Theo Schlossnagle wrote:
>
> On Feb 20, 2007, at 1:40 PM, Tom Lane wrote:
>
>> RPK <rohitprakash123@indiatimes.com> writes:
>>> I did not mean asking for undo from a life-time log. Since FlashBack
>>> Technology is already there, I just mean that world's most advanced
>>> database
>>> (PostgreSQL, as they say), must have an optimized way for undoing of at
>>> least a week changes.
>>
>> You're living in a dream world. Do you know any Oracle DBs who keep
>> enough rollback segments to go back a week?
>
> Ours go for a good 6 hours sometimes :-D
Eeven if it's just one hour, it's certainly better than nothing.
I fully agree that I'd not be acceptable to introduce performance
problems for _everyone_ by introducing flashback. But if you only
experience a drop in performance if you actually enable flashback
(by, let's say setting vacuum_min_deadtime=1h), then I don't see
why anyone would object to having support for some kind of flashback.
However, I just realized that doing this is much harder than I initially
thought, because catalog access always happens with SnapshotNow, and
e.g. "drop table" deletes datafiles at commit time, and not during vacuum.
Supporting "begin; drop table mytable; commit; begin; set transaction
flashback 1 hour; select * from mytable; commit" would therefore be
really hard...
greetings, Florian Pflug
