Skip to content

pgtap/mincut/compare.test.sql test data is flawed #2583

Open
@vslashg

Description

@vslashg

Problem
The two functionality tests in pgtap/mincut/compare.test.sql use graphs that are either vacuous or invalid, and the expected return values are not guaranteed.

This test case would benefit from being rewritten to use a graph designed for min-cut testing, but I am raising this issue because changes to Boost can cause it to fail.

To Reproduce
Compiling the test against Boost 1.80 or later (which uses a different implementation of Stoer-Wagner than earlier versions) causes failures.

Test 5

The graph for unit test "5: Mincut of edge 17" uses a graph with two nodes and one edge. Evaluating the inner query of the PREPARE stoerWagner6 statement (the one being passed to pgr_stoerWagner()) gives the following:

 id | source | target | cost | reverse_cost 
----+--------+--------+------+--------------
 17 |     14 |     15 |    1 |            1
(1 row)

Obviously this is the exact same graph as generated by the simpler stoerWagner5 subquery; in both cases the mincut is trivially to cut the only existing edge, so the comparison of stoerWagner5 and stoerWagner6 tests only that pgr_stoerWagner generates the same output given the same input, which was probably not what was intended.

Test 6

The subquery being passed to pgr_stoerWagner7 for the test "6: Compare the mincut of subgraph with actual result" results in the following graph:

 id | source | target | cost | reverse_cost 
----+--------+--------+------+--------------
  1 |      1 |      2 |    1 |            1
  2 |      2 |      3 |   -1 |            1
  3 |      3 |      4 |   -1 |            1
  4 |      2 |      5 |    1 |            1
  5 |      3 |      6 |    1 |           -1
  6 |      7 |      8 |    1 |            1
  7 |      8 |      5 |    1 |            1
  8 |      5 |      6 |    1 |            1
  9 |      6 |      9 |    1 |            1
 10 |      5 |     10 |    1 |            1
 11 |      6 |     11 |    1 |           -1
 12 |     10 |     11 |    1 |           -1
 13 |     11 |     12 |    1 |           -1
 14 |     10 |     13 |    1 |            1
 15 |      9 |     12 |    1 |            1
 16 |      4 |      9 |    1 |            1
(16 rows)

And expects that the mincut result is to cut the single edge with id 1.

There are three flaws with this test:

  1. The Stoer-Wagner algorithm requires all edges have non-negative costs, but edges 2 and 3 violate this.
  2. The Stoer-Wagner algorithm requires an undirected graph, but edges 2, 3, 5, 11, 12, and 13 have different costs in different directions.
  3. There are four distinct edges of equal weight that, each on its own, can be removed to partition the graph: edges 1, 6, 7, and 14. (Notice that nodes 1, 7, and 13 only have a single connection to the graph, and node 8 lies between node 7 and the rest of the graph.)

The test query for 6 asserts that the algorithm will select edge 1, but 6, 7, and 14 are valid (and edge 14 is selected when built against more modern versions of Boost).

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions