Re: Hammer: Transactional file updates

["Dmitri Nikulin" <dnikulin@xxxxxxxxx>]

On Sat, Aug 2, 2008 at 7:47 AM, Michael Neumann <mneumann@ntecs.de> wrote:
> That's not what I want. fsync only guarantees that the data is stored
> permanently. That's not my problem. I want to ensure that some
> write-operations are performed either-all-or-nothing.

I personally believe that Unix should have had a transactional file IO
API from the start, so that all modern file systems would implement it
and atomicity would be the standard, not a rare feature. Because right
now, even if a couple of file systems supported it, nobody would use
it because it's non-portable and the semantics aren't standardised.

Aatomic IO is pretty heavyweight compared to just streaming bytes onto
a block device. However, most new file systems are implementing it one
way or another, via journalling or COW or whatever, so obviously it's
a tradeoff we're willing to make. Having a proper standard API to
harness that potential would be a great advancement for Unix.

The same problem applies here... even if HAMMER had such an API and it
worked perfectly, relying on that API would forever tie your
implementation to HAMMER. If you decided to support UFS, ext3, ZFS,
etc. for storage, you'd have to write a custom transactional format
anyway, which becomes redundant when running on HAMMER. You'd have to
wait until an API becomes standard on at least the platforms you
intend to support.

-- 
Dmitri Nikulin

Centre for Synchrotron Science
Monash University
Victoria 3800, Australia