[rules-users] determinism with rulebase partitioning

[Edson Tirelli]

    I think that even with rulebase partitions, we should continue to
support current execution mode. So, we should keep a rulebase configuration
that basically allow the user to defines: either single-thread (as it is
today) or multi-thread (as we are trying to achieve) execution.

    Having that in mind, in the multi-thread mode:

(A) What does "parallel evaluation of a rulebase" mean? Is it designed
to optimise, for example, two threads processing a stateless and
stateful session?
   Means that rules that do not share nodes, being independent of each other
(from an evaluation perspective), will be evaluated in parallel. This is
very common scenario and a desidered feature in CEP engines.

(B) Are there only two partitions, both of which are invisible to the
user? Is there any value in allowing user-defined partitions?
    There will be as many partitions as the compiler can create for the
given set of rules. Rules that share more nodes, are more difficult to
partition, while rules that are independent from an LHS point of view, are
easier to parallelize. In my opinion, the only thing that may be helpful to
expose and allow the user to control is the maximum size of the thread pool
that is used to propagate facts. Even that I'm not sure is so helpful,
because it is complex to fine tune such things, since the partitioning is
completely dependent on the rules added to the rulebase.

(C) Does the partition used depend upon what type of session is used
(i.e. stateless always uses the partition without an agenda whereas
stateful always uses the partition with an agenda)?
    The partitioning of the rulebase is dependent upon the rules in the
rulebase and nothing more. But that is different from the agenda. The agenda
issue is much more complex, because even with partitions we can keep a
single deterministic agenda (as long as it is not in active mode -
runUntilHalt). Now, if the agenda is in active mode, or if we have multiple
agendas (1 per partition, for instance), then the engine behavior becomes
indeterministic. This is a common scenario in CEP systems that have multiple
different "queries" running over the same set of streams, trying to detect
and act upon them as soon as they are detected, and event streams are
indeterministic by their own nature. In common rules engines scenarios, I'm
not sure we can run in this indeterministic mode.

(D) Can a rule sometimes be deterministic and sometimes not (i.e.
depends upon the type of session)?
    It will always depend on the set of rules (the rulebase), not the type
of session. One rule is always deterministic when considered in isolation,
but two or more rules may or may not be deterministic in relation to each
other. Just remember Eisten's Relativity Theory... ;)

   []s
   Edson

2008/8/1 Anstis, Michael (M.) <manstis1 at ford.com>

> Hi Mark,
>
> A few questions:-
>
> (A) What does "parallel evaluation of a rulebase" mean? Is it designed
> to optimise, for example, two threads processing a stateless and
> stateful session?
>
> (B) Are there only two partitions, both of which are invisible to the
> user? Is there any value in allowing user-defined partitions?
>
> (C) Does the partition used depend upon what type of session is used
> (i.e. stateless always uses the partition without an agenda whereas
> stateful always uses the partition with an agenda)?
>
> (D) Can a rule sometimes be deterministic and sometimes not (i.e.
> depends upon the type of session)?
>
> Cheers,
>
> Mike
>
> -----Original Message-----
> From: rules-users-bounces at lists.jboss.org
> [mailto:rules-users-bounces at lists.jboss.org] On Behalf Of Mark Proctor
> Sent: 01 August 2008 07:05
> To: Rules Users List
> Subject: [rules-users] determinism with rulebase partitioning
>
> We have rulebase partitioning almost working, this allows parallel
> evaluation of a rulebase. For stateless lessions with no agenda this
> will allow for much faster executions, where you don't care about
> deterministic execution. However for deterministic execution its more
> complicated. The current plan is to have an agenda per parition, which
> means that we no longer have rulebase wide deterministic execution
> order, only with the partition itself. The user is unlikely to be aware
> of the created partitions, so won't be aware of the unditermistic
> behavour of their rulebase. Anyone have any input on mechanisms users
> can do to help the rulebase know what needs to be executed
> deterministically and what doesn't?
>
> Mark
> _______________________________________________
> 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
>



-- 
Edson Tirelli
JBoss Drools Core Development
JBoss, a division of Red Hat @ www.jboss.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/rules-users/attachments/20080801/b395e6f0/attachment.html