[ https://issues.apache.org/jira/browse/HBASE-9268?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13746169#comment-13746169 ]
Nicolas Liochon commented on HBASE-9268: ---------------------------------------- Hum. Different points: - 38 is about the number of puts that have failed with a SocketTimeout. As it's a multi put, it's likely to be a single message. It does not mean that the client retried 38 times. - we do a socket#setSoTimeout, but this is only for reads, not for write. - it's not possible to do write timeout in java w/o using nio API. - HDFS added SocketOutputStream back in HADOOP-2346, but HBase does not use it. - The API to use is NetUtils.getOutputStream(socket, timeout); Tested, it works. - We can use it, but the API does not allows to change the timeout on the fly as we do. - I'm not sure of the time needed by ZooKeeper to decide that the server was dead. The tests were strange. So, synthesis is: - Looking at the code, I don't think it's a new issue. JD, what do you think? - It seems we can fix or improve things here. I will give it a try. - I need to double check the zookeeper stuff. > Client doesn't recover from a stalled region server > --------------------------------------------------- > > Key: HBASE-9268 > URL: https://issues.apache.org/jira/browse/HBASE-9268 > Project: HBase > Issue Type: Bug > Affects Versions: 0.95.2 > Reporter: Jean-Daniel Cryans > Assignee: Nicolas Liochon > Fix For: 0.98.0, 0.95.3 > > Attachments: 9268-hack.patch > > > Got this testing the 0.95.2 RC. > I killed -STOP a region server and let it stay like that while running PE. > The clients didn't find the new region locations and in the jstack were stuck > doing RPC. Eventually I killed -CONT and the client printed these: > bq. Exception in thread "TestClient-6" java.lang.RuntimeException: > org.apache.hadoop.hbase.client.RetriesExhaustedWithDetailsException: Failed > 128 actions: IOException: 90 times, SocketTimeoutException: 38 times, -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira