kubernetes duplicate pod ip addresses - are pod ips supposed to be unique cluster wide?

I read this in the doc:

Every Pod gets its own IP address ... pods on a node can communicate with all pods on all nodes without NAT.

Should I read that as "every pod gets its own unique cluster wide IP address"?

I assumed this was the case, but reason I ask is I noticed pods with the same IP addresses just on different nodes just after I initialized a new cluster following the instructions here. The cluster has 3 nodes test-vm{4,5,6}, with test-vm4 as master, running on a local dummy network 10.1.4.0/16. I used flannel for the CNI and set it up like this:

kubectl patch node test-vm{4..6} -p '{ "spec": { "podCIDR": "10.244.0.0/16" } }' # Had to do this because didn't set it on cluster init. See https://stackoverflow.com/a/60944959/2038383.
kubectl apply -f https://raw.githubusercontent.com/coreos/flannel/master/Documentation/kube-flannel.yml

Notice 3 IPs occur twice for 2 different pods - 10.244.0.{2,3,4}:

$ kubectl get pods --all-namespaces -o wide -w
NAMESPACE     NAME                               READY   STATUS    RESTARTS   AGE     IP           NODE       NOMINATED NODE   READINESS GATES
default       curl                               1/1     Running   0          14m     10.244.0.4   test-vm6   <none>           <none>
default       my-nginx-cf54cdbf7-d6s9m           1/1     Running   0          17m     10.244.0.3   test-vm6   <none>           <none>
default       my-nginx-cf54cdbf7-twrvw           1/1     Running   0          17m     10.244.0.2   test-vm6   <none>           <none>
default       my-nginx-cf54cdbf7-xpff6           1/1     Running   0          17m     10.244.0.4   test-vm5   <none>           <none>
default       my-nginx-more-5f79688b9d-4c9jk     1/1     Running   0          3m10s   10.244.0.6   test-vm5   <none>           <none>
default       my-nginx-more-5f79688b9d-7htsn     1/1     Running   0          3m18s   10.244.0.5   test-vm5   <none>           <none>
default       my-nginx-more-5f79688b9d-gqz9b     1/1     Running   0          3m4s    10.244.0.7   test-vm5   <none>           <none>
default       nginx1                             1/1     Running   0          9s      10.244.0.8   test-vm5   <none>           <none>
kube-system   coredns-64897985d-kt82d            1/1     Running   0          41m     10.244.0.2   test-vm5   <none>           <none>
kube-system   coredns-64897985d-rd7gz            1/1     Running   0          41m     10.244.0.3   test-vm5   <none>           <none>
kube-system   etcd-test-vm4                      1/1     Running   0          41m     10.1.4.36    test-vm4   <none>           <none>
kube-system   kube-apiserver-test-vm4            1/1     Running   0          41m     10.1.4.36    test-vm4   <none>           <none>
kube-system   kube-controller-manager-test-vm4   1/1     Running   0          41m     10.1.4.36    test-vm4   <none>           <none>
kube-system   kube-flannel-ds-snkhk              1/1     Running   0          29m     10.1.4.38    test-vm6   <none>           <none>
kube-system   kube-flannel-ds-wtmqg              1/1     Running   0          29m     10.1.4.37    test-vm5   <none>           <none>
kube-system   kube-flannel-ds-x46xw              1/1     Running   0          29m     10.1.4.36    test-vm4   <none>           <none>
kube-system   kube-proxy-mjl69                   1/1     Running   0          41m     10.1.4.37    test-vm5   <none>           <none>
kube-system   kube-proxy-vz2p2                   1/1     Running   0          41m     10.1.4.36    test-vm4   <none>           <none>
kube-system   kube-proxy-xg4gg                   1/1     Running   0          41m     10.1.4.38    test-vm6   <none>           <none>
kube-system   kube-scheduler-test-vm4            1/1     Running   0          41m     10.1.4.36    test-vm4   <none>           <none>

Despite what the docs say, all pods can't communicate with each other. They can only communicate with pods on the same node and it's causing errors. Wondering whether this is a red flag that something is wrong or not, and looking for clarification on this one point about pod IP address uniqueness.

I figured it out. Firstly, yes pods are absolutely supposed to have a cluster wide unique IP address. It's fundamental to the way k8s works. The linked k8s doc is crap, and leaves the question a little open. Better worded sources:

Platforms like Kubernetes assume that each container (pod) has a unique, routable IP inside the cluster. The advantage of this model is that it removes the port mapping complexities that come from sharing a single host IP. -- https://github.com/flannel-io/flannel#networking-details

One of the core requirements of the Kubernetes networking model is that every pod should get its own IP address and that every pod in the cluster should be able to talk to it using this IP address. -- https://ronaknathani.com/blog/2020/08/how-a-kubernetes-pod-gets-an-ip-address/

Now the question is, why are my pods being assigned same IP addresses? Basically doing this is at flannel CNI init is wrong ( I copied this suggestion from this SO answer):

kubectl patch node test-vm{4..6} -p '{ "spec": { "podCIDR": "10.244.0.0/16" } }' # Had to do this because didn't set it on cluster init.

The podCIDR has to be unique for each node. This is how k8s ensures each scheduled pod has a unique IP address - each node assigns some IP in it's podCIDR. See this great blog post explaining it. The above is not equivalent to setting --pod-network-cidr on kubeadm init like you are supposed to. The --pod-network-cidr command line option actually corresponds to ClusterConfiguration networking.podSubnet. So if you need to set it after the fact you remove flannel then edit the cluster configuration (haven't actually tested this approach, I just re-initd with --pod-network-cidr set):

kubectl delete -f kube-flannel.yml
kubectl edit cm -n kube-system kubeadm-config # and add setting.

Once set:

the control plane will automatically allocate CIDRs for every node. -- https://kubernetes.io/docs/reference/setup-tools/kubeadm/kubeadm-init/.

If you going to set each node's podCIDR setting it must be unique for each node. You should avoid setting it manually if nodes are expected to be coming and going dynamically - which is the normal scenario.

UPDATE: The above method of setting the ClusterConfiguration networking.podSubnet after init does not actually work. It doesn't even work if you de-register and re-register all workers nodes which is annoying. AFAIK the only way to get automatic node podCIDR setting to work is to blow away your cluster and re-initialize with --pod-network-cidr set or networking.podSubnet set in the initial config (see --config option).