This prevents multiple runs of the same test from trampling on each other, both serially and in parallel.