I think the problem is PG does not know underlying data distribution and use one sampling strategy which works for normally distributed data.