200,000 requests a second. Your cluster does 500,000.

transcript

Clients send 200,000 operations a second. Your cluster does 500,000. Where did the rest come from?

Assume 200,000 operations a second. Half reads, half writes. Every key has three copies. A write goes to all three. 300,000 writes. A quorum read asks two replicas. 200,000 reads. Total: 500,000 operations a second inside the cluster.

Every client operation becomes about two and a half inside. That's the price of three copies. Now size the cluster for it.

Assume one node comfortably does 20,000 replica operations a second. 500,000 divided by 20,000. You need 25 nodes. And the data? About 3 terabytes over 25 nodes. Roughly 120 gigabytes each.

Here's the twist. 120 gigabytes a node isn't what limits you. Throughput is. The replica operations set the node count, not the bytes.

So never size a replicated store from client traffic. Count replica operations. That's where the work is.

Full breakdown in the lecture.