>
> > > What this does is to create a zero length file, and the rename unlinks
> > > the tablename file, and puts the zero-length file in it's place.
> > > rename() is atomic, so there is no time that the table file does not
> > > exist.
> > >
> >
> > Let
> > i1 be the inode of zz
> > i2 be the inode of tablename
> > before rename().
> >
> > Does this mean
> >
> > New backends read/write i1 inode and
> > backends that have the table open read/write i2 inode ?
> >
> > If so,it seems wrong.
> > All backends should see same data.
>
> Yes, I can see your point. It would show them different views of the
> table.
>
> So, as you were saying, we have no way of invalidating file descriptors
> of other backends for secondary segments.
It seems DROP TABLE has a similar problem.
It has been already solved ?
> Why does truncating the file
> not work? Any ideas?
>
I have gotten no bug reports for my trial implementation.
AFAIK,only Ole Gjerde has tested my patch.
Is it sufficient ?
Regards.
Hiroshi Inoue
Inoue@tpf.co.jp