How to successfully achieve bidirectional service reflection (endpointslice reflection) on namespaces with the same name on all kubernetes clusters

[jv4n5e]

I am currently facing the title's conundrum:
If I have a service being reflected from A to X, Y, and Z, and that is the only thing that is being reflected, how can I get a complete mesh between all pods on clusters A, X, Y and Z, if reflection is only occurring from cluster A to X, Y and Z? I then started by establishing bidirectional peering between A and X, but I could not get past that since having bidirectional peering is now 'overwriting' or conflicting endpointslices effectively cancelling each other.

This may be caused by declarative peering, as that is how I need to implement the peering, but I'm not sure. Please let me know what information you need to assist me, and what would be the correct way to achieve the initial goal of building a service mesh.
Since all I need is service reflection, I am leveraging virtual nodes on all Kubernetes clusters with create-node flag set to false

[cheina97]

Hi @jv4n5e, if I understand well A is your consumer cluster and X,Y,Z your providers. In that case, just offload a namespace, create a service targetting a deployment and after that you will be able to contact every pod from every cluster using the reflected service. Of course you cannot achieve direct connectivity between the clusters, your traffic will always pass through the consumer

[jv4n5e]

HI @cheina97 , thanks for your swift response. Can you please provide a bit more details?

• Offload a namespace on which cluster? Any?
• Doesn't matter how the namespace is offloaded?
• How would/should the scaling scenario be handled (e.g. what happens if I spawn new statefulsets on clusters that don't have reflection in place? How will the endpointslice propagate to other clusters?)?

Perhaps I need to first get a clear picture of consumers versus providers, although I am not interested in reflecting replicasets.

[jv4n5e]

@cheina97 I gave your response some thought. Are you saying that if I have a single consumer, then all pods on the offloaded namespace will be able to contact each other at all times, even when I spawn new statefulsets on the provider clusters? If I have the wrong idea, perhaps you can point me in the correct direction...

[cheina97]

@cheina97 I gave your response some thought. Are you saying that if I have a single consumer, then all pods on the offloaded namespace will be able to contact each other at all times, even when I spawn new statefulsets on the provider clusters? If I have the wrong idea, perhaps you can point me in the correct direction...

Yep, don't think about the topology. Just think that when you have a set of pods in the same namespace the can talk to each other, no matter where they are scheduled.