Isn't model polling already a part of TF Serving?

[switchfootsid]

I was reading through the documentation of the TF Serving architecture explicitly. It mentions that utilities for polling a file-system are made available out of the box.

New Sources could support new filesystems, cloud offerings and algorithm backends. TensorFlow Serving provides some common building blocks to make it easy & fast to create new sources. For example, TensorFlow Serving includes a utility to wrap polling behavior around a simple source. Sources are closely related to Loaders for specific algorithms and data hosting servables.

Did you try using these instead of Pub/Sub?