[rules-dev] Lazily Enabled Truth Maintenace (Project Propsoal)
Mark Proctor
mproctor at codehaus.org
Fri Sep 24 18:28:19 EDT 2010
On 23/09/2010 13:44, Leonardo Gomes wrote:
> Hi,
>
> I implemented the test cases you suggested, Mark, and started working
> on the counter. I discussed a bit with Edson yesterday and he
> clarified the difference between the counter for logically inserted
> facts and the fact id counter. This part seems to be fine, but I have
> a problem with the following scenario:
>
> After an initial activation of TMS I put everything in the equality
> map, then, due to retractions, the counter gets back to zero and I
> have to disable TMS for that specific type. When a new fact is
> logically inserted I have to look at the OTN memory and get everything
> that was inserted there since the last deactivation of TMS, to be be
> able to re-enable it. To do so I use the fact handle id that I'm
> supposed to have stored together with the counter. Fine.
>
> *Problem: *Since the OTN memory is an ObjectHashSet, I will have to
> iterate through it and filter (there's an ObjectFilter interface that
> I think I can implement) all the elements that have id greater than
> the one I stored. This will probably be inefficient and still likely
> to cause churn.
>
> *Possible solutions:*
>
> 1) Use a TreeMap? ** Don't think so, HashSet retrieval and insertion
> is O(1), on average, whereas TreeMap is O(log(n));
>
> 2) Use a LinkedHashSet (probably a homegrown implementation on top of
> ObjectHashSet)? ** Would allow me to get the ObjectEntry I want and
> then call getNext() up to the end, inserting all of them in the
> equality map;
>
> 3) Never disable TMS for a given type once it has been enabled? ** I
> think that lazily enabling TMS per type might be already enough an
> optimization, in a sense that the overhead with the calculation of the
> delta and all this logic to disable/re-enable might not be worthy.
> Moreover, switching from HashSet to LinkedHashSet may have some
> disadvantages.
Lets get it working with just lazy enabling to begin with, so for now it
can't be disabled once enabled. Once that is working we can look into
how we can make it disable and decide if it's worth doing or not.
Mark
>
> What do you think?
>
> Cheers,
> Leo.
>
>
> On Tue, Sep 21, 2010 at 12:09 AM, Mark Proctor <mproctor at codehaus.org
> <mailto:mproctor at codehaus.org>> wrote:
>
> On 20/09/2010 21:58, Leonardo Gomes wrote:
>> Ok, it took me much longer that I had thought, but I think it's
>> working now.. all tests passing except for
>> ReteooWorkingMemoryTest that does:
>>
>> FactHandle fd = workingMemory.insert( string );
>>
>> assertEquals( 1,
>> tms.getAssertMap().size() );
>>
>>
>> Which is a good sign, since tms is now empty as long as you
>> haven't done any logical insert :)
>>
>> I implemented just the first part of your suggestion, with a
>> boolean. I will try to implement the solution with the counter
>> later this week and maybe do some adjustments before submitting a
>> patch.
>>
>> Do you have any suggestion in terms of unit tests that you would
>> like to see or just the apparent lack of regressions is enough?
>> Well, I can also think of some tests by taking
>> ReteooWorkingMemoryTest as an example, in case you don't have
>> anything particular in mind.
> When you do some insertions cast to get some of the internal data
> structures and check the equality map is null, maybe check the OT
> confs are false, do a logical insertion check it can turn on for
> specific OTNs and not all and that it correctly adds all objects
> for the OTN. This will be important for the next stage of the work.
>
> Mark
>
>>
>> Cheers,
>> Leo.
>>
>>
>>
>>
>> On Mon, Sep 20, 2010 at 5:14 PM, Mark Proctor
>> <mproctor at codehaus.org <mailto:mproctor at codehaus.org>> wrote:
>>
>> On 20/09/2010 15:53, Leonardo Gomes wrote:
>>> I think I wasn't quite clear in my last email, so let me try
>>> to reformulate it:
>>>
>>> I also gave it a try to try to do what you suggested here
>>> (http://blog.athico.com/2010/09/lazily-enabled-truth-maintenace.html),
>>> Mark, and couldn't make it work due to the following situation:
>>>
>>> 1) Current code seems to rely on the equality map to know
>>> that a logical insert (insertLogical) for an object that has
>>> already been regularly inserted (insert) should be ignored;
>> You would need to lazily maintain an equality map. When the
>> first logical insertion is done we will have to first
>> populate that map from the Object Type Node set of FactHandles.
