Input data on the large problem is about 500M. I cannot reproduce on a small data set, I only see it on the larger problem.