Cluster Resource Forecast
Projects the Pod CPU and memory requests of a cluster against node capacity, with snapshots at 7 and 30 days, so you can add nodes or cut requests in time.
Requests reach 94% of node CPU, and the first warning is a Pending Pod
kubernetes.io, Reserve Compute Resources for System Daemons and Resource Management for Pods and Containers; Cluster Autoscaler FAQ; Wave Cluster Resource Forecast docs. 94% is the docs example and is illustrative here.
"The scheduler does not over-subscribe Allocatable" (kubernetes.io). When requests already take 94% of node CPU, the next Pods may not fit on any node. "If the scheduler cannot find any node where a Pod can fit, the Pod remains unscheduled until a place can be found" (kubernetes.io). The Pod stays Pending.
Cluster Autoscaler "periodically checks whether there are any pending pods" and then adds a node (Cluster Autoscaler FAQ). It acts after a Pod is already waiting.
Forecast line and the 7 and 30 day snapshots: Wave Cluster Resource Forecast docs. 94% of node CPU: docs example. Pending when no node fits: kubernetes.io. The curve and the other values are illustrative.
How Wave does it
- Podseach new Pod adds its requests
- requestsSchedulerplaces a Pod only where Allocatable has room
- no fitPodstays Pending with FailedScheduling
- pendingCluster Autoscaleradds a node after Pods are pending
- Wave Agentnode capacity and Pod requests
- gRPCWave Corekeeps at least 7 days of history
- trendForecastrequest trend at 7 and 30 days
- to 30dWave Consoleforecast line against node capacity
Learn the trend
Wave needs at least 7 days of history. It tracks node CPU and memory capacity next to the Pod requests and limits in the cluster.
Project requests forward
The CPU and memory charts add a Pod CPU Request (Forecast) and a Pod Memory Request (Forecast) line, with snapshots at 7 and 30 days. Where a forecast line climbs toward the capacity line is roughly when you would run short. There is no countdown; you decide what is too close.
Pick the cheapest fix
Add nodes when the growth is real. When requests are inflated, cut them first with Smart Sizing, and the forecast drops without new hardware. A Filter Group narrows the forecast to a subset of workloads.