Asynchronous commit

I was kinda hoping to read something like: "There was a meeting of PostgreSQL and kernel developers. Linux is going to get asynchronous commit."

Suppose I am using the Unix write-close-rename atomic file update idiom, and I write five versions of a single file (A, B, C, D, and E) in rapid succession, and then there's a crash. I usually don't care which version of A-E I end up with, as long as I get some version intact (or, if version A created the file in the first place, no version at all). In PostgreSQL, I can ask for asynchronous commit, which does more or less exactly that.

Unfortunately there's no way (that I know of) to ask the kernel for this and still be fast. I seem to have three bad choices: call fsync() where I want a write barrier, which slows down both the process that is writing the file and every other process sharing a disk with the writing process; or don't call fsync(), which may under some circumstances give me a file that is empty or broken; or put clunky workarounds in application code (e.g. use a tmpfs for the file and copy the file to a persistent disk with rsync in a loop running at whatever speed the disk can keep up with, or put the filesystem in a database server that implements asynchronous commit).

Ideally the kernel would provide some mechanism that preserves write order around user-supplied barriers, but does not otherwise affect the scheduling of writes or the process doing the writing, and still allows the kernel to freely reorder or coalesce writes between the barriers.