Generally that technique is something that was pretty common 20 years ago but is not a good strategy for either anonymizing or for testing/capturing variations.

Generally that technique is something that was pretty common 20 years ago but is not a good strategy for either anonymizing or for testing/capturing variations. It actually is a bad strategy for many/most anonymization use cases, and its use would count as a data breach. There is so much good tooling now for synthetic generation just at the data set level. And usually there are custom semantics "embedded" in the values and arities of data relationships in multi-dataset contexts, where a random "mixing" of real data would cause unhelpful impossible breakage. So- no, don't implement, better uses for your time.