Top 7 Kubernetes Rightsizing Tools 2026

Most Kubernetes rightsizing tools do one of three jobs: show changes, apply changes, or tie changes to cost. If your cluster is over-requested, lower requests can help with bin-packing, but you only save money when that leads to fewer nodes or less unused capacity.

I’d sum this article up like this:

• Goldilocks, Robusta KRR, and VPA fit teams that want recommendations first
• StormForge, CAST AI, ScaleOps, and PerfectScale fit teams that want policy-led automation
• Kubecost fits teams that need cost visibility, ownership, and rightsizing in the same place
• Rollout risk matters more than savings claims
• CPU and memory changes should be checked against HPA, PodDisruptionBudgets, and node scaling
• A lower request does not always mean a lower cloud bill

The article also makes one point very clear: rightsizing is not a stand-alone fix. It works with:

• VPA for pod sizing
• HPA and KEDA for replica counts
• Cluster Autoscaler or Karpenter for node count

A few numbers frame the problem fast:

• A 2026 CAST AI report says the average cluster is over-provisioned by 69% for CPU and 79% for memory
• The 2025 State of FinOps report puts workload optimisation and waste reduction at the top of the agenda
• CAST AI’s look-back window can be set from 3 hours to 7 days, with 24 hours as the default

If I were choosing from this list, I’d use a simple filter:

1. Do I want recommendations only, or auto-apply changes?
2. Do I need cost ownership and chargeback, or just pod sizing?
3. Can my workloads handle restarts, evictions, or in-place resize limits?
4. Will lower requests lead to fewer nodes and lower spend in £?

::: @figure {Kubernetes Rightsizing Tools Compared: Control Model, Automation & Cost Scope (2026)} :::

Kubernetes Rightsizing: Beyond CPU and Memory, with Andy Suderman | KubeFM

Quick comparison

Bottom line: I’d pick the tool by control model, not feature count. If you want a safe first step, start with read-only recommendations. If you want automation, make sure you can track throttling, OOM kills, latency, error rates, node use, and spend in £ across at least one full billing cycle.

The seven Kubernetes rightsizing tools

Goldilocks, Robusta KRR, and Kubernetes Vertical Pod Autoscaler

Goldilocks is a recommendation-only tool. It creates or exposes Vertical Pod Autoscaler objects for workloads and shows CPU and memory recommendations in a dashboard. By default, VPA update mode is set to Off, so nothing changes on its own.[5] Engineers review the output and then apply changes through manifests, Helm values, or GitOps.

That setup works well for teams that want a clear audit trail and human sign-off before anything goes into production. In practice, the quality of those recommendations depends on the VPA setup, the metrics you have, how much history is available, and how the workload behaves over time.

Robusta KRR runs as a CLI tool against your cluster, pulls historical use data from Prometheus, and produces a report with suggested request values across namespaces or even a whole fleet. It is read-only.

That makes KRR handy for scheduled fleet audits, CI/CD checks, or sorting workloads for review before a quarterly review. If you use filters, leave out system namespaces and critical workloads, compare the output with PodDisruptionBudgets, and check recommendations against peak usage rather than average usage. It also makes sense to confirm the release-specific metric window before using it in production.

Kubernetes VPA is the native starting point that many other tools build on or plug into. The recommender calculates target, lower, and upper bounds. The updater applies changes. The admission controller sets values on new pods.

If you run VPA in Off mode, you get recommendations without automatic changes. Switch to an automatic mode and the updater can evict pods to apply new values, which is where PodDisruptionBudgets, replica counts, readiness probes, and disruption tolerance start to matter. You should not run VPA and HPA on the same resource dimension without compatibility testing.[11]

The next group goes past pod settings and pulls cost data into the picture.

Kubecost and StormForge by CloudBolt

Kubecost is mainly a cost-visibility and allocation platform, not just a pod-tuning tool. Its rightsizing features sit next to cost allocation, budget tracking, showback, chargeback, and idle-resource analysis. So it can answer a question that recommendation-only tools usually cannot: which team owns the cost? Kubecost 3.1 added resource-quota rightsizing, with recommendations that can be applied straight from the UI.[10]

StormForge by CloudBolt uses historical workload data and machine-learning models to recommend request and limit values.[10] It starts in read-only mode.

One detail stands out here: its handling of HPA is quite exact. It can preserve the scale point for workloads that are actively scaling, treat idle or saturated HPAs differently, and optionally stop changes to resources that HPA already controls, all without changing the HPA configuration itself.[9] That matters because if you change CPU requests on a workload that HPA is scaling on CPU utilisation, you also change the scaling trigger point.

Namespace-level opt-in and opt-out controls, approval workflows, and policy-based rollout controls make gradual adoption far easier across a large cluster. Even so, it is worth checking edition-specific automation, supported autoscalers, and rollback behaviour against the release you plan to use.

These tools go further into policy-led automation and live workload control.

