Smart Sizing
Recommends a CPU and memory request and limit for each container from its real use, every 10 minutes, and applies them when you choose.
You pay for 10 CPU cores and use 1
CAST AI, 2025 Kubernetes Cost Benchmark Report; kubernetes.io, Resource Management for Pods and Containers.
Average CPU utilization across Kubernetes clusters is 10%, and memory utilization is 23% (CAST AI, 2025). Requests are set once, with a large margin, and rarely changed.
The scheduler places a Pod by its request, and the kubelet reserves that amount on the node, whatever the container really uses (kubernetes.io).
A request that is too small has its own cost. A CPU limit below what the container needs is enforced by throttling (kubernetes.io).
Bar heights are illustrative. The 10-minute cycle is from the Wave Sizing docs.
How Wave does it
- Engineerwrites requests in the manifest, once
- applykube-schedulerplaces the Pod by its request
- bindkubeletreserves the request on the node
- usagekubectl topshows real use, and the request does not change
- Wave Agentreads CPU and memory from cAdvisor on every node
- metricsSmart Sizing analysis algorithm7+ days of CPU, memory, and SLI metrics
- every 10 minNew request and limitper container, within its buffer, min, and max
- applyWave Coreapplies it by Resize Now or Realtime Resizing
Measure every container
The Wave Agent reads CPU and memory from cAdvisor on every node. First recommendations arrive in 24 to 48 hours and are stable after 7 days (Wave Sizing docs).
Recommend a request and a limit
The Smart Sizing analysis algorithm uses 7+ days of usage and SLI metrics. For workloads scaled by HPA or Autopilot, it uses a workload performance model instead. A new recommendation comes every 10 minutes.
Apply when you choose
Apply by hand with Resize Now, or turn on Realtime Resizing per workload and per container. Restore Original puts the first values back at any time.