Why our biggest printer stopped taking small jobs

We run three industrial FDM printers. Two are standard-format machines; one is the big one — a larger build plate and a hotter extruder, which makes it the only machine that can print oversized parts and high-temperature engineering materials. Every paid order that comes through the site gets assigned to a printer by software, automatically.

For months, that software made a decision that sounds perfectly sensible: send each new job to whichever printer looks least busy. It's what most scheduling systems do. And it was quietly wrong.

What the data showed

Every routing decision our system makes is logged — which printer, why, and what the queues looked like at that moment. When we analysed two months of those logs, the picture was stark. The big machine was busy 85% of the time, and jobs assigned to it were waiting nearly a full day on average before starting. Meanwhile the two standard machines sat idle for most of the working week.

Worse: nearly a third of what the big machine printed was ordinary small work — parts either standard machine could have handled. The one printer that oversized and high-temperature jobs must use was spending a third of its life on jobs that had two other homes.

The motorway problem

There's a well-known result in queueing theory that explains why this hurts so much. Delays don't grow steadily as a machine gets busier — they explode. A machine that's 50% busy barely queues at all. At 80%, waits multiply several times over. It's the same reason a motorway at 85% capacity is a traffic jam, not a fast road. Our measurements matched the textbook curve almost exactly, which meant we could predict what would happen if we changed the rules.

The fix: reserve the big machine

The scheduler now treats the big machine as what it is: a scarce resource. Small jobs go to the standard machines by default, even when the big machine happens to look free at that moment. The big machine only takes standard work when the small machines are genuinely jammed — we set that threshold at two hours of queueing, derived from the measured cost of occupying it.

There's a matching rule on the other side: if any machine is sitting idle while a job waits more than two hours somewhere else, the scheduler moves the job. The same two-hour figure governs both directions, so the system can never argue with itself about whether a move is worth it.

What it means if you order from us

Large-format parts and high-temperature engineering materials should start much sooner — our model says the average wait for those jobs drops by more than half. Ordinary parts won't be slower; the standard machines had plenty of spare capacity. And because every decision is still logged with its reasoning, we'll be checking the real numbers against the predictions over the coming weeks.

If you'd like the full detail — the queueing model, the measurements, and the design of the scheduling rules — we've written it up properly in a research note.