How often do you create fake data sets?

If it were easier to do the thing I want, I'd do it all the time. What I want is to be able to point to either a database with empty tables or an SQL file and fill in the tables correctly with bogus data while respecting all the constraints. I have not found a tool that does that.

It is trivial to write something where a tool takes a spec written in a way the tool understands (but possibly nothing else) and then spew out data without having to worry about, e.g., foreign key relationships.

If I have a choice, I wire together something Perl that does what I want.

Sometimes, I do not have a choice and have to use something like https://www.mockaroo.com/ or https://generatedata.com/ but then I have fiddle with stuff.

I am fond of David Golden's https://metacpan.org/pod/Data::Fake

Not often. The last time I had to do it was four years ago for two projects: financial data derived from real data with consent of the owner for one project for demos (consulting), and fitness tracker data for testing (wrote a sort of simulator that generated data from several mock trackers with different timestamps and payloads).

For discreet tests it's best to keep it small and do it by hand. Bulk data in my experience is problematic - if multiple testing relies on it it quickly becomes frozen due to potential conflicts. And then it needs updating as the product changes.