Cell-Based Architecture

The strategic move is converting availability from a probability into a bounded fraction. In a shared architecture, a bad deployment or a poison record affects everyone. With twenty cells, the same fault affects 5% of customers, and the incident becomes a degradation rather than an outage.

A cell is a full vertical slice — compute, data, cache, queues — with no shared state between cells. The only shared component is a thin routing layer that maps a customer to their cell, and that layer must be simple enough to be far more reliable than what it routes to, because it is the one remaining global dependency.

The properties this buys beyond blast radius are what make it worth the cost: deployments become naturally progressive, cell by cell; capacity scales by adding cells rather than by scaling components; and a cell provides a natural boundary for residency and tenancy requirements.

Shuffle sharding is the refinement that makes it stronger than it first appears. Assign each customer to a random subset of cells rather than one, and the probability that two customers share their entire set becomes very small — so a customer whose workload poisons a cell affects only the few who overlap with them, rather than everyone in a fixed partition.

The costs to state honestly: higher fixed cost from duplicated infrastructure, cross-cell operations become awkward or impossible, and operational tooling must work across all cells at once or the operational burden multiplies by cell count.