I wonder if snapshots on Linux inotify() event detection (or similar) would be easier. At the very least, skipping making snapshots that are exactly the same as previous versions would be a nice feature. I implemented an inotify() based angband cheat as perl script, once... no ZFS required! Does the average home directory change more than once per 5 minutes on average? I doubt it.
Zfs snapshots are extremely cheap. Its storage is structured as a tree, and changes are implemented though copy-on-write modifications from leaves to the root of the tree. That's how it e.g. eliminates the RAID5 write hole in software, for raidz.
Snapshots make use of this by taking another reference on that root block, so that it isn't disposed of (turned into free space) after the next change to the file system. That's why they're so cheap.
You should also see why this means that snapshots that are exactly the same as the previous snapshot should be no more expensive, in principle, than an extra hard link to a file. No more than an entry in some list somewhere, and an increment of a digit somewhere.
A snapshot mechanism based just on inotify would be way more expensive, without further help from the filesystem, like copy-on-write hard links. But then, almost all the magic is in the filesystem, not inotify.
Yes, ZFS is great, I use it myself. But technical efficiency does not equate to programming ease, and if I'm not mistaken on Linux you need root permissions to use ZFS snapshots right now. In addition, you are limited in scope to people's systems running ZFS, and who have home dirs on ZFS... which due to module-based compilation being the norm demands an initramsfs if you want to include root's home... and so on and so forth. It's not exactly 'no strings attached'. My point was, there's more than one way to do it (TMTOWTDI; the perl refrain) and inotify() actually has some benefits (portability across filesystems, non filesystem-wide scope potential, zero storage overhead in the case of no changes between two five minute periods due to a non-polling, event-driven architecture being some of the more obvious).
i'm work only on freebsd - so zfs was already there.
why snapshots every 5 minutes? it's cheap and i lost at most 5 minutes of my work if something bad happens.
