[BUG] Segmentation fault in cudf::conditional_inner_join with Zero-Row Input

[aocsa]

Describe the bug

There is a bug in cuDF where the cudf::conditional_inner_join operation triggers a Segmentation fault when one of the input values has num_rows = 0. This error occurs during the execution of the join operation.

Steps/Code to reproduce bug

• Create two cuDF DataFrames, one with zero rows.
• Perform cudf::conditional_inner_join.

--- a/cpp/tests/join/conditional_join_tests.cu
+++ b/cpp/tests/join/conditional_join_tests.cu
@@ -222,7 +222,10 @@ struct ConditionalJoinPairReturnTest : public ConditionalJoinTest<T> {
              std::vector<std::pair<cudf::size_type, cudf::size_type>> expected_outputs)
   {
     auto result_size = this->join_size(left, right, predicate);
-    EXPECT_TRUE(result_size == expected_outputs.size());
+    std::cerr << "result_size: " << result_size << std::endl;
+    std::cerr << "expected_outputs: " << expected_outputs.size() << std::endl;
+    // EXPECT_TRUE(result_size == expected_outputs.size());
     auto result = this->join(left, right, predicate);
     std::vector<std::pair<cudf::size_type, cudf::size_type>> result_pairs;
 };
 
+
+TYPED_TEST(ConditionalInnerJoinTest, TestLeftColumnIsEmpty)
+{
+  this->test({{}}, {{0}}, left_zero_eq_right_zero, {{}});
+};
+
+TYPED_TEST(ConditionalInnerJoinTest, TestRightColumnIsEmpty)
+{
+  this->test({{0}}, {{}}, left_zero_eq_right_zero, {{}});
+};
+

result_size: 0
expected_outputs: 1
Segmentation fault (core dumped)

Note: cudf::conditional_inner_join_size works fine, but cudf::conditional_inner_join ends with a segmentation fault.

Expected behavior

Return an empty DataFrame or a specific error.

Environment details

Method of cuDF install: source code
v24.06.00 branch release

[bdice]

@aocsa Thank you so much for the tests to reproduce this failure. I know that is a lot of work but it helps us tremendously.

I will start investigating this.

[bdice]

@aocsa I did some investigation. First, the tests seem to cover the empty case already:

cudf/cpp/tests/join/conditional_join_tests.cu

Lines 409 to 412 in b9a0b72

	TYPED_TEST(ConditionalInnerJoinTest, TestOneColumnLeftEmpty)
	{
	this->test({{}}, {{3, 4, 5}}, left_zero_eq_right_zero, {});
	};

Also, the test cases you provided seem to have a bug. Note that the expected value in the snippet above ({{}}) initializes to {{0, 0}}. This is a vector of a single pair of indices, rather than an empty vector (which is what we should have for an empty join result). Here's a minimal reproducer: https://godbolt.org/z/ErndbYK4v

If I fix your test cases to use {}, I do not get any test failures or segfaults.

For now, I have identified some improvements (additional missing test cases, etc.) and I am trying to track down how the test case you proposed (with {{}}) got to the point of causing a segfault. I will follow up with anything else I discover and will file a PR for those improvements.

[bdice]

I found a bug that would cause a segfault and have fixed it in #16094. When the right table has zero rows, conditional left anti-joins were returning a vector of indices containing garbage data. This is now corrected.

That PR will be evaluated for inclusion in a 24.06.01 hotfix release with some other fixes -- to be determined.