Abin Shahab created HDFS-8511:
---------------------------------

             Summary: Nodegroup locality does not pick the right data-local 
node for writes
                 Key: HDFS-8511
                 URL: https://issues.apache.org/jira/browse/HDFS-8511
             Project: Hadoop HDFS
          Issue Type: Bug
            Reporter: Abin Shahab
            Priority: Critical


Setting up Nodegroup locality destroys data-local writes if NM and DN have 
separate ip addresses.

When NM requests Namenode for 3 DN locations to write to(for replication=3), 
the Namenode returns a random order every time.

The reason is, Namenode wrongly assumes that it's list of DN's must include the 
NM's ip address. This assumption breaks down when you deploy NM and DN on 
separate machines/ip addresses.





--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to