3.5 Reset for Lab 2

Time: 2 min Changes: node A back to exclusive; shared-load deleted; forgotten-notebook kept

Lab 2 serves a latency-sensitive model. It must not land on a time-sliced node — and if it did, KEDA would happily stack replicas on the same card and you'd learn nothing about autoscaling. Put node A back the way you found it.

Delete the four small pods

kubectl delete -f modules/03-sharing/shared-load.yaml

Un-slice node A

Removing the label makes the config manager restart the plugin with no sharing config:

kubectl label node "$NODE_A" nvidia.com/device-plugin.config-
sleep 15
kubectl get nodes -l karpenter.sh/nodepool=gpu -o custom-columns='NAME:.metadata.name,GPU:.status.allocatable.nvidia\.com/gpu,PRODUCT:.metadata.labels.nvidia\.com/gpu\.product,REPLICAS:.metadata.labels.nvidia\.com/gpu\.replicas'

node/ip-10-0-23-77.ec2.internal unlabeled
NAME                         GPU   PRODUCT       REPLICAS
ip-10-0-23-77.ec2.internal   1     NVIDIA-L4     1
ip-10-0-5-140.ec2.internal   1     NVIDIA-A10G   1

Both nodes are exclusive again. Node A is empty; node B has the notebook.

Confirm the image cache is warm

kubectl get ds vllm-image-prepull

NAME                 DESIRED   CURRENT   READY   UP-TO-DATE   AVAILABLE   NODE SELECTOR
vllm-image-prepull   2         2         2       2            2           karpenter.sh/nodepool=gpu

READY 2/2: the vLLM image is on both GPU nodes. If node B still shows READY 1, it's mid-pull; it'll finish during the first minutes of Lab 2.

Where you are

Node	Instance	GPU	Running
A	g6.xlarge (or g5)	1, exclusive	nothing — warm image, waiting for vLLM
B	g5.xlarge (or g6)	1, exclusive	forgotten-notebook, 0% util

Two GPU nodes, one useful. Lab 2 will use node A and, when it scales, need a third.

If you fell behind

It's fine to arrive at Lab 2 with node A still sliced or the shared-load pods still running — but run this page's two commands first, or vLLM's replicas will pile onto the shared card. ./scripts/validate-lab1.sh still passing is the bar.

Next: Lab 2 →