Kubernetes Autoscaling with Karpenter: DevOps and Docker Live Show (Ep 173)

Starts at 3:40. Nirmal Mehta, a Principal Specialist Solution Architect at AWS, and a Docker Captain, joins me to discuss and demo Karpenter. Karpenter is an autoscaling solution that aims to simplify Kubernetes infrastructure by automating node scaling up and down... "the right nodes at the right time."
Support this show on Patreon! It's the #1 way to support this show, my podcast, and open source / bretfisher
Join the discussion on our Discord chat server devops.fan
Topic Links
==========
Starship Shell Prompt starship.rs/
Bret's favorite shell setup www.bretfisher...
Karpenter karpenter.sh/
Karpenter release blog aws.amazon.com...
K8s Scheduling Concepts kubernetes.io/...
Other types of autoscalers:
Horizontal Pod Autoscaler kubernetes.io/...
Vertical Pod Autoscaler github.com/kub...
Cluster Autoscaler github.com/kub...
Nirmal Mehta
==========
/ normalfaults
/ nirmalkmehta
Bret Fisher
=========
Best coupons for my Docker and Kubernetes courses at www.bretfisher...
Listen to my podcast of this show's past episodes at www.bretfisher...
#devops #kubernetes #aws

[@ArunKumar-xw6iw]

For example, the application has 10 replica pods. Karpenter gets the suitable single node based on the requests and limits of the app. In this case, if the node is having any issue, all 10 pods will be on pending state till the next new node is available. It is a single point of failure right?

[@nirmalmehta]

there are different strategies to use to ensure those 10 pods are running and at least some are available. check out the karpenter best practices guide here: aws.github.io/aws-eks-best-practices/karpenter/

[@ArunKumar-xw6iw]

Thanks for the discussion. Whenever the new node is provisioned there will be some daemon sets (log exporters, metric exporters etc)are deployed right? In this case, even though core pods are removed from the node, still the daemon set pods will run on that node right? So how can Karpenter terminate that node as the node will never be empty.

[@nirmalmehta]

the karpenter deprovisioning system considers a node empty when it has no active non-daemon set pods runnings. more info here: karpenter.sh/v0.18.1/tasks/deprovisioning/#how-karpenter-nodes-are-deprovisioned