On Mon, 24 Feb 2003, Larry Rosenman wrote:
>
>
> --On Monday, February 24, 2003 12:28:03 -0500 Tom Lane <tgl@sss.pgh.pa.us>
> wrote:
>
> > Dennis Gearon <gearond@cvc.net> writes:
> >> And what about all the stuff in memory, pending transactions, etc?
> >
> > If he's got an accurate snapshot of the WAL files, everything the
> > database has claimed to be committed will be committed. It's the same
> > scenario as a crash.
> >
> > The trouble with most "I'll just take a tar dump/rsync copy/whatever
> > of the $PGDATA directory" backup solutions is that you don't get a copy
> > of the WAL files that's consistent with what you copy out of the other
> > subdirectories --- and any time skew there will burn you. See recent
> > reports of duplicate rows for an example. But if NetApp can implement
> > an instantaneous copy (and offhand it seems that they should be able to,
> > since they sit between PG and the disk) then it should work.
> >
> > I'd still recommend some tests before trusting it ;-)
> the netapp facility is called SnapShot.
>
> What happens is that the original blocks of the file are write protected,
> and any changes are put elsewhere. (I don't recall whether it's the new or
> the old blocks that get saved, but it doesn't matter, the effect is the
> same).
>
> The snapshot stuff in Veritas' VxFS is very similar.
>
> It makes a consistent view of the filesystem available for backup.
Sounds like they're using MVCC for filesystems. :-)
