Twenty Little Boxes

Four million users share one stack. A single customer's traffic pattern exhausts a shared work queue and everyone goes down. The team moves to a cell-based architecture of 20 independent cells, each with its own queue, database, and compute, with customers routed to a cell by id. What has actually changed?

The same bug now takes down only the roughly 5% of customers in one cell instead of everyone

Availability rises across the board, because the 20 cells act as replicas of one another for each request

Blast radius is unchanged, because the routing layer in front of the cells still handles every request

Total capacity needs drop, because each cell provisions only for the share of traffic it owns