I opened this issue on the design repo https://github.com/ros2/design/issues/251 as a response to conversation on https://github.com/ros2/design/pull/239 (which describes a new architecture for intra-process communications within rclcpp)
There seems to be interest outside myself (@alsora @ivanpauno @raghaprasad @gbiggs, sorry if I missed anyone else), in coming up with a solution that makes this intra-process communication optimization available 1) to all clients e.g. rclpy, and 2) work more easily out of the box for topic-derived-features e.g. Services. It feels like we need a top down architecture that we can split up into concrete implementables and get contributors going on.
Not sure if here or the GitHub issue are the best place for the conversation but opening this for maximum visibility.