Preparing this page…
Your current navigation and account state stay available while the next page loads.
One latency-sensitive workload and one throughput workload must coexist on eight GPUs. Define who gets in, how batches form, when preemption is allowed, and how both promises survive a lost GPU. This independent, original scenario is designed for Research Engineer, ML / Inference Systems preparation. Work through the stages in order, explain each decision, and use the review dimensions to test whether the result remains defensible when constraints change.
Define what is admitted when the devices are full, and what happens to a long running batch when a latency-sensitive request arrives.
Does admission protect the promise that matters at that moment?Explain whether the throughput class may be preempted, what it costs when that happens, and how you would keep one class from starving the other over an hour.
Can the candidate quantify the cost of the policy it chose?One device leaves the pool while both classes have work in flight. Describe memory placement, what happens to that work, and how the lost capacity is accounted for in each promise.
Does the recovery path keep the promise accounting honest?Use these checks after a first draft. They are review prompts, not six more questions to answer.
Clarifies goals, constraints, and ambiguous requirements.
Protects invariants and handles boundary conditions.
Identifies realistic failures and designs recovery paths.
Makes ownership and transitions explicit.
Provides enough visibility to diagnose and run the system.
Explains what is gained, lost, and intentionally deferred.
Use a different system, project, or situation and check whether the structure still produces a specific, defensible answer.