Fwiw I used to work for large ML models for a well know big Ads company and we did indeed made sure to retain all the models on all historical logs within a tim

Fwiw I used to work for large ML models for a well know big Ads company and we did indeed made sure to retain all the models on all historical logs within a time bound after users have requested for their data to be deleted. This predated GDPR. It was expensive but we had to retrain models frequently anyway because of other reasons so this just added a more strict cadence and time bound checks but didn't really inflate costs dramatically