hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Andrew Purtell <apurt...@apache.org>
Subject Re: dfs.datanode.handler.count
Date Sun, 04 Jan 2009 09:23:45 GMT
Things are looking really good since the change. Load that
would cause failure after 4 hours can now be sustained for
24+. 

Setting the DFSClient timeout to 0 avoids datanodes being
marked as bad, but, if the handlers fall behind, there are
big negative consequences. For example, there's no sane
reason why the block containing the data file for META
should go missing... but when/if it does, that's it, game
over. I opened HBASE-1108 about that. 

Also, in particular, master state races are exposed, but
HBASE-1104 should address them. At least with DFS able to
respond in a timely manner on my clusters now, the window
for those races has apparently closed. 

  - Andy

> From: stack
>
> Thanks Andrew.
> St.Ack
> 
> Andrew Purtell wrote:
> > I have been able to improve the stability of DFS under
> > HBase by increasing the value of
> > dfs.datanode.handler.count from the default of 3. Seems
> > obvious in retrospect. Just thought I'd pass this along.



      

Mime
View raw message