The filament shuffle, and what CPUs taught us about it

Each of our printers has an automated material feeder holding four spools. Four sounds like plenty until you look at what customers actually order: over the last two months, jobs arrived in 32 different material-and-colour combinations. Twelve slots across the fleet, 32 things to put in them. Something always has to give, and that something is an operator walking over, unloading a spool, loading another, and losing half an hour of momentum.

It felt like this happened constantly. So, in the spirit of the scheduling work we published recently, we measured it instead of grumbling about it.

Thirteen a week

We don't log spool swaps directly, but we do log every job each printer ran, in order, with its material and colour. Replay that sequence against a four-slot feeder and you can count exactly when a swap must have happened. The answer: about thirteen changeovers a week across the fleet. Roughly two a day, every day. The feeling was right.

A feeder is a cache

Here's the reframe that unlocked it. A four-slot feeder holding a few materials out of a catalogue of 32 is exactly what computer scientists call a cache — a small fast store in front of a big slow one. Every time a job needs a material that isn't loaded, that's a cache miss, and the operator is the one who pays for it.

Caches have been studied to death since the 1960s, and there's a famous result: the best possible strategy, when you must evict something, is to evict the thing you won't need for the longest time. The catch is that a CPU can't know the future, so real processors use approximations. But a print farm isn't a CPU. We know the future — it's called the print queue. Which means the theoretically optimal strategy is one we can actually run.

What we built

The scheduler now produces a material plan for every machine. It reads the upcoming queue, checks what's loaded, and works out the swaps that will actually be needed — including which spool to remove each time. When it has to choose, it evicts whichever loaded material won't be needed for the longest, and if nothing in the visible queue settles it, it keeps the materials we sell most of and ejects the rare stuff. Replaying two months of history, that one rule would have removed about a quarter of all changeovers without reordering a single job.

Just as important: the plan comes batched. Instead of discovering one missing spool at a time — walk over, swap, walk back, repeat an hour later — the operator gets the whole list for the next stretch of queue in one go. Same swaps, one trip.

Why this matters beyond spools

None of this required new hardware or a bigger feeder. It required noticing that a problem we experienced as chores — the filament shuffle — was a problem computer science solved sixty years ago, and that our version is easier than the textbook one because our future is written down. Fewer interruptions for us means steadier throughput, which means more predictable turnaround for you.

The full write-up — the simulation method, the numbers, and the eviction rule — is in our research note on material changeovers, which pairs with the earlier note on print-farm scheduling.