[
https://issues.apache.org/jira/browse/HDFS-1431?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14597133#comment-14597133
]
Ming Ma commented on HDFS-1431:
-------------------------------
When I met with [~andrew.wang], [~ctrezzo], [~atm], [~cmccabe] the other day,
we had brief discussion about balancer. To make balancer use
BlockPlacementPolicy, alternatively we can run balancer inside namenode.
Namenode already has the necessary information. It needs to provide balancer
throttling with some refactoring. But overall it seems it shouldn't create much
overhead on namenode. It will be great to heard from others about this approach
on potential issues such as scale and performance.
> Balancer should work with the logic of BlockPlacementPolicy
> -----------------------------------------------------------
>
> Key: HDFS-1431
> URL: https://issues.apache.org/jira/browse/HDFS-1431
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: balancer & mover
> Affects Versions: 0.22.0
> Reporter: Scott Chen
> Assignee: Scott Chen
> Attachments: HDFS-1431.txt
>
>
> Currently Balancer does not obtain information from BlockPlacementPolicy so
> it can transfer the blocks without checking with BlockPlacementPolicy.
> This causes the policy break after balancing the cluster.
> There are some new policies proposed in HDFS-1094 and MAPREDUCE-1831 in which
> the block placement follows some pattern.
> The pattern can be broken by Balancer.
> I propose that we add the following method in BlockPlacementPolicy:
> {code}
> abstract public boolean canBeMoved(String fileName, Block block,
> DatanodeInfo source, DatanodeInfo destination);
> {code}
> And make Balancer use it in
> {code}
> private boolean isGoodBlockCandidate(Source source,
> BalancerDatanode target, BalancerBlock block)
> {code}
> What do you think?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
