On Sun, 2008-01-27 at 15:02 +0000, Gregory Stark wrote:
> It occurred to me the other day that synchronized scans could play havoc with
> clustered tables. When you dump and reload a table even if it was recently
> clustered if any other sequential scans are happening in the system at the
> time you dump it the dump could shuffle the records out of order.
>
> Now the records would still be effectively ordered for most purposes but our
> statistics can't detect that. Since the correlation would be poor the restored
> database would have markedly different statistics showing virtually no
> correlation on the clustered column.
>
> Perhaps we should have some form of escape hatch for pg_dump to request real
> physical order when dumping clustered tables.
>
Thank you for bringing this up, it's an interesting point.
Keep in mind that this only matters if you are actually running pg_dump
concurrently with another scan, because a scan will reset the starting
point after completing.
Regards,Jeff Davis