hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Vinayakumar B (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-6995) Block should be placed in the client's 'rack-local' node if 'client-local' node is not available
Date Sat, 20 Sep 2014 08:57:34 GMT

    [ https://issues.apache.org/jira/browse/HDFS-6995?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14141883#comment-14141883

Vinayakumar B commented on HDFS-6995:

Hi [~umamaheswararao], thanks for reviewing.
Scenario mentioned is very much possible. 
It was throwing NPE because, target selection was happened even though there was no node with
This is because, {{clusterMap.countNumOfAvailableNodes(
        scope, excludedNodes)}} was returning 1, for the scope of "/default-rack", it should
return 0 as no nodes available with "/default-rack".
I have posted a patch for the same in HADOOP-10131, along with one more scenario where {{NetworkTopology#countNumOfAvailableNodes(..)}}
was returning wrong value.
    Till now this API was not called as writer was considered as null if no DN exists in clients
   Please review HADOOP-10131 patch, once thats in, then a remote rack will be chosen in the
above mentioned case.

> Block should be placed in the client's 'rack-local' node if 'client-local' node is not
> ------------------------------------------------------------------------------------------------
>                 Key: HDFS-6995
>                 URL: https://issues.apache.org/jira/browse/HDFS-6995
>             Project: Hadoop HDFS
>          Issue Type: Bug
>          Components: namenode
>    Affects Versions: 2.5.0
>            Reporter: Vinayakumar B
>            Assignee: Vinayakumar B
>         Attachments: HDFS-6995-001.patch, HDFS-6995-002.patch, HDFS-6995-003.patch, HDFS-6995-004.patch
> HDFS cluster is rack aware.
> Client is in different node than of datanode,
> but Same rack contains one or more datanodes.
> In this case first preference should be given to select 'rack-local' node.
> Currently, since no Node in clusterMap corresponds to client's location, blockplacement
policy choosing a *random* node as local node and proceeding for further placements.

This message was sent by Atlassian JIRA

View raw message