Fascinating that 2-3 probes per request is a sweet spot, intuitively it seems like a lot of overhead.