Re: Gsoc2012 idea, tablesample

"Kevin Grittner" <Kevin.Grittner@wicourts.gov> writes:
> Florian Pflug <fgp@phlo.org> wrote:
>> Maybe one can get rid of these sorts of problems by factoring in
>> the expected density of the table beforehand and simply accepting
>> that the results will be inaccurate if the statistics are
>> outdated?
> Unless I'm missing something, I think that works for percentage
> selection, which is what the standard talks about, without any need
> to iterate through addition samples.  Good idea!  We don't need to
> do any second pass to pare down initial results, either.  This
> greatly simplifies coding while providing exactly what the standard
> requires.
>> I'm not totally sure whether this approach is sensible to
>> non-uniformity in the tuple to line-pointer assignment, though.

If you're willing to accept that the quality of the results depends on
having up-to-date stats, then I'd suggest (1) use the planner's existing
technology to estimate the number of rows in the table; (2) multiply
by sampling factor you want to get a desired number of sample rows;
(3) use ANALYZE's existing technology to acquire that many sample rows.
While the ANALYZE code isn't perfect with respect to the problem of
nonuniform TID density, it certainly will be a lot better than
pretending that that problem doesn't exist.
        regards, tom lane