Queueing basics for capacity: Little's law and why latency climbs before utilisation hits 100%

What it is

Little's law states that, for any system in a stable state, the average number of items inside it equals the average arrival rate times the average time each item spends inside: L = λW. It holds regardless of arrival pattern or service-time distribution. The Netflix concurrency-limits README applies it to services as Limit = Average RPS × Average Latency: 200 requests per second at 50 ms average latency means 10 requests in flight on average (arithmetic).

Utilisation is arrival rate divided by capacity. In the textbook single-server queue with Poisson arrivals and exponentially distributed service times (M/M/1), mean time in the system is S/(1 − ρ), where S is the mean service time and ρ the utilisation: at 50% utilisation a request takes 2S on average, at 80% 5S, at 90% 10S, at 95% 20S (arithmetic). The curve is nearly flat, then nearly vertical.

Why it matters

"Capacity" is not the point where the CPU reads 100%; it is the highest load at which latency still meets its target. The USE method page states that 100% utilisation is usually a sign of a bottleneck and that utilisation beyond about 70% can begin to be a problem: an average over seconds or minutes of 70% can hide bursts of 100%, and for resources such as disks that cannot be interrupted mid-operation, queueing delays can become more frequent and noticeable above 70%. The SRE book describes what follows: with insufficient capacity the server saturates its queues, latency increases and the queue consumes memory.

How to apply

• Use Little's law to convert between the three measurable quantities: in-flight requests = rate × latency. Size thread pools, connection pools and concurrency limits from measured rate and latency, not from a guess.
• Cross-check: a pool of 10 connections at 200 requests per second and 50 ms is full on average, so every burst waits.
• Measure queue time separately from service time (waiting for a worker versus being worked on). Rising queue time at flat service time is the signature of saturation, visible before utilisation graphs look alarming.
• Choose a target utilisation for the bottleneck resource from the latency tolerance; the 70% figure is a rule of thumb, not a law.
• Reduce variance: bursty arrivals and variable service times raise waiting at the same utilisation; batching and admission control buy latency without hardware.

Pitfalls

Little's law concerns averages over a stable period and says nothing about the tail. The M/M/1 formula assumes one queue, Poisson arrivals and exponential service times; real systems have several resources and dependent services, but the curve's shape carries over. Adding servers lowers ρ only if the load is spread.