On Wed, Dec 23, 2009 at 2:24 PM, Ken Weiner <[email protected]> wrote:

> HBase itself isn't complaining. I only see the error from Hadoop.  I wasn't
> sure if there are some side effects I'm not aware of.  Is it causing any
> problems to my application?


IIUC, nothing but some extra latency while the dropped socket is
reestablished.



> I see the error occur even when my system is
> under relatively light load with no MR jobs running.  I also forgot to
> mention that I saw this error with HBase 0.20.0/Hadoop 0.20.0 and I
> recently
> upgraded to HBase 0.20.2/Hadoop 0.20.1 and the error still occurs at the
> same rate after the upgrade.
>
> Yeah.  This would not be surprising.  This is hdfs behavior.

You could up the default timeout in hdfs if you wanted less
incidence: dfs.datanode.socket.write.timeout.  But, then sockets would never
timeout.

St.Ack



> Ken
>
> On Wed, Dec 23, 2009 at 11:49 AM, stack <[email protected]> wrote:
>
> > Yeah, see HADOOP-3831.   It looks like datanode timing out unused
> > connections.  As I understand it, later, when dfsclient wants to use this
> > block, it just sets up the socket again -- silently, transparently below
> > the
> > level at which the application can see.  Do I have it right?  Is hbase
> > itself complaining?
> >
> > St.Ack
> >
> >
>

Reply via email to