You can also, of course, read in the data, rewrite it with the deleted portion removed, and write it out again.
(Disclosure: I work at Google, though I don't know anything about how Google handles this)
That said, there's also data without PII (anonymized) that is not attached to any specific user, that of course is not subject to this data wipeout as Google wouldn't know what to delete. Such data can be used for general ml models.
Not sure how it's done at Google but anonymization is tricky and it's easier to gather two forms of information from the start, one with PII and one without PII. I think most large companies are doing it like that.
EDIT: Just remembered one thing. At Google, even when building general models data that is rare cannot be used because it could be used to identify specific people. For example: If search query is used by less than X people on a given day then these queries cannot be used for model building.