k8s-rightsizelisted

Right-size Kubernetes pod CPU/memory requests from real Prometheus usage data (per-pod p95 CPU, p98 memory over 7 days), compute how many nodes the cluster needs, and tune KEDA queue-based autoscaling. Use when the user asks to rightsize workloads, fix over/under-provisioned requests, cut node count or cloud cost, size Celery or background workers, tune KEDA listLength or HPA settings, or answer "how many nodes do we need". Works for any namespace with kube-prometheus and cAdvisor metrics.
smk-labs/claude-plugins · ★ 11 · DevOps & Infrastructure · score 79

Install: claude install-skill smk-labs/claude-plugins

# k8s-rightsize: data-driven pod request sizing from Prometheus Right-size CPU/memory **requests** for workloads in a namespace using real 7-day usage from Prometheus, then compute how many nodes the cluster needs. Built for queue-worker fleets (Celery and similar) plus web/API pods, but works for any namespace. ## Best-practice model (researched: Datadog, Kubernetes docs) - **Requests drive scheduling (bin-packing); limits enforce a runtime ceiling.** (K8s docs) - **CPU is compressible.** A CPU *limit* throttles via CFS even when the node has spare CPU, spiking p95/p99 latency. For **async/background workers (Celery and similar), do NOT set CPU limits**: let them burst. Set the **CPU request to observed steady-state (median to p95)**. - **Memory is incompressible.** It can only be OOM-killed, not throttled. Set **memory request to p95-p98**. Optionally add a **memory limit of about p98 x 1.25** to stop a leak from taking down a node. (Some teams deliberately run fully burstable with no memory limits; respect the owner's choice, but name the trade-off.) - **Spread replicas across nodes** (`topologySpreadConstraints`, `maxSkew: 1`, `topologyKey: kubernetes.io/hostname`, `whenUnsatisfiable: ScheduleAnyway`) so bursts scatter instead of stacking. This is what makes "no CPU limit" safe. - Celery: use `--max-tasks-per-child` to bound memory-leak creep. ### Percentile guidance (no CPU limit case) The request only controls *packing density*, not the ceiling. Higher percentile =