The Hours the Machine Works Alone

There is a part of this craft that is not craft at all. It is queueing. Nobody writes about it because there is nothing to say about it, which is roughly why we wanted to.

A time map of an episode

Six weeks, laid out by what is actually happening in them, produces three categories rather than the two you would expect.

Deciding. Perhaps a week of it, spread thinly across all six. Low-effort, high-consequence, and almost impossible to schedule.

Working. Around a hundred and fifty hours of hands on the thing. Drawing, layering, recording, mixing, assembling.

Waiting. Between forty and ninety hours of machine time on the final pass alone, plus every test render before it, plus the encodes, plus the upload. The forest episode described in the episode we deleted took ninety hours on its final pass and was then deleted, which is the purest available example of what this category is.

So a rough tenth of an episode's elapsed life is the machine working without us; in a bad week, a quarter. It is the one part of the process we cannot compress by caring more, working later, or being better at our jobs.

You cannot want a render to go faster. It is the only part of the work that is completely indifferent to you.

The wrong things to do while it runs

We have done all of these, most of them repeatedly.

1. 1

   Watch it

   The purest waste. A progress bar rewards attention with nothing, and it is specifically designed to look as though it might. Some part of the brain treats it as a task in progress and refuses to release the evening. We have each lost whole nights this way, awake in a room where the only event was a number going up.

2. 2

   Make small edits on a copy

   Feels productive. It is how we ended up, twice, with two project files that had both moved on and no clear answer to which one was the real one. Reconciling those cost more time than the render saved. Now the project is locked when the queue starts, and locked means locked.

3. 3

   Start the next thing at one in the morning

   The most seductive one, because it looks like discipline. Our informal count is that roughly three out of four decisions made after midnight during a render were reversed by ten the next morning — a colour, a room, a whole subject. The late hour does not produce bad ideas exactly. It produces confident ones, which is worse, because a confident bad idea survives to the following week.

4. 4

   Poke the machine

   Opening a preview. Checking a frame. Each of these steals from the job you are checking on, so the render gets slower in direct proportion to how much you want it finished. This is almost funny and never funny at two in the morning.

The one right thing is embarrassingly simple and took years to actually do: leave. Set the queue, close the door, go and be somewhere else. It is right for the same reason it is difficult — there is no version of staying that helps, and the mind does not accept that easily about a thing it started.

Trusting the queue, and the morning it fails

The whole arrangement depends on a bet: that the thing you set running at midnight will be correct at nine.

The bet loses often enough to be a real cost. Our worst instance was a long file that ran for forty-one hours with the rain layer switched off — the toggle had been left flipped during a check ten minutes before the queue went in, and nobody looked at the layer panel again because the file had been finished for two days. We found out on a Sunday morning. Forty-one hours of machine time, a day and a half of calendar, and one of the four missed publishing slots described in working slowly on a fast platform traces directly back to it.

The fix was not a better checklist, although we did write one. It was a ninety-second test render — twenty frames, spaced across the whole file — that has to complete and be looked at by a human before the long job is allowed to start. It costs a minute and a half. It has caught four errors in two years, at least one of which would have cost another weekend.

What the gap is actually good for

Here is the one genuine benefit, and we want to be careful not to oversell it, because the temptation to make waiting sound wise is exactly what this piece is trying to avoid.

A full playback at the end of a working day tells you very little. You have been inside the file for weeks; your eye has agreed to everything in it. A playback after forty hours of not being able to open it is a different instrument entirely. Things that were invisible on Thursday are obvious on Saturday — a colour that has crept, a rhythm that is too regular, an object that tells too much.

That distance is real and it is worth something. It is also entirely an accident of our tooling. If renders were instant we would have to invent a reason to leave the file alone for two days, and knowing us, we would not.

The bill for doing nothing

One thing that never appears in accounts of quiet work: the machine is not free while it waits.

A ninety-hour render on our setup draws roughly what a small heater does, continuously, for four days — somewhere around thirty-five or forty kilowatt-hours per long episode, about what our fridge uses in a month, spent before anybody has watched a second of it. Multiply by the test passes, the abandoned versions and the deleted forest episode, and a meaningful share of what this channel costs to run is electricity turning into heat in a room nobody is sitting in.

We bought a second machine partly to shorten this and it helped less than we expected. The long final pass does not split as neatly as the sales page implies; the queue went from about ninety hours to about fifty-five, not the half we had budgeted for. Now two machines heat the same room, there are two sets of updates to keep aligned, and one more thing that can be misconfigured at midnight. We would probably still buy it. It was not the purchase we thought it was.

We have not solved this

There is no ending to this one. Most of the waiting is simply boring, and boredom is not a hidden virtue — it is what it says it is. The nights spent watching a bar were not contemplative; they were wasted, and we would like them back.

What we have is a locked project file, a ninety-second test render, and a rule about going to bed that we keep about two-thirds of the time. That is the whole of our progress in three years on the part of the job that occupies more calendar than any other single stage.

The argument for why watching something in which almost nothing happens is different from watching a progress bar — a distinction we now take personally — is in the four-minute cat. The difference is whether the waiting is going anywhere.

The evening is warm here.