RDFS inference in Fuseki w/ TDB

[Dave Kolas <dk...@bbn.com>]

Hi there,

I'm currently using Fuseki to host a number of RDF datasets.  If one 
wants basic RDFS inference around a graph with good query performance, 
what is the suggested configuration?

I'm currently using a:

<http://jena.hpl.hp.com/2003/RDFSExptRuleReasoner>

wrapped around a TDB graph and dataset.  However, this seems to (a) 
dramatically increase the memory requirements and (b) destroy the query 
optimizer.  Is there a different way to configure things to get RDFS 
inference and better performance?  I don't actually have the need to 
track / eliminate inferred statements, so it would be ok to just forward 
chain on insertion into the same store.

Thanks,
--Dave--

Re: RDFS inference in Fuseki w/ TDB

[Dave Kolas <dk...@bbn.com>]

Andy,

Thanks for the reply.

By forward chain the data when loaded, I assume you mean manually doing 
so on the client side before inserting?

I think the use case follows the general parameters you outline below.  
The primary things really are just subClass and subProperty; I believe 
those are the only elements that we're depending on now. The schema is 
small, on the order of 50 classes.  I will check out the linked project.

Thanks!
--Dave--

On 3/6/2016 11:28 AM, Andy Seaborne wrote:
> On 01/03/16 19:20, Dave Kolas wrote:
>> Hi there,
>>
>> I'm currently using Fuseki to host a number of RDF datasets.  If one
>> wants basic RDFS inference around a graph with good query performance,
>> what is the suggested configuration?
>>
>> I'm currently using a:
>>
>> <http://jena.hpl.hp.com/2003/RDFSExptRuleReasoner>
>>
>> wrapped around a TDB graph and dataset.  However, this seems to (a)
>> dramatically increase the memory requirements and (b) destroy the query
>> optimizer.  Is there a different way to configure things to get RDFS
>> inference and better performance?  I don't actually have the need to
>> track / eliminate inferred statements, so it would be ok to just forward
>> chain on insertion into the same store.
>>
>> Thanks,
>> --Dave--
>>
>>
>
> Hi Dave,
>
> In the current codebase the only two options are using there in-memory 
> reasoner or forward-chaining the data when it is is loaded.  A 
> combination, whereby some inferences are done ahead of time and others 
> done via the in-memory reasoner at runtime is possible - I don't know 
> how much memory that saves though and it still might do bad things to 
> optimization.
>
>
> Adding some support for RDFS to TDB would make a lot of sense.
> I had a go with [1].
>
> To scope that, could you give some background on your RDFS needs here, 
> which as which parts of RDFS you use (and RDFS++ - all the single 
> property inferences such as symmetric properties) and whether the 
> assumption below fit your usage:
>
> 1/ It expands the subclass and subproperty chains once at startup so 
> there is no path following at query time,  just one hope for rdf:type.
>
> ?x rdf:type :C
> ==>
> ?x rdf:type ?X. ?X rdfs:"transitiveSubClassOf" :C .
>
> 2/ The schema is small - the workspace is in memory.
>
> 3/ The schema is fixed - any change would need to rebuild the sub* 
> data structures but also ...
>
> 4/ schema and data are separate (this is not a core assumption; I 
> wanted) to be be sure that worked well.
>
> 5/ No "messing with the furniture" i.e sub-properties of 
> rdfs:subClassOf, rdfs:subPropertyOf, rdf:type don't trigger inference. 
> I don't think it is fundamental in the design.
>
>     Andy
>
>
> [1]
> https://github.com/afs/jena-inf-engine
>
>

Re: RDFS inference in Fuseki w/ TDB

[Andy Seaborne <an...@apache.org>]

On 01/03/16 19:20, Dave Kolas wrote:
> Hi there,
>
> I'm currently using Fuseki to host a number of RDF datasets.  If one
> wants basic RDFS inference around a graph with good query performance,
> what is the suggested configuration?
>
> I'm currently using a:
>
> <http://jena.hpl.hp.com/2003/RDFSExptRuleReasoner>
>
> wrapped around a TDB graph and dataset.  However, this seems to (a)
> dramatically increase the memory requirements and (b) destroy the query
> optimizer.  Is there a different way to configure things to get RDFS
> inference and better performance?  I don't actually have the need to
> track / eliminate inferred statements, so it would be ok to just forward
> chain on insertion into the same store.
>
> Thanks,
> --Dave--
>
>

Hi Dave,

In the current codebase the only two options are using there in-memory 
reasoner or forward-chaining the data when it is is loaded.  A 
combination, whereby some inferences are done ahead of time and others 
done via the in-memory reasoner at runtime is possible - I don't know 
how much memory that saves though and it still might do bad things to 
optimization.


Adding some support for RDFS to TDB would make a lot of sense.
I had a go with [1].

To scope that, could you give some background on your RDFS needs here, 
which as which parts of RDFS you use (and RDFS++ - all the single 
property inferences such as symmetric properties) and whether the 
assumption below fit your usage:

1/ It expands the subclass and subproperty chains once at startup so 
there is no path following at query time,  just one hope for rdf:type.

?x rdf:type :C
==>
?x rdf:type ?X. ?X rdfs:"transitiveSubClassOf" :C .

2/ The schema is small - the workspace is in memory.

3/ The schema is fixed - any change would need to rebuild the sub* data 
structures but also ...

4/ schema and data are separate (this is not a core assumption; I 
wanted) to be be sure that worked well.

5/ No "messing with the furniture" i.e sub-properties of 
rdfs:subClassOf, rdfs:subPropertyOf, rdf:type don't trigger inference. I 
don't think it is fundamental in the design.

     Andy


[1]
https://github.com/afs/jena-inf-engine