We use various heuristic tunings to reach to different initial solutions, then we apply our local search process on each and pick the best result in the end. The problem is that whenever two different tunings yield the same solution after the heuristic, we still run the local search on both instances. Thus we're potentially doing a significant amount of extra work for nothing.

In practice, heuristic + local search happen in a lambda that is called across multiple threads, so we'd have to keep this setup for heuristic only, then filter out identical heuristic solutions after joining all threads, then parallelize the local search using another lambda.