Re: Gist fastbuild and performances

Ben:

On Fri, 8 Oct 2021 at 15:52, talk to ben <blo(dot)talkto(at)gmail(dot)com> wrote:

> I ran the same scipt on both sides at first.
> Then I switched to a COPY because the CREATE TABLE generates random data.

Logical, just pointing that you missed one side is generating, the
other is copying. Since I read a lot of things about selects updating
rows due to whatever bits affecting HOT and blah-blah-blah I suggested
doing the same script but with non-random data, hence the
generate+test-copy-13, test-copy-14 approach as an easy way to
eliminate potential discrepancies without having to look a lot of
dirty details.

> Since I got weird results, I wanted to be sure I had the same data on both versions.
> I ran the tests several times (and even asked a collegue to do it on his laptop) with the same results.

If you ran the test as written, you where comparing run times of
different sequence of operations. If you know the sequence of
operation does not matter all is fine and dandy. I do not, ( I suspect
is does not matter, but I do not know ) so I suggested to run the same
sequences ( as to me it seems to be easy & fast, given the timing you
gave and assuming no untimed operation was long ) and not worry about
differences.

> I totally agree that the query time are small and difficult to compare (but they are consistent across runs).
> I am just surprised that we have to access x17 pages for the same result on version 14. On bigger queries
> it could count. I just wanted to know if it's a know tradeoff of this new feature.

If you do not run the same sequences, you do not know. Note I do not
know what exact sequences you have tested, I write with only what I
have read as as input.

Francisco Olarte.