I think whether reassembling more coarse grained changes is realistic depends quite heavily on the schema and the existing usage. There's enough schemas, even p

I think whether reassembling more coarse grained changes is realistic depends quite heavily on the schema and the existing usage. There's enough schemas, even pretty normalized ones, where that's pretty simple for the frequently changing objects. In others it's pretty complicated and an approach like you describe, where the stream is just used to trigger a request to retrieve a resource, is more realistic. Especially if a somewhat efficient API to get data from the database already exists. In other cases infrastructure for transforming data from the normalized to denormalized data already exists. E.g. as part of the process of transforming the OLTP data into a data warehouse. In those cases it's not necessarily hard to add different transformations.