Walk me through how you would implement autoscaling for inference workloads with bursty traffic at Netflix.
