Kafka stores bytes; a schema system describes how applications interpret them. Backward compatibility asks whether a new reader can read older records. Adding a new required field without a default can break that promise because old records do not contain it. Changing an amount from number to string is another contract change. The local schema catalog models a narrow JSON field contract with string, number and optional types ending in ?. It is not Confluent Schema Registry or a complete Avro, Protobuf or JSON Schema validator. Evolve the requested subject while preserving the old fields. Inspect the new version and explain which old records remain readable.
Inspect the topics, partition logs and group offsets before changing the cluster.