[ 
https://issues.apache.org/jira/browse/HADOOP-1704?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12525470
 ] 

Doug Cutting commented on HADOOP-1704:
--------------------------------------

The problem is that applications can also easily delete large chunks of the 
filesystem already.  So I don't see this as a Trash-specific issue.  Perhaps we 
can open a new issue generically about namenode memory use for large deletes?  
I suppose the ultimate goal there would be to be able to delete the entire 
filesystem without growing the heap at all?  Would anything less satisfy folks? 
 Should we also file issues about adding or opening lots of files at once?  
Seriously, we could, based on the available heap and our knowledge of the size 
of various data structures, try to limit activity to always stay within the 
heap, returning application exceptions (TooManyFiles) when these may be 
exceeded.  Short of that, I don't see how we can really address stuff like this 
without a major re-write of the namenode, so that it does not use single-host, 
memory-resident datastructures for each file and block.

+1 for closing it.

> Throttling for HDFS Trash purging
> ---------------------------------
>
>                 Key: HADOOP-1704
>                 URL: https://issues.apache.org/jira/browse/HADOOP-1704
>             Project: Hadoop
>          Issue Type: Bug
>          Components: dfs
>            Reporter: dhruba borthakur
>
> When HDFS Trash is enabled, deletion of a file/directory results in it being 
> moved to the "Trash" directory. The "Trash" directory is periodically purged 
> by the Namenode. This means that all files/directories that users deleted in 
> the last Trash period, gets "really" deleted when the Trash purging occurs. 
> This might cause a burst of file/directory deletions.
> The Namenode tracks blocks that belonged to deleted files in a data structure 
> named "RecentInvalidateSets". There is a possibility that Trash purging may 
> cause this data structure to bloat, causing undesireable behaviour of the 
> Namenode.

-- 
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