err = addPodToHostNSIpset(pod, podIPs, &s.hostsideProbeIPSet)
	if err != nil {
		log.Errorf("failed to add pod to ipset: %s/%s %v", pod.Namespace, pod.Name, err)
		return err
	}

	log.Debug("calling CreateInpodRules")
	if err := s.netnsRunner(openNetns, func() error {
		return s.iptablesConfigurator.CreateInpodRules(&HostProbeSNATIP)
	}); err != nil {

	assert.NoError(t, err)
	fakeIPSetDeps.AssertExpectations(t)
}

// for tests that call `runtime.GC()` - we have no control over when the GC is actually scheduled,
// and it is flake-prone to check for closure after calling it, this retries for a bit to make
// sure the netns is closed eventually.
func assertNSClosed(t *testing.T, closed *atomic.Bool) {
	for i := 0; i < 5; i++ {
		if closed.Load() {
			return
		}

* It should not be client-specific.
  * In Istio sidecars, historically we had a lot of client-specific xDS. For example, putting the xDS-client's IP back into the xDS response. This makes efficient control plane implementation (most notably, caching), extremely challenging.
  * In practice, this largely means that references are fully qualified in the API. IP Addresses (generally) have a network associated with them, node names have a cluster associated with them, etc.