Tom Lane wrote:
> Simon Riggs <simon@2ndquadrant.com> writes:
> > On Tue, 2007-09-11 at 01:22 +0530, Pavan Deolasee wrote:
> >> I think the only difference is that the quick pruning does not mark
> >> intermediate tuples ~LP_USED and hence we may avoid WAL logging.
>
> > Sounds great.
>
> What it sounds is utterly unsafe. You can get away with not WAL-logging
> individual bit flips (that is, hint-bit-setting) because either state of
> the page is valid. If I read this proposal correctly it is to change
> t_ctid without WAL-logging, which means that a partial page write (torn
> page syndrome) could leave the page undetectably corrupted --- t_ctid
> is 6 bytes and could easily cross a hardware sector boundary.
OK, we certainly want to avoid WAL writes for pruning a single chain if
possible. One idea would be to just mark the item pointer as dead but
not repoint the first item pointer to point to the first live tuple.
Heikki, would that give us speedups similar to what you saw in your
earlier tests? It would if the majority of the cost of scanning the
page looking for the first live tuple was in doing the visibility tests
on every tuple.
Basically, when we are walking the chain via an index and we see the
tuple is dead we mark it as dead but don't repoint. We would repoint
and reclaim on a defragementation.
--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://www.enterprisedb.com
+ If your life is a hard drive, Christ can be your backup. +