly read all the data left in the socket.. Also, DataNode's closing of > the socket should not result in a TCP reset, otherwise I think DFSClient will > not be able to read from the socket. -- This message is automatically generated by JIRA. - You can reply to this email to add a comment to the issue online.