Kernel's fsyncing behavior is one thing, but just relying on a massive amount of fragile C code running in kernel is a significant liability, especially if your software is a centralized database and crashes, panics will bring down everything.
Yes, and also the traditional answer was that the kernel handles weird and complicated hardware and can talk to RAID controllers properly, but nowadays hardware has much less variance, and RAID is rare (and arguably unnecessary for a direct-IO database).
I think it'd be viable for an enterprise-y database to do IO directly over NVMe. Imagine the efficiency and throughput gains you could get from a database that (1) has a unified view of memory allocation in the system (2) directly performs its page-level IO on the storage devices.
