apacheGH-38183: [CI][Python] Use pipx to install GCS testbench

.github/workflows/cpp.yml

@@ -474,9 +474,7 @@ jobs:
         shell: bash
         run: |
           ci/scripts/install_gcs_testbench.sh default
-          echo "PYTHON_BIN_DIR=$(cygpath --windows $(dirname $(which python3.exe)))" >> $GITHUB_ENV
       - name: Test
         shell: msys2 {0}
         run: |
-          PATH="$(cygpath --unix ${PYTHON_BIN_DIR}):${PATH}"
           ci/scripts/cpp_test.sh "$(pwd)" "$(pwd)/build"

ci/docker/conda-cpp.dockerfile

@@ -42,17 +42,19 @@ RUN mamba install -q -y \
         valgrind && \
     mamba clean --all
 
+# We want to install the GCS testbench using the Conda base environment's Python,
+# because the test environment's Python may later change.
+ENV PIPX_PYTHON=/opt/conda/bin/python3
+COPY ci/scripts/install_gcs_testbench.sh /arrow/ci/scripts
+RUN /arrow/ci/scripts/install_gcs_testbench.sh default
+
 # Ensure npm, node and azurite are on path. npm and node are required to install azurite, which will then need to 
-# be on the path for the tests to run.  
+# be on the path for the tests to run.
 ENV PATH=/opt/conda/envs/arrow/bin:$PATH
 
 COPY ci/scripts/install_azurite.sh /arrow/ci/scripts/
 RUN /arrow/ci/scripts/install_azurite.sh
 
-# We want to install the GCS testbench using the same Python binary that the Conda code will use.
-COPY ci/scripts/install_gcs_testbench.sh /arrow/ci/scripts
-RUN /arrow/ci/scripts/install_gcs_testbench.sh default
-
 COPY ci/scripts/install_sccache.sh /arrow/ci/scripts/
 RUN /arrow/ci/scripts/install_sccache.sh unknown-linux-musl /usr/local/bin
 

ci/docker/conda-python.dockerfile

@@ -32,11 +32,6 @@ RUN mamba install -q -y \
         nomkl && \
     mamba clean --all
 
-# XXX The GCS testbench was already installed in conda-cpp.dockerfile,
-# but we changed the installed Python version above, so we need to reinstall it.
-COPY ci/scripts/install_gcs_testbench.sh /arrow/ci/scripts
-RUN /arrow/ci/scripts/install_gcs_testbench.sh default
-
 ENV ARROW_ACERO=ON \
     ARROW_BUILD_STATIC=OFF \
     ARROW_BUILD_TESTS=OFF \

ci/scripts/install_gcs_testbench.bat

@@ -19,7 +19,9 @@
 
 set GCS_TESTBENCH_VERSION="v0.36.0"
 
+python -m pip install pipx || exit /B 1
+
 @REM Install GCS testbench %GCS_TESTBENCH_VERSION%
-python -m pip install  ^
+pipx install ^
         "https://github.com/googleapis/storage-testbench/archive/%GCS_TESTBENCH_VERSION%.tar.gz" ^
         || exit /B 1

ci/scripts/install_gcs_testbench.sh

@@ -34,19 +34,24 @@ case "$(uname -m)" in
     ;;
 esac
 
-# On newer pythons install into the system will fail, so override that
-export PIP_BREAK_SYSTEM_PACKAGES=1
-
 version=$1
 if [[ "${version}" -eq "default" ]]; then
   version="v0.39.0"
-  # Latests versions of Testbench require newer setuptools
-  python3 -m pip install --upgrade setuptools
 fi
 
+export PIP_BREAK_SYSTEM_PACKAGES=1
+python3 -m pip install pipx
+
 # This script is run with PYTHON undefined in some places,
 # but those only use older pythons.
 if [[ -z "${PYTHON_VERSION}" ]] || [[ "${PYTHON_VERSION}" != "3.13" ]]; then
-  python3 -m pip install \
-    "https://github.com/googleapis/storage-testbench/archive/${version}.tar.gz"
+  pipx_flags=--verbose
+  if [[ $(id -un) == "root" ]]; then
+    # Install globally as /root/.local/bin is typically not in $PATH
+    pipx_flags="${pipx_flags} --global"
+  fi
+  if [[ ! -z "${PIPX_PYTHON}" ]]; then
+    pipx_flags="${pipx_flags} --python ${PIPX_PYTHON}"
+  fi
+  pipx install ${pipx_flags} "https://github.com/googleapis/storage-testbench/archive/${version}.tar.gz"
 fi

cpp/src/arrow/filesystem/gcsfs_test.cc

@@ -95,52 +95,52 @@ class GcsTestbench : public ::testing::Environment {
     if (const auto* env = std::getenv("PYTHON")) {
       names = {env};
     }
-    auto error = std::string(
-        "Could not start GCS emulator."
-        " Used the following list of python interpreter names:");
-    for (const auto& interpreter : names) {
-      auto exe_path = bp::search_path(interpreter);
-      error += " " + interpreter;
-      if (exe_path.empty()) {
-        error += " (exe not found)";
-        continue;
-      }
+    auto error = std::string("Could not start GCS emulator 'storage-testbench'");
 
