[ 
https://issues.apache.org/jira/browse/HDFS-3513?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13291123#comment-13291123
 ] 

Alejandro Abdelnur commented on HDFS-3513:
------------------------------------------

It is a bit more complicated than that, for some reason (if I recall correctly 
this was for some JT requirements) the FileSystem cache is per UGI but UGI uses 
object equality instead or username equality. so it you have 2 UGIs for the 
same user you end up with 2 FileSystem instances the FileSystem cache. (run 
into this problem with Oozie when hadoop security came to be and I was told 
that is the way it has to be). But going back to this patch, you won't have 
leaks, the entries in the fsCache in HttpFS remain in the cache by their 
FileSystem instances are closed after the time out. the worse case scenario I'm 
referring to, it is about the number of CachedFileSystem entires in the fsCache 
in HttpFS, this map that serves as cached is not being purged of entries, but 
the FileSystem instances in the entries are certainly closed after time out, 
thus no sockets  TIME_WAIT (specially because with this cache HttpFS is quite 
aggressive on closing the FileSystem instances).

Hope this clarifies.
                
> HttpFS should cache filesystems
> -------------------------------
>
>                 Key: HDFS-3513
>                 URL: https://issues.apache.org/jira/browse/HDFS-3513
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>    Affects Versions: 2.0.0-alpha
>            Reporter: Alejandro Abdelnur
>            Assignee: Alejandro Abdelnur
>         Attachments: HDFS-3513.patch, HDFS-3513.patch
>
>
> HttpFS opens and closes a FileSystem instance against the backend filesystem 
> (typically HDFS) on every request. The FileSystem caching is not used as it 
> does not have expiration/timeout and filesystem instances in there live 
> forever, for long running services like HttpFS this is not a good thing as it 
> would keep connections open to the NN.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: 
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Reply via email to