Expand it across racks
Now expand the previous paper design across racks. More servers add capacity, but shared dependencies can still take them down together.
Can the expanded service lose a rack
24 replicas × 20 = 480 requests/s
16 replicas × 20 = 320 requests/s
300 acceptable requests/s
The fictional requirement
Use twelve servers with two replicas each. Keep the assumed 20 acceptable requests per second per replica.
Three racks holding four servers each provide 480 requests per second normally. Losing one rack leaves 320, just above a 300-request target.
Check what all racks share
If routing, storage, or authentication depends on one unprotected component, rack separation does not remove that failure.
Also plan the restart. Eight returning replicas may load their models together while the surviving racks still serve customers.
For a separate illustrative power budget, four 2 kW servers plus 1 kW of support total 9 kW per rack. Facilities must validate real equipment and limits.
Your deliverable
Draw rack placement, request and storage paths, shared dependencies, and a staged recovery plan.
Check your answer
If peak demand rises to 340 requests per second, the assumed one-rack-down capacity no longer meets it. More verified capacity or a revised design is needed. “Redundant racks” does not change the arithmetic.