If your model's latency spikes in production at Netflix, what is your systematic approach to debugging the bottleneck?
