2:07 am: a kill switch in 5 seconds, with proof

transcript

2:07 am. A new payment flow is double charging users. On-call flips the kill switch. The CTO wants it everywhere in 5 seconds. And wants proof that it got there.

The obvious answers blow the budget. A 60-second poll. A CDN purge. Or every instance downloading a 5 megabyte ruleset.

The flip goes on a priority lane. It never queues behind bulk CI changes. 70 edge nodes get the notice, with the change inline. No second fetch. Each pushes it down about 50,000 open streams. Every SDK applies it in microseconds, then acks the version it holds.

Add up every hop. About 2 seconds. Inside the CTO's 5, with 3 to spare. Today's 60-second poll? Not even close.

Now the proof. After 3 seconds: 39,612 of 40,000 instances. 388 are on polling fallback. Due within 60 seconds. And here's the list, by host and SDK version.

The interviewer: so the answer to the CTO is not quite yes. Yes for every streaming SDK, with proof. And plainly: polling fallback, reconnecting streams and background apps may miss it.

Budget every hop. Prove it landed. Name who didn't.

Full breakdown in the lecture.