When the table is huge (i.e. has lots of blocks replica), it took >30 sec just to serialize the query plan, even though the actual planning process didn't take more than a sec. There's room for improvement in the FE-BE thrift structure to improve this timing.
For a medium size table with less than 100 blocks, we've also observed that the query plan serialization time can vary from 1sec to 15+sec. It's possible that it's due to Java GC.