[ 
https://issues.apache.org/jira/browse/HDFS-2006?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13040350#comment-13040350
 ] 

Matt Foley commented on HDFS-2006:
----------------------------------

In FS I'm familiar with, regular attributes are stored in the inode, while 
extended attributes require an additional disk access, usually considered a 
file read access.  How/where would you propose to store the extended attributes 
for HDFS?  The opening description above seems to imply the data would be 
stored on the NN, but what if it was in the DN metadata file for the file's 
first block, and Clients rather than NN could access it via xml or json?  This 
minimizes the complexity, and puts in the Client, which is presumably the only 
part of HDFS that cares about extended attributes anyway.  And each metadata 
file already allocates a whole 8KB or so, and uses almost none of it, so the 
storage is free.

bq. @Aaron: Since the NN's memory is dedicated to relatively few files and 
relatively many blocks...

That's not necessarily a good assumption.  Our average file size in production 
is something like 1.5 blocks per file (really big blocks).  Surprised me too, 
when I first heard it.  At any rate, need to assume that #files and #blocks are 
same order of magnitude.

> ability to support storing extended attributes per file
> -------------------------------------------------------
>
>                 Key: HDFS-2006
>                 URL: https://issues.apache.org/jira/browse/HDFS-2006
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>          Components: name-node
>            Reporter: dhruba borthakur
>            Assignee: dhruba borthakur
>
> It would be nice if HDFS provides a feature to store extended attributes for 
> files, similar to the one described here: 
> http://en.wikipedia.org/wiki/Extended_file_attributes. 
> The challenge is that it has to be done in such a way that a site not using 
> this feature does not waste precious memory resources in the namenode.

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira

Reply via email to