Why Kubernetes autoscalers keep fighting each other (and why the solution must be autonomous)
The New Stack Nikhil Kurup
Kubernetes autoscaling controllers produce contradictory replica and resource sizing changes because they each run independent reconciliation loops that concurrently mutate the same workload and cluster state. Average cluster CPU utilization is around 8% even while node provisioning stays high, reflecting inefficiencies from mismatched assumptions across HPA, VPA, KEDA, GitOps, and node autoscalers. The article argues for an autonomous unified control plane that jointly optimizes replicas and resource requests using multiple signals, keeps Git state aligned with live reality, and adds predictive, auditable decision-making to stop thrashing.
Why it matters
Ask an SRE a deceptively simple question: What is the right size for checkout-api, and how many replicas should it The post Why Kubernetes autoscalers keep fighting each other (and why the solution must be autonomous) appeared first on The New Stack.