Free tool

How much of your cluster is reserved and idle?

Your nodes are sized by what pods request, not what they use. Paste kubectl get pods -A -o json and your usage, from kubectl top pods -A or a Prometheus p95 CSV. See requested vs. used CPU and memory per workload, suggested requests, the node capacity they free, and a YAML patch for each workload.

  • No price list: add your node size and price if you want the dollars.
  • Every result shows its formula and inputs, so you can check the math.
  • Everything is calculated in your browser. Nothing you paste is sent anywhere.

Your cluster

Running and pending pods are counted; finished Jobs are left out. Pods are grouped into their Deployment, StatefulSet, DaemonSet or CronJob.

CSV columns: namespace, pod (or workload), container, CPU in cores and memory in bytes (any header with "cpu" and "mem" in it). The PromQL is below.

Added to used CPU. A p95 already leaves the top 5% to bursting.

Added to peak memory. Running out of memory means OOMKills, so keep some.

Your nodes (optional, for nodes freed and dollars)

Allocatable CPU and memory of one typical node, and what it costs you per hour (on-demand, or your discounted rate).

Before you change requests

Size from a week of peaks, not a snapshot

This is gotcha 1 in our free guide, 12 cloud bill gotchas, and how to get the money back: you pay for what you launch and request, not what you use.

  1. With at least a week of Prometheus retention, export p95 CPU and peak memory per container, and paste them as a CSV with columns namespace,pod,container,cpu,memory. Pods replaced during the week are matched to their workload by name.
    max by (namespace, pod, container) (
      quantile_over_time(0.95, rate(container_cpu_usage_seconds_total{container!="", image!=""}[5m])[7d:5m])
    )
    max by (namespace, pod, container) (
      max_over_time(container_memory_working_set_bytes{container!="", image!=""}[7d])
    )
  2. Or run the Vertical Pod Autoscaler with updateMode: "Off" (quoted) for at least a week, including your weekly peak, and read its target with kubectl get vpa -A.
  3. Lower requests one workload at a time, CPU before memory. If requests equal limits, lowering only the requests moves the pod from Guaranteed to Burstable QoS. A HorizontalPodAutoscaler measures utilization against requests, so adjust its target at the same time.
  4. After each change, watch for OOMKills, restarts and CPU throttling:
    kubectl get pods -A -o custom-columns='NS:.metadata.namespace,POD:.metadata.name,LAST:.status.containerStatuses[*].lastState.terminated.reason' | grep OOMKilled
  5. Smaller requests only shrink the bill when nodes leave. Check that Cluster Autoscaler scale-down or Karpenter consolidation is on, and that PodDisruptionBudgets or safe-to-evict: "false" annotations are not pinning half-empty nodes.

Privacy

Your cluster stays on this page

  • The right-sizer runs in your browser. It makes no request of its own, and sends nothing you paste or choose. A file you choose is read on your device, not uploaded.
  • Its Content-Security-Policy only allows connections back to this site, and it loads no third-party scripts, fonts or trackers.
  • A shared link keeps your settings after the #. Browsers never send that part of a link to a server. What you paste is never put in the link.
  • The only request with something you typed is the optional email signup. It sends your email address and the tool's name, nothing else.

Want us to look at the real thing?

Book a free 30-minute consult: we'll look at your bill with you, no obligation and no sales pitch. Or check your own cloud bill first with our free Cloud Bill Gotchas Checker.

Book a free 30-minute consult