11.C4 — kill one component and the pod stops at the exact frame that component owns¶
Claim — with the scheduler stopped a new pod persists but stays Pending at the scheduler frame that gives up; with every node cordoned it is Pending/unschedulable; with the CNI plugin binary renamed it reaches the node but sticks at ContainerCreating, Seam C's last frame failing. Each stall names a frame, and the frame is one you cited in Seam B or Seam C. One drill, demonstrated at two seams: the gap tells you which frame owns which transition.
Rests on — Seam B (the bind frame) and Seam C (the CNI frame). Do this after both are cited — the stalls are only legible against the frames they name.
Topology — pair, continued.
Read — the drill's chaos-table row and the manual-drills standard. By hand, and permanently so: the skill is reading the stall's location off the pod's own status and events.
Do — seam B, no node to bind: cordon every node and create a pod; it persists but never binds. Uncordon one and watch the exact frame fire.
kubectl cordon 10.10.10.130 10.10.10.131 2>/dev/null; \
for n in $(kubectl get no -o name); do kubectl cordon ${n#node/}; done
kubectl run stalled --image=nginx --restart=Never
kubectl get pod stalled -o wide # Pending; describe shows 'unschedulable'
kubectl describe pod stalled | sed -n '/Events/,$p'
kubectl uncordon $(kubectl get no -o name | head -1 | cut -d/ -f2) # binds now
Do — seam C, no CNI to wire the netns: on the worker, move the plugin binary aside and create a pod pinned there; it schedules (Seam B completes) but sticks at ContainerCreating — the CRI→CNI seam P6 named as the one it could only watch, now failing where you cited it.
WK=10.10.10.131
ssh zain@$WK 'sudo mv /opt/cni/bin/bridge /opt/cni/bin/bridge.off' # or the plugin in use
kubectl run cnistall --image=nginx --restart=Never --overrides='{"spec":{"nodeName":"10.10.10.131"}}'
kubectl get pod cnistall -o wide # ContainerCreating, not Running
kubectl describe pod cnistall | sed -n '/Events/,$p' # CNI ADD failing
Observe — scheduler-side, the pod is Pending/unschedulable and never leaves the scheduler; CNI-side, it is Scheduled but never Running, stuck at ContainerCreating. Two different stalls, two different frames, each the one the missing component owns.
Expect — Pending when the bind cannot happen and ContainerCreating when the netns cannot be wired. The stall's name is the diagnostic: it points at the exact seam, which is the whole reason the trace was worth citing.
Write down — for each kill, the pod phase, the events line, and the Seam-B or Seam-C frame it corresponds to.
Teardown — restore the CNI binary and uncordon everything (the repair is part of the drill — a pod must reach Running again), then delete the pods; the topology stays: