Thanks Ray, that's what I've been doing also.  I just now created a JIRA issue 
to
see if we can do it automatically in the future (maybe based on the number of 
regions, like we do for getSplits).

https://issues.apache.org/jira/browse/HIVE-1406

JVS

On Jun 14, 2010, at 9:53 AM, Ray Duong wrote:

Try setting the number of mappers base on your cluster size.

set mapred.map.tasks=XX;

Also, make sure to configure hive to hit multiple zookeepers.

-ray

On Mon, Jun 14, 2010 at 9:00 AM, Martin Fiala 
<[email protected]<mailto:[email protected]>> wrote:
Hello,

I am a newbie to Hive, but I'm already quite familiar with Hadoop/HBase.
I must appreciate the whole project and especially the new integration
with HBase, which is what we really need. :)

So back to the problem, I got Hive running with HBase, it works really
nice, gets data from HBase, computes something and returns results. But
even when I run it on a large table with hundreds of regions, it splits
data only into 2 maps, which means only 2 task trackers running on a 10
node cluster. When I run similar task written in Java+MapReduce and fire
it up, it splits input into hundreds of maps and the computation is
nicely distributed.

Is it some misconfiguration or why the Hive's InputSplit gives me only 2
maps?

Regards,
Martin Fiala


Reply via email to