What happens and why

Our factory is the same game

Compare the factory with the match game from «A little history»:

a bowl — a machine's buffer: the bar at the bottom of the machine shows how many parts are waiting;
matches — parts on the belts;
the die — the processing time: before each part the machine «rolls the die» and spends from 1/3.5 to 6/3.5 seconds on it, exactly a second on average;
the first player with the box — the source; only ours does not roll a die but releases exactly a part a second;
the warehouse — the sink, and the «made» display counts how many got there.

And, as on the hike, two things act in the factory at once:

dependent events — a machine cannot work on a part until the previous one has handed it over;
statistical fluctuations — every machine has fast parts and slow parts.

Each of them alone is harmless. Together they behave not the way common sense suggests.

Why fluctuations do not average out

It would seem that good and bad luck should cancel out: a six today, a one tomorrow — the same on average. But in a chain they have different fates.

Bad luck is passed on. A machine falls behind — the next one waits in the meantime. Its idle time cannot be won back, and every machine down the chain to the sink gets the delay.
Good luck is not passed on. A machine could have done a part quickly, but there is nothing in front of it — it was not fed. Its speed is wasted. Or it did it quickly, but the next one is busy, the buffer is full — the part stands and waits.

The chance to do more than average depends on the neighbours, the chance to do less does not. That is why a chain as a whole always works slower than the average speed of its links, and work in progress piles up between them.

Eight parts, three machines. Without dice everything runs evenly. With the die the slow parts pass their delay on, the fast ones gain nothing — and the last part is done later.

What our displays showed

Shift	Made by end of release	In progress	Last one done
no dice	about 115	about 13	about 140 s
1…6 die	about 106–111	about 20	about 155–160 s

The same 128 parts, the same machines, just as fast on average. But the second shift ended about 15 seconds later — that is 15 parts the factory could have made and did not — and it held twice as many parts inside.

Work in progress during the shift (our factory, test bench). Without dice it stays at 10–12. With the die it creeps up towards 20 — and after the release stops the factory takes longer to finish.

What follows

A balanced plant is a trap. If you fit the capacity of every section exactly to demand, fluctuations will eat part of the output, and the plant will miss its plan although «on paper» everything adds up.
Measure the output of the whole plant, not the load of every machine. Goldratt proposes three measures: throughput (how much money the sold products bring), inventory (everything invested in the parts inside) and operating expense. A machine that works «for the stockroom» raises its load, but not throughput.
The main thing is the weakest link. Every chain has a link that sets the pace for all — on the hike it is Herbie. An hour lost on it is lost for the whole plant; an hour saved anywhere else is a mirage.

On the hike Alex takes two steps: he puts Herbie at the head of the column — now nobody runs ahead, and the column stops stretching out; and he lightens Herbie's backpack, sharing his things among the others — and the whole troop walks faster. Later Goldratt turned this into the «drum — buffer — rope» method. These steps are the next kits of the series: the bottleneck, «drum — buffer — rope», breakdowns.