[
https://issues.apache.org/jira/browse/HDFS-5946?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13900671#comment-13900671
]
Jing Zhao commented on HDFS-5946:
---------------------------------
But for reading through webhdfs you also want to get the data locality. In that
case the client wants the local DN or the closest DN for reading. Currently the
sorting logic takes this part into account and the first DN in the list is the
best one. That's also why in the original/current implementation only web UI
uses random pick.
> Webhdfs DN choosing code is flawed
> ----------------------------------
>
> Key: HDFS-5946
> URL: https://issues.apache.org/jira/browse/HDFS-5946
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: namenode, webhdfs
> Affects Versions: 3.0.0, 2.4.0
> Reporter: Daryn Sharp
> Priority: Critical
>
> HDFS-5891 improved the performance of redirecting webhdfs clients to a DN.
> Instead of attempting a connection with a 1-minute timeout, the NN skips
> decommissioned nodes.
> The logic appears flawed. It finds the index of the first decommissioned
> node, if any, then:
> * Throws an exception if index = 0, even if other nodes later in the list are
> not decommissioned.
> * Else picks a random node prior to the index. Let's say there are 10
> replicas, 2nd location is decommissioned. All clients will be redirected to
> the first location even though there are 8 other valid locations.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
