There are better ways to snapshotting [LWN.net]

There are better ways to snapshotting

Posted May 7, 2009 20:54 UTC (Thu) by anton (guest, #25547)
Parent article: The two sides of reflink()

Many posters seem to think of snapshotting as a good use of reflinks, but snapshots should be created on the whole file system at once, atomically (so you get a snapshot that is consistent across files). File systems like btrfs have this functionality, and one can also use LVM for that (with a little fs support), so snapshots don't seem to be a good use and justification of reflink().

(Log in to post comments)

There are better ways to snapshotting

Posted May 10, 2009 0:05 UTC (Sun) by butlerm (subscriber, #13312) [Link]

One of the advantages of ZFS style snapshotting is it is a constant time
operation that is far more efficient than walking a directory tree and
creating thousands of new inodes. It doesn't work for making a (writable)
snapshot of a snapshot, however, and that is a very important use case for
virtualization, for example.

Copy-on-write filesystems could use a good interface (like this one) to make
efficient *copies* of files in the same (extended) filesystem. That is
something that ZFS and BTRFS both appear to lack, and which for portability
reasons is easy to downgrade to an ordinary file copy.

There are better ways to snapshotting

Posted May 10, 2009 18:36 UTC (Sun) by anton (guest, #25547) [Link]

One of the advantages of ZFS style snapshotting is it is a constant time operation that is far more efficient than walking a directory tree and creating thousands of new inodes. It doesn't work for making a (writable) snapshot of a snapshot, however

That may be a limitation of ZFS, but it's not a necessary limitation. LLFS can do it, and nothing I have read about Btrfs indicates that it cannot do it.

Copy-on-write filesystems could use a good interface (like this one) to make efficient *copies* of files in the same (extended) filesystem.

Yes, implementing cp more efficiently seems to be the main benefit I see from this system call.

There are better ways to snapshotting

Posted May 11, 2009 5:22 UTC (Mon) by butlerm (subscriber, #13312) [Link]

Netapp does writable snapshots of writable snapshots. The advantage of the
ZFS scheme is you can have an unlimited number of snapshots, where with
Netapp these days I believe you get 256.

The ZFS scheme is based on logical sequence numbers, Netapp uses block maps.
I understand that BTRFS allows nested writable snapshots, but that comes at a
general performance cost that ZFS doesn't have.