You’re still going to pay the cost of maintaining referential integrity — you’re just doing it on the app side. You can do it faster by being not-correct — eg you don’t need a lock if you ignore race conditions — but it’s not like the database is arbitrarily slow at doing one of its basic fundamental jobs.

Of course, you can just skip the validation altogether and cross your fingers and hope you’re correct, but it’s the same reasoning as removing array bounds checking from your app code; you’ve eked out some more performance and it’s great until it’s catastrophically not so great.

Your reasoning should really be inverted. Be correct first, and maintain excessive validation as you can, and rip it out where performance matters. With OLTP workloads, your data’s correctness is generally much more valuable than the additional hardware you might have to throw at it.

I’m also not sure why dropping/creating foreign keys is a big deal for migrations, other than time spent