DFSClient has a retry mechanism on block acquiring for read.  If the number of retries attends to a certain limit (defined by dfs.client.max.block.acquire.failures), DFSClient will throw a BlockMissingException back to the user application.  In the current implementation, DFSClient counts the failures across multiple block acquiring operations but the block acquiring operations are supposed to be independent.  HDFS-127 fixes this problem by counting the failures within a single operation.

I propose to commit HDFS-127 to 0.20 and above since this fix is safe and very useful.

Nicholas Sze

+1
On 10/19/09 2:34 PM, "Tsz Wo (Nicholas), Sze" <[EMAIL PROTECTED]> wrote:

DFSClient has a retry mechanism on block acquiring for read.  If the number of retries attends to a certain limit (defined by dfs.client.max.block.acquire.failures), DFSClient will throw a BlockMissingException back to the user application.  In the current implementation, DFSClient counts the failures across multiple block acquiring operations but the block acquiring operations are supposed to be independent.  HDFS-127 fixes this problem by counting the failures within a single operation.

I propose to commit HDFS-127 to 0.20 and above since this fix is safe and very useful.

Nicholas Sze

+1 This fix has been in limbo for a long time; thanks for finishing
it, Nicholas. -C

On Mon, Oct 19, 2009 at 2:34 PM, Tsz Wo (Nicholas), Sze
<[EMAIL PROTECTED]> wrote:
> DFSClient has a retry mechanism on block acquiring for read.  If the number of retries attends to a certain limit (defined by dfs.client.max.block.acquire.failures), DFSClient will throw a BlockMissingException back to the user application.  In the current implementation, DFSClient counts the failures across multiple block acquiring operations but the block acquiring operations are supposed to be independent.  HDFS-127 fixes this problem by counting the failures within a single operation.
>
> I propose to commit HDFS-127 to 0.20 and above since this fix is safe and very useful.
>
> Nicholas Sze
>
>

+1.

On Mon, Oct 19, 2009 at 3:55 PM, Chris Douglas <[EMAIL PROTECTED]>wrote:

> +1 This fix has been in limbo for a long time; thanks for finishing
> it, Nicholas. -C
>
> On Mon, Oct 19, 2009 at 2:34 PM, Tsz Wo (Nicholas), Sze
> <[EMAIL PROTECTED]> wrote:
> > DFSClient has a retry mechanism on block acquiring for read.  If the
> number of retries attends to a certain limit (defined by
> dfs.client.max.block.acquire.failures), DFSClient will throw a
> BlockMissingException back to the user application.  In the current
> implementation, DFSClient counts the failures across multiple block
> acquiring operations but the block acquiring operations are supposed to be
> independent.  HDFS-127 fixes this problem by counting the failures within a
> single operation.
> >
> > I propose to commit HDFS-127 to 0.20 and above since this fix is safe and
> very useful.
> >
> > Nicholas Sze
> >
> >
>

--
Connect to me at http://www.facebook.com/dhruba

+1

On Mon, Oct 19, 2009 at 2:34 PM, Tsz Wo (Nicholas), Sze <
[EMAIL PROTECTED]> wrote:

> DFSClient has a retry mechanism on block acquiring for read.  If the number
> of retries attends to a certain limit (defined by
> dfs.client.max.block.acquire.failures), DFSClient will throw a
> BlockMissingException back to the user application.  In the current
> implementation, DFSClient counts the failures across multiple block
> acquiring operations but the block acquiring operations are supposed to be
> independent.  HDFS-127 fixes this problem by counting the failures within a
> single operation.
>
> I propose to commit HDFS-127 to 0.20 and above since this fix is safe and
> very useful.
>
> Nicholas Sze
>
>

With five +1 (Suresh, Chris, Dhruba, Stack and myself) and no -1, the vote passed.  I am going to commitHDFS-127 to 0.20 and above.

Nicholas

----- Original Message ----
> From: "Tsz Wo (Nicholas), Sze" <[EMAIL PROTECTED]>
> To: [EMAIL PROTECTED]
> Sent: Mon, October 19, 2009 2:34:01 PM
> Subject: [VOTE] port HDFS-127 (DFSClient block read failures cause open DFSInputStream to become unusable) to hadoop 0.20/0.21
>
> DFSClient has a retry mechanism on block acquiring for read.  If the number of
> retries attends to a certain limit (defined by
> dfs.client.max.block.acquire.failures), DFSClient will throw a
> BlockMissingException back to the user application.  In the current
> implementation, DFSClient counts the failures across multiple block acquiring
> operations but the block acquiring operations are supposed to be independent.  
> HDFS-127 fixes this problem by counting the failures within a single operation.
>
> I propose to commit HDFS-127 to 0.20 and above since this fix is safe and very
> useful.
>
> Nicholas Sze