My reading was that columnar storage was a critical piece of Dremel. "Accessing the same data as Hive" doesn't sound like Impala is doing that. Is that optional then or is this just a query engine?

I can't say for sure, but I believe Doug Cutting's Trevni format is already in the beta but it's a slow version. Impala is initially focusing on architecture, I'm sure columnar format will have attention eventually.

https://github.com/cutting/trevni

But google's dremel actually doesnt use Map-Reduce at all. So is this using just the HDFS and a new layer along side map-r that is using the HDFS?

Impala does not seem to be using MR at all, see note: https://ccp.cloudera.com/display/IMPALA10BETADOC/Impala+Freq...