If you want to save money with your Kubernetes cluster you need to enable consolidation!! #kubernetes #aws #eks #karpenter
Thanks 🤘 Justin Garrison I have been using Kaprpenter for a while and have not seen this property let me read about it.
You had +300 pods in the first version and <200 in the second one 🤔
Any chance Todd is open sourcing his demo tool? It’s pretty powerful!
You still have +100 pods pending 👀
Indeed the karpenter is a perfect tool for those use cases, but I think you missed up, there's almost 100 pods that still pending !! 😅
In my opinion, inefficient bin packing on nodes is not what incurs the most cost. It's applications being over provisioned by requesting too much cpu and memory. Both karpenter and the CA can "consolidate" pods on nodes. Not trying to dunk on karpenter here, but this is not an unsolved problem. Implementing multi dimensional scaling that includes the VPA and the use of spot or reserved instances is what's going to save the most money.
Neither VPA nor HPA can ensure performance and save money. They are both threshold based and they have no insight into node capacity and usage. You need a tool like Turbonomic to give each pod exactly the resources it needs and to move pods to the most optimal node. Feel free to reach out to me if you need more details. https://www.ibm.com/products/turbonomic/kubernetes
Honestly i don't get it, why your cluster is spinning so many nodes with such a low load? If you use cluster autoscaler for EKS it should handle this. Also you collapse 49 nodes to 3 ?! Like if there were a reason to keep many nodes alive in case of HA requirements this might turn into hell pretty fast if you will pack everything in 3 nodes 😂
CIO at Netnod (ex-Spotify)
1yWhat is the cool CLI tool that shows the node utilization?