Re: GSoC idea - Simulated annealing to search for query plans

>
> I think someone already mentioned it, but it would be very neat if the
> optimizer could be pluggable. Then many different algorithms could be
> evaluated more easily.

Does it mean just to make the join order optimizer pluggable? If yes then
it is already pluggable as an extension. Is this the desired approach so
far?
I'm thinking of testing and improving SAIO as an extension before reaching
a satisfactory quality of code and returned plans.

Would then the destination be the /contrib and then main source tree or
would it ever stay as an extension?

Best regards,
Grzegorz

2015-02-27 15:03 GMT+01:00 ktm(at)rice(dot)edu <ktm(at)rice(dot)edu>:

> On Thu, Feb 26, 2015 at 10:59:50PM +0100, Grzegorz Parka wrote:
> > Dear Hackers,
> >
> > I'm Grzegorz Parka, BSc Engineer of Technical Physics and student of
> > Computer Science at WUT, Poland. Last year I've been a bit into
> > evolutionary algorithms and during my research I found out about GEQO in
> > Postgres. I also found out that there are plans to try a different
> attempt
> > to find optimal query plans and thought it could be a good thing to use
> it
> > as a project for GSoC.
> >
> > I'm interested in one of old TODO items related to the optimizer -
> > 'Consider compressed annealing to search for query plans'. I believe this
> > would be potentially beneficial to Postgres to check if such a heuristic
> > could really choose better query plans than GEQO does. Judging by the
> > results that simulated annealing gives on the travelling salesman
> problem,
> > it looks like a simpler and potentially more effective way of
> combinatorial
> > optimization.
> >
> > As deliverables of such a project I would provide a simple implementation
> > of basic simulated annealing optimizer and some form of quantitative
> > comparison with GEQO.
> >
> > I see that this may be considerably bigger than most of GSoC projects,
> but
> > I would like to know your opinion. Do you think that this would be
> > beneficial enough to make a proper GSoC project? I would also like to
> know
> > if you have any additional ideas about this project.
> >
> > Best regards,
> > Grzegorz Parka
>
> Hi,
>
> I think someone already mentioned it, but it would be very neat if the
> optimizer could be pluggable. Then many different algorithms could be
> evaluated more easily.
>
> Regards,
> Ken
>