Your link isn't viral. It got 5x slower anyway.

transcript

Your short link gets normal traffic. Nothing special. Today it's five times slower. Because of somebody else's link.

A normal key answers in 20 milliseconds, at p99. One link on this partition goes viral. Its own p99: 500 milliseconds. And the keys that just share its partition? 100. Three to five times worse.

Nothing moved. They share the partition's CPU, its I O, its queues. So they wait behind the viral key. That's contention.

But here's the catch. It gets worse. The overloaded partition fails its health checks. It's marked unhealthy. Traffic moves to its replicas. Which get the same hot key. And overload too. A cascade failure. From a single key.

The fix is layers. The edge, app memory, then Redis, all before the database. The database sees about 1 percent of hot traffic. p99 drops from 500 milliseconds to 20.

One hot key doesn't just hurt itself. It hurts everyone it lives with. A hot key is a partition problem, not a key problem.

Full breakdown in the lecture.