Separation of storage and compute

What did separating storage from compute buy, and what does every read now cost?

Separating storage from compute enables everything else in this article. It was unlocked by a hardware ratio and not by a business idea. Network bandwidth reached about 1 GB/s per host, which is the same order as a fully used SATA SSD (a common type of solid-state disk). From that moment, remote bytes were no longer categorically slower than local ones. Later, NVMe SSDs (faster solid-state disks attached directly to the computer's bus) outran the network again, which is exactly why local caching came back and did not disappear. The ratio decided both outcomes.

Cost consequence

Two meters became independently movable, so the storage-to-compute ratio became a free variable. This matters because in analytics the two grow at very different rates. Retained volume compounds (it only grows). Query volume follows headcount. With coupled purchasing (disks attached to the compute nodes), you buy CPU you will never use in order to hold bytes. Decoupling removes that waste.

The price is that every read now crosses a network. The per-request latency floor rises from about 100 µs (local SSD) to about 10 ms (remote object storage). A cache layer returns to recover the speed, and it brings an invalidation problem with it. This is cost conservation in plain view: the coupling cost was not removed. It was converted into cache complexity.

One dependency is worth noting: scale-to-zero is impossible without this separation. You cannot switch off a node that holds the only copy of the data. Everything below rests on it.