A DESCRIPTION OF THE PROBLEM :
Sum (relatively accurately) a list of differnet floating points numbers is hard, especially if these numbers are of different signs and different magnitudes. Add a stream of double with different orders, different results are assured. This may cause undesirable non-reproducibility.

Python provides a math.fsum method to support this. Similarly Ruby's sum method use a less accurate (but still much better than the naive method) algorithm to support this use case.

Correctly-rounded summation can be relatively fast and use relatively little memory if done with suitable algorithm.