hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "edward yoon" <edw...@udanax.org>
Subject Re: deleting files from the DFS
Date Fri, 14 Mar 2008 01:22:19 GMT
Can zookeeper locking services be used for these cases?

Thanks,
Edward.

On 3/13/08, Sanjay Radia <sradia@yahoo-inc.com> wrote:
> Ahmad Humayun wrote:
> > So does that mean nodes can possibly read files that have been "deleted"
> >
> If the name node entry has been deleted, new opens will fail.
> Already opened files may have their reads succeed if they have cached
> the block locations and also if the
> blocks have not been deleted.
>
> Hence, if you open a large file and someone deletes the file after you
> open it, at some point the reads may fail if
> the client needs to go to the namenode to get block locations.
> This not posix semantics but keeping the namenode stateless has
> advantages in scaling.
> > thanks :)
> >
> >
> > regards,
> >
> > On Wed, Mar 12, 2008 at 10:57 PM, Hairong Kuang <hairong@yahoo-inc.com>
> > wrote:
> >
> >
> >>> Case 1: Lets say there is a node which opens a file on the DFS to read.
> >>>
> >> From
> >>
> >>> my understanding, the open request will go to the namenode, which will
> >>>
> >> tell
> >>
> >>> it where the blocks of the file are. What if another node comes up and
> >>>
> >> asks
> >>
> >>> to delete that file. Of course this request will go to the namenode. How
> >>> will the namenode respond in such a situation?
> >>>
> >>>
> >> The namenode is stateless. So the delete request will go through. The
> >> first
> >> request may fail.
> >>
> >>
> >>> Case 2: Whenever a node wants to delete a file, I think the request
> >>>
> >> first
> >>
> >>> lands with the namenode? What happens to the replicas of the blocks of
> >>>
> >> that
> >>
> >>> file, as soon as that file gets deleted?
> >>>
> >>>
> >> The replicas are scheduled to be deleted by the namenode. But there may be
> >> some delay before they are actually deleted on the datanodes.
> >>
> >> Hairong
> >>
> >>
> >>
> >
> >
> >
>
>


-- 
B. Regards,
Edward yoon @ NHN, corp.

Mime
View raw message