Between Randomness and Arbitrariness: Some Lessons for Reliable Machine Learning at Scale