A giant problem is that people want to see reported what their 95% latency is. This is a complicated number that does not aggregate easily. Because as the article says, you need the whole dataset to calculate it.
By contrast we can ask what percentage of records is over 400 ms. This is a simple number. And if we simply keep track of how many requests there were and how many are slow, this is a simple thing to aggregate and a simple thing to calculate.
A giant problem is that people want to see reported what their 95% latency is. This is a complicated number that does not aggregate easily. Because as the article says, you need the whole dataset to calculate it.
By contrast we can ask what percentage of records is over 400 ms. This is a simple number. And if we simply keep track of how many requests there were and how many are slow, this is a simple thing to aggregate and a simple thing to calculate.