Around the cluster in 80ms : The journey of a packet

This submission has been added to the schedule

Around the cluster in 80ms : The journey of a packet

Submitted Nov 10, 2019

Section: Crisp talk (20 mins) Category: Network engineering

How does your packet flow in all the scary lanes of kubernetes’ abstracted layers and how to save it from getting delayed.

Basic understanding of how networking is handled in kubernetes
Thorough deep dive into all of the components and understanding how the packet flows through them
Common issues which have been faced in the networking layer and how to debug/solve them
How to customize and tweak the components so that latency can be reduced to minimum

Key takeaways

Kubernetes networking architecture
Figure out how to reduce latency
Figure out how to identify and fix intermittent delays

Intended audience

Any DevOps/SRE who have their infrastructure running in kubernetes, or in general wants to gain an understanding about how kubenertes handles networking out of the box
Any Devs who wants to understand/debug unknown network behavior in their microservices which are running in kubernetes

Outline

Starting with request generating from a Pod to a Service in the cluster, all the services/pods this request flows through

kernel : how does kernel perform lookups and thereby add latencies
dns lookup : how dns lookups is performed in kubernetes and how to debug it
service to pod ip translations : how service loadbalances over pods. When and how kube-proxy gets involved
reaching correct container : how request gets routed to correct node and in turn correct container

Requirements

Basic understanding of kubernetes concepts namely service and pod

Speaker bio

I’ve been working in SRE space since past 3 years. I’ve had the opportunity to setup and manage multiple kubernetes clusters in my organisation, MindTickle, along with developing internal tools for CI/CD, state store management etc.

I’ve faced a lot of networking horrors with K8s running in production for the past 2 years and took a deep dive into understanding how the networking works and how to tweak a lot of cluster components so as to derive optimal performance from it