We spend a lot of memory on column names, both transiently (during reads) and more permanently (in the row cache). Compression mitigates this on disk but not on the heap.
The overhead is significant for typical small column values, e.g., ints.
Even though we intern once we get to the memtable, this affects writes too via very high allocation rates in the young generation, hence more GC activity.
Now that CQL3 provides us some guarantees that column names must be defined before they are inserted, we could create a map of (say) 32-bit int column id, to names, and use that internally right up until we return a resultset to the client.