Numerical variables profiling in very large data set