GH-829: Stratification and evaluation for run-once rules

[afs]

Closes #829

• See this document rendered online here

[robert-david]

This is the correct URI for the githack render:
https://raw.githack.com/w3c/data-shapes/rules-run-once/shacl12-rules/index.html

[simonstey]

how would following corner cases be classified?

• A rule with an assignment element AND a negation element -> .once or .general?
• A rule with blank nodes in the head AND a recursive positive dependency -> .once or .general? If .once, the recursion is lost; if .general, blank nodes generate infinitely.
• A rule with no assignment elements and no blank nodes in head -> presumably .general, but not stated.
• A rule with BNODE() in an assignment expression but not in the head -> .once or .general?

[simonstey]

this def says a "stratum" it is "a pair of sets of rules": one for rules "executed once" (for assignments and blank nodes in head) and one for general rules. However, there is no normative definition anywhere in the spec that specifies how a rule is classified into .once or .general.

it's somewhat hinted at via: "for assignment and blank nodes created in a [=rule head=]" but I think we should add a precise normative statement such as:

A rule belongs to `P.once` if it contains an [=assignment element=] in its body or a blank node in its [=rule head=]; otherwise it belongs to `P.general`.

[afs]

See comment on updating the stratification algorithm.

[simonstey]

now that this paragraph got expanded, shouldn't the strat condition section be expanded accordingly as well?

https://github.com/w3c/data-shapes/pull/865/changes#diff-4cb699e56152bee095a8650906565f110dd7ed444551f33a7b1987a0a92fd214R876-R896:

   <section id="stratification-condition">
          <h4>Stratification Condition</h4>
          <p>
            [=Stratification=] is only defined when the following condition is
            satisfied. If a [=rule set=] does not meet this condition, then this
            specification does not define the outcome of [=rule set=] evaluation.
          </p>
          <dl>
            <dt><dfn>Stratification Condition</dfn></dt>
            <dd>
              The [=stratification condition=] requires that
              there is no [=recursive dependency=] involving a
              <em>negative</em> dependency
              in the [=dependency graph=] for a [=rule set=].
            </dd>
          </dl>
          <p>
            In other words, there is no `NOT` used in any rule that
            transitively depends on itself.
          </p>
        </section>

this still only forbids negative cycles, but says nothing about assignment elements or blank-node-generating rules. There is no condition that requires a rule with assignments or blank nodes in its head to be in a lower stratum than rules it depends on (or that such rules cannot be depending on itsedlf).

bug or feature?

[afs]

now that this paragraph got expanded, shouldn't the strat condition section be expanded accordingly as well?

Yes.

[simonstey]

maybe add:

## Partition each level into once and general
for i = 0 to maxStratum:
    let rules = stratumLevels.get(i)
    let once = { R in rules | R is a run-once rule }
    let general = rules \ once
    stratumLevels.set(i, pair(once, general))
endfor

[afs]

how would following corner cases be classified?

• A rule with an assignment element AND a negation element -> .once or .general?

"once" - that takes precedence.

• A rule with blank nodes in the head AND a recursive positive dependency -> .once or .general? If .once, the recursion is lost; if .general, blank nodes generate infinitely.

Reject.

• A rule with no assignment elements and no blank nodes in head -> presumably .general, but not stated.

Will update the stratification algorithm

• A rule with BNODE() in an assignment expression but not in the head -> .once or .general?

Once.

[recalcitrantsupplant]

one suggested wording for updating the stratification algorithm

[recalcitrantsupplant]

Suggested change

	</dd>
	<dd>
	In addition, every rule in `P.once` depends only on rules in lower
	[=stratification layers=].
	</dd>
	</dl>