I am just curious what is the optimal way to publish and consume messages, ignoring durability, persistence and similar things, but rather from the network perspective in a cluster?
If we publish a message over a connection opened to server 1 (s1), but the queues master-locator-node is on server 2 (s2), the server has to move that message from s1 to s2, right?
It would be optimal to always consume from queues that are "local" to the server we are connected on, meaning that all the queues we consume from over our connection are located on that server, wouldn't it?
Is this overcomplicating? Or would it be best to always publish to and consume from servers where the queue is located? I am dealing with somewhere around 3B messages daily, so I am trying to reduce latency and load as much as possible.
Yes, always publishing to and consuming from the queue master node is optimal. Your understanding of what happens when you connect to a non-master node is correct. Of course, this means you will have to make your applications aware of this information (from the HTTP API).
If you're not worried about message loss, there's little need for a cluster in this scenario.