Output validation using matching in SQL

[szarnyasg]

Will fix #205.

We can use the DuckDB appender to populate the tables.

Current validation scripts are in:

• https://github.com/ldbc/ldbc_graphalytics/tree/master/graphalytics-core/src/main/java/science/atlarge/graphalytics/validation
• https://github.com/ldbc/ldbc_graphalytics/tree/master/graphalytics-core/src/main/java/science/atlarge/graphalytics/validation/rule

A lot of time is spent parsing the results back from CSVs to Java data structures, this could also be improved by using DuckDB's COPY ... FROM 'filename.csv' (DELIMITER ' ', FORMAT csv) clause.

• https://github.com/ldbc/ldbc_graphalytics/blob/master/graphalytics-core/src/main/java/science/atlarge/graphalytics/validation/LongVertexValidator.java#L146
• https://github.com/ldbc/ldbc_graphalytics/blob/master/graphalytics-core/src/main/java/science/atlarge/graphalytics/validation/DoubleVertexValidator.java#L146

Validation tests (that are used to test the validation rules themselves) are in:

• https://github.com/ldbc/ldbc_graphalytics/tree/master/graphalytics-validation/src/main/java/science/atlarge/graphalytics/validation/algorithms.

Populating tables using the DuckDB appender and comparing WCC results

A snippet for using appenders (not sure whether it is useful):

try (DuckDBConnection conn = (DuckDBConnection) DriverManager.getConnection("jdbc:duckdb:")) {
    Statement stmt = conn.createStatement();

    // fill 'expected' table
    stmt.execute("DROP TABLE IF EXISTS expected");
    stmt.execute("CREATE TABLE expected(v bigint not null, x double not null);");
    DuckDBAppender expectedAppender = conn.createAppender("main", "expected");
    for (long vertexId : outputGraph.getVertices()) {
        expectedAppender.beginRow();
        expectedAppender.append(vertexId);
        expectedAppender.append(outputGraph.getVertexValue(vertexId));
        expectedAppender.endRow();
    }
    expectedAppender.close();

    // fill 'actual' table
    stmt.execute("DROP TABLE IF EXISTS actual");
    stmt.execute("CREATE TABLE actual(v bigint not null, x double not null);");
    DuckDBAppender actualAppender = conn.createAppender("main", "actual");
    for (long vertexId : outputGraph.getVertices()) {
        actualAppender.beginRow();
        actualAppender.append(vertexId);
        actualAppender.append(expected result);
        actualAppender.endRow();
    }
    actualAppender.close();

    ResultSet rs = stmt.executeQuery(
            "SELECT e1.v AS v, e1.x AS x, a1.x AS x\n" +
            "FROM expected e1, actual a1\n" +
            "WHERE e1.v = a1.v -- select a node in the expected-actual tables\n" +
            "  AND EXISTS (\n" +
            "    SELECT 1\n" +
            "    FROM expected e2, actual a2\n" +
            "    WHERE e2.v = a2.v   -- another node in expected-actual tables\n" +
            "      AND e1.x = e2.x   -- where the node is in the same equivalence class in the expected table\n" +
            "      AND a1.x != a2.x  -- but not in the actual table\n" +
            "  )\n" +
            ";");
    while (rs.next()) {
        System.out.format("%ld: %ld != %ld %n", rs.getLong(1), rs.getLong(2), rs.getLong(3));
    }
    rs.close();

Handling infinity values

Handling infinity necessitates special care as multiple values should be accepted:

if (low.equals("inf") || low.equals("+inf") || low.equals("infinity") || low.equals("+infinity")) {
    return Double.POSITIVE_INFINITY;
} else if (low.equals("-inf") || low.equals("-infinity")) {
    return Double.NEGATIVE_INFINITY;
}

Validation of completeness

The validation should not only check whether the results are correct, it should also check whether all vertices are included in the result set.