[
https://issues.apache.org/jira/browse/HDFS-8984?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14729201#comment-14729201
]
Daryn Sharp commented on HDFS-8984:
-----------------------------------
Generally paralleling an old patch I was working on, looks to be on the right
track. Regarding earlier comment, I don't think a bm null check is necessary.
The bm is initialized in the ctor and I'm not fond of allowing what should
never happen to go undetected.
I had tried to push all the repl queue logic into the bm. In this patch it's
still fractured across the bm and namesystem, so I'm a bit concerned about
possible races (if not now, need to analyze further, in the near future) with
sync'ed methods in the fsn making multiple calls into the bm, ex. regarding
state. Races may result if those bm methods are called directly elsewhere.
> Move replication queues related methods in FSNamesystem to BlockManager
> -----------------------------------------------------------------------
>
> Key: HDFS-8984
> URL: https://issues.apache.org/jira/browse/HDFS-8984
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Haohui Mai
> Assignee: Haohui Mai
> Attachments: HDFS-8984.000.patch, HDFS-8984.001.patch,
> HDFS-8984.002.patch, HDFS-8984.003.patch
>
>
> Currently {{FSNamesystem}} controls whether replication queue should be
> populated based on whether the NN is in safe mode or whether it is an active
> NN.
> Replication is a concept on the block management layer. It is more natural to
> place the functionality in the {{BlockManager}} class.
> This jira proposes to move the these methods to the {{BlockManager}}.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
