Re: [boost] RTL

From: Calum Grant (calum_at_[hidden])
Date: 2005-10-21 12:41:18

 Arkadiy Vertleyb wrote:
> "Calum Grant" <calum_at_[hidden]> wrote
> > I frankly have no idea how I would implement that problem 
> in SQL, but 
> > then this was a problem you chose to be very difficult to 
> solve using 
> > a SQL notation.
> 
> I am not an SQL expert, but should be something like this (if 
> I remember
> correctly):
> 
> SELECT number, name, city, COUNT(*) AS cnt
> FROM mytable a, mytable b
> WHERE [the long condition based on coordinates]
> GROUP BY a.number
> ORDER BY cnt DESC;
> 
> Then you hope that the optimizer figures out how to not 
> compare every possible pair ;-)

Needless to say, implementing a query analyser to do all that purely in
TMP would be rather difficult!  But I don't believe there are any hidden
costs in RML - what you see is what you get.

> This exact statement I modeled in RTL, but RTL also specifies 
> HOW to do things in addition to WHAT to do, so I was able to 
> do a smartrer
> (range-based) query.

Ah, but you can specify "how" using C++.

> I am thrilled with your performance results, though (although 
> I consider them the results of a by-hand solution).  Did you 
> completely avoid calculating the distance?  If so, I think 
> any such solution would provide only approximate result 
> (although it might happen to be a pretty good approximation).

I compare on Euclidean distance squared.  It does not actually matter
whether Euclidean distance, Euclidean distance squared, or surface
distance is used, the ordering will be preserved.

Best regards, Calum