Flowline

FlowlineLearn › Variability and queues

Kingman’s equation: why a balanced line still stalls

Give every station on a line exactly equal capacity and it will still stall, still starve, and still miss the output the arithmetic promised. Nothing is broken. The missing term is variability, and Kingman’s equation is how it gets priced.

CT_q ≈ [(c_a² + c_e²) / 2] × [u / (1 − u)] × t_e

Variability × Utilization × Time

The VUT form multiplies three factors: how variable arrivals and service are, a utilization term u/(1−u), and the process time itself. The middle factor is the one that ruins plans. At 80% utilization it is 4. At 90% it is 9. At 95% it is 19. Pushing a line from 90% to 95% busy roughly doubles its queue time while adding almost nothing to output.

This is why “run everything flat out” and “keep lead times short” are not two goals to balance but two ends of one dial, and why the high-utilization end of it is so much steeper than intuition suggests.

The three buffers

Variability does not disappear; it gets absorbed by one of three buffers, and choosing which is a real decision with real costs.

How variable is this simulation, actually

Measured process-time dispersion sits at roughly c² ≈ 0.80, less variable than exponential — so textbook theory predicts throughput should land above the Practical Worst Case curve rather than on it. Swept across Kanban card counts over five seeds at two hours each, it does, by 17–24%.

Throughput vs. WIP against Hopp & Spearman’s reference curves
WIP (cards)Measuredvs. Best Casevs. Practical Worst Case
15.83/min75%117%
48.15/min88%124%
129.02/min98%122%
409.18/min99.5%119%

A deterministic control run — zero variability — settles exactly on the Best-Case curve, which is what makes the gap above attributable to variability rather than to a modelling error. The first attempt at this measurement failed by about 34% at every level; the cause was the benchmark computing bottleneck rate in gross units while throughput counted only good ones, which is exactly four-station scrap compounding. Finding that instead of loosening the tolerance is the point of certifying against a real law.

What this does not claimKingman assumes an open, non-blocking queueing network. Flowline’s stations have bounded buffers that block upstream production, so real queue time can only come in under the VUT estimate. The suite asserts that direction, not equality — the estimate is an upper bound here, not a prediction.

Try it on a real line

Each of these is a fixed-seed puzzle factory in Flowline’s Challenge Mode. Every one is verified before it ships: the intended fix must win, doing nothing must fail, and a plausible wrong fix must also fail — so the puzzle isolates the idea it claims to teach. They run in your browser.

Open Challenge Mode

Keep reading