The ultimate answer is that you need to test your configuration with your expected workflow.
However, the thing that mitigates the remote IO factor (hopefully) is that the Solr HDFS stuff features a blockcache that should (when tuned correctly) cache in RAM the blocks your Solr process needs the most. Solr on HDFS currently doesn't have any sort of rack locality like there is with say HBase colocated on the HDFS nodes. So you can expect that even with Solr installed on the same nodes as your datanodes for HDFS, that there will be remote IO. Michael Della Bitta Senior Software Engineer o: +1 646 532 3062 appinions inc. “The Science of Influence Marketing” 18 East 41st Street New York, NY 10017 t: @appinions <https://twitter.com/Appinions> | g+: plus.google.com/appinions <https://plus.google.com/u/0/b/112002776285509593336/112002776285509593336/posts> w: appinions.com <http://www.appinions.com/> On Tue, Mar 24, 2015 at 2:47 PM, Joseph Obernberger <j...@lovehorsepower.com > wrote: > Hi All - does it make sense to run a solr shard on a node within an Hadoop > cluster that is not a data node? In that case all the data that node > processes would need to come over the network, but you get the benefit of > more CPU for things like faceting. > Thank you! > > -Joe >
