Craig Ringer <craig@postnewspapers.com.au> writes:
> Tom Lane wrote:
>> This is not correct; PG *never* overwrites an existing record (at least
>> not in any user-accessible code paths).
> That's what I always thought, but I encountered some odd behaviour while
> trying to generate table bloat that made me think otherwise. I generated
> a large table full of dummy data then repeatedly UPDATEd it. To my
> surprise, though, it never grew beyond the size it had at creation time
> ... if the transaction running the UPDATE was the only one active.
> If there were other transactions active too, the table grew as I'd expect.
> Is there another explanation for this that I've missed?
In 8.3 that's not unexpected: once you have two entries in a HOT chain
then a later update can reclaim the dead one and re-use its space.
(HOT can do that without any intervening VACUUM because only within-page
changes are needed.) However, that only works when the older one is in
fact dead to all observers; otherwise it has to be kept around, so the
update chain grows.
regards, tom lane