Thanks for response! That's basically what i ended up with!
Avro whenever possible and for python stick with pydantic. Except that at the same time i want to keep avro schema for those pydantic models and just convert them with fastavro, because then i dont have to rely on the quality of code/schema generators. A little sceptic about that though, as keeping the same schemas in 2 different technologies compatible and double (de)serialization might be troublesome.