How would you diagnose a production incident where API latency spiked significantly, focusing on network-layer causes?