I'm working on a fairly-high-throughput Python project (150-200k msg/s to start) that uses the kafka-python protocol via aiokafka. While profiling to determine where our bottlenecks are, Message instantiation, encoding, and decoding stand out as areas ripe for improvement.

This is an area I'd like to contribute, but the design of kafka.protocol seems more oriented towards being a readable reference implementation than a high-performance implementation. Are updates in this area desirable even if they lead to superficially uglier code? Likewise, how important is backwards compatibility to you?