cdsbalancer: priority LB name generation logic creates new names everytime when the EDS resource has no localities

[easwars]

This unit test can help repro the problem:

func (s) TestNameGenerator_Generate_EmptyLocalities(t *testing.T) {
    ng := newNameGenerator(0)
    // Repeated calls to generate with an empty list of localities should generate the same single name.
    for i := 0; i < 5; i++ {
        if diff := cmp.Diff(ng.generate([][]xdsresource.Locality{{}}), []string{"priority-0-0"}); diff != "" {
            t.Errorf("generate() = got: %v, want: %v, diff (-got +want): %s", ng.generate(nil), []string{"priority-0-0"}, diff)
        }
    }
}

[easwars]

There maybe other inconsistencies with our implementation that needs more debugging: grpc/proposal#541 (comment)

[easwars]

The fact that our name generator creates a new name everytime for EDS resources with no localities is problematic in tests that verify scenarios where an EDS response is NACKed. For example, in aggregate cluster tests where we have an aggregate cluster pointing to an EDS and DNS cluster, and the EDS resource is being NACKed, we expect the priority LB policy to failover to the DNS cluster. But because the go-control-plane server continuously re-sends resources that are being NACKed, and a NACKed EDS resource is represented by the xDS dependency manager as an EDS resource with no localities, the name generator in the CDS policy ends up creating a new name for every update that we receive from the management server. This results in the priority LB policy to constantly try to use the higher priority EDS child which is not expected to work.