Platfora is doing some interesting work with interactive, in-memory BI for Hadoop. They essentially do away with the traditional DW/ETL model and create ephemeral in-memory 'lenses' for querying and visualization.

Impala is married to Hadoop. What if your data infrastructure isn't built on hbase and too complex/large to integrate it easily? Would impala still serve that purpose?

Impala doesn't require HBase to operate, it can use raw HDFS. Simple example, if you had a few terabytes of TSV files, you could easily copy the raw data into HDFS and then create a simple schema around it. All queries on this data would be in parallel across all the nodes in the cluster, this is partly due to the distributed nature of HDFS.

If your data is too difficult to integrate into HDFS (doesn't have to be HBase) using existing Hadoop tools, I suspect you're going to have to do some work to use that data on any platform.