After we hit a certain number of frames/second, RabbitMQ became a bottleneck. My solution was to write the JPEG frame in the shared memory, then pass around shared memory ID, offset and frame length, instead of the whole frame. Only services that actually needed the frame would read it, others would just pass around a small JSON object. After that was implemented, the bottleneck disappeared.
I don't know if this qualifies as a "hack", but it certainly felt that way in the moment.
