Autonomous Driving Simulation Platform
Design a distributed platform that runs millions of reproducible driving simulations and produces comparable scorecards and replays.
Problem statement
For this exercise, assume engineers submit immutable simulator, policy, map, and scenario-set versions.
The platform fans out 30-second autonomous-driving scenarios, schedules heterogeneous CPU and GPU workers, captures deterministic seeds and artifacts, aggregates safety and comfort metrics, and supports replay and comparison across releases.
Exclude building the vehicle simulator physics engine and training the driving model.
What to build
- Create a reproducible simulation campaign from immutable software, map, scenario, and seed inputs.
- Schedule, retry, and cancel millions of isolated scenario runs with fair resource allocation.
- Build trustworthy scorecards, regression slices, and replay links across campaign versions.
At scale
- Run 20 million 30-second scenarios per day across CPU-only and GPU-enabled pools.
- Start 95 percent of admitted campaigns within 5 minutes while preserving team quotas.
- Retain compact metrics for one year and full replay artifacts for 30 days unless pinned.
- A retried scenario must remain reproducible and must not count twice in campaign aggregates.