OpenAIVerified interview challenge
Sora 2: Schedule video generation on a fixed GPU fleet
Design how a video service shares limited GPU capacity between competing workloads.
Unlock with Full AccessThe challenge
A Sora-like service receives more video requests during bursts than its GPU fleet can immediately process. Design the admission and scheduling policies for free, paid and internal workloads. Explain how the system responds as demand and available capacity change.
- Focus on scheduling
- Job creation, status updates and result storage already exist. Treat model inference as a black box.
What you'll practise
Capacity and admission
Relate incoming demand to the work the fleet can sustain.
Dispatch and fairness
Allocate workers across workload classes while accounting for contention and waiting time.
Changing fleet conditions
Handle workers joining, draining or failing, and assess how the policy performs.
Your canvas
Video workloadsDemand & priorities
GPU fleetCapacity & utilisation
The interview, drawing canvas and final report are included with Full Access.