Kanban Flow Lab: WIP Limits, Little's Law and Cycle Time

A physical Kanban board on a desk: cards flow from backlog through Develop, Review and Test while pawns work on them. Set WIP limits and see Little's law and queueing theory at work. Drag to orbit.

Switch to 2D
Controls

Presets

Parameters

Quality

Camera

Model checks

The science

The system. New work items arrive at random (exponentially distributed gaps, rate λ) into a backlog. A pull system moves each item through three columns, Develop, Review and Test; each column has c workers and every worker needs an exponentially distributed time with mean t to finish an item. A WIP limit K caps how many items may sit in a column (waiting, in service or finished and blocked); an item that is done but has no room in the next column keeps its worker slot blocked, exactly as on a real board. An optional CONWIP cap limits the total number of items on the board.

Little's law. For any stable system N = X · T: mean work in process N equals throughput X times mean cycle time T. It holds for any arrival and service distribution, so you can lower the cycle time only by lowering WIP (for a fixed throughput). The simulator shows the measured N, X and T and their ratio.

Queueing theory. A column with c workers and no limit is an M/M/c queue: utilisation ρ = λt/c, the Erlang C probability of waiting Pw = [ac/c! · 1/(1−ρ)] / [∑k<c ak/k! + ac/c! · 1/(1−ρ)] with a = λt, and the mean wait Wq = Pw/(c/t − λ). By Burke's theorem the departures of a stable M/M/c queue are again Poisson, so for unlimited columns the lead time is exactly the sum over the three columns of (t + Wq); the hockey-stick curve in the first graph is that sum as a function of λ. Close to capacity the wait grows like 1/(1−ρ), which is why a team at 95 % utilisation feels slow. With a CONWIP cap and one worker per stage the board is a closed Jackson network and the throughput is X(N) = G(N−1)/G(N) with Buzen's convolution G.

What the five checks compare. (1) The lead time of 16,000 simulated days against Erlang C plus Burke for three queues in series. (2) The mean wait of a 3-worker column against the Erlang C formula. (3) The throughput of a saturated CONWIP board (N = 3, 6, 12) against Buzen's exact result. (4) Little's law as an exact accounting identity (area under the WIP curve equals the sum of every item's own time on the board) and its long-run form N = X · T on the current settings. (5) The bottleneck law: a board with far more work than capacity delivers the capacity of the slowest column.

Simplifications and honest limits. Exponential times are a mathematical convenience: real task durations are heavier tailed, which makes real queues longer than M/M/c predicts. Workers are interchangeable inside a column, there is no rework, no priority classes, no work in the backlog being re-estimated, no sick days, and each item needs exactly one visit per column. The results of a single run are noisy; the displayed averages converge slowly near capacity, and the six-figure sample sizes in the checks are what pins them down. The theory curve ignores WIP limits and blocking (a limit moves waiting from the board into the backlog, it does not remove it). All numbers are illustrative, not data from a real team.

Loading model...
Loading Blender model...
-
-
Made with 🟠 Blender · Three.js