Upgrade wheels to use arrow 13 (#14339)

In #14330 we upgraded to Arrow 13. However, we only did so for conda packages. For wheels, we couldn't do the same because pyarrow 13 started supporting two manylinux versions, both 2.17 and 2.28. This results in ABI compatibility issues because cudf wheels bundle an identical libarrow and was previously configured to compile with the CXX11 ABI turned off for compatibility with the  libarrow in the 2.17 wheels. To address this, this PR modifies the CMake logic to only set that flag conditionally based on the glibc version on the host system.

Authors:
  - Vyas Ramasubramani (https://github.com/vyasr)

Approvers:
  - GALI PREM SAGAR (https://github.com/galipremsagar)
  - Robert Maynard (https://github.com/robertmaynard)
  - Ray Douglass (https://github.com/raydouglass)
  - Bradley Dice (https://github.com/bdice)

URL: #14339

.github/workflows/pr.yaml

@@ -101,6 +101,7 @@ jobs:
     uses: rapidsai/shared-workflows/.github/workflows/[email protected]
     with:
       build_type: pull-request
+      build-2_28-wheels: "true"
       script: "ci/build_wheel_cudf.sh"
   wheel-tests-cudf:
     needs: wheel-build-cudf

ci/build_wheel_cudf.sh

@@ -9,8 +9,7 @@ export SKBUILD_CONFIGURE_OPTIONS="-DCUDF_BUILD_WHEELS=ON -DDETECT_CONDA_ENV=OFF"
 
 ./ci/build_wheel.sh cudf ${package_dir}
 
-mkdir -p ${package_dir}/final_dist
 python -m auditwheel repair -w ${package_dir}/final_dist ${package_dir}/dist/*
 
 RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
-RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-upload-wheels-to-s3 ${package_dir}/final_dist
+RAPIDS_PY_WHEEL_NAME="cudf_${AUDITWHEEL_POLICY}_${RAPIDS_PY_CUDA_SUFFIX}" rapids-upload-wheels-to-s3 ${package_dir}/final_dist

ci/test_wheel_cudf.sh

@@ -3,8 +3,21 @@
 
 set -eou pipefail
 
+# Set the manylinux version used for downloading the wheels so that we test the
+# newer ABI wheels on the newer images that support their installation.
+# Need to disable pipefail for the head not to fail, see
+# https://stackoverflow.com/questions/19120263/why-exit-code-141-with-grep-q
+set +o pipefail
+glibc_minor_version=$(ldd --version | head -1 | grep -o "[0-9]\.[0-9]\+" | tail -1 | cut -d '.' -f2)
+set -o pipefail
+manylinux_version="2_17"
+if [[ ${glibc_minor_version} -ge 28 ]]; then
+    manylinux_version="2_28"
+fi
+manylinux="manylinux_${manylinux_version}"
+
 RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
-RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./dist
+RAPIDS_PY_WHEEL_NAME="cudf_${manylinux}_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./dist
 
 # echo to expand wildcard before adding `[extra]` requires for pip
 python -m pip install $(echo ./dist/cudf*.whl)[test]

ci/test_wheel_dask_cudf.sh

@@ -7,7 +7,20 @@ RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
 RAPIDS_PY_WHEEL_NAME="dask_cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./dist
 
 # Download the cudf built in the previous step
-RAPIDS_PY_WHEEL_NAME="cudf_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./local-cudf-dep
+# Set the manylinux version used for downloading the wheels so that we test the
+# newer ABI wheels on the newer images that support their installation.
+# Need to disable pipefail for the head not to fail, see
+# https://stackoverflow.com/questions/19120263/why-exit-code-141-with-grep-q
+set +o pipefail
+glibc_minor_version=$(ldd --version | head -1 | grep -o "[0-9]\.[0-9]\+" | tail -1 | cut -d '.' -f2)
+set -o pipefail
+manylinux_version="2_17"
+if [[ ${glibc_minor_version} -ge 28 ]]; then
+    manylinux_version="2_28"
+fi
+manylinux="manylinux_${manylinux_version}"
+
+RAPIDS_PY_WHEEL_NAME="cudf_${manylinux}_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-s3 ./local-cudf-dep
 python -m pip install --no-deps ./local-cudf-dep/cudf*.whl
 
 # Always install latest dask for testing

cpp/cmake/thirdparty/get_arrow.cmake

@@ -53,19 +53,35 @@ function(find_libarrow_in_python_wheel PYARROW_VERSION)
   find_package(Arrow ${PYARROW_VERSION} MODULE REQUIRED GLOBAL)
   add_library(arrow_shared ALIAS Arrow::Arrow)
 
