Prevent consumers from sending requests to servers outside of their zone.

[willgraf]

Is your feature request related to a problem? Please describe.
Multizone clusters are nice because they allow for redundant deployments. However, sending requests outside of the current zone comes with a fee (multizone egress fee). For the consumers, we are sending some large payloads across the network, so this can really add up.

Describe the solution you'd like
Ideally, the consumers will be able to send their requests ONLY to a tf-serving instance in their own zone.

[dylanbannon]

One idea is to have separate services that target pods using node information (every node has its zone encoded as an attribute), in addition to their standard targeting mechanisms:
https://kubernetes.io/docs/concepts/services-networking/service/#defining-a-service

If we have, say, tf-serving-zone1 and tf-serving-zone2 services, then we can somehow tell our redis-consumers to send their traffic to the appropriate service. I'm not totally sure how this would happen, though.