Kubernetes Cluster Autoscaler vs Karpenter
Update: Qovery officially supports Karpenter for all Managed AWS EKS clusters
Cluster Autoscaler is a tool that automatically adjusts the size of the Kubernetes cluster when one of the following conditions is true:
- some pods failed to run in the cluster due to insufficient resources.
- there are nodes in the cluster that have been underutilized for an extended period of time, and their pods can be placed on other existing nodes.
Depending on the cloud provider, it’s built-in or not (eg. GCP provides it by default, and AWS not).
The Kubernetes autoscaling SIG has done excellent work since then, but having also worked with Mesos, I was waiting for an initiative to optimize pod placement and reduce the billing cost to the maximum. This is where Karpenter entered the game, launched by AWS.
Welcome Karpenter
Karpenter automatically launches just the right compute resources to handle your cluster's applications. It is designed to let you take full advantage of the cloud with fast and simple compute provisioning for Kubernetes clusters.
Karpenter vs Cluster Autoscaler
Karpenter can look like a simple alternative to Cluster autoscaler, but it’s more than that. Both are really good, but before choosing one, you have to know their strengths and weaknesses. Karpenter is more useful when you start to have a significant workload, while cluster autoscaler is not the best in this case.
I’ve tried to summarize both solutions:
Conclusion
Both Cluster Autoscaler and Karpenter are interesting solutions. Cluster autoscaler is more stable, and more used, while Karpenter is the outsider. If you want to use Karpenter, you have to be sure the workload is adapted to it, otherwise, Karpenter won’t be able to do its job properly and you won’t get its benefits.
At Qovery we’re considering Karpenter as an alternative for a cost-effective solution. Nowadays, we strongly think Cluster Autoscaler may be more adapted to production usage (because of the common usage where application cache is useful), and Karpenter for pre-prod/staging/testing usage. We also hope other cloud providers will invest in Karpenter and see it becoming the new standard.
.svg)
.svg)
.svg)
Suggested articles
The cluster coming up is the easy part. What catches teams off guard is what happens six months later: certificates expire without a single alert, node pools run at 40% over-provisioned because nobody revisited the initial resource requests, and a manual kubectl patch applied during a 2am incident is now permanent state. Agentic control planes enforce declared state continuously. Monitoring tools just report the problem.
The instinct when setting up Kubernetes observability is to instrument everything and send it all to your APM vendor. That works fine at ten nodes. At a hundred, the bill becomes a board-level conversation. The less obvious problem is the fix most teams reach for: aggressive sampling. That is how intermittent failures affecting 1% of requests disappear from your monitoring entirely.
Scaling your deployments to zero is only half the battle. If your cluster autoscaler does not aggressively bin-pack and terminate the underlying worker nodes, you are still paying for idle metal. True environment sleeping requires tight integration between your ingress layer and your node provisioner to actually realize FinOps savings.
The structure, table, tool list, and code blocks are all worth keeping. The main work is fixing AI-isms in the prose, updating the case study to real metrics, correcting the FAQ format, and replacing the CTAs with the proper HTML blocks. The tool descriptions need the "Core strengths / Potential weaknesses" headers made less template-y, and the intro needs a sharper human voice.
For years, Red Hat OpenShift has been the safe choice for heavily regulated, on-premise environments. It operates as a secure fortress. But in the public cloud, that fortress acts as an expensive prison. Paying proprietary per-core licensing fees on top of your standard AWS or GCP compute bill is a redundant "middleware tax." Escaping OpenShift requires decoupling your infrastructure from your developer experience by running standard, vanilla Kubernetes paired with an agentic control plane.
Agentic Kubernetes resource reclamation is the practice of using an autonomous control plane to continuously identify, suspend, and delete idle infrastructure across a multi-cloud Kubernetes fleet. It replaces manual cleanup and reactive autoscaling with intent-based policies that act on business state, eliminating the configuration drift and cloud waste typical of unmanaged fleets.
Kubernetes focuses on container orchestration, but the reality on the ground is far less forgiving. Provisioning a single cluster is a trivial Day-1 exercise. The true operational nightmare begins on Day 2. Teams that treat multi-cloud fleets like isolated pets inevitably face crushing YAML configuration drift, runaway AWS bills, and severe scaling bottlenecks.
.webp)