3 ms·

fwiw, for non-html workloads we serve (nodejs) with shared-cpu (several 100 requests per second), the latencies (p50, p75) remain consistent.