Single-queue service

Why does one long queue serve people better than several short ones?

At the airport there is one snaking queue feeding a dozen desks. At the supermarket there are a dozen queues, one per till. Both arrangements are chosen deliberately by people who have thought about it, and travellers routinely say the single queue looks worse — it is visibly enormous. Yet it is the arrangement that spread. So what does the single queue give you that the dozen short ones cannot, and why does it feel like the opposite?

Hold the important things fixed first, or the comparison means nothing. Same number of servers, same rate of arrivals, same distribution of how long each transaction takes. All we are changing is the shape of the line. That is a strange thing to expect any effect from — nobody is working faster.

So look for what the shape can affect. Picture the parallel arrangement at a random moment. Is it possible for one till to have nobody at it while another has three people waiting? Obviously yes; it happens constantly. Now hold that thought, because it is the whole argument. In that moment a server is idle while a customer is waiting. Capacity is being produced and thrown away. The single queue makes that state impossible: as long as anyone is waiting, no server can be free, because the next person goes to whichever desk opens. Nothing got faster — less got wasted.

That gives us the first result, and it is the one people usually miss: pooling lowers the mean wait, not merely its spread. This is standard queueing theory. A single queue feeding many servers is the M/M/c system, whose waiting behaviour is given by the Erlang C formula; the alternative is c separate M/M/1 systems each getting a share of the arrivals. Compare the two at the same total load and the pooled system waits less. The gap is largest when the system is busy but not saturated, which is exactly when anyone cares.

Now the second effect, which is what people notice even when they cannot name it. In the parallel arrangement you pick a lane, and your fate is then tied to that one server and to whoever is in front of you. If the person at the head has a price check, a chequebook, or forty items and a fistful of coupons, you are stuck while the next lane sails past. Your wait has become one draw from a very wide distribution. In the single queue it depends on the average rate at which all twelve desks clear, and averages are far steadier than individual draws: one slow transaction delays everyone slightly instead of delaying six people enormously. So the variance drops sharply — which is also why the single queue feels fair, since what it eliminates is precisely being overtaken by people who arrived after you.

Keep the two effects apart. Lower mean comes from eliminating idle servers; lower variance comes from each customer being exposed to the whole pool's randomness rather than one server's. In principle you could have one without the other.

If the single queue wins on both, why is the supermarket still full of parallel lanes? Servers must be interchangeable for pooling to work, and a till with a loaded belt and a customer already unpacking cannot be redirected. A snaking line needs room, and the walk from its head to the desk adds seconds to every transaction, eroding the gain. The lane queue is lined with shelves of confectionery for a reason. And a single queue looks long to someone deciding from the door which shop to enter.

One more finding cuts against the whole argument. People often prefer parallel queues even when told the single queue is faster: choosing a lane feels like agency — reading the trolleys, judging the cashier, backing your own assessment — in a way a guaranteed average does not. Whether a service should honour that preference or override it is a design judgement, not a mathematical one.

The misconception to dislodge is that the single queue works by some psychological trick — that it merely feels better because you can see progress. It does feel better for that reason, but the improvement is real and mechanical, and it would still be there if nobody could see the queue at all.

A refinement that ties the pieces together: everything here depends on servers being substitutable. Pooling gains come entirely from the ability to send the next customer to whichever desk frees up, so any constraint that breaks substitutability — specialised counters, a loaded belt, a customer already seated — removes the benefit rather than reducing it. This is why the same logic reappears wherever resources can be shared, from hospital beds to call centres that cover for each other.

• Why a queueing system's wait rises steeply rather than smoothly as utilisation approaches its capacity.
• How occupying people during a wait — longer walks to baggage reclaim, mirrors by lifts — changes the experience without changing the delay.

Every term the collection defines is gathered in the glossary.