the renderer wakes when an episode is queued

the short version

the machine that makes your episode now switches off when nothing is waiting, and it starts again the moment something is queued. the page says waking the renderer for about half a minute, then it says rendering on the server, and the file is the same file it always was.

you do not press anything new. you finish the interview, or you press make today's episode, and you leave the page. if the machine was already on, the page goes straight to rendering. if it was off, you see waking the renderer first. that line is the whole of what changed on the screen, on the web and on the phone.

why it was on all the time

drawing an episode wants real cores. a seven-minute episode took eleven minutes on a slice of a shared processor, and a few minutes on a machine that is actually a machine. so the renderer has lived on its own machine, in london, and the app sends it the recipe and waits.

the awkward part was the waiting between episodes. most hours of a day, nobody is filming. the machine was still on, because the first version of the switch refused to turn the last one off. one machine, always, whether the queue held an episode or held nothing. that is a bill you pay for silence.

leaving it on was the safe reading of a real failure. if the machine is off when you finish talking, and the app decides the renderer has gone, it draws the episode itself, slowly, in the same place that serves the site. founders felt that as a file that took far too long, or as a page that sat on rendering with nothing moving. the rule from then on was: never let a stopped machine look like a missing one.

what you see

queue an episode and the app starts the machine immediately. it does not wait for a check that runs every few minutes. the episode page, the round-up page, and publish, when publish has to wait for a file, all say waking the renderer while that start is under way. the phone says the same words, and the line under them says the renderer was switched off, it is starting, about half a minute, and then this renders. you can leave the page.

a cold start is about ten to thirty seconds. the page keeps the waking line until the machine has the job, then it moves to rendering on the server and the clock you already know. a recap, an audio file, a portrait twin, a week in one cut: same words, same wait, same machine.

the wait has a limit. if the machine was asked to start and nothing has claimed the job after three minutes, the app draws the episode itself so you are not left on waking the renderer with no file coming. that is the slow path, and it is for a start that failed, not for a machine that was merely off. while the machine is on its way, the job stays queued for it.

what is queued

an episode is the usual thing, including the one the cut makes on its own when the takes are in, a podcast's audio, and the video that goes with a podcast. a week in one cut, the month in one film, and the week at a company are queued the same way. an agent's episode, when an agent is due, is queued the same way. an ad clip, once the answer has been heard and a stretch of it chosen, is queued the same way. a shot of your landing page, the one the changes page asks for, is queued the same way.

each of those starts a stopped machine at the moment it is queued. if a machine is already up, the start is a no-op and the job is taken on the next pass, a few seconds later. if several things are waiting, more than one machine can be up, up to the same ceiling as before. nothing about how many can run at once has changed.

when it switches off

when the queue has been empty for a few minutes, the machine finishes whatever it is drawing and then exits. a job in progress is never cut off so the bill can drop. a machine that still has an episode, an agent run, an ad clip or a page shot on it stays up. the check that sizes the fleet, every few minutes, will also stop a worker that is still running after that same empty wait, including the one that used to be kept on as a spare. it will not stop one that is mid-render.

the next thing you queue starts it again. there is no button for this and no setting. the off switch is the empty queue. the on switch is the next episode.

what did not change

the file is still made on that machine, with the same cut, the same captions, the same card. nothing is reordered and nothing is written that you did not say. the render is still a recipe, and making the file again still means the same button. raw takes are still yours to download, on every plan.

the co-host's voice, on a podcast or on camera, is a different small machine and it stays on. a conversation has nothing to queue. switching the renderer off does not touch it.

the size of the render machine is the same. this is not a smaller renderer. it is the same one, off between episodes, so the minutes you wait are the minutes it is actually drawing, plus about half a minute when it has to wake.

if it seems stuck

waking the renderer for much longer than half a minute means the start is slow, not that the episode has been dropped. leave the page. three minutes without a claim and the app draws it the slow way rather than waiting forever. if the page says the render didn't finish, the takes are safe and make today's episode runs it again, which starts the machine again.

a file that is already made is not made again because the machine woke. publishing still uses the file that matches the cut. a cover, the posts and a publish stamp do not send the episode back through the renderer.