Alex Ignatov <a.ignatov@postgrespro.ru> writes:
> You always should keep in mind that your application may run in test
> mode (future/past time) and maintain this code. While with my proposal
> you can always use some time function(now or localtimestamp or
> whatever) which you can freeze at anytime on DB level, not operation
> system(using some 3rd libs) or application(using wrappers and other hacks).
We're not really in the business of being Oracle --- which in this
particular context means not trying to duplicate tens of thousands of
bizarre little features with very narrow use-cases. If there's a
reasonable way for users to provide corner-case functionality for
themselves (and I'd say a wrapper function is a perfectly reasonable
way for this) then we don't really want to embed it in Postgres.
This particular feature seems like a seriously-poorly-thought-out
one, too. Time stops advancing across the whole DB? Really?
1. That would break all manner of stuff, for example the timestamps
in automatically-taken dumps, if you've got background jobs running
pg_dump. Just about everything except the session running the test
case would be unhappy, AFAICS.
2. Would this extend to, say, preventing autovacuum from running?
Or changing the timestamps of messages in the postmaster log, or
timestamps appearing in places like pg_stat_activity? Or causing
pg_sleep() to wait forever, because time isn't passing? If your
answer is "yes" across the board, that makes problem #1 an order
of magnitude worse, while if you want to be selective then you
have a bunch of nitty-gritty (and rather arbitrary) decisions to
make about what's frozen and what's not. And you've weakened the
argument that your test is actually valid, since potentially the
app would see some of the non-frozen values and misbehave.
3. While I can see the point of wanting to, say, test weekend behavior
on a weekday, I do not see how a value of now() that doesn't advance
between transactions would represent a realistic test environment for
an app with time-dependent behavior. As an example, you might
accidentally write code that expects two successive transactions to
see identical values of now(), and such a testbed wouldn't detect
the problem.
regards, tom lane