[ https://issues.apache.org/jira/browse/HDFS-7702?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14297657#comment-14297657 ]
Charles Lamb commented on HDFS-7702: ------------------------------------ Hi [~xiyunyue], I read over your proposal and have some high level questions. I am unclear about your proposal in the failure scenarios. If a source or target NN or one or more of the DNs fails in the middle of a migration, how are things restarted? Why use Kryo and not protobuf for serialization? Why use Kryo and not the existing Hadoop/HDFS protocols and infrastructure for network communications between the various nodes? Is the transfer granularity blockpool only? I infer that from this statement: bq. The target namenode will notify datanode remove blockpool id which belong to the source namenode, but then this statement: bq. it will mark delete the involved sub-tree from its own namespace leads me to believe that it's sub-trees in the namespace. Could you please clarify this statement: bq. all read and write operation regarding the same namespace sub-tree is forwarding to the target namenode. Who does the forwarding, the client or the source NN? > Move metadata across namenode - Effort to a real distributed namenode > --------------------------------------------------------------------- > > Key: HDFS-7702 > URL: https://issues.apache.org/jira/browse/HDFS-7702 > Project: Hadoop HDFS > Issue Type: New Feature > Reporter: Ray > Assignee: Ray > > Implement a tool can show in memory namespace tree structure with > weight(size) and a API can move metadata across different namenode. The > purpose is moving data efficiently and faster, without moving blocks on > datanode. -- This message was sent by Atlassian JIRA (v6.3.4#6332)