On Tue, Sep 23, 2008 at 6:06 PM, Billy Crook [email protected] wrote:
Journalling the files themselves would be redundant in XFS because new files and changes to files are Copy-on-write. The new sector is written first (even if the file isn't growing) then the metadata is updated. If any part of the write fails the metadata will be pointing back at the original sectors untainted by the failed write.
Is this a recent change? Because for a very long time, the XFS FAQ basically said "if it hurts, don't do it" when it comes to unclean shutdowns. I've never heard XFS described as copy on write.
Don't go setting filesystem requirements like how Journaltastic it has to be, unless you know the limits of usefulness and relevance of those requirements. If you're going to set requirements don't base them on the inner workings of the filesystem, but what you need to do with it. (Survive power failures without filesystem corruption) When it comes time to choose a filesystem, pick ext3. It's just the right choice unless you think you need more than 8TB of filesystem space. XFS can do that, but you're better off using multiple ext3 filesystems, and symlinks.
Ext3 it (at least as of today) Linux' filesystem. Period. There are thousands of choices. Only one right one (ext), and ext's current extremely stable version is 3.
I agree that ext3 is a great choice for stability and recovery. I've had plenty of unclean shutdowns without dataloss. There are, of course, psychopaths who want ext4 available in the next Ubuntu. But they're laughable, and the team tries to forward such requests to Theodore T'so's opinion on the subject.
Very shortly, there's going to be some new filesystems in order. btrfs and ext4 are both moving along and offer substantially different features. For example, btrfs does have copy on write semantics (http://lkml.org/lkml/2007/6/12/242).
Justin Dugger