This is wrong; you can implement atomic transactions in filesystems. As in prepare a change, and then publish it, such that the consumers of that event only see the before and after view, and not something in between. There are many ways to do this, on various scales, involving parts of a file, a whole file or multiple files. Filesystems really have no such disadvantage over databases.<p>Think about a remote git repo. Multiple users can push. It's based on files.<p>This HN site is based on files.
If this problem is as real and fundamental as presented, and the solution really is this equally fundamental new paradigm, then it should be a kernel module and not a product.