Spare capacity is part of usable capacity
When arrivals and service times vary, running a service near 100 per cent average utilisation can make waiting times explode. Spare capacity is not merely idle waste; it is the buffer that absorbs randomness. A system optimised for constant work can become unusable for work that arrives unevenly.
In a single-server queue, the server must on average work faster than jobs arrive or no steady state exists. Kingman’s heavy-traffic approximation shows mean waiting grows roughly with utilisation divided by one minus utilisation, multiplied by arrival and service variability. Moving from 80 to 90 per cent utilisation is therefore not a small change near the limit. Appointment systems, emergency departments, call centres and computing clusters add priorities, abandonment and multiple servers, but the qualitative trade-off remains. Predictable workloads can be scheduled more tightly; bursty ones require headroom, pooling or flexible demand.
Efficiency metrics often count visible idleness but ignore waiting, missed deadlines and recovery from disruption. The right capacity target depends on variability and the cost of delay. A fire engine standing still can be doing its job; a hospital bed kept open can be resilience rather than waste.
If every job arrived exactly when the previous one ended and took a known time, a server could be fully occupied without a queue. Random arrivals create clusters, and random service times create overruns. Capacity unused during a quiet interval cannot always be stored for the next burst. Headroom is the ability to catch up before the next fluctuation compounds the backlog.
Several flexible servers can share bursts better than isolated dedicated ones, because simultaneous peaks are less likely across all streams. Cross-training, common queues and load balancing can therefore improve both utilisation and waiting. Pooling has costs: specialised skill, travel, setup and priority rules may prevent full sharing. ‘Add slack’ and ‘pool demand’ are design options, not universal commands. That is the point at which a descriptive pattern becomes a research programme with observations capable of proving it wrong.
Can’t better forecasting remove the need for spare capacity under controlled conditions?
Forecasting can reduce uncertainty and schedule flexible work, but irreducible variation remains in failures, emergencies, job duration and correlated demand. A forecast also has error, which must be absorbed somewhere. Capacity, inventory, time buffers, demand control and service degradation are alternative shock absorbers; eliminating all of them merely transfers the queue.
A variable-arrival service operating indefinitely at full utilisation with bounded waiting and no hidden buffer would contradict the standard queueing premise.