Interesting. I really like the idea of using the UNIX sockets to be, essentially, a message queue before going to Kafka. This may be used on my next project.
There is a dedicated thread for monitoring the socket so clients don't block trying to write. In case of transient load spike or slowness in Kafka, the bulk of the queueing should occur inside the daemon, where you can see queue lengths via web interface. God luck with your project :-)