How traffic actually flows.
WHAT THE REQUIREMENT IS IN A CLUSTER
Every pod can reach every other pod, without translation.
WHY THAT MODEL WAS CHOSEN
It makes applications simpler: a pod's address is the same from anywhere.
WHAT IMPLEMENTS IT
A network plugin, chosen when the cluster is built.
WHAT THE APPROACHES ARE
An overlay network, encapsulating traffic between nodes Routing, where the underlying network knows pod addresses
WHAT OVERLAYS PROVIDE
Working anywhere, without changes to the underlying network.
WHAT THEY COST
Encapsulation overhead, and harder packet inspection.
WHAT ROUTED APPROACHES PROVIDE
Better performance, and visibility.
WHAT THEY REQUIRE
Cooperation from the network.
WHAT TO CHECK ABOUT YOUR PLUGIN
Whether it supports network policies.
WHY
Policies are silently ineffective without support.
WHAT SERVICE TRAFFIC ACTUALLY DOES
Gets redirected to one of the backing pods by rules on each node.
WHY THAT MATTERS FOR DIAGNOSIS
The service address is virtual, and never answers directly.
WHAT TO CHECK WHEN CONNECTIVITY FAILS
Resolution first Then whether the service has endpoints Then policies Then the plugin's own health
WHAT TO AVOID
Assuming the problem is the application.
