2022-01-09 17:52:55
Yesterday I shared this video in our chat and it looks like people liked it. So, I would like to share it here with the broader audience.
In this video Victor Farcic speaks about AWS Karpenter and its advantages comparing to good old cluster-autoscaler.
A few notable things:
- Karpenter is workload-aware. It means that it can determine, how many resources does your workload needs and scale up a cluster accordingly. So, if you need to place just a tiny pod, you’ll get a smaller node comparing to a situation if you need to run a few heavy tasks
- Karpenter is topology-aware. So, for example, you can schedule nodes for a given workload in a specific AZ only. It’s neat if you use EBS volumes or additional network interfaces
- It’s groupless, meaning that it doesn’t have a concept of “instance groups” like cluster-autoscaler (and many other autoscalers). So, cluster-autoscaler modifies parameters of instance group, Karpenter on another hand talks to AWS APIs directly. In theory, this should reduce scale-up and scale-down times
#kubernetes #aws
1.2K views14:52