> It'll always be stream-of-bytes, but you can emit in JSON or protobuf etc, if it's worth the cognitive overhead to do so.

This is the crux of it. That's great until you encounter a need for a schema, and then it's "schema-on-read" or some similar abomination. And the need might not manifest until you're pushing like 1TB/day or more of telemetry data with hundreds or thousands of engineers working on some >1MLoC monstrosity. Hard to dig out of that hole.

The situation is tragically optimal--we've achieved some kind of multiobjective local maximum on a rock in the sewer at the bottom of a picturesque alpine valley and declared victory. We should do better.

Or maybe I'm overly optimistic.