2:17 a.m.
Hi,
I could use some experienced guidance. I'm in the process of
evaluating Drools for use of using in a real-time transactional
environment to process about 3000 messages/second. I realize a lot of
this depends on the type and quantity of rules, hardware, etc. I'm
curious what steps others have taken to improve performance and if
there are any recommendations for my case detailed below.
A few specific questions I have are:
- Should each field in a message be a fact? (more info on my message
below) What fact granularity have you settled on in your usage and
why?
- Does the order of the conditions in a rule affect performance, the
execution order, or the structure of the Rete network?
- Does the order the facts are inserted into a stateless session (as
a list via the CommandFactory.newInsertElements) affect performance at
all?
The message is an EDI format and will typically have anywhere from 80
to 200 fields, potentially more. The message is divided into
transactions, then segments, then fields. We have an object model
that represents the message. We're using a stateless session and most
of the rules will modify fields or add fields and segments based on
the values of other fields. Currently, I flatten the object model
into a List containing the message, transactions, segments, and each
field, and then insert the list into the stateless session and fire.
I'm avoiding sequential mode for now until we have a better idea of
our requirements.
Here's a simplified example of what I'm doing now (using json-esque
syntax instead of the EDI format).
message {
segment {
SG:header
A0:agents
}
segment {
SG:agent
A1:000
A2:JAMES
A3:BOND
}
segment {
SG:agent
A1:86
A2:MAXWELL
A3:SMART
}
}
Each field has an id and a value. For this message, I would insert 14
facts: the message object, 3 segments objects, and 10 field objects.
rule "set James Bond's A1 to 007"
when
$a1 : Field(id == "A1", value != "007")
Field(id == "A2", value == "JAMES")
Field(id == "A3", value == "BOND")
then
$a1.setValue("007");
update($a1);
end
A relatively more complicated, but typical rule would be "set James
Bond's A1 to 007 iff he's the second agent in the message and one of
the other agents' first names does not contain AUSTIN and an agency
segment exists"
The above example assumes each segment and field is a fact and I think
it's a clean and flexible approach, but I'm concerned about the
overhead of inserting potentially 250 facts for each message. The
only other alternatives I can think of seem to have their own set of
problems:
1. limit the fields that rules can be written against to a limited
subset, which may not be feasible depending on how the requirements
evolve, and only assert those as facts. Doing this seems to double
the number of transactions per second in my nonscientific benchmark.
2. insert the Message object as a single fact and then write a slew of
accessor methods in that object to get at all possible fields in the
tree: getA1FromSecondAgentSegmentInFirstTransaction(). This seems
like it might perform well, but could be very messy.
3. provide an api in the message object model to find various
occurrences of fields in the messages, then use eval() in the rule.
like eval(msg.findSegment("agent", secondOccurrence).getField("A1")).
I've read that would be less efficient once the ruleset grows.
I'm sure many of you have dealt with this type scenario before, what
did you determine the best approach to be?
Thanks,
David
3:14 a.m.
David,
- Does the order of the conditions in a rule affect performance, the
execution order, or the structure of the Rete network?
Yes, quite a lot has been
written about this. The drools documentation
has a few paragraphs with good heuristics...
A popular read seems to be this paper, which provides good insight
into RETE performance aspects:
"Production Matching for Large Learning Systems (Rete/UL)"
by Robert B. Doorenbos
PhD thesis, Carnegie Mellon University, January 31, 1995
- Does the order the facts are inserted into a stateless session (as
a list via the CommandFactory.newInsertElements) affect performance
at
all?
All I can say that this is true (it has impact) for
statefull
sessions. The deal here is how often the agenda may get changed while
inserting objects.
I have no experience with stateless sessions and I hope others do;-)
- Should each field in a message be a fact? (more info on my message
below) What fact granularity have you settled on in your usage and
why?
I would suggest to start with a model which is most natural to your
rules (with the lowest WTF factor) and take it from there.
I assume that you need to correlate information from all segments of a
message?
