mindjump
systems

Find the bottleneck

Published October 30, 2026 · 3 min read

The output of any process is set by its slowest step. Improving every other step can feel productive and change nothing at all. Find the bottleneck first, then decide what to improve.

The idea

Picture a small workshop with four stations: cutting, welding, painting and packing. Cutting can handle 120 parts an hour and painting 90, but welding manages only 40. However fast the other stations are, finished parts leave the workshop at 40 an hour. Speeding up painting only makes painted parts wait longer; speeding up cutting only grows the pile in front of the welder.

120/hCutting40/hWelding90/hPainting150/hPackingwaitingoutput ofthe line: 40/hcapacity per station, parts per hour
The slowest station sets the pace for the whole line, and work piles up right in front of it.

Eliyahu Goldratt, a physicist who turned to manufacturing, built a method around this observation, the Theory of Constraints, popularised in his 1984 business novel The Goal. Its central claim is simple: an hour lost at the bottleneck is an hour lost for the whole system, while an hour saved anywhere else is often an illusion.

Case study · Computing

Amdahl’s law

In 1967 the computer architect Gene Amdahl pointed out a hard limit on parallel computing. If part of a task must run one step after another, adding processors speeds up only the rest. The serial part becomes the bottleneck.

The formula is short. If a fraction p of the work can be split across n processors, the best possible speed-up is:

speed-up = 1 / ( (1 − p) + p / n )

With 90% of the work parallel, even unlimited processors cannot make the task more than 10 times faster, because the remaining 10% still takes as long as before. With 95% parallel, the ceiling is 20 times.

0×5×10×15×20×141664256102495% parallel90% parallel50% parallelnumber of processors (log scale)speed-up
Each curve flattens toward its ceiling, 1 / (1 − p). Past a point, more processors buy almost nothing.

This is why software engineers measure where a program spends its time before optimising it. Speeding up a part that takes 5% of the run time can never save more than 5%.

Try it yourselfHow much faster can it get?Drag the sliders and watch the ceiling appear.

How to use it

  1. Map the flow from start to finish, step by step.
  2. Find where work waits. The bottleneck usually sits just after the biggest pile — of parts, tickets, requests or approvals.
  3. Exploit it: make sure the bottleneck never sits idle and never spends time on work someone else could do.
  4. Pace everything else to it. Feeding it more than it can handle only grows the queue.
  5. Only then add capacity at the bottleneck — and look again, because the constraint will move somewhere else.

Where it misleads

The busiest-looking step is not always the constraint; a step can look busy simply because work reaches it in bursts. Measure where work waits, not who looks stretched. The bottleneck is also often not a machine at all, but a rule, an approval step, or one person whose sign-off everything needs.

And bottlenecks move. Fixing one reveals the next, so treating this as a one-time fix misses the point. Finally, a step that is not the bottleneck can still be worth improving for cost, quality or safety — just do not expect it to raise output.

In one line: improve the slowest step first; until you do, everything else is decoration.

Next modelGoodhart’s lawReward a number and people will move the number — not always the thing it stood for.
mindjump · small thinking tools, written down so I actually use them.