[
https://issues.apache.org/jira/browse/HDFS-5828?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Buddy updated HDFS-5828:
------------------------
Attachment: HDFS-5828.patch
The problem was BlockPlacementPolicyDefault.chooseTarget was manually adding
the nodes in the results list to the excluded nodes list instead of using the
addToExcludedNodes method.
The addToExcludedNodes method is overridden by
BlockPlacementPolicyWithNodeGroup to also exclude other nodes in the same node
group.
> BlockPlacementPolicyWithNodeGroup can place multiple replicas on the same
> node group when dfs.namenode.avoid.write.stale.datanode is true
> -----------------------------------------------------------------------------------------------------------------------------------------
>
> Key: HDFS-5828
> URL: https://issues.apache.org/jira/browse/HDFS-5828
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: namenode
> Affects Versions: 2.4.0
> Reporter: Buddy
> Attachments: HDFS-5828.patch
>
>
> When placing replicas using the replica placement policy
> BlockPlacementPolicyWithNodeGroup, the number of targets returned should be
> less than or equal to the number of node groups and no node group should get
> two replicas of the same block. The Junit test
> TestReplicationPolicyWithNodeGroup.testChooseMoreTargetsThanNodeGroups
> verifies this.
> However, if the conf property "dfs.namenode.avoid.write.stale.datanode" is
> set to true, then block placement policy will return more targets than node
> groups when the number of replicas requested exceeds the number of node
> groups.
> This can be seen by putting:
> CONF.setBoolean(DFS_NAMENODE_AVOID_STALE_DATANODE_FOR_WRITE_KEY, true);
> in the setup method for TestReplicationPolicyWithNodeGroup. This will cause
> testChooseMoreTargetsThanNodeGroups to fail.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
