Why do Kubecost and CAST AI disagree about how full our cluster is?

November 18, 2026 5 answers
Asked by Ravdeli-6x

Kubecost reports about 10% idle capacity on our largest EKS cluster, but CAST AI suggests we could remove a lot of nodes. EC2 spending is up roughly a third year over year, traffic is stable, and the cluster has about a dozen more nodes than last spring.

Our Kubecost history is limited to 15 days. How can we work out what changed and whether either report reflects capacity we can actually remove?

Answers

~7n!quorvexa+

Those reports may be answering different questions. Check exactly how each one defines idle and calculates savings. A report based on resource requests can show nodes as full because pods have reserved most of their capacity, even if those pods usually consume much less. A recommendation based on observed usage may see room to reduce requests and repack workloads.

Compare CPU and memory requests with actual usage over a representative period, including peaks, not just averages. Also check whether the proposed savings depend on changing requests or on a different node mix. Low usage alone doesn't mean the scheduler can fit the current pods onto fewer nodes.

Trelvori+7b:

Memory peaks matter here too. A request that looks generous most of the time may still be protecting a workload during a spike.

Reply to this answer
c.742~plovervisor

Start with the AWS bill, not the sales meeting. Use Cost Explorer to look for when EC2 costs changed, filtering by cluster tags or node group where those are available. Compare node counts, instance types and purchasing options over the same period.

Then check whatever history you have in CloudWatch or a longer-running metrics system. Did the extra nodes arrive gradually, after a scaling event, or alongside a change in instance mix? Stable traffic doesn't rule out changes in workload requests, replicas, or infrastructure costs.

Reply to this answer
q.916+parcelwhistle

Check why the autoscaler isn't removing nodes before assuming they're needed. Pod disruption budgets, annotations that prevent eviction or disruption, and pods with local storage can all get in the way. Cluster Autoscaler logs or Karpenter disruption events may explain which nodes are blocked and why.

Ravdeli-6x:

That gives us something concrete to check against the nodes each report thinks are removable.

Reply to this answer
_g71~mivzelo.6

Look at the shape of the fleet as well as the total node count. A shift toward lots of smaller instances can increase the share of capacity consumed by system workloads on every node. The nodes might look busy individually while the overall mix is less efficient.

Compare instance sizes and node group or NodePool settings over time. If the mix changed, check what pending pods and constraints were driving those choices before changing it again.

Reply to this answer
4y!zunvori_682

Treat the savings estimate as a plan to test, not a number you can immediately take off the bill. Tighter bin packing can reduce costs, but it also leaves less room for bursts and may reduce workload spread.

Try request adjustments or consolidation on a limited set of workloads first. Check performance, disruption constraints, and failure tolerance before expanding the change.

Reply to this answer

Add your answer

Answers are reviewed before publication.

Related Discussions