[
https://issues.apache.org/jira/browse/HDFS-10289?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15243740#comment-15243740
]
Ming Ma commented on HDFS-10289:
--------------------------------
There was some discussion about moving balancer into namenode in
https://issues.apache.org/jira/browse/HDFS-1431. Maybe we can address the issue
[~kihwal] brought up, say to have some lightweight balancer CLI send NN
balancer commands via RPC. Even with that there will be complexity to move
balancer into namenode. Thus we would like to try out the "block movement
scheduling inside NN" idea as part of the migrator HDFS-8789 [~ctrezzo] is
working on.
> Balancer configures DNs directly
> --------------------------------
>
> Key: HDFS-10289
> URL: https://issues.apache.org/jira/browse/HDFS-10289
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: balancer & mover
> Affects Versions: 2.6.0
> Reporter: John Zhuge
> Assignee: John Zhuge
> Priority: Critical
>
> Balancer directly configures the 2 balance-related properties
> (bandwidthPerSec and concurrentMoves) on the DNs involved.
> Details:
> * Before each balancing iteration, set the properties on all DNs involved in
> the current iteration.
> * The DN property changes will not survive restart.
> * Balancer gets the property values from command line or its config file.
> * Need new DN APIs to query and set the 2 properties.
> * No need to edit the config file on each DN or run {{hdfs dfsadmin
> -setBalancerBandwidth}} to configure every DN in the cluster.
> Pros:
> * Improve ease of use because all configurations are done at one place, the
> balancer. We saw many customers often forgot to set concurrentMoves properly
> since it is required on both DN and Balancer.
> * Support new DNs added between iterations
> * Handle DN restarts between iterations
> * May be able to dynamically adjust the thresholds in different iterations.
> Don't know how useful though.
> Cons:
> * New DN property API
> * A malicious/misconfigured balancer may overwhelm DNs. {{hdfs dfsadmin
> -setBalancerBandwidth}} has the same issue. Also Balancer can only be run by
> admin.
> Questions:
> * Can we create {{BalancerConcurrentMovesCommand}} similar to
> {{BalancerBandwidthCommand}}? Can Balancer use them directly without going
> through NN?
> One proposal to implement HDFS-7466 calls for an API to query DN properties.
> DN Conf Servlet returns all config properties. It does not return individual
> property and it does not return the value set by {{hdfs dfsadmin
> -setBalancerBandwidth}}.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