-      bp::ipstream output;
-      server_process_ = bp::child(exe_path, "-m", "testbench", "--port", port_, group_,
-                                  bp::std_err > output);
+    auto testbench_is_running = [](bp::child& process, bp::ipstream& output) {
       // Wait for message: "* Restarting with"
-      auto testbench_is_running = [&output, this](bp::child& process) {
-        std::string line;
-        std::chrono::time_point<std::chrono::steady_clock> end =
-            std::chrono::steady_clock::now() + std::chrono::seconds(10);
-        while (server_process_.valid() && server_process_.running() &&
-               std::chrono::steady_clock::now() < end) {
-          if (output.peek() && std::getline(output, line)) {
-            std::cerr << line << std::endl;
-            if (line.find("* Restarting with") != std::string::npos) return true;
-          } else {
-            std::this_thread::sleep_for(std::chrono::milliseconds(20));
-          }
+      std::string line;
+      std::chrono::time_point<std::chrono::steady_clock> end =
+          std::chrono::steady_clock::now() + std::chrono::seconds(10);
+      while (process.valid() && process.running() &&
+             std::chrono::steady_clock::now() < end) {
+        if (output.peek() && std::getline(output, line)) {
+          std::cerr << line << std::endl;
+          if (line.find("* Restarting with") != std::string::npos) return true;
+        } else {
+          std::this_thread::sleep_for(std::chrono::milliseconds(20));
         }
-        return false;
-      };
+      }
+      return false;
+    };
 
-      if (testbench_is_running(server_process_)) break;
-      error += " (failed to start)";
-      server_process_.terminate();
-      server_process_.wait();
+    auto exe_path = bp::search_path("storage-testbench");
+    if (!exe_path.empty()) {
+      bp::ipstream output;
+      server_process_ =
+          bp::child(exe_path, "--port", port_, group_, bp::std_err > output);
+      if (!testbench_is_running(server_process_, output)) {
+        error += " (failed to start)";
+        server_process_.terminate();
+        server_process_.wait();
+      }
+    } else {
+      error += " (exe not found)";
+    }
+    if (!server_process_.valid()) {
+      error_ = std::move(error);
     }
-    if (server_process_.valid() && server_process_.valid()) return;
-    error_ = std::move(error);
   }
 
   bool running() { return server_process_.running(); }
 
   ~GcsTestbench() override {
     // Brutal shutdown, kill the full process group because the GCS testbench may launch
     // additional children.
-    group_.terminate();
+    try {
+      group_.terminate();
+    } catch (bp::process_error&) {
+    }
     if (server_process_.valid()) {
       server_process_.wait();
     }

python/pyarrow/tests/conftest.py

@@ -233,17 +233,16 @@ def minio_server_health_check(address):
 def gcs_server():
     port = find_free_port()
     env = os.environ.copy()
-    args = [sys.executable, '-m', 'testbench', '--port', str(port)]
+    exe = 'storage-testbench'
+    args = [exe, '--port', str(port)]
     proc = None
     try:
-        # check first if testbench module is available
-        import testbench  # noqa:F401
         # start server
         proc = subprocess.Popen(args, env=env)
         # Make sure the server is alive.
         if proc.poll() is not None:
             pytest.skip(f"Command {args} did not start server successfully!")
-    except (ModuleNotFoundError, OSError) as e:
+    except OSError as e:
         pytest.skip(f"Command {args} failed to execute: {e}")
     else:
         yield {

r/tests/testthat/test-gcs.R

@@ -116,12 +116,12 @@ test_that("GcsFileSystem$create() can read json_credentials", {
 })
 
 skip_on_cran()
-skip_if_not(system('python -c "import testbench"') == 0, message = "googleapis-storage-testbench is not installed.")
+skip_if_not(system('storage-testbench -h') == 0, message = "googleapis-storage-testbench is not installed.")
 library(dplyr)
 
 testbench_port <- Sys.getenv("TESTBENCH_PORT", "9001")
 
-pid_minio <- sys::exec_background("python", c("-m", "testbench", "--port", testbench_port),
+pid_minio <- sys::exec_background("storage-testbench", c("--port", testbench_port),
   std_out = FALSE,
   std_err = FALSE # TODO: is there a good place to send output?
 )