Cassandra doesn't store columns from different rows together to my knowledge. That means it is no way like an analytic column store where columns from all rows are stored together. This is not a bad thing, it just wasn't intended to be an analytic column store and makes design choices suited for other things.
Analytic column stores can exploit storing columns separately by using various encoding schemes to compress data that are far superior to generalized compression and they can use the sort to zero on relevant ranges for each column.
Ah ok. I understand why a columnar database would be better for analytical workloads, but I didn't realize that HBase and Cassandra aren't actually laid out in a columnar fashion on disk. I was confused by the 'column store' terminology. This makes Citus data's release even more compelling.
See http://wiki.apache.org/cassandra/ArchitectureSSTable
Analytic column stores can exploit storing columns separately by using various encoding schemes to compress data that are far superior to generalized compression and they can use the sort to zero on relevant ranges for each column.