Re: Expressions greedy advanceExact implementation

[Greg Miller <gs...@gmail.com>]

Reviving this thread with another thought…

I think we can improve on this last solution and lazily advance an
expression’s referenced double values without needing to push complexity
down into compiled expression.

What if we do something like this?
https://github.com/apache/lucene/pull/12560

Cheers,
-Greg



On Wed, Oct 26, 2022 at 11:06 Michael Sokolov <ms...@gmail.com> wrote:

> Thanks, yeah I thought so too. Merged
>
> On Wed, Oct 26, 2022 at 10:31 AM Robert Muir <rc...@gmail.com> wrote:
> >
> > I think deferring the advance call like this is fine and harmless,
> > only because this DoubleValues "caches" the result for the current
> > doc, so its idempotent anyway.
> >
> > Yes, about "advancing all the operands" as I mentioned, expressions
> > has no clue about this. If you wanted to change it, you'd have to push
> > both advancing AND caching down lower into the actual compiled
> > expression code.
> >
> > I think this would add way too much complexity, especially when it
> > would only improve the ternary "if" feature in such cases.
> >
> > On Wed, Oct 26, 2022 at 10:23 AM Michael Sokolov <ms...@gmail.com>
> wrote:
> > >
> > > see https://github.com/apache/lucene/pull/11878 ... it doesn't do what
> > > I initially asked for (still advances all of the operands), but it
> > > delays until doubleValue() is called, which is safe and could have
> > > some impact
> > >
> > > On Wed, Oct 26, 2022 at 9:58 AM Michael Sokolov <ms...@gmail.com>
> wrote:
> > > >
> > > > Hi, yes, makes sense Mikhail, that will address most of the problem.
> > > > But I also think, given the way Expressions work (they always return
> > > > true from advanceExact) there is no reason for them to advance their
> > > > operands. This shifts the burden/concern from the developer who no
> > > > longer has to think as hard about this :)  - let me post a PR that
> > > > shows
> > > >
> > > > On Wed, Oct 26, 2022 at 3:52 AM Mikhail Khludnev <mk...@apache.org>
> wrote:
> > > > >
> > > > > Hello, Michael.
> > > > > I suppose you can bind f2 to custom lazy implementation of
> DoubleValuesSource, which defer advanceExact() by storing doc num and
> returning true always, and actually advancing on doubleValue() only.
> > > > >
> > > > > On Tue, Oct 25, 2022 at 8:13 PM Michael Sokolov <
> msokolov@gmail.com> wrote:
> > > > >>
> > > > >> ExpressionFunctionValueSource lazily evaluates in doubleValues: an
> > > > >> expression like
> > > > >>
> > > > >>    condition ? f1 : f2
> > > > >>
> > > > >> will only evaluate one of f1 or f2.
> > > > >>
> > > > >> At the same time, the advanceExact() call is greedy -- when you
> > > > >> advance that expression it will also advance both f1 and f2. But
> > > > >> here's the thing: it always returns true, regardless of whether
> f1 and
> > > > >> f2 advance. Which makes sense from the point of view of the lazy
> > > > >> evaluation -- if condition is true we don't care whether f2
> advances
> > > > >> or not.
> > > > >>
> > > > >> My question is whether we could defer these child advanceExact
> calls
> > > > >> until ExpressionFunctionValues.doubleValue()?
> > > > >>
> > > > >>
> ---------------------------------------------------------------------
> > > > >> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> > > > >> For additional commands, e-mail: dev-help@lucene.apache.org
> > > > >>
> > > > >
> > > > >
> > > > > --
> > > > > Sincerely yours
> > > > > Mikhail Khludnev
> > >
> > > ---------------------------------------------------------------------
> > > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> > > For additional commands, e-mail: dev-help@lucene.apache.org
> > >
> >
> > ---------------------------------------------------------------------
> > To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> > For additional commands, e-mail: dev-help@lucene.apache.org
> >
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: dev-help@lucene.apache.org
>
>