list of filesystems with known issues

[Matt Mackall]

On Tue, 2009-03-03 at 21:38 +0000, Frank A. Kingswood wrote:
> Jens.Wulf at sew-eurodrive.de wrote:
>  > Hi all,
>  >
>  > every once in a while someone with a corrupted repo pops up on this 
> list. Sometimes it is because he used hg on a filesystem which has issues.
>  > Searching the bugtracker, wiki and mailing list didn't reveal that 
> much about filesystems which should be avoided, though. Maybe I just 
> didn't use the right keywords, but everything I found was
>  >   "There are known problems with XFS on Linux, but that's about it."
>  >
>  > Which ones cause problems?
> 
>  From recent experience, avoid a hard reboot on ext3 while committing 
> changes.

Unfortunately, this is a problem with basically all filesystems on all
desktop operating systems. Modern SATA devices offer no guarantee that
they'll respect flush-to-disk requests (eg via sync()) by actually
writing out their onboard caches. So short of some very heavy-handed and
massively performance-impairing barrier operations at the driver level
that basically no OS implements, all applications are at risk here.

Mercurial's exposure is relatively minor. Generally a repository hit by
a power failure can be cleanly recovered back to the pre-commit state
simply by cloning it.

But in discussion with major drive vendors, I've learned that power
failures can occasionally cause spurious writes to unrelated data. This
is one of the reasons why block-level checksums for Btrfs and ZFS are
considered a major feature.

-- 
http://selenic.com : development and support for Mercurial and Linux