How efficient does this thing need to be?
You can always just turn it into a two-dimensional sampling problem by thinking of the data as a function f(x=item), generating a random x=xr in [0,x], then generating a random y in [0,max(f(x))]. The xr is accepted if 0 < y <= max(f(xr)), or rejected (and another attempt made) if y > max(f(xr)). -- https://mail.python.org/mailman/listinfo/python-list
