A Pod object does not yet mean a running process
The cluster and component roles
API Server accepts the desired state, etcd stores it, the scheduler selects a Node, and kubelet ensures that containers start. A user's HTTP request does not pass through this control chain. A Node is a worker machine that provides resources. A Pod groups one or more application containers.
Requests and limits
A CPU request is a scheduling reservation. A limit restricts resource usage at runtime. Requests do not measure actual load or guarantee throughput. Memory is not modelled here; system CPU reservations are already excluded from the 2000m allocatable. 1000m equals one CPU. Allocatable is the capacity available to Pods after system reservations; here it is 2000m per Node.
Why Pending occurs
With a 1500m request, a Node with 2000m fits one Pod, even though 500m remains free. Three Pods require three such nodes. With a 500m request, four Pods fit. Do not reduce the reservation below the application's needs just to get a green status.
What we check
Choose the node count and reservation for three Pods, then compare Pending and capacity. A Namespace defines a logical scope for resources; a Node belongs to the cluster and is not inside a Namespace. Pending counts the Pods that have not yet been placed.
Task
Three Pods must serve 600 operations/s. Each Node has 2000m CPU available. Add nodes and align requests with application resource needs.
Each Node has 2000m allocatable CPU. Identical Pods are scheduled according to CPU requests; memory, affinity and taints are not modeled. The application requires 500m in this scenario. The rate of 200 operations/s is specified by the workload, not derived from requests.
- Add a Node.
- Compare requests with allocatable CPU.
- Eliminate Pending while preserving the application's reservation and staying within budget.