shapes-ISSUE-79 (Validation functions): Cleaner separation between value checking and property iteration [SHACL Spec]

shapes-ISSUE-79 (Validation functions): Cleaner separation between value checking and property iteration [SHACL Spec]

Raised by: Holger Knublauch
On product: SHACL Spec

I was never quite happy with one aspect of how SHACL templates (including the Core templates) were internally defined. To illustrate the problem, this is how sh:datatype is currently defined:

 a sh:ConstraintTemplate ;
 rdfs:subClassOf sh:AbstractPropertyConstraint ;
 sh:message "Values must have datatype {?datatype}" ;
 sh:sparql """
  SELECT ?this (?this AS ?subject) ?predicate ?object ?datatype
   ?this ?predicate ?object .
   FILTER (!sh:hasDatatype(?object, ?datatype)) .
  """ ;

The SPARQL query above does two things:
a) It iterates over all property values
b) It FILTERs each of these values.

This is a recurring pattern for many property constraints. I would like us to make this design pattern more explicit so that it becomes the following:

 a sh:PropertyValueConstraintTemplate ;
 rdfs:subClassOf sh:AbstractPropertyConstraint ;
 sh:message "Values must have datatype {?datatype}" ;
 sh:validationFunction sh:hasDatatype ;

Instead of pointing at a SPARQL query that does everything, such constraints only point at a Function, which must take a value as input and return a boolean. The engine can produce the surrounding SPARQL automatically, and can even directly inject the body of the sh:hasDatatype function.

This design has the following advantages:
- It makes the contract more explicit, modular and arguably cleaner
- It allows to focus on what really matters, reducing clutter
- It makes it easier to reuse the logic, especially for inverse property constraints (and sh:Arguments)
- It makes it easier to optimize execution - if only a boolean result is needed, then a code generator can more easily combine them into a single FILTER such as sh:hasDatatype(...) && sh:hasNodeKind(...)
- It lowers the implementation costs for other languages like JavaScript - these can focus on implementing the functions
- It raises the abstraction, e.g. in JavaScript these checks can be simple variable comparisons, regardless of how the surrounding iteration happened
- The key snippets are also reusable inside of SPARQL expressions, because they are also SPARQL functions.

The only disadvantage that I can think of is that there is a little bit more work for the engine implementers, because this requires a couple of new classes and properties to work correctly. However, I'd rather push the complexity to the engine developers and have a cleaner overall design for the end users. And, nobody is forced to use this new pattern - anyone can still use the current mechanism using sh:sparql.

I have implemented this on a test branch, and a new Turtle file can be found here:

I have not yet updated the textual documents - I'd love to hear from others on the general direction before I spend more time on this.

Received on Thursday, 13 August 2015 02:29:16 UTC