The kubelet podresources endpoint returns allocated exclusive resources to active pods.
The endpoint incorrectly returns resources allocated to terminated pods.

There are 2 factors which concur to create the bug

1. the podresources API depends on inner working of kubelet to retrieve the list of currently active pods. Previously, the related function incorrectly returned active and terminated pods.
2. if the podresources API incorrectly consider a terminated pod, we run into another issue in  memory manager. The memory manager collects stale resources (assignment to terminated pods) only in the allocation flow. Thus, if no pods manage to get admitted, the kubelet through podresources API incorrectly reports memory resources assigned to a terminated pods. This reporting is bogus as these resources are not reserved anymore, but the podresources API cannot know that. This does NOT affect the allocation flow (first thing it does is cleanup) but does affect the reporting, and this behavior is not fixed upstream.


why this affects only memory?

1. device assignment is explicitely cleaned by the podresources API endpoint
2. the cpu assignment is not (and it should) but it is automatically 
cleaned every cpuManagerReconcilePeriod seconds, so it automatically 
recovers

this breaks in an unrecoverable way numa aware scheduling.

4.18.z (any)
actually reproduced in Server Version: 4.18.0-0.nightly-2025-04-13-142946
    

100%    

    1. configure the kubelet with memory manager policy = Static
    2. run a job whose pod qualify for memory pinning (see example manifest below)
    3. query the podresources endpoint on the nodes. The endpoint is node-local exposed through a unix domain socket. It has to be queried programmatically. Probably the simplest option is to download the `knit` tool from https://github.com/openshift-kni/debug-tools/releases/tag/v0.2.1 and to use it like `knit podres` with root privileges.


example manifest:
```
apiVersion: batch/v1
kind: Job
metadata:
  labels:
    app: idle-gu-job-sched-stall
  generateName: generic-pause-
spec:
  backoffLimit: 6
  completionMode: NonIndexed
  completions: 2
  manualSelector: false
  parallelism: 2
  podReplacementPolicy: TerminatingOrFailed
  suspend: false
  template:
    metadata:
      labels:
        app: idle-gu-job-sched-stall
    spec:
      containers:
      - args:
        - 1s
        command:
        - /bin/sleep
        image: quay.io/openshift-kni/pause:test-ci
        imagePullPolicy: IfNotPresent
        name: generic-job-idle
        resources:
          limits:
            cpu: 100m
            memory: 256Mi
        terminationMessagePath: /dev/termination-log
        terminationMessagePolicy: File
      dnsPolicy: ClusterFirst
      restartPolicy: Never
      schedulerName: default-scheduler
      terminationGracePeriodSeconds: 30
      topologySpreadConstraints:
      - labelSelector:
          matchLabels:
            app: idle-gu-job-sched-stall
        matchLabelKeys:
        - pod-template-hash
        maxSkew: 1
        topologyKey: kubernetes.io/hostname
        whenUnsatisfiable: DoNotSchedule

```    

kubelet returns memory resources assigned to terminated pod

either:
1. kubelet does not return the terminated pod
2. kubelet return the terminated pod, but without any resource assigned to it

possibly affects older versions of openshift
solved kubernetes upstream by the pod workers refactoring: the podresources endpoint (correctly) ignores terminated pods and only lists active pods