Latency issues during the Airbnb k8s migration gave people the sads. #KubeCon 
Some of it was due to hardware changes. The generation of hardware that the app had been running on made a difference. Also the host OS. #KubeCon
They had some issues with noisy neighbors. One specific service was causing problems for others due to the CPU it was using. They hadn’t set CPU limits because they thought it might hurt performance. #KubeCon
The k8s scheduler. #KubeCon
The scheduler can work against you in some pathological cases, like when you have very large images. #KubeCon
Another mention about Autoscaling issues based on CPU load, which came up in the really fun Airbnb talk yesterday too. #KubeCon
Moving to a newer JVM helped their performance. #KubeCon
Traffic for NodePort services goes through IP tables, and gets rewritten to an IP from the overlay network. If you have 2 pods for the same service, it essentially users random load balancing which was causing traffic imbalances. #KubeCon
Consider using Envoy or cloud provider native IPs to avoid the overlay. #KubeCon
