[ 
https://issues.apache.org/jira/browse/HDFS-6133?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13981055#comment-13981055
 ] 

Daryn Sharp commented on HDFS-6133:
-----------------------------------

[~stack], I quickly skimmed HDFS-2576.  Favored placement appears to dovetail 
very nicely with replica pinning if the goal is to place a 2nd replica on a 
specific node in the event the primary region server fails.  Whether one or 
more replicas is pinned to preserve favored placement, it would appear to meet 
the requirement of this jira, correct?

[~zhaoyunjiong], ignoring the mid-long term goal of separating the BM service, 
the preconditions are simple and a valuable improvement to the NN:  untangle 
the namesystem / directory / BM architecture.  Each layer should have a 
specific role with no back references to other layers.  Namesystem is path 
based, directory is inode based, BM is just a BM.

By allowing no back references, it also opens the possibility of breaking up 
the global locking by  locking subsystems w/o fear of deadlocks.  In and of 
itself, the NN will benefit from a cleaner design and more targeted locking, 
and if done correctly, it's a small leap to run the BM as a separate service 
which in turn opens the door to allowing other services than the NN to utilize 
the block manager.  This is why I'm opposed (-1) to making the BM path aware.  
It ruins the ability to do any of the above.

Do you believe block pinning will address your issue?

> Make Balancer support exclude specified path
> --------------------------------------------
>
>                 Key: HDFS-6133
>                 URL: https://issues.apache.org/jira/browse/HDFS-6133
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>          Components: balancer, namenode
>            Reporter: zhaoyunjiong
>            Assignee: zhaoyunjiong
>         Attachments: HDFS-6133.patch
>
>
> Currently, run Balancer will destroying Regionserver's data locality.
> If getBlocks could exclude blocks belongs to files which have specific path 
> prefix, like "/hbase", then we can run Balancer without destroying 
> Regionserver's data locality.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Reply via email to