Re: [Exist-development] optimizing positional predicates for fast paging

[Michael S. <so...@if...>]

Thanks for your answer, Wolfgang - that's just the sort of thing I was
wondering about.

I walked through the logic in selectFollowing: I can see it would be
possible, but complicated to implement something similar for other axes and
for a somewhat more complex predicate (compare position against a constant
sequence).  Maybe that's worth doing - I'm not sure though.

Frankly the main cases I am interested in are not really axis selections,
but ordered sequences of nodes from different documents where the ordering
is determined either by a value (that may be indexed) or by the lucene
score.  These correspond to searches in our applications where the results
are ordered alphabetically by title, or by relevance.  So I asked the other
question mostly just to try and understand better what kind of optimization
strategies you are using.

I'll keep looking and let you know if I have anything helpful to offer!

I'm also curious about the answer to Joe's question re: the difference
between subsequence and predicate: if you have a moment could you elucidate?

-Mike

Also: my e-mails to this list seem to need moderation.  If you'd rather not
be bothered about that, I could move this discussion over to the exist-open
list?

> -----Original Message-----
> From: Wolfgang Meier [mailto:wol...@ex...] 
> Sent: Monday, December 14, 2009 3:37 AM
> To: Michael Sokolov
> Cc: exi...@li...
> Subject: Re: [Exist-development] optimizing positional 
> predicates for fast paging
> 
> > I wonder whether eXist is currently optimizing xquery 
> evaluations that 
> > use positional predicates, or similar.
> 
> I did implement shortcuts for the expensive axes: following 
> and preceding. For example, LocationStep.getFollowing checks 
> the property hasPositionalPredicate. If set to true, the 
> method evaluates the predicate in advance to get its numeric 
> value. This is then forwarded to the NodeSet method which 
> calculates the join:
> 
> return currentSet.selectFollowing(contextSet, position, contextId);
> 
> We could apply similar optimizations to other axes, though 
> their execution paths are more complex as you have more 
> possibilities to choose from.
> 
> However, my positional predicate optimization would only work 
> for //elem[1], not //elem[position() = 1]. But I guess it 
> would be possible to add a rewrite rule here to automatically 
> simplify the expression.
> 
> > collection("/")/ (//elem) [position() = (1 to 10)]
> 
> Note that using subsequence instead of the predicate can be a 
> lot faster sometimes.
> 
> Wolfgang
>