CAST AI Workload Autoscaler, ScaleOps, and PerfectScale

CAST AI Workload Autoscaler continuously generates recommendations from historical and forecast data, with policy controls for vertical rightsizing, horizontal scaling, optimisation targets, overhead, look-back periods, minimums, maximums, and apply thresholds.[2][3] The look-back period can be set from 3 hours to 7 days, with 24 hours as the default, so the right choice depends on whether the workload follows daily or weekly traffic patterns.[4][7]

A February 2026 update added two useful controls: recommendations now respect Kubernetes ResourceQuota hard limits, and there is an option to preserve the existing request-to-limit ratio when producing new values.[6] Immediate application can cause pod restarts. Deferred application avoids forced restarts. And on Kubernetes 1.33 and later, in-place resizing may let some changes happen without a restart at all.[7]

ScaleOps works on live workloads and keeps evaluation separate from application, while still giving teams policy-based rollout controls over workload resources and autoscaling behaviour. A common path is to start in read-only mode, measure possible savings and risk without touching production, and then turn on policy-based rollout controls for selected workloads or namespaces.

For variable workloads, the main settings are minimum and maximum resource bounds, change sensitivity, and cooldown periods that stop thrashing. Any capability tied to a specific ScaleOps plan, agent version, or supported autoscaler version should be checked before production use.

PerfectScale takes a similar efficiency-first route, but as a managed platform.

PerfectScale is a managed Kubernetes efficiency platform. It combines rightsizing recommendations with efficiency analysis and idle-resource visibility, which makes it a reasonable fit for teams that want a managed view across workload use and waste without building a full internal VPA or reporting workflow. It identifies idle workloads and efficiency waste, but removal or node reduction still depends on the product edition and automation settings.

Comparison table: how each tool fits

The big decision here is simple: do you want recommendations, automation, or cost control? The table below trims that down to the bits that matter most: control level, autoscaling fit, and cost scope.

Recommendation-only, automated, and cost-management categories

Goldilocks, KRR, and VPA in recommendation mode fit teams that want proof before making changes in production. That usually means more manual work for operators, but far less rollout risk.

StormForge, CAST AI, ScaleOps, and PerfectScale are a better match for teams that are ready to let a platform make changes within guardrails. The split between them comes down to scope. CAST AI reaches furthest into node and capacity choices. StormForge stands out when close coordination with HPA matters. ScaleOps and PerfectScale sit nearer to workload efficiency, without going as far into the full infrastructure layer.

Kubecost sits in a different lane. Its rightsizing features matter, but its main job is financial governance. If engineering and finance both need the same picture of Kubernetes spend, Kubecost is often the better fit.

In practice, the deciding factors are less about one headline feature and more about day-to-day operations: how mature your team is, how much pod disruption you can live with, how good your observability data is, and whether you're happy to give a platform write access.

That’s the lens to use before moving on to rollout safety: how much control are you willing to hand over?

Choosing and adopting a tool safely

Once you know which type of tool you need, the next call is rollout risk.

Matching tool to operating model

Pick the tool that fits the way your team makes changes, not the one with the longest feature list.

Recommendation-only tools work well for review-heavy teams where every change is checked before it reaches production. Automated tools are a different story. They need solid rollback steps, clear observability, and clear workload ownership before you let them touch live services.

The next job is to prove the choice safely in production.

Safe rollout and measuring results

Before you change anything, get a baseline.

Collect one representative week of CPU and memory data. That week should cover weekdays, weekends, batch windows, and peak traffic. Include p95 and p99 CPU figures, memory working-set and peak values, throttling counts, OOM-kill events, pod evictions, replica counts, and Kubernetes spend in £ by cluster, namespace, and workload. If the window is too short, a recommendation can look lean on paper while missing a monthly batch run or a seasonal traffic spike.

Start in read-only mode before applying any changes. For VPA, that means updateMode: Off. Review the proposed values first. Keep certain workloads out of scope at the start: stateful databases, message brokers, JVM-based services with tricky heap behaviour, latency-sensitive workloads, and anything without steady readiness, liveness, latency, or error-rate telemetry.

When you begin applying changes, set minimum and maximum CPU and memory bounds so recommendations stay inside tested limits. Put Pod Disruption Budgets in place, check that readiness probes are working, and start with a small group of stateless services that are easy to observe. Then expand by namespace or service owner only after that canary group meets your agreed service-level and cost targets.

One more thing: lower requests do not automatically cut your bill. You only see savings when lower requests lead to fewer nodes, better bin-packing, or less unused committed capacity. So track the numbers that matter: throttling, OOM kills, latency, error rates, node utilisation, and spend in £ across one full billing cycle.

Conclusion: match the tool to your control model

Rightsizing is one part of Kubernetes cost optimisation, not the whole answer. Savings also depend on node packing, idle workloads, scheduling, storage, and autoscaling.

So the final decision comes down to control, not feature count.

If your team wants recommendations that can be reviewed before anything changes, Goldilocks, Robusta KRR, and native VPA are a good fit. If you need rightsizing linked to allocation and chargeback, Kubecost makes more sense. And if your goal is policy-led automation, StormForge, CAST AI, ScaleOps, and PerfectScale sit at different points on the automation spectrum.

Once you’ve picked a tool, don’t get distracted by big savings claims. Safety matters more. Go with the option that gives clear, explainable recommendations and keeps changes inside policy limits your team can manage with confidence.

FAQs