[ https://issues.apache.org/jira/browse/HDFS-4288?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13531213#comment-13531213 ]
Daryn Sharp commented on HDFS-4288: ----------------------------------- Non-23 branches appear to have a rudimentary fix for this issue. However it appears (haven't tested) that if a node registers in safe mode with 500 blocks, then bounces and re-registers with only 100 blocks because maybe a disk died, the NN will ignore that BR if it's still in safemode. The NN may come out of safe mode w/o realizing it doesn't really have all the replicas it needs, and probably won't issue replications until the next full BR is sent which increases the possibility of data loss. > NN accepts incremental BR as IBR in safemode > -------------------------------------------- > > Key: HDFS-4288 > URL: https://issues.apache.org/jira/browse/HDFS-4288 > Project: Hadoop HDFS > Issue Type: Bug > Components: namenode > Affects Versions: 0.23.0, 2.0.0-alpha, 3.0.0 > Reporter: Daryn Sharp > Assignee: Daryn Sharp > Priority: Critical > > If a DN is ready to send an incremental BR and the NN goes down, the DN will > repeatedly try to reconnect. The NN will then process the DN's incremental > BR as an initial BR. The NN now thinks the DN has only a few blocks, and > will ignore all subsequent BRs from that DN until out of safemode -- which it > may never do because of all the "missing" blocks on the affected DNs. -- 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