Noverlay Networks for Health and Wellbeing
2020-03-28, 17:00–17:30, Room 1

As Kubernetes multi-cluster deployment scales now start to approach those of large IP internetworks of yesteryear, we see a myriad of network design alternatives mirroring the network design choices from the internets of old.

We will take a tongue-firmly-in-cheek yet factual look through some of the network history books at approaches that worked and those that didn't, including some architectural reasons for the flops, through the lens of a network engineer that grew up during the heydey of {i|I}nternet design.

We will draw parallels from history to what can be applied to the current design primitives for Kubernetes connectivity spanning services/pod networking, Ingress and external load balancing, especially when mingled with sophisticated traffic management at edge sidecar proxies.


How many load balancers does it take to replace a pod?

(Hint: include external/cloud LB's, Ingress, internal services or node ports, kube-proxy IPTables|IPVS rules, and often a sidecar proxy).

How do these interplay with pod connectivity approaches? Can there be optimizations?

Is it possible to optimize routing (or addressing) by zone or topology across various abstractions, or does
each successive abstraction get in the way of others?

We'll compare some of the options for these, and contrast with the lessons from network history books.