[BUG] explode leaks a host memory

[revans2]

Describe the bug
Spark has noticed host memory leaks when running large numbers of tests over a long period of time. I have been tracking them down and found that explode is leaking 32 bytes of memory in some cases when it is called.

Steps/Code to reproduce bug
I wrote a custom memory leak detector because java and valgrind are not happy together. valgrind does not really like CUDA all that much either, but if there are better tools I am happy to use them.

valgrind --leak-check=full ./gtests/LISTS_TEST ./gtests/LISTS_TEST --gtest_filter='ExplodeTest.Empty*'

Then look for the definitely leaked part (ignoring all of the warnings about IOCTLs and memory mappings...

==130808== 32 bytes in 1 blocks are definitely lost in loss record 227 of 1,274
==130808==    at 0x483BE63: operator new(unsigned long) (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==130808==    by 0x59B3108: std::unique_ptr<cudf::table, std::default_delete<cudf::table> > cudf::detail::gather<int const*>(cudf::table_view const&, int const*, int const*, cudf::out_of_bounds_policy, rmm::cuda_stream_view, rmm::mr::device_memory_resource*) (in .../libcudf.so)
==130808==    by 0x604D444: cudf::detail::(anonymous namespace)::build_table(cudf::table_view const&, int, cudf::column_view const&, cudf::device_span<int const, 18446744073709551615ul>, thrust::optional<cudf::device_span<int const, 18446744073709551615ul> >, thrust::optional<rmm::device_uvector<int> >, rmm::cuda_stream_view, rmm::mr::device_memory_resource*) (in .../libcudf.so)
==130808==    by 0x604DD4D: cudf::detail::explode(cudf::table_view const&, int, rmm::cuda_stream_view, rmm::mr::device_memory_resource*) (in .../libcudf.so)
==130808==    by 0x604DFA9: cudf::explode(cudf::table_view const&, int, rmm::mr::device_memory_resource*) (in .../libcudf.so)
==130808==    by 0x555F1D: ExplodeTest_Empty_Test::TestBody() (in .../gtests/LISTS_TEST)
==130808==    by 0x13B67150: void testing::internal::HandleExceptionsInMethodIfSupported<testing::Test, void>(testing::Test*, void (testing::Test::*)(), char const*) (in .../lib/libgtest.so)
==130808==    by 0x13B5B2C5: testing::Test::Run() (in .../lib/libgtest.so)
==130808==    by 0x13B5B424: testing::TestInfo::Run() (in .../lib/libgtest.so)
==130808==    by 0x13B5B54C: testing::TestSuite::Run() (in .../lib/libgtest.so)
==130808==    by 0x13B5BB02: testing::internal::UnitTestImpl::RunAllTests() (in .../lib/libgtest.so)
==130808==    by 0x13B5BD67: testing::UnitTest::Run() (in .../lib/libgtest.so)
==130808== 
==130808== 32 bytes in 1 blocks are definitely lost in loss record 228 of 1,274
==130808==    at 0x483BE63: operator new(unsigned long) (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==130808==    by 0x59B3108: std::unique_ptr<cudf::table, std::default_delete<cudf::table> > cudf::detail::gather<int const*>(cudf::table_view const&, int const*, int const*, cudf::out_of_bounds_policy, rmm::cuda_stream_view, rmm::mr::device_memory_resource*) (in .../libcudf.so)
==130808==    by 0x604D444: cudf::detail::(anonymous namespace)::build_table(cudf::table_view const&, int, cudf::column_view const&, cudf::device_span<int const, 18446744073709551615ul>, thrust::optional<cudf::device_span<int const, 18446744073709551615ul> >, thrust::optional<rmm::device_uvector<int> >, rmm::cuda_stream_view, rmm::mr::device_memory_resource*) (in .../libcudf.so)
==130808==    by 0x604E495: cudf::detail::explode_position(cudf::table_view const&, int, rmm::cuda_stream_view, rmm::mr::device_memory_resource*) (in .../libcudf.so)
==130808==    by 0x604E7B9: cudf::explode_position(cudf::table_view const&, int, rmm::mr::device_memory_resource*) (in .../libcudf.so)
==130808==    by 0x55603D: ExplodeTest_Empty_Test::TestBody() (in .../gtests/LISTS_TEST)
==130808==    by 0x13B67150: void testing::internal::HandleExceptionsInMethodIfSupported<testing::Test, void>(testing::Test*, void (testing::Test::*)(), char const*) (in .../lib/libgtest.so)
==130808==    by 0x13B5B2C5: testing::Test::Run() (in .../lib/libgtest.so)
==130808==    by 0x13B5B424: testing::TestInfo::Run() (in .../lib/libgtest.so)
==130808==    by 0x13B5B54C: testing::TestSuite::Run() (in .../lib/libgtest.so)
==130808==    by 0x13B5BB02: testing::internal::UnitTestImpl::RunAllTests() (in .../lib/libgtest.so)
==130808==    by 0x13B5BD67: testing::UnitTest::Run() (in .../lib/libgtest.so)

I have not seen leaks on join or other operations that use gather, so I suspect it is related explode itself.

Expected behavior
no leaks

[revans2]

As a side note I think I have tracked it down to the release being done.

cudf/cpp/src/lists/explode.cu

Line 68 in 8af4e84

std::vector<std::unique_ptr<column>> columns = gathered_table.release()->release();

When I comment out lines 70 to 99 in the file I still see leaks. When I then just return gathered_table there are no leaks.

[revans2]

Okay I figured it out. we are leaking a pointer to the table. gather_table.release() is releasing the pointer that is never freed. I'll put up a PR with a fix shortly.