[
https://issues.apache.org/jira/browse/HDFS-8204?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14508475#comment-14508475
]
Walter Su commented on HDFS-8204:
---------------------------------
Update: This issue is about "DN". HDFS-8147 is about "Storage". It's different.
To [~szetszwo],
>... It only makes balancer scheduling inefficient.
Yes, it is. A little, not much.
For example. A file with "cold" policy, 3 replicas.. Because one DN can only
have one "cold" StorageGroup. There is no chance one DN has 2 replicas. The
trunk code will reject it. So the 001 patch doesn't make balancer inefficient.
"hot", "allssd" are the same.
For example. A file with "onessd" policy, 5 replicas. So there is 1 ssd
replica, 4 disk replicas. Balancer is more likely choose 2 "disk" StorageGroups
to make a (source, target) pair. The trunk code will reject it. So the 001
patch doesn't make balancer inefficient.
Only when choosing ("ssd" StorageGroup, "disk" StorageGroup) to make a
(source, target) pair, *trunk will accept it while the 001 patch reject it.*
For example. A file with "onessd" policy, 3 replicas. So there is 1 ssd
replica, 2 disk replicas. What if the cluster is midium/huge? So many "disk"
StorageGroups, balancer can choose (source, target) whatever it want. So it wil
not be inefficient.
*What's more important, 2 replicas located on 1 DN doesn't match placement
policy.*
> Mover should not schedule two replicas to the same DN
> -----------------------------------------------------
>
> Key: HDFS-8204
> URL: https://issues.apache.org/jira/browse/HDFS-8204
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: balancer & mover
> Reporter: Walter Su
> Assignee: Walter Su
> Attachments: HDFS-8204.001.patch
>
>
> Balancer moves blocks between Datanode(Ver. <2.6 ).
> Balancer moves blocks between StorageGroups ( introduced by HDFS-6584) , in
> the new version(Ver. >=2.6) .
> function
> {code}
> class DBlock extends Locations<StorageGroup>
> DBlock.isLocatedOn(StorageGroup loc)
> {code}
> is flawed, may causes 2 replicas ends in same node after running balance.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
