Great article!  I use your AWS and Azure examples all the time with HealthShare and TrakCare deployments!!!  This GCP example will surely be beneficial as well!

Transparent HugePages (THP) are not the same as standard HugePages, and this is especially important for IRIS and its shared memory segment.  THP do not handle shared memory segments.  Please see my article that discusses this in detail:  https://community.intersystems.com/post/linux-transparent-hugepages-and-impact-intersystems-iris 

Message delivery is dependent on queuing, so if messages on the compute node(s) haven't completed through the production - especially on the outbound send to the Message Bank.  The Message Bank will only "bank" those messages that have been actually sent to it.  If there are still messages queued in the production, they will remain there in the production queues until those pods are restarted or PreStop hooks are used to allow the POD to have a grace period on container shutdown until all queues are empty.  An Interoperability Production is a stateful set, and the queues are required to support message delivery guarantees.