Member-only story
Reduce Amazon EKS cost by scaling node groups to zero
Amazon EKS just released the support for Kubernetes version 1.24. The new version supports a bunch of cool features. My favorite feature in this release is the ability to scale EKS managed node groups to (and from) zero.
Many customers I engage with have workloads that don’t run continuously. A good example is building software. Software build jobs run when new development teams push new code. Outside of business hours, the supporting infrastructure (like nodes) sits idle. Customers use autoscaling to scale down node groups, but managed node groups required a minimum of 1 node in a node group previously. That’s one node too many, especially when you need beefier and costly nodes with GPUs.
Scaling down to zero results in significant cost savings in such cases. In my opinion, you wouldn’t want to scale your entire cluster to zero. After all, you’d need some nodes to run Cluster Autoscaler and other shared services like Prometheus, AWS Load Balancer, CoreDNS, etc. You can use EKS on Fargate to run some of these services. But keep in mind that Prometheus requires a block storage, and AWS Fargate doesn’t support Amazon EBS yet.
You’d want to run a managed node group for your shared services, like Cluster Autoscaler, that run continuously. You can then add another node group for workloads that spawn periodically, and scale that node group to zero.
Cluster Autoscaler Managed Node group cache
The Kubernetes Cluster Autoscaler project added support for scaling node groups to and from zero in version 0.6.1. However, it only worked if you added specific tags to Auto Scaling groups. In other words, after creating a managed node group, you had to find out the associated Auto Scaling group and add Cluster Autoscaler tags yourself.
Starting Kubernetes version 1.24, you can create node groups (or tag existing node groups) with Cluster Autoscaler tags and Cluster Autoscaler will scale that node group to and from zero.
To enable scaling to and from zero, the awesome EKS team contributed a feature to the…
