Kerf

Kerf is an open-source (Apache-2.0), self-hosted Kubernetes rightsizing tool. It reads container usage from Prometheus or VictoriaMetrics, recommends CPU and memory requests and limits for every workload, prices the savings per node type, and, where a policy allows it, applies the change with an in-place pod resize or a rollout inside maintenance windows. Rolls back on OOM kills, restarts or CPU throttling. Everything runs in your clusters.

Long (for listings with room) Kerf finds the gap between what your Kubernetes workloads request and what they use, and closes it.

Recommends CPU and memory requests and limits per container from usage history (Prometheus or VictoriaMetrics; Thanos, Mimir and Amazon Managed Service for Prometheus also work). Prices the savings in dollars per node type, from the AWS price list or your own table. Applies changes only where a policy says so: off, audit or auto-patch per workload and per resource, in-place pod resize or rollout, inside maintenance windows, with guardrails and automatic rollback when a health watch sees OOM kills, restarts or CPU throttling. Every cluster starts in audit. Understands HPA-coupled workloads (requests and HPA target change together) and JVM containers (heap-aware memory). Fits GitOps: copy a recommendation as a patch instead of letting Kerf apply it. Hub and spoke: one central hub with the web UI and recommender, an agent per workload cluster that pulls recommendations, so the hub holds no write credentials to your clusters. One Helm chart, Apache-2.0, no SaaS, no telemetry to us. Version 0.x; EKS on Kubernetes 1.35 is the tested platform.