>> In the "insert" one of the first lines is:
>> ObjectTypeConf typeConf =
>> this.typeConfReg.getObjectTypeConf( this.entryPoint,
>>
>> object );
>>
>> So we get the ObjectTypeConf before we do anything with the
>> object itself, and we can check if TMS is being maintained
>> for that Object Type.
>>
>>>
>>> 2) If I apply the modifications that you suggested, from
>>> what I understood, things would start to be put in the
>>> equality map only when a logical insert is issued;
>>>
>>> *Problem*: How would I handle the situation described in
>>> item 1, if I don't have anything in the equality map at the
>>> moment a logical insert comes in and I have to "lazily
>>> activate" TMS?
>> final Rete source = this.ruleBase.getRete();
>> ClassObjectType cot = new ClassObjectType( MyClass.class );
>> Map<ObjectType, ObjectTypeNode> map =
>> source.getObjectTypeNodes( EntryPoint.DEFAULT );
>> ObjectTypeNode node = map.get( cot );
>> final ObjectHashSet memory = (ObjectHashSet)
>> workingMemory.getNodeMemory( node );
>>
>> That "memory" is the set of currently asserted objects for
>> that OTN. You can now iterate that and populate the equality
>> hash map. This is a one off as the flag will be set on the
>> ObjectTypeConf from then on and all objects will be added to
>> the equality map at the point of insertion.
>>
>> Does that help? All the information is there, you should have
>> to know how to retrieve it :) We don't currently maintain TMS
>> for anything other than the default entry point. Although I
>> think that was a mistake and we will probably move all
>> entrypoints to work with same, but for now you can ignore
>> that and just focus on the default.
>>
>> Btw this should not be confused with "equality" mode where
>> the equality map has to be maintained by default from the
>> start. Some would argue that a rule engine should only under
>> work and understand on the bases of equality and thus all
>> users must implement hashcode and equals correctly...
>>>
>>> --
>>>
>>> Moreover, I'm willing to attempt to implement the left and
>>> right un-linking and tried to start with this easier task to
>>> start to get familiar with drools-core. I already read the
>>> article you linked and your article. Would you have any
>>> document with an overview of the way drools implements rete?
>> not really no, Rete has already well documented in a number
>> of papers. The best thing to do is get onto irc and talk
>> directly to edson and I and we can walk you through classes.
>> http://www.jboss.org/drools/irc.html
>>
>> Mark
>>
>>>
>>> Thanks in advance!
>>>
>>> Cheers,
>>> Leo.
>>>
>>>
>>>
>>>
>>>
>>>
>>> On Sat, Sep 18, 2010 at 3:49 AM, Leonardo Gomes
>>> <leonardo.f.gomes at gmail.com
>>> <mailto:leonardo.f.gomes at gmail.com>> wrote:
>>>
>>> if you look at the AbstractWorkingMemory insert
>>> method you'll see one argument is whether it's a
>>> logical insertion or not. You'll also see it check
>>> the global maintainTMS configuration and also
>>> retrieve the ObjectTypeConf. So between those things
>>> someone should be able to get it working.
>>>
>>>
>>> Today, it enters a block where it operates on the
>>> equality map and also creates a default handle based on
>>> that TMS global option and *regardless* of whether it's
>>> a logical insert.
>>>
>>> If I'm *not* putting things in the equality map for
>>> regular inserts, when a logical insert comes in, but
>>> there were already stated inserts, how will I know that?
>>> I would create a new handle for the logical insert and
>>> do the tms.addLogicalDependency(...), even tough there
>>> were regular inserts before and this seems to be a wrong
>>> behaviour.
>>>
>>> Apparently, today, you can disable TMS and still use
>>> logical inserts in your drl, what, I believe, will lead
>>> to inconsistent behaviour, but you're at your own risk.
>>>
>>> Ideas? I feel that I missed something :)
>>>
>>> Cheers,
>>> Leo.
>>>
>>> P.S.: I reached the conclusions above based on the fact
>>> that LogicalAssertionTest started failing after I did
>>> the changes you suggested.
>>>
>>>
>>>
>>>
>>> On Wed, Sep 15, 2010 at 3:47 PM, Mark Proctor
>>> <mproctor at codehaus.org <mailto:mproctor at codehaus.org>>
>>> wrote:
>>>
>>> On 15/09/2010 14:35, Michael Anstis wrote:
>>>> Is this in drools-core; or drools-compiler?
>>>>
>>>> Whilst not undertaking to do the work; have a
>>>> purpose to nose through the code makes
>>>> understanding easier.
>>> It's all in DroolsCore.