-  # When using the libarrow inside a wheel we must build libcudf with the old ABI because pyarrow's
-  # `libarrow.so` is compiled for manylinux2014 (centos7 toolchain) which uses the old ABI. Note
-  # that these flags will often be redundant because we build wheels in manylinux containers that
-  # actually have the old libc++ anyway, but setting them explicitly ensures correct and consistent
-  # behavior in all other cases such as aarch builds on newer manylinux or testing builds in newer
-  # containers. Note that tests will not build successfully without also propagating these options
-  # to builds of GTest. Similarly, benchmarks will not work without updating GBench (and possibly
-  # NVBench) builds. We are currently ignoring these limitations since we don't anticipate using
-  # this feature except for building wheels.
-  target_compile_options(
-    Arrow::Arrow INTERFACE "$<$<COMPILE_LANGUAGE:CXX>:-D_GLIBCXX_USE_CXX11_ABI=0>"
-                           "$<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-D_GLIBCXX_USE_CXX11_ABI=0>"
+  # When using the libarrow inside a wheel, whether or not libcudf may be built using the new C++11
+  # ABI is dependent on whether the libarrow inside the wheel was compiled using that ABI because we
+  # need the arrow library that we bundle in cudf to be ABI-compatible with the one inside pyarrow.
+  # We determine what options to use by checking the glibc version on the current system, which is
+  # also how pip determines which manylinux-versioned pyarrow wheel to install. Note that tests will
+  # not build successfully without also propagating these options to builds of GTest. Similarly,
+  # benchmarks will not work without updating GBench (and possibly NVBench) builds. We are currently
+  # ignoring these limitations since we don't anticipate using this feature except for building
+  # wheels.
+  EXECUTE_PROCESS(
+    COMMAND ${CMAKE_C_COMPILER} -print-file-name=libc.so.6
+    OUTPUT_VARIABLE GLIBC_EXECUTABLE
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+  )
+  EXECUTE_PROCESS(
+    COMMAND ${GLIBC_EXECUTABLE}
+    OUTPUT_VARIABLE GLIBC_OUTPUT
+    OUTPUT_STRIP_TRAILING_WHITESPACE
   )
+  STRING(REGEX MATCH "stable release version ([0-9]+\\.[0-9]+)" GLIBC_VERSION ${GLIBC_OUTPUT})
+  STRING(REPLACE "stable release version " "" GLIBC_VERSION ${GLIBC_VERSION})
+  STRING(REPLACE "." ";" GLIBC_VERSION_LIST ${GLIBC_VERSION})
+  LIST(GET GLIBC_VERSION_LIST 1 GLIBC_VERSION_MINOR)
+  if(GLIBC_VERSION_MINOR LESS 28)
+    target_compile_options(
+      Arrow::Arrow INTERFACE "$<$<COMPILE_LANGUAGE:CXX>:-D_GLIBCXX_USE_CXX11_ABI=0>"
+                             "$<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-D_GLIBCXX_USE_CXX11_ABI=0>"
+    )
+  endif()
 
   rapids_export_package(BUILD Arrow cudf-exports)
   rapids_export_package(INSTALL Arrow cudf-exports)
@@ -408,22 +424,12 @@ function(find_and_configure_arrow VERSION BUILD_STATIC ENABLE_S3 ENABLE_ORC ENAB
 endfunction()
 
 if(NOT DEFINED CUDF_VERSION_Arrow)
-  # Temporarily use Arrow 12.0.1 in wheels and Arrow 13.0.0 otherwise
-  if(USE_LIBARROW_FROM_PYARROW)
-    set(CUDF_VERSION_Arrow
-        # This version must be kept in sync with the libarrow version pinned for builds in
-        # dependencies.yaml.
-        12.0.1
-        CACHE STRING "The version of Arrow to find (or build)"
-    )
-  else()
-    set(CUDF_VERSION_Arrow
-        # This version must be kept in sync with the libarrow version pinned for builds in
-        # dependencies.yaml.
-        13.0.0
-        CACHE STRING "The version of Arrow to find (or build)"
-    )
-  endif()
+  set(CUDF_VERSION_Arrow
+      # This version must be kept in sync with the libarrow version pinned for builds in
+      # dependencies.yaml.
+      13.0.0
+      CACHE STRING "The version of Arrow to find (or build)"
+  )
 endif()
 
 find_and_configure_arrow(

dependencies.yaml

@@ -242,16 +242,11 @@ dependencies:
           - cython>=3.0.0
           # TODO: Pin to numpy<1.25 until cudf requires pandas 2
           - &numpy numpy>=1.21,<1.25
-      - output_types: [conda]
+      - output_types: [conda, requirements, pyproject]
         packages:
           # Hard pin the patch version used during the build. This must be kept
           # in sync with the version pinned in get_arrow.cmake.
           - pyarrow==13.0.0.*
-      - output_types: [requirements, pyproject]
-        packages:
-          # Hard pin the patch version used during the build. This must be kept
-          # in sync with the version pinned in get_arrow.cmake.
-          - pyarrow==12.0.1.*
   build_python:
     common:
       - output_types: [conda, requirements, pyproject]
@@ -272,14 +267,10 @@ dependencies:
           - libarrow==13.*
   pyarrow_run:
     common:
-      - output_types: [conda]
+      - output_types: [conda, requirements, pyproject]
         packages:
           # Allow runtime version to float up to minor version
           - pyarrow==13.*
-      - output_types: [requirements, pyproject]
-        packages:
-          # Allow runtime version to float up to minor version
-          - pyarrow==12.*
   cudatoolkit:
     specific:
       - output_types: conda

python/cudf/pyproject.toml

@@ -8,7 +8,7 @@ requires = [
     "ninja",
     "numpy>=1.21,<1.25",
     "protoc-wheel",
-    "pyarrow==12.0.1.*",
+    "pyarrow==13.0.0.*",
     "rmm==23.12.*",
     "scikit-build>=0.13.1",
     "setuptools",
@@ -38,7 +38,7 @@ dependencies = [
     "pandas>=1.3,<1.6.0dev0",
     "protobuf>=4.21,<5",
     "ptxcompiler",
-    "pyarrow==12.*",
+    "pyarrow==13.*",
     "rmm==23.12.*",
     "typing_extensions>=4.0.0",
 ] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`.

python/cudf_kafka/pyproject.toml

@@ -5,7 +5,7 @@
 requires = [
     "cython>=3.0.0",
     "numpy>=1.21,<1.25",
-    "pyarrow==12.0.1.*",
+    "pyarrow==13.0.0.*",
     "setuptools",
     "wheel",
 ] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`.