[ https://issues.apache.org/jira/browse/HDFS-9650?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15102436#comment-15102436 ]
Frode Halvorsen commented on HDFS-9650: --------------------------------------- I have an addition; When I 'cleaned out' the datanode, it started out with 0 blocks, and did of course not do any harm to the name-node when restarting. Today I tried to restart it again, after it had recieved a few blocks, and the log on the namenode started getting those masseages again. This time, however, the datanode didn't have enough blocks to spam the namenode long enough to take i down. The namenode just reported those redundant addStoreBlock for the new blocks on the datanode, and finished up before the failovercontroller shut it down... So just now, I have a few datanodes I cannot restart... > Problem is logging of "Redundant addStoredBlock request received" > ----------------------------------------------------------------- > > Key: HDFS-9650 > URL: https://issues.apache.org/jira/browse/HDFS-9650 > Project: Hadoop HDFS > Issue Type: Bug > Reporter: Frode Halvorsen > > Description; > Hadoop 2.7.1. 2 namenodes in HA. 14 Datanodes. > Enough CPU,disk and RAM. > Just discovered that some datanodes must have been corrupted somehow. > When restarting a 'defect' ( works without failure except when restarting) > the active namenode suddenly is logging a lot of : "Redundant addStoredBlock > request received" > and finally the failover-controller takes the namenode down, fails over to > other node. This node also starts logging the same, and as soon as the fisrt > node is bac online, the failover-controller again kill the active node, and > does failover. > This node now was started after the datanode, and doesn't log "Redundant > addStoredBlock request received" anymore, and restart of the second name-node > works fine. > If I again restarts the datanode- the process repeats itself. > Problem is logging of "Redundant addStoredBlock request received" and why > does it happen ? > The failover-controller acts the same way as it did on 2.5/6 when we had a > lot of 'block does not belong to any replica'-messages. Namenode is too busy > to respond to heartbeats, and is taken down... > To resolv this, I have to take down the datanode, delete all data from it, > and start it up. Then cluster will reproduce the missing blocks, and the > failing datanode is working fine again... -- This message was sent by Atlassian JIRA (v6.3.4#6332)