Netflix runs large training and inference jobs on GPUs for personalization. These jobs run as gangs and stop if any GPU or network link fails. Basic cloud health checks miss GPU and network fabric faults. Netflix built layered health checks to catch these issues. Checks run every 30 seconds on each GPU machine. Faulty machines are automatically tainted or recycled. Jobs can now survive machine loss via checkpointing and elastic workers. The system uses Kubernetes-native node conditions for health signals. A fallback path reads kernel logs if driver tools fail. Netflix maintains its own fatal-error map for GPU families. Checks fail explicitly if dependencies are missing. The approach spans node health, scheduler, training loop, and observability layers.