Comment 8 for bug 633686

If fdatasync is the issue (as hinted in the original comment) that might
be avoided in RPM. But I/O can only be delayed, not avoided, when writing
content onto the file system. So all that can be done is to avoid
fdatasync, which will mean that I/O will likely proceed after
both yum (and rpmlib) have stopped executing.

Again only a strace can tell the root cause. I'm just guessing based on the
available info.