Depending on your budget and feelings on opensource there are two options to push the accelerator. I haven't dug into either one to talk to how they do it.
If you wanted to just roll your own I'd imagine it would be following the star schema methodology the Kimball is known for in Impala and then just leveraging Impala/HDFS caching to make it as fast as possible.
http://kylin.apache.org/or
http://www.atscale.com/
