While I agree there's always going to be micro spikes and keeping some headroom is smart, 33% may be too much of a headroom for all but the most latency sensiti

While I agree there's always going to be micro spikes and keeping some headroom is smart, 33% may be too much of a headroom for all but the most latency sensitive RPC services. Personally I aim for only 20% headroom.