Comment by dwattttt
> People ... aren’t machines that can run constantly on 100% utilization.
You also can't run machines at 100% utilisation & expect quality results. That's when you see tail latencies blow out, hash maps lose their performance, physical machines wear supra-linearly... The list goes on.
The standard rule for CPU-bound RPC server utilization is 80%. Any less and you could use fewer machines; any more and latency starts to take a hit. This is when you're optimizing for latency. Throughput is different.