Never realized how much subtlety there is in calculating, say, 95-th percentile of some value given a running stream of values. Datadog's approach makes a lot of sense and I love that they described it in a paper and shared a few open source implementations like this one.