[JENA-1813] Join optimization transform results in incorrect query results

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: Jena 3.13.1
• Fix Version/s: Jena 3.14.0
• Component/s: ARQ
• Assignee: Andy Seaborne
• Reporter: Shawn Smith
• Created: 09/Jan/20 03:07
• Updated: 03/Jun/22 20:23
• Resolved: 13/Jan/20 12:29

Description

I think I've found a query where TransformJoinStrategy incorrectly decides that a query is linear such that a "join" operation can be replaced by a "sequence" operation.  As a result, the query returns incorrect results.  Disabling optimizations with "qe.getContext().set(ARQ.optimization, false)" fixes the issue.

Here's the query: 

PREFIX  :  <http://example.com/>
SELECT ?a
WHERE {
  GRAPH :graph { :s :p ?a }
  GRAPH :graph {
    SELECT (?b AS ?a)
    WHERE { :t :q ?b }
    GROUP BY ?b
  }
}

Here's the data to test it with (two quads, as Trig): 

@prefix :      <http://example.com/> .
:graph {
    :s      :p      "a" .
    :t      :q      "b" .
}

I expected the query to return zero results because the two GRAPH clauses can't find compatible bindings for ?a.  But, in practice, Jena returns ?a="a" and logs a warning:

[main] WARN  BindingUtils - merge: Mismatch : "a" != "b"

Note the warning is actually coming from QueryIterProjectMerge.java, not BindingUtils.java.  With more complicated queries and datasets, this issue can result in thousands or millions of logged warnings.

The query plan before optimization looks like this:

(project (?a)
  (join
    (graph <http://example.com/graph>
      (bgp (triple <http://example.com/s> <http://example.com/p> ?a)))
    (graph <http://example.com/graph>
      (project (?a)
        (extend ((?a ?b))
          (group (?b)
            (bgp (triple <http://example.com/t> <http://example.com/q> ?b))))))))

Optimization replaces "join" with "sequence" which fails to detect conflicts on ?a:

(project (?a)
  (sequence
    (graph <http://example.com/graph>
      (bgp (triple <http://example.com/s> <http://example.com/p> ?a)))
    (graph <http://example.com/graph>
      (project (?a)
        (extend ((?a ?/b))
          (group (?/b)
            (bgp (triple <http://example.com/t> <http://example.com/q> ?/b))))))))

For convenience, here's Java code that reproduces the bug:

import org.apache.jena.query.ARQ;
import org.apache.jena.query.Dataset;
import org.apache.jena.query.DatasetFactory;
import org.apache.jena.query.QueryExecution;
import org.apache.jena.query.QueryExecutionFactory;
import org.apache.jena.query.ResultSet;
import org.apache.jena.riot.Lang;
import org.apache.jena.riot.RDFParser;
import org.junit.Test;

public class QueryTest {

    @Test
    public void testGraphQuery() {
        String query = "" +
                "PREFIX  :  <http://example.com/>\n" +
                "SELECT ?a\n" +
                "WHERE {\n" +
                "  GRAPH :graph { :s :p ?a }\n" +
                "  GRAPH :graph {\n" +
                "    SELECT (?b AS ?a)\n" +
                "    WHERE { :t :q ?b }\n" +
                "    GROUP BY ?b\n" +
                "  }\n" +
                "}\n";
        String data = "" +
                "@prefix :  <http://example.com/> .\n" +
                ":graph {\n" +
                "  :s  :p  \"a\" .\n" +
                "  :t  :q  \"b\" .\n" +
                "}\n";

        Dataset ds = DatasetFactory.create();
        RDFParser.fromString(data).lang(Lang.TRIG).parse(ds);

        try (QueryExecution qe = QueryExecutionFactory.create(query, ds)) {
            qe.getContext().set(ARQ.optimization, true);  // flipping this to false fixes the test
            ResultSet rs = qe.execSelect();
            if (rs.hasNext()) {
                System.out.println(rs.nextBinding());
                throw new AssertionError("Result set should be empty");
            }
        }
    }
}

Issue Links

• relates to: JENA-2332 Bad optimization transform when modifers used in SERVICE
• links to: GitHub Pull Request #670