[Bug] Http/GRPCRoute's backendref port didn't take effect when updating HttpRoute/GRPCRoute/k8Service

[Zijun Wang]

For following scenarios, Http/GRPCRoute's backendref ports didn't take effect and the controller will create lattice targets with unexpected ports:

• Update k8Service
• Update Http/GRPCRoute (unexpected targets ports can be created in a moment and then be de-registered immediately with draining, the reason is service_controller Watches HTTPRoute and GRPCRoute and do enqueueImpactedService() )

Root cause:
In this PR #351, we only assign &latticeTargetsModelBuildTask{ backendRefPort: port} in (t *latticeServiceModelBuildTask) buildTargets(), but miss to do that in (b *LatticeTargetsModelBuilder) Build().

Adding backendRefPort: port logic in the (b *LatticeTargetsModelBuilder) Build()` is quite likely could fix it.

Furthermore, build targets logic could diverge triggering from k8sService/Route/ServiceExport, should we consider to refactor and combine them?

[Doyoon Kim]

For ServiceExport, instead of using backendRef ports we are relying on annotations for determining ports - because at the moment of creating TGs we may not have HTTPRoute in the same cluster. Please try setting up "multicluster.x-k8s.io/port" annotation on ServiceExport.

[Zijun Wang]

Ok, I think only happen for updating HttpRoute/GRPCRoute/Service scenarios, no ServiceExport

[Doyoon Kim]

So your explanation is:

• k8s Service change does trigger all these - ServiceReconciler, HTTPRouteReconciler, and ServiceExportReconciler.
• ServiceReconciler is not aware of port information so tries to register all endpoints instead of watching backendRef.

I think it makes sense. Actually I don't get why ServiceReconciler even exists, the other two reconcilers are able to handle that anyways.

[Doyoon Kim]

Perhaps a better path forward is to remove ServiceReconciler and letting the other two reconcilers watch endpoint events instead.