[ 
https://issues.apache.org/jira/browse/HADOOP-993?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12481712
 ] 

dhruba borthakur commented on HADOOP-993:
-----------------------------------------

I would say that the the motivation behind this is to make the architecture 
cleaner and simple. The filesystem image need not store cluster configuration 
(i.e. datanodes). Regarding you points:

* "in case of a misconfiguration of a datanode, namenode can tell which 
datanode "should" have a certain storageid" -- You might be right here. But 
from my understanding, we have never had exploited this particular feature on 
our kryptonite cluster till now.
* "lastly - UI can report on missing datanodes, which proved immensely useful 
for administration" -- The preferred method to maintain cluster membership info 
is in the include/exclude files. The include/exclude file support was 
introduced recently.

I think it is architecturally sound to keep cluster membership info separate 
from filesystem namespace information. It might help in clustering the namenode 
(thru namespace partitioning) or in namenode failover scenarios in the future.

I agree with you that the cluster administrator will not be able to see dead 
data nodes in the UI anymore. This seems to be the biggest drawback of this 
patch. If the adminstrator wants to view all datanodes in the UI, he/she has to 
create a include file (manually) and list all datanodes in that file. Do you 
agree that this alternative is acceptable?

 

> Namenode does not need to store any data node info persistently.
> ----------------------------------------------------------------
>
>                 Key: HADOOP-993
>                 URL: https://issues.apache.org/jira/browse/HADOOP-993
>             Project: Hadoop
>          Issue Type: Improvement
>          Components: dfs
>            Reporter: Raghu Angadi
>         Assigned To: Sameer Paranjpye
>         Attachments: noDatanodesInFsimage.patch
>
>
> Namenode does not need to serialize datanode info. It will map datanode to 
> storageID when datanode register.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.

Reply via email to