Re: [PATCH] lock transaction-current separately from current

On Thu, Nov 08, 2007 at 06:06:12PM -0800, David Glasser wrote:
> > We could switch the lockfile to a new transaction-current-lock, sure,
> > though it'd be much better if we switched to:
> >
> > open transaction-current
> > lock
> > read
> > seek
> > write
> > fflush
> > close
>
> Hmm. Are you sure this is safe? What if the server crashed while
> halfway through writing out the line? (Or between the trunc that I
> suspect would be necessary there and the write?) Sure, in practice
> buffering is going to make the former case impossible, but is it
> really OK?
>

I believe that writes below the block size of the disk (as these will
be) are guaranteed atomic by POSIX, though I admit I haven't verified
that.

We don't need a trunc() since we're always going to overwrite with
something the same size or longer.

Regards,
Malcolm