On 10/12/2004 4:02 PM Tom Lane could be overheard saying::
>Marcos A Vaz Salles <msalles@gmail.com> writes:
>
>
>>In order to make index selection possible, we
>>have extended the PostgreSQL DBMS to allow the simulation of
>>hypothetical indexes. We believe these server extensions may be of
>>value for addition to the PostgreSQL code base.
>>
>>
>
>This would be of some value if the optimizer's cost estimates were
>highly reliable, but unfortunately they are far from being so :-(
>Without the ability to measure *actual* as opposed to estimated
>costs, I'm not sure you can really do much.
>
>
Is it possible for the backend to store performance data and try to
modify its cost estimates? I was thinking of the statistics analyzer
currently in use and whether (query) performance data could piggy back
on it or if it would need a different process, something that could try
and correlate cost estimates with actual costs. Given sample data the
calculation could result in an actual execution time estimation. Would
it be worth pursuing, or would it be too invasive?
I know the autotune project has similar goals from a different angle,
system/memory settings.
Thomas