I would stay well clear from your option 3 (eval()) - just because I
always do ;-)
Eval() simply defies RETE optimizations so it can be very sensitive to
both number of facts and number of rules (using eval).
If you know the field IDs (which I assume you would, using an EDI
formant) - I would see whether you can avoid the field facts and turn
them into slots (attributes) instead.
But not to the extend you outline in #2, that seems to be too awkward.
If you have buy another server ;-)
what did you determine the best approach to be?
I am pretty sure
that there is no generic "best approach".
I would proceed trying to understand the nature of the rules (are they
regular, ie following the same scheme, can they be generated?) and the
data and create
regression performance test-suite ASAP to monitor performance during
development.
In our application we for example have analyzed bottlenecks and found
that (no surprise) accessing properties like facta.factb.value is
slower compared to facta.facbtValue - that's in 4.07 - due to the
different mechanism the value is obtained. From there we have added
about three of these "shortcut accessors". However we have first
focused on optimizing the rules themself to avoid excessive matching
in the first place so there was no need for us to flatten our entire
model.
my 0.02EUR
-- Ingomar
Am 28.05.2009 um 02:17 schrieb David Zeigler:
Hi,
I could use some experienced guidance. I'm in the process of
evaluating Drools for use of using in a real-time transactional
environment to process about 3000 messages/second. I realize a lot of
this depends on the type and quantity of rules, hardware, etc. I'm
curious what steps others have taken to improve performance and if
there are any recommendations for my case detailed below.
A few specific questions I have are:
- Should each field in a message be a fact? (more info on my message
below) What fact granularity have you settled on in your usage and
why?
- Does the order of the conditions in a rule affect performance, the
execution order, or the structure of the Rete network?
- Does the order the facts are inserted into a stateless session (as
a list via the CommandFactory.newInsertElements) affect performance at
all?
The message is an EDI format and will typically have anywhere from 80
to 200 fields, potentially more. The message is divided into
transactions, then segments, then fields. We have an object model
that represents the message. We're using a stateless session and most
of the rules will modify fields or add fields and segments based on
the values of other fields. Currently, I flatten the object model
into a List containing the message, transactions, segments, and each
field, and then insert the list into the stateless session and fire.
I'm avoiding sequential mode for now until we have a better idea of
our requirements.
Here's a simplified example of what I'm doing now (using json-esque
syntax instead of the EDI format).
message {
segment {
SG:header
A0:agents
}
segment {
SG:agent
A1:000
A2:JAMES
A3:BOND
}
segment {
SG:agent
A1:86
A2:MAXWELL
A3:SMART
}
}
Each field has an id and a value. For this message, I would insert 14
facts: the message object, 3 segments objects, and 10 field objects.
rule "set James Bond's A1 to 007"
when
$a1 : Field(id == "A1", value != "007")
Field(id == "A2", value == "JAMES")
Field(id == "A3", value == "BOND")
then
$a1.setValue("007");
update($a1);
end
A relatively more complicated, but typical rule would be "set James
Bond's A1 to 007 iff he's the second agent in the message and one of
the other agents' first names does not contain AUSTIN and an agency
segment exists"
The above example assumes each segment and field is a fact and I think
it's a clean and flexible approach, but I'm concerned about the
overhead of inserting potentially 250 facts for each message. The
only other alternatives I can think of seem to have their own set of
problems:
1. limit the fields that rules can be written against to a limited
subset, which may not be feasible depending on how the requirements
evolve, and only assert those as facts. Doing this seems to double
the number of transactions per second in my nonscientific benchmark.
2. insert the Message object as a single fact and then write a slew of
accessor methods in that object to get at all possible fields in the
tree: getA1FromSecondAgentSegmentInFirstTransaction(). This seems
like it might perform well, but could be very messy.
3. provide an api in the message object model to find various
occurrences of fields in the messages, then use eval() in the rule.
like eval(msg.findSegment("agent",
secondOccurrence).getField("A1")).
I've read that would be less efficient once the ruleset grows.
I'm sure many of you have dealt with this type scenario before, what
did you determine the best approach to be?
