It sounds reckless—using randomness to count data. But HyperLogLog is one of the smartest approximations in big data, and it’s accurate within 2%. Coffee linked to significant new side effect, says ...