[ https://issues.apache.org/jira/browse/HDFS-7603?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14297302#comment-14297302 ]
Daryn Sharp commented on HDFS-7603: ----------------------------------- +1 Looks good, understood too hard to write a test to prove/disprove unfair lock starvation under large namespaces. Tested internally. > The background replication queue initialization may not let others run > ---------------------------------------------------------------------- > > Key: HDFS-7603 > URL: https://issues.apache.org/jira/browse/HDFS-7603 > Project: Hadoop HDFS > Issue Type: Bug > Components: rolling upgrades > Reporter: Kihwal Lee > Assignee: Kihwal Lee > Priority: Critical > Attachments: HDFS-7603.patch, HDFS-7603.patch > > > The background replication queue initialization processes configured number > of blocks at a time and releases the namesystem write lock. This was to let > namenode start serving right after a standby to active transition or leaving > safe mode. However, this does not allow others to run much if the lock > fairness is set to "unfair" for the higher throughput. > I propose adding a delay between unlocking and locking in the async repl > queue init thread. -- This message was sent by Atlassian JIRA (v6.3.4#6332)