There are some older fsync() bugs (as famously explored by Postgres developers: https://wiki.postgresql.org/wiki/Fsync_Errors ) but I'm not aware of any modern mainstream kernel where this is broken. If I'm wrong, please tell me!

An application that really wants confidence in a write—to the extent that the underlying device and drivers allow—should use O_DIRECT. Or maybe there is a modern equivalent with io-uring. But that is not easy engineering :)