Boundary reference architecture GCP

Hi! Trying out a POC of boundary using boundary-reference-architecture/deployment/gcp at main · hashicorp/boundary-reference-architecture · GitHub but one thing that looks strange from what I have read in general how the network should work.
If I understand correctly a client connecting needs direct network access to a worker (or thrue another worker with enterprise), but the API and control plane(?) can be loadbalanced statelessly.

In the GCP terraform:
Public IPs on the machines are only created if ssh access is enabled, a load balancer is put in front of api,controller and worker. Should it really be this way?