This is excellent. In the past when replicating via Debezium from a system making heavy use of cascade deletes I’ve had to write a layer that infers these deletes by introspecting the database schema, building a graph of all cascades (sometimes several layers) and identifying rows that should have corresponding delete records. These can then be excluded in whatever downstream system via an anti-join. It works but it will be better to not have to do that and instead have first class support for cascades.