How do you keep Kubernetes rightsizing from becoming a constant manual job?

November 19, 2026 5 answers
Asked by 5j~gimbalnotch_742

Our workloads and traffic change over time, but requests, replica counts, and node choices can get stale. How do you keep them aligned without overspending or trimming away useful headroom?

I'm interested in ways to combine rightsizing with tools like Karpenter and KEDA, and how much of the process you automate.

Answers

Nuvquaro-5m

Start with recommendations, not automatic changes. Look at CPU and memory usage over several weeks, compare it with requests, and review the outliers with the app owners. A monthly usage percentile plus a buffer can be a useful starting point, but check peaks and stress-test the service before cutting too close.

Keep some headroom deliberately, especially for workloads where a slow scale-up or memory pressure would hurt. Vertical Pod Autoscaler can provide recommendations, though those may need a bigger buffer than you're comfortable running with.

v.492!gasketheron:

I'd also make sure the observation window includes any recurring busy periods. A quiet month can make a request look safer to cut than it really is.

Reply to this answer
Tavrelo+4p

Watch the interaction between request changes and the HPA. If it scales on CPU utilization, the request is the denominator. A pod using 100m CPU is at 20% of a 500m request, but 50% of a 200m request. Cutting the request can therefore trigger more replicas even when traffic hasn't changed.

That matters when evaluating a rightsizing operator that mutates pods at admission. The Deployment manifest may still show the old request while the admitted pods get a different one. Check the actual pods, replica behavior, and node count before calling the change a saving.

5j~gimbalnotch_742:

That's a useful distinction. I'd want the actual admitted values visible alongside the recommendations.

Reply to this answer
Vella

Split the problem up. Karpenter can help provision nodes for pending workloads, while cost monitoring and regular metrics show where requests and capacity look wasteful. Neither removes the need to check whether the workloads are asking for the right resources.

Policy can enforce resource limits, and separate node pools can help isolate teams or workloads with different requirements. Send owners periodic recommendations rather than waiting for a full cluster cleanup. Estimates should include the effect on replicas and nodes, not just the reduction in requested CPU or memory.

Reply to this answer
9u!marblefunnel_284

Don't optimize every pool for maximum churn. Some workloads are sensitive to node replacement, so a protected pool with a slower rotation schedule can be a reasonable tradeoff. Pod disruption budgets and Karpenter's do-not-disrupt annotation are also worth considering when deciding how aggressively to consolidate or replace nodes.

_p92!ferruvia.3:

That extra capacity may cost more, but it makes the tradeoff explicit rather than treating every spare node as waste.

Reply to this answer
_a38!drimquelo.6

If you're looking at automated operators, demo them before handing over production changes. Check whether recommendations and mutations are visible to app teams and deployment tooling, and ask whether the control plane can run in your cluster or is managed externally. I'd start with a limited rollout and compare actual scaling and cost against the estimates.

Reply to this answer

Add your answer

Answers are reviewed before publication.

Related Discussions