Re: Solicitation of feedback

On Tue, 2010-07-20 at 17:10 +0100, Dave Reynolds wrote:
> [For some reason I couldn't read the original forwarded email so didn't
> see this question until Sandro's reply.]
> On Tue, 2010-07-20 at 08:54 -0400, Sandro Hawke wrote: 
> > On Mon, 2010-07-19 at 10:26 -0400, Chimezie Ogbuji wrote:
> > > 7.2 (Simple) RIF Core Entailment Regime
> > > 
> > > "It is unclear whether safe RIF-Core rules used to form combinations for
> > > this entailment regime guarantee uniqueness (up to RDF graph equivalence) on
> > > answer sets [...] Without strongly safe restrictions, there may be
> > > interoperability issues ... However, strong safety restrictions are only
> > > defined in the informative sections of the RIF-Core specification"
> > > 
> > > Can you give any background into why the strong safety characteristics are
> > > only an informative part of the specification that might help in informing
> > > conditions for preventing trivial infinite answers as appropriate for the
> > > RIF regime if the use of the strong safety criteria is not appropriate for
> > > this?
> > 
> > I'm going to have to let someone else answer this, or take some time to
> > swap this back in/figure this out, sorry.
> If I recall correctly the issue was that people treat rule systems,
> especially production rule systems, like programming languages. They use
> the expressivity of cyclic dependencies while, in practice, ensuring
> appropriate termination conditions. An artificial example being
> something like:
>    p(0) .
>    p(?x + 1) :- p(?x), ?x >= 0, ?x < 10 .
> The strong safety conditions would exclude such rule sets.
> We wanted Core to be a useful subset of both PR and BLD and felt that
> the restriction to strongly safe rules would eliminate too many rule
> sets used in practice (that would otherwise be within Core).

That sounds right, yes.

> I guess you could say that the SPARQL-RIF Core entailment regime is only
> defined over strongly safe rule sets and that interoperation is not
> guaranteed for other rule sets. 
> Or could you  say that interoperability is only guaranteed over rule
> sets which terminate (on the given proof engine) and that strong safety
> is one way to ensure that?

Yes, exactly.   Strong safety is one way to get certain guarantees, but
it's very limiting.   I have the impression that neither rule system
vendors nor rule systems users have any interest in sticking to
strongly-safe rules, and a lot of them are not interested in sticking to
RIF Core (or even BLD or PRD).   For the SPARQL WG to come in and
restrict them to a subset of what they want to do seems, well, let's
just say "sub-optimal".   

I'm pretty sure the right thing for SPARQL is to leave it to RIF, which
in this case means that systems doing RIF must implement RIF core, and
may implement additional features (eg PRD, or xml-data, or their own
built-ins.).   Rule authors have to decide what their audience is, and
pick the appropriate dialect; if they want the maximal audience, they
should stuck to RIF Core.

    - Sandro

Received on Tuesday, 20 July 2010 20:46:01 UTC