Idle Node Detection
Finds nodes that run only DaemonSet Pods, with no application Pod for about a week, and shows the CPU and memory they hold.
A node runs only DaemonSet Pods for 1 week and still costs full price
kubernetes.io, DaemonSet; Cluster Autoscaler FAQ (kubernetes/autoscaler on GitHub); Datadog State of Cloud Costs, 2024. The 1-week span is illustrative.
A DaemonSet ensures that all (or some) nodes run a copy of a Pod (kubernetes.io). So a node with no application Pod still runs node-exporter, fluentd, or calico-node. Kubernetes does not mark that node idle, and the cloud provider bills it like any other node.
Cluster Autoscaler considers a node for removal only when the Pod requests on it are below 50% of its allocatable, and it counts DaemonSet Pods in that sum by default. It does not scale down a node group that is already at its minimum size (Cluster Autoscaler FAQ).
83% of container costs are associated with idle resources, and about 54% of that wasted spend is cluster idle, the cost of overprovisioning cluster infrastructure (Datadog State of Cloud Costs, 2024).
Idle rule, about a week, and Pod Count: Wave Idle Node Detection docs. 50% check and minimum size: Cluster Autoscaler FAQ. The nodes and Pod counts are illustrative.
How Wave does it
- DaemonSetputs a Pod on every node, used or not
- requestsCluster Autoscalercounts DaemonSet Pods in its 50% check
- min sizeNode groupat its minimum size, not scaled down
- keptNodestays Ready and billed, with no flag
- Wave Agentreports the Pods on every node
- gRPCWave Coresorts system, DaemonSet, and application Pods
- classifyIdle windowno application Pod for about a week
- a weekWave Consoleidle CPU and memory per cluster
Sort every Pod
Wave sorts the Pods on each node into three groups: system Pods, DaemonSet Pods, and application Pods from Deployments, StatefulSets, and Jobs. One application Pod means the node is in use.
Wait out short states
A node is flagged only after it stays without application Pods for about a week by default. A scale-down delay or a batch Job that just finished does not flag it. A node that reads as completely empty is treated as a data gap, not as idle.
Show the waste
The Idle Nodes page shows the idle node count, idle CPU, and idle memory per cluster, and the Pod Count of each node by type. You decide what happens next: drain and delete the node, move workloads onto it, or keep it with the annotation waveautoscale.io/idle-node-detection="false".