Skip to content

Kerf and other rightsizing tools: details and sources

The table in the README is a summary of this page. Every fact about another tool comes from that tool's public docs, repository or pricing page, read on 2026-11-05. "Not stated" means we did not find it in their docs. It does not mean the tool lacks it. Products change: check the source before you rely on a cell, and open an issue or a pull request when something here is wrong or out of date.

Summary

Kerf VPA Goldilocks Robusta KRR IBM Kubecost StormForge PerfectScale CAST AI
License Apache-2.0 Apache-2.0 Apache-2.0 MIT (CLI) commercial; free tier up to 250 cores commercial; 30-day trial commercial; free tier (300 vCPU) commercial
Where it runs your clusters only your cluster your cluster CLI on your machine; optional SaaS UI self-hosted, or managed backend agent in-cluster, metrics sent to SaaS SaaS, or self-hosted in-cluster agent, SaaS console
Recommends requests and limits requests (limits keep their ratio) requests and limits (via VPA) requests and limits requests requests and limits requests and limits requests
Data source Prometheus, VictoriaMetrics, Thanos, Mimir, AMP Metrics API, history in checkpoints or Prometheus VPA Prometheus, Thanos, VictoriaMetrics, Mimir, managed Prometheus Prometheus (2.x); own agent (3.0) own forwarder own exporter metrics-server and own exporter
Applies changes agent, per policy: in-place, rollout, deferred eviction and admission webhook opt-in, through VPA ("not recommended") optional enforcer webhook, needs a Robusta UI account Kubescaler (alpha) applier, or patches for CI/CD eviction and webhook; GitOps notes immediate or deferred
In-place pod resize yes (default strategy) InPlaceOrRecreate (GA in VPA 1.6.0) through VPA not stated not stated yes (applier 2.12.0) yes (default since agent 1.0.26) yes
Next to an HPA on CPU/memory yes: requests and HPA target change together docs warn against it on the same metric not stated --allow-hpa flag not stated yes (vendor docs) HPA recommendations; docs advise against two vertical tools yes
Fleet UI yes no UI per cluster; multi-cluster in Fairwinds Insights through the Robusta UI Enterprise yes yes yes
Savings in dollars yes, per node type not stated not stated not stated yes yes yes yes
Rollback on bad health yes: OOM kills, restarts, CPU throttling not stated not stated not stated not stated yes not stated not stated
Maintenance windows yes, plus freeze windows not stated not stated not stated schedule settings yes yes not stated
JVM heap aware yes not stated not stated not stated not stated yes yes yes

Notes

  • VPA. In-place pod resize is GA in Kubernetes 1.35. VPA's InPlaceOrRecreate mode was alpha in 1.4.0, beta in 1.5.0 and GA in 1.6.0; an InPlace mode without eviction is alpha in 1.7.0. Auto is deprecated and behaves like Recreate.
  • Goldilocks creates VPA objects in Off mode and shows their recommendations; it inherits VPA's behaviour.
  • KRR. The README's own comparison lists autoscaling as future work, while the enforcer/ directory documents a mutating webhook. We describe the enforcer.
  • Kubecost was acquired by IBM (announced 2024-09-17). Kubecost created OpenCost (CNCF, Apache-2.0). Kubecost 3.0 no longer needs Prometheus.
  • StormForge was acquired by CloudBolt (announced 2025-03-31); the docs stay at docs.stormforge.io.
  • PerfectScale was acquired by DoiT (announced 2025-02-04). Its pricing page and a cached copy differ on the free tier (300 or 200 vCPU).
  • CAST AI. The JVM feature's status (early access or GA) isn't stated on the docs page.
  • The HPA, JVM and in-place claims for the commercial tools come from their documentation, not from our testing.

Kerf's own claims

Each Kerf cell is backed by this repository: modes and strategies in api/v1alpha1/common_types.go, guardrails and rollback thresholds in api/v1alpha1/rightsizingpolicy_types.go, HPA coordination and pricing in docs/architecture.md, metrics sources in docs/contracts/readiness.md.

Sources

  1. VPA: features, quick start, known limitations, releases, Kubernetes 1.35: in-place pod resize GA
  2. Goldilocks: repository, docs, advanced usage
  3. Robusta KRR: repository, enforcer
  4. IBM Kubecost: pricing, Kubescaler, OpenCost, acquisition
  5. StormForge: architecture, applier events, applier 2.12.0, schedule, product page
  6. PerfectScale: in-place rightsizing, automation CRD, Java containers, HPA, pricing
  7. CAST AI: workload autoscaling, in-place resizing, JVM, savings report, pricing