Hi Stamatis,

Thanks a lot for your reply. Yes, it seems the traits currently in Calcite
are used by the optimizer. I wonder whether we can extend it for other
use-cases. For example, I want to provide a way to the users that they can
set memory or cpu settings for an aggregate node from the user api. These
settings will only be used after optimization.

I haven't found other ways to achieve this, so maybe using trait is a neat
way?

Best, Hequn


On Fri, Jan 18, 2019 at 12:00 AM Stamatis Zampetakis <[email protected]>
wrote:

> Hi Hequn,
>
> I would describe traits as properties associated with RelNodes that provide
> useful information to the optimizer (rules etc.) in order to generate a
> plan.
>
> If the configuration you are referring to is meant to guide the optimizer
> in generating a plan then it seems ok to use traits. If not then probably
> you need something different.
>
> Can you elaborate a bit more on your usecase?
>
> Best,
> Stamatis
>
>
> On Tue, Jan 15, 2019, 10:57 AM Hequn Cheng <[email protected] wrote:
>
> > Hi Julian,
> >
> > Thanks a lot for your reply and the detailed explanation. It solves my
> > doubts well.
> > My custom trait only contains one value, so I think that there will not
> be
> > a problem.
> >
> > May I further the email with another question:
> > Is it ok or right to use a trait to pass configurations through RelNodes?
> > For example, a configuration set from api for the aggregate and used
> after
> > the optimization.
> > If not, are there any standard ways to achieve this?
> >
> > I haven't found any clear definition about trait. Only find comments in
> > code: *RelTrait represents the manifestation of a relational expression
> > trait within a trait definition.*
> >
> > Thank you!
> >
> > On Tue, Jan 15, 2019 at 3:12 AM Julian Hyde <[email protected]> wrote:
> >
> > > In most cases increasing the number of traits from one to two will
> > > increase the planning time by a negligible amount.
> > >
> > > But it can increase the size of the search space. Suppose a particular
> > > relational expression has 5 possible sort orders (order by x, order by
> x,
> > > y, order by (), order by z, order by x, z), and initially you have only
> > the
> > > collation trait enabled. A particular equivalence set might have 5
> > subsets,
> > > one for each sort order. Now let’s suppose you add the distribution
> trait
> > > to the mix, and there are 3 distributions (partition by (), partition
> by
> > x,
> > > partition by z). Now that subset will have 15 subsets, for the
> cartesian
> > > product of the traits.
> > >
> > > A larger search space could increase the planning time (and memory
> usage)
> > > significantly.
> > >
> > > But if each trait has only one or two values I doubt that there will
> be a
> > > problem.
> > >
> > > Julian
> > >
> > >
> > > > On Jan 14, 2019, at 1:38 AM, Hequn Cheng <[email protected]>
> wrote:
> > > >
> > > > Hi everyone,
> > > >
> > > > I want to pass properties through RelNodes via trait and I wonder if
> > the
> > > > number of traits in traitSet will affect the time of Volcano
> > > optimization.
> > > > For example, increasing one traitDef to two in VolcanoPlanner.
> > > >
> > > > I guess the answer is No. Is it correct?
> > > > As long as the search space is not increased, the Volcano
> optimization
> > > time
> > > > will not increase. And simply increasing the number of traits alone
> > does
> > > > not add complexity.
> > > >
> > > > Furthermore, besides time, are there any other side effects if I
> > increase
> > > > the number of traits?
> > > >
> > > > Thank you very much!
> > > >
> > > > Best,
> > > > Hequn
> > >
> > >
> >
>

Reply via email to