Randomized Backtracking for Heavy-Tailed Distributions
Some backtracking algorithms have a few (fruitless) branches that are very large, both deep and broad.
Algorithms which choose randomly at a split point will have a small probability of getting caught in one of these branches.
Therefore, some runs finish very quickly, most runs take some time, and a few runs take orders of magnitude more time than the median.
Solution: cut off long runs and reseed the randomizer.