[rules-users] Drools Expert scalability

[Dieter D'haeyere]

Well, the batch file wouldn't be delivered as a whole to the rule engine,
the app would parse the file and send cases one at a time to the rule
engine(s).

To dispatch to the rule engine instances, one would need to know what cases
are handled by the running engine(s) before one can decide to which engine
the new case can be sent
The criterion on actor id was just a sample, it could change over time ...
so that's why I think it is hard to say that scalability is done by adding
more server instances.

Load (for one application several are using the rule engine): +/- 100K
requests per hour.

Of course, if the engine is fast enough to evaluate everything in a single
instance, there is no problem at all :)
buy some faster hardware ? :)




2010/8/12 Wolfgang Laun <wolfgang.laun at gmail.com>

> Where is the limit for mutual case dependency? OK, you can't split a batch,
> but will the "multiple batch files per day" be independent of each other?
>
> If "actor" is the only criterion for dependencies, then you could split
> cases to different rule engines (on different CPUs) by actor id (odd-even)
> or some similar thing. Preprocessing to determine the server target could
> also be done by rules :-)
>
> I'd rather not try and go for engine pipelining. Transferring data from
> one stage to the next causes overhead, not only for the transport but also
> for WME bookkeeping and re-evaluation of patterns that are the same
> in stage N and N+1, and which could be factored into a single Rete.
>
> How many cases per day are there? How much delay is acceptable
> between starting a batch and obtaining the results? The 50k alone
> aren't sufficient to say that there is "an issue".
>
> -W
>
> 2010/8/12 Dieter D'haeyere <dieter.dhaeyere at gmail.com>
>
>> I have a question regarding the scalability of Drools Expert.
>>
>> Situation is as follows:
>> - Individual cases have to be evaluated by the rule engine (result is a
>> tarrif that could be refunded).
>> - the application receives the cases in flat files with 50K cases each.
>> The app parses them and sends them one by one to the rule engine
>> - the application receives multiple batch files per day.
>> - seperate cases can be dependant on each other.  Eg. if max 1 refund can
>> be done to an actor, two cases about the same actor can not run in parallel.
>> - loads will be huge, scalability is an issue
>>
>> So, what I see now
>> - Drools Server can be run as a server
>> - It is possible to have multiple instances of drools server, this would
>> allow to evaluate cases in parallel.  But ... this can cause problems (as
>> stated before): you can't run any two cases in parallel.  Preprocessing
>> could be done by the application (eg. determining the order in which to
>> present the cases to the rule engine) but over time extra constraints can
>> appear so preprocessing would have to be maintained continuously.
>>
>> So, this is an issue :)
>>
>> My questions are:
>> - is it eg possible to organise rules such that different rules can run on
>> different rule engines.  Like that you could have some kind of pipeline.
>> Maybe this could be defined with ruleflow ? Compare to the pipelines in
>> CPU's (for executing machine code).  Could you eg. relate one server
>> instance to one set of rules and another server instance to the sequential
>> next set of rules ?  How would you configure that (performance is of course
>> important).
>> - Or, is it possible to split rules but not as a pipeline but in
>> parallel... like a 'fork .. join' .  Again, this could be configured with
>> ruleflow?
>> - What do you see as the best way to solve this issue ?
>>
>>
>> Any help is welcome,
>> Dieter D'haeyere.
>>
>>
>> _______________________________________________
>> rules-users mailing list
>> rules-users at lists.jboss.org
>> https://lists.jboss.org/mailman/listinfo/rules-users
>>
>>
>
> _______________________________________________
> rules-users mailing list
> rules-users at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/rules-users
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/rules-users/attachments/20100812/38df3547/attachment.html