Step 1 · The pattern
A shared fleet spreads the load
Customers A, B and C send normal requests, a quarter of a second each. The router spreads them over three groups of servers, so the load is even and every request finds a free server. (All numbers are illustrative.)
- request
- served
In short
Cell-based architecture: how it works and how it breaks
A shared fleet serves every customer, and any request can land on any server. That spreads load evenly until one customer starts sending something expensive: a huge export, a pathological query, a broken integration retrying in a loop. Each of those requests holds a server for seconds, and because they land everywhere, they slowly occupy every server. Every customer's requests queue and fail, though only one customer did anything unusual. A cell-based architecture splits the service into cells: complete, isolated copies, each serving a fixed set of customers, with a thin router in front that sends each customer to its cell. The noisy customer can still overwhelm its own cell, but the damage stops there. The AWS guidance describes cells as fault-isolated boundaries that limit the impact of a failure; shuffle sharding refines the same idea.
What this simulation simplifies
- Each cell is drawn as one group of servers. A real cell contains every layer of the service, including its own data stores.
- Customers are assigned to cells by a fixed rule. Real systems keep the mapping in a small, highly available table and can move customers between cells.
- In the trap, requests are spread at random. Real load balancers often pick the least busy server, which spreads expensive requests even faster.
How this comes up in interviews
"You run a multi-tenant SaaS API. One tenant's traffic keeps degrading the service for everyone. How would you redesign it?"
Then they ask
- What exactly goes into a cell, and what stays shared?
- How do you assign tenants to cells, and what happens with a tenant too big for one?
- How do cells change the way you deploy?
A strong answer covers
- In a shared fleet any request can land anywhere, so one tenant's expensive traffic spreads over every server.
- Cells are isolated copies of the whole service, each serving a fixed set of tenants behind a thin router.
- A noisy tenant overwhelms only its cell; the blast radius is the tenants in that cell.
- Keep shared parts minimal, cap cell size from load tests, and give very large tenants their own cell.
- Deploy cell by cell, and consider shuffle sharding to shrink the impact further.
Points to cover, not a script. Say them in your own words and tie them to a system you have worked on.
On one page: the track cheat sheet.
Sources: AWS whitepaper: Reducing the scope of impact with cell-based architecture ↗ · Amazon Builders' Library: Workload isolation using shuffle-sharding ↗ · Report a problem
Diagrams are our own simulations: the timings, rates and counts in them are illustrative, not measurements. Captions are original wording.