I am merely a dabbler in this area and definitely not an expert, but my understanding is that columnar stores tend to be substantially more efficient for analytical operations over large sets of in memory data by virtue of the data being easier to operate on with vectorized instructions like SIMD.