The upgrade notes re: performance impact of the message format change in 0.10.0.0 are a bit wishy washy in terms of what you should do. They describe the potential impact and then say (at the end of a paragraph): "To avoid such message conversion before consumers are upgraded to 0.10.0.0, one can set". This should probably be written as a playbook for doing the upgrade without affecting performance and should be a recommendation, not presented as just an option. Nobody with an existing cluster wants the perf impact of switching their brokers to the new format while consumers are still using the old format unless they have a cluster with extremely light load.
Additionally, if we have some simple benchmark numbers on the performance impact on brokers, we should include that information. Today it is very unclear how bad that change will be – 5%, 20%, 90% worse?