AnthropicVerified interview challenge
Distribute a large model across a server fleet
Design how a 500 GB model reaches 1,000 servers over a constrained connection.
Unlock with Full AccessThe challenge
A 500 GB model is stored in an external repository, and all 1,000 servers need a complete copy. The connection to that repository is constrained, so the number of servers alone does not tell you how quickly the rollout can finish. Design the data path and explain which transfer limits determine the completion time.
The design also needs to account for servers receiving incomplete data or losing a transfer partway through. Work through the capacity reasoning alongside recovery and integrity, then defend the assumptions behind your architecture.
What you'll practise
Transfer capacity
Relate the amount of data that moves across each connection to the time required for the fleet to finish.
Distribution architecture
Explain how servers obtain the model and which shared resources can limit progress.
Recovery and integrity
Account for interrupted transfers and establish when a server has a usable copy.