Thanks,
David
_______________________________________________
rules-users mailing list
rules-users(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/rules-users
7:25 a.m.
I also think that the approach proposed by Greg of having a Message fact
type and various Segment fact types would let you write efficient and
maintainable rules. (The extreme splitting into Field facts has several
drawbacks, as has been pointed out, and also: 'value' is always a string,
so you lose type checking.) One thing I could not see is the effect of
the structural layer of transactions. And another one: are there several
(sub)types of Segment? If so, it will be possible of writing rules against
(abstract) base types, e.g.
abstract class Segment
class DebitSeg extends Segment
class CreditSeg extends Segment
rule notPositive
when
$s : Segment( amount <= 0 )
then
weird( $s );
Also, Segment subtypes might implement interfaces, and its
also possible to use them as fact names in patterns.
-W
On Thu, May 28, 2009 at 2:17 AM, David Zeigler <dzeigler(a)gmail.com> wrote:
Hi,
I could use some experienced guidance. I'm in the process of
evaluating Drools for use of using in a real-time transactional
environment to process about 3000 messages/second. I realize a lot of
this depends on the type and quantity of rules, hardware, etc. I'm
curious what steps others have taken to improve performance and if
there are any recommendations for my case detailed below.
A few specific questions I have are:
- Should each field in a message be a fact? (more info on my message
below) What fact granularity have you settled on in your usage and
why?
- Does the order of the conditions in a rule affect performance, the
execution order, or the structure of the Rete network?
- Does the order the facts are inserted into a stateless session (as
a list via the CommandFactory.newInsertElements) affect performance at
all?
The message is an EDI format and will typically have anywhere from 80
to 200 fields, potentially more. The message is divided into
transactions, then segments, then fields. We have an object model
that represents the message. We're using a stateless session and most
of the rules will modify fields or add fields and segments based on
the values of other fields. Currently, I flatten the object model
into a List containing the message, transactions, segments, and each
field, and then insert the list into the stateless session and fire.
I'm avoiding sequential mode for now until we have a better idea of
our requirements.
Here's a simplified example of what I'm doing now (using json-esque
syntax instead of the EDI format).
message {
segment {
SG:header
A0:agents
}
segment {
SG:agent
A1:000
A2:JAMES
A3:BOND
}
segment {
SG:agent
A1:86
A2:MAXWELL
A3:SMART
}
}
Each field has an id and a value. For this message, I would insert 14
facts: the message object, 3 segments objects, and 10 field objects.
rule "set James Bond's A1 to 007"
when
$a1 : Field(id == "A1", value != "007")
Field(id == "A2", value == "JAMES")
Field(id == "A3", value == "BOND")
then
$a1.setValue("007");
update($a1);
end
A relatively more complicated, but typical rule would be "set James
Bond's A1 to 007 iff he's the second agent in the message and one of
the other agents' first names does not contain AUSTIN and an agency
segment exists"
The above example assumes each segment and field is a fact and I think
it's a clean and flexible approach, but I'm concerned about the
overhead of inserting potentially 250 facts for each message. The
only other alternatives I can think of seem to have their own set of
problems:
1. limit the fields that rules can be written against to a limited
subset, which may not be feasible depending on how the requirements
evolve, and only assert those as facts. Doing this seems to double
the number of transactions per second in my nonscientific benchmark.
2. insert the Message object as a single fact and then write a slew of
accessor methods in that object to get at all possible fields in the
tree: getA1FromSecondAgentSegmentInFirstTransaction(). This seems
like it might perform well, but could be very messy.
3. provide an api in the message object model to find various
occurrences of fields in the messages, then use eval() in the rule.
like eval(msg.findSegment("agent",
secondOccurrence).getField("A1")).
I've read that would be less efficient once the ruleset grows.
I'm sure many of you have dealt with this type scenario before, what
did you determine the best approach to be?
Thanks,
David
_______________________________________________
rules-users mailing list
rules-users(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/rules-users
5453
days inactive
5453
days old
2 comments
3 participants
participants (3)
-
David Zeigler -
Ingomar Otter -
Wolfgang Laun