Lukas Smith <mls@pooteeweet.org> writes:
> Bruce Momjian wrote:
>
>> * Flush cached query plans when the dependent objects change,
>> when the cardinality of parameters changes dramatically, or
>> when new ANALYZE statistics are available
>
> Wouldn't it also make sense to flush a cached query plan when after
> execution it is determined that one or more assumptions that the
> cached query plan was based on was found to be off? Like the query
> plan was based on the assumption that a particular table would only
> return a hand full of rows, but in reality it returned a few
> thousand.
There is some merit to that.
I could also see it being sensible to flush a cached plan any time the
query took more than some [arbitrary/GUC-chosen] interval.
Supposing it took 20s to execute the query, it would surely seem
surprising for re-evaluating the plan to be expected to make up a
material proportion of the cost of the *next* invocation.
If we flush every plan that took >10s to evaluate, that offers the
possibility for it to be done better next time...
--
output = reverse("gro.mca" "@" "enworbbc")
http://www.ntlug.org/~cbbrowne/internet.html
Points are awarded for getting the last word in. Drawing the
conversation out so long that the original message disappears due to
being indented off the right hand edge of the screen is one way to do
this. Another is to imply that anyone replying further is a hopeless
cretin and is wasting everyone's valuable time.
-- from the Symbolics Guidelines for Sending Mail