Dear all, In our product env we find a need to do some datanode-level io isolation, and worked out a solution using favored node(HDFS-2576). Simply speaking, we arrange logical datanode group for different applications and make each application only write data to the relative datanode group, so failure of one datanode will only affect one single application.
HDFS-6009 is an umbrella for supplying a bunch of tools after datanodes are grouped, such as balancer within a group, and check and bring cross-group placed block back (favored node cannot control data replication in case of datanode failure). I've uploaded the patch for HDFS-6010, and patches for HDFS-6011/6012 are ready but depends on some new class introduced by HDFS-6010, so will upload them later. I'd really appreciate it if anyone could share his opinion on whether we need these tools in hdfs, or help review patch of HDFS-6010. Thanks. -- Best Regards, Yu Li
