Last year they flushed the cache at noon. 9 minutes down.

transcript

The interviewer asks: what does your cache do at exactly noon? Last year's answer: flush it. Nine minutes of errors.

At noon, price reads hit 1 point 5 million a second. The two replicas can take 20,000 misses a second. Refill 200,000 sale prices, one read each? 10 seconds of all of it.

So even done perfectly, a flush can't work. The candidate's move: what do we already know before noon?

Sale prices are locked at 6 pm the day before. So cache a timeline, not a price. 17,999 until noon. 9,999 until 2 pm. Then back to 17,999.

11:59:59. The pod reads version 42. Its clock says before noon. One second later. Same entry, same version. The clock says after. 9,999. Nothing in any cache changed.

Last year, noon was an invalidation. 200,000 keys cold at once. This year, zero database reads at noon. Clocks agree within 50 milliseconds. Every pod flips together.

If you know the change in advance, ship it in advance. Put the schedule in the cache, and noon is a non-event.

Full breakdown in the lecture.