[rules-users] @Rete Experts: the finer meshes of matching

[Mark Proctor]

On 08/04/2010 19:56, Wolfgang Laun wrote:
> It's common Rete wisdom to order patterns by expected frequency of
> matches, in descending order.
>
> Does this also apply to constraints within a single pattern? For
> example, is it better to write, for a New York address
>     Address( street == "Broadway", number == 1)
> than
>     Address( number == 1, street == "Broadway")
> or is this irrelevant?
>    
Drools will iterate to find the first == constraints, both literal and 
vairable constraints. So if != or < are first, that doesn't matter. For 
multiple literal constraints, should definitely have the strongest 
descriminator first - although it's not quite that simple. You also want 
the descriminator that provides the highest chances of node share 
amoung. == literal constraints are hashed so the wider the want out 
(alternative values) the more scalable performance you'll see. Drools 
actually attempts to hash the fanout of consequence child nodes in the 
alpha network, if they to are ==, but rarely is there any real gain from 
this as after the first level of descrimination there is general not a 
consequence wide fanout based on the previous constraints value.

There are deterministic algorithms that claim to analyse a network to 
re-order literal constraints to maximise node sharing, but that's not 
something we have, patches please  :)

If it's a variable constraint we support composite indexing. Drools 
produces a composite index based on the first 3 ==  constraints. Whether 
the order of those first 3 impacts perf, not sure. But if there was a 
4th variable constraint and that was the strongest descriminator, it 
would get left out of the index; when in reality it's probably the most 
important constraint as part of the index.

Mark
> Or does it depend on either constraint (on each own) also occurring in
> another rule without the other one?
>
> I guess this could also depend on the implementation, i.e., how
> fine-grainedly condition sharing is being done.
>
> Cheers
> Wolfgang
> _______________________________________________
> rules-users mailing list
> rules-users at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/rules-users
>
>
>