We're going to count up faucet hits and based on this, divide up the distribution at the end of the interval. I think we're going to start with a 1 day interval and shorten it if we can.
This method has the added advantage of always keeping us on target and scaling with load.
It will be interesting.