>>>
>>> It's a 5 minute hack for me and then 15 minute unit
>>> writing test. But I thought I'd write it up in a
>>> hope to bring someone else into the fold, we need
>>> more help writting the core engine someone else out
>>> there must want to work on current edge engine design :)
>>>
>>> if you look at the AbstractWorkingMemory insert
>>> method you'll see one argument is whether it's a
>>> logical insertion or not. You'll also see it check
>>> the global maintainTMS configuration and also
>>> retrieve the ObjectTypeConf. So between those things
>>> someone should be able to get it working.
>>>
>>> Mark
>>>
>>>>
>>>> On 14 September 2010 16:47, Mark Proctor
>>>> <mproctor at codehaus.org
>>>> <mailto:mproctor at codehaus.org>> wrote:
>>>>
>>>> Here is another project proposal, this time
>>>> simpler. I think this one has Wolfgang's name
>>>> on it ;)
>>>>
>>>> http://blog.athico.com/2010/09/lazily-enabled-truth-maintenace.html
>>>>
>>>> Three weeks ago I posted the project idea for
>>>> "Left and Right Unlinking"
>>>> <http://blog.athico.com/2010/08/left-and-right-unlinking-community.html>.
>>>> So far there are no takers, so if you are
>>>> interested let me know :)
>>>>
>>>> In the meantime I tried to think of a simpler
>>>> enhancement that we would like to see done.
>>>>
>>>> At the moment Drools has a user setting
>>>> "MaintainTMSOption" which can be true or false.
>>>> It's a small optimisation that when turned off
>>>> avoids using the equality hashmap that is
>>>> maintained for all inserted objects.
>>>>
>>>> It would be a much better idea to remove this
>>>> configuration setting, thus simplifying things
>>>> for end users and have TMS lazily enabled on
>>>> demand.
>>>>
>>>> For each object type there is an
>>>> "ObjectTypeConf" configuration object that is
>>>> retrieved every time a working memory action,
>>>> such as insert, is executed. The enabledTMS
>>>> boolean should be moved there, so there is one
>>>> per object type, by default it is false.
>>>>
>>>> When a working memory action occurs, like
>>>> insert, it retrieved the ObjectTypeConf and
>>>> checks the maintainTms boolean there, instead
>>>> of the current engine scoped configuration.
>>>> When a logical insertion occurs and the
>>>> ObjectTypeConf is retrieved if maintainTms is
>>>> false it sets the value to true and then
>>>> iterates the associated ObjectTypeNode memory
>>>> lazily adding all the objects to the TMS
>>>> equality map. From then on for that ObjectType
>>>> all inserted objects are added to that equality
>>>> map.
>>>>
>>>> With this you now have the advantage of TMS
>>>> being laziy enabled, so the minor hashmap
>>>> operation is no longer used and likewise a
>>>> small memory saving from not populating the
>>>> map. There is a further advantage that this is
>>>> now fine grained and when enabled only impacts
>>>> for that specific object type.
>>>>
>>>> A further enhancement could use a int counter,
>>>> instead of a boolean. Each logical insertion
>>>> for that object type increases the counter,
>>>> each retraction decreases the counter; even if
>>>> automatically retracted if the truth is broken
>>>> for that logical assertion. When the counter
>>>> reaches zero, TMS for that OTN can be disabled.
>>>> We do not however remove the objects from the
>>>> equality map, as this would cause "churn" if
>>>> TMS is continuously enabled and disabled.
>>>> Instead when TMS is disabled record the current
>>>> fact counter id. Then if TMS is disabled on a
>>>> retraction but there is a counter id, we can
>>>> check that counter id to see if the fact is
>>>> prior to TMS being disabled and thus would need
>>>> to be retracted from the equality map.
>>>>
>>>> _______________________________________________
>>>> rules-dev mailing list
>>>> rules-dev at lists.jboss.org
>>>> <mailto:rules-dev at lists.jboss.org>
>>>> https://lists.jboss.org/mailman/listinfo/rules-dev
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> rules-dev mailing list
>>>> rules-dev at lists.jboss.org <mailto:rules-dev at lists.jboss.org>
>>>> https://lists.jboss.org/mailman/listinfo/rules-dev
>>>
>>>
>>> _______________________________________________
>>> rules-dev mailing list
>>> rules-dev at lists.jboss.org
>>> <mailto:rules-dev at lists.jboss.org>
>>> https://lists.jboss.org/mailman/listinfo/rules-dev
>>>
>>>
>>>
>>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/rules-dev/attachments/20100924/42505636/attachment-0001.html
More information about the rules-dev
mailing list